Sie sind auf Seite 1von 218

Computational Modelling of Gas-Liquid

Flow in Stirred Tanks

A Thesis Submitted for the Degree of


Doctor of Philosophy
by
Graeme Leslie Lane
BE (Chem, Hons)

The University of Newcastle

Submitted November 2005


Revised submission August 2006

I hereby certify that the work embodied in this thesis is the result of original research
and has not been submitted for a higher degree to any other University or Institution.
(Signed) _________________________________________

ACKNOWLEDGEMENTS
I would like to express my gratitude to my supervisor, Professor Geoffrey Evans, for his
guidance and advice during the course of this project. I would also like to thank
Dr Phillip Schwarz (CSIRO), Dr Peter Witt (CSIRO) and Dr Greg Rigby (formerly at
the University of Newcastle) for their valuable assistance with various aspects of the
project. I would also like to thank my employer, CSIRO Minerals, whose sponsorship
has made this thesis possible.

ABSTRACT
This thesis describes a study in which the aim was to develop an improved method for
computational fluid dynamics (CFD) modelling of gas-liquid flow in mechanicallystirred tanks. Stirred tanks are commonly used in the process industries for carrying out
a wide range of mixing operations and chemical reactions, yet considerable
uncertainties remain in design and scale-up procedures. Computational modelling is of
interest since it may assist in investigating the detailed flow characteristics of stirred
tanks. However, as shown by a review of the literature, a range of limitations have been
evident in previously published modelling methods.

In the development of the modelling method, single-phase liquid flow was firstly
considered, as a basis for extension to multiphase flow. A finite volume method was
used to solve the equations for conservation of mass and momentum, in conjunction
with the k- turbulence model. Simulation results were compared with experimental
measurements for tanks stirred by a Rushton turbine and by a Lightnin A315 impeller.
Comparison was made between different methods which account for impeller motion.
Accuracy was assessed in terms of the prediction of velocities, power and flow
numbers, the presence of trailing vortices, pressures around the impeller, and the
turbulent kinetic energy and dissipation rate. The effect of grid density was investigated.

For gas dispersion in a liquid, the modelling method employed the Eulerian-Eulerian
two-fluid equations, again in conjunction with the k- turbulence model. The correct
specification of the equations was firstly reviewed. Different forms of the turbulent
dispersion force were compared. For the drag force, it was found that existing
correlations did not properly account for the effect of turbulence in increasing the
bubble drag coefficient. By analysing literature data, a new equation was proposed to
account for this increase in drag. For the prediction of bubble size, a bubble number
density equation was introduced, which takes into account the effects of break-up and
coalescence. The modelling method also allows for gas cavity formation behind
impeller blades.

Simulations of gas-liquid flow were again carried out for tanks stirred by a Rushton
turbine and by a Lightnin A315 impeller. Again, the impeller geometry was included
i

explicitly. A series of simulations were carried out to test the individual effects of
various alternative modelling options. With the final method, based on developments in
this study, simulation results show reasonable overall agreement in comparison with
experimental data for bubble size, gas volume fraction, overall gas holdup and gassed
power draw. In comparison to results based on previously published modelling
methods, a significant improvement has been demonstrated. However, a number of
limitations have been identified in the modelling method, which can be attributed either
to the practical limitations on computer resources, or to a lack of understanding of the
underlying physics. Recommendations have been made regarding investigations which
could assist with further improvement of the CFD modelling method.

ii

TABLE OF CONTENTS
Chapter 1. Introduction ...............................................................................................1
1.1 General background ............................................................................................1
1.2 Aim of the study..................................................................................................4
1.3 Scope of the study ...............................................................................................4
1.4 Organisation of the thesis....................................................................................5
Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks.....7
2.1 Introduction........................................................................................................7
2.2 Applications of gas-sparged stirred tanks ..........................................................7
2.3 Design of gas-sparged stirred reactors ..............................................................10
2.4 Characteristics of the flow in tanks stirred by a Rushton turbine .....................11
2.5 Dimensionless groups and correlations............................................................14
2.6 Alternative impeller designs .............................................................................18
2.7 Scale-up of stirred tank reactors........................................................................20
2.8 Advanced experimental methods .....................................................................21
2.9 Conclusions......................................................................................................23
Chapter 3. Review of Modelling Methods................................................................27
3.1 Introduction.......................................................................................................27
3.2 Basic principles of computational fluid dynamics ............................................27
3.3 Extension of the equations to two-phase flow ..................................................34
3.4 Review of simulations of single-phase flow in stirred tanks ............................37
3.5 Issues identified relating to single-phase modelling .........................................49
3.6 Review of simulations of gas-liquid flow in stirred tanks ................................52
3.7 CFD simulations of other systems with gas-liquid flow...................................58
3.8 Simulations of solids suspension in stirred tanks..............................................61
3.9 Differencing schemes for two-phase flow ........................................................63
3.10 Issues identified relating to two-phase modelling...........................................64
Chapter 4. CFD Simulations of Single-Phase Flow..................................................69
4.1 Introduction.......................................................................................................69
4.2 Simulations of single-phase flow with the Rushton turbine .............................69
4.3 Additional simulations of a tank stirred by a Rushton turbine ........................78
4.4 Prediction of detailed flow around the impeller...............................................82
iii

4.5 Prediction of turbulence ................................................................................... 88


4.6 Modelling of the Lightnin A315 impeller........................................................ 91
4.7 Conclusions....................................................................................................... 95
Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions ................... 135
5.1 Introduction..................................................................................................... 135
5.2 Approaches to modelling ................................................................................ 135
5.3 Averaging procedure for the two-fluid equations ........................................... 138
5.4 Closure method for the interfacial force ......................................................... 142
5.5 Comparison of models for the turbulent dispersion force .............................. 151
5.6 Evaluation of models for the turbulent dispersion force................................ 154
5.7 Added mass and lift forces.............................................................................. 156
5.8 Turbulence in two-phase flow ........................................................................ 161
5.9 Conclusions..................................................................................................... 165
Chapter 6. The Mean Drag Coefficient in Turbulent Flow..................................... 171
6.1 Introduction..................................................................................................... 171
6.2 Drag coefficient in stagnant flow.................................................................... 172
6.3 Previous studies of drag in turbulent flow ...................................................... 176
6.4 Development of a correlation for use in CFD simulations ............................. 184
6.5 Additional considerations for the CFD model ................................................ 193
Chapter 7. Modelling of Bubble Break-Up and Coalescence .................................. 211
7.1 Introduction.................................................................................................... 211
7.2 The population balance equation .................................................................... 212
7.3 Derivation of the bubble number density equation......................................... 213
7.4 Previously published literature relating to modelling of bubble size............. 216
7.5 Theory of bubble break-up.............................................................................. 218
7.6 Expressions for the break-up rate ................................................................... 222
7.7 Theories for bubble coalescence ..................................................................... 225
7.8 Efficiency term for coalescence...................................................................... 227
7.9 Modification of the coalescence efficiency expression .................................. 230
7.10 Prediction of ventilated gas cavities ............................................................. 233
7.11 Modelling within the framework of CFX4 .................................................. 236
Chapter 8. CFD Simulations of Gas-Liquid Flow .................................................. 241
8.1 Introduction..................................................................................................... 241
8.2 Data for validation of the model with the Rushton turbine ............................ 241
iv

8.3 Data for validation of the model with the Lightnin A315 impeller ...............245
8.4 Approach to development and validation .......................................................245
8.5 Modelling method for gas-liquid flow in tank stirred by Rushton turbine .....248
8.6 Modelling method for gas-liquid flow in tank stirred by Lightnin A315 .......252
8.7 Description of the modelling options..............................................................252
8.8 Simulation results for the tank stirred by a Rushton turbine..........................254
8.9. Results for simulations with the A315 impeller.............................................265
8.10 Conclusions...................................................................................................268
Chapter 9. Conclusions and Recommendations......................................................349
9.1 Introduction....................................................................................................349
9.2 Findings from the single-phase modelling.....................................................349
9.3 Findings from the two-phase modelling ........................................................350
9.4 Evaluation ......................................................................................................352
9.5 Recommendations ..........................................................................................357
Nomenclature ............................................................................................................359
References .................................................................................................................365
Relevant Papers Published by the Author.................................................................383
Appendix A: Summary of the Mathematical Model for Gas-Liquid Flow.............385
A.1 Introduction ...................................................................................................385
A.2 Equations for conservation of mass and momentum ....................................385
A.2 Reynolds stresses ...........................................................................................386
A.3 Interfacial forces.............................................................................................387
A.4 Bubble size model .........................................................................................390
A.5 Gas cavity model............................................................................................391

Chapter 1.

Introduction

Chapter 1. Introduction

1.1 General background


Mechanically-stirred tanks are widely used in the process industries, including
applications in production of chemicals, pharmaceuticals, foods, paper, minerals and
metals. Typical operations carried out in mixing tanks include blending of liquids,
contacting of a liquid with a gas or second immiscible liquid, solids suspension, and
chemical reactions. Despite many years of research and accumulated experience in the
design of this important type of equipment, the fluid flow behaviour of stirred tanks
remains a subject of active investigation. The design of a stirred tank needs to be
carefully matched to the particular operation, and due to the complex flow patterns
encountered, many uncertainties remain in design and scale-up procedures.

Operations involving multiphase mixtures, e.g. contacting of a liquid with a gas, another
immiscible liquid, particulate solids, or some combination of these, form a large
proportion of stirred tank applications. For multiphase operations, there are substantial
additional complexities which need to be addressed, compared with single-phase liquid
flow. Many of the uncertainties in design are related to multiphase aspects, and
therefore, the focus of this thesis is on multiphase flow. More specifically, this study
considers the case of gas-liquid contacting, which takes place in a significant proportion
of industrial stirred tank reactors.

Many experimental studies have been undertaken over the years to investigate the
characteristics of fluid flow in stirred tanks. Often, these studies have resulted in
empirical correlations which relate a global parameter, e.g. power draw, mixing time or
mass transfer rate, to the geometric configuration and operating conditions (Kresta &
Wood, 1991; Tatterson, 1991). These empirical correlations have been applied in the
design of mixing tanks, in combination with practical experience. For new processes,
the design is also generally optimised through studies of the process at the laboratory
scale. Then, for the full-scale reactor, a scale-up procedure is applied, which is generally
based on similarity criteria and various empirical scale-up rules (Bartels, 2002).

Chapter 1. Introduction

In such design approaches, empirical correlations may be limited in their applicability,


since extensive data are only available for standard tank configurations and common
impeller types. Also, design based on global quantities does not take into account the
non-uniform and complex three-dimensional flow in a stirred tank. Furthermore, due to
the approximate nature of scale-up procedures, the performance of the production-scale
equipment has often been found to be far from the optimum which was identified at the
laboratory level (Bartels, 2002). Therefore, for greater confidence in design, other
approaches are necessary, where better understanding of the fluid dynamics in stirred
tanks is obtained, including information about the internal flow structures and the
distributed properties of the multiphase dispersion (Bakker, 1992).

One approach to investigating the detailed internal flow is through experimental studies
at the laboratory scale, taking advantage of the considerable advances in techniques
which have been made in recent years. A range of visualisation and advanced
measurement methods, such as laser doppler velocimetry (e.g. Costes & Couderc, 1988;
Hockey, 1990; Petterson & Rasmuson, 1998), have been applied. However, while such
experimental methods provide valuable information, there are also various limitations.
For example, it is very difficult to apply experimental methods to full-scale industrial
tanks, and therefore the uncertainties of scale-up cannot be addressed. Experimental
methods also generally involve the use of model fluids (e.g. water and air) and cannot
be applied to real industrial processes, which potentially involve corrosive materials,
high temperatures and high pressure.

Hence, computer modelling offers an attractive alternative approach for investigating


stirred tanks. Computer modelling allows investigation of the detailed internal flow in
tanks of non-standard designs at actual process conditions. Compared to experimental
methods, computer modelling also offers advantages such as the ability to address scaleup issues or model full-scale reactors, and there is the potential to obtain data at a lower
cost in a shorter time frame (Fletcher, 1991). Computer simulation may also generate
data that are very difficult to obtain experimentally, and it may reduce the cost of pilot
plant development.

Computer simulation of stirred tanks can be achieved through the methods of


computational fluid dynamics (CFD). This is a method for obtaining numerical
2

Chapter 1. Introduction

solutions to the equations governing the flow of fluids and dispersed solids. Being based
on fundamental principles, the CFD method can be applied to new designs and new
geometries within given classes of problems, provided that the simulation method has
been sufficiently validated against representative test cases. Simulations of
mechanically-stirred tanks have been reported in the literature since the late 1970s
(Daskopoulos & Harris, 1996). The capabilities of CFD models have improved over the
years, due to continual improvements in the speed and memory capacity of computers,
and the on-going development of improved modelling procedures. However, accurate
simulation of this type of system is particularly challenging for CFD, and there are
many issues which must be addressed. This is even more so for the case of multiphase
problems. Therefore, developments in modelling stirred tanks have required on-going
efforts to refine and improve the method, and validation against experimental
measurements has remained necessary. Most modelling studies reported in the literature
have addressed only single-phase liquid flow, while studies considering multiphase flow
have been considerably fewer in number.

Procedures for CFD modelling of gas-sparged stirred tanks, as reported in the literature,
have presented a range of limitations. Earlier published studies adopted a simplistic
two-dimensional approach to the problem. While some later models have adopted a
more realistic three-dimensional geometry, predictive capabilities have been limited in
many cases since the impeller was not modelled explicitly. Instead, the impeller was
treated as a black box, with the fluid motion generated by the impeller being specified
by reference to empirical data. Only a small number of published studies have attempted
to include the impeller in an explicit way. Other typical simplifications of published
modelling methods have included the use of a single, fixed bubble size, but this ignores
the variations in bubble diameter due to bubble break-up and coalescence. Furthermore,
there has been a lack of agreement regarding the form of the equations governing the
two-phase flow. For example, various authors have applied different expressions for the
inter-phase forces on bubbles, and have adopted different approaches to calculating
turbulent dispersion of the gas. In many published studies, the accuracy and reliability
of results is limited, or else uncertain due to limited extent of validation against
experimental measurements. Due to these various limitations, there has been a clear
need for further development and validation of CFD modelling methods.

Chapter 1. Introduction

1.2 Aim of the study


The aim of this study was to investigate modelling methods for the improved numerical
simulation of gas dispersion in a liquid in mechanically-stirred tanks. In such an
improved modelling method, it was intended that the method should be as generalised
as possible, so as to be applicable to different tank and impeller designs, and offer
increased predictive capabilities compared to previously published methods. Therefore,
the model should include an explicit representation of the impeller geometry.
Furthermore, the modelling method should provide sufficient data for the design or
evaluation of a gas-sparged stirred tank. Therefore, outputs of the CFD model should
include velocities, flow patterns, gas volume fractions, bubble sizes, gas holdup and
power consumption. The accuracy of the modelling method should be assessed by
comparison with experimental data.

1.3 Scope of the study


With the aim of developing improved modelling methods, this study considers a range
of issues affecting modelling. However, for reasons of practicality in developing and
validating this modelling method, limitations must be set on the range of tank designs,
fluid properties and flow regimes considered. Thus, for the most part, investigations
have been limited to a configuration consisting of a standard design baffled tank
stirred by a Rushton turbine (i.e. a six-bladed disc turbine). This tank configuration is
the system which has been used most often in research. Therefore, modelling of this
configuration provides the greatest opportunity for assessing the accuracy of the
modelling method, since most of the available experimental information refers to a tank
with this type of impeller.

Likewise, development has been based on a turbulent air-water system, since this
corresponds to the model system for which laboratory data is available for validation.
Of course, real industrial systems involve other gases and liquids, and the liquid, in
particular, may exhibit different characteristics, such as a non-Newtonian viscosity. The
liquid may also be mixed with suspended solids. However, it is preferable firstly to
develop modelling for a simpler system before considering such complexities.

Chapter 1. Introduction

In industrial practice a wider range of impeller types are employed, since other impeller
designs are claimed to possess advantages such as improved energy efficiency and
flexibility in gas handling. This study also extends to modelling of one such impeller,
being the Lightnin A315, which is a wide-bladed hydrofoil suitable for gas dispersion.
The flow patterns generated by the Rushton turbine and the A315 are quite different,
and by including tanks stirred by both impeller types, a degree of generality can be
demonstrated in the model.

1.4 Organisation of the thesis


The structure of the remainder of this thesis is outlined as follows:

Chapter 2 provides a description of the applications and general design features of gassparged stirred tanks. A description is also given of the characteristics of the flow in
these systems, and the use of dimensionless groups and correlations to account for their
behaviour. Problems with design and scale-up procedures are outlined, and it is argued
that these problems might be addressed by CFD modelling.

Chapter 3 firstly summarises the general principles of computational fluid dynamics as


applied to stirred tanks. The literature is then reviewed relating to previous efforts in
modelling stirred tanks, firstly for single-phase flow, and then for the more complicated
case of gas-liquid flow. A number of issues are identified relating to modelling
procedures.

Chapter 4 describes work carried out to develop CFD modelling of single-phase flow in
a stirred tank, which provides a basis for the modelling of two-phase gas-liquid flow.
Different methods are compared for modelling the impeller motion, in terms of their
accuracy and computational requirements. The accuracy of modelling methods is
assessed in relation to velocities, the presence of trailing vortices, pressures near
impeller blades, and turbulence levels. Grid sensitivity is investigated.

Chapter 5 discusses the governing equations for two-phase flow. The derivation of
these equations is outlined in order to clarify the appropriate forms for terms such as
drag force, added mass, lift and turbulent dispersion force. The approaches of a number
5

Chapter 1. Introduction

of authors are considered for the specification of turbulent dispersion. The modelling of
turbulence in dispersed two-phase flow is also discussed.

Chapter 6 considers the specification of the drag coefficient for bubbles in turbulent
flow. The published literature on this topic is reviewed, from which it is found that drag
tends to increase due to turbulence. A unifying approach is then identified for
correlating data from several sources. This leads to an equation describing the effect of
turbulence on drag coefficient. Arguments are presented for the form of the equation
when extended to conditions for which experimental data are not yet available.

Chapter 7 describes modelling of bubble sizes in the tank using a bubble number
density equation. Models for break-up and coalescence are discussed and terms are
defined to take into account the efficiency and rates of break-up and coalescence. Also,
a modelling approach is outlined to account for gas cavity formation on impeller blades.

Chapter 8 describes the development of the CFD model for gas-liquid dispersion. The
sources of experimental data for validation are firstly described. A number of modelling
options are defined, so that comparison can be made between options which have been
used previously in the literature, and proposed new approaches based on developments
in this study. Results are presented for simulations at a number of operating conditions
with a tank stirred by a Rushton turbine and another tank stirred by a Lightnin A315
impeller. Simulation results are compared against experimental data available in the
literature, and it is shown that the preferred modelling options lead to significantly
improved agreement with data.

Chapter 9 summarises the main findings of this thesis, discusses some of the unsolved
problems and provides some recommendations for further work.

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

Chapter 2. Design and Fluid Flow Characteristics of GasSparged Stirred Tanks

2.1 Introduction
This chapter firstly illustrates some of the applications of gas-sparged stirred tanks in
industry. The design of such tanks is described in terms of typical configurations and
impeller types, and the flow characteristics of such tanks are discussed. The use of
global parameters and empirical correlations in design of stirred tanks is described, and
the issue of scale-up is outlined. It is seen that such approaches have a range of
limitations, which indicates the need for development of CFD modelling as a means of
assisting with the design of stirred tank reactors.

2.2 Applications of gas-sparged stirred tanks


As indicated in Chapter 1, mechanically-agitated tanks represent a very common and
important process operation across a wide range of process industries, including bulk
and fine chemicals production, food and beverages, pharmaceuticals, and minerals
processing and metals production. Mixing tanks are employed for a range of duties,
ranging from simple blending of liquids to complex chemical reactions.

Mixing

operations may be single-phase or multiphase (e.g. mixtures of liquids and solids, gas,
or a second immiscible phase). In multiphase operations the mixing vessel must meet
requirements such as suspension of solids, or break-up and dispersion of gas or liquid
phases as bubbles or droplets.

Indicating the importance of stirred tank reactors, it has been estimated (Butcher &
Eagles, 2002) that 50% of all chemical production takes place in batch stirred vessels,
representing an annual sales turnover value of US$1290 billion worldwide. Poor initial
design can lead to problems such as commissioning failures, production rates lower than
expected, and increased downstream processing costs, and these problems were
estimated as costing 0.53% of total turnover. Mixing problems also lead to additional
on-going maintenance and down-time costs. Furthermore, there may be costs due to
unnecessary overmixing (Tatterson, 1994). That is, because of uncertainty in design
7

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

procedures, it is necessary to overspecify the agitator power input or batch mixing time,
in order to ensure that the mixing is sufficient.

Gas-liquid contacting is an important industrial operation, since it has been estimated


(Tatterson, 1994) that about 25% of industrial reactions occur between a gas and a
liquid. A mechanically-stirred tank is often chosen for this purpose, although other
contacting methods are possible (Lee & Tsui, 1999). Examples of other contacting
methods include bubble columns, tray columns, and static mixers. The choice of
equipment is determined by factors such as the required residence times of gas and
liquid, the degree of conversion of reactants and selectivity for desired products, and
safety and flexibility of operation. A stirred tank may be preferred (Tatterson, 1994)
because it is possible to have a large inventory of liquid with a high degree of flexibility
over liquid residence time, and it may offer advantages such as a well-mixed
environment with relatively uniform reagent concentrations, pH and temperature.
Compared with equipment such as a bubble column, a stirred vessel can offer better
control over bubble size and spatial dispersion of the gas. Agitation also increases the
gas-liquid mass transfer rate. In addition to gas and liquid, a third solid phase is present
in some processing operations. In such a case, especially in minerals processing, a
stirred vessel is often used since agitation is also effective for keeping the solids in
suspension.

Gas-sparged stirred tanks are used for a variety of processes which may involve
reactions such as oxidation, hydrogenation, or chlorination. Some examples of such
industrial processes are as follows:

Aerobic Fermentation: Stirred tank fermenters are widely used in the food and
pharmaceutical industries, where micro-organisms are exploited to produce a variety
of products such as yeast, antibiotics, enzymes, amino acids, vitamins, flavour
enhancers, and thickening agents (Sengha, 1994; Benz, 2003). The use of a stirred
tank allows for suspension of the micro-organisms and facilitates uniform conditions
of pH, temperature, and nutrient and substrate concentrations, while sparging of air
provides for the oxygen requirements of the micro-organisms.

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

Hydrogenation: Industrially important hydrogenation reactions include the reaction


of hydrogen with vegetable oil for the production of margarine and related products
(Hasenhuettl, 1994). Mechanical agitation is required to suspend nickel catalyst
particles and disperse the hydrogen which is sparged into the bottom of the tank.
Typically, this takes place in a pressurised, high aspect ratio tank with multiple
impellers and internal heat transfer coils to control the exothermic reaction.

Pressure oxidation: Pressure oxidation processes are used in the minerals industry
to treat sulfidic ores, to liberate metals into solution for subsequent recovery by
solvent extraction and electrowinning. A typical application for pressure leaching is
treatment of refractory gold ores, where gold is bound up in the grains of pyrite or
arsenopyrite minerals (Thomas et al., 2002). In this case, reactions take place at high
temperature and pressure in an autoclave, which is typically a horizontal elongated
pressure vessel internally divided into a number of compartments separated by
weirs. Each compartment is fitted with its own agitator and oxygen sparger.

Bioleaching: This is another process used to extract metals from sulfide ores such as
sulfidic gold, copper and cobalt (Brierley & Briggs, 2002). Micro-organisms such as
bacteria and archaea species are introduced to the process, and these oxidise ferrous
ions or reduced sulfur species to obtain their metabolic energy, and in doing so
provide a pathway for leaching reactions. Air or CO2-enriched air is sparged into the
tank to meet the requirements of the microorganisms.

Mineral flotation: This process involves physical separation rather than a chemical
reaction, based on the exploitation of wettability differences of particles (Yarar,
1994). Applications are mainly in the minerals industry, but also include water
treatment and other applications. Air is introduced into a tank fitted with a radialstyle impeller, and particles are selectively attached to bubbles, which then rise to
the surface to form a froth, which is skimmed off.

It can be seen from these examples that many of the relevant industrial processes are
actually three-phase, since they involve gas, solids, and liquid. However, it is generally
the case that the most demanding requirement on the agitator (in terms of power

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

requirement and flow capacity) is the dispersion of gas, since solids are generally finely
ground. For example, in bioleaching, solids are typically mostly below ~75 micron in
particle diameter (Brierley & Briggs, 2002), so that solids suspension requires less
power than that for gas dispersion. Hence, gas dispersion is usually the major
consideration in these processes. Nevertheless, the impeller design must consider all
factors, and solids suspension may also be a factor in selecting the impeller type. An
additional consideration in processes involving micro-organisms is that shear rates must
be limited to avoid damage to the micro-organisms.

2.3 Design of gas-sparged stirred reactors


The primary consideration in design of a gas-sparged tank is the creation of a large
interfacial area (Tatterson, 1994), through a combination of sufficiently fine bubbles
and sufficiently high volumetric holdup. This must be done within constraints of
economic power input, and other process constraints such as sufficient heat transfer and
suspension of any solids present.

