Coupled Modelling of Hydrodynamics and Mass Transfer in Membrane Filtration

Coupled modelling of hydrodynamics and mass transfer in
membrane filtration
by
Gregory Peter Keir

BE / BSc (Hons I)
Submitted in fulfilment of the requirements for the degree of
Doctor of Philosophy
Deakin University
August, 2012
DEAKIN UNIVERSITY
ACCESS TO THESIS A
I am the author of the thesis entitled Coupled modelling of hydrodynamics and mass transfer in
membrane filtration submitted for the degree of Doctor of Philosophy.
This thesis may be made available for consultation, loan and limited copying in accordance with the
Copyright Act 1968.
'I certify that I am the student named below and that the information provided in the form is correct'
Full Name Gregory Peter Keir
Signed
Date 04/08/2012
DEAKIN UNIVERSITY
CANDIDATE DECLARATION
I certify that the thesis entitled Coupled modelling of hydrodynamics and mass
transfer in membrane filtration submitted for the degree of Doctor of
Philosophy is the result of my own work and that where reference is made to the
work of others, due acknowledgment is given.
I also certify that any material in the thesis which has been accepted for a degree
or diploma by any university or institution is identified in the text.
'I certify that I am the student named below and that the information provided in the form is correct'
Full Name Gregory Peter Keir
Signed
Date 04/08/2012
Abstract
The high-pressure membrane filtration processes of nanofiltration (NF) and
reverse osmosis (RO) are widely used in industry to remove dissolved
components from liquid streams, especially in the fields of water and wastewater
treatment. Despite their ubiquity, it is often difficult to predict the performance of
a membrane filtration system without expensive pilot studies. This is because the
flow of fluid through the membrane and the removal of dissolved components
depend both on the hydrodynamic conditions within the bulk of the membrane
module and the interactions between the dissolved components and the
microstructure of the membrane, which can be complex. This can be especially
apparent when multiple dissolved components are in solution.
This thesis outlines the development of a computational model to predict the
behaviour of such systems by combining a rigorous hydrodynamic model using
commercially available computational fluid dynamics (CFD) software with a
microscopic mass transfer model including the effects of convection, diffusion,
and electric migration to predict the removal of dissolved components by the
membrane. These fields of modelling in membrane applications have traditionally
been disparate; combining the two allows for a rigorous and fairly widely
applicable general predictive model of the NF and RO processes.
The CFD package ANSYS CFX is used for all simulations, allowing construction
of hydrodynamic models for arbitrary two-dimensional (2D) or three-dimensional
i
(3D) membrane channel geometries, and laminar or turbulent flow conditions.
The hydrodynamic model used here allows for the properties of the solution, such
as density, viscosity, solute diffusivity, and osmotic pressure to vary spatially
within the membrane channel depending on the local solute concentrations. This
allows for the formation of a concentration polarisation boundary layer near the
membrane surface, a commonly observed phenomenon which reduces the
effectiveness of NF and RO processes. In addition, the formulation presented here
allows for the permeate flux (the flow of fluid through the membrane) to vary
spatially depending on the local hydrodynamic conditions.
A substantial amount of additional program code has been written to couple the
CFD model to the microscopic mass transfer model, which is used to predict the
rejection of solutes and hence apply appropriate boundary conditions to the CFD
model. The mass transfer model is a refined version of the classic Donnan Steric
Partitioning Model (DSPM), which has been generalised to take molecular shape
effects into account for both charged and uncharged solutes. In addition, a method
has been described and implemented for numerical solution using the model for
an arbitrary combination of multiple charged and uncharged solutes. The
implementation of the coupled CFD-mass transfer model also allows for the
rejection of solutes to vary spatially throughout the membrane channel.
In summary, the novel research elements of this thesis are:
x Coupling of a commercial CFD code to a numerical solution of a refined
version of the DSPM to predict both membrane channel hydrodynamics
and mass transfer through NF and RO membranes;
ii
x Incorporation of both spatially variable permeate flux and solute rejection
into a CFD model of a membrane channel;
x Refinement of the classic DSPM to include molecular shape effects for
both charged and uncharged solutes; and
x Presentation of a numerical method of solution for the refined DSPM for
an arbitrary combination of multiple charged and uncharged solutes.
The coupled CFD-mass transfer model has been tested against experimental and
computational results from various sources, and found to give reliable results. In
addition, a series of numerical experiments have been carried out which highlight
some interesting phenomena predicted by the coupled model, including significant
spatial variation in predicted solute rejection, presence of an inverse concentration
polarisation layer for negatively rejected solutes, variation in near-wall
concentration gradients by solute, and significant variability in overall permeate
flux depending on solution composition. Finally, comments are made regarding
the suitability of the model to general application, and suggestions are made for
future research in extending the applicability of the model as computational
resources become more abundant.
iii
Acknowledgements
I would like to acknowledge the following for their support, hard work, and
guidance:
x Prof. Bas Baskaran, Mr. Mark Mitchell, Dr. Richard Yang, Mr. Steve
Bagshaw and others past and present within the School of Engineering at
Deakin University;
x Deakin University for providing a Deakin University Postgraduate
Research Scholarship;
x A/Prof. Siva Sivakugan, Prof. John Patterson, and others past and present
in the School of Engineering at James Cook University;
x The Graduate Research School at James Cook University for providing a
James Cook University Postgraduate Research Scholarship; and
x Dimuth Navaratna and my other fellow postgraduate students both at
Deakin University and James Cook University.
I must extend a special thank you to my thesis supervisor A/Prof. Jega
Jegatheesan for his tireless work, support, and guidance both as a supervisor and a
friend. I would also like to thank Dr. Li Shu for her guidance and patience.
Finally, and most importantly, I would like to thank my wife Merrin, and my
family, for their unconditional love and support.
iv
Table of contents
Abstract .................................................................................................................... i
Acknowledgements ................................................................................................ iv
Table of contents ..................................................................................................... v
List of figures .......................................................................................................... x
List of tables ....................................................................................................... xxvi
List of publications and presentations ................................................................ xxix
Chapter 1 Introduction............................................................................................. 1
1.1 Thesis outline ................................................................................................. 2
Chapter 2 Membrane filtration and the problem of flux decline ............................. 5
2.1 Membrane processes...................................................................................... 5
2.1.1 Membrane performance ........................................................................ 11
2.2 Membrane flux decline ................................................................................ 12
2.2.1 Concentration polarisation .................................................................... 13
2.2.2 Membrane fouling ................................................................................. 16
2.3 Methods to reduce flux decline.................................................................... 25
2.3.1 Pretreatment of feed solution ................................................................ 25
2.3.2 Modification of membrane properties................................................... 26
2.3.3 Control of membrane operating parameters.......................................... 27
2.3.4 Membrane cleaning ............................................................................... 35
2.4 Summary and directions for research .......................................................... 36
v
Chapter 3 The use of computational fluid dynamics in membrane filtration........ 38
3.1 Basic principles of CFD .............................................................................. 39
3.2 Modelling of membrane processes under laminar conditions ..................... 41
3.2.1 Flow through porous membranes.......................................................... 41
3.2.2 Concentration-polarisation effects ........................................................ 45
3.2.3 Cake formation effects .......................................................................... 50
3.2.4 The influence of osmotic pressure ........................................................ 55
3.3 Modelling of membrane processes under turbulent conditions ................... 58
3.3.1 Studies using Reynolds averaged Navier-Stokes methods ................... 59
3.3.2 Studies using direct numerical simulation methods.............................. 63
3.3.3 Summary of turbulent membrane models ............................................. 66
3.4 Prediction of membrane mass transfer ........................................................ 66
3.5 Comments on commercial membrane simulation packages........................ 74
3.6 Summary and directions for research .......................................................... 75
Chapter 4 Hydrodynamic modelling of membrane filtration ................................ 79
4.1 Introduction ................................................................................................. 79
4.2 2D flow through a porous membrane channel............................................. 80
4.2.1 Model geometry .................................................................................... 81
4.2.2 Boundary and initial conditions ............................................................ 82
4.2.3 Computational mesh, convergence and hydrodynamic verification ..... 84
4.2.4 Incorporation of concentration polarisation effects .............................. 86
4.2.5 Simulations ........................................................................................... 90
4.2.6 Concluding remarks ............................................................................ 103
4.3 Flow in spacer-filled channels ................................................................... 103
vi
4.3.1 Unsteady flow and turbulence in spacer-filled channels .................... 104
4.3.2 2D and 3D modelling .......................................................................... 105
4.3.3 Model geometry .................................................................................. 105
4.3.4 Boundary and initial conditions .......................................................... 107
4.3.5 Incorporation of concentration polarisation effects ............................ 108
4.3.6 Computational mesh and convergence criteria ................................... 109
4.3.7 Transient simulations .......................................................................... 111
4.3.8 Concluding remarks ............................................................................ 127
4.4 Additional remarks on 3D modelling ........................................................ 127
4.5 Conclusions................................................................................................ 134
Chapter 5 Prediction of membrane rejection ....................................................... 137
5.1 Introduction................................................................................................ 137
5.2 Approaches to describe separation in NF and RO processes .................... 138
5.3 Model theory and solution details.............................................................. 140
5.3.1 The extended Nernst-Planck equation ................................................ 140
5.3.2 Rejection of uncharged solutes ........................................................... 141
5.3.3 Rejection of charged solutes ............................................................... 147
5.3.4 Rejection of mixtures of charged and uncharged solutes ................... 151
5.3.5 A note on concentration polarisation and volumetric flux .................. 154
5.4 Computational and numerical considerations ............................................ 155
5.4.1 Use of existing numerical routines and libraries................................. 155
5.4.2 Program structure ................................................................................ 156
5.4.3 Determination of Donnan partitioning equilibrium constants ............ 157
5.5 Results and discussion ............................................................................... 158
vii
5.5.1 Data from Bowen and Mohammad ..................................................... 159
5.5.2 Data from Bowen and Mukhtar .......................................................... 167
5.5.3 Predictions for solutions with four or more charged solutes .............. 178
5.6 Conclusions ............................................................................................... 182
Chapter 6 A fully predictive membrane filtration model: coupled hydrodynamic

and predictive rejection modelling ...................................................................... 184
6.1 Introduction ............................................................................................... 184
6.2 Coupled modelling of solutions containing uncharged solutes ................. 185
6.2.1 Model details ....................................................................................... 185
6.2.2 Model verification ............................................................................... 188
6.2.3 Simulations for low transmembrane pressure ..................................... 193
6.3 Coupled modelling of solutions containing charged solutes ..................... 197
6.3.1 Model details ....................................................................................... 198
6.3.2 Special considerations for the coupled model .................................... 202
6.3.3 Model verification – two charged solutes ........................................... 203
6.3.4 Simulations – multiple charged solutes .............................................. 212
6.4 Coupled modelling of mixed solutions...................................................... 235
6.4.1 Model details ....................................................................................... 235
6.4.2 Simulations ......................................................................................... 238
6.5 Conclusions ............................................................................................... 246
Chapter 7 Simulations using the coupled hydrodynamic and rejection model ... 251
7.1 Introduction ............................................................................................... 251
7.2 Simulation approach .................................................................................. 252
7.2.1 Geometry............................................................................................. 253
viii
7.2.2 Boundary and initial conditions .......................................................... 254
7.2.3 Computational mesh and convergence criteria ................................... 255
7.3 Coupled simulation of single salt solution in spacer filled channel .......... 256
7.4 Coupled simulation of multiple salt solution in spacer filled channel....... 267
7.5 Coupled simulation of complex mixed solution in spacer filled channel .. 286
7.6 Conclusions................................................................................................ 300
Chapter 8 Conclusions......................................................................................... 304
8.1 Overview.................................................................................................... 304
8.2 Hydrodynamic modelling – conclusions and recommendations ............... 305
8.3 Mass transfer modelling for rejection prediction – conclusions and

recommendations ............................................................................................. 308
8.4 Coupled hydrodynamic-mass transfer modelling – conclusions and

recommendations ............................................................................................. 311
Appendix A Sample MATLAB numerical codes for predictive rejection model

............................................................................................................................. 316
Appendix B Sample FORTRAN numerical codes for coupled rejection model –

User Routine approach ........................................................................................ 334
ix
List of figures
Figure 2.1 - Schematic representation of: (a) typical membrane filtration process,
and (b) typical two-phase separation via membrane [after 2]. ................................ 7
Figure 2.2 - Typical flux decline during filtration run [after 2] ............................ 13
Figure 2.3 - Concentration polarisation profile under steady-state conditions [after
2]............................................................................................................................ 14
Figure 2.4 - Typical net-type spacer geometry, showing (a) use in spiral-wound
membrane [47], and (b) woven and non-woven spacers [35]. .............................. 30
Figure 2.5 - Flow regimes for gas-injected flow through membrane [35] ............ 31
Figure 2.6 - Helical screw thread inserts used to induce Dean vortices (left: one
start thread, right: three start thread) [57] ............................................................. 32
Figure 2.7 - Particle detachment mechanisms for ultrasonic agitation [68].......... 35
Figure 3.1 - Finite element mesh used to model concentration polarisation
phenomena [90] ..................................................................................................... 47
Figure 3.2 - Solution algorithm proposed by Richardson and Nassehi [91] ......... 49
Figure 3.3 - Solution algorithm for Lee and Clark [93] cake formation model .... 54
Figure 3.4 - Deviation of analytical and numerical velocity profiles for Pellerin et
al.'s model for (a) ultrafiltration, (b) microfiltration [5]. ....................................... 61
Figure 3.5 – Difference in ionic concentrations predicted in concentration
polarisation boundary layer using both Nernst-Planck and Fickian diffusion
models [103] .......................................................................................................... 73
x
Figure 4.1 - Model geometry for basic 2D model of laminar flow through porous
membrane channel (not to scale) ........................................................................... 82
Figure 4.2 - Comparison of dimensionless axial pressure drop for CFD model
with analytical solutions for hydrodynamic verification of model ....................... 86
Figure 4.3 - Representative detail of computational mesh for 2D channel with
refinement for concentration polarisation boundary layer at (a) channel inlet, (b)
membrane surface (not to scale) ............................................................................ 89
Figure 4.4 - Comparison of predicted permeate fluxes for CFD model with
experimental and modelled fluxes from [94] using experimentally determined
rejection factors ..................................................................................................... 93
Figure 4.5 - Comparison of dimensionless CP boundary layer profile for CFD
model with computational results of [94] using experimentally determined
rejection factors ..................................................................................................... 94
Figure 4.6 - Representative detail of variation in solute mass fraction in 2D
membrane channel including concentration polarisation boundary layer (sucrose
solution, TMP 1 MPa, NRe = 500 – not to scale) ................................................... 96
Figure 4.7 - Representative detail of velocity variation at inlet of 2D membrane
channel including concentration polarisation boundary layer (sucrose solution,
TMP 1 MPa, NRe = 500 – not to scale) .................................................................. 96
Figure 4.8 - Representative detail of velocity variation for entire 2D membrane
xi
Figure 4.9 - Representative detail of density variation for entire 2D membrane
Figure 4.10 - Representative detail of viscosity variation for entire 2D membrane
Figure 4.11 - Representative detail of osmotic pressure variation for entire 2D
membrane channel including concentration polarisation boundary layer (sucrose
solution, TMP 1 MPa, NRe = 500 – not to scale) ................................................... 98
Figure 4.12 - Comparison of predicted permeate fluxes for CFD model
incorporating osmotic pressure correction factor with experimental and modelled
fluxes from [94] using experimentally determined rejection factors .................. 101
model incorporating osmotic pressure correction factor with computational results
of [94] using experimentally determined rejection factors ................................. 102
Figure 4.14 – Model geometry for simulations of flow through 2D space-filled
channel ................................................................................................................ 106
Figure 4.15 – Representative detail of computational mesh for 2D spacer-filled
channel CFD model ............................................................................................. 110
Figure 4.16 – Comparison of predicted axial pressure drop as a function of
channel Reynolds number against computational results of [115] for spacer-filled
channel ................................................................................................................ 113
xii
Figure 4.17 – Comparison of predicted area-averaged wall shear stress as a
function of axial pressure drop against computational results of [115] for spacer-
filled channel ....................................................................................................... 114
Figure 4.18 – Comparison of predicted area-averaged salt mass fraction as a
function of channel Reynolds number against computational results of [115] for
spacer-filled channel............................................................................................ 114
Figure 4.19 – Predicted salt mass fraction distribution for spacer-filled channel for
conditions in [115] for steady flow at lower channel Reynolds numbers ........... 116
Figure 4.20 – Predicted fluid streamlines for spacer-filled channel for conditions
in [115] for steady flow at lower channel Reynolds numbers ............................. 117
Figure 4.21 – Predicted velocity vectors for spacer-filled channel for conditions in
[115] for steady flow at lower channel Reynolds numbers ................................. 118
Figure 4.22 – Predicted vorticity fields for spacer-filled channel for conditions in
[115] for steady flow at lower channel Reynolds numbers ................................. 119
Figure 4.23 – Predicted salt mass fraction distribution for spacer-filled channel for
conditions in [115] for unsteady flow at high channel Reynolds number (Rech =
800) at successive 6.1 ms timesteps (top to bottom) ........................................... 123
Figure 4.24 – Predicted fluid streamlines for spacer-filled channel for conditions
in [115] for unsteady flow at high channel Reynolds number (Rech = 800) at
successive 6.1 ms timesteps (top to bottom) ....................................................... 124
Figure 4.25 – Predicted velocity vectors for spacer-filled channel for conditions in
[115] for unsteady flow at high channel Reynolds number (Rech = 800) at
xiii
Figure 4.26 – Predicted vorticity fields for spacer-filled channel for conditions in
[115] for unsteady flow at high channel Reynolds number (Rech = 800) at
Figure 4.27 – Quarter unit-cell geometry corresponding to 2D model dimensions
used in [115] ........................................................................................................ 130
Figure 4.28 – Coarse 3D mesh (9.7 million elements) for unit-cell corresponding
to 2D model dimensions used in [115]................................................................ 131
Figure 4.29 – Representative detail of coarse 3D mesh (9.7 million elements) for
unit-cell corresponding to 2D model dimensions used in [115] ......................... 131
Figure 5.1 – Flowchart of calculation procedure for combination of charged and
uncharged solutes ................................................................................................ 153
Figure 5.2 - Comparison of experimental rejection data [141] and numerically
computed rejections as a function of flux ........................................................... 161
computed rejections as a function of flux for single salt solutions at fixed
concentration ....................................................................................................... 166
computed rejections as a function of flux for sodium chloride at varying
concentrations...................................................................................................... 166
Figure 5.5 – Variation of charge density as a function of bulk concentration using
experimental data from [141] .............................................................................. 167
Figure 5.6 - Comparison of experimental rejection data from Bowen and Mukhtar
[17] and numerically computed rejections for PES5 membrane for 10-3 M sodium
FKORULGHVROXWLRQXVLQJǻy/Ak and CX as fitting parameters ................................ 170
xiv
[17] and numerically computed rejections for PES5 membrane for 10-3 M sodium
FKORULGHVROXWLRQXVLQJǻy/Ak, CX, and rp as fitting parameters........................... 171
[17] and numerically computed rejections for PES5 membrane for 3×10-3 M
sodium sulphate solution using CX as fitting parameter ...................................... 173
Figure 5.9 - Variation of charge density as a function of bulk concentration using
experimental data from [17] ................................................................................ 175
Figure 5.10 - Comparison of experimental rejection data from Bowen and
Mukhtar [17] and numerically computed rejections for PES5 membrane for 1×10-3
M salt solution using membrane charge density isotherm for non-spherical pore
model with molar ratios of: (a) 0.2, (b) 0.4, (c) 0.6, (d) 0.8 (NaCl:Na2SO4) ...... 177
Figure 5.11 - Numerically computed rejections for PES5 membrane for 1×10-3 M
salt solution using membrane charge density isotherm for non-spherical pore
model with molar ratios of: (a) 0.333:0.333:0.333, (b) 0.25:0.25:0.5, (c)
0.25:0.5:0.25, (d) 0.5:0.25:0.25 (NaCl:Na2SO4:CaCl2)....................................... 179
Figure 5.12 - Numerically computed rejections for PES5 membrane for 1×10-3 M
salt solution using membrane charge density isotherm for non-spherical pore
model with molar ratios as listed in Table 5.11 .................................................. 181
Figure 6.1 – Flowchart of computation procedure for coupled hydrodynamic-mass
transfer model for uncharged solutes only .......................................................... 187
Figure 6.2 - Comparison of predicted permeate fluxes for CFD model with
experimentally determined fluxes of [94] using calculated spatially varying
rejection factors from pore model ....................................................................... 192
xv
model with computational results of [94] using calculated spatially varying
rejection factors from pore model ....................................................................... 193
Figure 6.4 - Axial variation in sucrose rejection for low transmembrane pressure
with spatially varying rejection factor ................................................................. 194
Figure 6.5 - Comparison of predicted permeate fluxes for CFD model for low
transmembrane pressures using calculated constant and spatially varying rejection
factors .................................................................................................................. 196
model for low transmembrane pressures using calculated constant and spatially
varying rejection factors ...................................................................................... 196
Figure 6.7 – Flowchart of computation procedure for coupled hydrodynamic-mass
transfer model for charged solutes only .............................................................. 201
Figure 6.8 – Comparison of predicted rejection and experimental rejection for
NaCl solution with CDNF50l membrane using fitted parameters from Table 6.5
............................................................................................................................. 207
Figure 6.9 – Comparison of experimental rejection (unfilled markers), predicted
rejection from pore model (solid line) and predicted rejection from coupled CFD-
pore model (filled markers) for NaCl solution with CDNF50l membrane using
fitted parameters from Table 6.5 ......................................................................... 209
Figure 6.10 – Predicted axial rejection variation from coupled CFD-pore model
for NaCl solution with CDNF50l membrane using fitted parameters from Table
6.5 ........................................................................................................................ 210
xvi
Figure 6.11 - Comparison of dimensionless CP boundary layer profile for coupled
CFD-pore model with computational results of [94] .......................................... 211
Mukhtar [17], numerically computed rejections from pore model, and area-
averaged predicted rejections from coupled hydrodynamic / pore model for PES5
membrane for 1×10-3 M salt solution with molar ratios of: (a) 0.2, (b) 0.4, (c) 0.6,
(d) 0.8 (NaCl:Na2SO4) ......................................................................................... 216
Figure 6.13 – Predicted axial rejection variation for PES5 membrane for 1×10-3 M
salt solution with molar ratios of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa ........ 218
Figure 6.14 – Predicted axial permeate flux variation for PES5 membrane for
1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4) ........................... 219
Figure 6.15 – Flowchart of computation procedure for coupled hydrodynamic-
mass transfer model for multiple charged solutes with simplified diffusion model
............................................................................................................................. 224
Mukhtar [17], numerically computed rejections from pore model, and area-
averaged predicted rejections from coupled hydrodynamic / pore model both with
and without concentration polarisation effects for PES5 membrane for 1×10-3 M
salt solution with molar ratio of 0.2 (NaCl:Na2SO4) ........................................... 228
Figure 6.17 – Predicted axial rejection variation for PES5 membrane both with
and without concentration polarisation effects for 1×10-3 M salt solution with
molar ratio of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa ...................................... 230
xvii
Figure 6.18 – Predicted axial permeate flux variation for PES5 membrane both
with and without concentration polarisation effects for 1×10-3 M salt solution with
molar ratio of 0.2 (NaCl:Na2SO4) ....................................................................... 231
Figure 6.19 – Detail of computed concentration fields (normalised relative to bulk
inlet concentration) of (a) Na+, (b) Cl-, (c) SO42- for PES5 membrane with
concentration polarisation effects for 1×10-3 M salt solution with molar ratio of
0.2 (NaCl:Na2SO4) for TMP of 200 kPa ............................................................. 233
Figure 6.20 – Predicted variation normal to membrane surface for solute
concentration normalised relative to inlet bulk concentration for PES5 membrane
with concentration polarisation effects for 1×10-3 M salt solution with molar ratio
of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa ......................................................... 234
Figure 6.21 – Flowchart of computation procedure for coupled hydrodynamic-
mass transfer model for multiple charged and/or uncharged solutes with simplified
diffusion model ................................................................................................... 237
Figure 6.22 - Comparison of area-averaged predicted rejections from coupled
hydrodynamic / pore model with concentration polarisation effects for PES5
membrane for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4) with
and without sucrose addition ............................................................................... 240
Figure 6.23 – Predicted axial rejection variation for PES5 membrane for 1×10-3 M
salt solution with molar ratio of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa with and
without sucrose addition...................................................................................... 243
Figure 6.24 – Predicted axial permeate flux variation for PES5 membrane both
with and without sucrose addition for 1×10-3 M salt solution with molar ratio of
0.2 (NaCl:Na2SO4) .............................................................................................. 244
xviii
Figure 6.25 – Detail of computed concentration fields (normalised relative to inlet
bulk concentration) of (a) Na+, (b) Cl-, (c) SO42-, (d) sucrose for PES5 membrane
for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4) with sucrose
addition for TMP of 200 kPa............................................................................... 245
Figure 6.26 – Predicted variation normal to membrane surface for solute
concentration normalised relative to inlet bulk concentration for PES5 membrane
for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4) with sucrose
addition for TMP of 200 kPa............................................................................... 246
Figure 7.1 – Model geometry for simulations of flow through 2D space-filled
channel................................................................................................................. 254
Figure 7.2 – Predicted salt mass fraction distribution for spacer-filled channel
using coupled-mass transfer model for single salt solution (NaCl, 0.002 kg kg-1,
equivalent to 34.1 mol m-3) ................................................................................. 257
Figure 7.3 – Predicted fluid streamlines for spacer-filled channel using coupled-
mass transfer model for single salt solution (NaCl, 0.002 kg kg-1, equivalent to
34.1 mol m-3) ....................................................................................................... 258
Figure 7.4 – Predicted velocity vectors for spacer-filled channel using coupled-
34.1 mol m-3) ....................................................................................................... 258
Figure 7.5 – Predicted vorticity fields for spacer-filled channel using coupled-
34.1 mol m-3) ....................................................................................................... 259
xix
Figure 7.6 - Predicted variation in mass transfer parameters within spacer-filled
channel with spatially varying permeate flux and solute rejection (NaCl, 0.002 kg
kg-1, equivalent to 34.1 mol m-3, Rech = 100) ...................................................... 261
channel with spatially varying permeate flux and solute rejection (NaCl, 0.002 kg
kg-1, equivalent to 34.1 mol m-3, Rech = 200) ...................................................... 262
Figure 7.8 - Predicted transverse variation in normalised solute concentration as a
function of non-dimensionalised distance from lower membrane wall within
spacer-filled channel with spatially varying permeate flux and solute rejection
(NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3) .............................................. 265
Figure 7.9 - Predicted transverse variation in normalised solute concentration as a
function of non-dimensionalised distance from upper membrane wall within
(NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3) .............................................. 265
Figure 7.10 – Predicted ionic mass fraction distribution for spacer-filled channel
using coupled-mass transfer model for multiple salt solution (NaCl:Na2SO4 with
molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3), Rech = 100 ................... 268
using coupled-mass transfer model for multiple salt solution (NaCl:Na2SO4 with
molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3), Rech = 200 ................... 269
mass transfer model for multiple salt solution (NaCl:Na2SO4 with molar ratio
0.2:0.8, overall ionic strength 34.1 eq m-3) ......................................................... 270
xx
channel with spatially varying permeate flux and solute rejection (NaCl:Na2SO4
with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, Rech = 100) ........... 275
channel with spatially varying permeate flux and solute rejection (NaCl:Na2SO4
Figure 7.17 - Predicted transverse variation in normalised solute concentration as
a function of non-dimensionalised distance from lower membrane wall within
(NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, Rech =
100)...................................................................................................................... 278
200)...................................................................................................................... 278
a function of non-dimensionalised distance from upper membrane wall within
xxi
100)...................................................................................................................... 279
200)...................................................................................................................... 279
Figure 7.21 – Difference in predicted ionic mass fraction distribution for spacer-
filled channel between constant rejection model and variable rejection model for
multiple salt solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic
strength 34.1 eq m-3), Rech = 100 ........................................................................ 281
Figure 7.22 – Difference in predicted ionic mass fraction distribution for spacer-
multiple salt solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic
strength 34.1 eq m-3), Rech = 200 ........................................................................ 283
Figure 7.23 – Difference in predicted permeate flux within spacer-filled channel w
between constant rejection model and variable rejection model (NaCl:Na2SO4
using coupled-mass transfer model for complex mixture solution (NaCl:Na2SO4
xxii
with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and additional sucrose
at 0.001 kg kg-1), Rech = 100 ............................................................................... 287
using coupled-mass transfer model for complex mixture solution (NaCl:Na2SO4
at 0.001 kg kg-1), Rech = 200 ............................................................................... 288
mass transfer model for complex mixture solution (NaCl:Na2SO4 with molar ratio
0.2:0.8, overall ionic strength 34.1 eq m-3, and additional sucrose at 0.001 kg kg-1)
............................................................................................................................. 289
mass transfer model for complex mixture solution (NaCl:Na2SO4 with molar ratio
............................................................................................................................. 289
mass transfer model for single complex mixture solution (NaCl:Na2SO4 with
molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and additional sucrose at
0.001 kg kg-1) ...................................................................................................... 290
Figure 7.30 – Predicted variation in mass transfer parameters within spacer-filled
channel with spatially varying permeate flux and solute rejection
(NaCl:Na2SO4:sucrose solution, Rech = 100) ...................................................... 292
channel with spatially varying permeate flux and solute rejection
(NaCl:Na2SO4:sucrose solution, Rech = 200) ...................................................... 293
xxiii
(NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and
additional sucrose at 0.001 kg kg-1, Rech = 100) ................................................. 294
Figure 7.36 – Difference in predicted solute mass fraction distribution for spacer-
complex mixture solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic
strength 34.1 eq m-3, and additional sucrose at 0.001 kg kg-1), Rech = 100 ........ 297
xxiv
Figure 7.37 – Difference in predicted solute mass fraction distribution for spacer-
complex mixture solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic
strength 34.1 eq m-3, and additional sucrose at 0.001 kg kg-1), Rech = 200......... 298
at 0.001 kg kg-1, Rech = 100) ............................................................................... 299
at 0.001 kg kg-1, Rech = 200) ............................................................................... 299
xxv
List of tables
Table 2.1 – Summary of common membrane processes [after 2] ........................... 8
Table 2.2 - Comparison of pressure-driven membrane processes [after 2] ............ 9
Table 2.3 - Summary of models derived for constant pressure filtration law for
fouled membranes [3]............................................................................................ 21
Table 2.4 - Summary of UF filtration models for fouled membranes [compiled
from 3] ................................................................................................................... 22
Table 2.5 - Summary of RO filtration models [compiled from 3] ........................ 24
Table 3.1- Summary of characteristics of discretisation methods [compiled from
79].......................................................................................................................... 40
Table 4.1 - Model input parameters for verification case ..................................... 85
Table 4.2 - Trial solute components and expressions for transport properties (after
[94]) ....................................................................................................................... 88
Table 4.3 – Model input parameters for simulations corresponding with Geraldes
et al. [94] ............................................................................................................... 90
Table 4.4 - Experimentally observed rejection factors for test compounds from
[94] for NRe = 2000 ................................................................................................ 91
Table 4.5 - Osmotic pressure correction factor expressions for test compounds
(after [94]) ............................................................................................................. 99
Table 4.6 – Model input parameters for simulations corresponding with Wardeh &
Morvan [115]....................................................................................................... 108
xxvi
Table 5.1 – Summary of existing numerical libraries and routines used in
development of predictive rejection model code ................................................ 155
Table 5.2 - Physical properties for uncharged solutes used in [141] .................. 160
Table 5.3 – Fitting results for uncharged solutes used in [141] .......................... 161
Table 5.4 - Physical properties for ionic solutes used in [141] ........................... 164
Table 5.5 – Fitting results for single salt solutions used in [141] ....................... 165
Table 5.6 - Physical properties for solutes from [17] .......................................... 169
Table 5.7 - Comparison of fitting for ǻy/Ak and CX for PES5 membrane used by
Bowen and Mukhtar [17] for 10-3 M sodium chloride solution .......................... 170
Table 5.8 - Comparison of ILWWLQJIRUǻy/Ak, CX, and rp for PES5 membrane used
by Bowen and Mukhtar [17] for 10-3 M sodium chloride solution ..................... 172
Table 5.9 - Comparison of fitting for CX for PES5 membrane used by Bowen and
Mukhtar [17] for 3×10-3 M sodium sulphate solution ......................................... 173
Table 5.10 – Fitted parameters for charge density isotherms using experimental
data from [17] ...................................................................................................... 175
Table 5.11 – Molar ratios used for simulation of six ion 1×10-3 M salt solution 180
Table 6.1 - Molecular properties for uncharged test compounds derived using
MOPAC2009 software ........................................................................................ 188
Table 6.2 - Predicted and experimentally observed rejection factors for test
compounds for CDNF50l thin-film composite NF membrane (Separem, Biela,
Italy) .................................................................................................................... 191
Table 6.3 – Minimum arguments to be passed from CFX solver to external
Fortran routine for calculating rejection of charged solutes................................ 198
xxvii
Table 6.4 – Estimates for permeate fluxes and wall concentrations for NaCl
solution based on previous CFD results .............................................................. 204
Table 6.5 – Fitted membrane parameters for CDNF50l membrane for NaCl
solution ................................................................................................................ 206
Table 6.6 – Model input parameters for simulations corresponding with Bowen &
Mukhtar [17] ....................................................................................................... 214
Table 6.7 – Properties for additional uncharged solute (sucrose) in coupled
simulation of charged-uncharged solute mixture ................................................ 238
Table 7.1 – Model input parameters for coupled simulations of 2D spacer-filled
channel ................................................................................................................ 255
Table 7.2 – Comparison of average permeate flux values for single salt (NaCl,
overall ionic strength 34.1 eq m-3) and multiple salt (NaCl:Na2SO4 with molar
ratio 0.2:0.8, overall ionic strength 34.1 eq m-3) cases........................................ 273
Table 7.3 – Fixed rejection values used for uncoupled model comparison case
(NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3) ...... 280
Table 7.4 – Comparison of average permeate flux values for single salt (NaCl,
overall ionic strength 34.1 eq m-3), multiple salt (NaCl:Na2SO4 with molar ratio
0.2:0.8, overall ionic strength 34.1 eq m-3), and complex mixture (NaCl:Na2SO4
at 0.001 kg kg-1) cases ......................................................................................... 290
Table 7.5 – Fixed rejection values used for uncoupled model comparison case
additional sucrose at 0.001 kg kg-1) .................................................................... 296
xxviii
List of publications and presentations
The following publications were derived from the work contained in this thesis:
G. Keir, V. Jegatheesan. Prediction of solute rejection and modelling of steady-
state concentration polarisation effects in pressure-driven membrane filtration
using computational fluid dynamics, Membrane Water Technology, Vol.3, No.2,
(2012).
The following presentations were derived from the work contained in this thesis:
G. Keir, V. Jegatheesan (2011), Prediction of solute rejection and modelling of
steady-state concentration polarisation effects in pressure-driven membrane
filtration using computational fluid dynamics, Conferences on the Challenges in
Environmental Science and Engineering 2011, Tainan, Taiwan.
G. Keir, V. Jegatheesan (2011), Modelling of nanofiltration and reverse osmosis
using coupled hydrodynamic and pore models, Institute for Technology Research
and Innovation Research Conference 2011, Geelong, Victoria.
xxix
G. Keir, V. Jegatheesan (2011), Modelling of Fouling Processes in Membrane
Filtration Using Computational Fluid Dynamics, Victorian Modelling Group
Meeting #6 for the Australian Hydraulic Modelling Association, Geelong,
Victoria.
xxx
Chapter 1
Introduction
Membrane filtration is a common industrial process for removing dissolved
components from liquid streams, especially in water and wastewater treatment.
While the process is widely used, predicting the behaviour and performance of a
particular membrane system is difficult. This is due to both the complex
hydrodynamic behaviour of the bulk flow which passes through membrane
modules, and the complex interactions between the dissolved constituents and the
membrane microstructure. Because of this, performance can vary widely based on
feed stream composition and operating conditions, and hence to determine
suitability for a particular application, pilot-scale studies are often required.
Development of a general predictive model for membrane filtration processes
would therefore be of significant benefit.
However, most research into membrane modelling has considered the
hydrodynamic processes within the membrane channel, and the mass transfer
processes governing removal of dissolved components by the membrane, in
isolation. The aim of this study is to reconcile these two approaches, and develop
an integrated computational model to describe both the membrane channel
hydrodynamics using a commercial computational fluid dynamics (CFD) code,
and the membrane mass transfer with appropriate numerical models.
1
1.1 Thesis outline
This thesis consists of eight chapters which are described in the following.
Chapter 1 (the present section) outlines the structure of the document, and
provides a concise statement of the research question and arguments for the
relevance and importance of the study.
Chapter 2 presents an overview of pressure-driven membrane filtration processes,
including microfiltration, ultrafiltration, nanofiltration, and reverse osmosis, and
describes the problem of flux decline inherent to all these processes. It describes
the mathematical models which have been developed to model these processes, as
well as presenting techniques that have been proposed to reduce or minimise this
flux decline.
Chapter 3 describes the basic principles of computational fluid dynamics, and
investigates the use of computational fluid dynamics techniques in describing the
membrane filtration process. The modelling of concentration polarisation effects
and membrane fouling is detailed in order to describe the process of membrane
flux decline. As well, the use of transient simulations and numerical turbulence
models is investigated.
Chapter 4 describes the development of a basic set of rigorous hydrodynamic
models of the membrane filtration process. These models are validated against
experimental and computational results presented in the literature, and provide a
2
useful method of predicting membrane hydrodynamic performance in the cases
where the rejection behaviour of the membrane is well known. A variety of
membrane channel geometric configurations are investigated, and relevant
hydrodynamic phenomena are discussed.
Chapter 5 presents a numerical method which has been developed in order to
predict the rejection of various components in solution by nanofiltration and
reverse osmosis membranes. The method, unlike most other methods presented in
the literature, is general enough to predict the rejection for all elements of a
complex mixture composed of arbitrary numbers of charged and uncharged
solutes. The method has been developed in order to be coupled with a rigorous
hydrodynamic simulation of the membrane filtration process as described in
Chapter 4, which allows for a complete predictive model of the membrane
filtration process to be established.
Chapter 6 describes the coupling of the hydrodynamic models presented in
Chapter 4 and the predictive rejection models of Chapter 5 to present a fully
predictive model of the membrane filtration process for nanofiltration and reverse
osmosis membranes. The model is verified against experimental and
computational results from the literature, and observations are made regarding the
implementation of the model in terms of numerical stability, general utility, and
areas for future refinement.
3
In Chapter 7, the coupled hydrodynamic-mass transfer model is used to perform
detailed simulations of realistic membrane channel geometries and hydrodynamic
conditions. A series of numerical experiments are made using the new model, and
salient hydrodynamic and mass transfer phenomena are identified and discussed.
Chapter 8 summarises the findings made in Chapter 4, Chapter 5, Chapter 6, and
Chapter 7, and presents a set of conclusions regarding the use of coupled
hydrodynamic and predictive rejection models for membrane filtration.
Comments are made regarding the suitability of this computational approach for
predicting the overall behaviour of membrane filtration systems in terms of
computational efficiency and general applicability. Finally, several suggestions
are made for future work in this area.
4
Chapter 2
Membrane filtration and the problem of flux decline
Membrane filtration plays an important role in modern life, and has grown in a
matter of years from a specialised laboratory tool into a widespread industrial
process with significant industrial, economic and environmental importance [1].
Membranes are used today for a variety of large scale applications, including:
x treatment of drinking water to remove contaminants and pathogens;
x production of potable water from seawater;
x treatment of domestic and industrial wastewater;
x recovery of valuable constituents from industrial effluent; and
x various medical applications, such as removal of urea and toxins from blood
streams by dialysis.
While there are a multitude of different membrane processes and applications,
there are certain underlying features common to all membrane processes, as
discussed in the following.
2.1 Membrane processes
Membrane processes typically consist of a feed stream consisting of several
phases (solvents and solute), which is then divided by a membrane module into a
5
permeate stream and a concentrate stream [2]. A membrane can be considered in
the general sense as a barrier which separates two phases and restricts the
transport of various chemical species in a specific manner [1]. Based on this
definition, a great number of materials and structures may be considered as
membranes. The primary method of classifying membranes is the distinction
between biological and synthetic membranes, though nearly all membranes of any
industrial importance are synthetic.
Mulder [2] identified three basic types of synthetic membranes:
x porous membranes – these separate phases by discriminating between particle
sizes;
x non-porous membranes – these separate phases through differences in
solubility and/or diffusivity, and can separate particles of similar size; and
x carrier membranes – these separate phases by use of very specific carrier
molecules which facilitate transport of specific components.
Separation characteristics of a membrane are largely determined by the
dimensions of the membrane structure for porous membranes, while selection for
other membrane types is mostly dictated by the material properties of the
membrane. However, it should be emphasised that there is no discrete transition
between membrane types, and some processes may be intermediate between these
groupings.
6
Transport through the membrane occurs as a result of some driving force, such as
a gradient in pressure, temperature, concentration or electric potential. A
schematic representation of the membrane process is shown in Figure 2.1. The
driving force for the membrane process varies depending on the specific
application. A list of common membrane processes and their driving mechanisms
is shown in Table 2.1.
membrane
feed stream concentrate

membrane
module
feed permeate
permeate driving
force
(a) (b)
Figure 2.1 - Schematic representation of: (a) typical membrane filtration process, and (b) typical two-phase
separation via membrane [after 2].
The membrane processes generally of interest to the environmental engineer, and
those that will be considered in this thesis, are the conventional pressure-driven
membrane processes. These are (in order of decreasing size of constituents
treated):
x microfiltration (MF);
x ultrafiltration (UF);
7
x nanofiltration (NF); and
x reverse osmosis (RO).
A comparison of typical parameters and applications for these processes is given
in Table 2.2.
Table 2.1 – Summary of common membrane processes [after 2]
Membrane process Phase 1 Phase 2 Driving force
Microfiltration Liquid Liquid Hydraulic pressure
Ultrafiltration Liquid Liquid Hydraulic pressure
Nanofiltration Liquid Liquid Hydraulic pressure
Reverse osmosis Liquid Liquid Hydraulic pressure
Piezodialysis Liquid Liquid Hydraulic pressure
Gas separation Gas Gas Vapour pressure
Vapour permeation Gas Gas Vapour pressure
Pervaporation Liquid Gas Vapour pressure
Electrodialysis Liquid Liquid Electrical potential
Membrane electrolysis Liquid Liquid Electrical potential
Dialysis Liquid Liquid Concentration
Diffusion dialysis Liquid Liquid Concentration
Membrane contactors Liquid Liquid Concentration
Gas Liquid Concentration / vapour pressure
Liquid Gas Concentration / vapour pressure
Thermo-osmosis Liquid Liquid Temperature / vapour pressure
Membrane distillation Liquid Liquid Temperature / vapour pressure
8
Table 2.2 - Comparison of pressure-driven membrane processes [after 2]
Characteristic Microfiltration Ultrafiltration Nanofiltration Reverse osmosis
Membrane type Symmetric/asymmetric Asymmetric Composite Asymmetric,

porous porous composite
Thickness 10-150 μm 150 μm Sublayer 150 μm, Sublayer 150 μm,

top layer <1 μm top layer <1 μm
Effective pore 0.05-10 μm 1-100 nm <2 nm <2 nm

size
Driving force Pressure (<2 bar) Pressure (1-10 Pressure (10-25 Pressure (12-25
bar) bar) bar brackish
water, 40-80 bar
seawater)
Separation Sieving mechanism Sieving Solution-diffusion Solution-diffusion

principle mechanism
Membrane Polymeric, ceramic Polymeric, Polyamide Cellulose

material ceramic triacetate,
aromatic
polyamide,
polyamide and
poly(ether urea)
Applications Water treatment, Dairy, food, Desalination, Desalination of

sterilisation, beverage metallurgy, micropollutant brackish water
clarification, analytical textile, removal, water and seawater,
applications pharmaceutical, softening, production of
automotive, water wastewater, dye ultrapure water,
treatment retention concentration of
juice, sugars and
milk.
Of the membrane filtration processes, MF most closely resembles conventional
coarse filtration, and is suitable for retaining suspensions and emulsions [2]. The
primary rejection mechanism for MF is physical sieving. In addition, the deposit
(cake) on the membrane acts to retain additional smaller particles than would be
expected for a given pore size [3].
9
UF is commonly used to remove colloidal particles and macromolecules. The
main rejection mechanism is physical sieving, although adsorption, charge
interactions, and the self-rejecting action of the deposited layer are also of
importance [3].
RO processes achieve membrane flux and separation by applying high
transmembrane pressure (TMP) to overcome the osmotic pressure of the solution.
RO removes both monovalent and multivalent ions, and is used primarily in
desalination applications. Rejection is largely due to the chemical affinity of the
solute to the membrane material, and typically increases with increasing hydrated
radius of the solute ion. However, interactions between mixtures of ions are
complex, and removal mechanisms are generally poorly understood [3].
NF processes lie between UF and RO processes, with both charge and size
playing important roles in NF rejection. For uncharged molecules, separation
occurs predominantly through physical sieving, while for ionic species, both
sieving effects and electrostatic interactions occur [4].
There are four main types of membrane module [5] which are:
x plate and frame modules, which consist of a thin channel in which one or
both of the walls are a membrane;
x spiral-wound modules, in which a thin channel is wrapped in a spiral
formation to increase membrane area per unit volume of filter;
10
x tubular and hollow fibre modules, which are cylindrical membrane tubes;
and
x capillary fibre modules, which are identical to hollow fibre modules but
with very small inner diameter.
2.1.1 Membrane performance
To describe the behaviour of a membrane filtration system, it is useful to first
define several parameters. The first of these is the flux, or flow though the
membrane, J, which is given by:
1 dV (2.1)
J
A dt
where V is the filtrate volume, t is time, and A is the surface area of the membrane
[3].
The rejection of a membrane Ri describes the efficiency with which the membrane
is able to separate phases of the feed mixture. The rejection is given by:
CP CB CP (2.2)
Ri 1
CB CB
where CB is the concentration in the bulk of the feed solution, and CP is the
concentration of the permeate. The value of Ri varies between zero (where both
solute and solvent pass through the membrane freely) and one (where the solute is
completely retained).
11
The flow through the membrane can also be characterised by the hydrodynamic
permeability LV, which is given by:
J (2.3)
LV
'P
where ¨3 is the TMP. These terms will be used in the following to describe
theoretical models of the membrane filtration process.
2.2 Membrane flux decline
The primary technical issue which has limited the application of membrane
filtration is that of flux decline –during a membrane separation process, the flux
through the membrane will inevitably decrease with time. The flux decline can be
substantial: in some processes: the process flux is often less than 5% of the pure
water flux [2]. Eliminating or reducing the extent of flux decline is thus one of the
primary aims of membrane research.
This flux decline can be attributed to two main factors: polarisation phenomena
(mainly concentration polarisation), and membrane fouling. A typical pattern of
flux decline over the duration of a membrane filtration run illustrating the
influence of these two factors is shown in Figure 2.2.
12
initial flux
flux decline
due to
polarisation
membrane
flux
flux J decline
due to
fouling
time t
Figure 2.2 - Typical flux decline during filtration run [after 2]
Flux decline due to polarisation occurs relatively rapidly, but is a reversible
process which represents only a small amount of the total flux decline. The
remainder of the flux decline is due to membrane fouling. However, fouling is not
completely independent of polarisation phenomena, and can be minimised by
restricting the extent of polarisation. It is useful to begin by distinguishing
between concentration polarisation (the polarisation phenomenon of interest in
pressure-driven membrane processes) and membrane fouling.
2.2.1 Concentration polarisation
Concentration polarisation occurs when rejected solutes accumulate next to the
membrane wall, forming a boundary layer [6]. This accumulation of solute causes
a diffusive flow back to the feed, which eventually reaches a steady-state
condition where the solute flow to the membrane surface equals the solute flux
through the membrane plus the diffusive flow from the membrane surface back to
13
the feed. This establishes a concentration profile within the boundary layer as
shown in Figure 2.3.
bulk of feed boundary membrane

layer
CW
CB
CP
boundary layer thickness į
distance from membrane surface x
Figure 2.3 - Concentration polarisation profile under steady-state conditions [after 2]
The primary consequence of concentration polarisation is the reduction in
permeate flux, by either increasing the osmotic pressure on the feed side of the
membrane or forming a gel layer at the membrane surface. Several models have
been developed to account for these phenomena, as discussed in the following.
2.2.1.1 Models of concentration polarisation
The Film Theory Model assumes that axial (crossflow) solute convection near the
membrane surface is negligible [7], and describes the concentration profile by a
one-dimensional mass balance equation. By integrating from the membrane wall
to the edge of the boundary later, the following is obtained:
14
DS CW C P (2.4)
J ln
G C B C P
where DS is the solute diffusivity, į is the boundary layer thickness, and CW is the
concentration at the membrane surface or wall.
The Gel Polarisation Model assumes that beyond a certain value of pressure
applied to the membrane, the permeation rate is limited by the presence of a gel
layer deposited on the membrane surface, formed as the concentration in the
boundary layer reaches the solubility limit [8]. Thus, assuming an upper limit on
the solute concentration at the membrane surface CG (the gel concentration), the
following equation can be obtained:
DS CG (2.5)
J ln
G CB
The Osmotic Pressure Model assumes that permeate flux is limited by the osmotic
pressure exerted by retained macromolecules on the surface of the membrane [2].
This osmotic pressure is normally neglected as it is small compared to the osmotic
pressure of low molecular weight solutes that pass through the membrane.
However, for high flux and rejection, the high concentration of macromolecules
on the membrane surface means the osmotic pressure can no longer be neglected,
and forms an osmotic pressure difference ¨Ȇ across the membrane. The flux is
then given by:
15
'P '3 (2.6)
J
KRM
where Ș is the dynamic viscosity of the solvent and RM is the clean membrane
resistance.
The Boundary Layer Resistance model assumes that the high concentration of
retained molecules in the boundary layer exerts a hydraulic resistance upon the
permeating solvent molecules. Assuming this resistance and the hydraulic
resistance of the membrane act in series, the following equation is obtained [9]:
'P (2.7)
J
K R M RCP
where RCP is the resistance exerted in the boundary layer due to concentration
polarisation. In RO applications, where osmotic pressure is significant, the
osmotic pressure effects are often incorporated into the resistance RCP.
2.2.2 Membrane fouling
Distinct from concentration polarisation, membrane fouling is defined as the
irreversible deposition of retained particles, colloids, emulsions, suspensions,
macromolecules, salts or other substances on or within the membrane [2].
Membrane fouling depends on an array of physical and chemical parameters (such
as concentration, temperature, pH, ionic strength and others) as well as other
specific interactions (e.g. Van der Waals forces, hydrogen bonding, dipole
16
interactions). As such, the phenomenon of fouling is extremely complex and not
fully understood.
Attempts have been made to classify fouling types by several methods, including:
x on the basis of fouling material;
x on the basis of the site at which fouling occurs; and
x on the basis of the underlying mechanism by which fouling occurs.
A brief summary of these classifications follows, though it is emphasised that
there are undoubtedly overlaps and linkages between the different classification
schemes.
2.2.2.1 Fouling materials
Several types of fouling material have been identified to occur in membrane
filtration systems. These include inorganic fouling or scaling, particulate and
colloidal fouling, organic fouling, and biological fouling (biofouling) [10].
Inorganic fouling is due to the deposition of inorganic compounds or scales of the
membrane surface [11]. This occurs when salts rejected by the membrane exceed
their solubility limit in the boundary layer, and are precipitated on the surface.
The inorganic foulants generally of practical concern are salts of calcium,
magnesium, carbonate, sulphate, silica and iron [3].
17
Particulate or colloidal fouling is the fouling of the membrane surface by
particulate material (suspended solids, colloids, and microorganisms) [12].
Particulate fouling is one of the most persistent types of fouling, and is
particularly prominent in the treatment of surface waters.
Organic fouling refers to the accumulation of organic compounds on or within the
membrane, and depends to a large extent on the organic characteristics. Large
organic molecules tend to cause more severe fouling, and mixtures of organic
compounds tend to display worse fouling than in isolation, due to solute-solute
interactions [3].
Biofouling occurs when the membrane surface is exposed to a liquid containing
microorganisms and dissolved organic matter (DOM). The organic molecules are
absorbed to the membrane surface, to which the microorganisms adhere to
produce a biofilm [10]. Biofouling is of significant concern in RO processes,
particularly those in tropical or sup-tropical regions [13].
2.2.2.2 Fouling locations
The classification of membrane fouling by the site of its occurrence distinguishes
between two primary fouling locations: the external surface (the build up of a
cake layer on the feed side of the membrane), and the membrane pores [14].
Fouling within the membrane pores can be further classified as complete pore
blocking (where pores are blocked by particles of approximately the pore size),
18
standard pore blocking (where pores are gradually narrowed by particles much
smaller than the pore size), and incomplete pore blocking (an intermediary
between the two). Internal fouling within the membrane pores is largely a function
of the physico-chemical parameters of the membrane system, while external
fouling is to a larger extent dictated by the hydrodynamic conditions (similar to
concentration polarisation).
2.2.2.3 Fouling mechanisms
The main mechanisms considered to govern membrane fouling are pore blocking
and cake formation (which are related to the site of fouling as discussed above).
Other mechanisms such as adsorption, particle deposition within pores, and
alterations to the cake layer are thought to essentially act by modifying either or
both of the major mechanisms [15].
A further distinction can be made between particle adsorption and deposition.
Macrosolute or particle adsorption is the interactions between particles and the
membrane which occur even in the absence of filtration (i.e. in the absence of a
TMP), and is usually responsible for irreversible fouling. Particle deposition, on
the other hand, is induced by filtration, and is the accumulation of particles on the
top surface of the membrane. This is reversible, non-adhesive fouling which is
analogous to external fouling or cake formation [15].
19
2.2.2.4 Models of membrane fouling
A multitude of mathematical models of membrane fouling have been developed.
A brief summary of the more notable models found in the literature follows
below.
The Resistance in Series Model describes the flux of a fouled MF membrane, and
is an extension of the Boundary Layer Resistance model previously discussed for
concentration polarisation. It adds two additional resistance terms for internal
fouling and external fouling (cake formation), denoted as RP and RC, respectively.
Again assuming the resistances act in series, the flux through the membrane is
then given by:
'P (2.8)
J
K R M RCP RC R P
A set of so-called filtration laws have been proposed [3] to further describe
fouling mechanisms. The general form known as the constant pressure filtration
law is given by:
d 2t § dt ·
n (2.9)
k¨ ¸
dV 2 © dV ¹
where k is a distribution coefficient, and n is a non-negative exponent. This was
used to develop four main filtration models, which are summarised in Table 2.3.
20
Table 2.3 - Summary of models derived for constant pressure filtration law for fouled membranes [3]
Name of model Governing equations Details
Complete Blocking 2
(2.10) Valid for particles with similar size to
d 2t § dt ·
Model k¨ ¸ membrane pores – particles seal the
dV 2 © dV ¹ pores and do not deposit on one
another. J0 is the initial flux through
which can be solved to give
the membrane.
V
J 0 1 e kt (2.11)
Standard Blocking 3/ 2
(2.12) Valid for particles much smaller than
d 2t § dt ·
Model k¨ ¸ membrane pores – particles pass
dV 2 © dV ¹ through pores and deposit on inner
pore surface. Pore volume decreases
proportionally with filtrate volume.
t kt 1 (2.13)

V 2 J0
Intermediate d 2t (2.14) Every particle reaching a pore

§ dt ·
Blocking Model k¨ ¸ contributes to blockage – particles
dV 2 © dV ¹ deposit on one another. Describes
which can be solved to give long term adsorption.
kV ln 1 ktJ 0 (2.15)
Cake Filtration (2.16) Valid for particles much larger than

d 2t
Model k membrane pores – particles deposit on
dV 2 surface without entering pores.
t kV 1 (2.17)

V 2 J0
A separate class of models has also been developed for UF processes, in which
mechanisms such as adsorption and charge interactions are more significant. A
summary of the major models for UF processes are summarised in Table 2.4.
21
Table 2.4 - Summary of UF filtration models for fouled membranes [compiled from 3]
Mechanical Sieving Ri >O 2 O @2 for O 1 (2.18) Describes hindered transport of solute

Model by convection and limited by steric
effects. Does not account for solute
Ri 1 for O t 1 (2.19) velocity drag, diffusion, or
concentration effects at membrane
surface.
O
d solute (2.20)
d pore
Modified Sieving (2.21) As per Mechanical Sieving Model,

d solute 2N 1
Rejection Model O except Ȝ includes Debye length term ț-
d pore 2N 1 1
to account for double layer thickness
surrounding charged solute.
Pore Flow Model (2.22) JS is the solute flux, rP is the pore

S'PWn p rp 4 radius, nP is the number of pores, ¨[ is
J
8K'x the membrane thickness and Ĳis the
tortuosity coefficient. Describes flow
Js 1 V JCB (2.23) through cylindrical membrane pores –
simplified model does not consider
frictional or diffusive forces.
Models for fouled NF processes vary again - the above UF models (and some RO
models to be discussed later) are applicable to NF processes to some extent,
though charge appears to be more significant than for other pressure-driven
processes [3]. NF behaviour for charged molecules can be described by the
extended Nernst-Plank equation [16], which is given by:
dCi ziCi Di , p d< (2.24)

Ji Di , p F K i ,cCiV
dx RT dx
where Ji is the flux of an ion i, Di,p is the ion diffusivity in the membrane, R is the
gas constant, F LVWKH)DUDGD\FRQVWDQWȌLVWKHHOHFWULFDOSRWHQWLDODQGKi,c is the
22
convective hindrance factor in the membrane. The most common model in the
literature, the Donnan steric partioning model (DSPM), first put forward by
Bowen and Mukhtar [17], is based upon Eq. (2.24). In the DSPM, rejection occurs
via both steric effects (when the molecule is physically excluded from the pore)
and by electrical (charge) effects, and the membrane is characterised by three
parameters: the membrane pore radius, the effective thickness/porosity of the
membrane, and the electrical charge inherent to the membrane. The DSPM is
described in detail later in Section 5.3.3. Further extensions to the DSPM have
also been made with the incorporation of dielectric exclusion effects [18, 19], by
which the dielectric constant within the membrane pores is less than that of the
bulk solution, though the details of these approaches are too involved to list here.
Finally, RO filtration models have been developed, which explicitly take into
account the effect of osmotic pressure. Notable models are summarised in Table
2.5.
23
Table 2.5 - Summary of RO filtration models [compiled from 3]
Preferential J B 'P '3 (2.25) ¨ȆLVWKHRVPRWLFSUHVVXUHGLIIHUHQFH

Sorption / across the membrane, and B is the
Capillary Flow pure water permeability of the
Model membrane. Based on the assumption
of a sorbed water layer at the
membrane surface, which controls
transport through the membrane.
Irreversible J LV 'P V'3W (2.26) LV is the hydrodynamic permeability

Thermodynamics of the membrane, LS¨ȆW is the
Model osmotic pressure difference between
the membrane wall and permeate, JS is
JS LS '3 1 V JCm.S (2.27) the solute flux, Cm,S is the average
solute concentration across the
membrane and ı is a reflection
'3 (2.28)
V{ for J 0 coefficient. Assumes coupled
'3 transport of solute and solvent.
Solution Diffusion DCm,W Vm,W (2.29) Cm,W is the concentration of water in

Model J 'P '3 the membrane, Vm,W is the partial
RT'x molar volume of water, ¨[ is the
membrane thickness, and k is a
kDM (2.30) distribution coefficient. Assumes that
J 'CS
'x solute and solvent dissolve in
membrane, and rejection of solute
depends on its ability to diffuse
through structured water inside the
membrane.
It is noted that these models generally consider only a single filtration mechanism
to be active at a given time, when in reality a multitude of separate models may be
required to adequately describe the overall behaviour of the membrane filter. In
addition, most models are only applicable to a single type of membrane process
(MF, UF, NF, RO). There is little work in the literature regarding coupling of
separate models to form a unified model [3].
24
2.3 Methods to reduce flux decline
A multitude of methods have been developed to attempt to reduce the flux decline
phenomenon, though the complex nature of fouling makes a general approach
difficult. The strategies employed fall largely into four groups [15]:
x pre-treatment of the membrane feed solution;
x modification of the membrane properties (surface and materials);
x control of membrane operating parameters; and
x membrane cleaning.
These methods are discussed in the following sections.
2.3.1 Pretreatment of feed solution
The pretreatment of feed solution to the membrane is an effective and well-
studied method of fouling control. Various pre-treatment methods have been used,
including:
x chlorination [e.g. 13, 20];
x coagulation [e.g. 21, 22, 23];
x pH adjustment [e.g. 21, 24, 25];
x activated carbon adsorption [e.g. 26, 27];
x chemical clarification [e.g. 28, 29];
x addition of complexing agents [e.g. 30]; and
25
x use of UF and MF as pre-treatment for RO processes [e.g. 31, 32, 33].
Pre-treatment methods have been extensively covered in the literature on
membrane fouling. For this reason, pre-treatment methods will not be studied in
this thesis. Thorough summaries of the topic may be found elsewhere [e.g. 15].
2.3.2 Modification of membrane properties
The properties of a membrane significantly influence fouling. Mulder [2] notes
that fouling is generally much more severe in porous membranes (UF, MF) than
dense membranes (RO), and that narrow pore size distributions may also reduce
fouling. Attempts have also been made to alter the surface chemistry of the
membrane to discourage fouling as described by Ma et al. [34], including:
x coating of water soluble polymers or charged surfactants onto the
membrane surface;
x forming ultrathin films on the membrane surface using Langmuir-Blodgett
techniques;
x coating of hydrophilic polymers onto the membrane surface; and
x grafting of monomers onto the membrane by electron beam or ultraviolet
radiation.
Details of these processes are beyond the scope of this discussion. An excellent
summary can be found in Hilal et al. [15]. Because of the considerable work in the
literature on this topic and the complexity of membrane materials science and
26
surface chemistry, not to mention the specialised apparatus required for
experimental study, membrane modification methods will not be considered in
this thesis.
2.3.3 Control of membrane operating parameters
Substantial reductions in membrane fouling and flux decline can be made by
manipulation of the operating parameters of the membrane process. This is
typically done with the aim of minimising the effect of concentration polarisation.
This can be achieved by either reducing membrane flux (which is usually
undesirable), or by increasing the mass transfer coefficient k of the system,
defined as:
DS (2.31)
k
G
The mass transfer coefficient characterises the hydrodynamics of the membrane
filtration system, i.e. the value of k is governed by the fluid properties (viscosity,
density, diffusivity), the fluid velocity, and the membrane geometry [2]. The mass
transfer coefficient can be related to some well known dimensionless numbers as
follows:
d (2.32)
kd h b c§
d ·
Sh a Re Sc ¨ h ¸
DS © L¹
27
dhv (2.33)
Re
Q
Q (2.34)
Sc
DS
where Sh is the Sherwood number, Re is the Reynolds number, Sc is the Schmidt
number, a,b,c and d are constants, dh is the hydraulic diameter, L is the length of
the channel, v is the flow velocity, and Ȟis the kinematic viscosity. Generally, an
increase in Reynolds number leads to an increase in the mass transfer coefficient,
due to increased local shear stress on the membrane surface which reduces
concentration polarisation and cake formation. The majority of methods to reduce
concentration polarisation thus focus on promoting turbulent conditions to
maximise shear stress and hence mass transfer. A brief summary of these methods
follows below.
2.3.3.1 Turbulence promoters
Turbulence promoters, or spacers, are a common method of improving membrane
mass transfer. However, the term is slightly misleading, as these devices do not
necessarily induce fully turbulent flow [35]. Spacers work by inducing increased
frictional and pressure losses by increasing resistance to axial flow, as well as
forming localised areas of stagnation (‘dead zones’). Spacers may be
manufactured in several forms, but generally consist of a net-like arrangements of
filaments, where flow disturbances are introduced at the filaments which cross the
28
axial flow direction (either transversely or at some angle). The spacer net can
either be woven, or non-woven, and are typically employed in spiral-wound
membrane modules, as shown in Figure 2.4.
To date, research into spacer design and optimisation has been quite extensive.
Numerous studies have been performed both experimentally [e.g. 36, 37-41] and
computationally [e.g. 42, 43-46] on this topic. It is considered by some authors
[35] that current spacer designs are so well optimised that little progress remains
to be made in this field. While this thesis will not specifically consider spacer
design, modelling of the flow through spacer-filled membrane modules will be
performed as part of developing a general model of the membrane filtration
process.
29
(a)
(b)
Figure 2.4 - Typical net-type spacer geometry, showing (a) use in spiral-wound membrane [47], and (b)
woven and non-woven spacers [35].
2.3.3.2 Gas sparging and bubbling
The injection of air or gases into the membrane feed solution has been
demonstrated by several authors to give significant improvements in membrane
flux [e.g. 48, 49, 50]. A range of flow regimes exists for such two-phase flows: at
high gas flow rates, mist flow or annular flow regimes are observed, and for low
30
gas flow rates, bubble flow is observed [51]. Between these two extremes lie
churning flow and slug flow, which is characterised by a quasi-periodic series of
elongated bullet-shaped bubbles known as Taylor bubbles [52]. The different flow
regimes are illustrated in Figure 2.5.
Figure 2.5 - Flow regimes for gas-injected flow through membrane [35]
The flux enhancement due to gas sparging has been attributed to several
mechanisms, primarily physical displacement of the concentration polarisation
boundary layer, and a reduction in the hydrodynamic resistance of the boundary
layer [49]. However, other significant mechanisms include pressure pulsing
caused by gas slugs, and bubble-induced secondary flow (i.e. flow around the
bubbles, which causes turbulence in the bubble wake) [53]. A significant number
of hydrodynamic and computational studies have been performed on this topic
[e.g. 51, 52, 54, 55, 56], though nearly all have focused exclusively on UF
processes. (These studies will be discussed in greater detail in Chapter 3). While
31
these techniques will not be specifically modelled in the work in this thesis, it
represents a feasible extension of the models developed herein for future work.
2.3.3.3 Dean vortices
Similar to the turbulence promoters discussed previously, Dean vortices have
been suggested as another method for increasing flux by minimising concentration
polarisation. Dean vortices are secondary flows which occur in flow through
coiled pipes due to centrifugal forcing. However, they may also be formed by use
of helical screw thread shaped inserts in tubular membranes, as shown in Figure
2.6. Bellhouse et al. [57] demonstrated substantial improvements in membrane
flux using these inserts both experimentally for UF, MF, and NF processes, and
using computational methods to predict the formation of the Dean vortices.
Figure 2.6 - Helical screw thread inserts used to induce Dean vortices (left: one start thread, right: three start
thread) [57]
Other authors have also demonstrated that the action of Dean vortices in curved
membrane tubes (typically helically or sinusoidally curved) is even more effective
32
in reducing the effects of concentration polarisation [e.g. 58, 59-61]. However, at
present such membrane geometries are not commercially available. Again, while
not specifically considered in this thesis, this remains another area that could in
future be investigated through use of the general computational models to be
developed in this thesis.
2.3.3.4 Pulsatile flow
The use of pulsatile flows (feed flows varying in velocity or pressure with time)
has also been demonstrated to enhance permeate flux. A variety of methods have
been used to generate pulsatile flows, including:
x collapsible-tube oscillators [62];
x infrasonic pulse generation [63];
x insertion of rotating perforated discs into the entrance of tubular
membranes [64]; and
x pneumatically activated jack / piston arrangements [65].
A number of mechanisms have been suggested for the flux increase afforded by
pulsatile flow, including increased shear, increased radial transport of particles
away from the membrane surface, backflushing though pores, cavitation, and
membrane flexing [62]. The shape of the pulsatile waveform has been observed to
significantly affect the efficiency of the process, with negative pressures (i.e.
backflushing) necessary to achieve the greatest reduction in membrane fouling
33
[66]. While a fair amount of research into pulsatile flows has occurred, the
majority of work has been experimental. There is some potential for further work
in this area using computational methods to further characterise and optimise the
mechanisms of flux enhancement by pulsation, though this is not specifically
modelled in this work.
2.3.3.5 Ultrasonic methods
The passing of ultrasonic waves through the filtration solution has been suggested
as a means for minimising membrane flux decline. As ultrasonic waves of
sufficiently high power are propagated through a liquid, cavitation bubbles can
occur, typically at frequencies of 20-1000 kHz [67]. Each cavitation bubble can
generate local temperatures of 4000-6000K and pressures of 100-200 MPa. The
inrush of fluid to fill the collapsed bubble forms a targeted jet which can be used
to increase shear stress at the membrane surface and hence minimise
concentration polarisation. A range of other potential mechanisms for particle
removal or detachment have been suggested, including cavitational phenomena
(microstreaming), acoustic streaming, and micro-jetting, which are shown in
Figure 2.7.
Most applications of ultrasonics to membrane filtration have focused on
membrane cleaning after filtration has occurred [e.g. 68, 69-71]. However, several
investigations have also been made into using ultrasonics during the filtration
process to reduce fouling with encouraging results [e.g. 72, 73]. However, there
34
has been experimental evidence of damage to the membrane surface by ultrasonic
treatment [e.g. 74, 75] . Overall, the use of ultrasonics in membrane filtration is
relatively unstudied and could be a potential application of a computational
membrane filtration model, though not considered further in this thesis.
Figure 2.7 - Particle detachment mechanisms for ultrasonic agitation [68]
2.3.4 Membrane cleaning
The methods outlined above are generally concerned with preventing or
mitigating the effects of flux decline via minimisation of concentration
35
polarisation and membrane fouling. Membrane cleaning methods, on the other
hand, are essentially a remedial method for dealing with the after-effects of
membrane fouling. It is considered in this case that the adage ‘prevention is better
than cure’ is fairly accurate; hence, this study will not consider membrane
cleaning methods in any great detail. Concise summaries of the common
membrane cleaning methods can be found elsewhere [e.g. 11, 76].
2.4 Summary and directions for research
It is apparent that considerable research effort has been invested on describing the
effects of various parameters on membrane flux decline. However, there is still
substantial work required to gain a full understanding of the fundamental
mechanisms of flux decline, including concentration polarisation and membrane
fouling. This is largely due to the complexity and variability of membrane
processes, and the difficulty in using analytical mathematical models to describe
these systems. The use of computational methods for simulating the mechanisms
of flux decline therefore represents an important direction for future study.
Computational simulations of the membrane filtration process also offer a
relatively inexpensive option for developing and optimising methods to reduce the
incidence of flux decline. A number of methods have been suggested or are
currently practiced, including:
x feed pretreatment;
36
x modification of membrane properties;
x membrane cleaning methods;
x control of membrane process operating parameters.
The first three classes of methods will not be considered in this study, as they
have generally been well-studied, are not always applicable to the general
problem of membrane flux decline, or are not suited to computational simulation.
However, control of process operating parameters appears to be a potentially
workable solution for minimising membrane flux decline. A valuable method for
investigating the control of these parameters is the use of computational
simulation using computation fluid dynamics techniques, which are discussed in
detail in the following chapter.
37
Chapter 3
The use of computational fluid dynamics in membrane
filtration
Computational fluid dynamics (CFD) refers to the use of numerical methods with
digital computers to solve fluid flow problems. The basic premise of CFD is
replacing the partial differential equations which govern viscous fluid flow with
discretised algebraic equations that approximate the partial differential equations
[77]. These equations are then solved numerically to obtain solutions for problems
where analytical solutions to the governing equations are not available.
CFD methods have relatively recently become a viable tool for analysis of
membrane filtration systems, due to increases in computing performance and
developments in numerical and mathematical methods. This chapter gives a brief
overview of the basic principles of CFD and current applications in fluid
dynamics, then discusses in detail various approaches to describing the behaviour
of membrane filtration systems using CFD.
38
3.1 Basic principles of CFD
Viscous fluid flow is governed by the basic principles of conservation of mass,
and conservation of linear momentum, which can be described by the following
equations, respectively:
wU (3.1)
Uv 0
wt
w Uv wP (3.2)
Uvu Pv SM y
wt wy
where ȡ is the density of the fluid, v is the component of the fluid velocity in the x
direction, and SMy is a source term in the y direction [35].
To solve the partial differential equations governing fluid flow, appropriate
numerical discretisation schemes must be employed. The three main classes of
methods are finite difference methods (FDM), finite element methods (FEM) and
finite volume methods (FVM). A detailed discussion of these methods can be
found elsewhere [78-80]. The main characteristics of each method are listed in
Table 3.1.
39
Table 3.1- Summary of characteristics of discretisation methods [compiled from 79]
Method Formulation considerations Mesh considerations Boundary condition

considerations
FDM Conceptually easy to Meshes must be Neumann boundary conditions

formulate structured in two or three (derivatives of variables specified
dimensions. Curvilinear at boundaries) cannot be strictly
meshes must be enforced, only approximated.
transformed to structured
Cartesian coordinates.
More difficult to
accommodate complex
geometries.
FEM Formulation less straight Easier to accommodate Neumann boundary conditions

forward complex geometries. are exactly enforced.
Non-Cartesian
coordinates and
unstructured meshes
possible.
FVM Can be formulated based on Depends on formulation Neumann boundary conditions

FEM or FDM (FEM or FDM). are exactly enforced.
Theoretically, it is possible to solve the governing equations of viscous flow via
computational methods for both laminar and turbulent flow regimes. However, in
turbulent flows, there is a much larger range of length and time scales than in
laminar flows, making the equations for turbulent flow more difficult and
computationally expensive to solve [80]. The problem of turbulence can be
approached in three main ways [79]:
x direct numerical simulation (DNS), where refined computational meshes
are used to resolve all scales of turbulence, large and small. This is
generally not feasible for problems of practical interest;
40
x Reynolds averaged Navier-Stokes (RANS) methods, where turbulent
motion is described in terms of time-averaged quantities, so that time-
dependent flow quantities are given in terms of a time-averaged mean
component and a time-varying fluctuating component. All scales of
turbulence are modelled (as opposed to directly computed) using these
methods;
x large eddy simulation (LES), where large scale turbulence is directly
computed and small-scale turbulence is modelled using time-averaged
methods. This method is becoming increasingly popular, as it offers
additional resolution of large-scale turbulence without the heavy
computational cost of DNS methods.
A detailed treatment of individual turbulence models is beyond the scope of this
thesis. Some of the further discussion will refer to certain turbulence models; the
reader is referred elsewhere for details [79, 80].
3.2 Modelling of membrane processes under laminar conditions
3.2.1 Flow through porous membranes
Investigations into the hydrodynamics of membrane channels largely began with
Berman’s [81] study of laminar flow in a rectangular channel with porous walls,
in which expressions were derived for fluid velocity and pressure as functions of
the channel dimensions, fluid properties and position within the channel. The
41
expressions were derived from the Navier-Stokes equations with the following
assumptions: (1) a steady state prevails; (2) the fluid is incompressible; (3) no
external forces act on the fluid; (4) the flow is laminar and (5) the velocity of the
fluid leaving the walls of the channel is independent of position [81]. The primary
results derived were that pressure drop in a porous channel is significantly less
than a non-porous channel, and that the velocity profile observed within the
channel was flatter near the centre and more curved near the walls than the
classical parabolic profile of Poiseuille flow.
Later studies were conducted by Friedman and Gillis [82] for a porous tube with a
viscous fluid, which were further refined by Mizushina et al. [83] for the case of
variable radial mass flux, and by Galowin and De Santis [84] for the case of
variable wall suction [all cited in 35]. Berwin’s approach was also extended by
Karode [85], who assumed that the velocity of fluid exiting the porous wall was
not constant, but was proportional to the local TMP. This revised analytical
solution showed good agreement with Berwin’s results for the case of constant
wall velocity.
Recent numerical treatments of laminar flow through a porous membrane have
followed a similar approach, by using the Navier-Stokes equations to describe the
motion of the fluid through the main ‘channel’ of the membrane module, and the
Darcy equation to describe the fluid flow through the porous walls of the
membrane module. For the case of flow through a cylindrical tube, the continuity
and Navier-Stokes equations can be written in cylindrical coordinates as:
42
wvr vr wvz (3.3)
0
wr r wz
wvr wv wp w § wv · (3.4)
Uvr Uv z z ¨ 2K r ¸
wr wz wr wr © wr ¹
wvr v w ª § wv wv ·º
2K 2K r2 «K ¨ z r ¸» Ug r 0
wr r wz ¬ © wr wz ¹¼
wv z wv wp w § wv z · (3.5)
Uvr Uv z z ¨ 2K ¸
wr wz wz wz © wz ¹
K § wv wv · w ª § wvz wvr ·º
¨ z r ¸ «K ¨ ¸» Ug z 0
r © wr wz ¹ wr ¬ © wr wz ¹¼
where r and z represent the radial and axial coordinates, respectively.
Similarly, the Darcy equation can be written as:
Pf wp (3.6)
vr 0
Lr wr
Pf wp (3.7)
vz 0
Lz wz
where μf is the viscosity of the filtrate, and Lr and Lv are the permeability of the
porous wall in the radial and axial directions, respectively.
An example of this approach is the work by Nassehi [86], who presented a
method for coupling the two sets of equations using FEM. The lack of second-
order derivatives in the Darcy equation (which are present in the Navier-Stokes
equations) presents difficulties in coupling the two sets of equations. Hence, the
43
Brinkmann equation which does contain second-order terms (shown below) has
sometimes been used in place of the Darcy equation, though this is not ideal, as
the Brinkmann equation is limited to high porosities generally not found in
membrane modules. The Brinkmann equation is given by:
P (3.8)
'P P 2v v
K
Nassehi resolved this issue by developing a Galerkin finite element scheme based
on the use of C0 Lagrange elements, in which the layer of elements joining the
porous wall to the rest of the solution domain are cast in such a form to represent
a combined flow field. That is, elements in the stiffness matrices corresponding to
the nodes located on the porous wall are replaced by the discretised form of the
Darcy equation, while the remainder of the matrices remain as per the Navier-
Stokes equations (the reader is referred to Nassehi’s paper for further details of the
finite element scheme). Results demonstrated the scheme was flexible in
application and accurately preserved flow continuity.
Another method of coupling the Navier-Stokes and Darcy equations was
demonstrated by Damak et al. [87] by assuming the permeability of the wall Lv to
be the same in the axial and radial direction, and assuming the viscosity of the
inflow and filtrate to be identical (denoted by Ș). This allows the flow through the
porous wall to be taken simply as a boundary condition for the Navier-Stokes
equations governing the free flow through the tube. The boundary condition can
be written as:
44
Lv 'P (3.9)
vr , vz 0 at r R, and 0 d z d L
K e
where e is the thickness of the porous wall, and R and L are the radius and length
of the tube, respectively. The justification for setting the axial velocity at the wall
to zero is given by work carried out by Schmitz and Prat [cited in 87, 88], which
demonstrated that the slip velocity at the porous wall surface was negligible. The
Navier-Stokes equations can then be solved by an implicit finite difference
method. The model was verified by comparison against velocity profiles
experimentally measured by Gouverneur [cited in 87, 89], and showed good
agreement. The results also demonstrated that the deviation for the parabolic
Poiseuille profile previously predicted by Berman [81] was directly dependent on
the wall suction (i.e. the local TMP).
3.2.2 Concentration-polarisation effects
The inclusion of concentration polarisation effects into the treatment of porous
walls introduces additional difficulties. The boundary layer formed by
concentration polarisation can be viewed as a kind of second porous wall with a
lower permeability than the membrane. The primary difficulty in modelling this
situation is determining appropriate concentration boundary conditions at the wall,
as concentrations will be continually changing at the wall, and the wall geometry
itself may change over time due to particle deposition. Several approaches have
been developed to account for this, and are discussed below.
45
A relatively simple approach to this problem was used by Huang and Morrissey
[90]. They used FEM to simulate the development of the concentration
polarisation boundary layer in a rectangular crossflow ultrafiltration membrane by
solving the Navier-Stokes equations to determine fluid velocities in a manner
similar to Berman [81]. These were then used to solve the convective-diffusive
mass transfer equation using FEM, which describes the formation of the boundary
layer:
wC v wC D w 2C (3.10)
u
wx H wO H 2 wO2
with the boundary C C0 at x 0 (3.11)
conditions
wC (3.12)
0 at O 0
wO
wC JW C (3.13)
at O 1
wO D
where JW is defined as per Equation (2.6). The finite element package Pdease2D®
was used, with a fine mesh near the membrane surface to accommodate the large
concentration gradient in the vicinity of the boundary layer, as shown in Figure
3.1. Results from the simulations were compared to experimental results using
bovine serum albumin (BSA) solutions. The results demonstrated a linear
relationship between the diffusion coefficient and the thickness of the
concentration polarisation boundary layer, as predicted by the classical Film
Theory and Gel Polarisation models. The model also corresponded well with
46
experimental predictions for the mass transfer coefficient and concentration on the
membrane surface.
Figure 3.1 - Finite element mesh used to model concentration polarisation phenomena [90]
Richardson and Nassehi [91] used a more sophisticated approach by assuming
that the solution does not remain evenly mixed in domains with porous walls. At
each time step, the physical properties of the fluid (viscosity and density) are
updated using relations for the concentration calculated at each time step, as
shown in Figure 3.2. In their model, the fluid flow in the bulk of the channel is
described by the Navier-Stokes equations, while the mass transfer through the
porous wall is described by the two-dimensional convective dispersion equation,
which is written as:
47
wC wC wC ªw º wC ª w º wC (3.14)
wt
vx
wx
vy
wy « wx DC » wx « wy DC » wy
¬ ¼ ¬ ¼
The solution of the flow and convective dispersion equations is then carried out in
a decoupled fashion (i.e. in separate steps, with the Navier-Stokes equations
solved to provide velocity and pressure fields, with the velocity field then used to
solve the convective dispersion equation). The density and viscosity are then
updated using current values of velocity and concentration, and the process
repeated until the solution converges. The density is calculated in terms of a solids
YROXPH IUDFWLRQ ĭ ZKLFK Fan be determined from the concentration, and the
densities of the solid and fluid (ȡs and ȡf, respectively), as written below:
U=U s )+U f ( 1-)) (3.15)
The viscosity is calculated by a two-step process, where the viscosity is first
updated using the following power law model:
K=K oJ n 1 (3.16)
where Șo is the consistency coefficient, n is the power law index, and J is the
shear rate. The viscosity is then updated based on the solids volume fraction via a
relationship such as the following:
Kf (3.17)
=1 /)
K
48
where Șf LVWKHYLVFRVLW\RIWKHFDUU\LQJIOXLGDQGȁLVDQHPSLULFal factor found
experimentally. The equations were solved by use of a streamline upwind Petrov-
Galerkin FEM technique, details of which can be found elsewhere. Results
indicated the model was capable of predicting both flow and mass transport
patterns for a variety of domains including flat and curved porous walls.
Figure 3.2 - Solution algorithm proposed by Richardson and Nassehi [91]
49
3.2.3 Cake formation effects
Geissler and Werner [92] simulated crossflow filtration in a flat-sheet membrane
system considering the formation of a cake layer upon the membrane. The growth
of the cake layer was modelled by considering both the convective transport of
particles towards the membrane and the diffusive transport in the other direction
due to concentration polarisation. The following assumptions were made in
developing the model:
x the flow is two-dimensional, incompressible, laminar, and Newtonian,
characterised by a local suspension viscosity Șsus;
x the particles within the fluid are neutrally buoyant;
x the permeate contains no particles; and
x the cake layer formed is immobile.
The equations thus derived to describe the particle transport within the system are
as follows:
w § * wC · w § * wC ·
(3.18)
¨ uC D ¸ ¨¨ vC D ¸ 0
wx © wx ¹ wy © wy ¸¹
2 wu (3.19)
D* Dˆ Iv rp
wy
'Px (3.20)
vF x
K >RM Rk x @
50
G k x (3.21)
Rk x
Kk
dG k x
>dm x dm
k x @dt
dif (3.22)
U p 1 H k dAx
where D* is the shear-induced diffusion coefficient, Dˆ Iv is the dimensionless
diffusion coefficient, rp is the particle radius, u and v are the velocities in the axial
and normal dimensions with respect to the membrane, Rk is the hydraulic
resistance of the cake, vF is the permeate velocity, Kk is the cake permeability, įk
is the local cake height, İk is the cake porosity, m is the mass flux and the
subscripts k and dif represent the cake and diffusion, respectively.
The model was verified against experimental observations of permeate flux and
cake layer height, and showed good agreement (within 15% of predicted values).
However, the model requires empirical parameters (particularly the hydraulic
resistance of the cake layer) to be determined by analysis of experimental data in
order to achieve useful results.
Lee and Clark [93] developed a more advanced model of cake formation for a
rectangular crossflow ultrafiltration system. Again, mass transfer is expressed by
the two-dimensional convective dispersion equation, and the steady-state
concentration field can be obtained by solving the following equations:
51
wC wC w 2C (3.23)
u v Dy
wx wy wy 2
with the boundary C x 0, y CB (3.24)
conditions:
§ wC · (3.25)
¨¨ ¸¸ 0
© wy ¹ y 0
§ wC · (3.26)
¨¨ vC D y ¸ 0
© wy ¸¹ y H0
where the half height of the channel is given by H0, and Dy is the diffusion
coefficient in the transverse direction. Equation (3.26) can be interpreted as
meaning there is no particle accumulation on the membrane surface (i.e. a steady-
state has been reached), which is valid for a membrane with complete rejection.
The steady-state model is then adapted to provide a step-wise pseudo-steady-state
model of the flux decline due to cake formation. This is done by assuming the
particle concentration at the membrane surface CW does not exceed a limiting
value Cmax U p 1 H k , and that the excess of particles will combine to form a
cake layer at the surface.
The solution algorithm proposed is as follows; first, the concentration profile is
obtained by solving the steady-state model using the velocity field at the previous
time step. Second, it is determined whether flux decline occurs by comparing the
concentration profile with the calculated Cmax – if greater than Cmax, a cake layer
52
will be formed at that location. The cake growth is calculated by a mass balance
between the mass convection rate and the mass accumulation rate at the
membrane surface. The permeate flux decline is then calculated using a
resistance-in-series model similar to Equation (2.8), so the local permeate flux is
known at each axial location. Third, a new velocity field is calculated using the
updated permeate flux as a boundary condition, and concentration profiles are
again calculated using the steady-state model. The second and third steps are then
repeated until CW d Cmax at every axial location. The solution algorithm is
summarised in Figure 3.3.
53
Figure 3.3 - Solution algorithm for Lee and Clark [93] cake formation model
The model results presented by Lee and Clark showed good agreement with
experimental results; however the model still requires the input of the specific
resistance of the cake deposit, which can only be determined by dead-end
filtration experiments.
54
3.2.4 The influence of osmotic pressure
Osmotic pressure plays an important role in RO and NF membranes. Several
authors have attempted to include the effects of osmotic pressure within CFD
models, by incorporating boundary conditions for the porous membrane wall
which can be expressed as a function of the osmotic pressure.
Geraldes et al. [94] developed a numerical model using FVM to simulate NF
processes in both spiral-wound and plate-and-frame systems, which can be
modelled as a slit-type channel configuration (i.e. the height of the channel is very
small compared to the length and width of the channel). These geometries mean
that the effect of the concentration polarisation boundary layer is a significant
operating issue. The model developed by Geraldes et al. [94] uses CFD for the
modelling of the fluid phase, with boundary conditions that take into account the
solute transport occurring within the membrane, including osmotic pressure
effects. The transport equations used for modelling of the fluid were as follows:
w Uu w Uv (3.27
0
wx wy
)
w Uuu w Uvu wp w § wu · w ª § wu wv ·º 2 w ª § wu wv ·º (3.28

2 ¨ P ¸ « P ¨¨ ¸» «P ¨ ¸»
wx wy wx wx © wx ¹ wy ¬ © wy wx ¸¹¼ 3 wx ¬ ¨© wx wy ¸¹¼
)
w Uuv w Uvv wp w § wv · w ª § wu wv ·º 2 w ª § wu wv ·º (3.29

2 ¨¨ P ¸¸ « P ¨¨ ¸» «P ¨ ¸»
wx wy wy wy © wy ¹ wx ¬ © wy wx ¸¹¼ 3 wy ¬ ¨© wx wy ¸¹¼
55
)
w UuZ A w UvZ A w ª § wZ ·º w ª § wZ ·º (3.30

« UD AB ¨ A ¸» « UD AB ¨¨ A ¸¸»
wx wy wx ¬ © wx ¹¼ wy ¬ © wy ¹¼ )
where ȦA is the mass fraction of species A within the solution, and DAB is the
binary mass-diffusion coefficient. Considering the channel to have length l and
half height h, the boundary conditions are given by:
x 0, y, u u0 , v 0, Z A Z A0 (3.31)
wu wv wZ A (3.32)
x l , y , 0, 0, 0
wx wx wx
wZ A (3.33)
y 0, x, u 0, Uv U p v p , UD AB U p v pZ A f c
wy
wZ A (3.34)
y 2h, x, u 0, v 0, 0
wy
where u0 is the uniform inlet velocity, ȦA0 is the uniform mass fraction at the inlet,
f’ is the intrinsic rejection coefficient and the p subscript denotes properties
relating to the permeate. The intrinsic rejection coefficient f’ can be described by
the Deen hindered transport model, which for the case of Pe << 1 can be
simplified as:
ª § 8u ·§ HD AB
0
·º
1 (3.35)
fc ¨
« ¨ 2 ¸¨
1 ¸¨ ¸»
¸
¬« © r0 ¹© 'P ¹¼»
56
where r0 is the average pore radius of the membrane, H is a hindrance factor for
solute diffusive transport and D AB