The typical components in a gas-sparged stirred reactor include a centrally mounted


shaft with one or more impellers, baffles at the wall, and a sparge pipe or sparge ring
near the tank bottom. Baffles are present to stop swirl in the tank, since this is
inefficient for mixing, although baffles may nevertheless be omitted in reactors such as
fermenters, where the presence of internals inhibits proper cleaning and sterilisation.
The aspect ratio of tanks varies. Often tanks of high aspect ratio are used, since this may
increase gas residence time, and in such cases multiple impellers on the same shaft are
used. In addition, internal heat exchange tubes may be present, either as a helical coil, or
in the form of vertical tube baffles. In some pressurised operations, horizontal
autoclaves with multiple compartments are used.

A wide range of impeller types are available, of which certain impeller types are
particularly suited for gas dispersion operations, while other impellers can only disperse
small amounts of gas. The most commonly encountered impeller for gas dispersion is a
radial flow turbine with six vertical blades mounted on a disc (Figure 2.1). This was
introduced by Rushton et al. (1950) as a standard design, and hence this is usually
known as the Rushton turbine. Advantages of this impeller type include the additional
10

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

strength given to the impeller by the disc compared, for example, to an open-bladed
paddle, and the disc also prevents short-circuiting of gas along the shaft (Smith, 1985).

A standard tank configuration has also evolved (Smith, 1985), as shown in Figure 2.2.
This is often used in experimental studies at the laboratory scale, and likewise in
numerical studies, thus providing a basis for comparing the results of different workers.
The standard tank configuration consists of a flat bottomed cylindrical tank filled to a
depth, H, equal to the tank diameter, T, with four full length baffles of width B = 0.1T.
Where the tank is fitted with a Rushton turbine, the impeller diameter D is normally
0.33T and is centrally mounted at a clearance, C, of 0.33H. The Rushton turbine has a
disc with diameter D2 = 0.75D, and the blade proportions are a length LB = 0.25D and a
width WB = 0.2D. This configuration has been used in the single-phase modelling
development in this thesis, and a very similar configuration has been used for the twophase modelling, but with an impeller clearance of 0.25T, in accordance with the
arrangement used by Barigou and Greaves (1992, 1996), whose data provided the main
basis for validation of the model.

In industrial practice, major deviations from the standard design are common. Examples
of such designs have been mentioned in Section 2.2, e.g. the use of high aspect ratio
tanks with multiple impellers for hydrogenation of vegetable oil, or the use of elongated
horizontal autoclaves for pressure oxidation. In such designs, the tank bottom is dished
rather than flat. Another variation is the use of a conical bottom. Also, in an effort to
overcome some of the perceived shortcomings of the Rushton turbine, alternative
impeller designs have been developed, and some of these have had commercial
acceptance. These include the Smith impeller, the Scaba impeller and the Lightnin A315
(see Section 2.6). Nevertheless, given the large amount of published data relating to
standard design tanks stirred by a Rushton turbine and the scarcity of data relating to
other impellers, a tank with a Rushton turbine has been chosen in this study as the basis
for most of the modelling development.

2.4 Characteristics of the flow in tanks stirred by a Rushton turbine


The flow characteristics of the Rushton turbine in the turbulent flow regime have been
studied extensively. As described by various authors (e.g. Nouri, 1988; Hockey, 1990),
11

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

the Rushton turbine generates a strong radial jet which emanates from the impeller and
impinges on the tank wall, and then divides into two wall jets, leading to the formation
of two recirculating ring vortices in the upper and lower parts of the tank (Hockey,
1990). Liquid returns to the impeller from above and below in the central region near
the impeller shaft. Measurements of mean velocities and turbulent fluctuating velocities
throughout the tank have been reported by a number of authors, e.g. Hockey (1990) and
Mavros et al. (1996), using laser doppler velocimetry. Measurements such as these
show that the flow pattern is also characterised by a wide variation in turbulence levels
in different parts of the tank, with turbulent kinetic energy being highest near the
impeller and in the impeller discharge stream. Further analysis of the turbulent energy
dissipation and other characteristics of the turbulence, such as the integral length scale,
has been reported by a number of authors, e.g. Wu and Patterson (1989), and these
quantities were also found to vary widely. Wu and Patterson estimated that the local
energy dissipation rate near the impeller tip was more than 20 times the average energy
dissipation rate, and the energy dissipation in the impeller region and discharge stream
accounted for 60% of the total.

An important feature of the flow in the immediate vicinity of the impeller blades is the
presence of trailing vortices, which are strongly swirling structures produced on the
trailing sides of impeller blades due to flow separation. Vant Riet & Smith (1975)
reported detailed measurements of velocities, pressure distributions, and the spatial
location of the trailing vortices formed behind blades of a Rushton turbine. They
observed pairs of vortices about one fourth of the blade height emanating from the top
and bottom edges of each impeller blade, with a strong reduction in pressure at the core
of the vortex. As well as radial flow impellers like the Rushton turbine, trailing vortices
are found in the flow produced by axial flow impellers (Bakker, 1992). For axial flow
impellers, a single tip vortex is produced (Smith, 1985).

Ranade and Joshi (1990) concluded that the trailing vortices play a central part in the
mechanism of energy dissipation of impellers, and are particularly important in
multiphase flows. For gas-liquid flow, it is found that due to the strong centrifugal
action and the low pressure at the vortex core, bubbles are drawn into the trailing
vortices near the blades, and then dispersed into the tank along the line of the vortex, as
it merges with the bulk flow in the tank (Smith, 1985).
12

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

While the basic flow pattern in a stirred tank has been described in many studies on the
basis of single phase flow, it is important to consider that the introduction of gas into a
stirred tank can have a major effect on the hydrodynamics, mainly because of the
manner in which the gas is drawn into the impeller (Smith, 1985). The interaction of the
gas with the impeller plays a crucial role in determining the flow characteristics in the
tank as a whole, especially in determining power consumption and circulation flow. Gas
introduced to an impeller tends to be drawn into the low pressure regions at the trailing
side of the blades, and it is found that at sufficiently high gas flow rates, ventilated gas
cavities are formed. It is found that an impeller operating in a gas-liquid mixture will
generally have a reduced power draw and pumping rate, due partly to the reduced
density of the mixture, and also due to the streamlining effect of the ventilated cavities
(Smith, 1985).

The pattern of gas dispersion will generally depend on a balance between the buoyant
energy of the gas and the power input of the impeller. Considering a Rushton turbine at
constant gas flow rate, at low impeller speed the buoyancy of the gas dominates the
flow, and the impeller is said to be flooded. At somewhat higher speeds, the impeller is
able to produce a radial dispersion action, and the impeller is said to be loaded.
Further increase in impeller speed leads to velocities in parts of the tank which are
sufficient to prevent bubbles rising, and then the recirculating regime is obtained
(Tatterson, 1991).

Several regimes have also been defined for the interaction of the gas with the Rushton
turbine and the formation of gas cavities on the blades. At low gas flow rates, the vortex
cavity regime is found, where the bubbles are drawn into the vortices and have the
appearance of a foam, while the single-phase structure of the vortices is maintained. As
the gas flow rate is increased, coalescence of the bubbles leads to growth of the cavities
and reduction in the volume of spinning liquid behind the blade. Eventually the gas
extends right up to the blade, and clinging cavities are formed. At higher gas flow rates
again, so-called large cavities are formed. A pattern of large and clinging cavities form
on alternate blades, known as the 3-3 configuration (Smith, 1985). An even higher gas
flowrates, large cavities form on all blades, and eventually the flooding point is reached,
where the flow of gas is so high that the impeller is ineffective for pumping liquid. Flow
13

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

maps have been produced to predict the gas cavity regime as a function of gas flow rate
and impeller speed, e.g. Warmoeskerken & Smith (1985). The reduction in power due
to gassing has been correlated with regard to the cavity regime (Tatterson, 1991).

2.5 Dimensionless groups and correlations


In design procedures for mechanically stirred tanks, there are a number of global
parameters or dimensionless groups which have been found useful in characterising
performance. Many experimental studies have aimed at determining the values of these
dimensionless groups and how they vary with the geometric configuration and operating
conditions. The most important basic characteristic of a stirred tank is usually
considered to be the power consumption, and this is often given in terms of a
dimensionless power number (sometimes called the Newton number), according to
(Tatterson, 1991):
NP =

l N 3D5

(2.1)

where P is the power consumption, l is the liquid density, N is the impeller speed (in
Hz or s-1) and D is the impeller diameter.

For single-phase flow, it has been found for stirred tanks that in general, NP is a
function of the impeller Reynolds number, Re, given by:
Re =

l ND 2
,
l

(2.2)

where l is the liquid viscosity. For agitators operating in the laminar flow regime (at
low Reynolds number), the power number is found to decrease with increasing
Reynolds number, while in a fully turbulent system (approximately Re > 104), the
power number becomes fairly constant for a given tank and impeller geometry. For
example, the power number of a standard design Rushton turbine in the turbulent
regime is about 5.0 (Tatterson, 1991).

Another useful characteristic is the primary flow rate of liquid produced by an impeller
at a given speed. This is expressed in terms of a dimensionless impeller flow number,
NQ, given by (Tatterson, 1991):

14

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

NQ =

Ql
ND 3

(2.3)

where Ql is the liquid flow rate through the discharge area swept by the impeller.
Many studies have been carried out to investigate how the power and flow numbers of
various impellers vary as functions of parameters such as impeller diameter, clearance
of the impeller from the tank bottom, impeller blade width, number of blades, angle of
blades etc. (Tatterson, 1991). By analysing the ratio of power to flow number, the
energy efficiency of different impeller designs for circulation of the liquid can be
compared.

When gas is introduced to a stirred tank (usually through a sparger located below the
impeller), a reduction in power is observed at constant impeller speed, with the effect
generally increasing with increasing gas flow rate. For a Rushton turbine, the loss of
power may be as much as 60%, before the flooding point is reached (Middleton, 1997).
A wide range of correlations have been proposed to account for the effect of gas on
power draw (Tatterson, 1991).

Two dimensionless groups which are commonly used in characterising the gas-liquid
dispersion are the gas flow number, Flg, and the Froude number, Fr. The gas flow
number (or aeration number) is given by:
Fl g =

Qg
ND 3

(2.4)

where Qg is the gas flow rate to the vessel, while the Froude number is given by:
N 2D
Fr =
.
g

(2.5)

where g is the acceleration due to gravity. These dimensionless groups have been used
by some workers to correlate the gassed power draw of the impeller (e.g. Bakker et al.,
1994). The aeration and Froude numbers have also been used to develop maps of the
flow regimes of a number of radial and axially pumping impellers (Middleton, 1997),
which indicate the formation of various types of cavity or where the flooding point is
reached.

15

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

Besides the power draw of impellers, the most important design parameters for a gassparged tank relate to the interfacial area created, and the consequent gas-liquid mass
transfer rate. Various correlations have been developed to predict the interfacial area, or
alternatively, separate correlations have been proposed for the gas holdup and average
bubble size, which together account for the interfacial area (Tatterson, 1991). Such
correlations are based on laboratory scale measurements for a limited number of
specific impellers and tank configurations, such as a Rushton turbine in a standard
design tank.
Most published correlations for the gas holdup, g, have taken the form:
Pg
g = C h
Vl

B
v sg ,

(2.6)

where Pg is the gassed power, Vl is the tank liquid volume, and vsg is the superficial gas
velocity. For example, Bakker et al. (1994) recommended an equation of this form, with
values for water-air systems being Ch = 0.16 0.04, A = 0.33 and B = 0.67.

Various authors (e.g. Calderbank, 1958; Lee & Meyrick, 1970; Bouaifi et al., 2001)
have proposed correlations for the average bubble size in stirred tanks. This is usually
represented by the Sauter mean diameter, d32, which is the bubble size which has the
same ratio of area to volume as the complete distribution. The Sauter mean diameter is
given as the ratio of the third and second moments of the bubble size distribution
function f(d), according to:
3

d 32 =

d f (d )d (d ) .
2
d f (d )d (d )

(2.7)

This is the most useful measure of average size since it is directly related to gas holdup
and interfacial area, a, according to (Barigou & Greaves, 1996):
a=

6 g
d 32

(2.8)

An example of a correlation for bubble size is that proposed by Calderbank (1958) for
coalescing systems:

16

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

d = 4.15

0.6
Pg

V
l

g0.5 + 0.0009 ,

0. 4

(2.9)

l0.2

where is the surface tension coefficient.

Correlations have also been proposed for the overall interfacial area, which results from
the combination of holdup and bubble size distribution. An example of such a
correlation, applicable to disc turbines, is that according to Hughmark (1980):
g
a = 1.38 l

Q
g
NV
l

2 4
N D
gWV 2 3
l

0.592

dN 2 D 4

V 2 3
l

0.187

(2.10)

where W is the impeller blade width.

Rather than calculate the interfacial area, correlations have also been proposed for the
combined mass transfer coefficient and interfacial area term, kla, as a function of tank
operating conditions. Hence according to Bakker et al. (1994):
Pg
k l a = C kla
Vl

b
v sg

(2.11)

where for air-water systems, the constants are given as Ckla = 0.015 0.005, a = 0.6 and
b = 0.6. The overall mass transfer rate can then be calculated according to:

dCl
= k l a Cl* Cl
dt

(2.12)

where Cl is the average concentration of gas dissolved in the liquid, and Cl* is the
saturation concentration.

As reviewed by Tatterson (1991), there are many other published correlations of this
type. These may cover different types of impeller, or in some cases, correlations have
been extended to cover a range of temperatures and pressures (e.g. Sridhar & Potter,
1980). However, correlations according to different authors may be functions of
different variables, and sometimes give conflicting results. Such correlation methods
may not work well in tanks of non-standard design, and are very difficult to generalise
to different liquids and gases. In particular, the bubble size is very sensitive to small
concentrations of species in the liquid such as electrolytes, surfactants, alcohols, oils

17

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

etc. (Middleton, 1997), making it very difficult to generalise these correlations for
different chemical systems. Sometimes, separate correlations are given for coalescing
and non-coalescing systems, but this approach is hardly likely to cover all possible
chemical systems.

Clearly, there are many situations where there will be no suitable correlation for a given
stirred tank reactor. On the other hand, these correlations may provide a useful guide to
how the behaviour of the gas-liquid dispersion will be affected by changing operating
conditions in an existing reactor. For example, according to equation 2.6, holdup will
increase with the power input as P0.33, and mean bubble diameter will decrease with
power input as P-0.4 according to equation 2.9.

2.6 Alternative impeller designs

Understanding of the flow produced by a gassed Rushton turbine has led to the
development of various alternative impellers for gas dispersion, which seek to improve
on the design of the Rushton turbine. The Rushton turbine has been criticised for two
main reasons. The first reason relates to the sharp drop-off of gassed power with
increasing gas flow rate, such that it is not uncommon in industrial operations to have a
power draw of 50% or less of the ungassed power (Nienow, 1990). It may be necessary
in the design to specify a motor large enough to mix the liquid in the case of zero gas
flow, even though most of the power capacity is not used under normal operating
conditions. Secondly, the Rushton turbine is relatively inefficient in terms of the liquid
flow produced for a given power input (Fraser et al., 1993). This is particularly a
problem where the mixing operation must also achieve some other duty such as solids
suspension or heat transfer, which are flow-controlled operations.

Other impeller types include those based on a disc turbine, but with curved, concave
blades (Bakker et al., 1994), which aim at reducing the strength of the trailing vortices
and therefore reducing cavity formation. Impellers of this type, such as the Scaba SRGT
hollow-blade type with parabolic blades (Nienow, 1990), have been shown to give
much flatter curves for power number as a function of gas flow number.

18

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

Gas-dispersion impellers of the downward axially-pumping hydrofoil type have also


been developed. Such impellers employ a high solidity ratio, where wide blades are
used to avoid gas shortcircuiting through the impeller. The hydrofoil shape gives
improved power to flow ratios, leading to smaller power input requirements. Designs of
this type have included the Prochem Maxflo T (Nienow, 1990), Lightnin A315 (Fraser
et al., 1993), Mixel TT, and the Narcissus impeller (Vlaev et al., 2002). Also, it has
recently been proposed to use upward-pumping wide-bladed hydrofoils. The use of such
impellers would be expected to lead to more stable behaviour, since the flows produced
by the impeller and by the buoyancy of the gas are working together in the same
direction, rather than being in opposed directions, which has been found to lead to flow
and torque instabilities (Nienow & Bujalski, 2004).

Of all these options, modelling of just one such impeller design has been included in
this study, being the Lightnin A315 impeller. This is a wide-bladed hydrofoil (see
Figure 2.3) designed to produce downward axial flow. Other common axial flow
impellers, e.g. the pitched bladed impeller or thin-bladed hydrofoils such as the Lightnin
A310 are only suited to dispersing relatively small amounts of gas. However, due to its
high solidity ratio (Bakker, 1992), the Lightnin A315 is able to disperse gas flow rates
of similar magnitude to a Rushton turbine without becoming flooded.

The Lightnin 315 is generally placed at a similar clearance to a Rushton turbine (about
1/3 of tank height) and pumps downwards towards the tank bottom, producing a single
recirculating flow pattern, as distinct from the two circulation loops of a Rushton
turbine. Compared with the Rushton turbine, the manufacturers (Lally, 1987; Fraser et
al., 1993) claim several advantages including the following:

Due to its hydrofoil shape, the Lightnin A315 has a much lower power number
(~0.7 compared with ~5.0 for a Rushton turbine), and this leads to a greater
hydraulic efficiency, so the impeller is able to recirculate gas bubbles more easily.
The lower power number also implies reduced torque, which leads to reduced cost in
the type of agitator motor required.

The Lightnin A315 maintains a relatively flat gassed power curve, so that, up to
about a gas flow number of 0.4, the power draw remains fairly constant, whereas for
19

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

the Rushton turbine the power drops off very rapidly. This has advantages since the
motor does not need to be overdesigned to cope with larger power requirement of
ungassed mixing (e.g. at start-up or in a plant upset), of which only a fraction is used
in normal operation.

The Lightnin A315 produces lower levels of shear, which is advantageous in


biochemical applications, where high shear rates may be destructive to
microorganisms.

Since the flow produced is directed toward the bottom, and since the impeller
discharge flow rate is higher, this impeller is more suited to solids suspension.

Laboratory measurements have indicated that the Lightnin A315 can produce mass
transfer rates at least 1015% higher for the same power input.

Since advantages such as these are claimed for the Lightnin A315, and since such
impellers are relevant to industrial practice, it has been thought worthwhile to include
modelling of the Lightnin A315 in this study. It was beyond the scope of this study to
assess claims of the manufacturer such as higher mass transfer. Rather, the modelling
has aimed at demonstrating the applicability of the method to a tank with this type of
impeller, so as to indicate a degree of generality in the modelling method.

2.7 Scale-up of stirred tank reactors

In the development or improvement of chemical processes, it is normal to carry out tests


at laboratory scale to determine, for example, reaction kinetics, residence times, product
yield, etc. A scale-up procedure is then required, to ensure that the reaction rate and
product yield achieved in a small tank (e.g. 110 litres) translate to economic results on
a much larger tank (possibly thousands of cubic metres in volume). Likewise, to study
an existing full-scale process, it is useful to be able to scale down to laboratory scale.
For a new process, an intermediate-size pilot plant is often built in order to reduce risk.
Poor scale-up procedures can have serious economic implications in terms of excessive
energy costs or less than expected production rates (Wernersson & Trgrdh, 1999).

20

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

The essential difficulty in scale-up (Oldshue et al., 1992) is that it is not possible, by any
scale-up technique, to maintain all of the properties of a stirred tank constant as the
physical scale is increased. For example, stirred tanks are often characterised in terms of
dimensionless groups such as the Reynolds number, Re, the Froude number, Fr, and the
Weber number, We. However, these are proportional to ND2, N2D and N2D3
respectively, so scaling while keeping any one of these groups constant means that the
other groups cannot be kept constant (Oldshue et al., 1992). An often-used scale-up
procedure is to maintain constant power per unit volume, but with this approach, as the
physical scale is increased, the distribution of fluid shear rates will change, since the
maximum shear rate will increase while the average shear rate will decrease. Also,
applying the principle of constant power per unit volume, mixing and circulation times
increase with scale.

Hence, scale-up is an uncertain process. According to Oldshue et al. (1992) the general
procedure should be to determine the feature of the process which is considered to be
controlling, and to scale up while keeping that characteristic constant. Other aspects of
the process will inevitably be different. For gas-liquid operations in stirred tanks, there
appears to be little agreement in the literature as to the best scale-up procedure.
Tatterson (1991) summarised some recommended procedures. For example, Westerterp
et al. (1963) recommended scale-up based on equal tip speed and D/T ratio. However,
Bourne (1964) recommended a basis of equal power per unit volume. Nishikawa et al.
(1981) recommended maintaining equal vvm (volumetric flow of gas per unit liquid
volume). According to several authors (Chandrasekharan & Calderbank, 1981; Oldshue,
1994), the usual assumption of maintaining geometric similarity is not necessarily the
best approach.

2.8 Advanced experimental methods

Earlier studies of stirred tanks were limited to measurements of a global nature, e.g.
power draw (e.g. using a torque meter) or overall interfacial area or kla (usually inferred
through measurement of some well-defined chemical reaction). Brief mention is made
here of more advanced experimental methods, since these provide one approach towards
gaining more detailed knowledge of the internal, three-dimensional flow in a stirred
tank.
21

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

For transparent, single-phase flow, advanced non-intrusive measurement techniques


such as particle image velocimetry (PIV) (e.g. Sharp & Adrian, 2001) and laser doppler
velocimetry (LDV) (e.g. Yianneskis et al., 1987; Costes & Couderc, 1988; Wu &
Patterson, 1989; Hockey, 1990; Mavros et al., 1996; Lee & Yianneskis, 1998) are
available, and these have been applied to obtain detailed measurements of liquid
velocities, flow patterns and turbulence parameters in stirred tanks.

For multiphase mixtures, there are limitations on the applicability of methods such as
LDV and PIV due to opacity of the mixture. However, one technique for two-phase
mixtures is phase doppler particle anemometry (PDPA), although this has only been
applied to very dilute two-phase mixtures, e.g. in PDPA measurements of a solids
suspension by Petterson and Rasmuson (1998), solids concentration was limited to
0.06%. More recently, PIV has been successfully applied at reasonable gas
concentrations. Aubin et al. (2004a) applied PIV to determine the mean velocities and
turbulent quantities in aerated vessels stirred by downward and upward pumping
pitched blade turbines. The dimensionless aeration number was 0.01 and the total gas
holdup was in the range 3.75. 8%.

There are other methods are available for investigating dense multiphase flows,
although these generally involve the use of an intrusive probe. For example, probes
have been developed to measure local gas volume fraction (e.g. Bakker, 1992; Barigou
& Greaves, 1996; Bomba et al., 1997) and local bubble size (e.g. Barigou & Greaves,
1992).

Another experimental method, which is non-intrusive and can be applied to opaque


liquids and multiphase flows, is electrical resistance tomography (ERT) (Mann et al.,
1996). This is based on using an array of electrodes positioned around the periphery of a
vessel to map out internal differences in electrical conductivity or resistivity over
different cross-sections in a tank. Therefore, it can be applied to various applications
where there are differences in conductivity. Mann et al. (1996) reported its use for
imaging mixing of strong brine solution into a lower concentration solution, imaging of
a vortex air-core inside a stirred vessel, and measuring the gas-voidage distribution in a
gas-sparged mechanically-stirred vessel.
22

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

All such experimental techniques are in general limited to laboratory or pilot-scale


tanks, since carrying out measurements at full-scale is generally too expensive or
impractical. In addition, the requirements of transparent liquids and ambient
temperatures and pressures limit most experimental investigations to model fluids (e.g.
air and water). While there is some potential for ERT to overcome this limitation,
measurements seem mainly limited to concentration fields, and issues remain relating to
spatial resolution and quantitative interpretation of the measured electrical signals.
Hence, in terms of experimental methods, it is difficult to investigate the effects of
scale-up or to investigate the behaviour of real, reacting systems.

2.9 Conclusions

In this chapter, typical applications of gas-sparged stirred tanks have been outlined, and
the design features and flow characteristics of these tanks have been discussed. It has
been shown that how the global or overall characteristics can be predicted using various
dimensionless parameters, empirical correlations and scale-up rules. However, such
approaches present a range of difficulties. One consideration is that the designs of
industrial tanks often differ from the laboratory configurations where correlations have
been developed. Also, correlations and scale-up rules only provide estimates of global
parameters, without any insight into the details of the fluid flow. The flow is very nonuniform in many ways, such as with respect to velocities, turbulence, and phase
distributions. In addition, transient behaviour may need to be described, e.g. with the
addition of a reagent to a stirred reactor which undergoes a fast chemical reaction, the
details of the mixing process may be important, rather than merely an estimate of the
mixing time. Scale-up procedures also present another problem, since it is often unclear
as to which parameter should be kept constant, e.g. superficial velocity or power per
unit volume, and it is not possible to keep all aspects of a process similar with changing
physical scale.

Hence, as emphasised by Bakker (1992), for proper understanding of the fluid dynamics
and reliable design and scale-up, it is necessary to investigate internal flow structures
and phase distributions. Experimental methods, such as LDV, PIV, ERT and various
probes, provide one approach towards obtaining the required information about internal
23

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

flow structures. However, such methods may be time-consuming and costly, they may
be limited to transparent liquids and model fluids, and are generally limited to the
laboratory scale. Hence, there is a clear need for reliable computational modelling
methods.

24

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

Figure 2.1. Six-bladed Rushton turbine (Mixing Equipment Co.).