0
is the binary mass-diffusion coefficient at
infinite dilution. The permeate velocity vp is dependent upon the TMP and the
osmotic pressure. However, studies by other researchers [cited in 94, e.g. 95, and
96] have established that permeate velocity declines with increased solute
concentration at the membrane surface (i.e. in the presence of a concentration
polarisation layer). The authors thus introduced a concentration dependent
correction factor ȁ to account for this, in a manner similar to Rosén and Trägårdh
[cited in 94, 97]. The permeate velocity is then given by:
vp /Z Am Ah 'P '3 (3.36)
'3
3 Z Am Z Ap (3.37)
where Ah is the hydraulic permeability of the membrane, and the subscripts m and
p refer to the membrane surface and the permeate, respectively. The above
equations were discretised using a finite volume formulation. Comparison of the
computational results showed excellent agreement with experimental results.
Essentially identical approaches to describe the permeate velocity, albeit without
the use of a correction factor, were also employed by Fletcher and Wiley [98] and
Miranda and Campos [99].
57
3.3 Modelling of membrane processes under turbulent conditions
The studies previously mentioned in this chapter have all concerned laminar flows
within the membrane module. However, it has been widely suggested that
turbulent conditions do occur in practice in UF systems [5, 100]. Further, the
mixing afforded by turbulent flow may provide significant benefits in reducing
concentration polarisation effects. Investigation of turbulent membrane processes
is an obvious application of CFD, though it poses significantly more theoretical
difficulties than the laminar case.
One of the obvious questions relating to the use of CFD to simulate membrane
processes is the question of when turbulent models need to be employed.
Typically, flow in narrow channels remains laminar up to Reynolds numbers of
approximately 2000. However, studies have indicated that this is not the case for
some types of membrane modules; for example, Schwinge et al.[44] examined the
transition to turbulence (i.e. the appearance of unsteady flow structures) in spiral-
wound membrane modules, concluding that transition may occur in spacer-filled
channels at Reynolds numbers of 300 to 400. Ranade and Kumar [101] performed
direct numerical simulations (DNS) on spacer-filled channels, which indicated a
transition to turbulence at a Reynolds number of approximately 350. Belfort and
Nagata [102] also showed the transition to turbulence in tubular channels is
delayed to a Reynolds number of approximately 4000. There is limited
information in the literature for other types of membrane modules.
58
However, the appearance of unsteady flow structures and the transition to
turbulence does not necessarily mean that employing conventional turbulence
models is the best approach. Turbulence models are suitable only for high
Reynolds numbers (generally greater than 30 000), at which point turbulence can
be assumed to be fully developed and isotropic [103]. Pellerin et al. [5] provide
some examples of UF and MF flow with Reynolds numbers of 20 000 and 30 000,
respectively, based on the guidelines of Belfort and Nagata [102]. However, the
majority of membrane flows are below this level; for example, it is reported that
for spiral-wound membrane modules, most flows lie in the range NRe = 1000 to
3000 [e.g. 42, 100]. Given this fact, other methods to describe unsteady flows in
these flow regimes may be more suitable, such as direct numerical simulation
(DNS), and we discuss this subject in greater detail later in Section 4.3.1. We first
begin however with a review of studies using conventional turbulence models.
3.3.1 Studies using Reynolds averaged Navier-Stokes methods
Pellerin et al.[5] proposed a general hydrodynamic model with the ability to
incorporate turbulent effects, as well as complex geometries and pressure-related
boundary conditions (to describe the mass transfer through the membrane
surface).. The transport equation used can be written in general terms as:
w UM
UuM *M M
(3.38)
SM
wt
59
ZKHUHīĳ is the exchange coefficient, Sĳ is the volumetric source rate, and ĳis a
general field variable. Turbulent effects in the field variable are represented using
a RANS approach via a combination of a time-averaged component M and a time-
varying component M c :
M M Mc (3.39)
The two equation k-İ turbulence model was used to resolve the turbulent motion,
where k is the turbulent kinetic energy and İ is the turbulent energy dissipation
rate. Further discussion on the k-İ model can be found in a multitude of sources
[e.g. 5, 80]. Boundary conditions at the membrane wall are given by the
transmembrane velocity vw using the Darcy equation:
vw A'P V'3 (3.40)
where A is a membrane-specific coefficient. The reflection coefficient ı is a
measure of the rejection of the membrane, and was set to one (very high
rejection). Finally, the concentration field of the solute can be written in the
generalised transport equation format as:
ª §U ·º (3.41)
w « U ¨¨ k ¸¸»
¬ © U ¹¼ ª §U
« Uu ¨¨ k
· §U
¸¸ DU¨¨ k
·º
¸¸» 0
wt ¬ © U ¹ © U ¹¼
where ȡk is the density fraction of species k. The above equations were discretised
in finite volume form using the computer code TURCOM, and simulations were
60
run to represent RO, UF and MF conditions. The results showed good agreement
with classical analytical solutions [5, 81, 104] for low Reynolds numbers.
However, significant deviation from the velocity profiles in the analytical
solutions was observed for high Reynolds number cases (as high as 50 000), as
shown in Figure 3.4.
(a) (b)
Figure 3.4 - Deviation of analytical and numerical velocity profiles for Pellerin et al.'s model for (a)
ultrafiltration, (b) microfiltration [5].
Ranade and Kumar [43] also carried out simulations using the k-İ turbulence
model to investigate spacer effects in spiral-wound membranes. The turbulence
model was only employed for Reynolds numbers greater than 300, based on
recommendations from the literature and the authors’ own direct numerical
simulations. The commercial package FLUENT was used for all simulations. The
61
authors successfully managed to model a variety of spacer geometries, with
adequate agreement with experimental results reported by Da Costa et al. [cited in
43, 105, 106]. They also noted that full CFD simulations of an entire membrane
module under turbulent conditions remain very computationally expensive, but
good results could be obtained by considering a representative unit ‘cell’ of the
module.
Cao et al. [107] also employed a RANS approach in simulating turbulence effects
caused by net-type spacers in a narrow channel. The authors made use of the RNG
k-İ turbulence model, which differs slightly in formulation from the classic k-İ
model used by Pellerin et al. [5]. The RNG model aims to describe the effects of
small eddies in terms of larger motions and a modified viscosity term in order to
eliminate these small-scale eddies from the governing equations. The RNG model
is particularly effective when small-scale flows are significantly anisotropic, as
such anisotropic flows tend to either become decoupled from the large-scale flow
or give their energy to large-scale eddies, rather than interacting directly with the
large-scale flow by increasing the dissipation of large eddies, as was demonstrated
by Sivashinsky and Yakhot [108]. The region adjacent to the membrane surface is
typically highly anisotropic, making the RNG approach suitable to this
application. Simulations were carried out with the FLUENT package for a variety
of geometric spacer configurations, details of which are not necessary to be
presented here.
62
3.3.2 Studies using direct numerical simulation methods
Direct numerical simulation (DNS) methods allow small-scale and large-scale
turbulent effects to be calculated directly, without resorting to the approximation
of turbulence models. This is considered to be a more useful approach for most
membrane filtration conditions, where Reynolds numbers are sufficiently low
(generally less than 30 000) that turbulence is not fully developed, and
conventional turbulence models are not strictly applicable. However, this
approach is extremely computationally expensive for most problems of practical
interest. Nonetheless, several studies have been undertaken using these methods.
Miyake et al. [109] made use of a DNS approach, where all scales of turbulence
are directly computed, to simulate turbulent flow with a periodic pressure gradient
in a rectangular channel with one porous wall. Using spectral methods developed
by Kim et al. [cited in 109, 110] to solve the three-dimensional Navier-Stokes
equations, a flow was simulated with periodic pressure boundary conditions on
the porous wall. The study indicated that for this flow, good results (compared to
the direct numerical solution) could be obtained by large-eddy simulation (LES)
models, as the periodic pressure gradient has less influence on small-scale eddies
compared to large eddies. This result is interesting, given the relative scarcity of
LES models in the membrane modelling literature. However, the applicability of
this result to more realistic membrane operating conditions is uncertain.
63
More recently, several investigations have been made using commercial CFD
codes to perform transient simulations with very small timesteps and fine
computational meshes to resolve unsteady flows in membrane channels. Some
examples of this approach can be seen in studies by Schwinge et al. [44, 111, 112]
and Fimbres-Weihs et al. [113, 114] using the ANSYS CFX code, and Koutsou et
al. [42] using FLUENT. These models were used to study the effect of spacers in
rectangular membrane channels, with all cases considered in these studies
simplified by the assumption of impermeable walls or the use of a ‘dissolving-
wall’ boundary condition, where the wall concentration is set to a high fixed value
to simulate the effects of concentration polarisation. Koutsou et al [42] found that
instabilities in the vicinity of the cylinder began to occur at a Reynolds number of
60, and further instabilities in the form of wall eddies began to occur at a
Reynolds number of 78. Schwinge et al.[44] also identified transition to unsteady
flow occurring at Reynolds numbers between 300 and 600, dependent on the
geometrical configuration of the spacers. These models provide useful insight into
the unsteady flow behaviour of these systems, but the simplified wall boundary
conditions employed are not entirely physically representative of a real membrane
system.
Newer studies have incorporated more realistic descriptions of membrane channel
permeation, such as that of Wardeh and Morvan [115]. This study simulated
unsteady flow in 2D spacer-filled channels for conditions similar to that of [44],
but incorporated a permeable wall condition using a Darcy’s law type expression.
64
It appears reasonable that further progress in this area should inevitably include
the effects of membrane permeation in this manner to develop a more fully
representative hydrodynamic model.
Finally, the simulation of flow in spacer-filled channels for these conditions has
also led to the use of 3D models becoming somewhat common. While 3D models
provide a more accurate representation of real flow conditions, they also incur
greatly increased computational expense. Consequently, initial studies used low
spatial resolutions [116] or neglected the membrane mass transport processes by
assuming impermeable walls [e.g. 101, 116, 117]. The most important and
common method of reducing computational expense for these 3D simulations
however is through use of a unit cell approach, using spatially periodic or
‘wrapped’ boundary conditions. This exploits the fact that spatially periodic (but
not necessarily temporally periodic) flow structures tend to occur in spacer-filled
channels between filaments, and hence simulation of only a single unit cell
between repeating filament structures is required. There are multiple examples of
this approach using the dissolving-wall boundary condition to ease computational
requirements [40, 118-121], with one study by Lau et al. [122] employing a true
permeable wall boundary condition. These studies have also almost universally
used the approach of direct numerical simulation of unsteady flows by use of
laminar models with very fine spatial and temporal discretisation. However, due
to the expense of 3D simulations, this approach is still limited to the unit-cell
approach, where spatially periodic flow structures occur. In this respect, direct
65
numerical simulation in this fashion remains more feasible in a general case by
use of 2D models.
3.3.3 Summary of turbulent membrane models
It is clear that nearly the entirety of research in the literature related to turbulence
modelling in membrane modules focuses on the modelling and optimisation of
spacers and turbulence promoters in spiral-wound membranes. In addition, most
applications of turbulence models have employed the standard k-İ model, or the
RNG k-İ model, which are suited only to high Reynolds numbers which may not
be common to a wide range of membrane applications. Little work has been done
in applying turbulent CFD techniques to other membrane configurations, and
correspondingly little work has been done using LES techniques, even though
these are becoming increasingly popular within the discipline of CFD at large.
However, the use of transient laminar flow simulations with very fine meshes and
timestepping has also been shown to well describe unsteady flow behaviour for
lower Reynolds numbers, which is where the majority of membrane applications
are likely to occur.
3.4 Prediction of membrane mass transfer
The majority of the CFD models discussed so far in this chapter have generally
attempted to model only the hydrodynamic conditions within the membrane
channel or module in question. Most neglect what is the most important property
66
of the membrane, which is the ability to selectively remove constituents from a
feed stream; that is, the transfer of mass through the membrane itself. A variety of
models have been developed to describe this mass transfer process, as outlined in
Section 2.2.2.4. By coupling a predictive mass transfer model to a CFD model, a
reasonably general model of the membrane filtration process would be potentially
realisable.
However, due to the different removal mechanisms in the different types of
pressure-driven membrane filtration, it is difficult for a single model to describe
membrane mass transfer generally. For the lower-pressure processes, where
particulate constituents in the feed stream are common and the membrane pores
are relatively large, one might surmise that it might be possible to explicitly model
the flow through the porous membrane microstructure within the CFD model
itself. That is, if a geometric model of the membrane pore structures could be
developed, standard Lagrangian particle transport models included in commercial
CFD packages could be used, in conjunction with some method for modelling
particle attachment and detachment within and on the membrane microstructure,
to predict the performance of these systems. In these particle transport models,
particles are essentially injected into the CFD flow domain, and then tracked
through the domain by solving differential equations describing the particulate
position, velocity, temperature and mass [123]. However, some serious difficulties
hinder this approach:
67
x The geometry of membrane microstructures is very complex, and can vary
significantly between different membranes, as well as spatially over a
single membrane, so it is difficult to construct a geometry for general use;
x Implementing a model to describe both particle-particle and particle-
membrane interactions is very difficult; a range of models have been
developed for other process engineering applications [123] but it is
unknown how applicable these are to membrane applications;
x Finally, and most significantly, the computational expense incurred by
such simulations is prohibitively large at present. This is especially so
when considering the combination of the complexity of particle-particle
interactions, particle-membrane interactions, and the membrane
microstructure geometry.
For now, at least, this method appears mostly infeasible. Considering the higher-
pressure processes (NF and RO), on the other hand, it is more conceivable that a
general predictive model may be developed. In these processes, the feed
constituents tend to be ions or molecules dissolved into solution, and the
membrane pores are relatively small. As these pore sizes begin to approach the
scale of individual water molecules, attempting to model the microstructure of the
membrane using CFD is futile, as conventional descriptions of hydrodynamics do
not apply at these scales. Instead, a more useful approach is to use a mathematical
model to predict the rejection of the solutes within the feed solution, which is then
applied as a boundary flux condition to a CFD model of the bulk of the membrane
68
channel. In this manner, the CFD model can be used to resolve the flow structures
and concentration polarisation effects within the membrane channel, even for
complex geometries, while the rejection model can be used to describe the mass
transfer through the membrane itself. A few attempts have been made using this
coupled CFD approach to describe membrane filtration, which are summarised in
the following.
Geraldes et al. [94] used a Deen hindered transport model to predict the rejection
of various charged and uncharged compounds in single-solute solutions by a NF
membrane in conjunction with a CFD model, as previously discussed in Section
3.2.4. In this approach, the mass transfer (rejection) of the solute depends on the
local pressure differential across the membrane, obtained from the CFD solution.
However, the rejection calculation, as per Eq. (3.35), depends on the fitted
parameter (8μ/r02)(HD0ABǻP), which is constant for a given membrane-single
component solution system. This makes the approach difficult to apply generally,
and even more so when multiple-component solutions are considered.
An additional significant difficulty in modelling multiple-component solutions in
CFD is describing the diffusion of each solute throughout the solution. The
classical way to deal with this in CFD is to use a modified version of Fick’s law,
where the rate of diffusion of each component is governed by a diffusivity
coefficient, assuming that each component in the solution diffuses independently
[103, 124, 125]. However, this method does not necessarily ensure
electroneutrality of the solution in the case of charged solutes. To overcome this, a
69
Fick’s law approach can be employed assuming that salts do not dissociate, and a
mean diffusion term is used for each individual salt.
In some cases, however, the diffusion of one component may depend on other
components. For this situation, a modification of Fick’s law was put forward by
Onsager [126], using a matrix of diffusion coefficients, effectively describing the
dependence of the diffusivity of each component on all other components. While
this approach is general and theoretically sound, determination of these diffusion
coefficient matrices is very difficult, and has only been performed for a handful of
solutions.
An alternative approach is the use of the Maxwell-Stefan diffusion model [103,
125], which considers diffusion to result from the superposition of different
frictional forces between the component molecules. The formulation of the
Maxwell-Stefan approach is complex and will not be treated in detail here, but has
been used for some membrane applications with success. However, it still
effectively requires use of a matrix of diffusion coefficients which are difficult to
determine, which makes it difficult to use for a general case.
However, simplification of the Maxwell-Stefan diffusion model leads to an
approach which has already been discussed in previous sections, which is the
Nernst-Planck model (see Section 2.2.2.4). As per Fick’s law, this requires only a
single diffusion coefficient for each component, but also takes into account
electrical effects [103]. Fimbres-Weihs and Wiley [103] describe a method using
70
this approach coupled with the commercial CFD package ANSYS CFX-10.0. This
package (and most other commercial packages) assume that multicomponent
species transport occurs via Fick’s law where each component diffuses
independently. This makes it impossible to directly specify the mass fluxes for
each component directly (via the extended Nernst-Planck equations). Instead, a
mass source term must be included in the model to cancel out the Fickian
diffusion term in the default transport equations, and then add the corresponding
diffusive term from the extended Nernst-Planck equations [103]. In this manner,
the components can be allowed to diffuse based on the diffusion rates of other
components, while guaranteeing the electroneutrality of the solution is conserved.
The Fickian diffusion value used in the inbuilt transport equations is thus a
dummy value, which should not affect the simulation as the term is cancelled out
by the additional source term [103].
Fimbres-Weihs and Wiley [103] then used this method to simulate the flow of a
NaCl:KCl solution through a 2D rectangular membrane channel. No-slip
boundary conditions were enforced along the channel walls, with either a fixed
wall concentration boundary condition or a constant flux boundary condition
imposed at the membrane surface. This method yielded converged solutions
which showed different behaviour to that predicted by the classcical Fick’s Law
approach. This is reflected in Figure 3.5, which shows the difference in ionic
concentrations within the boundary layer between the solution using the extended
Nernst-Planck diffusion model and the Fickian diffusion model. However, the use
71
of the more rigorous diffusion model presented some difficulties. In particular, the
inclusion of the source term into the governing equations for the commercial code
CFX for the entire flow domain caused the mass fraction transport equations to
become numerically stiff, particularly near the membrane surface. In addition, it
was found that the value of the dummy diffusion coefficient significantly affected
the simulation results, even though the term should have simply cancelled out
[103]. It is also probable that the inclusion of the source term and the associated
expense of solving the Nernst-Planck equations would have significantly
increased the overall computational expense. For these reasons, the authors
recommended that future work in coupling a multi-component diffusion model to
a CFD simulation over the entire flow domain should avoid the use of source
terms, and focus on instead directly specifying the diffusive terms within the
inbuilt transport equations. It is noted by the present author that to the best of their
knowledge this facility is not yet available in ANSYS CFX.
72
Figure 3.5 – Difference in ionic concentrations predicted in concentration polarisation boundary layer using
both Nernst-Planck and Fickian diffusion models [103]
Déon et al. [127] also adopted a coupled model of sorts, using the DSPM-DE (i.e.
including dielectric exclusion effects) to describe the mass transfer through the
membrane, and a 2D model using the extended Nernst-Planck equations to
describe the diffusion of multiple ions through the concentration polarisation
boundary layer. This approach is similar in principle to that of [103], except that
CFD is not used to determine the hydrodynamics of the membrane channel.
Instead, classical hydrodynamic equations for turbulent pipe flow are used to
determine the velocity profiles in the concentration polarisation layer. A similar
approach was also used by Bhattacharjee et al. [128], though using the
conventional DSPM and a linear velocity profile in the concentration polarisation
73
layer. This method was able to reproduce experimental rejection data well, and
has the advantages of providing a rigorous description of the transport of ions in a
multiple-ion solution both within the membrane itself and within the
concentration polarisation boundary layer. This is an improvement over the Fick’s
law approach, however, the formulation in [127] using classical 2D velocity
profiles within the concentration polarisation layers means the model is not
completely general in terms of the hydrodynamics. That is, in complex membrane
channel geometries, the concentration polarisation boundary layer hydrodynamics
may not follow the simplifying assumptions imposed by the use of turbulent pipe
flow velocity profiles. For this reason, the more general approach including the
coupled CFD model (capable of describing the hydrodynamics for arbitrary
geometries) used by Fimbres Weihs and Wiley [103] is potentially more widely
applicable. However, this method still suffers from numerical difficulties as the
extended Nernst-Planck equations become very stiff as discussed previously.
3.5 Comments on commercial membrane simulation packages
Several commercially available software packages for simulating membrane
performance exist, generally produced ‘in-house’ by major membrane
manufacturers. The most prominent examples of this are the ROPRO® software
developed by Koch Membrane Systems, the ROSA software developed by Dow
Water & Process Solutions, the TorayDS software developed by Toray
Membrane, and the IMSDesign software developed by Hydranautics [129].
74
As these packages are generally developed ‘in-house’ and are of significant
commercial value, there is relatively little information describing the internal
computational and numerical techniques implemented in these packages.
However, these packages in general are intended to aid process engineers in
designing appropriate membrane treatment trains, i.e. large-scale configurations of
membrane modules and arrays using ‘off-the-shelf’ commercially available
membrane modules (usually those of the manufacturer supplying the software). In
general, these packages provide a design solution encompassing array sizing,
operating pressures, scaling indices, and product and concentrate water qualities,
with predictions based on nominal membrane performance under steady-state
design conditions [129]. They are not intended for more fundamental studies of
membrane performance, either in terms of hydrodynamics or membrane mass
transfer. It is emphasised that the research direction in this thesis is not aimed at
recreating a general design tool such as these commercial packages, but rather to
provide a modelling tool which may reveal some more fundamental insights into
membrane channel behaviour, which in time may be able to contribute towards
describing membrane performance under highly variable conditions in such
design tools.
3.6 Summary and directions for research
It is apparent that CFD techniques are a useful and increasingly accessible method
for describing the behaviour of membrane filtration systems. However,
75
developing a general model for the pressure-driven membrane processes (MF, UF,
NF, and RO) presents some difficulties. The differences in removal mechanisms
between the lower-pressure processes (MF and UF) and the higher-pressure
processes (NF and RO) means that a general model of pollutant removal is
perhaps unattainable at present. In particular, the particulate nature of pollutants in
MF and UF processes requires a separate and considerably more expensive
approach to CFD modelling. In addition, the fouling behaviour of such systems is
also highly difficult to predict in a general sense.
On the other hand, if non-particulate pollutants only are considered (i.e. chemical
species which are mixed within the feed solution), then the possibility of a general
model becomes greater. Such a model would in principle be capable of describing
the high-pressure processes of NF and RO, as well as some UF processes. It is
also evident that these non-particulate pollutants will be present in the low-
pressure processes of MF and UF, though their removal will be significantly
lesser and their effect on membrane performance will be small. The converse
however is not true; high-pressure processes are generally not required to treat
such large pollutants which are removed by a range of pre-treatment processes.
In addition, modelling of a wide range of membrane processes introduces the
difficult topic of turbulence. While UF and MF processes may in some instances
operate in flow regimes where turbulence is fully developed (i.e. Reynolds
numbers greater than 30 000), the majority of processes lie in lower Reynolds
number regimes, where flow is either laminar (steady), or unsteady without
76
isotropic turbulence, and hence the use of conventional turbulence models cannot
be justified. These unsteady flows can be fully resolved, however, if appropriately
fine spatial and temporal discretisation is employed, which has been shown to be
feasible for conditions encountered in NF and RO.
The use of 2D versus 3D flow models is also of interest. 3D models should be
able to more fully describe the complex flow patterns in membrane modules, and
provide the most general implementation of the model possible. However, the
great computational expense incurred by 3D models has meant that they can only
be feasibly be employed for small representative portions of the membrane
channel where spatially periodic flow conditions exist, which is not the case in
general. The use of 2D models allows computational costs to be kept to a
reasonable level such that simulations can be performed for a general case, and
also include mass-transfer processes as discussed above.
Thus, the approach adopted in this thesis has been to develop a general
computational model for the filtration of feed solutions containing dissolved
solutes only. The model will include both a detailed description of the
hydrodynamic behaviour of the membrane for laminar conditions in 2D, including
the effects of concentration polarisation, and a model of the transport of these
dissolved solutes through the membrane itself. This differs from previous studies
in which these approaches have traditionally been separate, and allows for a
reasonably complete predictive model of the high-pressure processes of NF and
RO. With selection of appropriately fine computational meshes and timestep, the
77
laminar model is also able to directly resolve unsteady flows that may occur in
these high pressure processes, without recourse to turbulence models which are
generally not appropriate for these transitional flows. The use of 2D models also
makes the combined approach computationally feasible, and not restricted to
situations where spatially periodic flow structures exist, as in most other
applications of 3D simulations.
While it is not considered in this thesis, this approach can be generalised in the
future to address particulate modelling and turbulence modelling with a view
towards application to the low-pressure processes of MF and UF. In addition, the
approach can be easily generalised in the future to include 3D modelling as the
increases in computational resources make this a feasible prospect. Finally, such a
model could be used to investigate fouling reduction techniques such as those
discussed in Section 2.3.3, though this is also not addressed in this work.
78
Chapter 4
Hydrodynamic modelling of membrane filtration
4.1 Introduction
This chapter describes the development of a basic set of rigorous hydrodynamic
models of the membrane filtration process. The models are tested and verified
against experimental and computational data from the literature, and modelling
strategies and techniques particular to the problem of membrane filtration are
discussed. Several different geometric and flow configurations are investigated,
including 2D membrane geometries with and without hydrodynamic spacer
meshes, steady-state and transient simulations, as well as steady and unsteady
flow conditions. The simulations in this chapter are concerned primarily with the
describing the hydrodynamics of the membrane channel and the phenomena of
fouling due to concentration polarisation, but do not explicitly try to simulate the
mass transfer of solutes through the membrane structure itself. Thus, the purpose
of these simulations is to establish the usefulness of CFD as a basic modelling tool
for membrane filtration, which will be extended in Chapter 6 to provide a fully
predictive model of the membrane filtration process, with the addition of a
predictive rejection model to be described in Chapter 5.
79
4.2 2D flow through a porous membrane channel
The most basic hydrodynamic model of the membrane filtration process consists
of a two-dimensional (2D) rectangular membrane channel, with one or both walls
considered to be permeable. This configuration is general enough to locally
represent a range of membrane geometries for the purpose of investigation of CP
effects. The simplification to 2D is justifiable given the low aspect ratio of most
membrane feed channels, and is considerably less computationally expensive
[130]. The software package CFX 13.0 (Ansys, Inc., USA) is used for all
simulations of the hydrodynamics of the membrane channel. CFX is a general
purpose computational fluid dynamics package, which uses an element-based
finite volume method with a co-located grid layout such that the control volumes
are identical for all transport equations. A coupled solver is used, where the
equations for continuity and momentum are solved simultaneously as a single
system [123].
CFX is, in a strict sense, capable of three-dimensional (3D) modelling only;
however, 2D models may be constructed by the use of periodic pairing, which
creates symmetry boundary conditions on the lateral faces of the model which is
then only a single element thick. This effectively creates a 2D model, and makes it
trivial to extend the model later to three dimensions as necessary.
80
4.2.1 Model geometry
The model consists of a rectangular channel as shown in Figure 4.1. The flow
through the channel is oriented along the x-axis, and enters at x = 0 through the
inlet face. The flow exits the channel at x = L through the concentrate face, and
permeation occurs through either or both of the lower channel face at y = -h and
the upper face of the channel at y = h, which represents the membrane surfaces. It
should be noted that in Figure 4.1 that permeation is only shown through the
lower channel face for clarity. The left and right ‘walls’ of the channel are coupled
under the assumption of axisymmetric flow. The right wall of the channel is
located at z = -w/2, and the left face is located at z = +w/2. The width of the
channel w (i.e. the thickness of the single width element) typically should be of
the same order as the smallest cell within the face mesh, and was varied as such
depending on the face mesh size in later grid dependence studies. While simple
rectangular prismatic elements were used for this simple geometry, the model can
be generalised to any arbitrary 2D or 3D geometry.
81
y
feed flow
z x
permeate flow
w
2h
retentate flow
Figure 4.1 - Model geometry for basic 2D model of laminar flow through porous membrane channel (not to
scale)
4.2.2 Boundary and initial conditions
Boundary conditions are specified at the channel inlet (in the form of an axial
velocity profile) and the channel outlet (as an average static pressure). It is
assumed that a fully developed laminar flow profile exists at the channel inlet.
The bulk composition of the solution is also specified at the channel inlet in terms
of a solute mass fraction for each solute i, and initial conditions for the solution
are automatically generated by the CFX solver using this specified mass fraction
over the entirety of the membrane channel.
To represent the permeate flow through the membrane, a mass sink is applied at
the membrane surface. In this case, at each computational point along the
82
membrane surface a mass sink is applied as the product of the membrane flux as
defined by Eq. (4.1) and the local density of the fluid mixture:
JV LV 'P '3W (4.1)
where LV LV WKH K\GURG\QDPLF SHUPHDELOLW\ RI WKH PHPEUDQH DQG ǻȆW is the
osmotic pressure difference between the feed solution at the membrane surface
and the permeate. The TMP is not explicitly applied to the model, as doing so
decreases the resolution of small pressure gradients near the membrane surface
due to numerical rounding error, as has been noted previously [98]. The TMP is
instead simply applied as an input parameter to calculate the mass sink expression
representing the permeate flux. The solute mass fraction of the permeate is
specified at each computational point along the membrane surface according to
Eq. (4.2):
Zi , p Zi ,W 1 Ri (4.2)
where Ȧi,p and Ȧi,W are the mass fractions of solute i in the permeate stream, and
at the membrane surface, respectively, and Ri is the rejection of the solute i. The
rejection of a particular solute by a particular membrane is often determined
experimentally; however it would be desirable in many cases to predict the solute
rejection without recourse to experiment. A method for prediction of the
rejections Ri will be considered later in Chapter 5.
83
A symmetry condition is used for the outer ‘walls’ of the membrane channel
(those parallel to the x-y plane) to allow a 2D simulation. The remaining top and
bottom walls of the membrane channel are represented by using standard no-slip
boundary conditions in the x-direction, with a mass sink applied to represent the
permeation through the membrane surface as discussed previously. The velocity
in the y-direction is automatically determined by the CFX solver based upon the
local transport properties and the permeate flux imposed by the mass sink term,
which is a function of the local solute mass fraction and rejection.
4.2.3 Computational mesh, convergence and hydrodynamic verification
To ensure the basic hydrodynamics of the membrane channel were correct, the
model was first verified for the case where the solute has neutral properties, by
which we mean the solution viscosity, density, diffusivity, and osmotic pressure
are identical to those of water regardless of the solute concentration. The
analytical solutions of Berman [81] and Karode [85] where both upper and lower
walls of the membrane channel are permeable are used as verification cases for
the model. The model dimensions have been chosen as similar to those used by
Karode [85] for ease of verification. The relevant model input parameters are
shown in Table 4.1. The hydrodynamic behaviour of the model was verified by
observing the total axial pressure drop ǻ3axial between the membrane inlet and
outlet. The axial pressure drop was monitored to ensure that good agreement with
84
the analytical solutions was obtained, and to determine the mesh density at which
the calculated pressure drop did not change significantly.
Table 4.1 - Model input parameters for verification case
Parameter Value
Membrane channel length L 1m
Membrane channel half height h 1 mm
Membrane channel width w Variable depending on mesh density – equal to size of smallest mesh
element
Transmembrane pressure 3 bar
Membrane permeability LV 9.17 x 10-11 m Pa-1 s-1
Inlet Reynolds number NRe 250
Convergence was observed by monitoring the root mean squared (RMS) residuals
for mass, momentum, and solute mass fraction. The model was judged to be
converged once all residuals were smaller than 1 x 10-6. The hydrodynamic model
results indicated that good agreement with the analytical solutions was observed
using mesh resolutions in the order of 320-640 axial cells (in the x-direction) and
40 vertical cells (in the y-direction). The mesh size was then used as a starting
point for later refinement of the mesh needed to resolve the fine CP boundary
layer. A comparison of the non-dimensionalised axial pressure drop for the
converged numerical solution and the analytical solutions is shown in Figure 4.2.
Excellent agreement is shown between the CFD solution and the analytical
solution of Karode [85]; Berman’s solution [81] underpredicts the pressure drop
as this solution neglects the axial variation in permeate flux. The axial pressure
85
drop is, of course, lesser than that of an equivalent channel with impermeable
walls, shown for comparison.
50
CFD model
Dimensionless axial pressure drop
Impermeable walls
40
Berman (1953)
Karode (2001)
30
20
10
0
0 200 400 600 800 1000
Dimensionless axial length (x/h)
Figure 4.2 - Comparison of dimensionless axial pressure drop for CFD model with analytical solutions for
hydrodynamic verification of model
4.2.4 Incorporation of concentration polarisation effects
Concentration polarisation occurs when rejected solute accumulates next to the
membrane surface, forming a boundary layer [6]. This accumulation of solute
causes a diffusive flow back to the feed, which eventually reaches a steady-state
condition where the solute flow to the membrane surface equals the solute flux
through the membrane plus the diffusive flow from the membrane surface back to
the feed. The establishment of this boundary layer allows subsequent fouling of
86
the membrane layer to occur; minimisation or removal of the CP layer is thus a
primary goal of membrane research.
To model the effects of concentration polarisation at a basic level within the
membrane channel, the multicomponent flow option in CFX is used. This
assumes that the components of the fluid are mixed at a microscopic level so that
they share identical pressure, velocity and (where applicable) temperature fields,
and assumes that mass transfer occurs via convection and diffusion only. These
assumptions are reasonable for the majority of membrane filtration applications.
4.2.4.1 Spatial variation in transport properties
To describe the effects of concentration polarisation on the permeation of the
solute at varying points on the membrane surface, it is necessary for the transport
properties of the solution, such as diffusivity, density, and viscosity, to vary
spatially over the computational domain as a function of solute concentration A
simple approach and test case was derived from the work of Geraldes et al. [94] in
which the viscosity, diffusivity and osmotic pressure of aqueous solutions of
several compounds are described as polynomial functions of the solute mass
fraction. The compounds used in this study (sodium chloride, sodium sulphate,
sucrose and PEG1000) have been used as test compounds for the present model.
The solution density cannot be explicitly specified as a function of the solute mass
fractions due to software limitations, but is calculated using a harmonic average
87
based on the density of the individual components within the solution. The
osmotic pressure is calculated as an additional scalar parameter which varies
across the solution domain as a function of the solute mass fraction, and is then
used to determine a more accurate approximation of the permeate flux at the
membrane surface as per Eq. (4.1). The relevant expressions are shown in Table
4.2.
Table 4.2 - Trial solute components and expressions for transport properties (after [94])
Density Viscosity Diffusivity Osmotic pressure Ȇ

Compound
ȡ (g cm-³)* μ × 103 (Pa s) D × 109 (m2 sí1) × 10í5 (Pa)
1.61(1 í 14 Ȧi) for Ȧi
NaCl 2.165 0.89(1 + 1.63 Ȧi) < 0.006; 1.45 for Ȧi 805.1 Ȧi
>0.006
Na2SO4 2.664 0.89(1 + 3.52 Ȧi) 1.23(1 í 0.76 Ȧi 0.4) 337.8 Ȧi 0.95
72.18Ȧi(1 + 0.94Ȧi +
Sucrose 1.587 0.89(1 + 1.31Ȧi + 16.83Ȧ i 2) 0.52(1 í 1.33Ȧi)
2.93Ȧ i 2)
24.64 Ȧi (1 + 2.94 Ȧi
PEG1000 1.12 0.89(1 + 6.59 Ȧi + 120.8 Ȧi 3) 0.309
+ 19.25 Ȧi 2)
* Used for CFX internal calculation of solution density using harmonic average
4.2.4.2 Mesh refinement for resolution of CP layer
Significant mesh refinement is required in the vicinity of the membrane surface to
adequately resolve the fine CP boundary layer. Previous hydrodynamic models
had indicated cell aspect ratios of roughly 30 – 60 gave acceptable results. This
was used as a starting point for a grid independence study using the geometry of
Geraldes et al. [94], which is shown in Table 4.3. The vertical mesh resolution and
the degree of refinement near the membrane surface were both investigated in the
88
grid independence study. The final mesh configuration chosen was that of 100
axial cells by 75 vertical cells, with the mesh size varying in the vertical direction
such that the height of the largest mesh cell (in the centre of the channel) was 100
times that of the smallest mesh cell (adjacent to the membrane wall). An
illustrative portion of the computational mesh is shown in Figure 4.3.
(a)
(b)
Figure 4.3 - Representative detail of computational mesh for 2D channel with refinement for concentration
polarisation boundary layer at (a) channel inlet, (b) membrane surface (not to scale)
89
Table 4.3 – Model input parameters for simulations corresponding with Geraldes et al. [94]
Parameter Value
Membrane channel length L 200mm
Membrane channel width w Variable depending on mesh density – equal to size of smallest mesh
element
Transmembrane pressure 1, 2, 3, 4 MPa
Inlet Reynolds number NRe 500, 2000
In addition, initial model runs indicated that additional refinement was necessary
at the channel inlet to smoothly describe the ‘leading edge’ of the CP boundary
layer. The mesh was refined so that the first cells at the inlet were of unity aspect
ratio, growing progressively larger along the x-axis until x = 10 mm, after which
point the cells were of equal length for the remainder of the channel.
4.2.5 Simulations
The experimental data from Geraldes et al. [94] was used to provide the
membrane rejection factors for each solute at varying transmembrane pressures,
which are shown in Table 4.4. The membrane used in these experiments was the
CDNF50I thin-film composite NF membrane (supplied by Separem, Biela, Italy),
with a stated hydrodynamic permeability of 1.4×10-11 m2 s kg-1. Experiments were
conducted at several different crossflow velocities; the data used for this study are
for the extreme values (the lowest and highest studied velocities), which are
90
equivalent to entrance Reynolds numbers (NRe = 4u0Kȡ /Ș) of 500 and 2000,
respectively. These are equivalent to the stated Reynolds numbers of 250 and
1000 given by Geraldes et al. [94] due to a different formulation of the entrance
Reynolds number – the formulation used here is consistent with Berman [81] and
Karode [85]. The rejection values given were measured for the case where NRe =
2000.
Table 4.4 - Experimentally observed rejection factors for test compounds from [94] for NRe = 2000
7UDQVPHPEUDQHSUHVVXUHǻP
Compound 1 MPa 2 MPa 3 MPa 4 MPa
Robs* Robs Robs Robs
Sucrose 0.993 0.996 0.998 0.999
NaCl 0.403 0.572 0.649 0.715
Na2SO4 0.987 0.994 0.996 0.997
PEG1000 0.9981 0.9987 0.9993 0.9995
A comparison of the permeate fluxes predicted by the current model and the
experimental flux data and model results of Geraldes et al. [94] is shown in Figure
4.4. Values were compared at three collector sites (Collectors 1, 2 and 3) located
7.5 mm, 22.5 mm and 45 mm from the channel inlet, respectively – as no
experimental flux values were available for sodium chloride, these are not shown.
Reasonable agreement with the experimentally determined flux values is shown,
though the model consistently overpredicts the permeate flux. However, the
current model generally gives better agreement with the experimentally
91
determined fluxes than the model results of Geraldes et al. [94], except for the
case of sodium sulphate for NRe = 500.
The ability of the model to predict the formation of the CP boundary layer was
evaluated by comparing the predicted non-dimensional CP boundary layer
profiles with those presented by Geraldes et al. [94]. The non-dimensionalised CP
boundary layer height įW / h is plotted as a function of the dimensionless axial
length x / h in Figure 4.5. The profiles show good agreement with those
determined by Geraldes et al. [94]. As expected, a higher entrance Reynolds
number corresponds to a thinner CP boundary layer, due to increased shear
disrupting the CP boundary layer.
92
Sucrose, NRe = 500 Sodium sulphate, NRe = 500
6 6
Permeate flux (×10 -5 m3 m-2 s-1 ) 5 5
Permeate flux (×10 -5 m3 m-2 s-1 )

4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
Transmembrane pressure (MPa) Transmembrane pressure (MPa)
PEG1000, NRe = 500 Sucrose, NRe = 2000

6 6
5 5

4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
Sodium sulphate, NRe = 2000 PEG1000, NRe = 2000

6 6
5 5
4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
Collector 1 model Collector 2 model Collector 3 model

Collector 1 experimental Collector 2 experimental Collector 3 experimental
Collector 1 Geraldes et al. Collector 2 Geraldes et al. Collector 3 Geraldes et al.
Figure 4.4 - Comparison of predicted permeate fluxes for CFD model with experimental and modelled fluxes
from [94] using experimentally determined rejection factors
93
Sucrose Sodium chloride
0.05 0.06
NRe = 500 NRe = 500
0.05
0.04
Dimensionless CP layer height (įW / h)

0.04
NRe = 2000
0.03 NRe = 2000
0.03
0.02
0.02
0.01
0.01
0 -1.39E-17
0 10 20 30 0 10 20 30
Dimensionless axial length (x/h) Dimensionless axial length (x/h)
Sodium sulphate PEG1000

0.07 0.05
NRe = 500
0.06
NRe = 500
0.04
0.05
0.03
0.04
NRe = 2000
NRe = 2000
0.03
0.02
0.02
0.01
0.01
0 0
0 10 20 30 0 10 20 30
1 MPa model 1 MPa Geraldes et al.
Figure 4.5 - Comparison of dimensionless CP boundary layer profile for CFD model with computational
results of [94] using experimentally determined rejection factors
94
Some illustrative results from the simulations are also shown here. Figure 4.6
depicts the variation in solute mass fraction over the membrane channel,
illustrating the build up of the concentration polarisation boundary layer adjacent
to the membrane surface on the bottom of the channel. The velocity field in the
membrane channel can also be seen in Figure 4.7 near the channel inlet, and in
Figure 4.8 for the entire channel. The classic laminar velocity profile is observed
within the channel, though slightly flattened towards the lower half of the channel
due to the permeation through the lower channel wall.
The variation in the solution transport properties can also be observed from the
numerical solution. Representative examples of the variation in density, viscosity,
and osmotic pressure are shown in Figure 4.9, Figure 4.10, and Figure 4.11,
respectively. The density and viscosity increase in the vicinity of the membrane
surface as expected by relatively small amounts (in the order of several percent).
The osmotic pressure variation is more pronounced however, and can vary by
approximately an order of magnitude between the bulk flow in the centre of the
channel and adjacent to the membrane surface.
95
Figure 4.6 - Representative detail of variation in solute mass fraction in 2D membrane channel including
concentration polarisation boundary layer (sucrose solution, TMP 1 MPa, NRe = 500 – not to scale)
Figure 4.7 - Representative detail of velocity variation at inlet of 2D membrane channel including
96
Figure 4.8 - Representative detail of velocity variation for entire 2D membrane channel including
Figure 4.9 - Representative detail of density variation for entire 2D membrane channel including
97
Figure 4.10 - Representative detail of viscosity variation for entire 2D membrane channel including
Figure 4.11 - Representative detail of osmotic pressure variation for entire 2D membrane channel including
4.2.5.1 Osmotic pressure correction factor
The model consistently overpredicts the permeate flux compared to the
experimentally determined flux values. An attempt was made by Geraldes et al.
98
[94] to compensate for this over-prediction of flux by use of an osmotic pressure
correction factor ȁ, where the corrected permeate flux is given by:
J ȁ/V 'P '3W (4.3)
To see if this approach would account for the flux discrepancy in the current
model, a set of simulations was carried out using the osmotic correction approach
with expressions for ȁ as a function of the solute mass fraction as given by
Geraldes et al. [94], which are shown in Table 4.5. No correction factors were
provided for sodium chloride.
Table 4.5 - Osmotic pressure correction factor expressions for test compounds (after [94])
Compound Osmotic pressure correction factor ȁ *
Sucrose 1 - 0.375(1000Ȧi) 0.4
Na2SO4 1 - 0.012 (1000 Ȧi) 0.785
PEG1000 1 - 0.0384 (1000Ȧi) 0.31
* Expressions have been modified to account for different units for solute mass fraction
The permeate fluxes and CP boundary layer profiles obtained from these
simulations are shown in Figure 4.12 and Figure 4.13, respectively. The overall
agreement with the experimental flux values, when averaged over all
combinations of transmembrane pressure and collector location, is improved for
PEG1000 and sodium sulphate for both inlet Reynolds numbers. However, the
overall agreement is actually decreased for sucrose for both Reynolds numbers,
99
where the model tends to underpredict the flux, especially at higher
transmembrane pressures. In all cases, however, the model’s flux predictions are
significantly closer to the experimental results than the model results of Geraldes
et al. [94]. The modelled CP boundary layer profiles remain largely identical to
those obtained without the osmotic pressure correction factor.
100
Sucrose, NRe = 500 Sodium sulphate, NRe = 500
6 6
5 5

4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
PEG1000, NRe = 500 Sucrose, NRe = 2000

6 6
5 5

4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
Sodium sulphate, NRe = 2000 PEG1000, NRe = 2000

6 6
5 5
4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5

Figure 4.12 - Comparison of predicted permeate fluxes for CFD model incorporating osmotic pressure
correction factor with experimental and modelled fluxes from [94] using experimentally determined rejection
factors
101
Sucrose Sodium chloride
0.05 0.06
NRe = 500 NRe = 500
0.05
0.04

0.04
NRe = 2000
0.03 NRe = 2000
0.03
0.02
0.02
0.01
0.01
-2.78E-17 0
0 10 20 30 0 10 20 30
Sodium sulphate PEG1000

0.07 0.05
NRe = 500
0.06
NRe = 500
0.04
0.05
0.03
0.04
NRe = 2000
NRe = 2000
0.03
0.02
0.02
0.01
0.01
0 -2.78E-17
0 10 20 30 0 10 20 30

Figure 4.13 - Comparison of dimensionless CP boundary layer profile for CFD model incorporating osmotic
pressure correction factor with computational results of [94] using experimentally determined rejection
factors
102
4.2.6 Concluding remarks
It can be seen from comparison to analytical, experimental and computational data
that the present CFD modelling approach is capable of describing the flow
conditions for 2D membrane channels. Good agreement is observed in terms of
axial pressure drop, permeate flux, and the formation of the concentration
polarisation boundary layer. This gives reassurance that the basic hydrodynamic
model of the membrane filtration process is satisfactory, and we may now turn our
attention to more complex geometrical and hydrodynamic configurations. We first
consider the case of flow in spacer-filled channels.
4.3 Flow in spacer-filled channels
One area in which CFD methods have been reasonably prominent is in the
modelling of spiral-wound membrane (SWM) modules, as evidenced by a large
body of recent work in the area [36, 47, 113, 115, 119, 122, 131-136]. These
modules can exhibit complex hydrodynamic and mass transfer phenomena,
largely due to the effects of spacer meshes within the modules, which in addition
to separating the leaves of the membrane, are able to reduce the occurrence of CP
by promoting mixing and flow instabilities at the membrane surface, and hence
increase mass transfer through the membrane. This increase in mass transfer
comes at the expense of energy loss, though frequently the economic benefit of
increased mass transfer is greater than the disadvantage of energy loss. As direct
experimental observation of the flow field within SWM modules is difficult
103
without disturbing the flow structures of interest, CFD has become a powerful
tool for investigating the trade-off between optimisation of mass transfer and
minimisation of energy loss. Some of the primary considerations for CFD
modelling of spacer-filled channels are briefly reviewed in the following.
4.3.1 Unsteady flow and turbulence in spacer-filled channels
As discussed in Section 3.3, unsteady flow patterns can be observed in many
membrane filtration applications. This is particularly the case in spacer-filled
channels, especially at higher Reynolds numbers, as has been observed by several
authors [42, 137]. A casual evaluation of this might lead to the conclusion that use
of a turbulent CFD model may be beneficial for these cases. However, true fully-
developed, isotropic turbulence (generally for Reynolds numbers greater than 30
000) is only likely to be observed in the lower pressure membrane filtration
processes of MF and UF, and not for the high pressure NF and RO processes. In
particular, Reynolds numbers for SWM modules tend to be in the range 1000-
3000 [42, 100, 103]. Because of this, turbulence models (such as the common k-İ
model) which rely on the assumption of fully developed isotropic turbulence are
unsuitable for these conditions. Instead, a common approach is a form of direct
numerical simulation, using laminar flow models for transient simulation using
very fine computational meshes and timesteps [e.g. 42, 103, 111-114, 115, 137].
This is the approach that has been adopted in the subsequent simulations
discussed in Section 4.3.7.
104
4.3.2 2D and 3D modelling
As discussed in Section 3.3.2, both 2D and 3D models have been employed with
success to describe unsteady flow in spacer-filled membrane channels. However,
due to computational expense, 3D models are generally only currently feasible
using a unit-cell approach with ‘wrapped’ spatially periodic boundary conditions,
and with a simplified dissolving-wall boundary condition to represent membrane
mass transfer. The addition of a true permeable wall boundary condition adds
considerable computational expense, and is rare in the literature for 3D modelling.
With this in mind, a 2D model has been adopted for this investigation to reduce
computational requirements to manageable levels, and allow for the incorporation
of the permeable wall condition as employed in previous simulations. This also
allows for the further incorporation of predictive mass transfer modelling in a
coupled simulation approach (to be discussed in Chapter 6) within reasonable
computational constraints.
4.3.3 Model geometry
The model consists of a rectangular 2D channel filled with a ‘zigzag’ cylindrical
spacer arrangement identical to that used by Wardeh and Morvan [115], shown in
Figure 4.14. This zigzag arrangement is a reasonable 2D representation of a
woven spacer net, which is one of the most common spacer geometries. The flow
through the channel is oriented along the x-axis, and travels through an entrance
length equal to 12 times the spacer filament diameter df. It then travels past seven
105
spacer filaments which are followed by an exit length equal to 25 times the spacer
filament diameter. The filaments are separated by an axial distance of l, and the
channel height is h. For all simulations, the following configuration was used: h =
2 mm, l/h = 4, and df / h = 0.5.
The large entrance and exit lengths are necessary so that flow is fully developed
before the spacer-containing unit cells, and so that the outlet condition does not
impinge on the recirculation zones created by the final spacer [114]. The use of
only seven spacer filaments is a computationally economic way to represent the
spatially periodic velocity and pressure field in the spacer channel which in
practice is much longer. Simulations by Fimbres-Weihs et al. [114] showed at
least six filaments were necessary to obtain vortices in unsteady flow, and that
flow around the last spacer filament exhibited different behaviour due to exit
effects. In light of this, the region in the vicinity of the fifth and sixth spacers is
used as a representative monitoring location, as it is far enough downstream for
the periodic velocity and pressure flow structures to become established, but
sufficiently removed from exit effects as well [114].
spacer filaments
channel height h
diameter df permeate flow
feed retentate
entrance length permeate flow exit length

12 df 25 df
inter-filament distance l
Figure 4.14 – Model geometry for simulations of flow through 2D space-filled channel
106
Boundary conditions are specified at the channel inlet (in the form of a uniform
axial velocity) and the channel outlet (as an average static pressure of zero). The
average inlet velocity uavg is specified in terms of a channel Reynolds number as
used by several authors [44, 111, 114], which is given by:
Uuavg d h
Re ch (4.4)
K
where ȡ is the fluid density, Ș is the fluid viscosity, and dh is the hydraulic
diameter for a spacer-filled flow channel:
4H
dh
2 hch 1 H 4 d f
(4.5)
and where İ is the voidage of the spacer-filled channel:
Vtot VSP
H (4.6)
Vtot
where Vtot and VSP are the total volume per unit cell and volume occupied by
spacer filaments per unit cell, respectively. The model was run for four different
Reynolds numbers of 100, 200, 400 and 800 as per the work of Wardeh and
Morvan [115].
No-slip boundary conditions with no mass transfer are used along the channel
walls over the entrance and exit lengths, as well as on the curved surfaces of the
107
membrane spacers. The bulk composition of the solution is specified at the
channel inlet in terms of a solute mass fraction for each solute i; here, a simple
salt solution (sodium chloride) is considered. To represent the permeate flow
through the remainder of the membrane surface, a mass sink is again applied at
the membrane surface as previously described in Section 4.2.2. A constant
rejection value for the salt is used for all simulations. The relevant parameters for
the simulations are shown in Table 4.6, as per [115].
Table 4.6 – Model input parameters for simulations corresponding with Wardeh & Morvan [115]
Parameter Value
Transmembrane pressure 0.5 MPa
Membrane permeability LV 2 x 10-10 m Pa-1 s-1
Channel Reynolds number Rech 100, 200, 400, 800
Salt rejection R 0.95
Initial conditions are again automatically generated by the CFX solver, with the
initial solute concentration field set equal to that of the bulk solute mass fraction
throughout the domain.
4.3.5 Incorporation of concentration polarisation effects
As per Section 4.2.4.1, 87 the effects of concentration polarisation were included
by allowing the transport properties of the solution, such as diffusivity, density,
and viscosity, to vary spatially over the computational domain as a function of
solute concentration, using the relations provided by Geraldes et al. [94]. The
108
same approach was adopted by Wardeh and Morvan [115], though they were able
to directly specify the density using the relationship provided in [94]. This was not
possible in the present simulations, and again the harmonic average approach to
determine the solution density was employed as described in Section 4.2.4.1.
4.3.6 Computational mesh and convergence criteria
Due to the more complex geometry of the spacer-filled channel, unstructured
meshes were used for all simulations of the spacer-filled channel. As before,
significant mesh refinement is necessary in the vicinity of the membrane surface
to resolve the concentration polarisation boundary layer, and in this case
additional refinement is also necessary adjacent to the spacer filament surfaces. A
representative detail of a typical computational mesh in the vicinity of a spacer
filament is shown in Figure 4.15.
A mesh dependence study was conducted using a series of progressively finer
meshes to determine the mesh density after which variation in the solution was
considered to be negligible. The adopted mesh uses an element size of
approximately 0.05 mm in the bulk of the channel, growing progressively smaller
in the vicinity of the membrane and filament surfaces, where the smallest element
size is approximately 3 μm. This is similar to the mesh used by Wardeh and
Morvan [115], with a minimum element size of 5 μm and an element size in the
order of 0.02 mm in the bulk of the channel.
109
Convergence was again observed by monitoring the root mean squared (RMS)
residuals for mass, momentum, and solute mass fraction. The model was judged
to be converged once all residuals were smaller than 1 x 10-6. It should be noted
this is more stringent than the convergence criteria chosen in [115] (a RMS
residual value of 1 × 10-5).
Figure 4.15 – Representative detail of computational mesh for 2D spacer-filled channel CFD model
110
4.3.7 Transient simulations
A set of transient simulations were run for the conditions described in Table 4.6 to
verify the model against the results of Wardeh and Morvan [115]. Previously
steady-state simulations only were conducted in Section 4.2, as the relatively low
Reynolds numbers and simple channel geometry therein led to steady (invariant
with time) solutions being readily attainable. However, with the more complex
geometry of the spacer-filled channel and the comparatively large Reynolds
numbers, unsteady flow structures (i.e. varying with time) were found to exist by
Wardeh and Morvan [115], and also for similar conditions by other researchers
[e.g. 42, 113, 114]. For this reason, transient simulations were performed with
small time steps to accurately capture the unsteady flow behaviour. As per [114],
an approximate steady-state solution was chosen as the initial state for the
transient simulations.
Very small time steps are required to adequately resolve the unsteady flow
behaviour in such channels. A common way of representing this requirement is by
considering a dimensionless time step tstuavg / df; Koutsou et al. [42] used
dimensionless time steps smaller than 1.5 × 10-3, while Fimbres-Weihs et al. [114]
used time steps as small as 8.6 × 10-4. In light of this, the dimensionless time step
adopted for these simulations was set to the smallest dimensionless time step
found in the literature (8.6 × 10-4), corresponding to a physical time step varying
from 3.17 × 10-4 s with Rech = 100 to 3.96 × 10-5 s with Rech = 800. This is smaller
than the timestep chosen by Wardeh and Morvan [115] of 1 × 10-4 s with Rech =
111
800, but is reasonably similar. The simulations were run for slightly greater than
two residence times as per [114]. Because of the fine mesh and very small
timestep, these simulations incur considerable computational time (in the order of
weeks on an eight-CPU 64 bit machine with 16 GB of RAM).
Another way of evaluating the timestep selection is by examination of the
simulation Courant number. In all cases the chosen time step yielded a RMS
Courant number of approximately 0.2, with the maximum Courant number
observed less than 2.4. This is consistent with the values obtained in [114].
The results were verified against computational results presented in [115] for the
axial pressure drop, area-averaged wall shear stress, and area-averaged salt mass
fraction on the membrane surface in the spacer-filled channel. These results are
shown in Figure 4.16, Figure 4.17, and Figure 4.18, respectively.
Generally, good agreement is observed between the current simulation results and
that of Wardeh and Morvan [115]. The predicted variation in axial pressure drop
(shown in Figure 4.16) and wall shear stress (shown in Figure 4.17) in particular
is very close to that predicted in [115], though the current model predicts slightly
greater values of both. However, when considering the area-averaged salt mass
fractions (shown in Figure 4.18), the values predicted by the current model are
lesser, particularly at higher channel Reynolds numbers. The discrepancy between
the model predictions is thought be to due to several factors, including:
112
x The different treatment in specifying the variable density of the salt
solution, as discussed in Section 4.3.5;
x The small differences in computational mesh and timestep selection (i.e. a
smaller timestep was used for the present model, and smaller mesh sizes
adjacent to the membrane wall were used for the present model); and
x The differences in convergence criteria used (i.e. the current model has a
more stringent set of convergence criteria).
350
300
Axial pressure drop (Pa)
Current model
250 Wardeh
200
150
100
50
0
0 100 200 300 400 500 600 700 800 900
Channel Reynolds number
Figure 4.16 – Comparison of predicted axial pressure drop as a function of channel Reynolds number against
computational results of [115] for spacer-filled channel
113
1.4
1.2 Current model

Wall shear stress (Pa)
1 Wardeh
0.8
0.6
0.4
0.2
0
0 50 100 150 200 250 300 350
Axial pressure drop (Pa)
Figure 4.17 – Comparison of predicted area-averaged wall shear stress as a function of axial pressure drop
against computational results of [115] for spacer-filled channel
0.006
0.005
Salt mass fraction
0.004
0.003
Current model
0.002 Wardeh
0.001
0
0 100 200 300 400 500 600 700 800 900
Channel Reynolds number
Figure 4.18 – Comparison of predicted area-averaged salt mass fraction as a function of channel Reynolds
number against computational results of [115] for spacer-filled channel
In addition, the comparative decrease in mass fraction adjacent to the membrane
surface is qualitatively commensurate with the comparative increase in the wall
shear stress compared to the computational results in [115]. Considering these
factors, it may reasonably be judged that the model is producing reasonable
results and can be considered to be verified.
114
4.3.7.1 Steady flow structures
It is interesting to now examine the transient model results to observe the
evolution of any unsteady flow structures (if they exist) over time. Beginning first
with the lowest channel Reynolds number (Rech = 100), steady flow was observed
throughout the simulation period. That is, as time progressed through the
simulation period, there was no observable variation in the calculated flow
variable fields. The same was also true for channel Reynolds numbers of 200 and
400. This is consistent with previous computational results in the literature for
similar conditions [e.g. 111, 114].
Some representative graphs of the steady-state solutions for these lower channel
Reynolds number flows are shown in Figure 4.19 through Figure 4.22, all in the
vicinity of the fifth and sixth spacer filaments (chosen as a representative
monitoring location as discussed previously). In all instances (and all subsequent
graphical representations of the 2D spacer-filled channel), the bulk flow is from
left to right. Figure 4.19 shows the variation in salt mass fraction throughout the
channel, and clearly demonstrates the disruption of the concentration polarisation
layer due to the increased shear and recirculation induced by the spacer filaments.
The flow structures are the same as that described previously in the literature for
‘zig-zag’ channels [111, 112, 114], where the flow follows the zig-zag pattern
imposed by the spacers. The influence of increasing crossflow velocity (expressed
in terms of increasing channel Reynolds number) can also be seen in Figure
4.19;the formation of high concentration areas behind the filaments is very
115
apparent at Rech = 100. However, as the channel Reynolds number increases, the
increased recirculation and shear reduces the formation of these high
concentration zones. However, at higher channel Reynolds numbers, similar zones
are formed at the leading edge of the filaments due to recirculation.
Rech = 100
Rech = 200
Rech = 400
Figure 4.19 – Predicted salt mass fraction distribution for spacer-filled channel for conditions in [115] for
steady flow at lower channel Reynolds numbers
This can also be seen via examination of the fluid streamlines and velocity vector
plots, shown in Figure 4.20 and Figure 4.21 respectively. Large recirculation
zones are apparent on the downstream faces of the filaments, while smaller
116
recirculation zones occur on the upstream faces adjacent to the membrane walls.
At Rech = 400, the downstream recirculation zone extends nearly to the next
filament, growing progressively smaller with smaller channel Reynolds number.
Rech = 100
Rech = 200
Rech = 400
Figure 4.20 – Predicted fluid streamlines for spacer-filled channel for conditions in [115] for steady flow at
lower channel Reynolds numbers
117
Rech = 100
Rech = 200
Rech = 400
Figure 4.21 – Predicted velocity vectors for spacer-filled channel for conditions in [115] for steady flow at
This can also be visualised in terms of the vorticity, which is essentially the
circulation per unit area at each point in the flow domain. A plot of the vorticity in
the monitoring region (the fifth and sixth filaments) is shown in Figure 4.22. The
figure shows areas of high vorticity on the leading edge of the filaments in the
vicinity of the channel centre, which become increasingly pronounced with
increased channel Reynolds number. The recirculation areas behind filaments are
also clearly visible.
118
Rech = 100
Rech = 200
Rech = 400
Figure 4.22 – Predicted vorticity fields for spacer-filled channel for conditions in [115] for steady flow at
4.3.7.2 Unsteady flow structures
For the highest channel Reynolds number studied (Rech = 800), significant
unsteady flow behaviour was observed (i.e. the computed mass fraction, velocity,
and pressure fields varied throughout the simulation time). The existence of
unsteady flow at this channel Reynolds number is consistent with previous CFD
results in the literature [e.g. 114].
119
We can first examine the evolution of the salt mass fraction distribution over time
within the membrane channel, shown in Figure 4.23. The computed mass fraction
fields are shown at time increments of approximately 6.1 ms from top to bottom
(similar to that used in [114] to visualise such flows), and partially illustrate the
movement of vortices through the spacer-filled channel in this unsteady flow. A
fuller understanding of the flow can be obtained by examining these in
conjunction with successive plots of fluid streamlines (shown in Figure 4.24),
velocity vectors (shown in Figure 4.25), and vorticity (shown in Figure 4.26), at
identical time intervals.
In a manner similar to [114], it is useful to identify and numerically label some of
the individual vortices to track their movement through the spacer-filled channel
over time. Specifically, individual vortices are assigned a number when they
appear at the leftmost portion of the monitoring region, and this number is used to
track the movement of each particular vortex as it moves downstream (from left to
right) through the monitoring region with time (from the top to bottom of the
page). (It is not intended to perform an exhaustive study of vortex shedding
behaviour in spacer-filled channels, as excellent examples of this can be found in
other papers [e.g. 111, 112, 114, 137]. Instead, we merely provide a qualitative
demonstration of the phenomenon to show some salient features of spacer-filled
channel flow).
The movement of vortices is perhaps most easily visualised via examination of the
streamlines in Figure 4.24. Vortices are formed both in front of filaments (e.g.
120
vortices 4 and 6), and behind filaments adjacent to the membrane wall (e.g. vortex
7). Vortices formed at the leading edge of filaments are eventually ‘squeezed’
over the filament in a reasonably periodic fashion, where they form elongated
elliptical vortices adjacent to the main streamflow (e.g. vortices 1, 2, and 3).
These then travel downstream between filaments and coalesce (seen with vortices
2 and 3) before eventually reaching the next filament and ‘squeezing’ past.
Vortices formed behind filaments adjacent to the membrane wall (vortex 7, for
example) appear to also travel downstream below the larger vortices adjacent to
the main streamflow, and coalesce approximately halfway between filaments
adjacent to the membrane wall.
Examination of the velocity vector plots in Figure 4.25 also reveals that vortices
also rotate in different directions, depending on whether they are above or below
the main streamflow. Specifically, those vortices on the lower side of the channel
rotate clockwise, and those on the upper side rotate counter-clockwise. This is
consistent with the behaviour observed in [114] for a similar channel, and results
in the wall shear generated by these vortices being of opposite sign to the shear
generated by the main channel flow, as noted in [114].
The shearing behaviour of these flows has some effects on the concentration
polarisation layer and hence the salt distribution, as seen in Figure 4.23. The
greatest disruption to the concentration polarisation layer occurs directly opposite
the filaments on the opposing wall of the membrane, corresponding with the areas
of highest vorticity as shown in Figure 4.26. Regions of high salt concentration
121
also build up behind filaments as observed for the steady flow cases at lower
channel Reynolds number. However, the high concentration regions in front of the
filaments observed for low Rech in Figure 4.19 are not as prevalent for the
unsteady flow case, due to the vortex-induced movement of fluid, also consistent
with the observations in [114].
122
q
n o
p
r
q
n
s o,p
r
q
s n
o,p
r
s
n
t
s
n
t
Figure 4.23 – Predicted salt mass fraction distribution for spacer-filled channel for conditions in [115] for
unsteady flow at high channel Reynolds number (Rech = 800) at successive 6.1 ms timesteps (top to bottom)
123
q
n o
p
r
q
n
s o,p
r
q
s n
o,p
r
s
n
t
s
n
t
Figure 4.24 – Predicted fluid streamlines for spacer-filled channel for conditions in [115] for unsteady flow at
high channel Reynolds number (Rech = 800) at successive 6.1 ms timesteps (top to bottom)
124
q
n o
p
r
q
n
s o,p
r
q
s n
o,p
r
s
n
t
s
n
t
Figure 4.25 – Predicted velocity vectors for spacer-filled channel for conditions in [115] for unsteady flow at
125
q
n o
p
n q
o,p
s
r
q
s n
o,p
r
s
n
t
s
n
t
Figure 4.26 – Predicted vorticity fields for spacer-filled channel for conditions in [115] for unsteady flow at
126
4.3.8 Concluding remarks
The results presented for 2D flow in spacer-filled channels indicate that the
current model agrees well with computational data presented in the literature. In
addition, the current model includes a more physically realistic representation of
the flow by incorporating a true permeable wall boundary condition for the
membrane surfaces, which has largely been neglected by other researchers for
spacer-filled channel flow. The model results have also demonstrated the ability
of the transient laminar flow model to describe unsteady flows for moderate
channel Reynolds numbers, when used with appropriately fine spatial and
temporal discretisation. Using such a model, unsteady flow features are observed
at channel Reynolds of Rech = 800, and formation and movement of vortices can
be clearly observed in the membrane channel, consistent with previous results
presented in the literature. However, significant computational expense is incurred
by directly simulating the unsteady flow structures using the laminar flow model,
with model run times in the order of weeks on an eight-CPU 64 bit machine with
16 GB of RAM.
4.4 Additional remarks on 3D modelling
Previous sections have touched upon the large computational expense required for
3D modelling as opposed to 2D modelling. Given the not insignificant expense of
the transient simulation of unsteady 2D flows discussed in Section 4.3, it is
perhaps useful to comment further on the feasibility of 3D simulations.
127
The high cost of 3D simulations (generally only used for simulation of spacer-
filled channels) is largely compensated for by use of a unit-cell approach with
spatially periodic boundary conditions, as discussed previously. The assumption
of spatially periodic flow that this method relies on is valid in many spacer-filled
channels, though not in some instances. As part of this study, some preliminary
investigation into the computational and mesh requirements for 3D modelling was
undertaken, which is briefly presented here to give some context to, and
justification of the adoption of 2D models for this thesis.
The membrane channel configuration modelled by Wardeh and Morvan in 2D
[115], which has been reproduced and modelled in detail in Section 4.3, provides
a reasonable comparison point for the use of 2D and 3D models. To investigate
computational requirements for extending this simulation to a 3D case, model
geometries and meshes were constructed for the same channel geometry described
in Section 4.3.3, by assuming a 3D woven spacer net geometry. This was
conducted for two scenarios:
x a unit-cell approach, where the 3D computational domain extended for
two inter-filament distances in the x and z directions; and
x an entire channel approach, where the 3D computational domain was
extended for the entire channel length (i.e. over seven spacer filaments) as
modelled in the 2D case (including entrance and exit lengths), in the x
direction, and for two inter-filament distances in the z direction.
128
Unstructured meshes were generated for both of these geometries to examine
potential memory requirements and feasibility of performing 3D simulations for
these geometries. Meshes were generated on a dedicated 64-bit machine (8
processors, 16 GB of RAM), with considerably more computational power than a
typical desktop computer. It is useful to consider as a reference point the mesh
sizes established by the 2D grid independence study (mesh elements of 0.05 mm
in the bulk of the channel, progressively refined to elements of 3 μm in areas of
interest). To minimise memory requirements, meshes can be generated only for
the quarter unit cell geometry depicted in Figure 4.27, and the symmetry of the
unit cell thus exploited to construct the entire geometries by ‘stitching’ mesh
sections together. Even with this technique, an equivalently fine spatial resolution
for the unstructured 3D mesh was not achievable on this machine, as the meshing
process exceeded the memory limits of the machine for even the small unit cell
geometry.
The finest mesh that could be generated contained mesh elements of 0.05 mm in
the bulk flow with elements of 7.5 μm adjacent to the membrane surfaces. This
represents a mesh consisting of approximately 65.7 million elements, which by
current standards is a very large mesh. For comparison, the 2D mesh adopted in
Section 4.3.6 consists of approximately 295 000 elements; and 3D meshes for unit
cell problems found in the literature range from approximately this size (e.g.
[122]) up to large meshes of 15-20 million elements [118].
129
Figure 4.27 – Quarter unit-cell geometry corresponding to 2D model dimensions used in [115]
A coarser mesh was thus constructed with mesh elements of 0.1 mm in the bulk,
and wall elements of 20 μm, to yield a total mesh size of 9.7 million elements.
This is in the general range of mesh sizes presently investigated in the membrane
literature. This coarse mesh is shown in full in Figure 4.28, with a representative
detail shown in Figure 4.29.
130
Figure 4.28 – Coarse 3D mesh (9.7 million elements) for unit-cell corresponding to 2D model dimensions
used in [115]
Figure 4.29 – Representative detail of coarse 3D mesh (9.7 million elements) for unit-cell corresponding to
2D model dimensions used in [115]
131
A test case was run for the unit-cell model using the conditions previously
modelled in Section 4.3 with the coarse 3D mesh described above. Periodic
boundary conditions were established at the boundaries of the unit cell as
described in the literature by various authors [118, 121, 122]. As mentioned
previously, the majority of 3D unit cell models have considered the simplified
cases of impermeable membrane walls or the dissolving wall approximation. As a
first attempt, simulations were run using an impermeable wall boundary
conditions for a channel Reynolds number of 100. Using this approach, a well-
converged steady-state solution was able to be obtained. However, no detailed
results are presented, as the results could not be verified using a grid dependence
study due to the constraints on computational time.
However, applying the permeable wall boundary condition used in 2D simulations
led to convergence difficulties for the mass fraction equations, and a reliable
converged solution was not able to be obtained. It is probable that a more refined
mesh could in this case lead to a converged solution, but again this is limited by
the extensive computational cost, and no further efforts were expended in this area
due to time considerations.
Perhaps most importantly, computational times for the 3D steady-state solution
with impermeable walls were in the order of hundreds of times larger than for an
equivalent 2D steady-state solution. When one considers that the computational
times required for transient simulations for the 2D channel considered in Section
4.3.7 with permeable walls were in the order of weeks, it becomes apparent that
132
3D simulations with permeable walls would require very large computational
resources to be feasible.
Similarly, the larger model geometry required for the entire channel approach
proved prohibitively large for simulation at the same mesh resolution. It was not
possible to run these simulations on the computational facilities available due to
memory requirements, regardless of the boundary conditions adopted (permeable
or impermeable walls).
Hence, it may be concluded that for most practical purposes, full 3D simulations
of a membrane channel are as yet still not possible, and the 3D unit cell approach
remains the only viable option for the near future. However, this still imposes
limits on the level of mesh refinement available; while converged solutions may
be obtained using the simplifying assumption of an impermeable wall boundary
condition, convergence using a permeable wall boundary condition has proved to
be difficult. It appears that very fine mesh resolutions beyond the present
computational facilities available for this project are required to obtain converged
solutions when considering a true permeable wall boundary condition. Further, the
computational time requirements for such an undertaking are also prohibitively
large for most purposes, and indeed for the purposes of this thesis, at present.
133
4.5 Conclusions
The use of multicomponent CFD models using the commercial CFD code Ansys
CFX has been shown to well describe the hydrodynamic behaviour of membrane
channels. The inclusion of a true permeable wall boundary condition in these
models, rather than the impermeable wall or dissolving-wall approximations most
commonly encountered in the literature, has also been shown to match well with
experimental, analytical, and numerical predictions. This also allows for the
permeation rate to vary spatially throughout the membrane channel, introducing
an additional level of detail into the hydrodynamic predictions.
Concentration polarisation effects can also be well described using such
multicomponent models by allowing the transport properties of the solution
(viscosity, density, diffusivity, and osmotic pressure) to vary spatially according
to the local concentrations of each solution component. This is achieved by using
empirical relationships derived from the literature for the certain components
considered in this study; however, it is noted that a more general method of
predicting these properties would undoubtedly be beneficial. Comparison to
predicted concentration polarisation layer thicknesses from the literature produced
favourable results, and the model was hence able to predict the corresponding
(experimentally measured) reduction in permeate flux reasonably accurately.
Unsteady (time-varying) flow for the moderate channel Reynolds numbers
observed in high-pressure membrane filtration can also be well described by use
134
of transient laminar flow models, with appropriately fine temporal and spatial
discretisation to directly resolve unsteady flow structures (in effect, a form of
direct numerical simulation). This method was shown to agree well with
computational results previously presented in the literature, and produced
qualitatively similar behaviour to previous studies in terms of vortex formation
and transport for higher channel Reynolds numbers. Importantly, the transient
laminar flow approach is well suited to describing high-pressure membrane
filtration processes, where Reynolds numbers fall well below the generally
DFFHSWHGOLPLWIRUWKHXVHRIWXUEXOHQFHPRGHOV5H§
Finally, the models described in this chapter are suitable for implementation in
either 2D or 3D geometries, though simulations have been performed only for 2D.
The primary reason for this is computational expense, especially when the direct
simulation of unsteady flow effects using a transient laminar flow scheme is
considered. However, extension of the modelling approach to 3D is relatively
straightforward, but limited by present computational capabilities.
In summary, the CFD methods described here can well represent the
hydrodynamics of a membrane channel. However, they do not provide any
capacity for predicting the selective transfer of mass through the membrane itself
(i.e. the rejection of solutes), arguably the most important characteristic of a
membrane filtration system. Coupling of these hydrodynamic models to models of
the membrane mass transfer process would therefore provide a more complete
135
predictive model of the overall process. We now consider the numerical
implementation of a mass transfer model for this precise purpose in Chapter 5.
136
Chapter 5
Prediction of membrane rejection
5.1 Introduction
The primary purpose of membrane filtration is the separation of components from
a feed stream, which we have previously characterised using the concept of
rejection. The rejection of various components however can vary significantly,
depending on the nature of the components themselves, the nature of the
membrane, and the operating conditions of the filtration system, such as the
crossflow velocity, applied pressure etc. This great variability has meant that
historically, the most practical method for determining the rejection of a
component or components in a membrane filtration system has been through
experimental measurement, which may be expensive or impractical.
It is apparent that for a computational model of the membrane filtration process to
have general utility, a general method for predicting the rejection of feed streams
containing an arbitrary mixture of components, with an arbitrary membrane, and
arbitrary operating conditions is needed. The following chapter describes the
application of a computational method for this purpose, largely with regard to NF
and RO processes. This computational approach is then incorporated into rigorous
hydrodynamic simulations of the membrane filtration process, to be discussed in
detail in Chapter 6.
137
5.2 Approaches to describe separation in NF and RO processes
Two main descriptions of the solute separation process in NF and RO have
historically been used: the first are based on phenomenological (non-structural)
models in which the membrane is treated as a ‘black box’, and no characterisation
of the structural and electrochemical properties is required. Examples of this
approach include the early work of Spiegler and Kedem [138] and Kedem and
Katchalsky [139], and this approach has still been used in recent times [140]. In
these methods, membrane performance is characterised by use of fitting
parameters (salt permeability and a reflection coefficient) [e.g. 138, 139, 140].
These methods are limited in that they are only able to describe the behaviour of
binary salt systems (or in the limiting case, binary salt systems in the presence of a
completely rejected uncharged solute) [141].
The second approach are commonly termed ‘pore models’, or structural models,
and describe separation in terms of hydrodynamic permeation through a set of
(usually cylindrical) pores, in which the convective and diffusive transport of
solute is hindered by interactions between the solute and the membrane. It should
be noted that the use of pore models for NF and RO membranes does not imply
the physical existence of geometrically well-defined pores in these membranes;
rather, membranes can be characterised by an effective pore size which
approximates the hindrance in solute transport caused by the complex polymeric
microstructure of these membranes [17]. These models were initially proposed to
138
describe rejection of uncharged solutes in NF processes, but can be generalised to
describe transport of charged and uncharged solutes in most membrane types.
This is usually done by solving the extended Nernst-Planck equations using the
Donnan steric partitioning model (DSPM) to describe the transport of solutes
through the membrane [e.g. 16, 17, 127, 142, 143-150]. This approach has several
advantages: first, it is based on a more rigorous description of physical parameters
of the membrane system only, and all parameters used have a reasonably direct
physical interpretation. Secondly, it is capable of describing the transport of both
charged and uncharged solutes, and mixtures thereof.
However, the DSPM has been noted to have several limitations. In particular, it
has been shown that in some instances that the method fails to accurately predict
the rejection for divalent counterions [151]. Additionally, very high values for the
membrane charge can be obtained when used as a fitting parameter, which may
not be physically reasonable [e.g. 152, 153]. For this reason extensions to the
DSPM using the concept of dielectric exclusion have been proposed by several
authors [18, 19, 151], generally referred to as the DSPM-DE model. The dielectric
exclusion effect is caused by two factors: first, the difference in dielectric constant
between an aqueous solution and the surrounding membrane structure [151]; and
additionally, the difference in dielectric constant of the solution between the bulk
of the solution and in the confined membrane pores [154]. The DSPM-DE model
has indeed been shown to better predict rejection of multivalent counterions [e.g.
19, 151].
139
Despite its limitations, however, the DSPM represents a reasonable starting point
for developing a general numerical method to predict rejection, which will be
applied later (see Chapter 6) in conjunction with a CFD model of the membrane
channel hydrodynamics. For this reason, the extensions to the DSPM provided by
the dielectric exclusion models have been neglected for the current study, but
represent a useful and necessary extension for future work.
In this chapter, the use of a numerical method for solving these equations for a
general case using the DSPM is described, so the rejection of various components
in a mixed solution containing multiple charged or uncharged compounds can be
predicted.
5.3 Model theory and solution details
5.3.1 The extended Nernst-Planck equation
The extended Nernst-Planck equation can be used to describe the transport of
solutes through a membrane [17, 148], and hence predict the rejection of a solute
by a membrane based on several physically meaningful parameters. It can be
written as shown in Eq. (5.1), where the terms on the right hand side represent
diffusion, convection, and in the case of a charged solute, electric migration,
respectively, for some solute i:
140
dc i z i c i K i , d D i d\
ji K i , d Di K i ,c c i J V F (5.1)
dy RT dy
where ji is the solute flux, Di is the bulk diffusivity, zi is the electrical valence,
and ci is the concentration inside the membrane. The terms Ki,c and Ki,d represent
the hindered convection and diffusion of the solute inside the membrane,
respectively: the determination of these terms will be considered later. The overall
volumetric flux through the membrane is given by JV, and y is the axial coordinate
along the direction of the solution flux. The final term on the right hand side of
Eq. (5.1) representing transport due to electric migration depends on the Faraday
constant F, the ideal gas constant R, the absolute temperature T, and the electric
potential ȥ. The extended Nernst-Planck equation can thus predict the rejection
for uncharged solutes, charged solutes, or a mixture of the two.
5.3.2 Rejection of uncharged solutes
In the case of uncharged solutes, the electric migration term in Eq. (5.1)
disappears to yield the following:
dci
ji K i ,d Di K i ,c c i J V (5.2)
dy
It is assumed that the rejection of the uncharged solute can be modelled by a steric
partitioning coefficient ĭi which describes the accessibility of the pore to the
solute. The determination of the steric partitioning coefficient will be considered
141
later. This applies at the interface between the membrane and the external solution
both at the pore inlet (y = 0) and pore outlet (y = ǻy), such that:
ci , y 0 ĭi Ci ,W (5.3)
ci , y 'y ĭi Ci , p (5.4)
where ci is the concentration inside the membrane, Ci is the concentration of the
external solution, and the subscripts W and p represent the feed solution adjacent
to the membrane wall, and the permeate solution, respectively. The rejection of
solute i is defined as:
Ci , p
Ri 1 (5.5)
Ci ,W
By writing the flux of solute i as ji J V C i , p and integrating Eq. (5.2) along the
membrane pore between y = 0 and y = ǻy, the following solution can be easily
obtained:
ĭi K i ,c
Ri 1
1 1 ĭi K i ,c exp Pei
(5.6)
Pei K i ,c J V ǻ\ K i ,d Di (5.7)
where Pei is the Peclet number.
142
The next step is the determination of the steric partitioning coefficient ĭ and
hindrance coefficients Ki,c and Ki,d. These differ depending on the assumed shape
of the solute, i.e. whether the solute is assumed to be spherical or non-spherical.
The simplest case is that of a spherical solute as follows.
5.3.2.1 Spherical solutes
The steric partitioning coefficient ĭi describes the rejection of the solute due to
steric (molecular sieving) effects. The simplest method for describing steric
effects is to assume a spherical solute, which is sound for small solutes. The
partitioning coefficient can then be written as a function of the membrane pore
radius rp and the radius of solute i, typically approximated as the Stokes radius,
ri,s:
ĭi 1 Oi 2 , Oi ri , s rp (5.8)
It should be noted that the use of the membrane pore radius need not necessarily
mean that well-defined pores physically exist within the membrane; instead, as
noted in [17], it may be regarded as a pore size which yields equivalent hindrance
to solutes as that of the polymeric structure of the membrane.
The hindrance factors for diffusion (Kd) and convection (Kc) describe how drag
forces inside the confined space of the pore hinder the diffusion and convection of
143
the solute [155], and are also functions of the ratio of the solute radius to pore
radius, Ȝ. Here, the expressions presented by Bowen et al. [148] have been used:
K i ,d 1.0 2.30Oi 1.154Oi 0.224Oi

2 3
(5.9)
K i ,c 2 ĭi 1.0 0.054Oi 0.988Oi 2 0.441Oi 3 (5.10)
5.3.2.2 Non-spherical solutes
For larger solutes, the shape of the solute plays an important role in the steric
interactions between the pore and solute. Several methods for determining the
steric partitioning coefficient for a non-spherical solute have been presented in the
literature based on representing the molecule in terms of a simplified geometric
shape. Example shapes include cylinders [156, 157], rectangular parallelipeds
[158-160], and prolate revolution ellipsoids [161]. The method described by Kiso
et al. [158-160] has been adopted in the present model, and is described briefly in
the following.
Approximating the solute as a rectangular parallelepiped, the solute length Li,S is
given by the distance between the two most distant points on the solute
(effectively the distance between the two most distant atoms in the molecule in
addition to their van der Waals radii). The line that joins these two most distant
points is defined as the L-axis. The solute width is determined by projecting the
solute particle onto the plane perpendicular to the L-axis, and constructing the
144
smallest possible rectangle on that plane which will fully enclose the projection of
the solute particle. Defining the area of this rectangle as Si, the solute width Wi,S
can then be calculated:
1
Wi , S Si (5.11)
2
As the solute is considered to be non-spherical, the solute orientation affects the
proportion of the pore area that is accessible to the solute. The length of the
projection of the solute against the pore surface is called the effective solute
length, and is given by:
Li , p Li , S cos D 2Wi , S sin D (5.12)
where the angle between the L-axis and the pore surface is Į. The steric partition
factor for some angle Į is thus given by:
2
§ ai ·
ĭi D ¨ ¸ (5.13)
¨r ¸
© p¹
2
§ Li , p ·
ai rp ¨¨
2
¸¸ Wi , S (5.14)
© 2 ¹
A probabilistic approach is used to calculate an overall partition coefficient ĭi as
follows:
145
S /2
ĭi ³ ĭi D pD dD (5.15)
0
S /2
³ pD dD 1 (5.16)
0
where the probability function p(Į) represents the probability of an individual
soluWHEHLQJRULHQWHGDWDQDQJOHĮDQGPXVW satisfy Eq. (5.16). A common choice
of function is p(Į) = sin(Į).
A range of semi-empirical techniques are available to predict molecular geometry,
such as the MOPAC2009 software (Stewart Computational Chemistry, CO,
USA), which has been adopted for the compounds used in this study.
MOPAC2009 performs a series of optimisations to predict the most stable
conformation of a molecule, and returns a series of geometric parameters of the
optimised molecule, as well as Cartesian coordinates of each atom within the
molecule. Investigation of the MOPAC output indicated that the geometric
parameters output were sometimes unreliable, and hence direct processing of the
Cartesian output was required. For this purpose, a numerical routine was
developed to determine the solute length and width based on the Cartesian
coordinates of each atom and their associated van der Waals radii. The routine
employs a three-dimensional bounding box method to determine these geometric
parameters. The code for this routine and some sample output for several test
molecules (glucose, sucrose, cyanazine, glycerine, and Vitamin B12) is shown in
Appendix A.1.
146
Little information is available on determining hindrance factors for non-spherical
solutes. The approach adopted by Kiso et al. [159] uses the same expressions as
Eqs. (5.9) and (5.10); however, for the non-spherical solute the Stokes radius ri,s is
unknown. A correlation is thus used to approximate the Stokes radius (for use in
Eqs. (5.9) and (5.10) only) based on the solute width WS as follows:

ri , s u109 1.42 Wi , S u109 0.142 u10 9 (5.17)
5.3.2.3 Determination of membrane parameters for uncharged solutes
The only parameters thus needed to predict the rejection for a solution containing
uncharged solutes and an arbitrary membrane are the membrane pore radius rP
DQGSRUHOHQJWKǻy (sometimes given as ǻy /Ak, where Ak is the effective porosity
of the membrane). In principle the pore radius can be determined by inspection of
the membrane via methods such as atomic force microscopy [148]. However, in
practice, typically experimental measurement of the rejection of a single
uncharged solute is made, and rP and ǻy /Ak are then adjusted to achieve good
agreement with the experimentally measured rejection.
5.3.3 Rejection of charged solutes
For the case of charged solutes, the electric migration term is retained in Eq. (5.1),
which is no longer easily soluble by analytical means and must be solved
147
numerically. To do so, the electrical gradient through the pore, must be
determined, which for a system of n charged solutes is given by:
K i,c ci Ci, p
n
zi J V
d\
¦Ki 1 i , d Di (5.18)
dy F n 2
RT i 1
¦ zi ci
This is subject to the condition of electroneutrality in the bulk solution:
¦zC
i 1
i i 0 (5.19)
and that electroneutrality is also upheld within the membrane pore:
¦z c
i 1
i i zX CX (5.20)
where zX and CX are the valence and density of the fixed membrane charge. It is
assumed this charge is constant throughout the membrane pore.
Eq. (5.1) is then solved numerically as a boundary value problem using the
following boundary conditions based on the assumption of Donnan equilibrium
with the solution considered sufficiently dilute that the activity coefficients are
unity:
§ zF ·
ci , y 0 ĭi Ci ,W exp¨ i '\ D , y 0 ¸ (5.21)
© RT ¹
148
§ zF ·
ci , y 'y ĭi C i , p exp¨ i '\ D , y 'y ¸ (5.22)
© RT ¹
where '\ D , y 0 and '\ D , y 'y are the Donnan potential at the pore inlet and outlet,
respectively. The steric partition coefficients ĭi are identical in meaning to the
uncharged case, and may be determined in the same manner, either using a
spherical model (in the case of a monatomic ion) or non-spherical model (in the
case of a polyatomic ion) as appropriate.
For all charged solutes at the pore inlet, the Donnan potential is identical, and the
same condition applies for the pore outlet. The concentrations just inside the pore
inlet ci,y=0 are determined according to this equality condition, and according to
the conditions specified in Eqs. (5.19) and (5.20), so that a set of n non-linear
equations is obtained as follows:
n n
¦ zi ci, y 0 z X C X ¦ ziCi,W
i 1 i 1
0 (5.23)
1 §¨ C1,W ·¸ 1 §¨ Ci ,W ·¸
ln ln 0, i 2,..., n (5.24)
zi ¨© c1, y 0 ¸¹ zi ¨© ci , y 0 ¸¹
For the simple case of n = 2 with univalent solutes, analytical expressions can be
derived relatively easily to solve for the concentrations ci,y=0. For n > 2 and / or
multivalent solutes, this becomes more difficult, and the equations tend to be
highly non-linear. A numerical method using a variant of Powell’s dogleg method
[162] is thus used to obtain the concentrations for the general case.
149
Determining the outlet boundary condition presents an additional difficulty as the
permeate concentration is not yet known. This is overcome by writing the
permeate concentration for each solute in terms of the feed concentration and the
rejection, viz.
Ci , p 1 Ri Ci ,W (5.25)
so using Eq. (5.22) one can write:
§ zF ·
ci , y 'y ĭi 1 Ri Ci ,W exp¨ i '\ D , y 'y ¸ (5.26)
© RT ¹
It is now possible to solve the boundary value problem consisting of the n
ordinary differential equations defined by Eq. (5.1), and the 2n boundary
conditions defined by Eqs. (5.21) and (5.26). This is done numerically with the
rejections Ri as a set of additional unknown parameters that also must be solved
for in order to establish the Donnan potential at the pore outlet. Implementations
of this numerical solution have been written in both MATLAB and Fortran 90,
using the boundary value problem solvers developed for these languages by
Shampine et al. [163, 164]. The Donnan potential at the pore outlet is determined
within the boundary value problem solver by iteratively solving the system of
non-linear equations described by Eqs. (5.23) and (5.24).
150
5.3.3.1 Determination of membrane parameters for charged solutes
In addition to the parameters rP and ǻy /Ak used for uncharged solutes, the
additional parameters describing the charge of the membrane, zX and CX must be
determined for charged solutes. For virtually all practical membranes, the inherent
charge of the membrane is negative in sign, so zX = -1, though the numerical
implementation can handle either case. To determine CX, it should be possible to
perform first perform rejection experiments for uncharged solutes to obtain rP and
ǻy /Ak, and then perform additional rejection experiments for a single salt solution
and adjust CX to obtain a good fit. In the literature however, there are often
substantial differences between ǻy /Ak values between the uncharged and charged
cases [141, 148]. The reasons for this discrepancy are not fully clear, as discussed
later.
In addition, it has also been observed [17, 141, 147-149] that the charge density
CX also varies with the feed concentration of the salt solution. This can be
expressed in terms of an isotherm [17], which can be determined by rejection
experiments for a single salt solution at varying feed concentrations.
5.3.4 Rejection of mixtures of charged and uncharged solutes
Combining the approaches described previously to determine the rejections for a
solution containing charged and uncharged solutes is straightforward.
Calculations for any uncharged solutes which depend only on steric effects can be
151
done independently of the calculations for charged solutes. It is probable however
that many solutes may exhibit polar effects which may further complicate the
solute-solute and solute-membrane interactions. These interactions are not
considered here. A flowchart illustrating the combined procedure is shown in
Figure 5.1.
152
Figure 5.1 – Flowchart of calculation procedure for combination of charged and uncharged solutes
153
5.3.5 A note on concentration polarisation and volumetric flux
It is important to note that the model described here does not include the effect of
the ubiquitous phenomenon of concentration polarisation. This omission is
deliberate, as this numerical routine has been developed for the purpose of
inclusion in a coupled hydrodynamic and pore model using ANSYS CFX, as will
be discussed in Chapter 6. At this stage it is sufficient to note that the
concentration polarisation effects may be satisfactorily modelled outside of the
pore model (i.e. within a hydrodynamic model of the bulk flow in the membrane
channel) provided the hydrodynamic model maintains the electroneutrality
condition within the concentration polarisation boundary layer. In this way, the
conditions at the pore inlet can be obtained from the hydrodynamic solution
including the effects of concentration polarisation. The model presented here thus
predicts the real rejection of the membrane. It is emphasised all comparisons to
experimental data in this chapter are presented in terms of real rejection, which
have been obtained by other authors from observed rejections using correlations
for mass transfer coefficients [e.g. 17, 141, 148].
It is also assumed that the permeate volumetric flux JV is known, which can be
easily measured experimentally. For use in a predictive model of the membrane
filtration process using the coupled hydrodynamic / pore model approach, the flux
JV can be determined using a Darcy’s law formulation if the hydrodynamic
permeability of the membrane is known, or equivalently, a Hagen-Poiseuille
formulation.
154
5.4 Computational and numerical considerations
5.4.1 Use of existing numerical routines and libraries
As discussed in Section 5.3.3, numerical codes to allow these rejection
calculations have been developed both in MATLAB and Fortran 90, using already
available numerical routines where available. The use of robust public domain
numerical routines for general purpose calculations saves considerable time in
developing these codes, and gives confidence in the numerical predictions due to
the considerable testing and verification of such codes. A summary of the routines
used in the numerical codes is shown in Table 5.1.
Table 5.1 – Summary of existing numerical libraries and routines used in development of predictive rejection
model code
Routine / library name and language Details
fsolve (MATLAB) [165] Solvers for systems of non-linear equations. Used to solve
hybrd1 (Fortran 90, part of minpack Donnan partitioning problem for charged solutes at pore inlet and
library) [166] outlet (see Eqns. (5.18) through (5.26)).
quad (MATLAB) [165]

Numerical quadrature (integration) codes. Used to perform
qag (Fortran 90, part of quadpack
numerical integration to determine overall steric partitioning
library) [166]
coefficient for non-spherical solutes (see Eqn. (5.15)).
Boundary value problem solvers. Used to solve extended Nernst-

bvp4c (MATLAB) [164] Planck equations (see Eqns. (5.1), (5.21), and (5.26)) to determine
BVP_SOLVER(Fortran 90) [163] concentration profile throughout membrane pores and hence
determine solute rejections as unknown parameters.
155
5.4.2 Program structure
Some illustrative sample code for the MATLAB version of the numerical method
is shown in Appendix A. Briefly, there are five main routines:
x molprjn2 – this routine determines the molecular length and width for a
non-spherical solute based on the Cartesian coordinate output from the
MOPAC2009 software;
x kiso – this routine determines steric partitioning coefficients for non-
spherical solutes;
x doneqn3 – this routine calculates the Donnan partitioning equilibrium
constants for an arbitrary number of charged solutes so that the Donnan
potential for each solute can be calculated;
x np3n – this routine solves the extended Nernst-Planck equations for an
arbitrary number of charged solutes; and
x unchargedJv – this routine calculates the rejection for an arbitrary number
of uncharged solutes by solving a simplified version of the extended
Nernst-Planck equations.
There are several techniques of interest contained within these routines which will
be discussed briefly in the following.
156
5.4.3 Determination of Donnan partitioning equilibrium constants
The determination of the Donnan equilibrium constants for more than three
charged solutes must be done numerically, using existing routines to solve the
system of non-linear equations described by Eqns. (5.23) and (5.24). The most
important step to ensure these solvers work efficiently and produce the correct
solution is to provide an accurate initial estimation of the solution, which in this
case is the concentration of each solute inside the membrane pore. A reasonably
accurate estimation is made with the knowledge that the pore concentration of
counter-ions (solutes with the opposite charge to the membrane) is typically much
higher than the concentration in the bulk solution. Similarly, the concentration of
co-ions inside the pore is typically lower than that in the bulk solution. The initial
estimate of the concentration is thus given by:
zi zi
ci , y Ci , y CX , 0 (counter - ion) (5.27)
zX zX
zi 1 zi
ci , y Ci , y , !0 (co - ion) (5.28)
zX CX zX
Using this estimate was found to result in rapid convergence to a solution.
Confirmation that the true solution has been found can be obtained by using the
equilibrium constants to calculate the vector of Donnan potentials for the system,
which should be identical for every ion. The numerical routine uses this as a
check to ensure the correct solution has been found.
157
In a small proportion of cases, particularly where the membrane charge density CX
is large, the system of equations becomes close to singular, and the solver may
return a solution which yields a Donnan potential which is not equal for all ions.
In these cases, it is very difficult to provide a better estimate at the solution than
given by Eqns. (5.27) and (5.28). In this case, a crude but effective method is used
whereby the initial estimate at the solution is multiplied by a random number, and
then used as an updated initial estimate. The system of equations is again solved,
and the equality of the Donnan potential is checked. This process is repeated until
the Donnan potential equality condition is satisfied. In practice, the true solution
tends to be found within several iterations of this random multiplication method,
which is relatively small in terms of computational expense.
5.5 Results and discussion
A set of test cases for the model has been derived from experimental results
reported in the literature as follows: for rejection of uncharged compounds and
single salt solutions, the results of Bowen and Mohammad [141] have been used;
for rejection of single and multiple salt solutions, the results of Bowen and
Mukhtar [17] have been used. Finally, to demonstrate the general applicability of
the method, a theoretical investigation of the rejection of complex mixtures
containing four or more charged solutes has been undertaken.
158
5.5.1 Data from Bowen and Mohammad
5.5.1.1 Rejection of uncharged solutes
This study [141] used a CA30 cellulose acetate membrane (Hoescht Separation
Products, Germany) to investigate the rejection of a dye / salt solution. To
characterise the membrane pore radius using a spherical pore model, a series of
experiments was first performed with several uncharged compounds (glycerin,
glucose, sucrose, and Vitamin B12). These experimental results are used here to
validate the present model and present an improved characterisation of the CA30
membrane using the non-spherical pore model described previously.
The geometry of the uncharged compounds was determined using the
MOPAC2009 software (Stewart Computational Chemistry, CO, USA) using the
PM6 parameter optimisation [167]. The molecular length and width were then
determined by post-processing of the Cartesian coordinates generated by
MOPAC2009 and the associated van der Waal’s radius for each atom to obtain
the necessary molecular projections. The determined geometric parameters, as
well as the properties previously determined in [141] are shown in Table 5.2.
159
Table 5.2 - Physical properties for uncharged solutes used in [141]
Solute ri,S (nm) * Di × 109 (m2 s-1) * Li,S (nm) Wi,S (nm)
Glycerin 0.260 0.95 0.827 0.350
Glucose 0.365 0.69 1.067 0.441
Sucrose 0.471 0.52 1.138 0.515
Vitamin B12 0.740 0.33 2.244 1.109
* From [141].
In their characterisation of the membrane, Bowen and Mohammad [141] used
both the pore radius rp and the thickness-SRURVLW\ SDUDPHWHU ǻy/Ak as fitting
parameters, i.e. separate rp DQGǻy/Ak were fitted for each compound. An averaged
value of the rp was then adopted as the effective pore radius for subsequent
analysis with salt solutions. Clearly this is not ideal, as one might expect rp and
ǻy/Ak to be consistent regardless of the compound, as they are properties of the
membrane. Using a numerical fitting routine with a non-linear least-squares
algorithm, the analysis was undertaken here using a non-spherical pore model.
The results of the fitting are shown in Table 5.3, along with the previous fitting
results, and the membrane rejections as a function of flux are shown in Figure 5.2.
160
Table 5.3 – Fitting results for uncharged solutes used in [141]
Solute Previous fit * Current – non-spherical
rp (nm) ǻy/Ak (μm) rp (nm) ǻy/Ak (μm)
Glycerin 0.62 2.65 1.11 4.11
Glucose 0.61 1.92 1.11 4.11
Sucrose 0.66 0.74 1.11 4.11
Vitamin B12 0.87 0.39 1.48† 4.11†
* From [141] using spherical pore model.

†
Vitamin B12 pore radius fitted separately from other solutes.
100
90
80
70
60
Ri (%)
50
Bowen and Mohammad (1998)
40 Non-spherical pore model
Gly cerin
30 Glucose
Sucrose
Vitamin B1 2
20
10
0
0 5 10 15
6 3 -2 -1
Flux Jv × 10 (m m s )
Figure 5.2 - Comparison of experimental rejection data [141] and numerically computed rejections as a
function of flux
161
A reasonably good fit is obtained for glycerine, glucose and sucrose using fixed rp
DQG ǻy/Ak. This gives confidence that the use of a non-spherical pore model is
able to accurately characterise the membrane in terms of actual physical
parameters. It was found that it was difficult to fit the data for Vitamin B12 well
due to its large molecular size; this was dealt with in [141] by fitting a much
larger pore radius of 0.87 nm for Vitamin B12. Here, Vitamin B12 has been again
been treated separately for the sake of completeness with the same fixed ǻy/Ak as
the other solutes to obtain a larger pore radius of 1.48 nm. This should not be
interpreted as meaning that the pore radius would vary by solute. Instead, as noted
in [141], transmission of Vitamin B12 molecules probably occurs through a small
proportion of larger pores which are neglected in the current analysis by assuming
a constant pore size.
It can also be noted that the pore radius determined using the non-spherical pore
model (1.11 nm) is significantly larger than the averaged representative pore
radius of 0.63 nm previously determined [141] using a spherical pore model. This
is consistent with previous comparisons using both spherical and non-spherical
pore models [160], and the ratio of the spherical radius to the non-spherical radius
is also consistent with these results. As well, the ratio obtained for Vitamin B12
(1.48 nm : 0.87 nm) is similar.
162
5.5.1.2 Rejection of single salt solutions
In [141], solutions of sodium chloride, lithium chloride, and potassium chloride
were filtered using the same CA30 membrane to characterise the membrane in
WHUPVRIǻy/Ak and the membrane charge CX, using the representative rp = 0.63
QPREWDLQHGIURPDVSKHULFDOSRUHPRGHO,QWKLVFKDUDFWHULVDWLRQǻy/Ak differed
significantly from that obtained with uncharged solutes. As this should in theory
be a fixed physical property of the membrane, this is again somewhat less than
satisfactory.
Here, this data is revisited to characterise the membrane charge CX, using a
consistent rp DQGǻy/Ak from the characterisation using uncharged solutes with a
non-spherical pore model described in Section 5.5.1.1. For this particular case, all
the ionic solutes are monatomic, so the characterisation of the solute shapes as
spherical and hence the determination of the steric partitioning coefficient as per
Eq. (5.8) is justified. It is stressed however that the pore radius obtained using a
non-spherical pore model (1.11 nm) is used throughout so as to present a
consistent method for mixtures of monatomic (spherical) and polyatomic (non-
spherical) solutes. The physical properties of the ionic solutes are shown in Table
5.4.
163
Table 5.4 - Physical properties for ionic solutes used in [141]
Ion ri,S (nm) Di × 109 (m2 s-1)
Na+ 0.184 1.33
Cl- 0.121 2.03
K+ 0.125 1.95
Li+ 0.238 1.03
The model was first fitted in terms of a constant CX to rejection measurements for
the three salts at a fixed bulk concentration of 10 mol m-3, as shown in Figure 5.3.
A reasonably good fit was obtained with CX = 98.4 mol m-3 using the same
parameters as for the uncharged solute fitting (rp = QP ǻy/Ak = 4.11 μm).
The fitted membrane charge is significantly larger than that in [141], which is to
be expected as the pore radius is also significantly larger.
The model was then fitted to rejection data for sodium chloride at varying bulk
concentrations. In this case, the membrane charge can reasonably be expected to
vary with the ionic strength of the solution. Using the same procedure and rp and
ǻy/Ak values, the results obtained from the model fitting are shown in Table 5.5
and Figure 5.4.A reasonably good fit is again obtained using a consistent set of
parameters between the uncharged and charged solute cases. As expected, the
fitted CX are larger than the previous fit due to the larger pore radius. The
variation in CX with the bulk concentration is shown in Figure 5.5, and was found
164
to fit a linear isotherm well, as in [141]. The isotherm for the current fitted
parameters is given by:
CX 8.23CB 24.06 (5.29)
Table 5.5 – Fitting results for single salt solutions used in [141]
Previous fit Current fit

CB
Salts (mol m- rp ǻy/Ak CX rp ǻy/Ak CX
3
)* (nm) (μm) (mol m- (nm) (μm) (mol m-
3 3
) )
NaCl, KCl, 10 0.63 6.2 39.2 1.11 4.11 98.4

LiCl
NaCl 0.5 0.63 6.2 10.7 1.11 4.11 27.6
1.0 0.63 6.2 11.4 1.11 4.11† 29.3
5.0 0.63 6.2 23.6 1.11 4.11† 59.4
50.0 0.63 6.2 184.1 1.11 4.11† 468
100.0 0.63 6.2 333.6 1.11 4.11† 832
* Bulk concentration - equivalent to the wall concentration Ci,W assuming no concentration polarisation.
165
20
18
16
14
12
Ri (%)
10
6 KCl
LiCl
NaCl
4 Fitted model
2 4 6 8 10 12
Flux Jv × 106 (m3 m-2 s -1)
function of flux for single salt solutions at fixed concentration
0.5 mol m -3
1.0 mol m -3
-3
60 5.0 mol m
10 mol m -3
50 mol m -3
-3
100 mol m
50 Fitted model
40
Ri (%)
30
20
10
2 4 6 8 10 12
Flux Jv × 106 (m3 m-2 s -1)
function of flux for sodium chloride at varying concentrations
166
1000
900
800
700
600
CX (mol m-3)
500
400
300 Fitted data points

Fitted isotherm
200
100
0 20 40 60 80 100 120
-3
Bulk concentration CB (mol m )
Figure 5.5 – Variation of charge density as a function of bulk concentration using experimental data from
[141]
5.5.2 Data from Bowen and Mukhtar
This study [17] differs from the previously mentioned study [141] as different salt
solutions were used to characterise a NF membrane in terms of rpǻy/Ak, and CX,
without initial use of uncharged solutes to determine rp. The same approach is
applied here to demonstrate the consistency of the computational method.
167
5.5.2.1 Rejection of single salt solutions
In this study [17], six different NF membranes were used in a stirred cell
arrangement and rejections were experimentally measured for solutions of sodium
chloride and sodium sulphate. These observed rejections were converted to real
rejections (as used throughout here) by use of a mass transfer coefficient to
describe the effects of concentration polarisation. The membranes were initially
characterised by assuming a uniform pore radius of 1 nm and adjusting CX and
ǻy/Ak to obtain a good fit between the experimental results and their solutions to
the extended Nernst-Planck equations shown in Eq. (5.1). It should be noted that
in all simulations it was assumed by Bowen and Mukhtar [17] that steric effects
were not significant (i.e. the steric partitioning coefficients are unity), and that all
membranes have a negative effective charge (zX = -1). A slightly different
expression was also used to calculate the hindrance coefficients Ki,c and Ki,d;
however, the expressions given by Eqs. (5.9) and (5.10) have been retained here
as they are applicable for a wider range of Ȝ. Relevant physical parameters for the
sodium, chloride and sulphate ions are shown in Table 5.6, including molecular
length and width parameters derived for the polyatomic sulphate ion.
168
Table 5.6 - Physical properties for solutes from [17]
Ion ri,S (nm) * Di × 109 (m2 s-1) * Li,S (nm) Wi,S (nm)
Na+ 0.1840 1.333 - -
Cl- 0.1207 2.031 - -
SO42- 0.2309 1.062 0.566 0.271
* From [17].
Using the same approach, the present solution method was applied to model the
experimental results presented. A comparison of the experimental results of
Bowen and Mukhtar for a single membrane (the PES5 membrane) with a 10-3 M
solution of sodium chloride [17] with the rejection predicted by the present
method, both with and without steric effects, is shown in Figure 5.6, and the
resulting fitted parameters are shown in Table 5.7. Again, the steric partitioning
coefficients were determined for a spherical solute as the ions are monatomic. The
parameters were adjusted by way of a numerical fitting routine using a non-linear
least-squares algorithm. A slightly improved fit with the experimental data is
observed compared to the parameters obtained previously in [17], due to the
different expressions for the hindrance coefficients.
169
30
25
20
Ri (%)
15
10
Numerical - no steric effects
Numerical - with steric effects
5 Numerical - parameters from Bowen and Mukhtar (1996)
Experimental
5 10 15 20 25 30 35 40
6 3 -2 -1
Figure 5.6 - Comparison of experimental rejection data from Bowen and Mukhtar [17] and numerically
computed rejections for PES5 membrane for 10-3 M sodium chloride solution using ǻy/Ak and CX as fitting
parameters
Table 5.7 - Comparison of fitting for ǻy/Ak and CX for PES5 membrane used by Bowen and Mukhtar [17] for
10-3 M sodium chloride solution
Configuration ǻy/Ak (×10-5 m) CX (eq m-3) rp (nm) *
No steric effects 1.92 4.35 1
Steric effects – spherical pore model 1.47 2.85 1
Parameters from [4] – no steric effects 2.00 4.39 1
* Assumed value – not fitted.
170
The results in Table 5.7 show significant differences in the fitted parameters
depending on whether steric effects are included, indicating that steric interactions
may be important in this system. It is also possible to further characterise the
membrane in terms of the pore radius, which was previously assumed to be 1 nm.
Allowing the pore radius to be fitted as an additional parameter using the least-
squares algorithm yields the fitting parameters shown in Table 5.8. A graph of the
fit is shown in Figure 5.7.
30
25
20
Ri (%)
15
Numerical - no steric effects