Drive shaft
Liquid level

Baffle

6 bladed disc
turbine

LB

WB
D2
C
D

Figure 2.2. Layout of a standard configuration tank with a Rushton turbine (notation and relative
dimensions given in the text).

25

Chapter 2. Design and Fluid Flow Characteristics of Gas-Sparged Stirred Tanks

Figure 2.3. Lightnin A315 impeller (from Post Mixing website:


http://www.postmixing.com/mixing%20forum/impellers/impellers.htm).

26

Chapter 3. Review of Modelling Methods

Chapter 3. Review of Modelling Methods

3.1 Introduction

This chapter reviews previous efforts at CFD modelling of stirred tanks. Before
proceeding with this review, the basic principles of CFD modelling have been
summarised, to provide some background to the modelling issues. Then, the published
literature has been reviewed, considering both single-phase flow and gas-liquid
dispersion in stirred tanks. Also, some mention is made of relevant simulations in other
related situations, such as solids suspension in stirred tanks or gas-liquid flow in bubble
columns. By reviewing the literature, the prior state of development of relevant
modelling methods has been assessed. Several principle issues were identified relating
to shortcomings of previously published modelling methods, or the need for further
clarification where different authors have proposed different approaches.

3.2 Basic principles of computational fluid dynamics

Computational fluid dynamics is a method for simulating the flow of liquids, gases and
particulate solid particles, either separately or in some multiphase combination, along
with other associated phenomena such as chemical reactions and mass and heat transfer.
This is done through the numerical solution of the basic equations governing
conservation of mass, momentum and enthalpy in a fluid, where the fluid is assumed to
be a continuum without regard to the details of its molecular structure.

For an isothermal flow, the governing fundamental equations are the equations of
conservation of mass and momentum, which are given in vector notation by (Bird et al.,
1960):

+ ( u) = 0
t

(3.1)

( (

))

( u )
+ ( uu) = p + L u + u T + g
t

(3.2)

where is fluid density, u is the velocity vector, p is the dynamic pressure, t is time, and

L is the laminar viscosity.

27

Chapter 3. Review of Modelling Methods

Equation 3.2 is often referred to as the Navier-Stokes equation. The application of


equations 3.1 and 3.2 is most straightforward when applied to laminar flow of a singlephase fluid. However, practical flow problems generally involve a range of
complexities, such as the presence of turbulence or the dispersion of one phase in
another. The same equations still apply, and application to such complex flows without
further modification is referred to as direct numerical solution (DNS). However, the
DNS approach is generally impractical due to the huge computer resources required.
Therefore, averaged forms of these equations are usually solved. In such cases where
the equations have been averaged, empirical closure relations must be applied, and the
appropriate forms of these closures remain an active subject of investigation.

Whether using the exact form or an averaged form of the conservation equations, it is
usually not possible to obtain an analytical solution to these equations, except for a
limited number of simple geometries, and therefore a numerical routine is required to
find a solution (Fletcher, 1991). The method involves representation of the governing
partial differential equations by a set of algebraic equations which can be solved on a
computer (Fletcher, 1991). The geometry of interest is discretised into a number of
nodes or cells and the equations are solved over these cells using a range of iterative
techniques. Methods available for spatial discretisation include spectral, finite element
and finite volume methods (Fletcher, 1991). Finite volume meshes are most commonly
used. Within this category, meshes may be regular or body fitted, and structured or
unstructured. The simulations carried out for this study are based on body-fitted, blockstructured finite volume meshes, which is typically the case in previously published
work related to CFD simulation of stirred tanks.

Numerical solution of the Navier-Stokes equations involves a great deal of complexity,


and therefore, to assist in carrying out such CFD simulations, generalised software
packages have been developed and commercialised, and these are commonly used as a
basis for setting up a model, which may then be customised within the limits allowed by
the software supplier, e.g. through user-supplied subroutines. Commonly used
commercial software packages include Fluent, CFX4, and Star-CD. Many of the
modelling studies mentioned here in reviewing the literature were carried out using one
of these codes as a basis, although custom-developed codes were used in other cases,
and sometimes the particular code is not mentioned. The modelling work carried out for
28

Chapter 3. Review of Modelling Methods

this thesis was performed within the framework of CFX4, using additional user-written
subroutines as required.

The conservation equations, equations 3.1 and 3.2, are applied most readily to laminar
flow. However, the majority of practical flow problems involve turbulent flow, meaning
that the flow exhibits unsteady time fluctuations and related turbulent eddy structures.
In principle, direct numerical simulation may be attempted, wherein all temporal and
spatial scales of the flow are resolved. However, this approach tends to lead to an
enormous demand on computer memory and computation time, beyond the practical
capabilities of most computers. Hence, for practical simulations it is normal to use
approximate methods based on averaged forms of the equations, leading to the
introduction of a turbulence model. An intermediate approach known as Large Eddy
Simulation (LES) involves averaging over the smaller turbulence structures while still
calculating the large turbulent structures. However, in most cases, the equations
governing fluid flow are time averaged following a procedure called Reynolds
averaging (Fletcher, 1991; CFX4 Solver Manual, 2002).

In this approach, the instantaneous values of variables such as velocity u are expressed
as the sum of a mean and a turbulent fluctuating component, U and u, according to:
u = U + u .

(3.3)

These expressions are substituted in equations 3.1 and 3.2, and the equations are then
averaged. This leads to the following form of the equations:

+ ( U) = 0 ,
t

(3.4)

( U)
+ ( U U) = P + ( L (U + (U) T uu) + g ,
t

(3.5)

where each capitalised variable represents an averaged mean quantity. It should be


emphasised that since the numerical method solves for mean values of variables, the
resulting flow field will look considerably smoother compared to an actual snapshot
of the flow for a given instant in time, as would be obtained for example by
experimental methods such as particle image velocimetry (PIV).

It is seen that the Reynolds-averaged equation of momentum conservation contains an

additional term, uu , called the Reynolds stress, which arises because the
29

Chapter 3. Review of Modelling Methods

product of fluctuating velocity components is non-zero. The Reynolds stress tensor


represents the additional momentum transport due to the cascade of turbulent eddies.
The averaged equations cannot be solved unless a closure expression is specified for the
Reynolds stress, and this requires a turbulence model.

Turbulence modelling is a large field in itself and remains a major challenge for the
science of fluid mechanics. Turbulence models vary in their level of complexity, and
accordingly in the level of detail at which they attempt to model the phenomena of
turbulence. Simpler models include the Prandtl mixing length model, which however
requires a priori knowledge of the mixing length. The most popular models are the twoequation models, such as the k model.
A first step in turbulence models such as k is to invoke the Boussinesq eddy viscosity
concept (Bakker, 1992; CFX4 Solver Manual, 2002), where the Reynolds stress is
assumed to be given in terms of a turbulent viscosity, T, and the mean flow velocity
gradients according to:

u u = T U + U T

2
kI ,
3

(3.6)

where k is the turbulent kinetic energy per unit mass, which can be expressed in terms
of the fluctuating velocity component, u, as:
k = 1 u 2 ,
2

(3.7)

and I is the identity tensor.


The turbulent viscosity, T, is calculated according to:

T = C

k2

(3.8)

where is the rate of dissipation of turbulent kinetic energy per unit mass and C is a
constant whose value is usually given as 0.09 (CFX4 Solver Manual, 2002).
The variables k and in the turbulence model are calculated through conservation
equations according to (CFX4 Solver Manual, 2002):


( k )
+ ( U k ) = L + T k + ,
t
k

30

(3.9)

Chapter 3. Review of Modelling Methods


( )

2
+ ( U ) = L + T + c1 c2
,
t

k
k

(3.10)

where c1, c2, k and are model constants (values for these are given in Table 8.3), and

is the production of turbulent kinetic energy by mean velocity gradients given by:

2
= ( L + T )U U + U T U(( L + T ) U + k ) .
3

(3.11)

In applying boundary conditions at the walls in turbulent flow calculations with the k-
model, it is usual to use wall functions to model the boundary layer, since otherwise a
highly refined grid is necessary to resolve the steep gradients of velocity and other
variables in the boundary layer. The wall function approach is formulated using the
concept of a universal law of the wall (Launder and Spalding, 1974; Lathouwers, 1999;
CFX4 Solver Manual, 2002). This concept has been supported by experimental
evidence for a range of flows and assumes that there is constant stress in the near-wall
region and the eddy length scale is proportional to the distance from the wall. These
assumptions lead to a logarithmic velocity profile near the wall. A relationship is
obtained between the shear stress at the wall, w, and the velocity component parallel to
the wall in the adjacent node, ut ,adj , according to:
1

w =

C 4 k adj2
+
ln( Eyadj
)

ut ,adj ,

(3.12)

where is the von Karman constant, E is the roughness constant, C is a constant with
the usual value of 0.09, kadj is the turbulent kinetic energy in the cell adjacent to the
+
is the scaled distance to the wall, which is given by:
wall, and y adj

+
y adj

C 4 k w 2
=
y adj ,

(3.13)

where yadj is the distance from the adjacent cell centre to the wall. The equation for the
wall stress provides a boundary condition for the tangential velocity component. The
value of the turbulent kinetic energy in the cell adjacent to the wall is obtained by
solving the transport equation for k, but using a special treatment for the production
term (Lathouwers, 1999; CFX4 Solver Manual, 2002). The transport equation for

31

Chapter 3. Review of Modelling Methods

turbulent energy dissipation rate is not solved adjacent to walls, but instead, its value is
obtained through the relation (Lathouwers, 1999):

3
4k 2
adj

y adj

(3.14)

For the wall function approach to be valid, it is necessary that the first cell centre
adjacent to the wall should lie within the part of the boundary layer described by the
logarithmic velocity profile. Therefore, for the wall function approach to work properly,
it is required that the non-dimensionalised distance to the wall of each neighbouring cell
+
< 300 (Bartels et al., 2002).
should fall in the range 11 < y adj

The approach described thus far, employing the k model, is very commonly used. It is
favoured for its generality and its ability to be used in a wide range of applications.
However, situations sometimes arise where the model is not sufficiently accurate, and
therefore a range of other two-equation turbulence models have been used in various
circumstances. These include modified k models, such as the low-Reynolds number
k model, where instead of wall functions, the velocity profile in the boundary layer is
resolved using a sufficiently fine grid, and wall dampening functions are applied in the
transport equations for k and . Other two-equation turbulence models include the
Renormalised Group (RNG) k model, and the k model ( being the turbulence
frequency, given by = /k) (CFX4 Solver Manual, 2002).

A problem with two-equation turbulence models is the assumption of a scalar turbulent


viscosity, whereby the shear stress component in each direction is linearly proportional
to the mean velocity gradient, with the same constant of proportionality in each
direction. In real turbulent flows, particular those with swirl or anisotropic turbulence,
this assumption often does not hold and the relationship between the components of the
Reynolds stress tensor is more complex. This may be the situation with stirred tanks,
where there is swirl (such as in the trailing vortices), and measurements have shown
(e.g. Wu & Patterson, 1989; Bakker, 1992) that the turbulence near the impeller is
anisotropic.

32

Chapter 3. Review of Modelling Methods

More complicated turbulence models have been developed, such as the Reynolds Stress
model, where the modelling equations attempt to calculate all the normal and shear
stresses individually. The drawback of such a model is that instead of two differential
equations, seven are now required. Therefore, the Reynolds stress model has the
disadvantage of requiring considerable additional computational effort, yet because of
the complexity of this model and the need for additional empirical constants, the
performance of the model may not necessarily be any better than a two-equation model.
Another model which avoids the need for differential equations is the Algebraic Stress
Model. This has been used occasionally with varying degrees of success (e.g. Bakker,
1992).

Besides the choice of turbulence model, a range of other issues may need consideration
in a CFD model. These include the choice between two-dimensional and threedimensional representations of the geometry. The specification of grid density and the
distribution of grid points may be important in determining the accuracy of the solution.
Other aspects of CFD modelling which may affect accuracy are related to the numerical
method by which the equations are solved. A variety of iterative procedures have been
developed, and these vary in their accuracy and the computational demand required. An
important factor in the solution method may be the discretisation methods relating to
spatial and temporal gradients.

Spatial discretisation relates to the estimation of the convective terms in the governing
equations. To calculate the convective fluxes, an algebraic function is required
describing the spatial variation of variables between nodes or cell centres in the
discretised grid, and this algebraic approximation is referred to as the differencing
scheme (Casey & Wintergerste, 2000; CFX4 Solver Manual, 2002). A range of schemes
are possible and these vary in accuracy depending on the form of the algebraic function
and the number of grid points involved in the calculation. The degree of accuracy of
these is often described in terms of the order (first order, second order etc.). First order
schemes may introduce a large truncation error which is known as numerical diffusion,
and therefore higher order schemes should be preferred. However, the order of a method
does not necessarily have a direct connection to the accuracy of a solution on a given
grid (Casey & Wintergerste, 2000), and on relatively coarse grids lower order upwind
methods may give better solutions than higher order schemes. Furthermore, the choice
33

Chapter 3. Review of Modelling Methods

of differencing schemes is complicated by the fact that the more accurate schemes tend
to be significantly less robust and more computationally expensive (CFX4 Solver
Manual, 2002).

Similarly to the differencing scheme for spatial discretisation, when a transient


simulation is carried out (e.g. a sliding grid simulation of a stirred tank), then a temporal
discretisation scheme is needed to approximate the time derivative (Casey &
Wintergerste, 2000). Schemes include implicit first and second order schemes, and
again the choice may affect the accuracy of the solution.

Differencing schemes for the convective terms include upwind, central differencing,
hybrid, power law, second order upwind, quadratic upwind differencing (often referred
to as QUICK) and total variation diminishing (TVD) schemes (CFX4 Solver Manual,
2002). In the CFX4 code, the default spatial differencing option is called the hybrid
scheme. In this scheme, central differencing is used if the mesh Peclet number number
(i.e. the ratio of convection to diffusion) is less than 2. Otherwise, upwind differencing
is used but without diffusion (CFX4 Solver Manual, 2002). This is better than a simple
upwind scheme, since central differencing or upwind differencing is chosen where most
appropriate.

In the literature, various differencing schemes have been used by different workers in
simulations of stirred tanks, although the choice of scheme has not always been stated.
In a few cases, alternative differencing schemes have been evaluated. The choice of
differencing schemes by various workers is reviewed briefly later in this chapter.

3.3 Extension of the equations to two-phase flow


Besides the complexity due to turbulence, further complexity arises once multiphase
flow is considered. A number of approaches are available for modelling two-phase or
multiphase flows. For modelling of stirred tanks and other process equipment such as
bubble columns, it is quite impractical, except in extremely dilute flows, to attempt to
model the bubbles or particles in an individual sense. Hence, the most appropriate
approach is to apply the Eulerian-Eulerian equations. Here, an averaging method leads
to an approach where each phase is described in terms of a local volume fraction, i,
34

Chapter 3. Review of Modelling Methods

and separate equations are solved simultaneously for the conservation of mass and
momentum of each phase. The terms included in these equations vary according to
different authors in the literature. However, a generalised form can be stated as follows,
where the equations for phase i (where i = 1 for the continuous phase and i = 2 for the
dispersed phase) are:

( i i )
+ ( i i U i + (D i i )) = S i
t

(3.15)

( i i U i )
+ ( i i U i U i ) = i Pi + i i g
t
+ ( i ( L,i + T ,i )(U i + (U i ) T ))

(3.16)

+ Fi + A i + L i + Ti + S i U i

where Si is a source term, e.g. at an inlet or at the gas sparger, g is the gravity vector in
the buoyancy force term, 0 is a reference density, D i is the turbulent diffusivity
coefficient for volume fraction, Fi is the drag force, Ai is the added mass force, Li is the
lift force and Ti is the turbulent dispersion force.
As in the single-phase case, there is a Reynolds stress in each phase, which may be
modelled by an eddy viscosity, T,i, which is calculated from a turbulence model.
Approaches to modelling the turbulence include using the standard k- model, or some
modified two-phase turbulence model which takes into account additional source terms
due to the presence of the dispersed phase. For gas-liquid dispersions, the turbulence in
the gas phase is generally neglected.

Compared to the single-phase equations, there are a number of additional terms. In the
equation for conservation of mass, an additional dispersive term may be present in the
two-phase form, equation 3.15, where D i is the turbulent dispersion coefficient. This
dispersive term is found mainly in earlier simulations reported in the literature, and it is
an alternative to including a dispersion force in the momentum conservation equation.
In all cases where more rigorous ensemble averaging is applied (as discussed in Chapter
5), there is no dispersive term in the mass conservation equation (Lathouwers, 1999). In
the equation for conservation of momentum, equation 3.16, additional terms include the
gravity or buoyancy force, and the interphase forces, which may consist of drag force,
Fi, added mass, Ai, lift force, Li, and turbulent dispersion force, Ti. While all models
35

Chapter 3. Review of Modelling Methods

usually include the drag force, the other forces are not always included. In some cases,
there is no dispersive term, either in the mass conservation equation or as a turbulent
dispersion force.

Of the several forces acting between gas and liquid in a dispersed bubbly flow, the drag
force is usually the most important, since in the absence of acceleration, a balance
between drag and buoyancy forces determines the slip velocity of a bubble or particle.
This in turn is the most important factor for determining gas holdup and distribution.
The drag force is given by:
C
3
F2 = F1 = 2 1 D U 2 U1 (U 2 U1 ) .
4
d

(3.17)

The correct calculation of this drag force is an important consideration. To calculate the
drag force, it is necessary to calculate the drag coefficient, CD, which must be based on
an empirical correlation. Various authors have adopted a range of different approaches
to calculating CD.
The specification of the turbulent dispersion is also important in terms of predicting the
distribution of the dispersed phase, since due to random turbulent velocity fluctuations,
there is a tendency for bubbles to spread out. As well as the choice between dispersive
terms in the mass or momentum conservation equations, these terms take a number of
different forms depending on the author.

Another important issue in modelling gas-liquid dispersion is the local bubble size, d. A
range of bubble sizes will exist in a stirred tank, where the bubble size will show a
probabilistic distribution which varies with position and time throughout the tank.
Bubble size can be modelled by prescribing a number of different size classes. More
commonly, the bubble size is specified in terms of a mean diameter. The mean diameter
may be varied spatially throughout the tank, or an overall average diameter may be
used. It can be seen from the literature that the simplest assumption has been the one
most commonly used, i.e. a uniform average bubble size has been prescribed based on
empirical data. A more sophisticated approach, used by a smaller number of workers,
has been to include a sub-model to predict local bubble size based on the dynamics of
bubble break-up and coalescence. Regardless of the approach taken for predicting
bubble size, the usual practice is to calculate the gas phase flow with a single equation
36

Chapter 3. Review of Modelling Methods

for momentum conservation using a mean diameter. This means that all bubbles must
move at a single, average slip velocity. A more sophisticated approach would be to
model the gas with several size classes, each with its own transport equations.

The derivation and form of the two-phase equations are discussed in greater detail in
Chapter 5. Here, the intention has been just to outline these equations, since in
reviewing the various published studies to date, it is found that there are numerous
differences in the treatment of the two-phase equations according to different authors,
and these differences will be highlighted.

3.4 Review of simulations of single-phase flow in stirred tanks


The majority of published work relating to CFD modelling of mechanically-stirred
tanks has been concerned with single-phase liquid flow. This subject has been reviewed
firstly, since although this represents a simpler situation compared to multiphase flow,
the modelling methods developed for single-phase flow form a basis for multiphase
simulations.

The development of CFD modelling of mechanically-stirred tanks has been reviewed by


Daskopoulos and Harris (1996), and Brucato et al. (1998a). Daskopoulos and Harris
traced the origins of the application of CFD to stirred tanks back to the late 1970s. The
earliest published simulations were comparatively simplistic. However, simulation
methods have progressed over the years in response to increasing computer resources
and more sophisticated modelling techniques. Several common themes can be identified
in terms of the challenges related to modelling of stirred tanks. One problem relates to
the very large demand on computer memory and computation time, which is necessary
to resolve the three-dimensional geometry and the complex, time-dependent threedimensional flow structures in a stirred vessel. Related to this problem, various authors
have made different choices such as between two- or three-dimensional representation
of geometry, and steady-state or transient calculation. Grid density is also an issue
affecting accuracy. A major issue impeding progress has been the need to develop
techniques to deal with impeller rotation in a baffled tank, since because of impeller
motion, there is no single fixed frame of reference in which to calculate the flow field.

37

Chapter 3. Review of Modelling Methods

Other considerations have included the choice of turbulence model. Various


compromises have been made by different investigators in response to these difficulties.

In early efforts at simulation of single-phase liquid flow (e.g. Harvey & Greaves, 1982;
Placek et al., 1986, Pericleous & Patel, 1987), the tank geometry was approximated as
axisymmetric and was represented by a two-dimensional axial-radial plane. Thus, the
effect of the impeller blades and baffles was smeared out and was represented by
sources or sinks of momentum in a steady-state calculation. The calculated overall flow
patterns demonstrated qualitative agreement with experimental data, although this type
of model is clearly limited by its inability to resolve three-dimensional flow structures.

Three-dimensional simulations were first reported by Middleton et al. (1986), and many
further such studies have since been reported. In simulations such as those of Middleton
et al., the baffles and three dimensional flow structures were resolved, but in their study
and those of many other authors, a practice has been continued where the impeller is
treated as a black box (e.g. Brucato et al., 1989, 1990; Gosman et al., 1992; Ranade et
al., 1989, 1990; Kresta & Wood, 1991). That is, the impeller motion is modelled by
some sort of momentum source in the volume swept by the impeller; or alternatively, an
impeller boundary condition method is applied, where velocity and turbulence
parameters are specified on the surfaces bounding the impeller region. Such data must
be obtained by some empirical approach or by reference to experimental measurements,
such as laser doppler velocimetry (LDV) measurements. This approach lacks generality,
since empirical data is always required for specifying the impeller. It has been
determined further (Brucato et al., 1998a) that the impeller boundary conditions are not
a property of the impeller but depend also on the environment surrounding the impeller,
e.g. clearance to the tank bottom and presence of other impellers, so one set of boundary
conditions can only be applied to tank geometries that are the same or only vary
slightly. Also, since impeller boundary conditions are generally averaged in the
azimuthal direction, trailing vortices generated by impeller blades are not predicted. As
discussed in Chapter 2, the trailing vortices are an important flow feature, especially
influencing gas-liquid flow.

An example of a study employing the impeller boundary condition method is the work
reported by Brucato et al. (1989, 1990). They used the commercial package CFDS38

Chapter 3. Review of Modelling Methods

FLOW3D (later called CFX4) to model a tank with a radial flow turbine. They imposed
azimuthally-averaged profiles of radial and tangential velocity components, k and , on
the vertical cylindrical surface bounding the impeller-swept region. The transport
equations for mass and momentum were solved for the whole vessel in conjunction with
the k- turbulence model, including the interior of the impeller zone. The approach of
Gosman et al. (1992) was somewhat different, in that the circumferential velocity
component and the values of k and were imposed throughout the volume swept by the
impeller blades by reference to experimental data, while all other variables were solved
for in the impeller zone in the normal manner. Other simulations using impeller
boundary conditions include those of Ranade et al. (1989, 1990) who simulated both a
radial flow turbine and an axial-flow pitched blade turbine. Bakker (1992) also used an
impeller boundary condition method similar to that of Brucato et al. (1989, 1990) to
simulate the flow produced by a Rushton turbine and a pitched blade turbine (see also
Bakker & van den Akker, 1994b). Simulations by Kresta and Wood (1991) used a
swirling radial jet model to specify the velocity components and turbulence quantities
on the vertical boundary swept by the impeller.

Another approach which has been used in a few cases is to adopt a rotating frame of
reference for the tank, so that the impeller geometry can be resolved. In such a case, the
baffles should be modelled as a momentum sink. Harvey et al. (1995) used a rotating
frame of reference, but they simply disregarded the relative motion between impeller
and baffles, while restricting themselves to laminar flow conditions. Such methods do
not appear satisfactory in describing the real effects of baffling.

Several methods have been developed to allow explicit simulation of the whole flow
field, including within the impeller, without reference to empirical data. Some of these
methods resolve the time-dependent motion of the impeller blades, while other methods
use a steady-state approximation. An example of the latter approach is the inner-outer
method first proposed by Brucato et al. (1994). In this approach, the tank is divided into
two partly overlapping zones, the inner domain being the region containing the
impeller and the outer domain being the remaining bulk of the tank including the
baffles. The computational procedure involves firstly simulating the flow in the inner
domain in a frame of reference rotating with the impeller. From a first trial solution in

39

Chapter 3. Review of Modelling Methods

this region, distributions of velocity and turbulence quantities are obtained which are
used as a first estimate of the boundary condition on the inner surface of the outer
region. The flow in the outer region is then calculated in the normal inertial reference
frame, to obtain an estimate of the velocity and turbulence quantities on the outer
boundary of the inner region. A second simulation of the inner region is then carried
out, and then the procedure is continued until satisfactory convergence is achieved.
Each of the inner and outer simulations was conducted under steady-state assumptions,
and exchanged values were azimuthally averaged and corrected for the change in
reference frame. Brucato et al. (1994) showed that the method gave quite good
predictions of the flow field, although trailing vortices from the impeller blades were
not resolved.

Another approach to account for the impeller motion is a moving-deforming mesh


technique, as implemented by Perng and Murthy (1993). A single mesh and reference
frame was used, with the grid region around the impeller rotating, leading to
deformation of the mesh near the interface. When this became too severe, the grid was
regenerated. Results were not compared to experimental data.