10 Numerical - with steric effects
Experimental
5 10 15 20 25 30 35 40
6 3 -2 -1
computed rejections for PES5 membrane for 10-3 M sodium chloride solution using ǻy/Ak, CX, and rp as
fitting parameters
171
Table 5.8 - Comparison of fitting IRUǻy/Ak, CX, and rp for PES5 membrane used by Bowen and Mukhtar [17]
for 10-3 M sodium chloride solution
Configuration ǻy/Ak (×10-5 m) CX (eq m-3) rp (nm)
No steric effects 1.96 4.31 1.01
Steric effects – spherical pore model 1.60 3.00 1.14
Bowen and Mukhtar [17] then fitted experimental data for sodium sulphate at a
feed concentration of 3×10-3 M, and observed a larger dependence of the rejection
behaviour with the pore radius due to the larger size of the sodium sulphate
molecule. They obtained a pore radius of 0.72 nm, which was then used to refit
the data for sodium chloride at a concentration of 10-3 M. Here, the data for
sodium sulphate (3×10-3 M) is fitted with the same rp DQG ǻy/Ak as for sodium
chloride (10-3 M), which should be consistent for both solutions. However, the
charge density CX is allowed to vary, as it is expected to be a function of ionic
strength. Additionally, the non-spherical pore model can also be used for the
sulphate ion with the characteristic length and width given in Table 5.6. From this
point onwards the combination of the spherical and non-spherical pore models is
referred to simply as the non-spherical pore model. The variation in rejection is
shown in Figure 5.8, with the parameters obtained shown in Table 5.9.
172
100
90
80
70
60
Ri (%)
No steric effects
50 Steric effects - spherical
Steric effects - non-spherical
Experimental
40
30
20
10
0
0 5 10 15 20 25
6 3 -2 -1
computed rejections for PES5 membrane for 3×10-3 M sodium sulphate solution using CX as fitting parameter
Table 5.9 - Comparison of fitting for CX for PES5 membrane used by Bowen and Mukhtar [17] for 3×10-3 M
sodium sulphate solution
Configuration ǻy/Ak (×10-5 m) rp (nm) CX (eq m-3)
No steric effects 1.96 1.01 22.0
Steric effects – spherical pore model 1.60 1.14 13.8
Steric effects – non-spherical pore model 1.60 1.14 16.5
173
Several comments can be made regarding the fitting of the sodium sulphate
solution: first, a good fit is possible using the same pore radius as obtained in
Figure 5.7 for sodium chloride both with and without the inclusion of steric
effects (1.01 and 1.14 nm, respectively). This is in contrast with the results in
[17], where a significantly smaller pore radius (0.72 nm) was required to fit the
sodium sulphate data. Secondly, using a non-spherical pore model for the sulphate
ions gives very similar predicted rejections as for a spherical pore model with
consistent pore radius and membrane thickness parameters. This indicates that use
of a combination of spherical and non-spherical pore models is appropriate when
such mixtures occur, and shows the internal consistency of the approach treating
the pore radius and membrane thickness as fixed physical parameters.
Bowen and Mukhtar [17] also measured the rejections of both salts at various
concentrations for a fixed flux to determine how CX varied with concentration.
Using the same data, this procedure was repeated to obtain two isotherms
describing the variance in CX with the concentration when expressed in units of
equivalents per cubic metre: one isotherm using the spherical pore model, and the
other using the combined spherical and non-spherical pore models as described
above. The resulting isotherms are shown in Figure 5.9, and as per [17] are of the
type:
log10 C X s log10 C B q (5.30)
174
1.8
1.6 Spherical model

Non-spherical model
Spherical model isotherm
1.4 Non-spherical model isotherm
1.2
Log10 of CX (mol m-3)
0.8
0.6
0.4
0.2
0
-1 0 1 2
Log10 of bulk concentration CB (mol m-3)
Figure 5.9 - Variation of charge density as a function of bulk concentration using experimental data from [17]
The fitted parameters s and q are shown in Table 5.10. This allows for the
prediction of the rejection of mixed salt solutions of arbitrary ionic strength,
which is addressed in the following.
Table 5.10 – Fitted parameters for charge density isotherms using experimental data from [17]
Configuration s q
Steric effects – spherical pore model 0.509 0.676
Steric effects – non-spherical pore model 0.523 0.699
175
5.5.2.2 Rejection of multiple salt solutions
Bowen and Mukhtar [17] measured the rejection of mixtures of sodium chloride
and sodium sulphate at various molar ratios, and then predicted the rejection for
each ion using an isotherm to describe the charge density CX for each molar ratio.
This is an interesting test for the model as the experimental data display widely
variable rejection behaviour, including the phenomenon of negative rejection.
The current method has been applied to reproduce these results. For brevity, the
results presented here are only for the case where the combined spherical and non-
spherical model is used (depending on whether the ions are monatomic or
polyatomic). Using the membrane parameters described in Table 5.9 and Table
5.10 for the non-spherical pore model (rp = 1.14 nm, ǻy/Ak = 1.60 × 10-5 m, s =
0.523, q = 0.699), the resulting model predictions are shown in Figure 5.10 for
molar ratios of 0.2, 0.4, 0.6 and 0.8 (NaCl:Na2SO4). Very good agreement is
observed between the numerical model and the experimental data, including the
negative rejection of the Cl- co-ion.
176
100 100
(a) (b)
80 80
60 60
40 40
20 20
Ri (%)
Ri (%)
0 0
-20 -20
-40 -40
-60 -60
0 10 20 30 40 0 10 20 30 40
Flux Jv × 106 (m3 m-2 s -1) Flux Jv × 106 (m3 m-2 s -1)
100 100
(c) (d)
80 80
60 60
40 40
20 20
Ri (%)
Ri (%)
0 0 +
Na model
-
Cl model
-20 -20 SO2- model
4
+
Na experimental
-40 -40 -
Cl experimental
2-
SO4 experimental
-60 -60
0 10 20 30 40 0 10 20 30 40
computed rejections for PES5 membrane for 1×10-3 M salt solution using membrane charge density isotherm
for non-spherical pore model with molar ratios of: (a) 0.2, (b) 0.4, (c) 0.6, (d) 0.8 (NaCl:Na2SO4)
177
5.5.3 Predictions for solutions with four or more charged solutes
The model presented above is in theory capable of predicting the rejection for a
mixture containing arbitrary numbers of charged and uncharged solutes. Presented
in the following are some predictions for the rejection behaviour for solutions
containing four or more charged solutes using the combined spherical and non-
spherical pore models. It should be noted that these predictions are for theoretical
interest only as little detailed experimental data is available for such complex
mixtures.
The first solution to be modelled is a 1×10-3 M salt solution of sodium chloride,
sodium sulphate, and calcium chloride containing four different charged solutes
(Na+, Cl-, SO42-, and Ca2+). The membrane parameters used, including the charge
density isotherm, are identical to that in Figure 5.10, with four different molar
ratios (NaCl:Na2SO4:CaCl2) of 0.333:0.333:0.333, 0.25:0.25:0.5, 0.25:0.5:0.25,
and 0.5:0.25:0.25. The predicted rejections are shown in Figure 5.11. The
rejection curves are broadly similar to that observed for the three-ion case (Figure
5.10), with the addition of the Ca2+ counter-ion (rp = 0.236 nm, Di = 0.236 × 109
m2 s-1), which experiences greater rejection than the Na+ counter-ion due to its
greater valence and larger ionic radius. Negative rejection is still predicted for the
Cl- co-ion at low fluxes in all cases.
178
100 100
(a) (b)
80 80
60 60
40 40
20 20
Ri (%)
Ri (%)
0 0
-20 -20
-40 -40
-60 -60
0 10 20 30 40 0 10 20 30 40
100 100
(c) (d)
80 80
60 60
40 40
20 20
Ri (%)
Ri (%)
0 0
-20 -20 Na model

+
-
Cl model
-40 -40 SO model
2-
4
2+
Ca model
-60 -60
0 10 20 30 40 0 10 20 30 40
Figure 5.11 - Numerically computed rejections for PES5 membrane for 1×10-3 M salt solution using
membrane charge density isotherm for non-spherical pore model with molar ratios of: (a) 0.333:0.333:0.333,
(b) 0.25:0.25:0.5, (c) 0.25:0.5:0.25, (d) 0.5:0.25:0.25 (NaCl:Na2SO4:CaCl2)
179
A final demonstration of the model has been made for a solution comprising
sodium chloride, sodium sulphate, calcium chloride, potassium chloride and
lithium chloride. This solution therefore contains six separate ions (Na+, Cl-, SO42-
, Ca2+, K+, and Li+). Simulations were performed for a 1×10-3 M salt solution
using identical membrane parameters including the charge density isotherm as for
Figure 5.10 and Figure 5.11, using the molar ratios listed in Table 5.11. The
predicted rejections are shown in Figure 5.12. Again, some negative rejection is
predicted at low flux for the Cl- co-ion, with positive rejections predicted for the
SO42- co-ion and all counter-ions.
Table 5.11 – Molar ratios used for simulation of six ion 1×10-3 M salt solution
Molar fraction
Configuration
NaCl Na2SO4 CaCl2 KCl LiCl
(a) 0.2 0.2 0.2 0.2 0.2
(b) 0.4 0.15 0.15 0.15 0.15
(c) 0.15 0.4 0.15 0.15 0.15
(d) 0.15 0.15 0.4 0.15 0.15
(e) 0.15 0.15 0.15 0.4 0.15
(f) 0.15 0.15 0.15 0.15 0.4
180
100 100
(a) (b)
80 80
60 60
Ri (%)
Ri (%)
40 40
20 20
0 0
-20 -20
0 10 20 30 40 0 10 20 30 40
100 100
(c) (d)
80 80
60 60
Ri (%)
Ri (%)
40 40
20 20
0 0
-20 -20
0 10 20 30 40 0 10 20 30 40
100 100
(e) (f)
Na + model
80 80 Cl- model
SO24 - model
60 60 Ca
2+
model
K+ model
Ri (%)
Ri (%)
Li+ model
40 40
20 20
0 0
-20 -20
0 10 20 30 40 0 10 20 30 40
Figure 5.12 - Numerically computed rejections for PES5 membrane for 1×10-3 M salt solution using
membrane charge density isotherm for non-spherical pore model with molar ratios as listed in Table 5.11
181
5.6 Conclusions
The use of the extended Nernst-Planck equation in conjunction with steric pore
models for both spherical and non-spherical solutes provides a general method for
predicting the rejection performance of nanofiltration (as well as reverse osmosis)
membranes for both charged and uncharged solutes. The analysis of several sets
of experimental data from the literature showed that combining spherical and non-
spherical pore models to predict steric hindrance allows for membranes to be
characterised and modelled using a consistent set of physical parameters (namely,
the membrane pore radius rp and the membrane thickness parameter ǻy/Ak). This
allows for accurate prediction of solutions containing arbitrary mixtures of
uncharged (non-spherical) solutes, and charged solutes, both monatomic
(spherical) and polyatomic (non-spherical). The fact that good agreement can be
achieved with experimental data using consistent membrane parameters for
different solutes gives good confidence that the chosen physical parameters are
indeed a good choice for membrane characterisation. The use of isotherms to
describe the variation in membrane charge density CX in terms of the bulk
concentration (expressed in units of equivalents per cubic metre) also has been
shown to provide a robust method of predicting rejection of multiple ion
solutions.
Additionally, the use of a numerical solution to the Donnan partitioning model for
charged solutes allows for prediction of the rejection for solutions containing
182
large numbers of charged solutes with the extended Nernst-Planck equations,
which to the authors knowledge has not been described in the literature for
mixtures of more than four unique solutes. This can easily be extended to very
complex mixtures of charged and uncharged solutes as the uncharged solute
rejections can be predicted independently of those of the charged solutes.
It is also acknowledged that the DSPM, of which the present model is a small
refinement, has some limitations which can be potentially addressed by the further
inclusion of dielectric exclusion effects (i.e. the DSPM-DE). However, the
numerical techniques used for solving the equations of the present model are
directly applicable to the DSPM-DE equations, and the future extension of the
model to include these effects is a feasible and worthwhile avenue of study.
Finally, this method is suitable to be incorporated into a coupled hydrodynamic
and pore model, so that the entire predictive model of a NF or RO membrane
system may be reasonably constructed. The development of this coupled model
approach is described in Chapter 6.
183
Chapter 6
A fully predictive membrane filtration model: coupled
hydrodynamic and predictive rejection modelling
6.1 Introduction
The models described in Chapter 4 have been successful in predicting the
hydrodynamic behaviour and growth of the concentration boundary layer within a
membrane channel. However, as true predictive models they are not entirely
useful, as the permeate flux through the membrane wall is a function of the
osmotic pressure, which depends on the rejection factor of a particular solute for a
particular membrane at a particular transmembrane pressure. In all the cases
modelled so far, this rejection factor has been provided based on experimental
observations of physical membrane systems.
It would seemingly be extremely useful to be able to predict the rejection of an
arbitrary solute or combination of solutes for an arbitrary membrane without
resorting to experimental studies, and hence provide a generalised computational
model for the membrane filtration process. The numerical method for
characterisation of membranes and prediction of membrane rejection presented in
Chapter 5 allows for this aim to be achieved when coupled with a rigorous
184
hydrodynamic model, at least for NF and RO systems (and potentially some UF
systems).
This chapter describes the development of a general model which couples the
hydrodynamic models described in Chapter 4, and the predictive rejection models
described in Chapter 5, to describe the membrane filtration process in
considerable detail, and for as general a case as feasible.
6.2 Coupled modelling of solutions containing uncharged solutes
For uncharged solutes, analytical solutions to the Nernst-Planck equations are
available (as described in Section 5.3.2), and hence predicting the rejection is
fairly straightforward in computational terms. Because of this, it is possible to
perform all of the calculations for a coupled hydrodynamic-predictive rejection
model for uncharged solutes entirely within the CFX environment, using the
native Command Expression Language (CEL) to perform all calculations. The
following section describes the development of a CEL algorithm for coupling the
uncharged solute rejection model with the hydrodynamic model.
6.2.1 Model details
As described in Section 5.3.2, the prediction of uncharged solute rejections is
fairly simple. The most involved numerical procedure in the method is the
numerical integration to find the overall steric partitioning coefficient in the case
185
of a non-spherical solute, as in Eq. (5.15). This can be achieved relatively simply
using native CEL expressions within CFX, using a Simpson’s rule quadrature
method. CEL expressions are then calculated for the rejection at each
computational point on the membrane surface, using local values obtained from
the hydrodynamic solution. A flowchart illustrating the computational procedure
for the coupled model with uncharged solutes is shown in Figure 6.1.
186
Figure 6.1 – Flowchart of computation procedure for coupled hydrodynamic-mass transfer model for
uncharged solutes only
187
6.2.2 Model verification
Using the coupled model, verification was undertaken using the experimental
conditions measured by Geraldes et al. [94] for two uncharged compounds
(sucrose and PEG1000). All simulation conditions were identical to that used in
Section 4.2.5, except that now with the coupled rejection model, the solute
rejection is calculated within the hydrodynamic model based on the local
hydrodynamic conditions, rather than being explicitly specified. In addition, this
approach allows for the solute rejection to vary spatially over the computational
domain (i.e. over the membrane surface) depending on the local hydrodynamic
conditions.
The MOPAC2009 software was used to predict the solute molecule geometry for
the two test compounds as shown in Table 6.1. For all simulations, no osmotic
pressure correction factor was used, in order to properly elucidate the effect of the
spatially variable rejection model on the predicted permeate fluxes, if any.
Table 6.1 - Molecular properties for uncharged test compounds derived using MOPAC2009 software
Molecular
Solute length LS Solute width WS
Compound Molecular formula weight
(nm) (nm)
(g mol-1)
Sucrose C12H22O11 342.30 0.973 0.480
PEG1000 H-(OCH2CH2)n-OH* ~ 1000 5.459 1.549
* Generic formula for polyethylene glycol – here n = 22 to correspond with listed average molecular weight of 1000 g mol-1
188
Detailed information on the geometric properties of the CDNF50l membrane was
unavailable; however de Pinho et al. [168] performed a series of experiments in
which the average pore radius for this membrane was characterised as 0.52 nm.
This value was obtained by comparing results from a similar pore model,
assuming a spherical molecule, and experimental studies. Using a similar
approach, a pore radius of 0.79 nm gave good agreement with the experimentally
measured rejection factors for the present non-spherical model. As the rejection in
the current formulation can vary axially with the local osmotic pressure and TMP,
which depends on the hydrodynamic solution, a representative rejection value
averaged over the membrane surface was compared against the measured
rejection in order to determine the pore radius. This was done for NRe = 2000 to
correspond with the experimental conditions under which the rejection was
measured by Geraldes et al. [94].
That the use of a non-spherical model yields a larger apparent pore radius than a
spherical pore model is a fairly obvious result, which is consistent with other
results in the literature [159, 160]. In addition, the ratio of the calculated pore
radius for the spherical model to the non-spherical model (~0.6) is similar to that
obtained by Kiso et al. [159, 160]. If the Hagen-Poiseuille equation is used to
predict the pure water flux through the membrane with an assumed porosity of 0.1
as per Kiso et al. [159], a pore length of 8.9 × 10-8 m yields the same
hydrodynamic permeability as measured by Geraldes et al. [94]. This is broadly
consistent with other stated pore lengths in the literature [e.g. 159].
189
The rejection factors calculated by the pore model in conjunction with the
hydrodynamic model are shown in Table 6.2. Rejection values shown are
averaged over the membrane surface for both inlet Reynolds numbers, using the
pore radius of 0.79 nm determined using the experimental data for the higher inlet
Reynolds number of NRe = 2000. Good agreement is shown between the
experimentally determined rejection factors and the averaged values obtained
through the present model. For PEG1000, the modelled molecular dimensions are
sufficiently large to ensure complete rejection along the entire axial length (that is,
the modelled molecular shape is larger than the entrance to the membrane pores
regardless of its orientation). The experimental rejection data indicates that
PEG1000 should have rejection slightly less than unity, which suggests that either
the modelled molecular shape is too large; or, more likely, that there is some
variability in pore size within the membrane so that a small proportion of solute is
permitted through the larger pores. This issue can be addressed with the use of a
pore size distribution rather than a uniform pore size to represent the porous
membrane structure as described by Kiso et al. [160], which could be included in
future work. The axial variation in rejection for both compounds is largely
negligible, as the variation in osmotic pressure and hydrodynamic pressure is
small compared to the applied TMP for the test cases.
190
Table 6.2 - Predicted and experimentally observed rejection factors for test compounds for CDNF50l thin-
film composite NF membrane (Separem, Biela, Italy)
7UDQVPHPEUDQHSUHVVXUHǻP
Compound 1 MPa 2 MPa 3 MPa 4 MPa
Robs* Rcal † Robs Rcal Robs Rcal Robs Rcal
0.993 0.993 0.993 0.993

Sucrose 0.993 0.996 0.998 0.999
0.993 0.993 0.993 0.993
1.000 1.000 1.000 1.000

PEG1000 0.9981 0.9987 0.9993 0.9995
1.000 1.000 1.000 1.000
* Observed rejection data from [94] for NRe = 2000

†
Calculated rejection data averaged over membrane surface using pore model derived from [159] with pore size of 0.79 nm
for NRe = 500 and NRe = 2000, respectively
The permeate fluxes obtained from these simulations are shown in Figure 6.2. The
modelled permeate fluxes are almost identical to those obtained using a constant
experimentally obtained rejection factor as in Figure 4.4. As well, the predicted
CP boundary layer profiles, shown in Figure 6.3, are largely identical to those
obtained with the experimental rejection factors shown in Figure 4.5.
191
Sucrose, NRe = 500 Sucrose, NRe = 2000
6 6
5 5

4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
PEG1000, NRe = 500 PEG1000, NRe = 2000
6 6
5 5
4 4
3 3
2 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5

Figure 6.2 - Comparison of predicted permeate fluxes for CFD model with experimentally determined fluxes
of [94] using calculated spatially varying rejection factors from pore model
192
Sucrose PEG1000
0.05 0.05
NRe = 500
NRe = 500
0.04 0.04

0.03 NRe = 2000 0.03
NRe = 2000
0.02 0.02
0.01 0.01
-2.78E-17 -2.78E-17
0 10 20 30 0 10 20 30

Figure 6.3 - Comparison of dimensionless CP boundary layer profile for CFD model with computational
results of [94] using calculated spatially varying rejection factors from pore model
6.2.3 Simulations for low transmembrane pressure
To see if the axial variation in rejection would have any effect on fluxes or CP
layer thicknesses using the experimental conditions of Geraldes et al. [94], a trial
was conducted only for a lower range of TMPs ( 0.125 – 0.5 MPa), using both an
axially varying rejection, and a constant rejection obtained from the average of the
prior simulation. This exercise has been performed for the sucrose solution only,
as the modelled pore dimensions ensure complete rejection of PEG1000
193
regardless of TMP. For this case, the axial variation in rejection is shown in
Figure 6.4 as a function of the non-dimensional axial length. The predicted
rejection decreases along the channel length, as the osmotic pressure increases and
the hydrodynamic pressure decreases. No results are shown for the higher
Reynolds number case for the TMPs of 0.125 MPa and 0.25 MPa, as the osmotic
pressure in these cases is large enough to preclude positive flux through the
membrane.
Sucrose, NRe = 500

1
Predicted rejection Rcal
0.99
0.98
0.97
0 20 40 60 80 100
Sucrose, NRe = 2000
1
0.99
0.98
0.97
0 20 40 60 80 100
0.125 MPa 0.25 MPa 0.375 MPa 0.5 MPa
Figure 6.4 - Axial variation in sucrose rejection for low transmembrane pressure with spatially varying
rejection factor
194
Comparisons of the predicted membrane fluxes and CP layer profiles for the low
TMP cases are shown in Figure 6.5 and Figure 6.6, respectively. The predicted CP
layer profiles and fluxes are identical whether the rejection is varying or constant,
primarily due to the very small variations in rejection as shown in Figure 6.4.
However, if conditions were such that the axial rejection variation were greater, it
might be expected that the predicted CP layer thickness for the varying rejection
simulation is slightly decreased compared to the constant rejection. The
discrepancy in CP layer thickness would presumably grow larger as the axial
distance increases: as the rejection factor decreases along the channel length, the
hydrodynamic pressure decreases and osmotic pressure increases. Consequently,
as more solute is permitted through the membrane at increasing axial lengths, the
solute concentration at the membrane surface and hence the height of the CP
boundary layer would, in theory, decrease. Similarly, one might expect a slight
decrease in the modelled fluxes when the rejection is allowed to vary, which
becomes more significant as the axial distance along the membrane channel
increases, though this is not observed for the conditions modelled here.
195
Sucrose, NRe = 500 Sucrose, NRe = 2000
7 7
6 6

5 5
4 4
3 3
2 2
1 1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.3 0.4 0.5 0.6
Collector 1 varying Collector 2 varying Collector 3 varying

Collector 1 constant Collector 2 constant Collector 3 constant
Figure 6.5 - Comparison of predicted permeate fluxes for CFD model for low transmembrane pressures using
calculated constant and spatially varying rejection factors
Sucrose
0.08
0.07
NRe = 500
0.06
0.05
0.04
0.03
NRe = 2000
0.02
0.01
0
0 20 40 60 80 100
0.125 MPa varying 0.125 MPa constant

Figure 6.6 - Comparison of dimensionless CP boundary layer profile for CFD model for low transmembrane
pressures using calculated constant and spatially varying rejection factors
196
The inability of the spatially variable rejection factor to address the discrepancy
between modelled and experimentally measured fluxes suggests that further
mechanisms not included in the current model may be operating to reduce the
flux, separate from any variation in the intrinsic rejection of the membrane.
Determining the nature of these mechanisms and quantifying how they may be
represented without resorting to the empirical approach of the osmotic pressure
correction factor may be a useful direction of investigation for future work.
6.3 Coupled modelling of solutions containing charged solutes
To incorporate the rejection model described in Chapter 5 for charged solutes, a
significant amount of additional numerical calculations are required within the
hydrodynamic model. To do so within the native Command Expression Language
(CEL) of the ANSYS CFX software is impractical, hence the capabilities of user
Fortran are required. This is a facility of the CFX software where external code
can be written (in Fortran 77 and / or Fortran 90) and compiled to interface with
the CFX solver [123]. This allows for a great deal of extensibility to the
hydrodynamic model, though the actual coding process is to some extent tedious,
not entirely straightforward and requires significant investment of time. This
section will briefly describe the development of the user Fortran interface
developed to couple the hydrodynamic model and the rejection model.
197
6.3.1 Model details
The simplest method available to interface the CFX solver with external code is
the use of so-called User Routines. In this method, CEL expressions within the
CFX model are defined in terms of User Functions, which call User Routines
consisting of libraries compiled from the external Fortran code. The arguments
that are required to be passed to the external routine are described in Table 6.3.
Table 6.3 – Minimum arguments to be passed from CFX solver to external Fortran routine for calculating
rejection of charged solutes
Arguments Comments
Parameters specific to each solute
Mass fraction Determined at the membrane surface by the hydrodynamic model
Molecular weight User-supplied property
Valence User-supplied property
User-supplied property that can be over-written if molecular length

Stokes radius
and width are supplied
User-supplied property that can be over-written if molecular length

Diffusivity
and width are supplied
Molecular length User supplied property; if not supplied spherical pore model used
Molecular width User supplied property; if not supplied spherical pore model used
Parameters specific to the membrane
Membrane pore radius User-supplied property
Membrane charge valence User-supplied property
Membrane charge density User-supplied property
Hydrodynamic parameters
Volumetric flux Determined at the membrane surface by the hydrodynamic model
198
This method has some limitations on its generality as follows:
x The User Functions defined within the CFX model must have a fixed
number of scalar arguments. This means, for example, that separate User
Functions (and effectively separate CFX models) must be developed for
systems with two uncharged solutes, three uncharged solutes and so on.
x The User Routine is allowed to return only one scalar argument, which in
this instance is the solute rejection for one of the charged solutes. For a
system of only two charged solutes, this is feasible as the rejection of both
solutes must be identical to maintain electroneutrality, and so the second
solute rejection can simply be set to that of the first solute. However, for a
system of n > 3 solutes, the rejections are in general not identical, and so
to return the n unique rejections, n separate User Functions must be
constructed and the numerical routine must be run n times. This is clearly
not ideal as it makes the model construction convoluted, and more
importantly adds significantly to the computational expense.
These limitations could be overcome if the User Routines were allowed to accept
and return array-valued (vector) arguments, as per the Matlab sample code shown
in Appendix A. However, at this stage this functionality is unfortunately not
possible in ANSYS CFX. Nevertheless, the User Routine approach can be made
to work satisfactorily, examples of which are presented in the following.
199
The general programming approach is similar to the MATLAB model presented
in Appendix A, with separate Fortran subroutines (written in Fortran 90) used to
separately determine steric partitioning coefficients, Donnan equilibrium
coefficients, and solve the BVP numerically. To interface with the CFX solver, a
‘wrapper’ subroutine must be written in Fortran 77 to pass the input arguments to
the Fortran 90 subroutines and finally return the calculated rejection to the CFX
solver. Sample code for calculating the rejection of a sodium chloride solution is
shown in Appendix B, and a flowchart of the computational process is shown in
Figure 6.7. The computational procedure is broadly similar to that for the
uncharged case, except that the external Fortran routine is called to calculate
rejection, and additional algebraic variables must be created to record the
calculated rejections and permeate fluxes within the CFX results file, as results
returned from User Routines are not automatically incorporated as such.
200
Figure 6.7 – Flowchart of computation procedure for coupled hydrodynamic-mass transfer model for charged
solutes only
201
6.3.2 Special considerations for the coupled model
Several points should be made regarding the successful coupling of the
hydrodynamic model and the rejection model. These issues largely concern the
passing of the solute mass fractions (obtained from the hydrodynamic model
solute including concentration polarisation effects) to the external rejection
routine. Perhaps most important is the physical locale in the computational
domain where the rejection calculations are actually performed. In principle, the
coupled approach is capable of determining the rejection separately for every
single computational cell adjacent to the membrane surface, depending on the
local hydrodynamic conditions. Previous simulations in Section 6.2.2 indicated
that when permitted, solute rejection did not vary appreciably over the membrane
surface for the relatively high transmembrane pressures common in nanofiltration
and reverse osmosis applications. In addition, solving the BVP defined by the
Nernst-Planck equations to perform the rejection calculations separately at each
computational point on the membrane surface is more expensive. Despite these
drawbacks, for these coupled simulations, the rejection has been allowed to vary
spatially as described in the interests of obtaining the most general and rigorous
solution possible. If desired, however (perhaps for reasons of computational
expense or when negligible variation in rejection can be expected), an averaged or
area-averaged solute mass fraction may be passed to the external rejection routine,
so that the rejection is calculated as a constant value across the entire membrane
surface.
202
Another point of interest in coupling the models lies in the iterative solution
scheme used by CFX when solving a multicomponent fluid problem. During the
first ‘false time-step’, the solute mass fractions at the membrane surface may be
uniformly zero if the solute mass fractions are not explicitly specified throughout
the domain in the initial conditions (i.e. the membrane channel is effectively
initially filled with clean water). This causes numerical difficulties for the coupled
rejection routine. In some cases it may be desirable to specify a ‘clean-water’
initial condition, so to avoid the numerical difficulties this causes, the rejection
routine is simply bypassed until an appreciable (non-zero) solute mass fraction
appears at the membrane surface within the hydrodynamic model.
6.3.3 Model verification – two charged solutes
Using the code presented in Appendix B, the coupled model was run to reproduce
the results of Geraldes et al. [94], previously modelled using an user-assigned
rejection in Chapter 4. To do so, it was first necessary to fit appropriate membrane
parameters to the experimental rejection data so that the membrane parameters in
the pore model could be assigned.
Real rejection data for the CDNF50l membrane used by Geraldes et al. were given
for the two charged test compounds of sodium chloride and sodium sulphate [94]
as functions of the applied transmembrane pressure. The rejection values for
sodium sulphate are all close to unity, while those for sodium chloride vary
between about 0.4 and 0.7. Therefore, the data for sodium chloride were used to
203
fit the membrane parameters, as the near-unity rejections for sodium sulphate
mean that the model is relatively insensitive to variation in the parameters.
An additional difficulty in fitting the membrane parameters for this case is that no
experimental flux data was given by Geraldes et al. [94] for the sodium chloride
experiments. Instead, the applied transmembrane pressures are listed, and it is
stated that solute wall concentrations from their CFD model, using the
experimentally determined permeate fluxes as a computational boundary
condition, are used to predict the real rejections. In this case, neither the
experimental permeate fluxes for sodium chloride or the wall concentrations
determined by Geraldes et al. [94] are known. To work around this, computational
results from the CFD model presented in Chapter 4, averaged over the entire
membrane area, have been used to estimate the permeate fluxes and wall
concentrations, as shown in Table 6.4
Table 6.4 – Estimates for permeate fluxes and wall concentrations for NaCl solution based on previous CFD
results
Transmembrane pressure
Parameter
1 MPa 2 MPa 3 MPa 4 MPa
-1 -3 -3 -3
NaCl mass fraction (kg kg ) 2.39×10 3.28×10 4.33×10 5.70×10-3
NaCl concentration (mol m-
3 40.8 56.0 73.9 97.3
)
Permeate flux (m3 m-2 s-1) 1.14×10-5 2.54×10-5 3.94×10-5 5.34×10-5
Using the values shown in Table 6.4, the non-linear least squares fitting procedure
described in Chapter 5 was again used to fit membrane parameters to the rejection
204
data for sodium chloride. The obtained parameters are shown in Table 6.5, and a
comparison between the model and experimental data is shown in Figure 6.8.
Some comments are necessary regarding the fit, as the adopted parameters are
unusual: firstly, the fitted value for the pore radius is very small (0.29 nm), and is
inconsistent with the fitted value previously obtained for the same membrane
using uncharged solutes (~0.79 nm). Secondly, the membrane charge density is
very low compared to previous fitting results for other membranes. A good fit to
the experimental data was not possible using larger pore radius or membrane
charge density values. The very low membrane charge density is perhaps most
easily explainable: it may simply be that this particular NF membrane does not
exhibit significant charge effects.
However, the small pore radius is more problematic; it would be expected to be
similar to the pore radius to that previously obtained using uncharged solutes. One
explanation for the discrepancy is that the estimates used for the wall
concentration and permeate flux, obtained from previous CFD model results, do
not correspond well with those used by Geraldes et al. [94]. However, as the
values used by Geraldes et al. are unknown, these values represent a reasonable
‘best guess’. A simple sensitivity analysis was conducted on the wall
concentration and permeate flux values by increasing and decreasing the value of
each by ± 100 %, as shown in Table 6.5. The analysis indicated that this large
variation in the concentration and flux values had no effect on the fitted pore
radius, and a relatively small effect on the thickness parameter ǻy/Ak. However, the
205
fitted charge density CX was somewhat sensitive to the variation in wall
concentration. From this it is reasonable to conclude that the fitted value for the
pore radius of 0.29 nm, while very small, is acceptable, and that the previous
fitting for uncharged solutes is less applicable, as the rejections for the uncharged
solutes are generally near unity and hence less sensitive to the fitting process.
Table 6.5 – Fitted membrane parameters for CDNF50l membrane for NaCl solution
Parameters ǻy/Ak (×10-6 m) rp (nm) CX (×10-3 eq m-3)
Fitted values 4.34 0.29 1.3
Fitted values, flux +100% 1.93 0.29 0.95
Fitted values, flux -100% 7.05 0.29 0.75
Fitted values, wall 4.34 0.29 54.4

concentration+100%
Fitted values, wall concentration- 4.34 0.29 1.1

100%
206
100
90
Numerical - with steric effects
Experimental
80
70
60
Ri (%)
50
40
30
20
10
0
1 2 3 4 5 6
5 3 -2 -1
Figure 6.8 – Comparison of predicted rejection and experimental rejection for NaCl solution with CDNF50l
membrane using fitted parameters from Table 6.5
The coupled model was then run using these fitted parameters for the lower
entrance Reynolds number of 500 for transmembrane pressures of 1, 2, 3, and 4
MPa as per previous simulations.
The results of the simulations are shown in Figure 6.9 in terms of the predicted
rejection and permeate flux when averaged over the membrane area. The spatial
variation in the predicted rejection is also shown in Figure 6.10. A small decrease
in rejection is apparent along the axial flow direction, which becomes less
significant as the transmembrane pressure is increased. This rejection variation
207
along the channel is in the order of 5 % for an applied transmembrane pressure of
1 MPa, decreasing to approximately half a percent for the 4 MPa case. The
variation is less apparent at higher transmembrane pressures as the retarding effect
of the osmotic pressure developed within the concentration polarisation boundary
layer is less significant compared to the applied pressure.
The rejections predicted by the coupled model correspond almost exactly to the
previous predictions by the pore model only, which confirms the external
rejection routine is producing reasonable results. The comparison to the
experimental rejection data is essentially the same as before, and produces a
reasonable fit.
It is interesting to note however that the permeate fluxes predicted by the coupled
model for the given transmembrane pressures are slightly less than the permeate
fluxes extrapolated from previous (uncoupled) CFD model results, as shown in
Table 6.4. The reason for this disparity is due to the previous CFD model results
assuming a constant rejection for the entire membrane channel, whereas the
current (coupled) model allows for the rejection to vary spatially across the
membrane surface. However, refitting the pore model with the modified fluxes
predicted by the coupled model produces almost no variation in the fitted
membrane parameters, as would have been expected based on the sensitivity
analysis previously conducted (shown in Table 6.5).
208
100
90
80
70
60
Ri (%)
50
40
30 Numerical - with steric effects

Experimental
20 CFD model results
10
0
1 2 3 4 5 6
Flux Jv × 105 (m3 m-2 s -1)
Figure 6.9 – Comparison of experimental rejection (unfilled markers), predicted rejection from pore model
(solid line) and predicted rejection from coupled CFD-pore model (filled markers) for NaCl solution with
CDNF50l membrane using fitted parameters from Table 6.5
209
1 MPa
0.490
0.480
0.470
0.460
0.450
0.440
0 10 20 30 40 50 60 70 80 90 100
2 MPa
0.615
0.610
0.605
0.600
0.595
0.590
0.585
0 10 20 30 40 50 60 70 80 90 100
3 MPa
0.660
0.655
0.650
0.645
0.640
0 10 20 30 40 50 60 70 80 90 100
4 MPa
0.678
0.676
0.674
0.672
0.670
0.668
0 10 20 30 40 50 60 70 80 90 100
Figure 6.10 – Predicted axial rejection variation from coupled CFD-pore model for NaCl solution with
CDNF50l membrane using fitted parameters from Table 6.5
The results can also be analysed once again in terms of the observed concentration
polarisation boundary layers as per Section 4.2.5. A comparison of the
concentration polarisation layer thicknesses predicted by the present coupled
model and the model in [94] is shown in Figure 6.11. Despite the discrepancy
210
between the predicted and experimental rejections, the concentration polarisation
layer thicknesses remain close to that predicted by Geraldes et al. [94] – i.e. the
polarisation layer is less sensitive to the rejection than it is to the hydrodynamic
conditions within the membrane channel. Nevertheless, the coupled approach is
surely a step forward in a rigorous method for simulating concentration
polarisation and mass transfer, and this comparison confirms that the coupled
model approach still allows for the concentration polarisation boundary layer to
be developed accurately.
0.06
0.05
0.04
0.03
1 MPa model
1 MPa Geraldes et al.
2 MPa model
0.02 2 MPa Geraldes et al.
3 MPa model
4 MPa model
0.01
0
0 10 20 30
Figure 6.11 - Comparison of dimensionless CP boundary layer profile for coupled CFD-pore model with
computational results of [94]
211
6.3.4 Simulations – multiple charged solutes
The coupled model is also capable of predicting the hydrodynamic and rejection
behaviour of systems of an arbitrary number of charged solutes. However,
detailed hydrodynamic measurements of such multicomponent systems (such as
the concentration polarisation boundary layer formation and the variation in solute
concentrations through the membrane) are not available in the literature. In
addition the costs of specialised apparatus to measure and visualise such systems
accurately are prohibitively expensive for this project. Thus, a series of numerical
experiments have been performed for solutions with multiple charged solutes for
several reasons:
x To verify the rejection calculations produced by the coupled model for
multiple charged solutes agree with previous calculations and
experimental data; and
x To investigate general trends and hydrodynamic phenomena that arise
from the coupled hydrodynamic / multiple solute approach, which have
not been previously investigated.
We first attempt to verify the coupled hydrodynamic rejection model against the
rejection results of Bowen and Mukhtar [17] for mixtures of sodium chloride and
sodium sulphate solutions using the PES5 membrane. These were previously
modelled using the rejection model only in Section 5.5.2.2.
212
6.3.4.1 Verification (without concentration polarisation)
To use the coupled model, the hydrodynamics of the membrane channel must be
described. The experimental setup used in [17] was that of a cylindrical stirred
cell (an Amicon Model 8200 cell of 200 mL capacity) – the stirring of the cell
minimises the formation of the concentration polarisation layer. To explicitly
model the hydrodynamics of the cell caused by the motion of the stirrer would be
prohibitively difficult, so it is assumed that the cell can be well described by a 2D
membrane channel with a permeable lower wall, with negligible concentration
polarisation occurring. The lack of concentration polarisation is modelled by
simply setting the transport properties of the solutes (the ions) to be identical to
that of the solvent (water) for the bulk hydrodynamic simulation.
The channel has equivalent height to the cell (approximately 66 mm), and a
channel length equal to that of the cell diameter (62 mm). The fluid velocity
across the membrane surface within the stirred cell may be assumed to be small,
and so can be modelled by setting a small inlet velocity for the fictive 2D channel.
For these simulations, the inlet Reynolds number was set to an arbitrary small
value of 10. It is stressed that the intent of these simulations is not to construct an
exhaustive hydrodynamic model of the stirred-cell arrangement. Rather, the intent
is to provide broadly similar hydrodynamic conditions by which the external
rejection routine may be tested by comparing the average rejections across the
membrane surface. For these purposes, the simplified geometric and flow
configuration described here is sufficient.
213
In addition, the hydrodynamic permeability of the membrane is required. The
permeability has been obtained from pure water flux experiments conducted by
Bowen and Mukhtar [17]; our analysis of the data indicates the permeability of
the PES5 membrane to be approximately 7.76 × 10-11 m².s kg-1.
The model was then run with the parameters previously described in Section
5.5.2.2 for a range of transmembrane pressures as used in [17]. A detailed listing
of the simulation parameters is shown in Table 6.6.
Table 6.6 – Model input parameters for simulations corresponding with Bowen & Mukhtar [17]
Parameter Value
Membrane channel length L 62 mm
Membrane channel width w Variable depending on mesh density – equal to size of

smallest mesh element
Transmembrane pressure 50, 100, 200, 300, 400 kPa
Inlet Reynolds number NRe 10
Membrane pore radius rp 1.14 nm
Solution bulk concentration CB 1 mol m-3
Molar ratio (NaCl:Na2SO4) 0.2, 0.4, 0.6, 0.8
Membrane charge density CX 6.80, 6.39, 5.96, 5.50 mol m-3 *
0HPEUDQHWKLFNQHVVSDUDPHWHUǻy/Ak 1.60 × 10-5 m
* Variable according to solution ionic strength; described by isotherm in Eqn. (5.30) with parameters
s = 0.523, q = 0.699.
214
6.3.4.2 Verification results (without concentration polarisation)
An initial verification of the coupled model results can be obtained by comparing
the rejections measured and predicted in [17] to the rejections predicted by the
coupled model, when averaged over the membrane surface. This comparison is
shown in Figure 6.12, with the area-averaged rejections from the coupled
hydrodynamic / rejection model shown with filled markers. The results indicate
good agreement with the experimental data, indicating that the simplifying
assumptions made about the cell hydrodynamics are justifiable, and that the
coupled rejection routine returns reasonable results for multiple charged solutes.
There is a small discrepancy between the flux values predicted by the coupled
model and the experimentally measured values for each transmembrane pressure.
The values of the transmembrane pressures used for the experimental
measurements in [17] are not explicitly stated, but it may be reasonably assumed
that the pressures were as listed in Table 6.6 (50, 100, 200, 300, and 400 kPa),
since they correspond with the values listed in [17] used to determine the
hydrodynamic permeability of the membrane. In any case, this discrepancy in flux
is not significant in this case, as the intent is simply to verify the validity of the
rejection routine within a reasonable amount of accuracy. It is likely that the flux
discrepancy simply reflects the simplifying assumptions made about the
hydrodynamics of the stirred cell. Further, it is reasonable to infer that if a
rigorous attempt to model the stirred cell hydrodynamics were made, it would
215
result in higher velocities across the membrane surface, thus increasing the
predicted permeate flux, which would reduce the apparent discrepancy.
100 100
(a) (b)
80 80
60 60
40 40
20 20
Ri (%)
Ri (%)
0 0
-20 -20
-40 -40
-60 -60
-80 -80
0 10 20 30 40 0 10 20 30 40
6 3 -2 -1 6 3 -2 -1
Flux Jv × 10 (m m s ) Flux Jv × 10 (m m s )
100 100
(c) (d)
80 80
60 60
40 40
20 20
Ri (%)
Ri (%)
Na + model
Cl- model
0 0
SO2- model
4
-20 -20
+
Na experimental
-
Cl experimental
SO2-
-40 -40 4
experimental
Na + CFD
-60 -60 Cl- CFD
SO2-
4
CFD
-80 -80
0 10 20 30 40 0 10 20 30 40
6 3 -2 -1 6 3 -2 -1
Flux Jv × 10 (m m s ) Flux Jv × 10 (m m s )
Figure 6.12 - Comparison of experimental rejection data from Bowen and Mukhtar [17], numerically
computed rejections from pore model, and area-averaged predicted rejections from coupled hydrodynamic /
pore model for PES5 membrane for 1×10-3 M salt solution with molar ratios of: (a) 0.2, (b) 0.4, (c) 0.6, (d)
0.8 (NaCl:Na2SO4)
216
Even though concentration polarisation effects are neglected for this model, some
spatial variation in rejection over the membrane surface is apparent for the
relatively low transmembrane pressures encountered here. The variation in
rejection for these cases is generally in the order of several percent. This is due to
the axial decrease in flux along the membrane channel due to pressure drop from
the permeable wall condition (as discussed in Section 4.2.3). Some representative
graphs of this rejection variation are shown for a TMP of 200 kPa, with a molar
ratios (NaCl:Na2SO4) of 0.2 in Figure 6.13. The variation shown here is typical of
the other cases simulated.
Additionally, the pressure drop through the membrane channel should cause an
axial decrease in flux along the membrane surface. This turns out to be almost
negligible for these conditions in the absence of concentration polarisation – the
variation is typically less than half a percent of the average flux. A representative
example of the spatial variation in permeate flux is shown in Figure 6.14 for the
molar ratio of 0.2, which is typical of the other cases simulated.
217
Na+
0.800
0.795
0.790
0.785
0.780
0.775
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
SO42-
0.900
0.895
0.890
0.885
0.880
0.875
0.870
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Cl-
-0.200
-0.220
-0.240
-0.260
-0.280
-0.300
-0.320
-0.340
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 6.13 – Predicted axial rejection variation for PES5 membrane for 1×10-3 M salt solution with molar
ratios of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa
218
25
20 50 kPa
100 kPa
200 kPa
300 kPa
Predicted permeate flux J V × 10 6 (m3 m-2 s-1 )
15
10
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 6.14 – Predicted axial permeate flux variation for PES5 membrane for 1×10-3 M salt solution with
molar ratio of 0.2 (NaCl:Na2SO4)
The results of the verification study in the absence of concentration polarisation
indicate the coupled rejection model is working as expected The non-negligible
axial variation in solute rejection indicates that use of the spatially variable
coupled rejection model is desirable in order to perform a rigorous simulation of
219
the membrane system, particularly at lower transmembrane pressures. The
coupled model may now be extended fully by considering the effects of
concentration polarisation by re-including spatially variable, concentration-
dependent transport properties as described in the following section.
6.3.4.3 Simulations with concentration polarisation effects
We now perform a series of numerical experiments for the coupled multiple
charged solute model with concentration polarisation effects. First, we revisit the
simple hydrodynamic model of the stirred cell used in [17] as described in Section
6.3.4.2, but now add a description of the mixed solution transport properties to
include the phenomenon of concentration polarisation. This allows us to compare
the relative effect of concentration polarisation for a known case. A more
extensive set of simulations for more realistic membrane channel geometries will
be performed in Chapter 7. This allows for identification of general features of the
hydrodynamics and mass transfer predicted by the coupled model for a typical
membrane channel.
The use of mixed salt solutions provides some additional difficulty in estimating
the transport parameters of the mixed solution. As discussed in Section 3.4,
rigorous simulation of the diffusion of components in solution using the Maxwell-
Stefan approach [125] or the Onsager approach [126] is possible, but largely
infeasible due to the lack of experimental data for determining diffusion
coefficient matrices. In addition, the use of the Nernst-Planck equations to
220
describe the diffusion of components within the bulk solution has been shown to
cause severe numerical difficulties when coupled with a commercial CFD model
[103]. In particular, the only way such a multi-component diffusion model can be
incorporated for the entire flow domain in ANSYS CFX is through use of an
additional source term as used by Fimbres-Weihs and Wiley [103], which was
shown to cause numerical difficulties. For these reasons, a simplified Fick’s law
approach has been adopted for these simulations, and is detailed in the following.
Previously relationships established by Geraldes et al. [94] were used to describe
the viscosity, osmotic pressure, and diffusivity of sodium chloride and sodium
sulphate solutions as functions of the salt mass fractions. It is difficult to find
references in the literature that describe similar relationships for mixed solutions
of sodium chloride and sodium sulphate. For these simulations some empirical
approximations have been made to estimate these relationships as follows.
The first property to consider is the diffusivity of the solution. In principle, the
diffusivity of each ion in the system could be described independently as per the
Fick’s law approximation; however, preliminary tests indicated that this made it
difficult to ensure the electroneutrality condition within the bulk solution and at
the membrane surface was met. Instead, the multicomponent fluid was considered
to consist of four separate ions (Ion 1: Na+, Ion 2: Cl-, Ion 3: Na+, Ion 4: SO42).
The diffusivity of Ions 1 and 2 was given by the relationship derived by Geraldes
et al. [94] for sodium chloride, and that of Ions 3 and 4 was given by the similar
relationship for sodium sulphate. That is:
221
DIons 1,2 u109 (m 2s -1 ) 1.61>1 - 14 ZIon 1 ZIon 2 @ for ZIon 1 ZIon 2 < 0.006;
(6.1)
1.45 for ZIon 1 ZIon 2 > 0.006
>
DIons 3,4 u109 (m 2s -1 ) 1.23 1 - 0.76 ZIon 3 ZIon 4
0.4
@ (6.2)
Effectively this represents the scenario where the salts do not dissociate in
solution, and each salt diffuses independently of the other. While this is of course
not a rigorous representation of the real ionic transport within the bulk flow, it is a
reasonable approximation to make in light of the difficulties encountered in a full
multi-component diffusion model for the bulk flow. This ensures that the
electroneutrality condition is upheld, which allows the predictive rejection model
described in Section 5.3.3 to determine the mass transfer through the membrane in
terms of a source applied at the boundary only (rather than throughout the entire
domain). If possible, further work to determine the variation in mixed salt solution
diffusivity as a function of concentration would be beneficial.
To describe the solution viscosity, an average is made of the viscosities calculated
using the relationships described by Geraldes et al. [94]:
0.89>1 - 1.63 ZIon 1 ZIon 2 @ 0.89>1 - 3.52 ZIon 3 ZIon 4 @

Psolution u103 (Pa s) (6.3)
2
Again, this is satisfactory for this investigation, though further description of the
viscosity variation of multicomponent solutions would be of future interest.
222
Finally, the osmotic pressure exerted by the mixed salt solution is simply taken as
the sum of the osmotic pressures exerted by the sodium chloride and sodium
sulphate solutions in isolation:
3 solution u 10-5 (Pa ) 805.1ZIon 1 ZIon 2 337.8ZIon 3 ZIon 4