Various simulation methods have been developed in which a multi-block finite volume
mesh is used to represent the tank, and the section of the grid around the impeller is
allowed to rotate in discrete steps. An early example of this approach was reported by
Takeda et al. (1993), who applied their dynamic multiblock method to simulate the flow
in baffled tanks stirred by paddles and pitched blade impellers, but they did not provide
any comparisons with experimental data. Luo et al. (1993) modelled a tank stirred by a
Rushton turbine using the STAR-CD code. The equations governing fluid motion were
solved in time-dependent form over two domains, with the inner domain rotating with
the impeller. At the interface, the mesh was allowed to shear and/or slide to
accommodate the relative motion. Agreement of their results with experimental velocity
data was found to be better than that obtained with a steady-state black box model.

Since 1994, sliding mesh models have been implemented in commercial codes such as
Star-CD, Fluent and CFX4 (formerly CFDS-FLOW3D). Brucato et al. (1998a)
described how the method is implemented in CFX4. The flow domain is divided into
two cylindrical, non-overlapping sub-domains, each gridded as a separate block. The
40

Chapter 3. Review of Modelling Methods

outer domain is fixed while the inner one rotates with the impeller. The flow equations
in the rotating domain are still formulated in the fixed inertial reference frame, although
the rotation of the grid results in acceleration terms equivalent to the Coriolis and
centrifugal forces arising in a non-inertial frame. The two regions are coupled together
implicitly using an interpolation method that takes relative motion into account.
Simulations are carried out in transient mode, with the inner domain being rotated in
small increments, until periodically repeating results are obtained.

Various authors have evaluated the sliding mesh method. Daskopoulos and Harris
(1996) found results to be somewhat unsatisfactory, and reported severe underprediction
of the turbulence using the sliding mesh approach. However, the quality of their results
may have been diminished by insufficient grid resolution. Tabor et al. (1996) reported
on simulations of a tank with a Rushton turbine using a sliding mesh option in the
STAR-CD code. They obtained good agreement with experimental data for mean
velocities, and were able also to predict the structure of the trailing vortices.

Brucato et al. (1998a) carried out simulations of a tank stirred by a Rushton turbine
using three different methods, being an impeller boundary condition approach, the
inner-outer method and the sliding mesh method, all of which were implemented with
the CFDS-FLOW3D code. They compared the usefulness and accuracy of the three
methods. The impeller boundary condition approach was found to give reasonable
results in terms of velocity predictions in the bulk of the tank, although they commented
on the limited predictive capability of this approach due to its dependence on
experimental data of sufficient accuracy. The inner-outer method, as well as being more
predictive, was found to be more accurate than the impeller boundary method. The
sliding mesh method was found to be the most accurate with respect to agreement with
experimental velocity measurements, although it underpredicted the turbulence. Neither
the inner-outer method nor the sliding mesh method generated the impeller trailing
vortices in their simulations. Comparisons were made of the computational demands.
Compared to the impeller boundary method, the inner-outer method was 1.54.5 times
more demanding in computer time. The sliding grid simulations were 712 times longer
than those using the impeller boundary method.

41

Chapter 3. Review of Modelling Methods

Another alternative for calculating the flow in the whole of the tank is the multiple
frames of reference (MFR) method. This is a steady-state method which avoids the very
large computation requirements associated with the sliding mesh method. The flow
domain is divided into two zones, with the inner zone being in a rotating frame of
reference attached to the impeller, and the outer zone being in the fixed inertial
reference frame. In the inner zone the centrifugal and Coriolis forces are included in the
momentum conservation equation. The flow is calculated in both domains
simultaneously, with information being exchanged at the interface taking into account
the change in reference frame. Unlike the inner-outer method, there is no overlap
between zones. A simulation of this type was first reported by Luo et al. (1994), who
applied the method to the same tank geometry as previously used in their moving grid
simulation. Predicted velocities showed reasonable agreement with measurements, at
least better than obtainable with a momentum source method. Compared to the moving
grid method, the application of the MFR approach led to a saving of over ten times in
computer time.

A multiple frames of reference approach was also used by Wechsler et al. (1999) in
simulations of a tank stirred by a pitched blade impeller. By using high-performance
parallel vector computers, they could employ grids of up to about 1 million cells
covering a 90 periodic section, enabling them to test the grid sensitivity of the
modelling results. Even with a smaller grid of ~125,000 cells, the trailing vortices
vortices were resolved and their axes were in good agreement with experimental
measurements. They found that results with a MFR approach were in excellent
agreement with those obtained with a sliding mesh method, but the MFR method
required an order of magnitude less in computation time.

The multiple frames of reference approach has since been incorporated as a standard
option in commercial codes such as Fluent and CFX4. Two approaches are available in
CFX4. In the first, data are exchanged across the interface on a cell-to-cell basis.
Therefore, the results represent a snapshot of the flow for a given impeller position, and
results should preferably be averaged for several different impeller positions. In an
alternative approach, data exchanged at the interface is azimuthally averaged (see CFX4
Solver Manual, 2002).

42

Chapter 3. Review of Modelling Methods

Another steady-state technique is the snapshot approach, which was first developed
for simulating flow generated by a pitched bladed turbine (Ranade & Dommeti, 1996)
and later applied to the Rushton turbine (Ranade, 1997a), with the assertion that the
method can be applied to any impeller geometry. The snapshot approach calculates the
flow in the tank for a fixed position of the impeller relative to the baffles, by applying
sources and sinks on the surfaces of the impeller blades. Mass sources and sinks are
applied on the front and back blade surfaces respectively, to model the suction and
ejection of fluid. Sink terms are also applied at the back surface for all other variables,
but source terms are not added on the front side. Simulation results showed generally
good agreement with experimental velocity data and the energy dissipation in the tank
was only moderately underpredicted. The axis of flow from the Rushton turbine blades
agreed well with the axis of the trailing vortices, although the trailing vortices
themselves were not resolved.

Other developments of the CFD method have considered the prediction of turbulence
levels. Lee et al. (1996) carried out simulations of a tank stirred by a Rushton turbine
using the sliding mesh method. They compared simulation results for finite volume
meshes with different numbers of cells, and found that for grids of 100,000 cells or
more (covering a 180 tank sector assuming symmetry), the velocity field over most of
the tank was in good agreement with experimental data. However, they found that
turbulent kinetic energy was underpredicted, especially in the impeller discharge stream,
with the extent of underprediction being about 50%. Increasing the grid size to 240,000
cells showed only a slight improvement.

Ng et al. (1998) assessed the accuracy of predictions of turbulent kinetic energy in the
impeller discharge stream of a Rushton turbine, in simulations using the sliding mesh
method. They also found that the peak in the turbulence near the impeller tip was
substantially underpredicted. They repeated the simulation with higher grid resolution
and an embedded grid around the impeller blades, and found that improved predictions
of k were obtained, probably due to better prediction of velocity gradients around the
impeller. However, at a grid density of 400,000 cells, the prediction of the peak k value
was still only about 50% of the measured value.

43

Chapter 3. Review of Modelling Methods

Most simulations reported in the literature have modelled turbulence using the k-
model. However, recognising the limitations of this model in a swirling flow which is
not strictly isotropic (Wu & Patterson, 1989), some authors have investigated other
turbulence models. In simulations using either a body force model or a sliding mesh
approach, Ahlstedt and Lahtinen (1996) compared the performance of the standard k-
model, the Renormalisation Group (RNG) k- model, and the algebraic stress model.
They found that differences in results were quite small, and due to limited experimental
data they could not distinguish between them in terms of accuracy. In simulations of
Rushton turbine using the sliding mesh approach, Jaworski et al. (1997) compared the
standard k- model and the RNG k- model. They found no difference between the
models in predictions of mean velocities.

Turbulence levels were substantially

underpredicted in both cases, although predicted values were somewhat better using the
standard k- model.

Bartels et al. (2002) carried out simulations of a tank stirred by a Rushton turbine using
the multiple frames of reference method. With two different supercomputers, they were
able to employ grid sizes of about 2 million cells. With the k- model, they found that
turbulent kinetic energy was still underpredicted, even with such a high grid resolution.
They also carried out direct numerical simulations and found that this approach yielded
accurate results very close to experimental findings. However, the Reynolds number for
the direct numerical simulation was only 7,275, and they pointed out that at higher
Reynolds numbers, the amount of computer time would be exorbitant.

Simulations in the literature have mainly considered simple impeller shapes such as the
Rushton turbine and pitched blade turbine. On the other hand, modelling of more
complex impeller shapes, such as the various hydrofoil designs, is of interest since they
are favoured in industrial installations, for reasons such as improved hydraulic
efficiency. For simulations of impellers such as hydrofoils, published simulations have
mostly used a black box approach, e.g. Bakker (1992), who modelled a tank with a
Lightnin A315 impeller using the impeller boundary condition method. For simulation
using the sliding mesh or MFR methods, meshing of such impellers is more complex
and leads to non-orthogonal grids, which may adversely affect convergence of a
simulation.
44

Chapter 3. Review of Modelling Methods

Weetman (1997) described simulations in which a Lightnin A310 impeller was


explicitly represented in a structured mesh. They obtained solutions using the Fluent
code with either the sliding mesh or MFR methods. Comparison with data was limited
to showing that both methods gave the correct impeller discharge flow. Koh and Wu
(1998) also modelled a tank with a Lightnin A310 impeller, using a structured multiblock mesh and the sliding grid method. They found good agreement with experimental
velocity measurements and the impeller flow number. Naude et al. (1998) modelled a
baffled tank stirred by a LUMPP LB impeller. They used the MFR method in
conjunction with an unstructured mesh of ~49,000 cells for the whole tank, and
obtained reasonable agreement with velocity measurements and global parameters such
as the power number. Jaworski et al. (1998) investigated modelling of an APV-B2
impeller, which has four wide overlapping hydrofoil blades. In this case however, they
concluded that it was not possible to generate a satisfactory mesh using a structured grid
approach, probably due to excessive skewness, and they reverted therefore to the
impeller boundary condition approach.

3.4.1 Discretisation schemes for single-phase flow


Various choices have been made by different workers for the spatial and temporal
differencing schemes in the solution of the governing equations, and these are illustrated
by a few examples here. However, the choice of scheme has not always been stated.

Bakker (1992) carried out simulations of stirred tanks with various impellers using the
impeller boundary condition method. For spatial discretisation, he compared the Power
Law scheme (a first-order accurate scheme) with the QUICK scheme, and found that the
difference in simulation results was small. Therefore, for further simulations, only the
Power Law scheme was used, since the additional computation time with the QUICK
scheme did not seem justified.

In simulations of a tank stirred by a Rushton turbine using a sliding mesh method, Luo
et al. (1993) employed a second order centred scheme for spatial discretisation, and an
Euler implicit scheme for temporal discretisation.

45

Chapter 3. Review of Modelling Methods

Tabor et al. (1996) carried out simulations of stirred tanks using the MFR and sliding
mesh methods. They adopted a blended differencing scheme weighted as 0.8 central
differencing and 0.2 upwind differencing.

Brucato et al. (1998) compared the hybrid scheme with the third-order QUICK scheme
in their simulations of a standard-design tank stirred by a Rushton turbine. Based on the
impeller boundary method and using a grid of 97,440 cells covering a 60 tank section,
they found that predictions of velocity were not appreciably different with either
differencing scheme, which indicated that numerical diffusion effects associated with
the upwind scheme were not significant.

Naude et al. (1998) reported on simulations of a tank with a propeller using an


unstructured, tetrahedral mesh. According to them, the first-order upwind scheme is
mainly suitable for flows which align with the grid directions, but with their
unstructured mesh, this is never the case, and therefore they used a second-order
discretisation scheme.

Aubin et al. (2004b) carried out CFD simulations of single-phase flow in vessels stirred
by upward- or downward-pumping pitched blade turbines. The impeller motion was
accounted for using either the sliding grid or a multiple frames of reference approach. In
their modelling work, they investigated the effect on simulation results due to
alternative spatial discretisation schemes. The schemes tested were upwind, higher order
upwind, and QUICK. With the higher order schemes, they employed the almost secondorder accurate Van Leer limiter scheme for the turbulence variables. They found that the
choice of discretisation scheme had no significant effect on mean radial and axial
velocities in the tank, while the difference in the tangential velocities was fairly small.
With regard to turbulence, they found a greater difference between schemes. All
discretisation schemes underpredicted the turbulent kinetic energy (which is typical of
all reported simulations in the literature), but the extent of underprediction was not as
great with higher upwind or QUICK as it was with the upwind scheme.

46

Chapter 3. Review of Modelling Methods

3.4.2 Large eddy simulations


As well as simulations employing the Reynolds-averaged Navier Stokes (RANS)
equations in conjunction with a turbulence model, simulations of turbulent flow in
stirred tanks have been carried out using the large eddy simulation (LES) approach.
Simulations using this approach have been reported by authors such as Eggels (1996),
Revstedt et al. (1998), Derksen et al. (1999), and Hartmann et al. (2004).

In such an approach, an unsteady calculation is carried out in which all turbulent


fluctuations at a spatial scale greater than the grid spacing are resolved. Spatial filtering
of the Navier Stokes equations is applied, and a subgrid-scale model is introduced for
the unresolved scales of turbulence. This approach is more computationally demanding
than using the RANS equations, but less demanding than direct numerical simulation
(DNS). Such simulations offer the prospect of more accurate prediction of the turbulent
flow in stirred tanks, and offer a much more detailed picture of the actual turbulence,
rather than providing just the mean flow and simple statistics of the turbulence.
However, large eddy simulations are highly computationally demanding compared to
simulations bases on the RANS equations, since not only must they be carried out in a
time-stepping manner with small time steps, but LES must be continued for sufficient
length of time so as to obtain the statistics of the flow field, such as the mean flow field
and r.m.s. velocities.

According to Revstedt et al. (1998), the LES method can overcome various difficulties
with RANS equations and standard turbulence models. Such difficulties include an
inability to capture slowly varying structures in the tank. Also, two-equation turbulence
models are not suited to flows with anisotropy and streamline curvature, and whereas
Reynolds stress models are not so restricted, they are only suitable for turbulent flows at
high Reynolds numbers. Reynolds stress models have a large number of model
parameters whose values are not universal, and convergence difficulties sometimes
occur.

Revstedt et al. (1998) applied LES to model a baffled tank stirred by a centrally-located
Rushton turbine. A multigrid approach was taken, where the finest global grid had 643
cells. The motion of the impeller was modelled by adding source terms to the
momentum equations. The subgrid-scale stress was modelled implicitly as the
47

Chapter 3. Review of Modelling Methods

truncation error, using an upwind finite differencing scheme for the convective terms.
The simulation was carried out with time steps equivalent to 0.5 rotation, and a total of
30,000 time steps were completed, requiring 340 CPU hours on a workstation with a
120 MHz processor. Their results indicated qualitative agreement with experimental
data, although further improvements in accuracy were needed, for which they suggested
that an improved treatment of the impeller motion was required.

In simulations such as those of Eggels (1996), Derksen et al. (1999) and Hartmann et al.
(2004), a Lattice-Boltzmann method has been adopted for solving the governing
equations, rather than using a finite volume method. This approach was chosen for its
computational efficiency and its suitability to parallel computing. In the LatticeBoltzmann approach, particles reside on a lattice and are allowed to move from one site
to another during time steps. Collisions occur at lattice sites, and the effects on the
particles are calculated with a collision operator which obeys the laws of mass and
momentum conservation. A cubic, uniform lattice was employed, where to model the
boundaries of a cylindrical tank and a moving impeller, a force-field algorithm was
implemented.

Derksen et al. (1999) applied LES in conjunction with the Lattice-Boltzmann method to
simulate the flow in a tank stirred by a Rushton turbine. A conventional Smagorinsky
model was applied for the subgrid-scale eddy viscosity. The calculation was firstly
carried out using a coarse grid, and then continued on a grid of 1803 lattice cells, and the
calculation was continued on a four-processor parallel computer for almost one month.
Results demonstrated good agreement with experimental data for the phase-averaged
mean velocity field, and also demonstrated good predictions of the turbulent kinetic
energy. The positions of the trailing vortices were also well predicted.

Hartmann et al. (2004) carried out further simulations of a tank stirred by a Rushton
turbine, using both the RANS equations and LES based on the Lattice-Boltzmann
method. They compared two alternative models for the subgrid-scale eddy viscosity,
and made comparisons between the RANS and LES methods, and comparisons with
experimental data. They found that both RANS and LES methods gave accurate
representations of the mean flow, although the predictions of turbulent kinetic energy
were considerably more accurate using the LES approach.
48

Chapter 3. Review of Modelling Methods

3.5 Issues identified relating to single-phase modelling


From this review of CFD simulations of single-phase flows, several main points can be
drawn out. Firstly, the review demonstrates that there are a number of methods available
to address the problem of modelling the impeller in the presence of baffles. In earlier
published simulations, a black box approach has been taken, yet through the
investigations of other authors it becomes clear that this approach is rather limited in its
predictive capabilities. For a more general simulation method, this approach should be
avoided. Other methods should be preferred where the impeller is included explicitly in
the flow domain. The choice remains between a fully unsteady calculation using the
sliding mesh method, or a steady-state approximation. The sliding mesh method has the
advantage of capturing the full time-dependent flow field, but the necessary computer
resources are potentially excessive. This is particularly the case when extending the
simulation method to multiphase flow. As a compromise in a sliding mesh simulation, it
may be necessary to reduce the number of grid points to speed up calculation, but this
approach may diminish the accuracy of results due to insufficient resolution of flow
structures. Therefore, a steady-state approximation may be preferable. Of the several
approaches available, the multiple frames of reference approach is probably the best,
since it has been shown to be nearly as accurate as the sliding mesh method in terms of
predicting the mean flow field.

It may also be important that the modelling method captures the trailing vortex
structures produced by the impeller blades, since these may be an important feature in
modelling gas-liquid flow, where they are involved in the capture and dispersion of gas
by the impeller. These have only been identified in several studies, such as those using
the sliding mesh method, but in simulations using methods such as imposed impeller
boundary conditions, or the inner-outer or snapshot approaches, the trailing vortices
were not resolved.

Regarding the grid density, it appears that fairly accurate predictions of the bulk mean
velocity field and global parameters such as power and flow numbers can be obtained
using about 100,000200,000 cells for half of a tank (where stirred by a Rushton
turbine, and assuming symmetry). For resolution of fine flow details, higher grid
resolution may be necessary. Also, for accurate prediction of the turbulence, it would
appear that a much greater grid resolution is required, but this is possibly beyond the
49

Chapter 3. Review of Modelling Methods

practical capacity of most current computers. Even using a supercomputer and a grid
with about ten times as many cells as in other simulations reported in the literature,
turbulence was underpredicted (Bartels et al., 2002). Where modelling complex
impeller shapes, it appears that careful attention needs to be given to the grid quality,
and in some cases (Naude et al., 1998) an unstructured mesh may be the only option.
Reported simulations of hydrofoils have not been extensively validated thus far.

Regarding the issue of turbulence models, quite a few papers have been published
which review different models such as k-, RNG k-, and RSM. However, in many
cases, such studies are quite inconclusive regarding the choice of model, since the effect
of the turbulence equations is obscured by other inadequacies in the model. For
example, Bakker (1992) compared the k- and algebraic stress models, but used a
momentum boundary condition method, which was shown by Brucato et al. (1998a) to
be less accurate than other methods. Wechsler et al. (1999) showed that a very fine grid
is necessary for accurate prediction of turbulence, so that unsatisfactory results can be at
least partly attributed to insufficient grid resolution rather than the choice of turbulence
model. Overall, there appears to be no evidence to date to show that any of the
alternative turbulence models perform any better than the most frequently used model,
being the standard k- model.

In all simulations using the RANS equations with a turbulence model for closure, the
turbulent kinetic energy and turbulent energy dissipation rate have been underpredicted.
One approach to this difficulty might be to avoid the RANS equations, and use DNS or
LES instead. However, the work of Bartels et al. (2002) using DNS has indicated that
the computational demand is extremely high for fully turbulent flow. Simulations of
stirred tanks have also been reported which use large eddy simulation, and these have
generally shown that both the mean flow field and the turbulent kinetic energy are well
predicted. Such an approach would therefore seem attractive in overcoming the problem
of underprediction of turbulence by various turbulence models, although the
computational demands of LES are very high, both in terms of the grid sizes required
and the computing time, even using the relatively efficient Lattice-Boltzmann method.
In terms of two-phase modelling, the computational demand will be considerably higher
again. Due to this likely high computational demand, it was decided that LES was not

50

Chapter 3. Review of Modelling Methods

feasible for the development of a modelling method for gas-liquid flow. Therefore, for
practical calculation of two-phase flow, the RANS equations were used, and the k-
turbulence model was employed, since as stated above, alternative turbulence models
have thus far not demonstrated any improvement.

Regarding the choice of differencing schemes for spatial discretisation, the literature
indicates that a range of different schemes have been used for the convective terms in
the governing equations. In some cases first-order schemes such as hybrid or power law
have been used, while other workers have adopted higher order schemes, presumably in
order to be more certain of accuracy in their results. Yet where comparisons have been
made between different schemes (Bakker, 1992; Brucato et al, 1998a; Aubin et al.,
2004b), it has been found in all cases that the choice of a higher order differencing
scheme has yielded a negligible improvement in the prediction of mean velocities. The
only significant difference reported is an improvement in the prediction of turbulent
kinetic energy using schemes such as higher order upwind, although, like in all other
reported simulations in the literature, the predicted turbulent kinetic energy still remains
underpredicted. Given these findings, it has seemed appropriate to carry out the
simulation work in this thesis using the default hybrid scheme in CFX4, since the use of
a higher order scheme would add to the computational demand and might reduce the
stability or convergence rate, probably with little benefit. The review of the literature
has suggested that it is likely that other factors would have a much greater impact on the
accuracy of the solution, e.g. the choice of modelling method for the impeller and the
grid resolution.

The choice of temporal discretisation scheme (for the sliding grid method) has been
stated in only a few cases, e.g. Luo et al. (1993) used an implicit (backward) Euler
scheme, which is first order. There is no evidence in the literature that this scheme is
inadequate, and since a second order scheme would be more computationally
demanding, it would seem adequate to use a first order scheme similar to Luo et al.
(1993).

Modelling of the single-phase flow in a stirred tank forms a basis for extension to twophase flow. As such, the development of modelling of single-phase flow which has been
undertaken in this study is described in Chapter 4. Some of the issues identified in this
51

Chapter 3. Review of Modelling Methods

review are explored further, in order to establish the practicality and accuracy of the
method.

3.6 Review of simulations of gas-liquid flow in stirred tanks


Compared to simulations of single-phase flow, two-phase or multiphase flows present
considerable additional complexities. The number of equations increases since the flow
fields and volume fractions of two or more interpenetrating phases need to be
computed, and a variety of issues arise related to the coupling between the phases
through interfacial forces, and modelling of turbulence in the multiphase mixture. In gas
dispersions, additional complexity occurs since bubble sizes are not fixed, but vary as
the gas flows through the tank, due to the processes of break-up and coalescence. As
well as being present as a dispersion, gas can be present as a continuous phase in the
form of cavities on impeller blades.

For two-phase or multiphase simulations of stirred tanks, the number of published


studies have been much fewer compared to simulations of single-phase flow. This is
despite the fact that most industrial applications are multiphase, and reflects the greater
level of difficulty and complexity of such simulations. The review in this section has
been limited mainly to simulations of gas dispersion, since that is the main focus of this
study. However, some of the issues in modelling gas-sparged tanks are similar to those
in other situations, especially other gas-liquid flows such as bubble columns, and also
solids suspension in stirred tanks. Therefore, several simulations of these flows have
been considered where they help to highlight the relevant problems.

It will be seen from the following review that a range of approaches have been adopted
in efforts to model gas-sparged mechanically-stirred tanks. Previously reported studies
show differences which may be related to various simplifications of the problem to
make the computation feasible, or the differences may relate to assumptions made about
the two-phase flow and the form of the equations describing the underlying physics.
Several main issues can be identified, including accuracy of representation of the
geometry (i.e. two- or three-dimensional grid, and grid resolution); approach to
modelling the impeller rotation; the assumed form of the modelling equations including
definition of the interphase forces and modelling of turbulence; approach to modelling
52

Chapter 3. Review of Modelling Methods

the bubble size; and the extent of validation and demonstrated degree of accuracy of the
modelling method. The review here attempts to identify the approach to these issues in
each published study.

The first published CFD modelling study of a gas-sparged stirred vessel was by Issa and
Gosman (1981). In their model of a vessel equipped with a Rushton turbine, a black
box approach was taken to represent the action of the impeller. A uniform small bubble
size was used throughout the tank, and the force between the bubbles and the liquid was
specified in terms of a drag force with a drag coefficient given by standard correlations
for bubble rise in stagnant flow. A three-dimensional grid was used to represent the tank
geometry, covering a 90 section of the tank using periodic boundaries, although the
grid used was extremely coarse (8 5 7 interior nodes), which was probably due to the
limitations of computers available at the time. Results were not validated due to lack of
suitable experimental data.

Pericleous and Patel (1987) carried out simulations of gas-sparged tanks with a number
of different impeller types. The flow field was taken to be axisymmetric, so that the grid
was two-dimensional. An approximate momentum source method was used to account
for the impeller action, while the baffles were modelled as a momentum sink. The
bubble size was assumed uniform with a constant bubble slip velocity. The turbulence
model used was a one-equation model based on the Prandtl mixing length hypothesis.
Simulation results were not compared to experimental data.