0.95
(6.4)
The extension to more complex multicomponent solutions is fairly obvious,
provided relationships are available for the variation in these transport properties
as functions of the mass fractions of each solute in isolation. Clearly this is not as
rigorous a description of the transport of ions within the bulk membrane flow as
would be provided by a Nernst-Planck approach as per Fimbres-Weihs and Wiley
[103], but on the other hand it avoids the large computational expense in doing so,
and the associated numerical difficulties as the governing equations become very
stiff. A flowchart of the computational procedure when using this simplified
diffusion model for multicomponent flow is shown in Figure 6.15.
223
multiple charged solutes with simplified diffusion model
224
Besides the approach for describing the solution transport properties shown
above, all parameters (e.g. geometries, boundary conditions etc) used for the
simulations with concentration polarisation are identical to those in Section
6.3.4.1. For brevity the simulation results presented here are only for Case (a),
with a molar ratio (NaCl:Na2SO4) of 0.2.
6.3.4.4 Simulation results (with concentration polarisation)
The coupled model including concentration polarisation effects was generally able
to produce well-converged solutions. However, the increased stiffness of the mass
transport equations due to the inclusion of the concentration polarisation effects
leads to a fairly significant increase in computational expense: simulations with
identical meshes and computational parameters generally required at least three
times (and sometimes up to 20 times) as many iterations to converge to a steady-
state solution, compared to the case with no concentration polarisation.
However, it was difficult to obtain a converged solution for the highest
transmembrane pressure simulated of 400 kPa, even with very refined meshes and
small (false) timestepping. Internal model computations appeared to indicate for
this transmembrane pressure that the rejection of the chloride co-ion varied axially
between relatively large negative and positive rejection values (approximately
-0.28 to 0.25). This appears to have caused severe numerical difficulties for the
CFX solver in using the rejection to calculate the mass source / sink term to be
applied at the membrane boundary. That is, with the large axial variation in
225
rejection causing both positive and negative flux of chloride ions along the same
boundary in the hydrodynamic model, the mass transfer equations appear to have
become very stiff. It should be noted that this particular scenario (where large
variations in both positive and negative rejection are observed for the same ion
within a membrane channel depending on the axial position) is in part reflective
of the low crossflow velocity adopted for these simulations, and is probably
unlikely to occur in practical membrane channels. However, further work to
determine the cause of the numerical failure of the solver routine would be
beneficial.
The computational results indicate that the effect of the concentration polarisation
boundary layer is very significant at the low crossflow velocity investigated here.
A comparison of the predicted area-averaged rejections from the coupled CFD
model both with and without concentration polarisation effects is shown in Figure
6.16. While the predicted fluxes are slightly less for a given transmembrane
pressure, there is a significant reduction in rejection when concentration
polarisation effects are included. This is due to the increased viscosity and
osmotic pressure near the membrane surface, which retard the permeation through
the membrane; and due to the increased solute mass fractions at the membrane
surface, which reduce the real rejection of the membrane.
(It should be noted that the results from the coupled CFD model with
concentration polarisation are not intended to be compared to the experimental
results in this case. The large discrepancy between the predicted rejections using
226
the coupled CFD model with concentration polarisation, and the experimental
results, is not a reflection of any failing on the part of the CFD model. This is
because the CFD model has not attempted to model the complex hydrodynamics
of the stirred cell arrangement as previously discussed, which in the absence of
concentration polarisation is an acceptable approximation. Clearly, when CP
effects are included, this approximation is invalid. Instead, this comparison is
intended to illuminate the effect of including concentration polarisation on
rejection behaviour by comparing a given (theoretical) case without concentration
polarisation, and the same (theoretical) case with concentration polarisation.)
227
100
(a)
(d)
80
60
40
20
Ri (%)
Na + model
0 Cl- model
SO24 - model
Na + experimental
-20 Cl- experimental
SO24 - experimental
Na + CFD
-40 Cl- CFD
SO24 - CFD
Na + CFD with CP
-60 Cl- CFD with CP
SO24 - CFD with CP
-80
0 5 10 15 20 25 30 35 40
6 3 -2 -1
Figure 6.16 - Comparison of experimental rejection data from Bowen and Mukhtar [17], numerically
computed rejections from pore model, and area-averaged predicted rejections from coupled hydrodynamic /
pore model both with and without concentration polarisation effects for PES5 membrane for 1×10-3 M salt
solution with molar ratio of 0.2 (NaCl:Na2SO4)
It is also instructive to examine the axial variation in the predicted rejection and
permeate flux with concentration polarisation included in the CFD model.
Previously, in the absence of concentration polarisation, the axial rejection
variation was in the order of several percent. With concentration polarisation
accounted for, the axial variability in rejection is much more pronounced at the
channel inlet as the concentration polarisation layer begins to develop. In some
228
cases, the variation is as much as 25 to 30 percent for the positively rejected ions
(Na+ and SO42-). Some representative graphs of this rejection variation are shown
for a TMP of 200 kPa, in Figure 6.17.
The predicted axial variation in permeate flux for varying transmembrane
pressures is shown in Figure 6.18. Compared to the results without concentration
polarisation, the decrease in flux along the axial length of the channel is much
more pronounced. The variation is as great as approximately 10% of the average
permeate flux for the lowest simulated transmembrane pressure (50 kPa). This
indicates that the inclusion of concentration polarisation effects can have
significant effects on the predicted permeate flux, as well as the predicted
rejection, and can further explain the large variability in predicted rejection in the
case. (That is, due to the axial decrease in permeate flux, the predicted rejections
also decrease axially in accordance).
229
Na+
0.850
0.800
0.750
0.700
0.650
0.600
0.550
0.500
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
SO42-
0.950
0.900
0.850
0.800
0.750
0.700
0.650
0.600
0.550
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Cl-
-0.200
-0.250
-0.300
-0.350
-0.400
With CP
-0.450
Without CP
-0.500
-0.550
-0.600
-0.650
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 6.17 – Predicted axial rejection variation for PES5 membrane both with and without concentration
polarisation effects for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa
230
25
300 kPa
20
200 kPa
15
With CP
Without CP
10
100 kPa
50 kPa
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 6.18 – Predicted axial permeate flux variation for PES5 membrane both with and without
concentration polarisation effects for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4)
231
It is also interesting to examine the computed concentration fields for each ion,
including the concentration polarisation boundary layer, in the presence of a
negatively rejected ion. A selection of representative graphs of the variation in
concentration in the vicinity of the membrane surface (normalised relative to the
bulk inlet concentration) are shown in Figure 6.19. For the positively rejected
ions, the classical concentration polarisation boundary layer is formed above the
membrane surface as expected. However, for the negatively rejected chloride co-
ion, an ‘inverse’ concentration polarisation boundary layer is formed, where the
concentration at the membrane surface is less than that in the bulk of the cell (the
spectrum between blue and purple representing the inverse concentration
polarisation layer).
232
(a)
(b)
(c)
Figure 6.19 – Detail of computed concentration fields (normalised relative to bulk inlet concentration) of (a)
Na+, (b) Cl-, (c) SO42- for PES5 membrane with concentration polarisation effects for 1×10-3 M salt solution
with molar ratio of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa
233
This is shown in more detail in Figure 6.20, which shows the variation in solute
concentration (normalised relative to the inlet bulk concentration) in the normal
direction to the membrane surface, at the centre of the cell. The concentration of
the negatively rejected chloride co-ion is seen to increase with increasing distance
from the membrane surface. Figure 6.20 also illustrates that the thickness of the
concentration polarisation layer also varies slightly between ions due to their
varying transport properties. This suggests that, while not as rigorous as a full
ionic diffusion model for the bulk flow (used with partial success by [103]), the
coupled CFD-mass transfer model using the diffusivities of individual salts may
offer a more physical representation of the concentration polarisation layer than a
simple empirical correlation (e.g. a film layer model).
1.700
Normalised mass fraction (Ȧi /Ȧi,B)
1.500
Cl-
Na+
1.300
SO42-
1.100
0.900
0.700
0.001 0.01 0.1
Dimensionless normal distance from membrane (y/2h)
Figure 6.20 – Predicted variation normal to membrane surface for solute concentration normalised relative to
inlet bulk concentration for PES5 membrane with concentration polarisation effects for 1×10-3 M salt solution
with molar ratio of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa
234
6.4 Coupled modelling of mixed solutions
We have seen how the coupled hydrodynamic / rejection model approach is able
to describe in detail the behaviour of membrane systems with arbitrary numbers of
charged or uncharged solutes. For a practical feed solution, there may be a
combination of charged and uncharged solutes; in this section, the previous
modelling approaches are combined with further User Fortran to provide a general
predictive model which addresses this situation.
6.4.1 Model details
The implementation of the combined charged-uncharged solute model is fairly
straightforward. Again, to ensure electroneutrality conditions are met and
avoiding the numerical effort and difficulty of a full diffusion model for the entire
flow domain, a simplified Fick’s law approach is used to describe diffusion of
salts as described in Section 6.3.4.3. Similarly for uncharged solutes, the diffusion
of each uncharged solute is assumed to occur independently of other solutes
(charged or uncharged), and hence the diffusivity coefficient in water for each
uncharged solute is used to describe this. The independent diffusion of an
arbitrary number of uncharged solutes therefore does not violate the
electroneutrality requirement for the charged solute rejection routine. Little data is
available to describe the diffusion of multiple uncharged solutes in a single
solution, so this approach is a reasonable approximation. It is noted however that
in principle the diffusion of charged and uncharged solute could also be described
235
by a rigorous diffusion model for the entire flow domain, as in [103], though the
computational difficulties in this approach remain prohibitive. In addition, this is a
potential area for future research which may yield some benefit for membrane
CFD modelling.
The approach used to describe the viscosity of the solution and the osmotic
pressure generated is the same as that of Section 6.3.4.3. That is, expressions for
the viscosity and osmotic pressure as a function of the mass fraction for single-
solute solutions (corresponding to the components of the multicomponent
solution) are used, and:
x The calculated viscosities are averaged to obtain a representative viscosity
for the multicomponent solution; and
x The exerted osmotic pressures are summed to obtain an overall osmotic
pressure.
A flowchart demonstrating the computational procedure for this general model
with a mixed solution of charged and uncharged solutes is shown in Figure 6.21.
236
multiple charged and/or uncharged solutes with simplified diffusion model
237
To demonstrate the model, the scenario used in Section 6.3.4.3 has been revisited
(the simplified stirred-cell arrangement of Bowen and Mukhtar [17] including
concentration polarisation effects for a NaCl:Na2SO4 mixture). However, an
additional uncharged solute (sucrose) has been included to demonstrate the
efficacy of the numerical routine and to briefly examine the predicted effects on
the membrane channel behaviour. The properties used for the additional
uncharged solute (sucrose) are shown in Table 6.7.
Table 6.7 – Properties for additional uncharged solute (sucrose) in coupled simulation of charged-uncharged
solute mixture
Parameter Value
Solute length LS 1.137 nm
Solute width WS 0.513 nm
Bulk mass fraction + 0.0001 kg kg-1 (~ 0.291 mol m-3)
Single-solute viscosity * 0.89(1 + 1.31Ȧi + 16.83Ȧ i 2) × 103 (Pa s)
Single-solute diffusivity * 0.52(1 í 1.33Ȧi) × 109 (m2 sí1)
Single-solute osmotic pressure * 72.18Ȧi(1 + 0.94Ȧi + 2.93Ȧ i 2) × 10í5 (Pa)
* Relationship obtained from [94].

+
Value selected to be similar to concentration of salts.
6.4.2 Simulations
Simulations were run for the conditions previously described in Section 6.3.4.3
with the addition of sucrose as an uncharged solute. For brevity, the results
presented here are only for the molar ratio (NaCl:Na2SO4) of 0.2. Well converged
solutions were obtained for all cases, even for the highest transmembrane pressure
of 400 kPa, which previously proved difficult to obtain converged solutions for in
238
the case examined in Section 6.3.4.3. The reason for the improved numerical
stability with the addition of an uncharged solute is unknown and difficult to
speculate upon.
A comparison of the area-averaged predicted rejections and permeate fluxes for
the various transmembrane pressures (50 kPa, 100 kPa, 200 kPa, 300 kPa, 400
kPa) is shown in Figure 6.22. Here a comparison is made to the case presented in
Section 6.3.4.4 to examine the additional effect of the uncharged sucrose molecule
on the rejection of the other (charged) components and the volumetric permeate
flux. Examination of Figure 6.22 reveals that the addition of sucrose has a
retarding effect on the permeate flux in all cases. This is an expected result, as the
higher viscosity within the concentration polarisation layer from the addition of
sucrose, as well as the additional osmotic pressure exerted by the sucrose
molecules, both act to reduce the permeate flux in accordance with Eq. (4.1).
239
100
80
60
40
Sucrose
Na +
20 Cl
-
Ri (%)
SO2 -
4
Na + without sucrose
0 -
Cl without sucrose
SO2 - without sucrose
4
-20
-40
-60
-80
0 5 10 15 20 25 30 35 40
6 3 -2 -1
Figure 6.22 - Comparison of area-averaged predicted rejections from coupled hydrodynamic / pore model
with concentration polarisation effects for PES5 membrane for 1×10-3 M salt solution with molar ratio of 0.2
(NaCl:Na2SO4) with and without sucrose addition
The variation in the predicted rejection between the two cases is also
commensurate with this trend. If the rejection model is considered in isolation, the
calculated ionic rejections should remain constant with the addition of any
number of uncharged solutes. However, when considered in the coupled
hydrodynamic-mass transfer model, the differing hydrodynamic conditions within
the membrane channel and particularly within the concentration polarisation layer
present a different set of inlet boundary conditions to the ionic rejection model.
240
That is, the inclusion of the uncharged solutes within the hydrodynamic model
yields slightly different velocity, pressure, and concentration fields, which affect
the coupled rejection calculations. This is expressed in the simulation results by
the magnitudes of the predicted rejections in the presence of sucrose being slightly
less than predicted for a purely electrolytic solution. Specifically, the rejections of
the positively rejected ions (Na+ and SO42-) are smaller positive values when
sucrose is added, and the rejections of the negatively rejected Cl- co-ion are
smaller negative values when sucrose is added. Finally, as expected, the predicted
rejection of sucrose monotonically increases with increasing transmembrane
pressure.
An examination of the axial variation in the predicted rejections is shown in
Figure 6.23, again only for the transmembrane pressure of 200 kPa. The general
trend in axial rejection variation with sucrose addition is very similar to that
without sucrose, though the addition of sucrose tends to slightly reduce the
rejection as previously noted. There is also some axial variation in the rejection of
sucrose itself, though this variation is small in magnitude compared to that of the
charged solutes.
A comparison of the axial variation in permeate flux is also shown in Figure 6.24.
Again, the rapid initial decrease in flux near the channel in inlet is observed,
followed by a very gradual decrease along the remainder of the channel. As
previously noted, the addition of sucrose causes a slight decrease in flux due to
241
the higher viscosity of the mixed solution and the greater osmotic pressure
exerted.
A representative detail of the concentration fields predicted for each solute, again
normalised relative to the inlet concentration, in the vicinity of the membrane
surface is shown in Figure 6.25. Again, the inverse concentration polarisation
layer for the negatively rejected chloride co-ion is observed, while the relative
variation in concentration is observed to be smaller for sucrose than the other
components. The variation in solute concentration (normalised relative to the inlet
concentration) in the normal direction to the membrane surface, at the centre of
the cell is also shown in Figure 6.26. This shows the variation in the thickness of
the concentration polarisation for sucrose compared to the other components more
clearly. The variation for the other components remains similar to that observed in
the absence of sucrose.
242
Na+
0.730
0.710
0.690
0.670
0.650
0.630
0.610
0.590
0.570
0.550
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
SO42-
0.800
0.750
0.700
0.650
0.600
0.550
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
-
Cl
-0.550
-0.560
-0.570 With sucrose

-0.580 Without sucrose
-0.590
-0.600
-0.610
-0.620
-0.630
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Sucrose
0.708
0.707
0.706
0.705
0.704
0.703
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 6.23 – Predicted axial rejection variation for PES5 membrane for 1×10-3 M salt solution with molar
ratio of 0.2 (NaCl:Na2SO4) for TMP of 200 kPa with and without sucrose addition
243
35
With sucrose
Without sucrose
30 400 kPa
25
300 kPa
20
15 200 kPa
10
100 kPa
5
50 kPa
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 6.24 – Predicted axial permeate flux variation for PES5 membrane both with and without sucrose
addition for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4)
244
(a)
(b)
(c)
(d)
Figure 6.25 – Detail of computed concentration fields (normalised relative to inlet bulk concentration) of (a)
Na+, (b) Cl-, (c) SO42-, (d) sucrose for PES5 membrane for 1×10-3 M salt solution with molar ratio of 0.2
(NaCl:Na2SO4) with sucrose addition for TMP of 200 kPa
245
1.700
Normalised mass fraction (Ȧi /ȦiB)
1.500
Cl-
Na+
1.300
SO42-
Sucrose
1.100
0.900
0.700
0.001 0.01 0.1
Dimensionless normal distance from membrane (y/2h)
Figure 6.26 – Predicted variation normal to membrane surface for solute concentration normalised relative to
inlet bulk concentration for PES5 membrane for 1×10-3 M salt solution with molar ratio of 0.2 (NaCl:Na2SO4)
with sucrose addition for TMP of 200 kPa
6.5 Conclusions
A method for coupling the CFD simulations of membrane channel hydrodynamics
(discussed in Chapter 4) and the mass-transfer model for prediction of solute
rejection (discussed in Chapter 5) has been successfully developed and verified.
The method is capable of providing a reasonably general and rigorous model of
the high-pressure membrane filtration processes (NF and RO) that is fully able to
predict the salient behaviours of such a system; i.e. both the channel
hydrodynamics, which significantly influence concentration polarisation and
fouling behaviour and hence permeate flux, and the mass transfer of solutes
through the membrane, which affects permeate quality. However, these processes
246
do not act in isolation, and interact in complex and subtle ways. The ability of the
coupled model to capture these interactions makes it a powerful predictive tool for
modelling of membrane filtration.
Simulations for several simple cases were carried out to verify the predictions of
the coupled model, which revealed some interesting (predicted) phenomena. The
model allows for both the permeate flux and the rejection of individual solutes to
vary spatially over the computational domain, according to the local
hydrodynamic conditions. Axial reduction in permeate flux through a membrane
channel has been widely described and observed, but axial reduction in solute
rejection less so. Model results indicated that axial rejection variation was so
small to be negligible in some cases, while quite significant in other cases at very
low crossflow velocities. While these velocities were generally considerably
smaller than encountered in practical applications, the utility of including spatially
variable solute rejection is apparent.
The use of the coupled rejection model in conditions where negative ionic
rejection was observed also allowed prediction of an inverse concentration
polarisation layer (i.e. where concentrations at the membrane surface are less than
that of the bulk flow). While this is a fairly obvious consequence of negative
rejection, this has not been widely described in CFD models of membrane
filtration.
247
The coupled model also accounts in part for the interplay in rejection and
permeate flux between uncharged and charged solutes. If a DSPM-style rejection
model is considered in isolation, the rejections of charged and uncharged
molecules are essentially independent. However, by coupling the hydrodynamic
model predictions of the concentration polarisation layer formation to the
rejection routine, the addition of uncharged solutes to a salt solution reduces both
the overall permeate flux and rejection of the charged solutes due to the increased
viscosity of the concentration polarisation layer, and the increased osmotic
pressure exerted within. This would also similarly affect other uncharged solutes.
In this manner, this gives a more physical representation of the complex
concentration polarisation layer mass transfer behaviour than has typically been
presented in CFD simulations of the membrane filtration process.
Of course, several simplifying assumptions have been made in the development of
the model, either for reasons of numerical stability or mathematical convenience.
These represent natural areas of further progression for the modelling approach
which would be valuable to investigate in further studies. The first of these
assumptions (discussed previously in Chapter 5) is the neglect of dielectric
exclusion effects in the modified DSPM used to predict solute rejection. As
highlighted earlier, inclusion of these effects would provide a more general
rejection model (at the expense of some further mathematical and modelling
difficulty), though the numerical method used in the coupled model is well-suited
to extension to include these effects.
248
The method to describe the diffusion of solutes throughout the bulk flow in the
membrane channel is also relatively simplified, using a modified Fick’s law
approach to ensure electroneutrality is conserved throughout the domain. The
reasoning for this approach is as follows: while rigorous diffusion models for
mixed solutions exist, they suffer from either a lack of experimental data to
determine diffusion coefficient matrices (e.g. an Onsager [126] or Maxwell-Stefan
[125] type approach), or numerical difficulties when implemented via a fictive
source term in commercial CFD codes (e.g. the extended Nernst-Planck approach
attempted in [103]). The combination of mean salt diffusion coefficients for
charged solutes and individual diffusion coefficients for uncharged solutes
ensures electroneutrality conditions are conserved, and incurs comparatively small
computational expense and numerical difficulty. The downfall of this approach is
it relies on empirical combination of experimentally determined relations
describing the transport properties of single-component solutions, and is as such
not completely general. However, model results indicated that this approach was
still able to simulate the formation of component-specific concentration
polarisation layers, which is thought to provide a more physical representation of
the concentration polarisation phenomena than simple film theory style
approximations. Further work in both predicting transport and rheological
properties of mixed solutions, and implementing a rigorous electro-diffusive
model for the bulk channel flow, would be important steps towards a more general
model.
249
Finally, some numerical instabilities were observed for an isolated case using the
coupled model that mesh and timestep refinement were unable to resolve. The
reason for this is unknown and is difficult to speculate upon, but appears to be
related to the spatially variable rejection routine predicting both positive and
negative flux for a single ion over a single boundary source. This condition is
unlikely to occur in realistic operating conditions, but further investigation and
refinement of the numerical routine if necessary would be worthwhile.
We may now turn our attention to using the coupled hydrodynamic-mass transfer
model to conduct some more detailed simulations of high-pressure membrane
filtration channels in Chapter 7.
250
Chapter 7
Simulations using the coupled hydrodynamic and rejection
model
7.1 Introduction
In Chapter 6, a fully predictive method coupling a hydrodynamic (CFD) model
and a rejection (pore) model was described, with verifications and some
illustrative simulations carried out for simple hydrodynamic cases. These
verifications were only possible for such simple cases due to the inherent
difficulties in observing and measuring both the flow structures for practical
membrane geometries, and the concentration fields within the membrane channel
with complex feed solutions. Verifications were thus performed for these simple
cases using larger-scale spatially averaged parameters (such as the rejection and
permeate flux), which indicated the coupled model performed well.
The coupled model is also able to predict the behaviour of the membrane system
in a level of detail that is (for most purposes) not presently feasible by
experimental observation. In this chapter the coupled model is used to conduct
some detailed numerical experiments for more realistic membrane channel
geometries and hydrodynamic conditions. The purpose of these simulations is to
identify some of the salient features of the hydrodynamics and mass transfer
251
predicted by the coupled model for a typical membrane channel. In the absence of
detailed experimental data at these fine scales for such systems, such simulations
may provide a useful tool for predicting the behaviour and performance of
membrane systems.
7.2 Simulation approach
All simulations conducted in this chapter are for a 2D model of a spacer-filled
membrane channel. Previously in Section 4.3, a hydrodynamic model of such a
channel was developed using the geometry and flow conditions previously
modelled by Wardeh and Morvan [115]. These conditions provided for several
complex hydrodynamic flow features to be observed, and are thus an interesting
hydrodynamic case to investigate further. In the following simulations, we use the
previous hydrodynamic model geometry and flow conditions, but extend the
model by inclusion of the coupled mass transfer model to predict the rejection of
solutes by the membrane.
In the previous simulations of the spacer-filled channel, a simple sodium chloride
solution was considered with a high fixed rejection (R = 0.95) specified at the
membrane boundary. It is noted that this value was used as an arbitrary value in
numerical simulations of a (fictive) membrane only. To provide an insight into the
behaviour of the coupled model, we now consider a case where the mass-transfer
behaviour of the system is more variable; i.e. where the solute rejection varies
over a wider range of values. To do so, we adopt membrane parameters similar to
252
that used in Chapter 5 for the PES5 membrane used in experiments by Bowen and
Mukhtar [17].
Finally, we perform the simulations for several different feed solutions, including
single salt solutions, multiple salt solutions, and multicomponent mixtures
containing both salts and organic (uncharged) solutes. In this way, we aim to
provide a general investigation of the utility of the coupled model, identify
relevant hydrodynamic and mass transfer phenomena, and evaluate areas for
future investigation beyond this thesis. A brief revision of the simulation
parameters is contained in the following.
7.2.1 Geometry
As per Section 4.3.3, the model geometry used by Wardeh and Morvan [115] is
employed, consisting of a 2D channel with long entrance and exit lengths. The
channel is filled with seven cylindrical spacers oriented normal to the main
channel flow, which are arranged in a zig-zag fashion (alternating between the top
and bottom of the channel) to emulate a woven spacer net. Both upper and lower
walls of the channel are permeable membranes, with the exceptions of the
surfaces of the spacer filaments and the entrance and exit sections. The geometry
is used to represent the periodic flow that occurs in spiral-wound membrane
(SWM) modules; the use of seven filaments is sufficient to develop the periodic
flow structure in the vicinity of the fifth and six filaments, as previously noted. A
253
recap of the model geometry is shown in Figure 7.1, with the dimensions h = 2
mm, l/h = 4, and df / h = 0.5.
spacer filaments
channel height h
diameter df permeate flow
feed retentate
entrance length permeate flow exit length

12 df 25 df
inter-filament distance l
Figure 7.1 – Model geometry for simulations of flow through 2D space-filled channel
The same boundary and initial conditions as described in Section 4.3.4 are
enforced here, with the average velocity (in terms of the channel Reynolds
number Rech) and solute mass fraction(s) specified at the channel inlet, and an
average static zero pressure specified at the outlet. No-slip walls with no mass
transfer are used over the entrance and exit lengths, and on the filament surfaces.
The permeate flow through the membrane surfaces is again specified as a mass
sink, though the solute fluxes are now determined by the coupled mass transfer
model described in Chapter 6, which calculates the rejection of each individual
solute independently for each computational point on the membrane surface. The
properties of the membrane are derived from the parameters used in Chapter 5 for
the PES5 membrane used in experiments by Bowen and Mukhtar [17].
254
Finally, only the two lower channel Reynolds numbers previously considered
(Rech = 100 and 200) are simulated here, for reasons of computational expense.
This is because higher channel Reynolds numbers (Rech ZHUH VKRZQ LQ
Section 4.3.7.2 to cause unsteady flow structures, greatly increasing
computational expense, even with an uncoupled (hydrodynamics only) model.
Due to the much greater computational expense again of the coupled
(hydrodynamics and mass transfer) modelling approach, a decision was made to
investigate steady flows only for the coupled model, though future investigation
of unsteady flows would indeed be worthwhile. Initial conditions are again
automatically generated by the CFX solver, with the initial solute concentration
field set to be globally equal to that of the bulk solute mass fraction. The model
parameters common to all subsequent simulations are thus shown in Table 7.1.
Table 7.1 – Model input parameters for coupled simulations of 2D spacer-filled channel
Parameter Value
Transmembrane pressure 0.5 MPa
Membrane permeability LV 2 x 10-10 m Pa-1 s-1
ǻy/Ak 1.60 ×10-5 m
CX 15 eq m-3 (fixed ionic strength for all simulations)
rp 1.14 nm
ZX -1 (negatively charged membrane)
7.2.3 Computational mesh and convergence criteria
An unstructured 2D mesh was used for all simulations, with significant refinement
of the mesh in the vicinity of the channel walls and the surfaces of the spacer
255
filaments. Previous mesh refinement studies for the same geometry without the
coupled mass transfer model indicated a spatial resolution of approximately 0.05
mm in the bulk of the channel was required, decreasing down to approximately 3
μm against the membrane walls and filaments. However, with the inclusion of the
coupled mass transfer model, a slightly more refined mesh was required to obtain
accurate solutions. The adopted mesh consists of 0.02 mm elements in the channel
bulk, with refinement down to 2 μm against the membrane walls and filaments.
The refined mesh consists of approximately 358 000 elements (cf. 295 000
elements for the mesh previously adopted in Section 4.3.6).
Convergence was judged to occur once the mass, momentum, and solute mass
fraction residuals were all smaller than 1 x 10-5. This convergence target is relaxed
compared to previous simulations as it proved difficult to obtain convergence to
the very tight previous tolerance of 1 x 10-6 with the coupled model.
7.3 Coupled simulation of single salt solution in spacer filled channel
The first set of simulations was conducted for a single salt solution of sodium
chloride using the parameters outlined in Table 7.1 and Figure 7.1. An inlet salt
mass fraction of 0.002 kg kg-1 was used for these simulations, representing an
ionic strength of approximately 34.1 eq m-3 (i.e. an inlet sodium chloride
concentration of 34.1 mol m-3). Simulations were performed for channel Reynolds
numbers of 100 and 200; at these Reynolds numbers, the flow observed was
steady, and there was no variation of the solution with time.
256
A selection of representative results from the simulations are shown here, again
all in the vicinity of the fifth and sixth spacer filaments (chosen as a representative
monitoring location). As before, the predicted salt mass fraction (Figure 7.2), fluid
streamlines (Figure 7.3), velocity vectors (Figure 7.4), and vorticity fields (Figure
7.5) are shown. The salt mass fraction shown in Figure 7.2 has been normalised
relative to the inlet (bulk) salt mass fraction for comparison with later cases.
Rech = 100
Rech = 200
Figure 7.2 – Predicted salt mass fraction distribution for spacer-filled channel using coupled-mass transfer
model for single salt solution (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3)
257
Rech = 100
Rech = 200
Figure 7.3 – Predicted fluid streamlines for spacer-filled channel using coupled-mass transfer model for
single salt solution (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3)
Rech = 100
Rech = 200
Figure 7.4 – Predicted velocity vectors for spacer-filled channel using coupled-mass transfer model for single
salt solution (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3)
258
Rech = 100
Rech = 200
Figure 7.5 – Predicted vorticity fields for spacer-filled channel using coupled-mass transfer model for single
salt solution (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3)
The general features of the flow are similar to that previously observed in Section
4.3.7.1. Large recirculation zones are formed on the downstream face of the
filaments, and smaller zones on the upstream faces, in which correspondingly
high salt concentrations are observed due to concentration polarisation. These
zones grow larger with increasing channel Reynolds number. However, as the salt
rejection for the current case (R varying between approximately 0.017 – 0.018) is
significantly less than the case in Section 4.3.7.1 (R = 0.95), the magnitude of the
variation in salt concentration throughout the channel is greatly reduced. The
concentration polarisation effect is still observable, however.
259
Of more interest is the variation in mass transfer parameters within the monitoring
area due to the inclusion of spatially variable rejection and permeate flux. The
predicted variation in the salt rejection, permeate flux, and wall shear are shown in
Figure 7.6 and Figure 7.7, for Rech = 100 and Rech = 200, respectively. The
predicted salt mass fraction and velocity streamlines are shown for spatial
reference. Values are shown both at the upper and lower (permeable) surfaces of
the membrane channel.
260
Salt mass fraction
Velocity streamlines
Rejection R
0.01784
Lower wall
0.01776 Upper wall
0.01768
0.01760
Flux JV (× 10-5 m³ m² s-1)

6.65
Lower wall
6.55 Upper wall
6.45
6.35
Wall shear (Pa)

0.40
Lower wall
0.30 Upper wall
0.20
0.10
0.00
Figure 7.6 - Predicted variation in mass transfer parameters within spacer-filled channel with spatially
varying permeate flux and solute rejection (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3, Rech = 100)
261
Salt mass fraction
Rejection R
0.01750
Lower wall
0.01730 Upper wall
0.01710
0.01690
Flux JV (× 10-5 m³ m² s-1)

6.30
Lower wall
6.20 Upper wall
6.10
6.00
5.90
Wall shear (Pa)

1.00
Lower wall
0.80
Upper wall
0.60
0.40
0.20
0.00
varying permeate flux and solute rejection (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3, Rech = 200)
262
Considering first the predicted wall shear, the maximum shear is observed as
expected opposite to the spacer filaments. That is, the maximum shear for the
lower wall is observed at the same axial location as the filament on the opposing
(upper) wall, and vice versa. This result corresponds with previous observations
for similar channels [e.g. 114]. A local increase in wall shear is also observed
adjacent to the recirculation zones on the downstream side of the filaments, due to
the recirculation motion previously described (clockwise rotation adjacent to the
lower wall, and counter-clockwise adjacent to the upper wall). A similar local
increase in shear might be expected adjacent to the upstream filament edge due to
the smaller recirculation zone observed here; however, for the relatively low
channel Reynolds numbers examined here, the increase in shear in this region is
not significant.
The axial variation in permeate flux, however, follows a different trend from that
of the wall shear. Flux decreases on the downstream side of an adjacent filament,
as the recirculation zone separates from the membrane wall However, flux then
increases towards the upstream side of an adjacent filament, as the recirculation
zone reattaches to the membrane wall. The region of highest flux (i.e. the highest
mass transfer) is located slightly downstream of the reattachment point, which is
consistent with the results of Fimbres-Weihs et al. [114]. The flux variation
implies, as noted in [114], that separation points and reattachment points are
associated with regions of low and high mass transfer, respectively.
263
Perhaps somewhat surprisingly, the spatial variation in salt rejection follows an
almost identical trend to that of the permeate flux. Examining the rejection
variation in conjunction with the predicted salt mass fraction distribution, this can
be succinctly explained by the reasoning presented in [114]; at reattachment
points, low-concentration fluid flows towards the (high-concentration) wall.
Simultaneously, flow within the boundary layer adjacent to the wall moves away
from the reattachment point, creating a low salt concentration and thereby
increasing the predicted rejection. Similarly, a high salt concentration is formed at
separation points, decreasing the predicted rejection. However, the variation in
rejection is relatively small in comparison to the variation in permeate flux.
The formation of the concentration polarisation boundary layer can also be
visualised more clearly by examining the variation in salt mass fraction
(normalised relative to the inlet concentration) in the direction normal to the
membrane wall. This has been performed for a point midway between filaments
in the monitoring zone in terms of non-dimensionalised distance from the
membrane surface for both the upper and lower membrane surfaces. The variation
is shown in Figure 7.8 for the lower membrane surface, and Figure 7.9 for the
upper membrane surface. Clearly in the single salt case this transverse variation in
concentration will be identical for both ions; however it is useful for examining
the formation of varying concentration polarisation layers for different ions in a
multicomponent solution, as will be discussed in Section 7.4.
264
1.014
Rech = 100
1.012 Rech = 200
1.010
1.008
1.006
1.004
1.002
1.000
0.001 0.01 0.1
Dimensionless normal distance from membrane (|y|/2h)
Figure 7.8 - Predicted transverse variation in normalised solute concentration as a function of non-
dimensionalised distance from lower membrane wall within spacer-filled channel with spatially varying
permeate flux and solute rejection (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3)
1.014
Rech = 100
1.012 Rech = 200
1.010
1.008
1.006
1.004
1.002
1.000
0.001 0.01 0.1
dimensionalised distance from upper membrane wall within spacer-filled channel with spatially varying
permeate flux and solute rejection (NaCl, 0.002 kg kg-1, equivalent to 34.1 mol m-3)
265
As expected, Figure 7.8 and Figure 7.9 show the magnitude of the transverse
variation in concentration to be greatest at the lower channel Reynolds number.
That is, the effect of concentration polarisation is greatest for the lower Reynolds
number. Additionally, the concentration polarisation effect is more pronounced
for the upper membrane wall (i.e. on the wall opposite to the immediately
downstream filament). This reflects the disparity in permeate flux between the
lower and upper walls at the filament midpoint shown in Figure 7.6 and Figure
7.7.
Finally, to examine the effect of allowing salt rejection to vary spatially using the
coupled model, an additional set of simulations was conducted using the same
parameters, but with a manually specified constant rejection (i.e. a non-coupled
model). In these cases, the rejection was specified to be equal to the rejection
determined by the coupled model simulations, averaged over the monitoring area
in the vicinity of the fifth and sixth filaments. That is, for the Rech = 100 case, the
rejection was set to a constant value of 0.0178, and for the Rech = 200 case, a
value of 0.0172.
Examination of the results of the constant rejection (non-coupled) simulations
revealed the results to be practically identical to that of the coupled model, and
therefore the results are not reproduced here. This is due to the relatively low
spatial variation in rejection for these conditions.
266
7.4 Coupled simulation of multiple salt solution in spacer filled
channel
A second set of simulations was then conducted for a multiple salt solution
composed of sodium chloride and sodium sulphate with an arbitrarily chosen
molar ratio of 0.2:0.8 (NaCl:Na2SO4). The inlet mass fraction was adjusted to
maintain a consistent ionic strength with the previous simulation for a single salt
solution in Section 7.3 (approximately 34.1 eq m-3); this yields inlet
concentrations of approximately 3.79 mol m-3 (NaCl) and 15.16 mol m-3
(Na2SO4). (Maintaining a consistent ionic strength between simulations allows for
the use of a consistent membrane volumetric charge density CX for easier
comparison between cases).
Simulations were again run for two channel Reynolds numbers of 100 and 200,
and again steady flow was observed in all cases. Of particular interest is the fact
that negative rejection of the chloride co-ion is predicted for these conditions,
similar to the conditions presented in Section 5.5.2.2. Ionic mass fraction
distributions are shown, normalised relative to the inlet mass fraction for each ion,
in Figure 7.10 and Figure 7.11 for Rech = 100 and Rech = 200 respectively. These
clearly indicate the negative rejection of the chloride co-ion and the formation of
the ‘inverse’ concentration polarisation layer – the colour spectrum between pink
and blue on these figures represents areas where the local concentration is less
than the inlet concentration (i.e. where the normalised mass fraction is less than
267
one). As before, for the positively rejected ions (Na+ and SO42-), the classical
concentration polarisation layer is formed near the membrane surface, and areas
of high concentration are formed adjacent to the spacer filaments due to the
recirculation zones in these regions. Again, the high concentration zones are more
prominent on the downstream face of the filaments due to the relatively larger size
of the corresponding recirculation zone. Corresponding low concentration regions
for the negatively rejected Cl- ion occur in these recirculation zones.
Na+
Cl-
SO42-
Figure 7.10 – Predicted ionic mass fraction distribution for spacer-filled channel using coupled-mass transfer
model for multiple salt solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3),
Rech = 100
268
Na+
Cl-
SO42-
model for multiple salt solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3),
Rech = 200
Predicted streamlines, velocity vectors, and vorticity plots are again shown in
Figure 7.12, Figure 7.13, and Figure 7.14, respectively. These show the same
general hydrodynamic trends as observed for the single salt case discussed in
Section 7.3.
269
Rech = 100
Rech = 200
multiple salt solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3)
Rech = 100
Rech = 200
Figure 7.13 – Predicted velocity vectors for spacer-filled channel using coupled-mass transfer model for
270
Rech = 100
Rech = 200
Figure 7.14 – Predicted vorticity fields for spacer-filled channel using coupled-mass transfer model for
We can again examine the spatial variation in mass transport in terms of solute
rejection, permeate flux, and wall shear, as well. These are shown in Figure 7.15
and Figure 7.16 for Rech = 100 and Rech = 200, respectively, with fluid
streamlines shown for spatial reference. Examining first the solute rejections, a
significant variation is predicted depending on axial position between the
filaments. For the positively rejected ions (Na+ and SO42-), these follow a similar
trend to that observed for the rejection for the single salt solution discussed in
Section 7.3. Rejection decreases on the downstream side of an adjacent filament,
increases on the upstream side of an adjacent filament, and reaches a minimum
slightly upstream of an adjacent filament. Somewhat differently to the single salt
results, however, the rejection for these ions is at a maximum adjacent to an
271
opposing filament, rather than the previously observed location slightly
downstream of the reattachment point for the downstream filament recirculation
zone. This is more apparent at the higher channel Reynolds number as shown in
Figure 7.16. A local maximum is however still observed at the reattachment point
for the downstream filament recirculation zone for the lower channel Reynolds
number, as shown in Figure 7.15.
For the negatively rejected co-ion (Cl-), the opposite of this trend is observed, as
might be expected. Maximum rejection occurs at the upstream filament
recirculation zone, decreasing rapidly towards the filament face, then increases
rapidly at the downstream face of the filament. The rejection then decreases
gradually, reaching a minimum adjacent to the opposing filament, again more
prominently at the higher channel Reynolds number. A local minimum
corresponding to the reattachment point of the downstream filament recirculation
zone is also observable for the Rech = 100 case.
Inspection of the predicted wall shear variation reveals very similar results to that
for the single salt solution, both in terms of general variation and shear magnitude.
This is perhaps expected as the bulk hydrodynamic conditions are in general very
similar.
More interestingly, and perhaps initially counter-intuituively, is the variation in
permeate flux. While this follows the same general trend as described for the
single salt solution in Section 7.3 (similar to that of the rejection trend for
272
positively rejected ions), the overall magnitude of the permeate flux is greater for
the multiple salt solution, as shown in Table 7.2. This represents a flux increase of
approximately 6 – 7 %, which is not insignificant.
Table 7.2 – Comparison of average permeate flux values for single salt (NaCl, overall ionic strength 34.1 eq
m-3) and multiple salt (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3) cases
Solution Averaged permeate flux JV × 10-5 m3 m-2 s-1
Rech = 100 Rech = 200
Single salt (NaCl) 6.51 6.16
Multiple salt (NaCl:Na2SO4) 6.92 6.58
This occurs despite identical operating parameters (inlet velocity, applied TMP)
and ionic strength being imposed. This result seems even more unintuitive
considering the considerable increase in the magnitude of rejections (in the order
of 0.2 – 0.25) for the multiple salt case, compared to the single salt case (in the
order of 0.02) – it might be expected that the higher rejection would lead to a
more highly developed concentration polarisation boundary layer, retarding the
permeate flux due to viscous effects and an increase in osmotic pressure.
However, the flux increase can be explained by several factors: first, the constant
ionic strength (expressed in eq m-3) used between simulations and the inclusion of
the multivalent sulphate ion means that the inlet salt concentration for the multiple
salt case (expressed in mol m-3) is less than the single salt case. This implies that
the osmotic pressure developed in the multiple salt case would be reduced, and
273
hence have a lesser retarding effect on the permeate flux. In addition, the overall
solution viscosity would be reduced to have a similar effect. More significant
perhaps, is the negative rejection of the chloride ion, which causes the formation
of the ‘inverse’ concentration polarisation layer, where the concentration of
chloride decreases in increasing proximity to the membrane surface. This reduces
solution viscosity and osmotic pressure near the membrane surface, which may
also lead to the increase in overall permeate flux. In any case, the permeate flux
increase illustrates the fact that subtle changes in the mass transfer behaviour of
the near-wall region may have significant effects on the flux behaviour of the
membrane which are not immediately or intuitively obvious without use of a
rigorous simulation tool.
274
Na+ rejection
0.210
Lower wall
0.205 Upper wall
0.200
0.195
0.190
Cl- rejection
-0.180
Lower wall
-0.190 Upper wall
-0.200
-0.210
-0.220
SO42- rejection
0.250
Lower wall
0.240 Upper wall
0.230
0.220
0.210
Flux JV (× 10-5 m³ m² s-1)

7.10
Lower wall
7.00 Upper wall
6.90
6.80
6.70
Wall shear (Pa)

0.40
Lower wall
0.30 Upper wall
0.20
0.10
0.00
varying permeate flux and solute rejection (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1
eq m-3, Rech = 100)
275
Na+ rejection
0.210
Lower wall
0.200 Upper wall
0.190
0.180
Cl- rejection
-0.170
Lower wall
-0.180
Upper wall
-0.190
-0.200
-0.210
-0.220
SO42- rejection
0.250
Lower wall
0.240 Upper wall
0.230
0.220
0.210
Flux JV (× 10-5 m³ m² s-1)

7.00
Lower wall
6.80
Upper wall
6.60
6.40
6.20
6.00
Wall shear (Pa)

1.00
Lower wall
Upper wall
0.50
0.00
varying permeate flux and solute rejection (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1
eq m-3, Rech = 200)
276
Examining the transverse variation in solute concentration can also give a clearer
picture of the concentration polarisation predictions of the model. Figure 7.17
through Figure 7.20 illustrate the variation in the normalised ionic concentrations
with the non-dimensionalised transverse distance. As previously discussed, this
analysis has been performed for both the upper and lower channel walls at a point
midway between the fifth and sixth spacer filaments. As discussed in Section
6.3.4.4, these results can be loosely interpreted as describing the overlapping
concentration polarisation layers of different thickness for each individual ion, or
more precisely, the varying concentration gradients for each individual ion within
the concentration polarisation layer. In all cases, the greatest variation in
concentration is observed for the SO42- ion. The inverse concentration polarisation
layer for the negatively rejected Cl- co-ion is also apparent. The magnitude of
variation in the non-dimensionalised concentration is also greater for the lower
channel Reynolds number, i.e. the concentration polarisation effect is stronger at
lower Reynolds number, as expected.
In addition, the variation is more pronounced (the concentration polarisation
effect is stronger) for the upper membrane wall. That is, the concentration
polarisation effect midway between filaments is stronger on the opposing wall
from the downstream filament than the adjacent wall. This is commensurate with
the permeate flux predictions shown in Figure 7.15 and Figure 7.16, and the
previous results for the single salt solution.
277
1.200
1.150 Cl-
Na+
1.100
SO42-
1.050
1.000
0.950
0.900
0.850
0.001 0.01 0.1
permeate flux and solute rejection (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3,
Rech = 100)
1.200
1.150 Cl-
Na+
1.100
SO42-
1.050
1.000
0.950
0.900
0.850
0.001 0.01 0.1
Rech = 200)
278
1.250
1.200 Cl-
Na+
1.150
SO42-
1.100
1.050
1.000
0.950
0.900
0.850
0.001 0.01 0.1
Rech = 100)
1.250
1.200 Cl-
Na+
1.150
SO42-
1.100
1.050
1.000
0.950
0.900
0.850
0.001 0.01 0.1
Rech = 200)
279
Again, to examine the effect of the spatial rejection variation inherent in the
coupled model approach, a set of simulations was conducted using fixed constant
rejections derived from the averaged predicted rejections from the coupled model.
A summary of the constant rejection values used in these simulations is shown in
Table 7.3.
Table 7.3 – Fixed rejection values used for uncoupled model comparison case (NaCl:Na2SO4 with molar ratio
0.2:0.8, overall ionic strength 34.1 eq m-3)
Component Rejection
Rech = 100 Rech = 200
Na+ 0.200 0.200
Cl- -0.205 -0.201
SO42- 0.233 0.232
Previous results for the single salt solution in Section 7.3 had indicated negligible
difference between the coupled simulation results with variable rejection, and the
uncoupled results with constant rejection. This was ascribed to the low variation
in rejection, and low overall rejection, predicted for the conditions in Section 7.3.
In the multiple salt case considered here, though, an observable difference in some
results is apparent. This is due to the higher rejection magnitudes predicted for the
present case, and hence the greater effect exerted by the concentration polarisation
effect. A comparison of the difference in normalised mass fraction for each ion
between the constant rejection case and variable rejection case is shown in Figure
280
7.21 and Figure 7.22 for Rech = 100 and Rech = 200, respectively. The variation is
relatively small (in the order of one percent of the normalised mass fraction).
Na+
Cl-
SO42-
Figure 7.21 – Difference in predicted ionic mass fraction distribution for spacer-filled channel between
constant rejection model and variable rejection model for multiple salt solution (NaCl:Na2SO4 with molar
ratio 0.2:0.8, overall ionic strength 34.1 eq m-3), Rech = 100
Some brief comments can be made about the differences in predicted ionic
concentration fields shown in Figure 7.21 and Figure 7.22. The main region where
differences are observed is in the vicinity of the filaments, which is unsurprising
considering the relatively rapid variation in rejection predicted by the coupled
model for these regions. Mass fractions predicted by the variable rejection model
281
are generally lower in the vicinity of these recirculation zones for the positively
rejected ions, and higher for the negatively rejected Cl- ion.
A small difference is also apparent along the remainder of the membrane surface
outside the filament recirculation zones: the mass fractions predicted by the
variable rejection model for the positively rejected ions are generally greater in
the area opposing the filaments, and lesser in the remaining area adjacent to the
filaments. The opposite of this trend is apparent for the negatively rejected Cl- ion.
This trend is more apparent for the higher channel Reynolds number, and for the
Cl- and SO42- ions. The overall difference is also greater for the higher channel
Reynolds number, which corresponds to the increasing hydrodynamic influence of
the filament recirculation zones.
282
Na+
Cl-
SO42-
Figure 7.22 – Difference in predicted ionic mass fraction distribution for spacer-filled channel between
constant rejection model and variable rejection model for multiple salt solution (NaCl:Na2SO4 with molar
ratio 0.2:0.8, overall ionic strength 34.1 eq m-3), Rech = 200
However, this variation in the ionic concentration field is not significant enough
to cause significant variation in the computed velocity and pressure fields between
the constant rejection and variable rejection solutions. It does, however, impact to
a small degree on the distribution of the permeate flux at the membrane surface, as
shown in Figure 7.23 and Figure 7.24 for Rech = 100 and Rech = 200, respectively.
However, the overall effect on the permeate flux over the monitoring area is so
small as to be negligible (as would be expected when using constant rejections
283
derived from the area-averaged predictions of the coupled model). The variation
in wall shear between the cases is also negligible, and not shown.
Flux JV (× 10-5 m³ m² s-1)
7.15
Lower wall
7.10 Upper wall
Lower wall constant rejection