Trgrdh (1988) modelled a fermentation vessel equipped with two radial turbines. He
assumed that the flow was axisymmetric, and modelled the impellers and baffles using
the source term method according to Harvey and Greaves (1982). In the boundary
conditions for the bottom impeller adjacent to the sparger, the pumping capacity was
reduced in proportion to the reduction in power due to gassing, as given by an empirical
equation. A constant bubble size was assumed throughout the tank, which was obtained
from an empirical correlation by an iterative procedure, since the average bubble size is
a function of the gas holdup. The results appeared qualitatively reasonable, but no
comparisons were made with experimental data.

53

Chapter 3. Review of Modelling Methods

Gosman et al. (1992) reported on simulations of gas-liquid dispersion in a tank stirred


by a Rushton turbine. This work was an extension of the earlier work reported by Issa
and Gosman (1981). Simulations were based on a three-dimensional grid with
27 20 15 cells in the axial, radial and cirumferential directions respectively (8,100 in
total). To model the effect of impeller rotation, they specified experimentally-based
values of the tangential velocity and the turbulence parameters k and throughout the
volume swept by the impeller, while solving for the radial and axial velocity
components in the normal way. They ignored variation in bubble size and used a
constant bubble size of 4 mm, which was justified on the basis that small variations in
bubble size will not be expected to affect the rise velocity of the bubbles, which is
almost constant for diameters 26 mm in still liquid. Special emphasis was placed on
the development of the equations governing two-phase flow. Their version of the twophase equations included the added mass force as well as the drag force (but no lift
force) and terms for turbulent dispersion originating from averaging of the drag and
added mass forces. Modelling of the liquid phase turbulence was done by a k model
which was modified to include extra source terms due to interphase momentum
exchange. Results were compared with experimental gas volume fraction measurements
at several points in the tank, and in general the gas volume fractions were
underpredicted.

In the modelling work carried out by Bakker (1992), single-phase modelling using the
commercial code Fluent was extended to two-phase flow, by developing a code called
GHOST! (Gas Holdup Simulation Tool see also Bakker & van den Akker, 1994a).
The simulation procedure was based on the principal that one-way coupling could be
adopted, i.e. the flow of gas is defined by the liquid flow field, but gas flow is assumed
not to affect the liquid flow pattern. The only modification to this was that the liquid
flow field is reduced in proportion to the reduction in power caused by gassing, as
determined empirically. For the liquid flow field, the rotation of the impeller was
accounted for using an impeller boundary method, where the values of axial, radial and
tangential velocity components, and the values of the turbulence parameters, k and , are
fixed over the outflow surface swept by the impeller, such velocities having been
determined previously by laser doppler velocimetry (LDV) measurements. Three
dimensional grids were used and three different impellers were modelled, being the
54

Chapter 3. Review of Modelling Methods

Rushton turbine, a pitched blade turbine and the Lightnin A315 impeller. Two different
ring spargers were modelled.

While previously published methods had adopted a constant bubble size, Bakker (1992)
introduced a parameter called the bubble number density, as a means of predicting the
mean bubble size at each point in the flow domain. The bubble number density was
calculated through an additional conservation equation. In this equation, a source term
was introduced to represent the average rate of change of bubble size due to processes
of coalescence and break-up, with the bubble number density tending towards an
equilibrium value dependent on a critical Weber number.

For the estimation of the drag force on the bubbles, Bakker (1992) attempted to account
for the effect of turbulence, which he expected to lead to an increase in drag coefficient
compared to that predicted by the usual correlations for bubble rise in a stagnant liquid.
In his approach, the drag coefficient was still calculated based on a correlation for rise
through a stagnant liquid, but a modified Reynolds number was used, in which the
liquid viscosity is increased by adding a term proportional to the turbulent viscosity.

Overall gas holdup and overall mass transfer coefficient, kla, were well predicted for
each impeller type, though the number of operating conditions was limited.
Comparisons with data for local gas void fraction and local bubble size also showed fair
agreement, although Bakker commented that, in the outflow of the impellers, model
predictions and experimental data tended to separate, due to a lack of accuracy of the
impeller modelling method.

Djebbar et al. (1996) modelled a gas-sparged baffled vessel stirred by an axial


downward pumping Lightnin A315 impeller. They assumed that the presence of the gas
bubbles does not have any effect on the pattern of the liquid flow field. However, all
liquid velocities were reduced in proportion to the reduced power due to gassing, as
determined experimentally. For prediction of local average bubble size, they used a
bubble number density equation, similar to the approach proposed by Bakker (1992).
The impeller was modelled by the impeller boundary method, fixing the liquid velocity
and turbulence parameter profiles in accordance with experimental measurements.
However, for the boundary condition for , estimates according to two different authors
55

Chapter 3. Review of Modelling Methods

were tested. A two-dimensional axisymmetric solution was obtained using the finite
element code FIDAP. The number of grid cells was not mentioned. Discussion of the
accuracy of simulation results was limited only to consideration of the overall mass
transfer coefficients. The mass transfer coefficient was calculated from the gas volume
fractions and local gas volume fractions (giving an estimate of the interfacial area) in
combination with a correlation for kL. Reasonable agreement with measured mass
transfer coefficients was obtained. There was no discussion regarding the accuracy of
features such as the internal gas distribution.

Zhu and Stokes (1998) modelled a gas-sparged tank with a dished bottom stirred by a
Rushton turbine. They adopted an axisymmetric approach. In the area occupied by the
impeller blades, the tangential velocity component was taken as that of a solid rotating
body, whilst in the area occupied by the baffles, the tangential velocity component was
taken as zero. By this method they avoided the use of experimental data for the impeller
momentum source. They used a grid of 2,000 triangular elements, and solved the
governing two-phase equations using the general partial differential equation solver
Fastflo. Results were compared with experimental data with reference to velocities only.
Radial velocities of air were generally in good agreement. However, axial velocities of
air were generally over-predicted, and this was blamed in part on a difference between
the numerical and experimental averaging methods. Comparisons of gas distribution
and gas holdup were not considered.

Morud and Hjertager (1996) carried out modelling of a tank stirred by a Rushton turbine
using an axisymmetric, two-dimensional approach. The impeller and baffles were
modelled as momentum sources and sinks following the approach of Pericleous and
Patel (1987). The specification of the interfacial force between gas and liquid was
limited to the drag force. By assuming that the bubbles were in the distorted or churnturbulent regimes, they avoided the need to specify the bubble diameter, since the drag
coefficient is independent of size according to the correlation of Ishii and Zuber (1979)
for bubbles in a stagnant liquid. Simulations were carried out on a grid of 47 24 cells
(1,128 in total) in the axial and radial directions. They obtained good agreement with
gas holdup measurements. Results were also compared with gas velocity measurements
obtained by LDV, and qualitative agreement was generally obtained, although the axial
gas velocity was overpredicted in many places.
56

Chapter 3. Review of Modelling Methods

Jenne and Reuss (1997) modelled a tank stirred by a Rushton turbine, once again using
a two-dimensional axisymmetric approach. The impeller boundary condition method
was applied, using circumferentially-averaged radial and tangential velocities as
measured experimentally. However, since gas sparging reduces the impeller pumping
capacity, the radial velocity component was reduced in proportion to the reduced gassed
power, as determined from a correlation. A two-way coupled two-fluid model was used,
with turbulence in the liquid phase accounted for using the Chen-Kim modified k-
model. The interphase momentum coupling term was given as the drag force only, and
there was no mention of a model for turbulent dispersion of the bubbles. Bubble size
was not calculated explicitly, but was assumed to be in the ellipsoidal range 26 mm,
where by applying the correlation of Ishii and Zuber (1979), the drag coefficient is
independent of bubble size. By testing several grid sizes, they determined a suitable grid
size as 43 225 cells in the radial and axial directions respectively (9,675 in total).
Results for a single operating condition were compared in terms of the total holdup
predicted by a correlation, and very good agreement was obtained.

Ranade and Deshpande (1999) reported on simulations of a gas-sparged tank with a


Rushton turbine. They used three-dimensional meshes of 17 32 32 (17,408 in total)
and 35 63 80 cells (176,400 in total), with grid density being higher in the impeller
region. The flow induced by the impeller was modelled by the snapshot approach,
where sources and sinks are specified on the front and rear of each blade to simulate the
suction and ejection of fluid by the blades. Simulations were carried out within the
commercial code Fluent, where a coupled two-fluid model was used, where the
interphase momentum exchange term was specified in terms of the drag force only,
although the method of calculation of drag coefficient was not detailed. No term for
turbulent dispersion appears to have been included in the equations. Bubbles were
assumed to be uniformly 2 mm in diameter. Results were only assessed qualitatively.
Impeller pumping rate and power draw were found to reduce as the gas flow rate was
increased. Simulation results showed a tendency of gas to accumulate behind impeller
blades, although the extent of accumulation was underpredicted since trailing vortices
were not predicted by the simulation method.

57

Chapter 3. Review of Modelling Methods

Deen et al. (2002) modelled gas dispersion in a tank stirred by a Rushton turbine. They
used a three-dimensional grid of 46 56 36 cells (92,736 in total) in the radial, axial
and circumferential directions covering a 180 section of the tank, and another denser
grid of 370,944 cells. A fixed bubble size was used in the simulations, and the bubble
drag coefficient was determined from standard correlations for a distorted or spherical
shaped bubble. Turbulence was accounted for by the standard k- model, with an
additional turbulent viscosity component to account for bubble-induced turbulence,
according to Sato and Sekoguchi (1975). However, it would appear that turbulent
dispersion of gas bubbles was neglected in the model. Impeller rotation was modelled
by the sliding mesh method, and they calculated 20 impeller revolutions for a complete
simulation. Compared to earlier published simulations, this computation would appear
to be considerably more computationally intensive, due to the high grid cell number and
use of the transient sliding mesh method. Nevertheless, computation times remained
fairly reasonable, since they reported that their simulation on the fine grid took 54 hours
of CPU time on a SGI Origin 2000 computer, and this reflects the progressive
improvements in available computer speed and memory. In their results, they were able
to observe formation of gas cavities on each of the blades, despite the assumed uniform
bubble size. Comparison with experimental data was made in terms of the gas and
liquid velocities as determined by particle image velocimetry (PIV). Reasonable
agreement was found for the gas radial velocity component values, although the axial
velocity component of the gas phase was considerably overpredicted. No comparisons
were made with overall gas holdup.

3.7 CFD simulations of other systems with gas-liquid flow


Besides mechanically-stirred tanks, modelling of a range of other systems with gasliquid dispersed flow has been reported in the literature, such as bubble columns and
bubble plumes. These studies are of some interest, since, as illustrated by the following
examples, there are some similar issues in modelling of these systems, in particular with
respect to the specification of the two-fluid equations governing fluid flow, and related
sub-models for drag coefficient, turbulent dispersion force etc.

Davidson (1990) carried out axisymmetric two-dimensional simulations of a cylindrical


liquid bath with a central bubble plume. The form of the Eulerian two-fluid equations
58

Chapter 3. Review of Modelling Methods

followed the ensemble-averaged form of Drew (1983), where the momentum


conservation equation contains terms for the pressure difference between microscopic
phase interfaces and the bulk phases; the drag force; added mass and lift forces in the
mathematically objective form given by Drew and Lahey (1987); and a turbulent
dispersion force based on the mean drift velocity as derived from averaging of the drag
force. Drag coefficient was calculated according to the correlation of Ishii and Zuber
(1979), and a uniform bubble size was set equal to the empirically-determined value of
3.5 mm. For the Reynolds stress a constant turbulent viscosity was assumed so as to
avoid the uncertainties of a two-phase k- model. Davidson obtained reasonable
agreement with measurements of gas voidage and velocity. In his model, the pressure
difference terms and virtual mass terms were found to enhance numerical stability,
while lift and dispersion terms were important for obtaining the correct void fraction
profiles.

Ranade (1997b) simulated the flow in a bubble column using an axisymmetric twodimensional mesh and the Eulerian two-fluid equations, which in this case were timeaveraged rather than ensemble-averaged. With this averaging method, they obtained a
diffusive term in the mass conservation equation. Also, additional volume fractionvelocity correlation terms were obtained in the momentum conservation equation,
which would suggest that the volume fraction consists of mean and fluctuating
components. This approach to the two-phase equations therefore differs from the more
usual ensemble averaging, where the volume fraction is simply a mean quantity arising
from the averaging. In the momentum conservation equation, the added mass and lift
forces were not included, however, extra force terms were included to account for
bubble wake effects and wall effects. The standard k- model was used to account for
turbulent flow. For the drag coefficient, a constant ratio of CD/d was assumed, as typical
of bubbles with diameters 38 mm. However, Ranade noted that the high turbulence
levels found in bubble columns may be expected to lead to higher values of drag
coefficient (and correspondingly lower slip velocities) compared to those found for the
rise of a single bubble through a quiescent liquid. Therefore, the drag coefficient value
was increased by a factor of 2.2 to obtain agreement with experimental data. Using this
modelling approach, Ranade found good agreement with experimental data for gas
voidage, mean flow and turbulent velocity fluctuations.

59

Chapter 3. Review of Modelling Methods

Mudde and Simonin (1999) carried out simulations of a bubble plume in a rectangular,
narrow tank, where gas enters through a porous sparger off-set from the centre and
covering only a portion of the bottom of the vessel. With this arrangement, the bubble
plume shows an oscillatory behaviour, for which experimental measurements of the
amplitude and period are available for model validation. They applied a form of the
two-fluid Eulerian equations based on conditional ensemble-averaging and closure
relations developed by Simonin and Viollet (1990) and Bel FDhila and Simonin
(1992). In these equations, the averaging procedure leads to a number of terms in the
momentum conservation equation representing turbulent dispersion, which are written
as functions of various characteristics of the phases such as characteristic time scales
(this approach to turbulent dispersion is discussed in Chapter 5). Mudde and Simonin
used a modified k- model containing extra source terms related to interfacial transfer.
For the bubble size, a fixed diameter of 3 mm was assumed. For the interfacial forces,
they firstly only allowed for the drag force, with the drag coefficient based on a
correlation for spherical particles in stagnant flow, but this did not lead to satisfactory
results. By also including the added mass, they were able to obtain good agreement with
experimental measurements.

Modelling of the same oscillating bubble plume was investigated further by Oey et al.
(2003). They carried out a sensitivity study which investigated the effects of a number
of different modelling options. They found that interfacial closure expressions based on
the work of Gosman et al. (1992) and Simonin and coworkers (e.g. Bel Fdhila &
Simonin, 1992) gave similar results, with oscillations periods and amplitudes that were
more or less the same, and in fair agreement with experimental observations. In contrast
to the earlier work of Mudde and Simonin (1999), they found that inclusion of the drag
force alone was sufficient to reproduce the correct oscillating behaviour. Inclusion of
other effects, including the influence of bubbles on turbulence, the turbulent dispersion
force and the added mass force, had a secondary tuning effect on their results. The
choice of differencing scheme for convective transport was found to be more important,
since with some schemes the numerical diffusion was too high and adversely affected
their results. They did not include lift forces in their simulations, but suggested that the
force may sometimes be significant for a bubble column.

60

Chapter 3. Review of Modelling Methods

3.8 Simulations of solids suspension in stirred tanks


Simulations of solid-liquid two-phase flow in stirred tanks have also been carried out,
and these are of some interest since there are some similarities in the modelling
approach and related issues. One such simulation was carried out by Gosman et al.
(1992), who modelled a dilute suspension in water of glass particles with mean diameter
232.5 m, in a vessel stirred by a Rushton turbine. The impeller motion was treated
using imposed impeller boundary conditions. The finite volume grid contained about
16,000 cells covering a 90 periodic section. An Eulerian-Eulerian two-fluid model was
adopted, where the interphase forces consisted of drag force and several terms
describing turbulent dispersion. For the drag coefficient, the standard drag curve was
assumed to apply. Predicted velocities of the dispersed phase showed good agreement
with experimental measurements. A distinct vertical gradient in the solids volume
fraction was found, however, no comparison with experimental data was made for the
solids concentration values.

Micale et al. (2000) carried out CFD simulations of a stirred tank with suspended solid
particles. They modelled a laboratory scale tank stirred by a Rushton turbine, with glass
particles of diameter 355425 m and solids volume fractions in the range 0.0040.04.
The impeller was included explicitly, using the Inner-Outer approach to account for
impeller motion in the baffled tank, and the total number of cells in the grid was about
54,000. Two approaches for modelling the solids phase were tested, being a simple
settling velocity model, where the particles are assumed to be transported as a passive
scalar with a superimposed settling velocity, and a more complex Eulerian-Eulerian
two-fluid model. Turbulence was modelled by a homogeneous k- model, where both
phases are assumed to share the same k and values. In the two-fluid model, the
interphase force term was assumed to consist of the drag force only, with no allowance
for turbulent dispersion of particles. For the drag coefficient, the effect of freestream
turbulence on the value calculated from a standard drag curve was taken into account by
applying the correlation of Brucato et al. (1998b), where the drag coefficient increases
as a function of dp/K, where dp is the particle diameter and K is the average
Kolmogorov length scale of the turbulence over the whole tank. This correction led to
substantial reductions in terminal settling velocity, e.g. from 23.0 mm/s to 13.4 mm/s
for an impeller speed of 1200 rpm. Results were compared with experimental
61

Chapter 3. Review of Modelling Methods

measurements of axial solids concentration profiles. Fair agreement was found using the
settling velocity model, and better agreement was obtained using the more complex
two-fluid model. The use of the correlation of Brucato et al. (1998b) to correct the drag
coefficient was found in all cases to give improved agreement with experimental data,
since otherwise the concentration variation from top to bottom was exaggerated, due to
higher slip velocities.

Montante et al. (2001) modelled solids suspension in a high aspect ratio tank fitted with
four pitched blade impellers. The solids considered were dilute suspensions of monosized glass spheres, with diameters 0.1, 0.33 and 0.79 mm. The impellers were included
explicitly using the sliding grid method to account for impeller motion. A grid of about
95,000 cells was used, covering a 90 periodic section. An Eulerian-Eulerian two-fluid
model was used, where the interphase forces was represented by the drag force only.
However, turbulent dispersion was included through a diffusive term (proportional to
gradient of particle volume fraction) in the mass conservation equation, with the
turbulent Schmidt number assumed to be 0.8. Turbulence was modelled by a
homogeneous k- model. For the drag coefficient, a correction was made to allow for
the effect of freestream turbulence using a correlation due to Pinelli et al. (1996), where,
similar to the correlation of Brucato et al. (1998b), the drag coefficient correction is a
function of dp/K. Comparison of simulation results with experimental data for the
vertical solids concentration profile demonstrated very good agreement over a range of
conditions. The correction of the drag coefficient for turbulence was found to be
important in obtaining good results, especially for the largest particles considered.

Ljungqvist and Rasmuson (2001) used the CFX4 code to model a baffled tank stirred by
a downward pumping pitched blade turbine. They applied an Eulerian-Eulerian model
for solids suspension in the tank, in which the impeller boundary condition method was
used. Drag, added mass, lift and dispersion forces on the particles were included. They
also carried out measurements of particle slip velocities in very dilute mixtures (solids
0.0010.02 % v/v), using Phase Doppler Anemometry, which they compared with their
CFD results. Values of axial slip in the CFD results were similar to the particle terminal
velocity in stagnant fluid, while PDA measurements showed that the axial slip was
substantially smaller. The authors suggested that the reduced slip velocities may be due

62

Chapter 3. Review of Modelling Methods

to the effect of turbulence in increasing particle drag. They attempted to apply the
correlation of Brucato et al. (1998b) to account for the reduced axial slip, however, in
this case the correlation was found to have negligible effect.

Sha et al. (2002) modelled solids suspension in vessels stirred by pitched blade
impellers using an Eulerian-Eulerian two-fluid model. Impeller motion was modelled
using the sliding grid method, and turbulence was modelled through the k- model. The
interphase force term consisted of terms describing drag, virtual mass, lift, wall
lubrication and turbulent dispersion. For the drag coefficient, standard literature
correlations were assumed to apply. A wide size range of particles was represented by
six separate phases, with diameters ranging from 50 to 900 m. Comparison with
experimental measurements of the vertical suspension density profile showed only
approximate agreement.

3.9 Differencing schemes for two-phase flow


Various differencing schemes have been used by different workers for spatial and
temporal discretisation in the solution method for two-phase flow, although the choice
of schemes is not always mentioned. The choice of schemes is reviewed briefly here,
with the review limited to simulations of stirred tanks.

Issa and Gosman (1981) used a hybrid scheme for spatial discretisation. The same
scheme was presumably used later by Gosman et al. (1992).

Bakker (1992) employed a first-order power law scheme in his one-way-coupled


simulations of gas-liquid flow in stirred vessels.

Morud and Hjertager (1996) employed a hybrid spatial scheme and an implicit
(presumably first-order) temporal scheme in their axisymmetric two-dimensional
simulations.

Ljungqvist and Rasmuson (2001) simulated solids suspension in a tank with a pitched
blade turbine using the impeller boundary method. They used the CCCT scheme for
spatial discretisation , which is a modification of the QUICK scheme.
63

Chapter 3. Review of Modelling Methods

Deen et al. (2002) used a hybrid scheme for spatial discretisation in their simulations of
a tank stirred by a Rushton turbine. They used the sliding grid method but did not
report the temporal discretisation scheme.

Micale et al. (2004) simulated solids suspension in a stirred tank with the sliding grid
method. They used the hybrid scheme for spatial discretisation of the convective terms,
but did not report the temporal scheme used.

No comparisons between differencing schemes were identified in the literature for the
case of two-phase (gas-liquid or solid-liquid) flow.

3.10 Issues identified relating to two-phase modelling


This review of the existing literature, relating to modelling of gas-liquid dispersion in
stirred tanks, reveals that a range of different modelling approaches have been taken by
different workers in the field. The examples given for modelling of related flows
(bubble columns and solids suspension in stirred tanks) also highlight the range of
possible modelling approaches. No general agreement is evident as to the best way to
model this type of flow, and indeed, while some authors have proposed various
refinements, e.g. in the way that bubble size and drag on bubbles should be calculated,
later authors have not adopted these approaches. Some of the main issues are
summarised as follows:

Number of dimensions: Quite a few workers have used two-dimensional grids to


approximate a stirred tank. Presumably, this limitation has been imposed by
available computer resources, but the associated approximations, such as momentum
sources and sinks for impellers and baffles, would appear to be gross
oversimplifications. Several authors have reported good results in terms of
comparison with experimental data for gas holdup or correlations for mass transfer.
Yet, these models are not very realistic, and such accuracy is possibly due to tuning
of the model, or a matter of good fortune. In more recent publications a threedimensional grid has been shown to be feasible, and clearly, the three-dimensional
approach should be adopted for a fully predictive model.

64

Chapter 3. Review of Modelling Methods

Grid resolution: The number of grid cells has tended to increase over time,
reflecting the progressive improvement in available computer resources. The use of
coarse grids has probably obscured the performance of several models. If at all
possible, the finite volume grid should be dense enough to achieve grid
independence, and this should be demonstrated. However, grid independence might
still not be possible due to constraints on computer resources. In such a case, it
would seem advisable to make the grid as dense as possible. The work of several
authors suggests that a non-uniform grid should be used with a higher grid density
near the impeller, since the flow pattern is more complex near the impeller,
involving trailing vortices.

Impeller representation: Nearly all simulation methods reported in the literature


have adopted a black box approach to the impeller, typically using empirically
determined impeller boundary conditions. This avoids the need to model the flow
within the impeller, but puts severe limitations on the predictive ability of the model.
Additionally, the boundary condition data are usually only available from singlephase measurements, and a correction must be introduced to account for the effect of
gassing. The form of this correction represents an uncertainty. Other reported
simulations have included the impeller explicitly, either through the snapshot
approach, or the sliding grid method. In such cases, the model must be expected to
predict gas cavity formation and the related drop in power draw and impeller
discharge flow. The accumulation of gas behind impeller blades appears to have
been demonstrated qualitatively, but not quantitatively. Also, published models have
treated the gas in the cavity as an assembly of discrete bubbles, rather than a
continuum.

Formulation of the two-phase equations: Previously reported simulations reveal a


considerable variety in the way that the two-phase equations are specified.
Differences include those due to the averaging method, which can be time averaging
or ensemble averaging, and the treatment of turbulent dispersion depends on the
chosen form of averaging. In the momentum equation, there is variation in the
choice of interphase forces to include. In some cases only the drag force has been
included, while in other models the interphase force includes terms such as added
65

Chapter 3. Review of Modelling Methods

mass, lift and turbulent dispersion. The turbulent dispersion force varies in form,
while in some cases there is no explicit term at all in the model to account for
turbulent dispersion. Clarification is needed as to the correct approach to closure of
the two-phase equations.

Differencing schemes for the discretised equations: Review of the literature has
indicated that in most cases for two-phase simulations of stirred tanks, a first-order
differencing scheme such as the hybrid scheme has been used for spatial
discretisation. The scheme for temporal discretisation has hardly been mentioned,
but presumably the fairly standard implicit backward Euler scheme has been used
(where the sliding grid method was employed). Review of the literature has not
revealed any comparisons between alternative differencing schemes for multiphase
flow. However, from the review of single-phase flow simulations, it would appear
that higher order schemes may provide little benefit, at least as far as stirred tank
simulations are concerned. Since the hybrid (upwind-central) differencing scheme is
less computationally demanding and more robust, this scheme has seemed the most
appropriate for development of CFD modelling methods for gas-liquid flow. Since
the two-phase simulations have been implemented using a MFR method to account
for impeller motion, the temporal differencing scheme is not relevant.

Turbulence modelling: Turbulence has been most commonly modelled using the k
model, and since fairly reasonable results have been obtained with this model, it
would seem appropriate to use the k model in this study as well. While a Reynolds
stress model might be more accurate, it may lead to an impractically high
computational demand. Other models have been proposed such as the Chen-Kim
modified k model, yet any benefit from such modifications remain unclear due to
other deficiencies, e.g. the use of two-dimensional grids and impeller momentum
sources. The use of two-phase turbulence models has been suggested, particularly in
the case of modelling bubble columns, where additional source terms have been
introduced to represent the turbulence production due to the dispersed phase. The
present modelling method should also consider the modification of turbulence by the
presence of gas bubbles.

66

Chapter 3. Review of Modelling Methods

Specification of bubble drag coefficient: An important consideration in the twophase model is the specification of drag coefficient on bubbles, and the related
bubble slip velocity. Various authors have adopted one or another correlation for
drag coefficient based on published correlations for bubble rise in a stagnant liquid.
However, in some other cases, it has been considered important to modify the drag
coefficient to account for effects of freestream turbulence on drag, whereby slip
velocities tend to be reduced. The effect of turbulence has been considered by a
number of authors in relation to modelling gas-sparged stirred tanks, and also in
modelling of bubble columns and tanks with suspended solids. In a number of
studies where this modification has not been included, it has been noted that
predicted vertical slip velocities of bubbles or particles were too great in comparison
to measurements, or else the predicted gas holdup in the tank was too low. Further
investigation of the drag coefficient formulation seems warranted.

Prediction of bubble size: In the majority of published methods, the complexity of


the CFD method has been reduced by assuming a fixed, uniform bubble size.
Clearly, this assumption is at odds with reality, where there is a population of
bubbles of different sizes, with the population statistics varying throughout the tank.
Where the aim of the model has been to predict gas volume fractions and velocities,
but not interfacial area, this assumption has been justified by reference to the fact
that, for the expected size range, the distorted regime is assumed to apply and slip
velocity is independent of bubble size. Yet, this justification is made in reference to
correlations for drag coefficient in stagnant flow, which may not be applicable in the
turbulent flow inside a stirred tank. Prediction of bubble size is preferable, and this
has been demonstrated to be possible using a population balance model or a bubble
number density equation. Further investigation should be made into the prediction of
bubble sizes.

Impeller type: Simulations in which the impeller is explicitly included have thus far
been limited to the case of a Rushton turbine. Although gas-liquid flow has been
simulated with other impellers, this has only been done using a black box
approach. Yet other impeller types are also used in industry for gas dispersion, and
therefore a fully predictive approach, including the impeller explicitly, should be
investigated for these impellers also.
67

Chapter 3. Review of Modelling Methods

Validation: Some form of validation of the modelling results has been attempted in
most published studies, though not all. Validation has often been quite limited in
extent, often considering only one aspect of the two-phase flow. Such limits on
validation have often been due to lack of suitable experimental data for comparison.
In further development of modelling, the possibilities for more extensive validation
should be examined, and either suitable data should be generated, or modelling
geometry and conditions should be matched to those for which data is already
available. Simultaneous comparison with several aspects of the flow would be
preferable in order to judge the overall performance of the model.

The various issues, which have been identified in this chapter, have been considered in
the modelling development carried out for this study. Some of the issues are similar for
both single and multiphase flow, e.g. impeller representation and turbulence modelling,
and therefore development for the simpler case of single-phase flow is considered firstly
in the next chapter, as a basis for developing modelling methods for two-phase flow.

68

Chapter 4. CFD Simulations of Single-Phase Flow

Chapter 4. CFD Simulations of Single-Phase Flow

4.1 Introduction
It was necessary firstly to investigate the capability for modelling single-phase liquid
flow in a stirred tank, since this forms a basis for extending the modelling method to
gas-liquid flow. This chapter discusses various aspects of modelling single-phase flow,
including the ability to simulate the velocity field and predict basic global parameters
such as power number. Some aspects of the single-phase simulation results are
identified where limitations in the accuracy of predicting various parameters will have a
follow-on effect on the accuracy of two-phase modelling. In addition, the practicality of
the method is considered in terms of the computational effort required, bearing in mind
that even greater computational effort will be required for two-phase flow.

4.2 Simulations of single-phase flow with the Rushton turbine


4.2.1 Tank configuration
The simulations carried out for this part of the study were based on a standard
configuration baffled tank stirred by a Rushton turbine. The tank geometry and impeller
speed were set up to be the same as in the laboratory tank of Hockey (1990), who
obtained velocity measurements using laser doppler velocimetry (LDV). Simulation of
this tank has also been carried out by Luo et al. (1993, 1994). The tank diameter, T, is
0.294 m, and the tank is stirred by a Rushton turbine of diameter D =

3T

, which has

six equally-spaced vertical blades on a horizontal disc, with a blade height of


blade length of

4D

and a disc diameter of

4D.

5D,

The Rushton turbine is mounted on a

central shaft in a flat-bottomed cylindrical tank at a clearance of

3T

, and there are four

full-length baffles attached to the wall at equal spacing, with width

1 T
10

. The liquid

depth, H, is equal to the tank diameter, T, and the liquid was specified as water, with a
density of 1000 kg/m3 and a viscosity of 0.001 Pa s. The impeller rotational speed was
set to 300 rpm, corresponding to a Reynolds number of 48,000 (giving flow in the
turbulent regime).

69

Chapter 4. CFD Simulations of Single-Phase Flow

4.2.2 Modelling procedure using the sliding mesh method


For the single-phase flow in the stirred tank, the equations to be solved are the
conservation equations for mass and momentum, as given in Chapter 3, equations 3.4
and 3.5. For closure of the equations, the k- turbulence model was used (equations 3.9
and 3.10). Numerical solution of these equations was carried out using the commercial
code CFX4, and the method for solving the conservation equations is described in detail
in the manual (CFX4 Solver Manual, 2002). Briefly, the solution strategy is based on a
finite-volume method with all variables defined at the centres of cell volumes. The
SIMPLEC algorithm is used for velocity-pressure coupling and the Rhie-Chow
interpolation method is used to avoid checkerboard oscillations. Terms in the equations
are discretised in space using second-order central differencing, except for the advection
terms, where the user may select from a range of differencing schemes. Here, the
default option was chosen, which is a hybrid method in which central differencing is
used if the mesh Peclet number is less than 2, or otherwise upwind differencing is used.
For time derivatives, a fully-implicit backward differencing method was used. To
improve the stability of the solution method, under-relaxation factors were applied to
each variable.

It should be noted that the pressure which is solved for in the CFX4 code is a modified
pressure in which the hydrostatic pressure is removed, and the isotropic component of
the Reynolds stress tensor (as derived from the eddy viscosity hypothesis) is added.
Therefore, the modified pressure, p, is given by:
p = p + 2 k (g x ) ,
3

(4.1)

where p is the true pressure, k is the turbulent kinetic energy per unit mass, is the
liquid density, g is the gravity vector, and x is the position vector. The pressure is
calculated relative to a reference pressure, which in all simulations was set to a value of
zero for a reference location at the bottom, centre of the tank. This definition of pressure
was taken into account when comparing results with experimental data (see Section
4.4.2).

To represent the tank geometry, a finite volume mesh was specified, and this was
composed of two domains with an unmatched grid interface between them. The inner
rotating domain extends to a radius of 0.06 m and the outer stationary domain extends
70

Chapter 4. CFD Simulations of Single-Phase Flow

from 0.06 m to the wall (at a radius of 0.147 m), with each domain extending vertically
from the bottom to the surface. To account for the motion of the impeller, the sliding
mesh option within CFX4 was used, where the inner domain is allowed to rotate by a
certain number of degrees at each time step, and the flow is calculated with a
modification to the velocity terms in the conservation equations which takes into
account the motion of the grid (CFX4 Solver Manual, 2002).

The finite volume mesh was set up in cylindrical coordinates, with 48, 38 and 60 cell
divisions in the axial, radial and azimuthal directions respectively (109,440 cells), with
the grid resolution being finest in the region of the impeller, where the blade surface
was resolved by 6 10 cells in the horizontal and vertical directions respectively. To
reduce the computational time required, a 180 section of the tank was modelled, on the
assumption that flow is symmetrical for the other half of the tank. Figure 4.1 shows a
surface plot of the mesh, and also illustrates the rotation of the inner domain.

The baffles, impeller disc, and impeller blades were treated as zero thickness walls
(thin surfaces) and the impeller shaft is treated as a solid zone. The surface of the tank
was treated as a flat, zero-stress surface, whereas all other walls are treated as zero-slip
boundaries. At the walls, the boundary layer is modelled by applying a wall function, as
described in Chapter 3. This removes the need for a highly refined grid near the wall,
which would otherwise be necessary to resolve the steep gradients of velocity and other
variables in the boundary layer.

The flow was calculated by means of a transient simulation, with the initial condition
taken as zero velocity at all points. At each time step, the inner block was rotated by a
certain angle and the flow field was recalculated. The simulation was run in two stages.
In the first stage, an approximate solution was obtained using coarse time steps
corresponding to 30 rotation (i.e. a time step of 0.0167 seconds) for 10 revolutions of
the impeller. The solution was then refined by calculations with steps of 12 (i.e. a time
step of 0.0067 seconds) for a further 3 revolutions. About 20 iterations were used for
each time step, with under-relaxation factors set to 0.65 for the axial, radial and
azimuthal velocity components, 0.7 for k and , and 1.0 for pressure. The simulation

71

Chapter 4. CFD Simulations of Single-Phase Flow

was run until the developed flow pattern became periodically repeatable, indicating that
a steady state was reached.

4.2.3 Results using the Sliding Mesh method


The results of the simulation are illustrated by Figures 4.24.10. Figure 4.2 shows a
vector plot of the velocities for a vertical cross-section of the tank, through the middle
of the tank, half way between baffles and half way between impeller blades. The flow
field exhibits the characteristic pattern of a Rushton turbine, with radial discharge from
the impeller, which splits into upper and lower circulation zones, with liquid returning
axially to the top and bottom of the impeller. The flow field is three-dimensional and
shows various degrees of swirl, and interactions with the baffles, as illustrated by the
plot in a horizontal plane near the top of the tank, in Figure 4.3. Figure 4.4 shows the
velocity vectors in a horizontal plane intersecting the impeller. In Figure 4.5, the
pressure distribution is plotted for a horizontal plane through the impeller, showing the
regions of high pressure in front and low pressure behind each blade. Figure 4.6 shows
the distribution of turbulent kinetic energy in a vertical cross-section of the tank, while
Figure 4.7 shows the corresponding distribution of turbulent energy dissipation rate.
Here, it is seen that relatively high levels of turbulence are associated with the radial
impeller discharge stream, while turbulence decays away to lower levels at higher and
lower positions in the tank. Figures 4.8 and 4.9 show the turbulent kinetic energy and
rate of dissipation of turbulent energy in a horizontal plane intersecting the impeller
centre. The highest levels of turbulence in the impeller region are seen to be in regions
on the trailing sides of the blades and can be associated with the trailing vortices, as
discussed in more detail further on.

Since wall functions were used to deal with the steep gradients in the boundary layers, it
was necessary to check that the results were consistent with the wall function approach
as implemented in the CFX4 code. In Figure 4.10, the values of y+ in the cells adjacent
to walls have been plotted to check the suitability of the wall function approach for
modelling the boundary layer. Values are found to range from 15 to 160, and these
values are within the allowable range, confirming that the grid used is suitable in this
respect.

72

Chapter 4. CFD Simulations of Single-Phase Flow

For purposes of validation, simulation results have been compared to experimental data
for mean velocities at different positions in the tank as obtained by Hockey (1990). It
was necessary firstly to calculate values on the same basis. Using the sliding mesh
method, the final solution represents an instantaneous solution for a particular
orientation of the impeller with respect to the baffles. However, velocities may vary
according to impeller position. For the experimental Laser Doppler measurements, the
velocity measurements represent a time average for all impeller positions. To make the
simulation results comparable, mean velocity data were obtained by averaging over 24
positions of the impeller relative to the baffles.

Comparison with Hockeys experimental data has been made for radial profiles of the
velocity, from the tank centre to the wall, at several axial positions over the height of the
tank. Figures 4.11, 4.12 and 4.13 show comparisons of simulated and experimental
radial profiles, for the time-averaged axial, radial and tangential velocities at heights of
0.02 m (near the tank bottom), 0.065 m (just below the impeller), 0.12 m (just above the
impeller) and 0.18 m (about 2/3 of the tank height). All velocities are in a plane half
way between baffles. The tangential velocity components at 0.12 m and 0.18 m were
not reported in the experimental data and so comparison is not possible for these cases.
Although there are some discrepancies, the results show reasonably good agreement for
all velocity profiles.

Another approach to validating the simulation is to calculate global parameters such as


the power number, NP, and flow number, NQ, as defined in Chapter 2 (in equations 2.1
and 2.3, respectively).

The impeller discharge flow was estimated by calculating a summation over a


cylindrical area spanning the width of the impeller blades at a diameter of 1.02D. Thus,
the total discharge flow, QL, is calculated according to:
QL = 2 vi Ai ,

(4.2)

where vi is the radial velocity at the ith cell node, Ai is the area of radial flow of the ith
cell, and the factor of 2 takes into account the 180 periodicity. Using this method, the
estimated flow number, NQ, was calculated to be 0.71. This agrees very well with the
value of 0.73 reported in the literature (Costes & Couderc, 1988).
73

Chapter 4. CFD Simulations of Single-Phase Flow

The power input to the tank, P, was calculated from the torque, , on the impeller shaft
(as suggested by Harvey and Rogers, 1996), where torque was estimated from the
pressure differential on the blades and the shear stress on the disc according to:
= ( p1 p 2 ) i ri Ai + j r j A j ,
i

(4.3)

where the summation is over the control cells i corresponding to each blade and j
corresponding to the disc. The total torque is calculated as twice the torque on the three
blades in the 180 tank section. Power is then calculated according to:
P = 2N .

(4.4)

Using this method, the calculated power number based on simulation results was 4.5,
which is in very good agreement with experimental data, being only 4% lower than the
experimentally determined value of 4.67 reported by Hockey (1990).

Despite the apparent good agreement of the power number, the ability of the CFD
method to predict the power number demands further scrutiny. Here, the impeller is
represented by thin surfaces, which is the preferred approach since the grid refinement
necessary to resolve the blade thickness would lead to an undesirably large number of
cells in the grid. Other workers (e.g. Bartels et al., 2002) have also used thin surfaces to
represent the impeller disc and blades, and have also reported good predictions of the
power number. However, it has been shown in a number of studies (e.g. Bujalski et al.,
1986; Rutherford et al., 1996) that the power number of a Rushton turbine is actually
dependent on the thickness of the blades, and as the thickness decreases, the power
number increases. Therefore, the question arises as to what actual thickness is being
represented by a grid with thin surfaces.

In the paper by Rutherford et al. (1996), the authors present a plot (their Figure 4)
showing how power number varies with blade thickness. This plot actually corresponds
to the conditions used in the experiments of Hockey (1990), and the data point for a
speed of 300 rpm (as simulated in this study) and a thickness to impeller diameter ratio
of 0.0306 (as used by Hockey) is plotted (NP = 4.67). The plot shows that as thickness
decreases, the power number increases, with a limiting value for very small thickness of
about 6.6. In the CFD method, the impeller blades are represented by zero thickness
thin surfaces, and therefore it might be expected that the CFD simulation would yield

74

Chapter 4. CFD Simulations of Single-Phase Flow

a value of power number approaching 6.6. Instead, the power number is slightly
underestimated, at a value of 4.5. This value is likely to be a function of the grid
resolution. Here, the pressure distribution around the blade is represented by a relative
coarse grid, and this results in an underprediction of the actual pressure drop across a
very thin blade, thus yielding a predicted value of power number which is seemingly in
good agreement with measurements for a thicker blade. Thus, the apparent good
agreement of the CFD and experimental values is somewhat fortuitous. It would
perhaps be better to regard the CFD prediction of power number as having a fairly wide
error band, since the value does not correspond to impeller blades of any precise
thickness. For more reliable results, a much higher resolution grid would be required, in
which the actual blade thickness is resolved. For many applications, this will lead to
excessive computer resources. In particular, in the development of the model for gasliquid flow (see Chapter 8), there were limitations on the practical grid density.

Some simulations were also carried out using denser grids (see Section 4.4.2), although
the blades were still represented by thin surfaces in most cases. Again, the apparent
good agreement obtained should be regarded as being somewhat fortuitous. In just one
case, the actual blade thickness was represented. However, the blade thickness was only
spanned by four cells, and the effect on power number was only slight.

These comments on the prediction of power number apply equally well to the prediction
of the flow number, NQ. Again, Rutherford et al. (1996) reported that the flow number
increases with decreasing blade thickness. Thus, the apparent good agreement between
CFD and measurements is somewhat fortuitous, since the CFD result could be regarded
as an underprediction of the discharge flow of very thin blades, due to the coarseness of
the grid.

4.2.4 Modelling procedure using the multiple frames of reference method


The simulation of the standard configuration tank with a Rushton turbine was repeated
using the Multiple Frames of Reference (MFR) method. This is an approximate steadystate method which was first outlined by Luo et al. (1994) and Tabor et al. (1994). Like
the sliding mesh method, this method of simulation is capable of modelling the flow
throughout the tank including the impeller region, without reference to empirical data as
in the impeller boundary condition method. However, the Multiple Frames of Reference
75

Chapter 4. CFD Simulations of Single-Phase Flow

method avoids the need for time stepping, and thus presents the possibility of carrying
out simulations with a much shorter computation time. This feature may be particularly
important for simulating gas-liquid flow, where the computational demands are
expected to be considerably greater. For gas-liquid flow, there are approximately twice
as many equations to be solved, and convergence of the solution is slowed due to the
complexity of the multiphase physics and the requirement to allow the distribution of
the dispersed phase to evolve.

At the time when this work was carried out, the MFR method was not an option in
CFX4, and hence, the method was implemented by writing a User Fortran sub-routine.
The method has since become a standard feature, in later releases of CFX4.

To apply the MFR method, the tank is divided into two domains, a zone surrounding the
impeller and a zone for the rest of the tank, similarly to the sliding mesh method.
Instead of moving the grid in the impeller zone, two frames of reference are used for
solving the equations governing fluid motion. In the impeller zone a rotating frame of
reference is applied in which the impeller appears stationary, whereas in the bulk of the
tank, the normal inertial frame of reference is used. In the rotating frame of reference,
there is a body force term added to the momentum conservation equation, consisting of
the Coriolis and centrifugal forces according to:

B = 2 U ( x) ,

(4.5)

where is the angular velocity vector, U is the velocity vector and x is the position
vector. A steady-state solution is calculated with the impeller fixed in one position
relative to the baffles, and therefore in the equations of conservation of mass and
momentum, the terms containing derivatives with respect to time are neglected. At the
interface between the domains, the velocities are corrected by an implicit coupling
method, where at cell centres adjacent to the interface, the azimuthal velocity seen
across the interface is corrected for the change in reference frame, by adding or
subtracting r (where is the angular frequency of the impeller and r is radius at the
cell centre). Spatial derivatives of the azimuthal velocity should also be corrected, to
avoid fictitious large production terms in the equations for k and . However, this could
not be readily done in CFX4, so instead, a smoothing procedure was used at the
interface to avoid erroneous values of k and near the interface.
76

Chapter 4. CFD Simulations of Single-Phase Flow

Whereas in the sliding mesh method, the inner domain extended vertically from bottom
to surface (a choice made as a matter of simplicity), here the interface between rotating
and stationary frames of reference was set at an axial distance 0.25D from the impeller
centreline. In the radial direction, the interface between domains was set radially at
r = 0.75D. These positions of the interface between the domains were chosen according

to the recommendation of Luo et al. (1994). The simulation was carried out using the
same grid as for the sliding mesh simulations described above, with 48, 38 and 60 cell
divisions in the axial, radial and azimuthal directions respectively. The same hybrid
differencing scheme was used and boundary conditions on surfaces are the same as
before. The simulation was run until the sum of mass residuals was reduced by at least
three orders of magnitude, which required approximately 800 iterations, with underrelaxation factors set to about 0.6 for the components of velocity and 0.5 for k and .

4.2.5 Results of MFR method and comparison with sliding mesh method

Results of the simulation using the MFR method are illustrated by Figures 4.144.19.
Note that in the vector plots, the vectors in the rotating frame of reference have been
converted back to the stationary frame of reference. In most respects, the results look
very similar to those obtained with the sliding mesh method. The only noticeable
difference is that near the top of the tank (compare Figures 4.3 and 4.15) the swirling
motion is not as pronounced with the MFR results.

The MFR method was found to be much more computationally efficient, requiring only
a fraction of the number of iterations compared to the sliding mesh method. For the
same number of grid cells, the solution was obtained in a computer processing time
which was about one tenth of that required for the sliding mesh method. This is the
same level of saving in computer time as reported by Luo et al. (1994).

Some more detailed analysis has been carried out to check on the accuracy of the MFR
method. Other authors who have investigated the accuracy of this method include Luo
et al. (1994), who found better agreement with velocity measurements when compared
to a black box approach. Here, a comparison is made specifically between the sliding
mesh and MFR methods.

77

Chapter 4. CFD Simulations of Single-Phase Flow

Results have been compared for the sliding mesh method, MFR method, and the data
from Hockey (1990) and Wu and Patterson (1989). Radial profiles of the mean
velocities at several axial positions are compared, as illustrated in Figures 4.204.25.
For the mean velocity components, it is found that both methods give good agreement
with experimental data, with only small discrepancies at certain positions, and there is
little to distinguish between the sliding mesh and MFR methods in terms of accuracy.

In summary, it is found here that the MFR method provides a level of accuracy which is
similar to the sliding mesh method in prediction of the mean fluid flow. However, the
MFR method has the advantage that it allows results to be obtained in about one tenth
of the computation time. Of course, a possible disadvantage of the MFR method is that
the impeller is frozen at one position, whereas the flow pattern might vary depending
on the relative positions of the impeller and baffles. Therefore, in some circumstances it
might be necessary to calculate the flow for several different impeller positions. Also, if
the details of the transient impeller-baffle interaction are important, such as in
calculating the mixing of a tracer, then the MFR method might not be suitable.
However, the modelling of gas-liquid flow using impellers with diameter ~ 1 T (as in
3
the simulations in this study), it is fair to assume that the impeller-baffle interaction will
be relatively weak, and in this case the average flow field for one impeller position
should suffice, for purposes such as determining the gas distribution in the vessel.
Therefore, given the much greater computational efficiency of the MFR method, this
approach is particularly attractive for modelling gas-liquid flow, where there will be
considerable additional computational requirements.

4.3 Additional simulations of a tank stirred by a Rushton turbine


Several additional simulations were carried out to investigate various aspects of the
flow in a tank stirred by a Rushton turbine, including the ability of the CFD method to
predict the trailing vortices behind impeller blades, the pressures in the impeller region,
and the turbulence in the tank. Simulations with higher grid density were carried out
with the aim of improving the accuracy of the CFD simulations with respect to these
features of the single-phase flow.

78

Chapter 4. CFD Simulations of Single-Phase Flow

One aspect which was investigated was the pressure distribution on the surface of
impeller blades. Experimental measurements were available for comparison (Manning,
1993), however, the measurements were made in a square tank. Therefore, a new multiblock grid was generated to match the geometry, using the CFX Meshbuild software, as
illustrated by Figure 4.26. The tank was a baffled square tank with side length of 0.58 m
and fluid depth of 0.58 m. The diameter of the disc turbine impeller was 185 mm, and
the impeller was located at a height of 0.20 m from the base of the tank. The impeller
speed was 263 rpm and the working fluid was water, giving an impeller Reynolds
number of 1.5 105. The finite volume grid consisted of a total of 186,360 cells. To
reduce computational requirements, one half of the tank was modelled assuming
symmetry and applying a periodic boundary condition for the planes at the high and low
ends of the grid in the azimuthal direction. Baffles, impeller disc and blades were
modelled as thin surfaces. The Multiple Frames of Reference method was used, and
boundary conditions and solution procedure were similar to the previous simulation
using the MFR method. Results were examined in terms of the pressures on the front
and rear surfaces of the impeller blades, as discussed in Section 4.6.2.

To obtain improved predictions of the pressures on blade surfaces, another approach


was taken with the aim of obtaining greater grid resolution around the impeller. Thus,
the flow domain was limited to a region surrounding just one impeller blade (and not
extending to the wall, bottom or surface). A 60 section including one impeller blade
was modelled, covering the same zone as the rotating domain in the simulation of half
of the tank, namely an axial extent of 0.25D from the impeller centreline, and a radial
extent of r = 0.75D. A grid of 291,840 cells was used, which provided resolution of the
blade by a surface grid 24 40 cells (see Figure 4.27). Periodic boundaries were applied
in the azimuthal direction. The upper and lower horizontal boundaries were specified as
inlets, with fixed boundary conditions for velocity, k and , based on an interpolation of
the values obtained in the simulation of half of the tank. Since fluid is discharged in a
radial/tangential direction, the outer vertical boundary was treated as a mass flow
boundary. A steady-state simulation was run until the mass residual was reduced by
three orders of magnitude. Predicted pressures were compared against experimental
measurements, as discussed in Section 4.6.2.