7.05 Upper wall constant rejection
7.00
6.95
6.90
6.85
6.80
6.75
6.70
Figure 7.23 – Difference in predicted permeate flux within spacer-filled channel w between constant rejection
model and variable rejection model (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3,
Rech = 100)
284
Flux JV (× 10-5 m³ m² s-1)
6.90
Lower wall
Upper wall
6.80 Lower wall constant rejection
Upper wall constant rejection
6.70
6.60
6.50
6.40
6.30
6.20
6.10
Rech = 200)
285
7.5 Coupled simulation of complex mixed solution in spacer filled
channel
We now perform a final set of simulations using the multiple salt solution
described in Section 7.4, but with the addition of an uncharged solute (sucrose) to
demonstrate the applicability of the model to describe mixtures of charged and
uncharged solutes. The salt inlet mass fractions are maintained as per previous
simulations to maintain a consistent ionic strength, but the additional sucrose
component is added at an inlet concentration of 0.001 kg kg-1.
The results indicate that the addition of the uncharged component (sucrose) has
several effects on the behaviour of the membrane channel, primarily in terms of
mass transfer through the membrane. The high sucrose concentration in the
concentration polarisation layer (seen in the predicted normalised mass fraction
distributions in Figure 7.25 and Figure 7.26 for Rech = 100 and 200, respectively)
has a retarding effect on both the permeate flux and rejection of charged solutes.
This is due to the high osmotic pressure exerted by the sucrose molecules, as well
as the correspondingly higher solution viscosity due to the presence of sucrose
within the concentration polarisation layer. (This trend was also observed
previously in simulations of the simple stirred cell arrangement in Section 6.4).
The average reduction in flux (values averaged over the monitoring area are
shown in Table 7.4) due to the addition of sucrose is approximately 9 %.
286
Na+
Cl-
SO42-
Sucrose
model for complex mixture solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-
3
, and additional sucrose at 0.001 kg kg-1), Rech = 100
The presence of sucrose also slightly decreases the predicted magnitude of the
rejections of the ionic solutes (averaged solute rejections for the current case are
shown in Table 7.5). This is true for both the positively rejected ions and the
negatively rejected Cl- co-ion. However, this effect is smaller (in the order of one
to two percent), as previously observed in Section 6.4.
287
Na+
Cl-
SO42-
Sucrose
model for complex mixture solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-
3
, and additional sucrose at 0.001 kg kg-1), Rech = 200
The uncharged solute addition however does not significantly alter the bulk
hydrodynamics of the channel, and the general flow features are very similar to
that observed previously. Representative plots of the predicted fluid streamlines
(Figure 7.27), velocity vectors (Figure 7.28), and vorticity fields (Figure 7.29) are
shown in the following.
288
Rech = 100
Rech = 200
complex mixture solution (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and
additional sucrose at 0.001 kg kg-1)
Rech = 100
Rech = 200
Figure 7.28 – Predicted velocity vectors for spacer-filled channel using coupled-mass transfer model for
289
Rech = 100
Rech = 200
Figure 7.29 – Predicted vorticity fields for spacer-filled channel using coupled-mass transfer model for single
Table 7.4 – Comparison of average permeate flux values for single salt (NaCl, overall ionic strength 34.1 eq
m-3), multiple salt (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3), and complex
mixture (NaCl:Na2SO4 with molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and additional sucrose at
0.001 kg kg-1) cases
Solution Averaged permeate flux JV × 10-5 m3 m-2 s-1
Rech = 100 Rech = 200
Single salt (NaCl) 6.51 6.16
Multiple salt (NaCl:Na2SO4) 6.92 6.58
Complex mixture (NaCl:Na2SO4:sucrose) 6.31 6.00
The spatial variation in solute rejection, permeate flux, and wall shear are shown
in Figure 7.30 and Figure 7.31 for Rech = 100 and 200, respectively. Similar
trends are again observed for permeate flux and the rejection of positively rejected
290
solutes, with maximum values observed opposite the spacer filaments. A local
maximum is also again apparent at the reattachment point of the downstream
filament recirculation zone at Rech = 100, but disappears with the higher crossflow
velocity at Rech = 200. Minimum flux and rejection occur at the separation points
of the recirculation zones upstream and downstream of the filaments.
Conversely, for the negatively rejected Cl- co-ion, minimum values (i.e. the
negative rejection of the greatest magnitude) are observed opposite to the spacer
filaments. Maximum values (the smallest magnitude of the negative rejection) are
observed at the recirculation zone separation points upstream and downstream of
the filaments. A local minimum is also apparent at the lower channel Reynolds
number only for the reattachment point of the downstream filament recirculation
zone. The wall shear results, being largely a function of the bulk channel
hydrodynamics, remain essentially unchanged from previous model results.
Again considering a location midway between the fifth and sixth filaments, the
transverse variation in the normalised solute concentrations is shown in Figure
7.32 through Figure 7.35. The variation for the ionic solutes is similar to that
observed previously in Section 7.4, however, this variation is dwarfed in
magnitude by that of the additional sucrose. Wall concentrations for the sucrose
component are in the order of two to three times the bulk concentration, due to the
high rejection of sucrose (R §The variation is again more pronounced for
the lower channel Reynolds number due to the lower wall shear.
291
Na+ rejection
0.205
Lower wall
0.200
Upper wall
0.195
0.190
0.185
0.180
Cl- rejection
-0.170
Lower wall
-0.180 Upper wall
-0.190
-0.200
-0.210
SO42- rejection
0.250
Lower wall
0.240 Upper wall
0.230
0.220
0.210
Sucrose rejection
0.770
Lower wall
0.765 Upper wall
0.760
0.755
Flux JV (× 10-5 m³ m² s-1)
7.00
Lower wall
6.50 Upper wall
6.00
5.50
Wall shear (Pa)
0.40
Lower wall
0.30 Upper wall
0.20
0.10
0.00
Figure 7.30 – Predicted variation in mass transfer parameters within spacer-filled channel with spatially
varying permeate flux and solute rejection (NaCl:Na2SO4:sucrose solution, Rech = 100)
292
Na+ rejection
0.210
Lower wall
0.200 Upper wall
0.190
0.180
Cl- rejection
Lower wall
-0.180
Upper wall
-0.200
-0.220
SO42- rejection
0.250
Lower wall
0.240
Upper wall
0.230
0.220
0.210
0.200
Sucrose rejection
0.770
Lower wall
0.765
Upper wall
0.760
0.755
0.750
0.745
Flux JV (× 10-5 m³ m² s-1)
6.70 Lower wall
Upper wall
6.20
5.70
5.20
Wall shear (Pa)
1.00
Lower wall
Upper wall
0.50
0.00
varying permeate flux and solute rejection (NaCl:Na2SO4:sucrose solution, Rech = 200)
293
2.250
2.050 Cl-
Na+
1.850 SO42-
Sucrose
1.650
1.450
1.250
1.050
0.850
0.001 0.01 0.1
and additional sucrose at 0.001 kg kg-1, Rech = 100)
2.050
Cl-
1.850
Na+
SO42-
1.650 Sucrose
1.450
1.250
1.050
0.850
0.001 0.01 0.1
294
2.850
2.650
Cl-
2.450 Na+
SO42-
2.250
Sucrose
2.050
1.850
1.650
1.450
1.250
1.050
0.850
0.001 0.01 0.1
3.350
Cl-
2.850 Na+
SO42-
Sucrose
2.350
1.850
1.350
0.850
0.001 0.01 0.1
295
Finally, simulations were again made using fixed constant rejections to elucidate
the effect of the spatially varying rejection capability of the coupled approach. A
summary of the adopted constant rejection values derived from the averaged
rejections predicted by the coupled model is shown in Table 7.5.
Table 7.5 – Fixed rejection values used for uncoupled model comparison case (NaCl:Na2SO4 with molar ratio
Component Rejection
Rech = 100 Rech = 200
Na+ 0.197 0.196
Cl- -0.197 -0.198
SO42- 0.229 0.228
Sucrose 0.763 0.759
As per the comparison in Section 7.4, the results do differ slightly between the
constant rejection case and the variable rejection (coupled) case. The difference in
normalised solute mass fractions is shown in Figure 7.36 and Figure 7.37 for Rech
= 100 and Rech = 200, respectively. The general trend is similar to that previously
observed, with predicted mass fractions for positively rejected solutes in the
variable rejection case being higher in the vicinity of the membrane opposite to
spacer filaments, and lower elsewhere in the vicinity of the membrane. In
particular, the comparative decrease in mass fraction is again most pronounced in
the recirculation zones immediately before and after the filaments. This effect is
more significant for the higher channel Reynolds number, and especially so for
the strongly rejected sucrose component (R §7KLVLVGXHWRWKHDGGLWLRQDO
296
osmotic and viscous effects of the high sucrose concentration. In addition, a small
area of increased concentration is now observed directly after the fifth filament,
which is thought to represent a locally high concentration zone corresponding to
the separation point of the downstream filament recirculation zone. The converse
of these trends is observed for the negatively rejected Cl- co-ion, as previously
reported.
Na+
Cl-
SO42-
Sucrose
Figure 7.36 – Difference in predicted solute mass fraction distribution for spacer-filled channel between
constant rejection model and variable rejection model for complex mixture solution (NaCl:Na2SO4 with
molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and additional sucrose at 0.001 kg kg-1), Rech = 100
297
Na+
Cl-
SO42-
Sucrose
Figure 7.37 – Difference in predicted solute mass fraction distribution for spacer-filled channel between
constant rejection model and variable rejection model for complex mixture solution (NaCl:Na2SO4 with
molar ratio 0.2:0.8, overall ionic strength 34.1 eq m-3, and additional sucrose at 0.001 kg kg-1), Rech = 200
Once more, though, the bulk hydrodynamics, nor the overall permeate flux are not
significantly affected by this mass-transfer variation. However, a small variation
between the constant and variable rejection cases is again apparent, most notably
in the vicinity of the spacer filaments, as shown in Figure 7.38 and Figure 7.39.
298
Flux JV (× 10-5 m³ m² s-1)

6.90
Lower wall
Upper wall
6.50
6.30
6.10
5.90
5.70
Flux JV (× 10-5 m³ m² s-1)

6.80
Lower wall
Upper wall
6.40
6.20
6.00
5.80
5.60
5.40
5.20
5.00
299
7.6 Conclusions
Using the coupled hydrodynamic – mass transfer model, a series of numerical
experiments have been carried out for a 2D spacer-filled membrane channel. The
more complex geometry of such a channel leads to more complex hydrodynamic
and mass transfer effects, and allows for a more realistic application of the
coupled model than the simplified verification cases presented in Chapter 6.
Simulation results showed that the coupled model was successfully able to
reproduce the general hydrodynamic behaviour of the membrane channel
predicted by an equivalent (non-coupled) constant rejection model. This included
effects such as recirculation zones located upstream and downstream of the spacer
filaments, and formation of the concentration polarisation boundary layer
(including an inverse concentration polarisation layer in the presence of
negatively rejected solutes).
However, the coupled model approach affords some additional scope for
predicting the mass-transfer behaviour in the vicinity of the membrane walls.
Examination of the permeate flux and solute rejection, both allowed to vary as a
function of the local hydrodynamic conditions in the couple model, showed some
predictable patterns. In general, the magnitude of both the permeate flux and
solute rejection:
300
x decrease on the downstream face of an adjacent filament, as the associated
recirculation zone separates from the membrane wall;
x increase towards the upstream side of an adjacent filament, as the previous
recirculation zone reattaches to the membrane wall; and
x at low channel Reynolds numbers, reach a local maximum slightly
downstream of the reattachment point.
These trends are consistent with previous observations in the literature that these
reattachment and separation points are associated with regions of high and low
mass transfer, respectively. At higher channel Reynolds numbers, the magnitudes
of the flux and solute rejections reach a maximum directly opposite the spacer
filaments, which is consistent with the present simulation results showing this
location to coincide with the maximum wall shear. The two channel Reynolds
numbers (Rech = 100, 200) chosen for these simulations thus appear to straddle the
point where the bulk hydrodynamic flow features, rather than the near-wall
recirculation regions, begin to govern the permeate flux behaviour for this
particular system.
The results also revealed the transverse gradient in solute concentration (when
normalised appropriately) to vary between individual solutes in a multicomponent
solution. Considering an axial location midway between filaments, this effect was
shown to be greater for the membrane wall opposing the downstream filament,
commensurate with the predicted trends in flux and rejection variation.
301
While the bulk hydrodynamic behaviour and overall permeate flux remained
essentially constant between cases using constant rejection and varying rejection
(i.e. the coupled model), some variation was observed in the near-wall region both
in terms of the solute mass fraction distribution, and the distribution of the
permeate flux. This variation became more apparent at higher Reynolds number,
and where higher solute rejections were predicted.
An interesting effect was also observed for the overall magnitude of permeate flux
when compared between a single salt solution, a multiple salt solution, and a
complex mixed solution (multiple salts with an additional uncharged solute).
Although constant ionic strength and hydrodynamic forcing parameters (inlet
velocity and transmembrane pressure) were maintained across all simulations,
significant variations in the overall magnitude of the permeate flux were
observed: a change in the salt composition from pure NaCl to a 0.2:08
NaCl:Na2SO4 mixture increased the overall flux by approximately six to seven
percent, while further adding sucrose to the salt mixture reduced the overall flux
by approximately nine percent. These results give some justification to the use of
such a coupled model, in that accounting for the subtle changes in the mass
transfer behaviour of the near-wall region. This evidently may have significant
effects on the flux behaviour of the membrane which are not intuitively apparent
without use of a rigorous simulation tool. Future numerical studies of this effect
using a wider range of solution compositions (i.e. a greater number of different
302
solutes, and range of ionic strengths) would be useful, and could be relatively
easily verified by experimental studies of the overall permeate flux magnitude.
In summary, the coupled model approach can be seen to be able to well describe
the behaviour of a membrane channel over a range of conditions. The principal
limitation of the method however remains its computational expense: as noted in
Chapter 6, the coupled method introduces at least an order of magnitude increase
in computational expense compared to a constant rejection model. In real terms, a
steady-state solution for one of the coupled models simulated in this chapter
requires in the order of several days to a week of computational time on a typical
desktop computer, depending on the complexity of the feed solution. This is for
most purposes prohibitively large to be used as a practical tool for rapidly
predicting membrane system behaviour, let alone considering the use of transient
models to describe the filtration history of a system. Nevertheless, the coupled
approach may prove to be a useful tool in examining the fundamental behaviour
of membrane channels, provided sufficiently detailed experimental data can be
gathered to test the model predictions; and, as computational abilities inevitably
increase, may become more widely accessible.
303
Chapter 8
Conclusions
8.1 Overview
This research has presented an integrated approach for predictive modelling of
membrane filtration, by coupling mass transfer models to predict removal of
pollutants by the membrane with computational fluid dynamics simulations of the
membrane channel hydrodynamics. The objective of the study to develop a
general predictive model of the membrane filtration process has been realised, at
least for the high-pressure processes of NF and RO. This has been achieved by the
use of rigorous multicomponent hydrodynamic models, including true permeable
wall boundary conditions allowing spatially variable permeate flux and solute
rejection, in conjunction with numerical methods for predicting rejection for
mixtures of arbitrary numbers of charged and uncharged solutes, including the
effects of molecular shape. The coupling of these traditionally disparate modelling
approaches has yielded a model which should more comprehensively describe the
complex behaviours in a membrane channel, including the common phenomenon
of concentration polarisation.
In the following, conclusions drawn from this research are presented first
regarding the use of computational fluid dynamics (CFD) techniques to describe
hydrodynamics in membrane filtration, and suggestions are made for future work
304
in this area. Secondly, concluding arguments are made on the use of mass transfer
models to predict rejection in membrane filtration, along with recommendations
for future extensions of the present work. Finally, a set of conclusions and
recommendations for future work using the coupled hydrodynamic-mass transfer
approach are presented, including potential future applications of the coupled
modelling approach.
8.2 Hydrodynamic modelling – conclusions and recommendations
The commercial CFD code Ansys CFX was employed to predict the basic
hydrodynamic behaviour of a range of membrane channel configurations typically
encountered in practical applications. This approach was shown to describe the
hydrodynamic behaviour of these channels well, and was verified against
analytical, experimental and computational data from the literature. The primary
features of the basic hydrodynamic simulations conducted were as follows:
x Multicomponent fluid models were successfully employed with
concentration-dependent, spatially variable transport properties. This
means that the relevant transport properties of the solution (the viscosity,
density, diffusivity, and osmotic pressure) can vary spatially within the
membrane channel. Importantly, this allows for the concentration
polarisation boundary layer common to high-pressure membrane filtration
processes to be simulated; and
305
x True permeable wall boundary conditions at the membrane surface were
imposed, which allows for a more realistic description of the
hydrodynamics near the membrane surface than the more simplistic
impermeable wall or dissolving-wall approximations traditionally used.
The other benefit of this approach is that, in conjunction with the spatially
variable transport properties of the multicomponent fluid approach, it
allows for the permeate flux to vary spatially throughout the membrane
channel, and hence provide a more physical description of the channel
hydrodynamics.
The resulting implementation is general enough to describe most practical
situations for the high pressure membrane filtration processes of NF and RO.
However, the methods used to determine the spatially varying transport properties
are derived from empirical relationships presented in the literature for a limited
range of compounds. Further research to develop more general methods for
predicting the variation in these properties would therefore be a significant
improvement.
In addition, the multicomponent fluid approach employed does not attempt to
describe the behaviour of solutions containing particulate components, as
observed in lower-pressure membrane filtration (MF and UF). Future extension of
the modelling approach to include this behaviour may be a worthwhile endeavour.
306
Laminar flow models were successfully used to predict membrane channel
hydrodynamics in the flow regimes typically observed for NF and RO
membranes. For higher channel Reynolds numbers, transient laminar flow models
with very fine spatial and temporal discretisation were shown to be effective at
directly simulating unsteady flow structures without recourse to turbulence
models. Turbulence models are generally unsuitable for the relatively low
Reynolds numbers encountered in NF and RO, though they may be applicable for
the higher velocities of MF and UF. Investigation of the application of turbulence
models for these low-pressure processes may provide a useful avenue of future
investigation.
Finally, simulations were performed for several simplified 2D geometries
representative of a range of practical membrane geometries. It was demonstrated
that these 2D simulations can describe the behaviour of membrane channels in
reasonable detail, both with and without hydrodynamic spacers in the channel.
Some investigation was carried out into the use of 3D models of membrane
channels. It was concluded that, given the currently available computational
resources, 3D modelling is still largely infeasible, especially considering the very
fine spatial discretisation required to resolve the fine concentration polarisation
boundary layer, and, in transient simulations of unsteady flow, the extremely
small temporal discretisation also required.
In summary, it is recommended that future investigation into the hydrodynamic
modelling of membrane channels focus on the following areas:
307
x Development of general methods for predicting spatial variation in
transport properties for arbitrary multicomponent mixtures;
x Incorporation of particulate transport effects to potentially describe UF
and MF processes;
x Study of turbulence modelling applications for UF and MF processes; and
x Use of 3D modelling as computational resources advance.
8.3 Mass transfer modelling for rejection prediction – conclusions and
recommendations
A refined version of the classic Donnan Steric Partitioning Model (DSPM) has
been developed to present a consistent approach for predicting the rejection of
both charged and uncharged solutes. The method is based on the numerical
solution to the extended Nernst-Planck equations, assuming that solute transport
through membrane pores occurs through the processes of convection, diffusion,
and electric migration. The membrane is characterised in terms of three physically
meaningful parameters: the pore radius, an effective pore length or membrane
thickness, and a volumetric charge density. The method presented here differs
from previous implementations in that:
x The influence of molecular shape on the steric partitioning of the solute is
considered for both charged and uncharged molecules. Solute molecule
308
geometry is represented by a characteristic length and width, rather than
assuming a spherical solute. Previous models have only employed
molecular shape effects for large uncharged solutes: here, this is extended
for all uncharged and charged solutes, with the exception of monatomic
ions (where the spherical assumption remains well justified);
x Numerical methods have been used to solve the system of non-linear
equations describing the Donnan partitioning at the pore inlet and outlet,
which have historically only been solved analytically for small numbers of
univalent solutes. The numerical method employed here allows the
solution to be generalised to arbitrary numbers of charged solutes of
arbitrary valence; and
x The numerical solution to the Donnan partioning problem is used in
conjunction with a numerical solution of the boundary value problem
posed by the extended Nernst-Planck equations to iteratively determine the
concentration profile of solutes throughout the membrane pores and hence
predict solute rejection. The general nature of the numerical solution
means that rejection may be predicted for complex solutions of more than
three charged solutes, which is uncommon in the literature. In addition, the
rejection of an arbitrary number of uncharged solutes may also be
predicted simultaneously.
309
The refined model was shown to well predict the rejection of solutions containing
uncharged compounds, single salt solutions, and multiple salt solutions. The
model was also able to predict rejections well for a variety of feed solutions for a
given membrane using a consistent set of fixed membrane parameters (pore radius
and membrane thickness) and an isotherm to describe the variation in the
membrane volumetric charge density with the ionic strength of the feed solution.
This is an improvement over similar previous model implementations by other
authors where these parameters are adjusted between different feed solutions to
provide an adequate fit. Instead, the fact that these parameters can be maintained
at constant values for a given membrane indicates that the membrane
characterisation method is robust, and the chosen parameters are indeed
physically representative of the membrane.
In some cases with large uncharged solutes, agreement between experimental data
and the refined model predictions was less than ideal. In particular, complete
rejection was predicted for some compounds where experimental data indicated
slightly less than unity rejection. This discrepancy is believed to be due to the
existence of a pore size distribution in the physical membrane which is not
accounted for in the model. Future work on extending the model should include
such a pore size distribution within the mathematical model, as well as developing
a method by which this distribution can be characterised using experimental data.
Finally, the refined DSPM presented here does not include dielectric effects,
which have relatively recently been shown to be important in predicting rejection
310
in the case of multivalent counter-ions. However, dielectric effects can be
successfully incorporated into a DSPM model, and the numerical methods
required for solution would be substantively similar to those presented here.
Therefore, it is recommended that future work to include dielectric exclusion
effects into the general numerical solution with molecular shape effects would be
a definite improvement in predicting solute rejection for these processes.
8.4 Coupled hydrodynamic-mass transfer modelling – conclusions and
recommendations
Using the hydrodynamic models of the membrane channel and the mass-transfer
models to describe solute rejection, a coupled modelling approach was developed
and successfully employed. Numerical routines were developed using Fortran to
interface with the CFX solver, and predict rejections for multicomponent
solutions to be applied as boundary conditions for the coupled hydrodynamic
model. The coupled model was successfully run for a range of conditions, and
indicated several interesting phenomena not commonly discussed in the
hydrodynamic modelling literature, including:
x Spatial variation in solute rejection over the membrane surface and along
the membrane channel;
311
x Solute-dependent concentration polarisation boundary layers (i.e. varying
concentration gradients in the near-wall region for different solutes in
multicomponent solutions);
x The formation of an inverse concentration polarisation boundary layer in
the presence of negatively rejected solutes; and
x Significant variation in the overall permeate flux magnitude for single-salt
solutions, multiple salt solutions, and complex mixtures (multiple salts and
uncharged solutes), even though consistent ionic strength was maintained
throughout the simulations.
The predicted effect on permeate flux magnitude can be ascribed to the coupled
model’s ability to more fully describe the variation in solution viscosity and
osmotic pressure in the near-wall region and concentration polarisation boundary
layer (which influence permeate flux to a significant extent in these processes). A
useful area for future investigation would be to carry out a more detailed
numerical study of this effect using a wider range of solution compositions, which
could be relatively easily verified by experimental studies of the overall permeate
flux magnitude.
Some additional conclusions can be made regarding the coupled model approach,
in particular the treatment of the diffusion of solutes within the bulk flow of the
hydrodynamic model. In the uncoupled hydrodynamic only approach, individual
diffusion coefficients could be employed for each solute; however, in the coupled
312
approach this could not guarantee electroneutrality conditions at the membrane
surface, necessary for the solution of the mass-transfer model to predict solute
rejection. A simplified approach using diffusivities of salt ion-pairs, rather than
individual ions, was successfully employed to enforce electroneutrality
throughout the hydrodynamic model. This avoids the severe numerical difficulties
encountered by incorporating a full electro-diffusion model (such as the extended
Nernst-Planck equation) into the bulk flow of the hydrodynamic model as
reported by other authors. However, overcoming these numerical difficulties to
provide a rigorous coupled model with global incorporation of electro-diffusive
effects, not just within the membrane pores, remains a pressing goal for this
modelling approach. Further research efforts would be well expended in this area.
Other recommendations for future work on a coupled hydrodynamic-mass transfer
model have largely been referred to in the previous sections on hydrodynamics
and mass transfer separately. These include:
x Development of a general method for prediction of mixed solution
transport properties;
x Extension of the models to 3D geometries if computational resources
allow;
x Incorporation of other effects to make the model more generally applicable
to the lower-pressure processes of MF and UF, such as turbulence
modelling and particulate modelling; and
313
x Refinement of the mass-transfer model to include dielectric exclusion
effects.
Along with the recommendations presented for future work shown above, it is
also useful to consider the potential future applications of the model. An area in
which the coupled model approach would be particularly useful is in the
modelling of fouling reduction techniques, such as pulsatile flow and use of
‘turbulence’ promoters (either traditional spacer nets or Dean vortex promoters).
The ability of the model to simulate the formation of the concentration
polarisation layer in detail lends itself well to evaluating the applicability of such
techniques, which are largely aimed at minimising or eliminating the
concentration polarisation phenomenon. In addition, the use of transient models to
simulate the most likely unsteady flow in such situations would be appropriate.
Extension of the modelling approach to include the physics of other fouling
reduction techniques (such as multiphase flow to describe air sparging, or acoustic
propagation to describe ultrasonic fouling reduction) may also provide a future
application of the coupled model approach.
Finally, while a general entirely predictive model of membrane filtration
processes may be an overly optimistic goal for the near future, it is possible that
computational resources will advance to the point where full 3D simulations of
entire membrane module are possible. In this case, the coupled hydrodynamic-
mass transfer modelling approach, with the appropriate extensions suggested here,
314
may finally realise the goal of a completely predictive model of membrane
filtration.
315
Appendix A
Sample MATLAB numerical codes for predictive rejection
model
A.1 Molecular length and width calculations
The following code is used to determine the molecular length and width of a
solute based on the Cartesian coordinate output from the MOPAC2009 software
using a three-dimensional bounding box method. A graphical sample of the code
output for five test molecules (glucose, sucrose, cyanazine, glycerine, and Vitamin
B12) is also shown.
function [L,MWd] = molprjn2(xyz,atoms)

% Calculate molecular length L and width MWd as per Kiso et al
% Lookup table for VDW radius

VDWlookup = {'Ag' 1.72
'Ar' 1.88
'As' 1.85
'Au' 1.66
'Br' 1.85
'C' 1.7
'Cd' 1.58
'Cl' 1.75
'Co' 2.00
'Cu' 1.4
'F' 1.47
'Ga' 1.87
'H' 1.2
'He' 1.4
'Hg' 1.55
'I' 1.98
'In' 1.93
'K' 2.75
'Kr' 2.02
'Li' 1.82
'Mg' 1.73
316
'N' 1.55
'Na' 2.27
'Ne' 1.54
'Ni' 1.63
'O' 1.52
'P' 1.8
'Pb' 2.02
'Pd' 1.63
'Pt' 1.72
'S' 1.8
'Se' 1.9
'Si' 2.1
'Sn' 2.17
'Te' 2.06
'Tl' 1.96
'U' 1.86
'Xe' 2.16
'Zn' 1.39};
% Loop over structure 'atoms' to look up VDW radius for each

atom
for i = 1:length(atoms)
% Loop through structure 'VDWlookup'

for j = 1:length(VDWlookup)
if isequal(atoms(i),VDWlookup(j,1))
VDW(i) = VDWlookup(j,2);
break
end
end %j
end %i
% Convert cell array to numeric array

VDW=cell2mat(VDW');
% Create cloud of n points at radius VDW around each point in

xyz
% Create n random angles
n = 100;
[randangle] = randomAngle3d(n);
clouds = [];
for i = 1:length(xyz)
for j = 1:length(randangle)
317
clouds = [clouds; xyz(i,:) +
VDW(i)*sph2cart2(randangle(j,:))];
end %j
end %i
% Find length L - very inefficient but it works...

L=0;
for i = 1:(length(clouds)-1)
for j = (i+1):length(clouds)
% Start at second atom

% Distance between atoms including VDW radius
distance=sqrt(sum((clouds(j,:)-clouds(i,:)).^2));
% Find greatest distance

if distance>L
L = distance;
first = i;
second = j;
end %if
end %j
end %i
% Find width W
W = 0;
for k = 1:(length(clouds)) % Check each point in clouds
% Point to test whether in cylinder or not

testpt = clouds(k,:);
% Endpoints of cylinder
pt1 = clouds(first,:);
pt2 = clouds(second,:);
% Length of cylinder squared

lengthsq = L^2;
% Translate so pt1 is origin

d = pt2 - pt1;
% Vector from pt1 to test point

p = testpt - pt1;
% Dot the d and pd vectors to see if point lies behind the

% cylinder cap at pt1.x, pt1.y, pt1.z
dotpd = dot(p,d);
318
% distance squared to the cylinder axis including VDW
radius:
dsq = dot(p,p) - dotpd^2/lengthsq;
if sqrt(dsq) > W
W = sqrt(dsq);
third = k;
end %if
end %k
% Create plane containing points first, second, third

plane = createPlane(clouds(first,:), clouds(second,:),
clouds(third,:));
% Transform projected points so that they all lie on global

planes
% Create transformation matrix from plane to global plane
Tform = createBasisTransform3d('global',plane);
% Determine angles of rotation of plane
[PHI, THETA, PSI] =
rotation3dToEulerAngles(createBasisTransform3d(plane,'global')
);
% Transform projected points onto plane
cloudsTform = transformPoint3d(clouds, Tform);
% Find bounding box on transformed plane

BOXTform = point3dBounds(cloudsTform);
% Get dimensions of bounding box
L1 = BOXTform(2)-BOXTform(1);
% Get centre of bounding box
C1 = 0.5*(BOXTform(1)+BOXTform(2));
boxcentreTform = [C1 C2 C3];
% Transform centre back to global plane

boxcentre = transformPoint3d(boxcentreTform,
createBasisTransform3d(plane,'global'));
% Get first, second and third atoms

firstxyz = floor(first / n);
secondxyz = floor(second / n);
thirdxyz = floor(third / n);
% Draw clouds, atom centres, and spheres

%drawPoint3d(clouds,'.')
hold on
for i = 1:length(xyz)
319
drawSphere(xyz(i,:),
0.5*ones(size(VDW(i))),'FaceAlpha',0.5,'FaceColor','r');
drawSphere(xyz(i,:), VDW(i),'FaceAlpha',0.1);
end
% Plot enclosing box

CUBOID = [boxcentre L1 L2 L3 PHI THETA PSI];
drawCuboid(CUBOID,'FaceAlpha',0.05)
drawPoint3d(boxcentre)
axis equal
% Calculate molecular width

S = L2*L;
MWd = 0.5*sqrt(S);
end % molprjn2
320
321
322
A.2 Non-spherical steric partition coefficient calculation
The following code is used to determine steric partitioning coefficients for non-
spherical solutes by numerical integration.
%% Steric hindrance calculator for non-spherical solute

% Based on model of Kiso et al. (2010)
% phi = kiso(rp,L,MWd)
% Arguments are:
%
% rp - membrane pore radius [m]
% L - molecular length including VDW radius [m]
% MWd - molecular width including VDW radius [m]
%
% Returns:
%
% phi - column vector of steric partition coefficients
%
%%
function phi = kiso(rp,L,MWd)
% Sub-function for evaluating overall solute partition

coefficient
function F = int(alpha) % alpha is angle of molecular
orientation [rad]
% Effective molecular length Lp [m]
Lp = L(j).*cos(alpha)+2.*MWd(j).*sin(alpha);
% Solute-accessible radius in the pore a [m]
a = (rp(j).^2-(Lp./2).^2).^0.5-MWd(j);
if isreal(a)==0
a = 0;
end
% Overall solute partition coefficient
F = (a./rp(j)).^2.*sin(alpha);
end %int
for j=1:length(L)
% Evaluate integral of function F over 0:pi/2
phi(j) = quad(@int,0,pi/2);
end %j
phi=phi';
end %kiso
323
A.3 Donnan partitioning solver
The following code is used to solve the Donnan partitioning problem for charged
solutes, using a numerical solver to find the roots of a system of non-linear
equations.
%% Donnan equilibrium constant root finder

% APPEARS TO WORK GENERALLY!!!!
% Calculates roots of expression for Donnan equilibrium
constant
%
% Syntax: K = doneqn3(z,C,zX,CX,phi)
%
% Arguments are:
%
% z - column vector of valences of ions 1 to n
% C - column vector of solution concentrations of ions i to n
[mol m^-3]
% zX - valence of fixed membrane charge
% CX - concentration of fixed membrane charge [mol m^-3]
% phi - column vector of steric partition coefficients
function K = doneqn3(z,C,zX,CX,phi)
% Electroneutrality in solution means sum(z.*C) = 0

% Electroneutrality in membrane means sum(z.*y) + zX*CX = 0
% where y is the vector of concentrations of ions i to n
INSIDE THE
% MEMBRANE [mol m^-3]
% We use the combination of these two conditions as the first
equation of
% our function F
% i.e. sum(z.*y) + zX*CX + sum(z.*C) = 0
% As Donnan potential is identical for each ion i etc:

% 1/z1*log(C1/(y1*phi1))=1/zi*log(Ci/(yi*phii))
% therefore for the jth ion:
% 1/z1*log(C1/(y1*phi1)) - 1/zj*log(Cj/(yj*phij)) = 0 for each
j, j ~=1:
% This will give us another (n-1) equations for our function
F, giving
% us n equations with n unknowns (y1,y2,...,yn)
% Nested subfunction for fsolve to solve

function F = obj(y)
F = zeros(length(z),1);
% j represents all ions besides the first
324
j = 2:length(z);
% Electroneutrality in membrane and solution
F(1) = sum(z.*y) + zX*CX + sum(z.*C); %sum(Jv*C.*(1-
rejection).*z)
% % Equality of Donnan potential for each ion
% F(j) = 1./z(1).*log(C(1)./(y(1).*phi(1))) -
1./z(j).*log(C(j)./(y(j).*phi(j)));
% Equality of Donnan potential for each ion
F(j) = 1./z(1).*log(y(1)./(C(1).*phi(1))) -
1./z(j).*log(y(j)./(C(j).*phi(j)));
end
zcheck = z/zX;
for i=1:length(zcheck)
if zcheck(i) < 0 % counter-ion
guess(i) = C(i)*abs(zcheck(i)*CX);
else % co-ion
guess(i) = C(i)*abs(zcheck(i)/CX);
end %if
end %i
% Set options for root-finding function fsolve

options = optimset('Display','off','TolFun',1e-15,'TolX',1e-
12);
% Solve using root-finding function fsolve with the initial

guess that
% y/C = guess(i) for each ion i
[x,fval] = fsolve(@obj,guess',options);
% Calculate K
K = C.*phi./real(x);
% Check all Donnan voltages are equal

Donnancheck = log(1./K)./z;
origx = x;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%
% fsolve may return negative roots which are not physically
meaningful
% Check to see if this is the case and if negative roots were
found solve
% again using random number multiplied by absolute value of
original
% solution as starting guess
while min(x)<0 || abs(max(Donnancheck) - min(Donnancheck)) >
1E-3
newx = abs(origx).*rand(1);
325
[x,fval] = fsolve(@obj,newx,options);
% Recalculate K
K = C.*phi./real(x);
% Recheck all Donnan voltages are equal

Donnancheck = log(1./K)./z;
end
%
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%
%
end
326
A.4 Nernst-Planck equation solver
The following code is used to solve the extended Nernst-Planck equations for
charged solutes, using a numerical solver to find the roots of a system of non-
linear equations.
%% Nernst-Planck equation solver

% FULLY WORKING
% Solves extended Nernst-Planck equation for arbitrary number of
ions in
% solution
%
% Syntax:
%
% rejection = np3n(C,z,D,rs,Jv,rp,deltax,zX,CX)
% rejection = np3n(C,z,D,rs,Jv,rp,deltax,zX,CX,steric)
% rejection = np3n(C,z,D,rs,Jv,rp,deltax,zX,CX,steric,guess)
%
% Arguments are:
%
% C - column vector of feed concentrations of ions [mol m^-3]
% (note 1 M = 1 mol L^-1 = 1000 mol m^-3)
% z - column vector of valences of ions
% D - column vector of diffusivities of ions [m^2 s^-1]
% rs - column vector of Stokes radii of ions [m]
% NOTE: if non-spherical solutes, specify rs = 0
% Jv - volume flux [m s^-1]
% deltax - membrane pore length [m]
% zX - effective membrane charge valence
% CX - effective membrane charge density [equiv m^-3]
%
% Optional arguments are:
%
% steric - logical value. 0 - no steric effects, 1 - steric
effects.
% L - column vector of molecular lengths [m] - used if non-
spherical model
% desired
% MWd - column vector of molecular widths [m] - used if non-
spherical model
% desired
% NOTE: if spherical solutes, specify L = MWd = 0
% guess - vector of guesses for rejection of ions
%
% Calls the following user functions:
%
327
% doneqn3.m
% kiso.m (if non-spherical model desired)
%
function rejection =
np3n(C,z,D,rs,Jv,rp,deltax,zX,CX,steric,L,MWd,guess)
% Set steric effects to be included if not specified

if nargin < 10, steric = 1; end
% Check if molecular length and width parameters have been

specified
if nargin == 12, nonsph = 1; else nonsph = 0; end
% Set guess at rejections to 0.5 for all ions if not specified

if nargin <13, guess = 0.5*ones(size(C)); end
% tic
% Define physical constants

% Faraday constant [C mol^-1]
F = 96485.3365;
% Temperature [K]
T = 293;
% Ideal gas constant [J mol^-1 K^-1]
R = 8.3144621;
% Check input vectors are same length

if length(C) == length(z) && length(C) == length(z) &&
length(C) ==...
length(D) && length(C) == length(rs)
else
error('Input vectors must be same length!')
end
% Define rejections as unknown parameters to pass to solver

parameters = guess;
% Calculate ratio of Stokes radius to pore radius, lambda

lambda = rs/rp;
if steric == 1;
% Steric partioning is included
% If molecular length and width are specified, use non-
spherical pore
% model
if nonsph == 1;
for i = 1:length(rs)
if rs(i) == 0 % no radius specified, use non-
spherical model
phi(i,1) = kiso(rp,L(i),MWd(i));
328
% Estimate rs based on MWd to calculate
hindrance factors
% for non-spherical solute
rs(i) = (1.42*(MWd(i)*1E9)-0.142)*1E-9;;
else % radius specified, use spherical
model
phi(i,1) = (1 - lambda(i))^2;
end %if
end %i
else
% Calculate spherical steric partition factors for all
molecules
phi = (1 - lambda).^2;
end % if
else
% Steric partioning not included
% Leave all steric partition factors equal to one
phi = ones(size(C));
end
% Calculate hindrance factors for diffusion using relationship

% Bowen et al. (1996)
% Kd = -1.705.*lambda+0.946;
% Deen (1987)
Kd = 1-2.30*lambda+1.154*(lambda.^2)+0.224*(lambda.^3);
% Calculate hindrance factors for convection using

relationship
% Bowen et al. (1996))
% Kc = -0.301.*lambda+1.022;
% Deen (1987)
G = 1+0.054*lambda-0.988*(lambda.^2)+0.441*(lambda.^3);
Kc = (2-phi).*G;
% Calculate diffusivity in pore [m^2 s^-1]

Dp = D.*Kd;
% Calculate Donnan potential at inlet [V] which is constant

for all ions
% Use external function doneqn3 to calculate vector of inlet
Donnan
% voltages
Donnaninletvector =
R.*T./(z.*F).*log(1./doneqn3(z,C,zX,CX,phi));
% At the outlet the Donnan voltage will be given by:

% R.*T./(z.*F).*log(1./doneqn3(z,C.*(1-parameters),zX,CX,phi))
% This is incorporated into the boundary conditions as
parameters is as
% yet unknown
% Calculate concentrations just inside pore inlet
329
inlet = C.*phi.*exp(z.*F/(R*T).*Donnaninletvector);
% BVPINIT is used to form an initial guess for a mesh of 100

equally spaced
% points. The guesses for the concentrations within the pore
% are evaluated in NPINIT. A guess for unknown parameters (the
rejection
% of each ion) is the last argument of BVPINIT.
solinit = bvpinit(linspace(0,deltax,100),@npinit,parameters);
% The BVP solver returns the structure 'sol'. The computed

rejections are
% returned in the field sol.parameters.
sol = bvp4c(@npode,@npbc,solinit);
rejection = sol.parameters;
Donnanoutletvector = R.*T./(z.*F).*log(1./doneqn3(z,C.*(1-
rejection),zX,CX,phi));
% Check zero current condition

I = sum(Jv*C.*(1-rejection).*z);
% ------------------------------------------------------------
-----------
function dydx = npode(x,y,parameters)
% Derivative function. Variables provided by the outer
function
%dydx = Jv./Dp.*(Kc.*y-C.*phi.*(1-parameters))-
z.*y.*F/(R*T)*sum(Jv./Dp.*z.*(Kc.*y-C.*phi.*(1-
parameters)))/(F/(R*T)*sum(z.*z.*y));
%dydx = Jv./Dp.*(Kc.*y-C.*phi.*(1-parameters))-
z.*y.*sum(Jv./Dp.*z.*(Kc.*y-C.*phi.*(1-
parameters)))/(sum(z.*z.*y));
dydx = Jv./Dp.*(Kc.*y-C.*(1-parameters))-
z.*y.*sum(Jv./Dp.*z.*(Kc.*y-C.*(1-
parameters)))/(sum(z.*z.*y));
end
% ------------------------------------------------------------
function res = npbc(ya,yb,parameters)
% Boundary conditions. parameters is a required
argument.
res = [ ya - inlet
%yb - C.*phi.*(1-parameters).*exp(-
z.*F/(R*T).*R.*T./(z.*F).*(log(1./doneqn3(z,C.*phi.*(1-
parameters),zX,CX))))];
yb - C.*phi.*(1-
parameters).*exp((log(1./doneqn3(z,C.*(1-
parameters),zX,CX,phi))))];
end
330
% ------------------------------------------------------------
-------------
% Auxiliary function -- initial guess for the solution - OK
function yinit = npinit(x)
yinit = inlet.*(1-(x./deltax).*parameters);
end
% ------------------------------------------------------------
-------------
% toc
end % np3n
331
A.5 Uncharged solute rejection calculator
The following code is used to find the analytical solution for the rejection of
uncharged solutes.
%% Rejection calculator for uncharged solutes

%
% Syntax: rejection = unchargedJv(L,MWd,rp,Jv,deltax)
%
% Arguments are:
%
% L - molecular length including VDW radius [m]
% MWd - molecular width including VDW radius [m]
% Jv - volumetric flux [m s^-1]
%
% Calls the following user functions:
%
% kiso.m
%
function rejection = unchargedJv(L,MWd,rp,Jv,deltax)
% Define physical constants

% Faraday constant [C mol^-1]
F = 96485.3365;
% Temperature [K]
T = 293;
% Ideal gas constant [J mol^-1 K^-1]
R = 8.3144621;
% Thickness of a water molecule [nm]
d = 0.28;
% Boltzmann constant k [J?K^?1]
k = 1.3806504E-23;
% Viscosity of bulk solution eta0 [Pa s]
eta0 = 0.001;
% Calculate steric partioning coefficient phi and molecular

width MWd by
% calling external function kiso
phi = kiso(rp,L,MWd);
% Calculate approximate Stokes radius

rs = (1.42*(MWd*1E9)-0.142)*1E-9;
% Calculate ratio of Stokes radius to pore radius, lambda

lambda = rs./rp;
332
% Calculate hindrance factors for diffusion using relationship
from Bowen
% et al. (1997)
Kd = 1-2.30*lambda+1.154*(lambda.^2)+0.224*(lambda.^3);
% Calculate lag coefficient using relationship from Bowen et

al. (1997)
G = 1+0.054*lambda-0.988*(lambda.^2)+0.441*(lambda.^3);
% Calculate hindrance factors for convection using

relationship from Bowen
% et al. (1997)
Kc = (2-phi).*G;
% Calculate viscosity of water in pore eta [Pa s]

eta = (1+18*(d*1E-9./rp)-9*(d*1E-9./rp).^2)*eta0;
% Estimate bulk diffusivity [m^2 s^-1]

D = (rs.*eta0).^-1*(k*T/(6*pi));
% Calculate diffusivity in pore [m^2 s^-1]

Dp = D.*Kd;
% Calculate Peclet number

Pe = Jv.*Kc.*deltax./Dp;
% Calculate rejection
rejection = 1-phi.*Kc./(1-(1-phi.*Kc).*exp(-Pe));
end % unchargedJv
333
Appendix B
Sample FORTRAN numerical codes for coupled rejection
model – User Routine approach
B.1 Wrapper subroutine
The following code is used to pass the arguments from the CFX solver to the
other subroutines used to perform the numerical calculations.
#include "cfx5ext.h"
dllexport(np3nwrapper)
SUBROUTINE NP3NWRAPPER (
& NLOC, NRET, NARG, RET, ARGS, CRESLT, CZ,DZ,IZ,LZ,RZ )
CC
CD User routine: template for user CEL function
CC
CC --------------------
CC Input
CC --------------------
CC
CC NLOC - size of current locale
CC NRET - number of components in result
CC NARG - number of arguments in call
CC ARGS() - (NLOC,NARG) argument values
CC
CC --------------------
CC Modified
CC --------------------
CC
CC Stacks possibly.
CC
CC --------------------
CC Output
CC --------------------
CC
CC RET() - (NLOC,NRET) return values
CC CRESLT - 'GOOD' for success
CC
CC --------------------
334
CC Details
CC --------------------
CC
CC
CC
CC============================================================
==========
C
C ------------------------------
C Preprocessor includes
C ------------------------------
C
C
C ------------------------------
C Global Parameters
C ------------------------------
C
C
C ------------------------------
C Argument list
C ------------------------------
C
INTEGER NLOC,NARG,NRET
C
CHARACTER CRESLT*(*)
C
REAL ARGS(NLOC,NARG), RET(NLOC,NRET)
C
INTEGER IZ(*)
CHARACTER CZ(*)*(1)
DOUBLE PRECISION DZ(*)
LOGICAL LZ(*)
REAL RZ(*)
C
C ----------------------------------------------------------
C Input arguments are:
C
C 'Na plus.mf' is stored in ARGS(1:NLOC,1)
C 'Cl minus.mf' is stored in ARGS(1:NLOC,2)
C 'Na plus.Molar Mass' is stored in ARGS(1:NLOC,3)
C 'Cl minus.Molar Mass' is stored in ARGS(1:NLOC,4)
C 'rsNa plus' is stored in ARGS(1:NLOC,5)
C 'rsCl minus' is stored in ARGS(1:NLOC,6)
C 'DNa plus' is stored in ARGS(1:NLOC,7)
C 'DCl minus' is stored in ARGS(1:NLOC,8)
C 'zNa plus' is stored in ARGS(1:NLOC,9)
C 'zCl minus' is stored in ARGS(1:NLOC,10)
C 'LNa plus' is stored in ARGS(1:NLOC,11)
C 'LCl minus' is stored in ARGS(1:NLOC,12)
C 'MWdNa plus' is stored in ARGS(1:NLOC,13)
335
C 'MWdCl minus' is stored in ARGS(1:NLOC,14)
C 'zX' is stored in ARGS(1:NLOC,15)
C 'CX' is stored in ARGS(1:NLOC,16)
C 'rp' is stored in ARGS(1:NLOC,17)
C 'deltax' is stored in ARGS(1:NLOC,18)
C 'FluxVar' is stored in ARGS(1:NLOC,19)
C
C Output arguments are:
C
C 'Rejection Na plus' is stored in RET(1:NLOC,1)
C ----------------------------------------------------------
C
C ------------------------------
C External routines
C ------------------------------
C
C
C ------------------------------
C Local Parameters
C ------------------------------
C
C
C ------------------------------
C Local Variables
C ------------------------------
C
REAL REJ
REAL ARGSOUT(NARG)
REAL REXX((NARG-5)/7)
INTEGER I,J
INTERFACE
SUBROUTINE NP3NVECTOR(ARGSOUT,NARG,REXX)
REAL :: ARGSOUT(:),REXX(:)
INTEGER NARG
END SUBROUTINE NP3NVECTOR
END INTERFACE
C
C ------------------------------
C Stack pointers
C ------------------------------
C
C=============================================================
==========
C
C ---------------------------
C Executable Statements
C ---------------------------
C
C------------------- Loop over locations NLOC ----------------
----------
336
DO I=1,NLOC
DO J = 1,NARG
ARGSOUT(J) = ARGS(I,J)
END DO
CALL NP3NVECTOR(ARGSOUT,NARG,REXX)
C Set rejection to value contained in REXX
RET(I,1) = REXX(1)
END DO
C------------------- End loop over locations NLOC ------------

----------
C Print results for last NLOC for debugging purposes
CALL USER_PRINT_REAL('Na plus.mf',ARGSOUT(1))

CALL USER_PRINT_REAL('Cl minus.mf',ARGSOUT(2))
CALL USER_PRINT_REAL('Na plus.Molar Mass',ARGSOUT(3))
CALL USER_PRINT_REAL('Cl minus.Molar Mass',ARGSOUT(4))
CALL USER_PRINT_REAL('rsNa plus',ARGSOUT(5))
CALL USER_PRINT_REAL('rsCl minus',ARGSOUT(6))
CALL USER_PRINT_REAL('DNa plus',ARGSOUT(7))
CALL USER_PRINT_REAL('DCl minus',ARGSOUT(8))
CALL USER_PRINT_REAL('zNa plus',ARGSOUT(9))
CALL USER_PRINT_REAL('zCl minus',ARGSOUT(10))
CALL USER_PRINT_REAL('LNa plus',ARGSOUT(11))
CALL USER_PRINT_REAL('LCl minus',ARGSOUT(12))
CALL USER_PRINT_REAL('MWdNa plus',ARGSOUT(13))
CALL USER_PRINT_REAL('MWdCl minus',ARGSOUT(14))
CALL USER_PRINT_REAL('zX',ARGSOUT(15))
CALL USER_PRINT_REAL('CX',ARGSOUT(16))
CALL USER_PRINT_REAL('rp',ARGSOUT(17))
CALL USER_PRINT_REAL('deltax',ARGSOUT(18))
CALL USER_PRINT_REAL('FluxVar',ARGSOUT(19))
CALL USER_PRINT_REAL('Rejection',REXX(1))
CALL USER_PRINT_REAL('Rejection',REXX(2))
C
C Set success flag.
CRESLT = 'GOOD'
C
C=============================================================
==========
337
END
338
B.2 Steric partition subroutine
The following code is used to determine steric partitioning coefficients for
spherical or non-spherical solutes by numerical integration.
subroutine kiso(rp,L,MWd,rs,phi)
! uses QAG (included in quadpack.f90) to find the steric
partioning
! coefficient by numerical integration
implicit none
real, parameter :: llim = 0.0E+00

real abserr
real rlim
real, parameter :: epsabs = 0.0E+00
real, parameter :: epsrel = 0.001E+00
integer ier
integer, parameter :: key = 6
integer neval
real, parameter :: pi = 3.141592653589793E+00
real result
real :: rp,L(:),MWd(:),rs(:),phi(:)
integer j
real Lj,MWdj
rlim = pi/2.e0
do j = 1,size(L)
Lj = L(j)
MWdj = MWd(j)
! If no molecular dimensions supplied, assume

spherical
if (Lj .eq. 0) then ! spherical
phi(j) = (1-(rs(j)/rp))**2
else ! non-spherical
call qag ( phif, llim, rlim, epsabs, epsrel,

key, result, abserr, neval, ier )
phi(j) = result
rs(j) = (1.42*(MWd(j)*1E9)-0.142)*1E-9
339
end if
end do
contains
function phif ( alpha )

! phif is the integrand function
real :: phif,alpha,Lp,a
! Effective molecular length Lp [m]

Lp = Lj*cos(alpha)+2*MWdj*sin(alpha)
! Solute-accessible radius in the pore a [m]

if (rp<(Lp/2)) then
a = 0
else
a = (rp**2-(Lp/2)**2)**0.5-MWdj
end if
! Overall solute partition coefficient

phif = (a/rp)**2*sin(alpha)
return
end function phif
end subroutine kiso
340
B.3 Donnan partitioning solver
The following code is used to solve the Donnan partitioning problem for charged
solutes, using a numerical solver to find the roots of a system of non-linear
equations.
subroutine doneqn3(z,C,zX,CX,phi,K)
! uses HYBRD1 (included in minpack.f90) to solve Donnan
partitioning problem
implicit none
integer ( kind = 4 ) n
real ( kind = 8 ), allocatable :: fvec(:),y(:),origy(:)

integer ( kind = 4 ) :: iflag,info,i
real ( kind = 8 ) tol
real :: z(:),C(:),zX,CX,phi(:),K(:)
real, allocatable :: zcheck(:),Donnancheck(:)
real :: miny,diffDonnancheck,harvest,Donnantol
interface
subroutine hybrd1 ( f02, n, y, fvec, tol, info )
real ( kind = 8 ) y(n)
real ( kind = 8 ) fvec(n)
real ( kind = 8 ) tol
integer ( kind = 4 ) info
external f02
end subroutine hybrd1
end interface
n = size(C)
allocate(fvec(n))
allocate(y(n))
allocate(origy(n))
allocate(zcheck(n))
allocate(Donnancheck(n))
zcheck = z/zX
! Initial guess for solution

do i = 1, n
if (zcheck(i) < 0) then !counter-

ion
341
y(i) = C(i)*abs(zcheck(i)*CX)
else !co-ion
y(i) = C(i)*abs(zcheck(i)/CX)
end if
end do
K = y
! Set options for hybrd1

iflag = 1
tol = 1.0D-6
! Call hybrd1 to solve equations

call hybrd1 ( f02, n, y, fvec, tol, info)
! Calculate K
K = C*phi/abs(y);
! Tolerance for Donnan voltage check

Donnantol = 1E-3
! Check all Donnan voltages are equal

Donnancheck = log(1.0/K)/z;
miny = minval(abs(y))
diffDonnancheck = abs(maxval(Donnancheck) -
minval(Donnancheck))
if ( diffDonnancheck>Donnantol ) then
origy = y
call random_seed
end if
! hybrd1 may not return the correct solution

! Check to see if this is the case and if so solve
! again using random number multiplied by absolute value
of original
! solution as starting guess
do while ((miny .lt. 0.0) .or.
(diffDonnancheck>Donnantol))
call random_number(harvest)
y = abs(origy)*harvest;
! Call hybrd1 to solve equations

call hybrd1 ( f02, n, y, fvec, tol, info)
! Recalculate K
342
K = C*phi/abs(y);
! Recheck all Donnan voltages are equal

Donnancheck = log(1.0/K)/z;
miny = minval(y)
diffDonnancheck = abs(maxval(Donnancheck) -
minval(Donnancheck))
end do
deallocate(fvec)
deallocate(y)
deallocate(origy)
deallocate(zcheck)
deallocate(Donnancheck)
contains
subroutine f02 (n, y, fvec, iflag)
implicit none
real ( kind = 8 ) y(n)
real ( kind = 8 ) fvec(n)
integer ( kind = 4 ) iflag
integer ( kind = 4 ) j
! Electroneutrality in solution and membrane

fvec(1) = sum(z*abs(y)) + zX*CX + sum(z*C)
! Equality of Donnan potential:

do j = 2, n
fvec(j) = (1.0/z(1))*log(abs(y(1))/(C(1)*phi(1)))
- (1.0/z(j))*log(abs(y(j))/(C(j)*phi(j)))
end do
end subroutine f02
end subroutine doneqn3
343
B.4 Nernst-Planck equation solver
The following code is used to solve the extended Nernst-Planck equations for
charged solutes, using a numerical solver to find the roots of a system of non-
linear equations.
subroutine np3nvector(ARGSOUT,NARG,rejection)
use BVP_M
implicit none
TYPE(BVP_SOL) :: SOL
INTEGER NARG
REAL :: ARGSOUT(:),rejection(:)
INTEGER NODE,NPAR,LEFTBC,RIGHTBC
! Define input arguments from vector ARGSOUT

real,allocatable ::
L(:),MWd(:),rs(:),D(:),z(:),C(:),mf(:),mm(:)
real rp,deltax,zX,CX,Jv
! Define physical constants

real, parameter :: F = 96485.3365E0
real, parameter :: T = 293.0E0
real, parameter :: R = 8.3144621E0
real, allocatable :: phi(:),phiout(:),Kd(:),lambda(:)

real, allocatable ::
G(:),Kc(:),Dp(:),Donnaninletvector(:),Kin(:)
real, allocatable :: inlet(:),Kout(:),Donnanoutletvector(:)
real Icheck
double precision :: xmesh(10)
double precision, allocatable :: rejectionguess(:)
integer j
! Interface block - necessary to ensure arrays are passed

! in Fortran 90 style, rather than Fortran 77 -
! see http://w3.pppl.gov/~hammett/comp/f90_arrays.txt
interface
subroutine kiso(rp,L,MWd,rs,phi)
real :: rp,L(:),MWd(:),rs(:),phi(:)
end subroutine kiso
subroutine doneqn3(z,C,zX,CX,phi,K)
real :: z(:),C(:),zX,CX,phi(:),K(:)
end subroutine doneqn3
344
end interface
! Define global variables for the number of differential

! equations, NODE, the number of unknown parameters, NPAR,
! and the number of boundary conditions at the left end of
! the interval, LEFTBC. The number of boundary conditions
! at the right end is NODE+NPAR-LEFTBC.
! For this particular problem:
! NODE = NPAR = LEFTBC = RIGHTBC = (NARG-5)/7, which is the
! number of ions present
!INTEGER, PARAMETER :: NODE=2
!INTEGER, PARAMETER :: NPAR=NODE
!INTEGER, PARAMETER :: LEFTBC=NODE
!INTEGER, PARAMETER :: RIGHTBC=NODE
NODE = (NARG-5)/7
NPAR=NODE
LEFTBC=NODE
RIGHTBC=NODE
! Allocate vectors
allocate(L(NODE),MWd(NODE),rs(NODE),D(NODE),z(NODE),C(NODE),mf
(NODE),mm(NODE))
allocate(phi(NODE),phiout(NODE),Kd(NODE),lambda(NODE))
allocate(G(NODE),Kc(NODE),Dp(NODE),Donnaninletvector(NODE),Kin
(NODE))
allocate(inlet(NODE),Kout(NODE),Donnanoutletvector(NODE))
allocate(rejectionguess(NODE))
! Get input arguments from vector ARGSOUT

do j = 1,NODE
mf(j) = ARGSOUT(j)
mm(j) = ARGSOUT(1*NODE+j)
rs(j) = ARGSOUT(2*NODE+j)
D(j) = ARGSOUT(3*NODE+j)
z(j) = ARGSOUT(4*NODE+j)
L(j) = ARGSOUT(5*NODE+j)
MWd(j) = ARGSOUT(6*NODE+j)
end do
! First check if mass fractions are zero (first computional

step)
! If so, set rejection to zero and return
do j = 1,NODE
if (mf(j) .eq. 0.0) then
rejection = 0.0
return
end if
345
end do
Jv = ARGSOUT(size(ARGSOUT))
deltax = ARGSOUT(size(ARGSOUT)-1)
rp = ARGSOUT(size(ARGSOUT)-2)
CX = ARGSOUT(size(ARGSOUT)-3)
zX = ARGSOUT(size(ARGSOUT)-4)
! Calculate concentrations C from mass fractions mf and molar

masses mm
C = mf/mm*997.0*1000
! Use external subroutine kiso to calculate vector

! of steric partition coefficients phi and modify vector of
! Stokes radii rs as necessary
call kiso(rp,L,MWd,rs,phi)
! Calculate ratio of Stokes radius to pore radius, lambda

lambda = rs/rp
! Calculate hindrance factors for diffusion

Kd = 1-2.30*lambda+1.154*(lambda**2)+0.224*(lambda**3)
! Calculate hindrance factors for convection

G = 1+0.054*lambda-0.988*(lambda**2)+0.441*(lambda**3)
Kc = (2-phi)*G
! Calculate diffusivity in pore [m^2 s^-1]

Dp = D*Kd
! Calculate Donnan potential at inlet [V] which is constant

for all ions
! Use external subroutine doneqn3 to calculate vector of inlet
Donnan
! voltages
call doneqn3(z,C,zX,CX,phi,Kin)
Donnaninletvector = (R*T/(z*F))*log(1/Kin)
! Calculate concentrations just inside pore inlet

inlet = C*phi*exp(z*F/(R*T)*Donnaninletvector)
! Guess for rejections

rejectionguess = 0.5
rejection = rejectionguess
xmesh = BVP_LINSPACE(0.0D0,DBLE(deltax),10)
! BVP_INIT is used to form an initial guess for a mesh of 10

equally spaced
346
! points from zero to deltax (supplied in xmesh). The initial
guess is that
! of the inlet concentrations. A guess for unknown parameters
(rejectionguess)
! is supplied in an array as the last mandatory argument of
BVP_INIT.
SOL =
BVP_INIT(NODE,LEFTBC,xmesh,DBLE(inlet),rejectionguess,MAX_NUM_
SUBINTERVALS = 1000)
SOL = BVP_SOLVER(SOL,FSUB,BCSUB,TRACE=-1,TOL=1.0d-3,METHOD=4)
rejection = SOL%PARAMETERS
Donnanoutletvector = (R*T/(z*F))*log(1.0/Kout)
! Check zero current condition

Icheck = sum(Jv*C*(1-rejection)*z)
CALL BVP_TERMINATE(SOL)
! Dellocate vectors
deallocate(L,MWd,rs,D,z,C,mf,mm)
deallocate (phi,phiout,Kd,lambda)
deallocate (G,Kc,Dp,Donnaninletvector,Kin)
deallocate (inlet,Kout,Donnanoutletvector)
deallocate(rejectionguess)
contains
! System of ODEs (vector notation)

SUBROUTINE FSUB(X,Y,P,DYDX)
DOUBLE PRECISION :: X,Y(NODE),P(NPAR),DYDX(NODE)
dydx = Jv/Dp*(Kc*y-C*(1-P))-z*y*sum(Jv/Dp*z*(Kc*y-
C*(1-P)))/(sum(z*z*y));
END SUBROUTINE FSUB
! Boundary conditions for ODEs (vector notation)

SUBROUTINE BCSUB(YA,YB,P,BCA,BCB)
DOUBLE PRECISION :: YA(NODE),YB(NODE),P(NPAR),&
BCA(LEFTBC),BCB(RIGHTBC)
! First, return Donnan equilibrium coefficients at

outlet in Kout
call doneqn3(z,real(C*(1-P)),zX,CX,phi,Kout)
! Inlet boundary conditions
347
BCA = YA - inlet
! Outlet boundary conditions
BCB = YB - C*phi*(1.0D0-P)*exp(log(1.0D0/Kout))
END SUBROUTINE BCSUB
end subroutine np3nvector
348
Bibliography
[1] M.C. Porter, Handbook of industrial membrane technology, in, Noyes Publications,
New Jersey, 1990.
[2] M. Mulder, Basic principles of membrane technology, Kluwer Academic, Dordrecht,

Netherlands, 1991.
[3] A.I. Schäfer, Natural organics removal using membranes : principles, performance
and cost, Technomic Publishing, Lancaster, Pa., 2001.
[4] A.E. Childress, M. Elimelech, Relating Nanofiltration Membrane Performance to

Membrane Charge (Electrokinetic) Characteristics, Environ. Sci. Technol., 34 (2000)
3710-3716.
[5] E. Pellerin, E. Michelitsch, K. Darcovich, S. Lin, C.M. Tam, Turbulent transport in

membrane modules by CFD simulation in two dimensions, J. Membrane Sci., 100 (1995)
139-153.
[6] J.M. Benito, I.U. Richard, Modeling Concentration-Polarization in Reverse Osmosis

Spiral-Wound Elements, J. Environ. Eng., 122 (1996) 292-298.
[7] S. Kim, E.M.V. Hoek, Modeling concentration polarization in reverse osmosis

processes, Desalination, 186 (2005) 111-128.
[8] S.S. Sablani, M.F.A. Goosen, R. Al-Belushi, M. Wilf, Concentration polarization in

ultrafiltration and reverse osmosis: a critical review, Desalination, 141 (2001) 269-289.
[9] J.G. Wijmans, S. Nakao, J.W.A. Van Den Berg, F.R. Troelstra, C.A. Smolders,
Hydrodynamic resistance of concentration polarization boundary layers in ultrafiltration,
J. Membrane Sci., 22 (1985) 117-135.
[10] M. Al-Ahmad, F.A. Abdul Aleem, A. Mutiri, A. Ubaisy, Biofouling in RO

membrane systems Part 1: Fundamentals and control, Desalination, 132 (2000) 173-179.
[11] A. Al-Amoudi, R.W. Lovitt, Fouling strategies and the cleaning system of NF
membranes and factors affecting cleaning efficiency, J. Membrane Sci., 303 (2007) 4-28.
[12] S.n.F.E. Boerlage, Scaling and particulate fouling in membrane filtration systems, in,
Swets & Zeitlinger, Lisse, 2001.
349
[13] J.S. Baker, L.Y. Dudley, Biofouling in membrane systems -- A review, Desalination,
118 (1998) 81-89.
[14] T.V. Knyazkova, A.A. Maynarovich, Recognition of membrane fouling: testing of

theoretical approaches with data on NF of salt solutions containing a low molecular
weight surfactant as a foulant, Desalination, 126 (1999) 163-169.
[15] N. Hilal, O.O. Ogunbiyi, N.J. Miles, R. Nigmatullin, Methods Employed for Control
of Fouling in MF and UF Membranes: A Comprehensive Review, Separation Science and
Technology, 40 (2005) 1957 - 2005.
[16] W. Mohammad, L. Pei, A. Kadhum, Characterization and identification of rejection

mechanisms in nanofiltration membranes using extended Nernst–Planck model, Clean
Technologies and Environmental Policy, 4 (2002) 151-156.
[17] W.R. Bowen, H. Mukhtar, Characterisation and prediction of separation performance

of nanofiltration membranes, J. Membrane Sci., 112 (1996) 263-274.
[18] S. Bandini, D. Vezzani, Nanofiltration modeling: the role of dielectric exclusion in

membrane characterization, Chemical Engineering Science, 58 (2003) 3303-3326.
[19] A. Szymczyk, P. Fievet, Investigating transport properties of nanofiltration

membranes by means of a steric, electric and dielectric exclusion model, J. Membrane
Sci., 252 (2005) 77-88.
[20] L. Heng, Y. Yanling, G. Weijia, L. Xing, L. Guibai, Effect of pretreatment by

permanganate/chlorine on algae fouling control for ultrafiltration (UF) membrane system,
Desalination, 222 (2008) 74-80.
[21] A. Maartens, P. Swart, E.P. Jacobs, Feed-water pretreatment: methods to reduce

membrane fouling by natural organic matter, J. Membrane Sci., 163 (1999) 51-62.
[22] Y. Chen, B.Z. Dong, N.Y. Gao, J.C. Fan, Effect of coagulation pretreatment on
fouling of an ultrafiltration membrane, Desalination, 204 (2007) 181-188.
[23] W. Ma, Y. Zhao, L. Wang, The pretreatment with enhanced coagulation and a UF
membrane for seawater desalination with reverse osmosis, Desalination, 203 (2007) 256-
259.
[24] R. Ghosh, Z.F. Cui, Fractionation of BSA and lysozyme using ultrafiltration: effect
of pH and membrane pretreatment, J. Membrane Sci., 139 (1998) 17-28.
350
[25] L. Masse, D.I. Massé, Y. Pellerin, The effect of pH on the separation of manure
nutrients with reverse osmosis membranes, J. Membrane Sci., 325 (2008) 914-919.
[26] M.M.T. Khan, W. Jones, A. Camper, S. Takizawa, H. Katayama, F. Kurisu, S.

Ohgaki, Powdered activated carbon and biofiltration improve MF performance: Part I,
Membrane Technology, 2007 (2007) 7-10.
[27] H.K. Shon, S. Vigneswaran, J. Cho, Comparison of physico-chemical pretreatment

methods to seawater reverse osmosis: Detailed analyses of molecular weight distribution
of organic matter in initial stage, J. Membrane Sci., 320 (2008) 151-158.
[28] M. Gryta, Chemical pretreatment of feed water for membrane distillation, Chemical
Papers, 62 (2008) 100-105.
[29] S.L. Kim, J. Paul Chen, Y.P. Ting, Study on feed pretreatment for membrane
filtration of secondary effluent, Separation and Purification Technology, 29 (2002) 171-
179.
[30] B. Van der Bruggen, D. Segers, C. Vandecasteele, L. Braeken, A. Volodin, C. Van

Haesendonck, How a Microfiltration Pretreatment Affects the Performance in
Nanofiltration, Separation Science and Technology, 39 (2005) 1443 - 1459.
[31] C.K. Teng, M.N.A. Hawlader, A. Malek, An experiment with different pretreatment
methods, Desalination, 156 (2003) 51-58.
[32] J. Xu, G. Ruan, X. Chu, Y. Yao, B. Su, C. Gao, A pilot study of UF pretreatment
without any chemicals for SWRO desalination in China, Desalination, 207 (2007) 216-
226.
[33] G.K. Pearce, The case for UF/MF pretreatment to RO in seawater applications,
Desalination, 203 (2007) 286-295.
[34] H. Ma, C.N. Bowman, R.H. Davis, Membrane fouling reduction by backpulsing and
surface modification, J. Membrane Sci., 173 (2000) 191-200.
[35] R. Ghidossi, D. Veyret, P. Moulin, Computational fluid dynamics applied to

membranes: State of the art and opportunities, Chemical Engineering and Processing, 45
(2006) 437-454.
[36] J. Schwinge, D.E. Wiley, A.G. Fane, Novel spacer design improves observed flux, J.
Membrane Sci., 229 (2004) 53-61.
351
[37] V. Geraldes, V. Semião, M.N. de Pinho, The effect of the ladder-type spacers
configuration in NF spiral-wound modules on the concentration boundary layers
disruption, Desalination, 146 (2002) 187-194.
[38] V. Geraldes, V. Semião, M.N. de Pinho, Flow management in nanofiltration spiral

wound modules with ladder-type spacers, J. Membrane Sci., 203 (2002) 87-102.
[39] V. Geraldes, V. Semião, M. Norberta Pinho, Hydrodynamics and concentration

polarization in NF/RO spiral-wound modules with ladder-type spacers, Desalination, 157
(2003) 395-402.
[40] F. Li, G.W. Meindersma, A.B. de Haan, T. Reith, Optimization of non-woven

spacers by CFD and validation by experiments, Desalination, 146 (2002) 209-212.
[41] F. Li, W. Meindersma, A.B. de Haan, T. Reith, Experimental validation of CFD

mass transfer simulations in flat channels with non-woven net spacers, J. Membrane Sci.,
232 (2004) 19-30.
[42] C.P. Koutsou, S.G. Yiantsios, A.J. Karabelas, Numerical simulation of the flow in a
plane-channel containing a periodic array of cylindrical turbulence promoters, J.
Membrane Sci., 231 (2004) 81-90.
[43] V.V. Ranade, A. Kumar, Comparison of flow structures in spacer-filled flat and
annular channels, Desalination, 191 (2006) 236-244.
[44] J. Schwinge, D.E. Wiley, D.F. Fletcher, A CFD study of unsteady flow in narrow
spacer-filled channels for spiral-wound membrane modules, Desalination, 146 (2002)
195-201.
[45] M. Shakaib, S.M.F. Hasani, M. Mahmood, Study on the effects of spacer geometry
in membrane feed channels using three-dimensional computational flow modeling, J.
Membrane Sci., 297 (2007) 74-89.
[46] D.E. Wiley, D.F. Fletcher, Computational fluid dynamics modelling of flow and
permeation for pressure-driven membrane processes, Desalination, 145 (2002) 183-186.
[47] Y.-L. Li, K.-L. Tung, CFD simulation of fluid flow through spacer-filled membrane
module: selecting suitable cell types for periodic boundary conditions, Desalination, 233
(2008) 351-358.
[48] C. Cabassud, S. Laborie, J.M. Lainé, How slug flow can improve ultrafiltration flux
in organic hollow fibres, J. Membrane Sci., 128 (1997) 93-101.
352
[49] S.R. Bellara, Z.F. Cui, D.S. Pepper, Gas sparging to enhance permeate flux in
ultrafiltration using hollow fibre membranes, J. Membrane Sci., 121 (1996) 175-184.
[50] Z.F. Cui, K.I.T. Wright, Flux enhancements with gas sparging in downwards
crossflow ultrafiltration: performance and mechanism, J. Membrane Sci., 117 (1996) 109-
116.
[51] T. Taha, Z.F. Cui, CFD modelling of gas-sparged ultrafiltration in tubular

membranes, J. Membrane Sci., 210 (2002) 13-27.
[52] T. Taha, Z.F. Cui, Hydrodynamic analysis of upward slug flow in tubular
membranes, Desalination, 145 (2002) 179-182.
[53] Q.Y. Li, Z.F. Cui, D.S. Pepper, Effect of bubble size and frequency on the permeate
flux of gas sparged ultrafiltration with tubular membranes, Chemical Engineering
Journal, 67 (1997) 71-75.
[54] R. Ghosh, Z.F. Cui, Mass transfer in gas-sparged ultrafiltration: upward slug flow in
tubular membranes, J. Membrane Sci., 162 (1999) 91-102.
[55] S. Smith, T. Taha, Z. Cui, Enhancing hollow fibre ultrafiltration using slug-flow -- a
hydrodynamic study, Desalination, 146 (2002) 69-74.
[56] S.R. Smith, R.W. Field, Z.F. Cui, Predicting the performance of gas-sparged and
non-gas-sparged ultrafiltration, Desalination, 191 (2006) 376-385.
[57] B.J. Bellhouse, G. Costigan, K. Abhinava, A. Merry, The performance of helical

screw-thread inserts in tubular membranes, Separation and Purification Technology, 22-
23 (2001) 89-113.
[58] M.E. Brewster, K.-Y. Chung, G. Belfort, Dean vortices with wall flux in a curved
channel membrane system : 1. A new approach to membrane module design, J.
Membrane Sci., 81 (1993) 127-137.
[59] M. Bubolz, M. Wille, G. Langer, U. Werner, The use of dean vortices for crossflow
microfiltration: basic principles and further investigation, Separation and Purification
Technology, 26 (2002) 81-89.
[60] R. Moll, P. Moulin, D. Veyret, F. Charbit, Numerical simulation of Dean vortices:

fluid trajectories, J. Membrane Sci., 197 (2002) 157-172.
353
[61] P. Moulin, D. Veyret, F. Charbit, Dean vortices: comparison of numerical simulation
of shear stress and improvement of mass transfer in membrane processes at low
permeation fluxes, J. Membrane Sci., 183 (2001) 149-162.
[62] W. Wang, C.D. Bertram, D.E. Wiley, Effects of collapsible-tube-induced pulsation

vigour on membrane filtration performance, J. Membrane Sci., 288 (2007) 298-306.
[63] P. Czekaj, W. Mores, R.H. Davis, C. Güell, Infrasonic pulsing for foulant removal in
crossflow microfiltration, J. Membrane Sci., 180 (2000) 157-169.
[64] E. Spiazzi, J. Lenoir, A. Grangeon, A new generator of unsteady-state flow regime in

tubular membranes as an anti-fouling technique: A hydrodynamic approach, J. Membrane
Sci., 80 (1993) 49-57.
[65] P. Blanpain-Avet, N. Doubrovine, C. Lafforgue, M. Lalande, The effect of

oscillatory flow on crossflow microfiltration of beer in a tubular mineral membrane
system - Membrane fouling resistance decrease and energetic considerations, J.
Membrane Sci., 152 (1999) 151-174.
[66] H.-y. Li, Bertram, C.D., Wiley, D.E., Mechanisms by which pulsatile flow affects
cross-flow microfiltration, AIChE Journal, 44 (1998) 1950-1961.
[67] H.M. Kyllönen, P. Pirkonen, M. Nyström, Membrane filtration enhanced by

ultrasound: a review, Desalination, 181 (2005) 319-335.
[68] M.O. Lamminen, H.W. Walker, L.K. Weavers, Mechanisms and factors influencing
the ultrasonic cleaning of particle-fouled ceramic membranes, J. Membrane Sci., 237
(2004) 213-223.
[69] M.O. Lamminen, H.W. Walker, L.K. Weavers, Cleaning of particle-fouled

membranes during cross-flow filtration using an embedded ultrasonic transducer system,
J. Membrane Sci., 283 (2006) 225-232.
[70] X. Chai, T. Kobayashi, N. Fujii, Ultrasound-associated cleaning of polymeric

membranes for water treatment, Separation and Purification Technology, 15 (1999) 139-
146.
[71] S. Muthukumaran, S. Kentish, S. Lalchandani, M. Ashokkumar, R. Mawson, G.W.

Stevens, F. Grieser, The optimisation of ultrasonic cleaning procedures for dairy fouled
ultrafiltration membranes, Ultrasonics Sonochemistry, 12 (2005) 29-35.
[72] X. Chai, T. Kobayashi, N. Fujii, Ultrasound effect on cross-flow filtration of

polyacrylonitrile ultrafiltration membranes, J. Membrane Sci., 148 (1998) 129-135.
354
[73] T. Kobayashi, X. Chai, N. Fujii, Ultrasound enhanced cross-flow membrane
filtration, Separation and Purification Technology, 17 (1999) 31-40.
[74] I. Masselin, X. Chasseray, L. Durand-Bourlier, J.-M. Lainé, P.-Y. Syzaret, D.

Lemordant, Effect of sonication on polymeric membranes, J. Membrane Sci., 181 (2001)
213-220.
[75] R.-S. Juang, K.-H. Lin, Flux recovery in the ultrafiltration of suspended solutions
with ultrasound, J. Membrane Sci., 243 (2004) 115-124.
[76] E. Zondervan, B. Roffel, Evaluation of different cleaning agents used for cleaning
ultra filtration membranes fouled by surface water, J. Membrane Sci., 304 (2007) 40-49.
[77] B.R. Munson, D.F. Young, T.H. Okiishi, Fundamentals of fluid mechanics, 4th ed.
ed., Wiley, New York, 2002.
[78] M.B. Abbott, D.R. Basco, Computational fluid dynamics : an introduction for
engineers, Longman Scientific & Technical, Harlow, 1989.
[79] T.J. Chung, Computational fluid dynamics, Cambridge University Press, Cambridge,
U.K., 2002.
[80] J.H. Ferziger, M. Peric, Computational methods for fluid dynamics, Springer-Verlag,
New York, 1996.
[81] A.S. Berman, Laminar Flow in Channels with Porous Walls, J. Appl. Phys., 24
(1953) 1232-1235.
[82] M. Friedman, J. Gillis, Viscous flow in a pipe with absorbing walls, J. Appl. Mech.,
34 (1967) pp. 819–827.
[83] T. Mizushina, S. Takeshita, G. Unno, Study of flow in a porous tube with radial
mass flux, J. Chem. Eng. , 4 (1971) pp. 135–142.
[84] L.S. Galowin, M.J. De Santis, Investigation of laminar flow in a porous pipe with
variable wall suction, AIAA, 12 (1974) pp. 1585–1594.
[85] S.K. Karode, Laminar flow in channels with porous walls, revisited, J. Membrane
Sci., 191 (2001) 237-241.
[86] V. Nassehi, Modelling of combined Navier-Stokes and Darcy flows in crossflow

membrane filtration, Chemical Engineering Science, 53 (1998) 1253-1265.
355
[87] K. Damak, A. Ayadi, B. Zeghmati, P. Schmitz, A new Navier-Stokes and Darcy's
law combined model for fluid flow in crossflow filtration tubular membranes,
Desalination, 161 (2004) 67-77.
[88] P. Schmitz, M. Prat, 3-D Laminar stationary flow over a porous surface with suction:
Description at pore level, AIChE Journal, 41 (1995) 2212-2226.
[89] C. Gouverneur, Thèse de Doctorat, in, Institut National Polytechnique de Toulouse,

1991.
[90] L. Huang, M.T. Morrissey, Finite element analysis as a tool for crossflow membrane
filter simulation, J. Membrane Sci., 155 (1999) 19-30.
[91] C.J. Richardson, V. Nassehi, Finite element modelling of concentration profiles in

flow domains with curved porous boundaries, Chemical Engineering Science, 58 (2003)
2491-2503.
[92] S. Geissler, U. Werner, Dynamic model of crossflow microfiltration in flat-channel

systems under laminar flow conditions, Filtration & Separation, 32 (1995) 533-537.
[93] Y. Lee, M.M. Clark, Modeling of flux decline during crossflow ultrafiltration of
colloidal suspensions, J. Membrane Sci., 149 (1998) 181-202.
[94] V. Geraldes, V. Semião, M.N. de Pinho, Flow and mass transfer modelling of
nanofiltration, J. Membrane Sci., 191 (2001) 109-128.
[95] M. Niwa, H. Ohya, E. Kuwahara, Y. Negishi, Reverse osmotic concentration of

aqueous 2-butanone (methyl ethyl ketone), tetrahydrofuran and ethyl acetate solutions,
Journal of Chemical Engineering of Japan, 21 (1988) 164-171.
[96] S. Tanimura, S.-I. Nakao, S. Kimura, Transport analysis of reverse osmosis of

organic aqueous solutions, Journal of Chemical Engineering of Japan, 24 (1991) 364-371.
[97] C. Rosén, C. Trägårdh, Computer simulations of mass transfer in the concentration

boundary layer over ultrafiltration membranes, J. Membrane Sci., 85 (1993) 139-156.
[98] D.F. Fletcher, D.E. Wiley, A computational fluids dynamics study of buoyancy
effects in reverse osmosis, J. Membrane Sci., 245 (2004) 175-181.
[99] J.M. Miranda, J.B.L.M. Campos, Concentration polarization in a membrane placed

under an impinging jet confined by a conical wall -- a numerical approach, J. Membrane
Sci., 182 (2001) 257-270.
356
[100] G. Belfort, Fluid mechanics in membrane filtration: Recent developments, J.
Membrane Sci., 40 (1989) 123-147.
[101] V.V. Ranade, A. Kumar, Fluid dynamics of spacer filled rectangular and curvilinear
channels, J. Membrane Sci., 271 (2006) 1-15.
[102] G. Belfort, N. Nagata, Fluid mechanics and cross-flow filtration: some thoughts,
Desalination, 53 (1985) 57-79.
[103] G.A. Fimbres-Weihs, D.E. Wiley, Review of 3D CFD modeling of flow and mass
transfer in narrow spacer-filled channels in membrane modules, Chemical Engineering
and Processing: Process Intensification, 49 (2010) 759-781.
[104] B.K. Gupta, E.K. Levy, Symmetrical laminar channel flow with wall suction,
American Society of Mechanical Engineers, (1975).
[105] A.R. Da Costa, A.G. Fane, C.J.D. Fell, A.C.M. Franken, Optimal channel spacer
design for ultrafiltration, J. Membrane Sci., 62 (1991) 275-291.
[106] A.R. Da Costa, A.G. Fane, D.E. Wiley, Spacer characterization and pressure drop
modelling in spacer-filled channels for ultrafiltration, J. Membrane Sci., 87 (1994) 79-98.
[107] Z. Cao, D.E. Wiley, A.G. Fane, CFD simulations of net-type turbulence promoters
in a narrow channel, J. Membrane Sci., 185 (2001) 157-176.
[108] G. Sivashinsky, V. Yakhot, Negative viscosity effect in large-scale flows, Physics

of Fluids, 28 (1985) 1040-1042.
[109] Y. Miyake, K. Tsujimoto, H. Beppu, Direct numerical simulation of a turbulent

flow in a channel having periodic pressure gradient, International Journal of Heat and
Fluid Flow, 16 (1995) 333-340.
[110] J. Kim, P. Moin, R. Moser, Turbulence statistics in fully developed channel flow at
low Reynolds number, Journal of Fluid Mechanics Digital Archive, 177 (1987) 133-166.
[111] J. Schwinge, D.E. Wiley, D.F. Fletcher, Simulation of the Flow around Spacer
Filaments between Narrow Channel Walls. 1. Hydrodynamics, Industrial & Engineering
Chemistry Research, 41 (2002) 2977-2987.
[112] J. Schwinge, D.E. Wiley, D.F. Fletcher, Simulation of the Flow around Spacer
Filaments between Channel Walls. 2. Mass-Transfer Enhancement, Industrial &
Engineering Chemistry Research, 41 (2002) 4879-4888.
357
[113] G.A. Fimbres-Weihs, D.E. Wiley, Numerical study of two-dimensional multi-layer
spacer designs for minimum drag and maximum mass transfer, J. Membrane Sci., 325
(2008) 809-822.
[114] G.A. Fimbres-Weihs, D.E. Wiley, D.F. Fletcher, Unsteady Flows with Mass
Transfer in Narrow Zigzag Spacer-)LOOHG &KDQQHOVௗ $ 1XPHULFDO 6WXG\ ,QGXVWULDO
Engineering Chemistry Research, 45 (2006) 6594-6603.
[115] S. Wardeh, H.P. Morvan, CFD simulations of flow and concentration polarization
in spacer-filled channels for application to water desalination, Chemical Engineering
Research and Design, 86 (2008) 1107-1116.
[116] S.K. Karode, A. Kumar, Flow visualization through spacer filled channels by
computational fluid dynamics I.: Pressure drop and shear rate calculations for flat sheet
geometry, J. Membrane Sci., 193 (2001) 69-84.
[117] C.P. Koutsou, S.G. Yiantsios, A.J. Karabelas, Direct numerical simulation of flow
in spacer-filled channels: Effect of spacer geometrical characteristics, J. Membrane Sci.,
291 (2007) 53-69.
[118] G.A. Fimbres-Weihs, D.E. Wiley, Numerical study of mass transfer in three-
dimensional spacer-filled narrow channels with steady flow, J. Membrane Sci., 306
(2007) 228-243.
[119] C.P. Koutsou, S.G. Yiantsios, A.J. Karabelas, A numerical and experimental study
of mass transfer in spacer-filled channels: Effects of spacer geometrical characteristics
and Schmidt number, J. Membrane Sci., 326 (2009) 234-251.
[120] F. Li, W. Meindersma, A.B. de Haan, T. Reith, Novel spacers for mass transfer
enhancement in membrane separations, J. Membrane Sci., 253 (2005) 1-12.
[121] J.L.C. Santos, V. Geraldes, S. Velizarov, J.G. Crespo, Investigation of flow patterns
and mass transfer in membrane module channels filled with flow-aligned spacers using
computational fluid dynamics (CFD), J. Membrane Sci., 305 (2007) 103-117.
[122] K.K. Lau, M.Z. Abu Bakar, A.L. Ahmad, T. Murugesan, Feed spacer mesh angle:
3D modeling, simulation and optimization based on unsteady hydrodynamic in spiral
wound membrane channel, J. Membrane Sci., 343 (2009) 16-33.
[123] ANSYS CFX-Solver Theory Guide, ANSYS, Inc., Canonsburg, PA, 2010.
[124] D. Bhattacharyya, S.L. Back, R.I. Kermode, M.C. Roco, Prediction of

concentration polarization and flux behavior in reverse osmosis by numerical analysis, J.
Membrane Sci., 48 (1990) 231-262.
358
[125] R. Krishna, J.A. Wesselingh, The Maxwell-Stefan approach to mass transfer,
Chemical Engineering Science, 52 (1997) 861-911.
[126] L. Onsager, THEORIES AND PROBLEMS OF LIQUID DIFFUSION, Annals of

the New York Academy of Sciences, 46 (1945) 241-265.
[127] S. Déon, P. Dutournié, L. Limousy, P. Bourseau, The two-dimensional pore and

polarization transport model to describe mixtures separation by nanofiltration: Model
validation, AIChE Journal, 57 (2011) 985-995.
[128] S. Bhattacharjee, J.C. Chen, M. Elimelech, Coupled model of concentration

polarization and pore transport in crossflow nanofiltration, AIChE Journal, 47 (2001)
2733-2745.
[129] J. Kucera, Reverse osmosis : design, processes, and applications for engineers,
Scrivener Pub. ;
Wiley, Salem, Mass.
Hoboken, N.J., 2010.
[130] P. Schausberger, N. Norazman, H. Li, V. Chen, A. Friedl, Simulation of protein

ultrafiltration using CFD: Comparison of concentration polarisation and fouling effects
with filtration and protein adsorption experiments, J. Membrane Sci., 337 (2009) 1-8.
[131] J.S. Vrouwenvelder, C. Picioreanu, J.C. Kruithof, M.C.M. van Loosdrecht,

Biofouling in spiral wound membrane systems: Three-dimensional CFD model based
evaluation of experimental data, J. Membrane Sci., 346 (2010) 71-85.
[132] M. Shakaib, S.M.F. Hasani, M. Mahmood, CFD modeling for flow and mass
transfer in spacer-obstructed membrane feed channels, J. Membrane Sci., 326 (2009) 270-
284.
[133] C. Picioreanu, J.S. Vrouwenvelder, M.C.M. van Loosdrecht, Three-dimensional

modeling of biofouling and fluid dynamics in feed spacer channels of membrane devices,
J. Membrane Sci., 345 (2009) 340-354.
[134] Y.-L. Li, K.-L. Tung, M.-Y. Lu, S.-H. Huang, Mitigating the curvature effect of the
spacer-filled channel in a spiral-wound membrane module, J. Membrane Sci., 329 (2009)
106-118.
[135] Y.-L. Li, K.-L. Tung, The effect of curvature of a spacer-filled channel on fluid
flow in spiral-wound membrane modules, J. Membrane Sci., 319 (2008) 286-297.
359
[136] S. Ma, L. Song, S.L. Ong, W.J. Ng, A 2-D streamline upwind Petrov/Galerkin
finite element model for concentration polarization in spiral wound reverse osmosis
modules, J. Membrane Sci., 244 (2004) 129-139.
[137] J. Schwinge, D.E. Wiley, D.F. Fletcher, Simulation of Unsteady Flow and Vortex
Shedding for Narrow Spacer-Filled Channels, Industrial & Engineering Chemistry
Research, 42 (2003) 4962-4977.
[138] K.S. Spiegler, O. Kedem, Thermodynamics of hyperfiltration (reverse osmosis):

criteria for efficient membranes, Desalination, 1 (1966) 311-326.
[139] O. Kedem, A. Katchalsky, Permeability of composite membranes. Part 1.-Electric

current, volume flow and flow of solute through membranes, Transactions of the Faraday
Society, 59 (1963) 1918-1930.
[140] B. Van der Bruggen, C. Vandecasteele, Modelling of the retention of uncharged

molecules with nanofiltration, Water Research, 36 (2002) 1360-1368.
[141] W.R. Bowen, A.W. Mohammad, Diafiltration by nanofiltration: Prediction and

optimization, AIChE Journal, 44 (1998) 1799-1812.
[142] J.M.M. Peeters, Characterization of nanofiltration membranes, in, Enschede, 1997,

pp. 159.
[143] J. Palmeri, P. Blanc, A. Larbot, P. David, Theory of pressure-driven transport of

neutral solutes and ions in porous ceramic nanofiltration membranes, J. Membrane Sci.,
160 (1999) 141-170.
[144] J. Schaep, C. Vandecasteele, A. Mohammad, W. Bowen, Analysis of the Salt

Retention of Nanofiltration Membranes Using the Donnan–Steric Partitioning Pore
Model, Separation Science & Technology, 34 (1999) 3009.
[145] J.M.M. Peeters, J.P. Boom, M.H.V. Mulder, H. Strathmann, Retention

measurements of nanofiltration membranes with electrolyte solutions, J. Membrane Sci.,
145 (1998) 199-209.
[146] C. Diaper, V. Correia, S. Judd, Characterisation of zirconium/poly(acrylic acid) low

pressure dynamically formed membranes by use of the extended Nernst-Planck equation,
J. Membrane Sci., 138 (1998) 135-140.
[147] X.-L. Wang, T. Tsuru, S.-i. Nakao, S. Kimura, The electrostatic and steric-
hindrance model for the transport of charged solutes through nanofiltration membranes, J.
Membrane Sci., 135 (1997) 19-32.
360
[148] W.R. Bowen, A.W. Mohammad, N. Hilal, Characterisation of nanofiltration
membranes for predictive purposes -- use of salts, uncharged solutes and atomic force
microscopy, J. Membrane Sci., 126 (1997) 91-105.
[149] W.R. Bowen, A.W. Mohammad, A theoretical basis for specifying nanofiltration
membranes -- Dye/salt/water streams, Desalination, 117 (1998) 257-264.
[150] W.R. Bowen, J.S. Welfoot, Modelling of membrane nanofiltration--pore size

distribution effects, Chemical Engineering Science, 57 (2002) 1393-1407.
[151] D. Vezzani, S. Bandini, Donnan equilibrium and dielectric exclusion for

characterization of nanofiltration membranes, Desalination, 149 (2002) 477-483.
[152] C. Labbez, P. Fievet, A. Szymczyk, A. Vidonne, A. Foissy, J. Pagetti, Retention of

mineral salts by a polyamide nanofiltration membrane, Separation and Purification
Technology, 30 (2003) 47-55.
[153] J. Schaep, C. Vandecasteele, A. Wahab Mohammad, W. Richard Bowen,

Modelling the retention of ionic components for different nanofiltration membranes,
Separation and Purification Technology, 22–23 (2001) 169-179.
[154] M.S. Hall, V.M. Starov, D.R. Lloyd, Reverse osmosis of multicomponent
electrolyte solutions Part I. Theoretical development, J. Membrane Sci., 128 (1997) 23-
37.
[155] W.M. Deen, Hindered transport of large molecules in liquid-filled pores, AIChE
Journal, 33 (1987) 1409-1425.
[156] B. Van der Bruggen, J. Schaep, W. Maes, D. Wilms, C. Vandecasteele,

Nanofiltration as a treatment method for the removal of pesticides from ground waters,
Desalination, 117 (1998) 139-147.
[157] B. Van der Bruggen, J. Schaep, D. Wilms, C. Vandecasteele, Influence of

molecular size, polarity and charge on the retention of organic molecules by
nanofiltration, J. Membrane Sci., 156 (1999) 29-41.
[158] Y. Kiso, T. Kon, T. Kitao, K. Nishimura, Rejection properties of alkyl phthalates

with nanofiltration membranes, J. Membrane Sci., 182 (2001) 205-214.
[159] Y. Kiso, K. Muroshige, T. Oguchi, M. Hirose, T. Ohara, T. Shintani, Pore radius

estimation based on organic solute molecular shape and effects of pressure on pore radius
for a reverse osmosis membrane, J. Membrane Sci., 369 (2011) 290-298.
361
[160] Y. Kiso, K. Muroshige, T. Oguchi, T. Yamada, M. Hhirose, T. Ohara, T. Shintani,
Effect of molecular shape on rejection of uncharged organic compounds by nanofiltration
membranes and on calculated pore radii, J. Membrane Sci., 358 (2010) 101-113.
[161] J.L.C. Santos, P. de Beukelaar, I.F.J. Vankelecom, S. Velizarov, J.G. Crespo, Effect
of solute geometry and orientation on the rejection of uncharged compounds by
nanofiltration, Separation and Purification Technology, 50 (2006) 122-131.
[162] M.J.D. Powell, A Fortran Subroutine for Solving Systems of Nonlinear Algebraic
Equations, in: P. Rabinowitz (Ed.) Numerical methods for nonlinear algebraic equations,
Gordon and Breach Science Publishers, 1970.
[163] L.F. Shampine, P.H. Muir, H. Xu, A User-Friendly Fortran BVP Solver, Promacon,
Athens, GRECE, 2006.
[164] L.F. Shampine, M.W. Reichelt, J. Kierzenka, Solving Boundary Value Problems
for Ordinary Differential Equations in MATLAB with bvp4c, in, 2000.
[165] MATLAB Release Notes, The MathWorks, Inc., 2009.
[166] J. Burkardt, Source Codes in Fortran 90, in.
[167] J. Stewart, Optimization of parameters for semiempirical methods V: Modification

of NDDO approximations and application to 70 elements, Journal of Molecular
Modeling, 13 (2007) 1173-1213.
[168] M.N. de Pinho, V. Semião, V. Geraldes, Integrated modeling of transport processes

in fluid/nanofiltration membrane systems, J. Membrane Sci., 206 (2002) 189-200.
362

Coupled Modelling of Hydrodynamics and Mass Transfer in Membrane Filtration

Cargado por

Información del documento

Derechos de autor

Formatos disponibles

Compartir este documento

Compartir o incrustar documentos

Opciones para compartir

¿Le pareció útil este documento?

¿Este contenido es inapropiado?

Copyright:

Formatos disponibles

Coupled Modelling of Hydrodynamics and Mass Transfer in Membrane Filtration

Cargado por

Copyright:

Formatos disponibles

Coupled modelling of hydrodynamics and mass transfer in

Gregory Peter Keir

Submitted in fulfilment of the requirements for the degree of

membrane filtration submitted for the degree of Doctor of Philosophy.

Full Name Gregory Peter Keir

or diploma by any university or institution is identified in the text.

Full Name Gregory Peter Keir

The high-pressure membrane filtration processes of nanofiltration (NF) and

reverse osmosis (RO) are widely used in industry to remove dissolved

treatment. Despite their ubiquity, it is often difficult to predict the performance of

microstructure of the membrane, which can be complex. This can be especially

apparent when multiple dissolved components are in solution.

This thesis outlines the development of a computational model to predict the

behaviour of such systems by combining a rigorous hydrodynamic model using

commercially available computational fluid dynamics (CFD) software with a

microscopic mass transfer model including the effects of convection, diffusion,

and electric migration to predict the removal of dissolved components by the

membrane. These fields of modelling in membrane applications have traditionally

applicable general predictive model of the NF and RO processes.

of hydrodynamic models for arbitrary two-dimensional (2D) or three-dimensional

as density, viscosity, solute diffusivity, and osmotic pressure to vary spatially

membrane surface, a commonly observed phenomenon which reduces the

effectiveness of NF and RO processes. In addition, the formulation presented here

spatially depending on the local hydrodynamic conditions.

an arbitrary combination of multiple charged and uncharged solutes. The

rejection of solutes to vary spatially throughout the membrane channel.

In summary, the novel research elements of this thesis are:

x Coupling of a commercial CFD code to a numerical solution of a refined

version of the DSPM to predict both membrane channel hydrodynamics

and mass transfer through NF and RO membranes;

into a CFD model of a membrane channel;

x Refinement of the classic DSPM to include molecular shape effects for

both charged and uncharged solutes; and

x Presentation of a numerical method of solution for the refined DSPM for

an arbitrary combination of multiple charged and uncharged solutes.

some interesting phenomena predicted by the coupled model, including significant

spatial variation in predicted solute rejection, presence of an inverse concentration

polarisation layer for negatively rejected solutes, variation in near-wall

concentration gradients by solute, and significant variability in overall permeate

flux depending on solution composition. Finally, comments are made regarding

future research in extending the applicability of the model as computational

resources become more abundant.

x Deakin University for providing a Deakin University Postgraduate

in the School of Engineering at James Cook University;

x The Graduate Research School at James Cook University for providing a

James Cook University Postgraduate Research Scholarship; and

x Dimuth Navaratna and my other fellow postgraduate students both at

Deakin University and James Cook University.

I must extend a special thank you to my thesis supervisor A/Prof. Jega

family, for their unconditional love and support.

Table of contents ..................................................................................................... v

List of figures .......................................................................................................... x

List of tables ....................................................................................................... xxvi

List of publications and presentations ................................................................ xxix

1.1 Thesis outline ................................................................................................. 2

Chapter 2 Membrane filtration and the problem of flux decline ............................. 5

2.1 Membrane processes...................................................................................... 5

2.1.1 Membrane performance ........................................................................ 11

2.2 Membrane flux decline ................................................................................ 12

2.2.1 Concentration polarisation .................................................................... 13

2.2.2 Membrane fouling ................................................................................. 16

2.3 Methods to reduce flux decline.................................................................... 25

2.3.1 Pretreatment of feed solution ................................................................ 25

2.3.2 Modification of membrane properties................................................... 26