79

Chapter 4. CFD Simulations of Single-Phase Flow

Some further development has also been carried out to investigate the effect of grid
resolution, without limiting the simulation to just the impeller zone. Here, the aim was
to improve the accuracy of the CFD simulations, with respect to the trailing vortices and
prediction of turbulence. Further simulations were set up for just a 60 sector with a
single blade and single baffle, assuming periodicity in the tank. This allowed the grid
density to be increased without excessively increasing the total number of cells, so that
the computational demand was kept within the limits of available computer resources.
This approach implies that there are six baffles in the tank rather than four. However,
the purpose of the baffles is to stop swirl, and it is expected that once there is sufficient
baffling to prevent swirl, the flow pattern will not change much with the addition of
more baffles, except in the local region surrounding the baffles.

For a 60 sector, the grid has been set up with two different grid densities, being a grid
with about 56,000 cells (equivalent to 168,000 cells over half of a tank), and a denser
grid with the total number of cells increased almost by a factor of 8, to about 440,000
cells, which represented the practical upper limit at the time in terms of available
memory space. Thus, the first grid is similar in density to the original simulations of a
180 sector, while the second grid is considerably denser. These finite volume grids, as
illustrated in Figures 4.28 and 4.29, were generated as multi-block grids using the CFX
Meshbuild software, with a non-uniform grid aimed at optimising the resolution around
the impeller. Also, with the aim of increasing accuracy of results, the grid was
constructed to resolve the actual thickness of the impeller blade and the disc, whereas
previously these were thin surfaces. In the lower density grid, the blade surface is
resolved by 8 9 cells and azimuthally there are 10 cells on either side of the blade
(similar to the simulation of a 180 sector), while in the higher density grid, the blade
surface is resolved by 18 20 cells, and there are 20 cells on either side of the blade.
The grid on the surface of the blade and disc are illustrated in Figures 4.30 and 4.31.
The higher density grid is equivalent to a grid of about 1.3 million cells covering half of
a tank.

In developing the CFD model for higher grid resolution, it was considered that the main
region requiring the denser grid would be the impeller region. Ideally, to limit the
computational demand, it would be desirable to increase the grid density in the impeller
80

Chapter 4. CFD Simulations of Single-Phase Flow

region as much as possible without adding grid cells elsewhere in the tank. However,
several attempts were made to create grids of this type, but it was found that when the
difference in cell sizes between impeller region and the bulk became too large,
convergence could not be achieved. It seems that a smooth transition of cell sizes is
needed, and by trial and error the grid as illustrated by Figure 4.29 was obtained as a
compromise. The need for localised grid refinement might be addressed by using a
method such as an embedded grid, as has been reported by Ng et al. (1998) in their
simulations of stirred tanks using the STAR-CD code. However, the modelling in the
present study was restricted to the use of CFX4, which does not have the option of
embedded grids.

For the simulations with the 60 tank sector, tank dimensions were the same as in the
previous simulations of the tank used by Hockey (1990), and the impeller speed was
again 300 rpm. The MFR method was employed to keep computation time to a
minimum. The boundary conditions and solution method remained similar to previous
simulations, with the under-relaxation factors again set to 0.6 for velocity components
and 0.5 for turbulence quantities. With the lower density grid, convergence was again
achieved in about 800 iterations. However, it was found that with the higher density
grid, convergence was somewhat slower, and about 1400 iterations were required to
reduce the sum of mass residuals by three orders of magnitude. Therefore, for denser
grids, there is additional computational demand beyond that due to the additional
number of cells.

Results for the simulations with the 60 sector are illustrated in Figures 4.324.34. It
was found that with the grid of ~56,000 cells, the predicted flow field is generally very
similar to the previous results obtained for a 180 sector, using the MFR method. With
the higher density grid of ~440,000 cells, the flow pattern and velocities are again
generally the same (Figures 4.354.37), but greater detail is obtained. Also, as discussed
in the following section of this chapter, greater detail is obtained near the impeller
blade, which allows for better prediction of the trailing vortices, and predictions of
pressure and turbulence near the impeller are improved.

81

Chapter 4. CFD Simulations of Single-Phase Flow

The values of y+ in the cells adjacent to walls were checked by plotting on the walls, as
shown in Figures 4.34 and 4.37. It is seen that as the grid density increases, y+ values
decrease, but still remain within the proper range (i.e. about 11300) at the higher grid
density. As pointed out by Bartels et al. (2002), with further refinement of the grid a
point will be reached where the wall function approach is no longer valid, and then
another approach will be needed, such as the use of the low-Reynolds number k-
model, where wall functions are not used.

For the simulations with the 60 sector, axial and radial components of velocity have
been compared against the experimental measurements of Hockey (1990), for selected
radial profiles at a position half way between baffles, as shown in Figures 4.384.43. It
can be seen that, in the simulation with the lower grid density, the values of these
velocity components are very similar to the previous results for a 180 sector using the
MFR method, with generally good agreement with experimental data, but with same
deviations from the data at various points. With the higher grid density of ~440,000
cells, the agreement with experimental data is again good, and predictions are actually
improved at several places, particularly in the radial velocity component at the impeller
tip and at a height of 0.12 m, towards the wall. Hence, it may be concluded firstly, that
the use of a 60 sector, with the implied assumption of six baffles rather than four, does
not significantly affect the flow pattern or velocities in the tank. Secondly, increasing
the grid density leads to improved agreement with experimental data for velocities.
However, it has also been shown here that quite good predictions of the velocity field
can be obtained at lower grid densities with substantially reduced computational
demand. This approach, using a 60 sector and about 60,000 cells, provides a means of
minimising grid size, and this approach was pursued in the modelling of gas-liquid
dispersion, to reduce the computation time for the two-phase calculations.

4.4 Prediction of detailed flow around the impeller


Thus far, the discussion has concentrated on prediction of the bulk flow pattern.
Additional analysis of the simulation results has been carried out to assess the ability of
the model to predict finer details of the flow around the impeller. To investigate these
features, and the effect of increased grid resolution, additional simulations were carried
out, as outlined above.
82

Chapter 4. CFD Simulations of Single-Phase Flow

4.4.1 Prediction of the trailing vortices

The production of trailing vortices, due to separation of flow over impeller blades, has
been described in Chapter 2. This aspect of the tank hydrodynamics is considered
important in the production of turbulence, and is relevant to applications such as gas
dispersion. Gas is drawn into the vortices, and depending on the prevailing operating
conditions, the accumulation of gas in the vortices may lead to the formation of
ventilated cavities of a number of types. The trailing vortices have been studied
experimentally, by flow visualisation methods (e.g. vant Riet & Smith, 1975), and by
laser doppler velocimetry measurements (e.g. Yianneskis et al., 1987).

In modelling of stirred tanks by CFD, some simulation methods reported in the


literature have apparently been incapable of predicting the trailing vortices. In the
impeller boundary condition method, azimuthally-averaged data is used at the impeller
boundary, so flow structures between blades are not resolved. Also, Brucato et al.
(1998a) reported that trailing vortices were not resolved by the Inner-Outer method.
With the sliding mesh method, however, the positions of the trailing vortices were first
identified by Tabor et al. (1996). They provided a broad indication of the position of the
trailing vortices by producing a map of various isosurfaces of the magnitude of vorticity
around the impeller blade. The simulation results of this study have also been examined
in terms of the ability to predict trailing vortices (further details can be found in Rigby
et al., 1998a; 1998b).

The pressure field in the plane through the centre of the impeller has been plotted in
Figure 4.5, indicating the low pressure region formed behind each impeller blade. The
pressure field is also plotted for an axial-azimuthal plane at the blade edge in Figure
4.44. Here, vectors have been added to the plot. These vectors are calculated in the
frame of reference of the rotating impeller, revealing a region of recirculation in the
wake, originating from the upper and lower edges of the blade.

For each pressure contour at the rear of the blade, the point furthest out from the
impeller is roughly aligned with the outflow from the impeller. A locus formed by
joining together these low pressure peaks forms a basis for predicting the axis of the
trailing vortices. This can be done for angles behind the blade up to about 30, but
83

Chapter 4. CFD Simulations of Single-Phase Flow

beyond that point the vortices are destroyed and the flow merges with the bulk flow.
The axis of the upper vortex has been plotted in Figure 4.45, and comparison with the
published loci of Yianneskis et al. (1987) and Stoots and Calabrese (1995) shows
reasonable agreement, although some difference is seen with the earlier visualisation of
vant Riet and Smith (1975).

Hence, it can be seen that using the sliding mesh method and a grid of about 100,000
cells over half of the tank, it is possible to predict the circulating motion of the trailing
vortices, along with the correct axis of the vortices. However, grid resolution remains
limited around the impeller blades, so that it remains somewhat doubtful that the full
effect of the vortices is predicted, e.g. the correct values of velocity gradients, the
resulting vorticity, and the values of pressures.

As described above, a further simulation of a 60 sector with a grid of about 440,000


cells was carried out, providing improved grid resolution around the impeller. Using
this grid, the trailing vortices are resolved in much greater detail. Figure 4.46 shows a
vector plot in a plane normal to the axis of the vortices, revealing the circulation in
greater detail. Figure 4.47 shows a plot of streamlines to visualise the upper trailing
vortex. With higher resolution, the simulations would be expected to provide better
prediction of the high velocity gradients which are evident in these vortices, and related
to prediction of these gradients, better prediction of pressures and turbulence production
near the impeller blades should be expected.

4.4.2 Prediction of pressures around the impeller blades

Some investigations were carried out regarding the ability of the CFD method to
correctly predict the pressure field in the tank, particularly pressures at the surfaces of
the impeller blades, and the pressure contours in the region of the impeller. The
pressures on the blade surfaces are of interest for calculating the torque and hence the
power number of the impeller. More detailed information about the distribution of
pressures may be of interest for investigating the mechanical stability of the agitator,
since uneven pressure distribution may cause bending moments and stresses on the
blades. Knowledge of the pressure at impeller blade surfaces is also important in
applications where a gas-inducing impeller is employed. In such cases, gas is drawn
down from the headspace through a hollow shaft connected to hollow blades with
84

Chapter 4. CFD Simulations of Single-Phase Flow

orifices on the trailing sides, and the flow of gas obtained depends on the pressure
differential between the low pressure side of the blade and the headspace.

Work was carried out to compare CFD predictions with available experimental data for
the pressure distribution on the leading and trailing sides of the blade of a Rushton
turbine. Pressure measurements were made by Manning (1993) using an impeller fitted
with a hollow blade. The hollow blade had an array of 46 pressure tappings evenly
spaced over one vertical face. For each experiment, all tappings were sealed and
machined flush except for a single tapping, where the pressure was transmitted via a
continuous liquid line though the blade and a hollow shaft to an external differential
pressure cell (further details of this work can be found in Lane et al., 2001). Pressures
were reported as the difference between the dynamic pressure and the stationary
hydrostatic pressure, and were expressed as a pressure coefficient, CP, defined
according to:
CP =

p l g ( x x0 )
,
1 2
vtip
2

(4.6)

where the hydrostatic pressure at a height x, relative to the reference pressure at x0, has
been subtracted from the total local pressure, p, and vtip is the impeller tip velocity.
Experimental values of the pressure coefficient are shown in Figures 4.48 and 4.49 for
the leading and trailing blade faces, respectively.

The tank used by Manning (1993) in these experiments was simulated by CFD as
described in Section 4.5. For the simulation of half of the tank, predicted pressure
coefficients are shown in Figures 4.50 and 4.51. Good agreement was obtained on the
leading face. On the trailing face, the pattern of the pressure coefficients is similar to
that obtained experimentally, and it reflects the adjacent roll vortices, but the agreement
of values is rather poor in some places, being too low in the middle of the blade. With
the aim of improving the simulation results, the region surrounding one blade was
modelled, as described in Section 4.5, giving a higher grid resolution around the blade.
The predicted pressure coefficients are shown in Figures 4.52 and 4.53. It can be seen
that despite the much higher grid resolution, pressure coefficients show only slightly
better agreement with data for the centre region of the blade.

85

Chapter 4. CFD Simulations of Single-Phase Flow

Despite the errors in predicting local pressure coefficients, the prediction of power
number, based on the overall pressure difference over the impeller blade according to
equations 4.3 and 4.4, was quite good. The simulation of the 180 degree section of the
tank gave a power number of 5.2, while the simulation of the zone surrounding one
impeller blade gave a power number of 5.0. By comparison, experimental
measurements with a torque meter gave a power number of 5.1.

The prediction of the pressure field in the liquid surrounding the impeller blades has
also been investigated, and comparison is made with the experimental measurements of
vant Riet and Smith (1975). Figure 4.54 shows the pressure coefficients in a horizontal
plane through the axis of the trailing vortex according to vant Riet and Smith, as
obtained at a similar Reynolds number to that in the simulations. The experimentallydetermined pressure coefficients can be compared with those obtained from the
simulation of the 60 degree tank section at a grid density of ~56,000 cells, as plotted in
Figure 4.55. In the simulation results, the underpressure in the region of the trailing side
of the blade, where the trailing vortices occur, only reaches a value of CP of about -0.8,
while according to vant Riet and Smith (1975), the minimum pressure coefficient is
about -3.5. Also, the minimum pressure in the simulation results is found at the blade
surface, while the experimental data indicate that the minimum pressure is off the blade,
at the core of the vortex. Hence, the prediction of pressure in this zone is quite poor in
the CFD model. In modelling of the impeller with the hollow blade, the minimum
pressures were also on the surface, which may be part of the reason why pressure
coefficients on the surface did not match very well with data.

The pressure coefficients in the horizontal plane of the vortex axis are plotted in
Figure 4.56 for the results obtained with a 60 tank sector with the higher grid density of
~440,000 cells. With this higher grid resolution, the minimum pressure coefficient is
now about -1.0, and the shape of the pressure contours matches those of vant Riet and
Smith (1975) more closely. However, the minimum pressure still lies on the blade, and
the minimum pressure is still only about

of the measured value. The calculation of

power number for this high resolution grid gave a value of 4.55, only slightly higher
than for the results using a thin surface.

86

Chapter 4. CFD Simulations of Single-Phase Flow

Hence, it is found that the pressures are not very well predicted at the trailing side of the
impeller blades, both on the blade and in the adjacent fluid where the trailing vortices
form, even at the highest grid resolutions tested. To improve the predictions, it may be
necessary to use an even finer grid, in order to fully resolve the velocity gradients
associated with the trailing vortices. Improvement might also be possible by using a
different turbulence model such as the Reynolds stress model, to allow for the
anisotropic turbulence in this region. Improvement might also be possible by better
resolution of the boundary layer on the blade, possibly avoiding the use of wall
functions. However, in terms of the gas-liquid modelling considered in this study, such
approaches were not considered feasible due to the additional computational demand
which would be required.

It can be noted here that while the details of the pressure distribution around the
impeller blades have not shown good agreement in some places, it has still been
possible to calculate the power to a fair degree of accuracy based on the pressure
difference over the blades of the Rushton turbine, both for the configurations of
Manning (1993) and Hockey (1990) (see also Section 4.2.3). For the Lightnin A315, the
prediction of power draw was also fairly reasonable (Section 4.6). It would seem that
the CFD method is able to calculate the overall pressure drop over the blades fairly
adequately as part of a converged solution of the momentum conservation equation,
even though the pressures are locally smeared out and the grid resolution is not high
enough to reveal the correct details of the local pressure distribution. However, it should
again be emphasised that the agreement of the power estimates with experimental
values is somewhat fortuitous (as discussed in Section 4.4.2), since the power varies
with blade thickness, but, with one exception, the actual blade thickness has not been
resolved by the grid. Thus, all such estimates of power can be expected to have a fair
margin of error associated with them.

The predicted values of pressure, particularly near the trailing side of the impeller
blades, are important in modelling gas-liquid flow, since the amount of gas collected
here and the related size of cavities will depend on these pressures. The results here
indicate that, at the level of grid resolution which is practical for development of the
two-phase model, these pressures will not be as low as in reality. This must be taken
into account in the modelling of gas cavities (as described in Chapter 5), where
87

Chapter 4. CFD Simulations of Single-Phase Flow

adjustable constants are applied in order to match the size of cavities and the related
gassed power draw with experimental data. These constants may in fact be sensitive to
grid density, since the predicted pressure field is grid sensitive.

4.5 Prediction of turbulence


Often in CFD studies, accuracy of results has only been assessed with respect to mean
velocities. However, the turbulent parameters such as the turbulent kinetic energy, k,
and the rate of dissipation of turbulent energy, , are also of great importance in
modelling of mixing tanks for a range of applications, since various mathematical submodels are typically formulated as functions of k or . Examples include modelling
equations for micromixing or heat transfer to surfaces. In modelling gas-liquid flow, the
values of k and are used in sub-models such as equations for turbulent dispersion, and
break-up or coalescence of bubbles. Hence, as well as comparing against mean velocity
measurements, it is also important to obtain accurate predictions of turbulence
quantities. Results have been compared here against data for k and obtained from a
couple of sources (Wu & Patterson, 1989; Deglon et al., 1998). These data were
obtained using a tank stirred by a Rushton turbine, however, measurements were made
using different tank sizes and impeller speeds. Here, it is assumed that comparison can
be made with values that have been scaled by making them non-dimensional with
respect to appropriate reference quantities. The turbulent kinetic energy, k, was scaled
by the square of the impeller tip speed, and the energy dissipation rate, , was scaled by
the overall power input per unit volume.
Comparison with experimental k and values has been carried out for the results from
the simulation of a 180 tank section using the sliding mesh and MFR methods. Values
at two positions in the tank are illustrated by Figures 4.57 and 4.58. In the impeller
discharge stream, it was found that values of turbulence parameters are substantially
underpredicted by both methods. There is slight improvement with the MFR method,
though predicted values near the impeller tip are still only about 50% of measured
values. Closer to the wall, the two simulation methods give similar results and
agreement with measurements improves. At the position in the bulk of the tank below
the impeller, values of k according to the simulation methods are in fair agreement with

88

Chapter 4. CFD Simulations of Single-Phase Flow

experimental data at positions towards the wall, but simulation predictions separate
from the data near the centre. Comparison with experimental values of in the bulk of
the tank indicates substantial underprediction by both simulation methods.
To some extent, the accuracy of the experimental data for might be called into
question, since this parameter could not be measured directly, but depended on
estimates of the integral scale of turbulence (see Wu & Patterson, 1989). However, it is
also possible to consider the total turbulent energy dissipation in the tank as obtained by
integration of over the liquid volume of the whole tank, VTOT, according to:
P = 0 TOT dV ,
V

(4.7)

where V is volume. In fully turbulent flow, this term should account for nearly all of the
energy dissipation (with a small additional energy dissipation due to mean flow
gradients), so using this value it should be possible to calculate the dimensionless power
number. However, applying this calculation with the results from the sliding mesh
method gave a value for the power number of only 2.5 (compared with 4.7
experimentally). This indicates that the values of in the simulation results are
substantially underpredicted in many positions.

As outlined in Chapter 3, in other simulations reported in the literature (e.g. Lee et al.,
1996; Jaworski et al., 1997; Ng et al., 1998) using the sliding mesh or MFR methods,
underprediction of turbulence has also been reported. For example, Ng et al. (1998)
found that even with a grid of 400,000 cells, the turbulent kinetic energy in the impeller
discharge stream was still underpredicted by 50%. A full explanation for the
underprediction of turbulence has not yet been established, however according to
Bartels et al. (2002), insufficient grid resolution around the impeller is likely to be a
contributing factor, since it leads to poor prediction of the large velocity gradients
associated with the trailing vortices. If the CFD model fails to resolve the velocity
gradients fully, then the production terms in the k and equations will be too low, since
these are calculated from gradients of mean velocity. It may also be that the use of the
k- model is part of the problem. Bartels et al. (2002) obtained much better predictions

of k using direct numerical simulation (DNS), thus avoiding the use of a turbulence
model. However, DNS remains impractical for most CFD simulations of stirred tanks,
especially at high Reynolds numbers. Other turbulence models such as the RNG k-
89

Chapter 4. CFD Simulations of Single-Phase Flow

model have been tested (e.g. Jaworski et al., 1997), but have not shown any
improvement.

It has been found that the Multiple Frames of Reference method gives slightly improved
predictions of k and in the impeller stream compared with the sliding mesh method,
without increasing the grid density. This is probably due to better convergence, since
600 iterations were calculated for a single impeller position. With the sliding mesh
method, the level of convergence may be lower, since only a small number of iterations
(e.g. 2030) are typically used at each time step, before advancing the impeller to the
next position.

The effect of grid resolution on prediction of turbulence quantities was investigated by


considering the results obtained in the simulations of a 60 tank sector. Figures 4.59 and
4.60 show plots of turbulent kinetic energy in a horizontal plane intersecting the
impeller centre, where it is seen again that the highest turbulence is associated with the
trailing vortex, similarly to Figure 4.8, and it is seen that the maximum value of k
increases with increased grid density. Figures 4.61 and 4.62 provide a comparison of the
radial profiles of k and at the impeller centre line, as obtained experimentally
according to Wu and Patterson (1989), and as obtained in the simulations of the 180
sector and the 60 sector with the lower and higher grid densities. It was found that
values obtained with the 60 tank sector at the lower grid density are similar to those
obtained for the 180 sector. At the higher grid resolution, substantially increased values
were obtained in the impeller discharge stream, although k and are still underpredicted
by as much as ~35% and ~25% respectively. It is likely that further increases in grid
density would lead to even better prediction of turbulence quantities. On the other hand,
Bartels et al. (2002) reported that with a grid of about 2 million cells using the k-
model, k was still underpredicted. Thus, there is probably also some effect related to the
choice of turbulence model. It must be taken into account that the k- model is not very
well suited to modelling anisotropic turbulence, as has been shown to be present in the
flow in the impeller region (Wu & Patterson, 1989; Bakker, 1992), and the model is
also not well suited to swirling flows. Therefore, to obtain more accurate predictions of
the turbulence, it may be necessary to use a different turbulence model such as the
Reynolds stress model.
90

Chapter 4. CFD Simulations of Single-Phase Flow

In terms of modelling gas-liquid flow, it was not practical to use a very dense grid, nor
was it practical to use a turbulence model such as the Reynolds stress model, due to the
high computational demand. Thus, in the development of the two-phase model, the
deficiencies in prediction of k and have remained. This points to the need for caution,
since in gas-liquid flow it is necessary to use these values in other modelling equations.

Despite the underprediction of turbulent kinetic energy and turbulent energy dissipation
rate, the earlier sections of this chapter have demonstrated that it is still possible to
predict the velocity field with a reasonable level of accuracy. The explanation for this
would be that, although the individual values of k and are too low in most places, the
CFD method still obtains reasonable values of the ratio k2/, upon which the turbulent
viscosity is based. The turbulent viscosity is the only input from the turbulence model
into the momentum conservation equation, and since this is being predicted fairly well,
the correct velocities can be obtained.

Although the values of turbulence parameters are too low, it can be seen from the results
(e.g. Figures 4.64.9) that the general pattern of turbulence according to the simulations
appears quite reasonable, i.e. the highest turbulence levels are in the trailing vortices and
the impeller discharge stream, with turbulence decaying away to lower levels in the tank
bulk. This has suggested that the underprediction of turbulence could be dealt with by
applying global correction factors to k and where they are used in other equations
such as in modelling turbulent dispersion or bubble coalescence. The correction factors
were estimated based on comparisons with the values expected based on data in the
literature. Thus, these correction factors represent an empirical input to the model,
which cannot be avoided at present. Unfortunately, this approach can only offer limited
improvement, since the extent of underprediction varies depending on the position, but
only single global correction factors were applied.

4.6 Modelling of the Lightnin A315 impeller


In addition to the Rushton turbine, a model was also set up for a baffled tank stirred by a
Lightnin A315 impeller. The tank geometry in the model has been set up to be
essentially the same as that used by Bakker (1992), who obtained some experimental
91

Chapter 4. CFD Simulations of Single-Phase Flow

data for a tank with this impeller type, both for single-phase flow and for gas-liquid
dispersion. Bakker also modelled this tank, although the impeller was represented in his
modelling method using an impeller source term, or black-box approach. Here, the
impeller has been modelled explicitly.

The tank considered here has a diameter, T, of 0.444 m, and is stirred by an A315
impeller of diameter 0.4T mounted at

of the tank height. There are four baffles of

width 0.1T equally spaced at the walls. The tank is also fitted with a ring sparger located
below the impeller (this is included since the same mesh was used for gas-liquid
modelling, as reported in Chapter 8).

The grid-generating software CFX Meshbuild (CFX4 Solver Manual, 2002) was used to
generate a structured multi-block finite volume grid, including an explicit representation
of the A315 impeller. Compared to the Rushton turbine, grid generation is not as
straightforward, since the blades are wide and curved and have an overlap. It is
particularly difficult to generate the grid near the impeller hub, since due to the
cylindrical geometry, the mesh needs to taper as it converges at the inner radius. If the
mesh diverges too far from orthogonality, then convergence problems may result. Care
was therefore taken to ensure that the mesh generated did not contain cells with poor
orthogonality or excessive skew or taper. Due to the overlap of blades, it was not
possible to divide the tank into periodic sections, and therefore the whole of the tank
was modelled. A mesh consisting of 52 blocks was generated in Cartesian coordinates,
but subsequently transformed into cylindrical coordinates. The mesh is illustrated by a
surface plot in Figure 4.63. A closer view of impeller, as generated in the Meshbuild
program, is shown in Figure 4.64. The blade thickness was neglected, and instead the
impeller blades are treated as thin surfaces.

Initially, the multiple frames of reference method was applied for the impeller motion,
but several attempts at running the model in this way failed to give a converged
solution. It is believed that the problem relates to the non-orthogonality of the grid.
Instead, it was necessary to use the sliding mesh method. The comparatively high
computational demand of this method led to the restriction of the grid cell number to
~180,000 cells, which covers the whole of the tank, so that grid density was not as high

92

Chapter 4. CFD Simulations of Single-Phase Flow

as in simulations of the Rushton turbine. The simulation procedure is similar to that


outlined in Section 4.2.2, using similar boundary conditions and again employing the k-

turbulence model.
Results are illustrated by Figures 4.654.68. As indicated by the vector plot in Figure
4.65, the flow pattern consists of a single recirculating loop, with downward axial flow
produced by the impeller. Figure 4.66 shows the pressure distribution in a horizontal
slice through the impeller. Figure 4.67 shows the pattern of turbulent kinetic energy,
with highest values in the discharge of the impeller and at the centre of the recirculation
loop at approximately the level of the impeller. The pattern of turbulent energy
dissipation is similar, as shown in Figure 4.68.

From the simulation results, the dimensionless power and flow numbers have been
calculated, following the procedure outlined in Section 4.2.3. The power number, NP,
and the flow number, NQ, have values of 0.76 and 0.74 respectively, according to
Bakker (1992), while according to Fraser et al. (1993), the values are 0.75 and 0.73
respectively. Based on the CFD results, the power number was estimated as 0.9, which
is about 20% higher than reported values, while the flow number was estimated to be
0.82, which is about 10% higher than reported values. Thus, for these global quantities
the agreement is not as good as with the Rushton turbine, although it can still be said
that there is fair agreement between CFD results and experimental data. The poorer
agreement might possibly be due to the non-orthogonality of the finite volume mesh for
the A315 impeller. On the other hand, a certain margin of error can be expected, since
the impeller was represented by thin surfaces, and the power and flow numbers of
impellers are known to depend on the actual thickness of the impeller blades (as
discussed in Section 4.2.2).

Results for the simulation of the A315 have been compared against experimental data
for the axial component of the mean and r.m.s. turbulent fluctuating velocity, as
obtained by Bakker (1992) using laser doppler velocimetry measurements, at an
impeller speed of 360 rpm. Radial profiles of the mean axial velocity at four heights in
the tank are compared in Figures 4.694.72. Compared to Bakkers data, there is quite
good agreement, though some discrepancies can be noted. In particular, in the discharge
flow immediately below the impeller (Figure 4.70), the data points of Bakker indicate
93

Chapter 4. CFD Simulations of Single-Phase Flow

velocities 3040% larger than in the CFD results. However, integration of the CFD data
to estimate the flow number has given a flow number of 0.82, in comparison to the
value of 0.750.76 as quoted by Bakker (1992) and Fraser et al. (1993). If the discharge
axial velocities were as large as those in the experimental data, then the calculated flow
number will be even higher. Therefore, the experimental velocity data are not consistent
with the flow number as reported by Bakker and Fraser et al. This suggests that the
LDV data may have overestimated the liquid velocities in the immediate discharge of
the impeller, and the velocities predicted by the CFD simulation are probably more
accurate than the LDV values. In summary, there is a good consistency between the
results using CFD and available experimental data, but the accuracy of the experimental
data is somewhat questionable.

The ability of the model to predict turbulence in a tank stirred by an A315 impeller has
also been assessed, by comparing experimental LDV data with simulation results for the
root mean square turbulent velocity, u0, which is obtained from predicted values of k
according to:
u 0 = 2k .

(4.8)

Again, the data for comparison comes from Bakker (1992), who reports measurements
of the axial component of the r.m.s. turbulent velocity. For purposes of comparison with
the simulation results, isotropic turbulence is assumed and the overall r.m.s. velocity is
calculated according to:
u 0 = u 2 + v 2 + w 2 = 3u .

(4.9)

Radial profiles of u0 are compared in Figures 4.734.75, from which it may be


concluded that, overall, there is fair agreement between simulation and measurements.
Agreement is particularly good in the immediate discharge of the impeller, and in other
places the values generally vary by about 30%. The values in the experimental data
would themselves have some error band associated with them, although the error was
not estimated by Bakker (1992). Also, the accuracy of the simulation values would be
limited by the assumption of isotropic turbulence, since measurements by Bakker near
the impeller have indicated that the turbulence is anisotropic.

94

Chapter 4. CFD Simulations of Single-Phase Flow

Predicted values of the turbulent energy dissipation rate have been assessed by
integrating over the whole volume of the tank, in accordance with equation 4.6, to
obtain the total turbulent energy dissipation. A value of 131 W was obtained from the
simulation results, which compares very well with a power input of 132.5 W, as
obtained using equation 2.1 and assuming a power number of 0.75.

Hence, it is found that when the CFD simulation method is applied to a tank stirred by a
Lightnin A315 impeller, quite good predictions of k and are obtained within a certain
error band, in contrast to the substantial underprediction found with the Rushton
turbine, and whereas with the Rushton turbine very dense grids seem to be needed, here
the turbulence seems to be predicted fairly well with a grid of only ~180,000 cells
covering the whole tank. Probably the difference relates to the difference in flow
patterns in the impeller region. With a Rushton turbine, strong trailing vortices form on
each blade, and associated with these are very high velocity and pressure gradients.
Unless the grid density is very high, the CFD model will have trouble in resolving these
gradients and the associated turbulence production. In the case of the A315 impeller, the
hydrofoil shape leads to flow generation with much lower shear rates (Fraser et al.,
1993), and trailing vortices would be comparatively weak. Therefore, a lower grid
resolution seems adequate for predicting turbulence generation near the impeller.

4.7 Conclusions
Development of the CFD method for single-phase flow in stirred tanks has
demonstrated that the method can, in general, give good predictions of the velocities,
flow patterns (including trailing vortices), and power and flow numbers. Similar results
can be obtained using either the sliding mesh or MFR method. However, the MFR
method requires only about one tenth of the computation time.

The simulation method was also found to have some deficiencies. It was found that the
pressures at the trailing side of the blades of the Rushton turbine were not as low as
indicated by experimental data, and turbulence quantities were found to be
underpredicted at many positions. It was demonstrated that improvements in predicting
turbulence quantities and pressure can be achieved through use of a higher grid density.
However, the practical density of the finite volume mesh is limited by considerations
95

Chapter 4. CFD Simulations of Single-Phase Flow

such as available computer resources. In two-phase calculations, the additional


computational demand puts further limits on grid refinement. Therefore, the deficiencies
in prediction of turbulence quantities and pressure field cannot be removed, and instead,
it has been necessary to take these into account in the two-phase model. The two-phase
model must therefore contain adjustable constants which may need to be changed if the
grid resolution is substantially increased. This is discussed further in Chapter 8.

96

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.1. Surface plot of the finite volume mesh for the tank stirred by a Rushton turbine, illustrating
rotation of the inner domain as carried out in the sliding mesh method.

97

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.2. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method: mean
velocity vectors (m/s) in a vertical plane intersecting the centre of the tank, half way between the impeller
blades and half way between baffles.

98

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.3. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method: mean
velocity vectors (m/s) in a horizontal plane near the top of the tank (0.27 m from base).

Figure 4.4. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method: velocity
vectors (m/s) in a horizontal plane through the centre of the impeller.

99

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.5. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method: pressure
distribution (Pa) in horizontal plane intersecting the centre of the impeller.

Figure 4.6. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method:
distribution of turbulent kinetic energy, k (m2/s2), in a vertical plane through the centre of the tank, half
way between the impeller blades and half way between baffles.

100

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.7. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method:
distribution of turbulent energy dissipation rate, (W/kg), in a vertical plane through the centre of the
tank, half way between the impeller blades and half way between baffles.

Figure 4.8. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method:
distribution of turbulent kinetic energy, k (m2/s2), in a horizontal plane through the centre of the impeller.

101

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.9. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method: distribution
of turbulent energy dissipation rate, (W/kg), in a horizontal plane through the centre of the impeller.

Figure 4.10. Simulation of the tank stirred by a Rushton turbine using the sliding mesh method: values of
the dimensionless distance to the wall in adjacent cells,
160.

102

+
y adj
, as plotted on walls. Values range from 15

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.11. Radial profiles of the time-averaged axial velocity component: results using the sliding mesh
method (+) and comparison with data of Hockey (1990) (*).

103

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.12. Radial profiles of the time-averaged radial velocity component: results using the sliding
mesh method (+) and comparison with data of Hockey (1990) (*).

104

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.13. Radial profiles of the time-averaged tangential velocity component: results using the sliding
mesh method (+) and comparison with data of Hockey (1990) (*). Measurements at 0.12 m and 0.18 m
are not available.

105

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.14. Simulation of the tank stirred by a Rushton turbine using the Multiple Frames of Reference
(MFR) method: mean velocity vectors (m/s) in a vertical plane intersecting the centre of the tank, half
way between the impeller blades and half way between baffles.

106

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.15. Simulation of the tank stirred by a Rushton turbine using the Multiple Frames of Reference
(MFR) method: mean velocity vectors (m/s) in a horizontal plane near the top of the tank (0.27 m from
base).

Figure 4.16. Simulation of the tank stirred by a Rushton turbine using the Multiple Frames of Reference
(MFR) method: velocity vectors (m/s) in plane through the centre of impeller.

107

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.17. Simulation of the tank stirred by a Rushton turbine using the Multiple Frames of Reference
(MFR) method: pressure distribution (Pa) in plane at the centre of the impeller.

Figure 4.18. Simulation of the tank stirred by a Rushton turbine using the Multiple Frames of Reference
(MFR) method: distribution of turbulent kinetic energy, k (m2/s2), in a vertical plane through the centre of
the tank, half way between the impeller blades and half way between baffles.

108

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.19. Simulation of the tank stirred by a Rushton turbine using the Multiple Frames of Reference
(MFR) method: distribution of turbulent energy dissipation rate, (W/kg), in a vertical plane through the
centre of the tank, half way between the impeller blades and half way between baffles.

109

Normalised Axial Velocity (m/s)

Chapter 4. CFD Simulations of Single-Phase Flow

0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

Radius (m)

0.1

0.15

Figure 4.20. Comparison of axial velocity components (normalised by impeller tip speed) obtained by
each simulation method and experimental data ( experimental (Hockey, 1990); + sliding mesh

Normalised Axial Velocity (m/s)

simulation; MFR method). Axial location at x = 0.065 (below the impeller).


0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

Radius (m)

0.1

0.15

Figure 4.21. Comparison of axial velocity components (normalised by impeller tip speed) obtained by
each simulation method and experimental data ( experimental (Hockey, 1990); + sliding mesh

Normalised Axial Velocity (m/s)

simulation; MFR method). Axial location at x = 0.12 (just above the impeller).
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

Radius (m)

0.1

0.15

Figure 4.22. Comparison of axial velocity components (normalised by impeller tip speed) obtained by
each simulation method and experimental data ( experimental (Hockey, 1990); + sliding mesh
simulation; MFR method). Axial location at x = 0.18 (about 2/3 of tank height).

110

Chapter 4. CFD Simulations of Single-Phase Flow

Normalised Radial Velocity


(m/s)

0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
0

0.05

0.1

0.15

Radius (m)

Figure 4.23. Comparison of radial velocity components (normalised by impeller tip speed) obtained by
each simulation method and experimental data ( experimental (Hockey, 1990); + sliding mesh
simulation; MFR method). Axial position at x = 0.065 (below impeller).

Normalised Radial Velocity


(m/s)

0.80
0.60
0.40
0.20
0.00
-0.20
0

0.05

0.1

0.15

Radius (m)

Figure 4.24. Comparison of radial velocity components (normalised by impeller tip speed) obtained by
each simulation method and experimental data ( experimental (Wu & Patterson, 1989); + sliding
mesh simulation; MFR method). Axial position at x = 0.098 (at impeller mid-plane).

Normalised Radial Velocity


(m/s)

0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

0.1

0.15

Radius (m)

Figure 4.25. Comparison of radial velocity components (normalised by impeller tip speed) obtained by
each simulation method and experimental data ( experimental (Wu & Patterson, 1989); + sliding
mesh simulation; MFR method). Axial position at x = 0.12 (just above impeller).

111

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.26. Surface plot of finite volume grid for a square tank stirred by a Rushton turbine (half of tank
has been modelled assuming symmetry).

Figure 4.27. Surface plot of finite volume grid for the region surround one impeller blade.

112

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.28. Finite volume grid for 60 sector of tank stirred by a Rushton turbine, with one impeller
blade and one baffle (assuming periodic boundary conditions in azimuthal direction and 6 baffles in tank).
Grid with density of ~56,000 cells.

Figure 4.29. Finite volume grid for 60 sector of tank stirred by a Rushton turbine, with one impeller
blade and one baffle (assuming periodic boundary conditions in azimuthal direction and 6 baffles in tank).
Higher density grid with ~440,000 cells.

113

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.30. Surface plot of grid on the impeller blade and disc, for simulation of a 60 tank section
with a Rushton turbine, with total grid size of ~55,000 cells.

Figure 4.31. Surface plot of grid on impeller blade and disc, for simulation of a 60 tank section with a
Rushton turbine, with total grid size of ~440,000 cells.

114

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.32. Single blade/single baffle set-up (60 sector) with ~56,000 cells: velocity vectors (m/s) in a
vertical plane half way between impeller blades (vectors plotted on actual grid).

115

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.33. Single blade/single baffle set-up (60 sector) with ~56,000 cells: velocity vectors (m/s) in
horizontal plane near surface, 0.27 m from bottom (vectors plotted on actual grid).

Figure 4.34. Single blade/single baffle set-up (60 sector) with ~56,000 cells: values of the dimensionless
distance to the wall in adjacent cells,

116

+
y adj
, as plotted on walls. Values in the range 11150.

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.35. Single blade/single baffle set-up (60 sector) with ~440,000 cells: velocity vectors (m/s) in a
vertical plane half way between impeller blades (vectors plotted with reduced grid density for clarity).

117

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.36. Single blade/single baffle set-up (60 sector) with ~440,000 cells: velocity vectors (m/s) in
horizontal plane near surface, 0.27 m from bottom (vectors plotted with reduced grid density for clarity).

Figure 4.37. Single blade/single baffle set-up (60 sector) with ~440,000 cells: values of the
dimensionless distance to the wall in adjacent cells,
(note different scale compared to lower density grid).

118

+
y adj
, as plotted on walls. Values in range 1080

Normalised Axial Velocity (m/s)

Chapter 4. CFD Simulations of Single-Phase Flow

0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

0.1

0.15

Radius (m)

Figure 4.38. Comparison of axial velocity components (normalised by impeller tip speed) obtained in
simulations of a 60 tank sector and experimental data ( experimental (Hockey, 1990), MFR,

Normalised Axial Velocity (m/s)

~56,000 cells, + MFR, ~440,000 cells). Axial location at x = 0.065 (below the impeller).

0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

0.1

0.15

Radius (m)

Figure 4.39. Comparison of axial velocity components (normalised by impeller tip speed) obtained in
simulations of a 60 tank sector and experimental data ( experimental (Hockey, 1990), MFR,

Normalised Axial Velocity (m/s)

~56,000 cells, + MFR, ~440,000 cells). Axial location at x = 0.12 (slightly above the impeller).

0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

0.1

0.15

Radius (m)

Figure 4.40. Comparison of axial velocity components (normalised by impeller tip speed) obtained in
simulations of a 60 tank sector and experimental data ( experimental (Hockey, 1990), MFR,
~56,000 cells, + MFR, ~440,000 cells). Axial location at x = 0.18 (about 2/3 of tank height).

119

Chapter 4. CFD Simulations of Single-Phase Flow

Normalised Radial Velocity


(m/s)

0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
0

0.05

0.1

0.15

Radius (m)

Figure 4.41. Comparison of radial velocity components (normalised by impeller tip speed) obtained in
simulations of a 60 tank sector and experimental data ( experimental (Hockey, 1990), MFR,
~56,000 cells, + MFR, ~440,000 cells). Axial location at x = 0.065 (below the impeller).

Normalised Radial Velocity


(m/s)

0.8
0.6
0.4
0.2
0.0
-0.2
0

0.05

0.1

0.15

Radius (m)

Figure 4.42. Comparison of radial velocity components (normalised by impeller tip speed) obtained in
simulations of a 60 tank sector and experimental data ( experimental (Wu & Patterson, 1989), MFR,
~56,000 cells, + MFR, ~440,000 cells). Axial location at x = 0.098 (at impeller mid-plane).

Normalised Radial Velocity


(m/s)

0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
0

0.05

0.1

0.15

Radius (m)

Figure 4.43. Comparison of radial velocity components (normalised by impeller tip speed) obtained in
simulations of a 60 tank sector and experimental data ( experimental (Hockey, 1990), MFR,
~56,000 cells, + MFR, ~440,000 cells). Axial location at x = 0.12 (slightly above the impeller).

120

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.44. Vectors in a vertical plane intersecting the tips of adjacent impeller blades (based on results
using sliding mesh method and grid of 109,440 cells), plotted in rotating frame of reference and showing
prediction of the trailing vortices.

Figure 4.45. Axis of the trailing vortices produced by the Rushton turbine, as determined by experimental
measurements, and as predicted using the sliding mesh method, following the calculation procedure
outlined in Rigby et al., 1998b.

121

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.46. Trailing vortices as predicted from the simulation of a 60 tank sector containing a single
impeller blade, with a higher density grid (~440,000 cells). Velocity vectors coloured by pressure (range
-500 to 500 Pa) in a plane approximately normal to the trailing vortex axis, located at the blade tip.

Figure 4.47. Single blade/single baffle set-up (60 sector) with high density grid (~440,000 cells):
streamline visualisation of the upper trailing vortex on a Rushton turbine blade (surfaces coloured by
pressure (Pa)).

122

Chapter 4. CFD Simulations of Single-Phase Flow

-1.1

0.3

-1.0

0.6

0.5

-0.4
-0.9

0.7

0.2

0.4

Inner
Edge

Outer
Edge

0.6
0.3

-0.8

0.0

0.7
Inner
Edge

0.0

0.1

-0.5

0.1

0.6

0.5

-0.3

Outer
Edge

-0.5

-0.1
-0.6
-0.9

-0.7

0.6 0.5

-0.1

-0.7

-1.2
-1.1

0.5

0.3

-0.7

-0.7

-1.4

0.2

-1.2

Figure 4.48. Pressure coefficients on leading side

Figure 4.49. Pressure coefficients on trailing side

of blade: experimental measurements.

of blade: experimental measurements.

0.6

-1.4

0.7

-1.2

0.4
-1.0

-0.9

0.3
0.2

0.5

0.6

Outer
Edge

0.7

Inner
Edge

-0.6

-1.1

-1.1

0.6
0.1
Inner
Edge

-1.3

-1.0

-0.7

-0.9

-0.5
-0.9

-0.8

-0.8

-1.0
-0.9

0.8

Outer
Edge

-0.4
-0.6

-1.1

0.4
-1.2
-1.4

-1.3

Figure 4.50. Pressure coefficients on leading side

Figure 4.51. Pressure coefficients on trailing side

of blade: CFD simulation of half of square tank

of blade: CFD simulation of half of square tank

(lower grid resolution).

(lower grid resolution).

0.6

-1.5

-1.4 -1.3

0.7

-1.2
-1.0

0.4

-1.0

0.6
0.1
Inner
Edge

-0.9

0.3
0.2

0.5

0.6

Outer
Edge

0.7

Inner
Edge

-0.9

0.8

-0.6

-0.6

-0.7
-0.8

-0.7

-0.5

-1.0

-1.3

Outer

-0.4 -0.1 Edge


-0.3

-0.9
-1.1

-1.4

-0.2

-0.8

-0.6

-1.0
-1.5

-0.9

-0.9

-0.8

0.4

-0.5

-1.1

-0.8

-1.2

Figure 4.52. Pressure coefficients on leading side

Figure 4.53. Pressure coefficients on trailing side

of blade: CFD simulation of zone around blade

of blade: CFD simulation of zone around blade

(higher grid resolution).

(higher grid resolution).

123

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.54. Pressure coefficient, Cp, plotted as isobars, from 0.5 to 3.5 in steps of 0.5, in the horizontal
plane through the axis of the trailing vortex, according to vant Riet & Smith (1975), for a Rushton turbine
at Reynolds number of 6 104.

Figure 4.55. Contours of dimensionless pressure coefficient at one-quarter blade height, for simulation
with single blade/single baffle set-up (60 sector) with 55,000 cells.

124

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.56. Contours of dimensionless pressure coefficient at one-quarter blade height, for simulation of
single blade/single baffle set-up (60 sector) with high density grid (~440,000 cells).

125

Chapter 4. CFD Simulations of Single-Phase Flow

Normalised turbulent kinetic energy,


x = 0.098

0.1

25
20

/ av

k /V tip

0.08

Normalised energy dissipation rate,


x = 0.098

0.06

15

0.04

10

0.02

0
1.0

1.5

2.0

2.5

3.0

1.0

1.5

2.0

R/Ri

2.5

3.0

R/Ri

Figure 4.57. Comparison of k and values obtained in simulations of 180 tank section and experimental
data ( experimental [9], + sliding mesh simulation, z Multiple Frames of Reference simulation). Axial
location at impeller mid-plane.

0.008

Normalised turbulent kinetic energy,


x = 0.049

0.6
0.5
0.4
0.3
0.2
0.1
0

/ av

k /V tip

0.006

Normalised energy dissipation rate,


x = 0.049

0.004
0.002
0
0.0

0.5

1.0

1.5
R/Ri

2.0

2.5

3.0

0.0

0.5

1.0

1.5

2.0

2.5

3.0

R/Ri

Figure 4.58. Comparison of k and values obtained in simulations of 180 tank section and experimental
data ( experimental [8, 10], + sliding mesh simulation, z Multiple Frames of Reference simulation).
Axial location below impeller in bulk flow, half-way between the impeller and the tank bottom.

126

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.59. Single blade/single baffle set-up (60 sector) with 55,000 cells: distribution of turbulent
kinetic energy, k (m2/s2), in a plane through the centre of the impeller.

Figure 4.60. Single blade/single baffle set-up (60 sector) with high density grid (~440,000 cells):
distribution of turbulent kinetic energy , k (m2/s2), in a plane through the centre of the impeller (note
different scale: predicted values are higher with this grid).

127

Chapter 4. CFD Simulations of Single-Phase Flow

0.1

0.09

0.08

Experimental

0.07

0.06

k/V t

Sliding mesh, 180 deg


0.05

MFR, 180 deg


0.04

MFR 60 deg, 56 000


cells

0.03

0.02

MFR 60 deg, 441 000


cells

0.01

0
0

0.2

0.4

0.6

0.8

1.2

(r -R i )/(R t -R i )

Figure 4.61. Prediction of turbulent kinetic energy, k, in plane at impeller centre line: comparison of different
simulation approaches. Turbulent kinetic energy normalised by square of impeller tip velocity, Vt, and radius
normalised by impeller tip radius Ri and tank radius Rt.

30

25

3 2
/(N D )

20

Experimental data
MFR, 180 deg

15
MFR 60 deg, 56 000
cells

10
MFR 60 deg, 441 000
cells

0
0

0.2

0.4

0.6

0.8

1.2

(r -R i )/(R t -R i )

Figure 4.62. Prediction of turbulent energy dissipation rate, (W/kg), in plane at impeller centre line: comparison
of different simulation approaches. Energy dissipation rate normalised by power input per unit volume, and radius
normalised by impeller tip radius Ri and tank radius Rt.

128

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.63. Cut-away surface plot of finite volume mesh for tank stirred by Lightnin A315 impeller.

129

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.64. Lightnin A315 impeller (surface plot of a CFD finite volume mesh)

Figure 4.65. Velocity vectors (m/s) in a vertical plane through centre of a tank stirred by a Lightnin
A315 impeller.

130

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.66. Pressure (Pa) in a horizontal plane intersecting the mid-height of the Lightnin A315
impeller.

Figure 4.67. Turbulent kinetic energy (m2/s2) in a vertical plane through centre of tank stirred by
Lightnin A315 impeller.

131

Chapter 4. CFD Simulations of Single-Phase Flow

Figure 4.68. Turbulent energy dissipation rate (W/kg) in a vertical plane through centre of tank stirred
by Lightnin A315 impeller.

3.0
2.0

U (m/s)

1.0
0.0
-1.0
-2.0
-3.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.69. Comparison of axial velocity component for the Lightnin A315 (obtained using Sliding
Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial profile at a
vertical position 0.083 m below impeller centre.

132

Chapter 4. CFD Simulations of Single-Phase Flow

3.0
2.0

U (m/s)

1.0
0.0
-1.0
-2.0
-3.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.70. Comparison of axial velocity component for the Lightnin A315 (obtained using Sliding
Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial profile at a
vertical position 0.035 m below impeller centre.

2.0

U (m/s)

1.0

0.0

-1.0

-2.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.71. Comparison of axial mean velocity component for the Lightnin A315 (obtained using
Sliding Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial
profile at a vertical position 0.030 m above impeller centre.

2.0

U (m/s)

1.0

0.0

-1.0

-2.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.72. Comparison of axial mean velocity component for the Lightnin A315 (obtained using
Sliding Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial
profile at a vertical position 0.107 m above impeller centre.

133

Chapter 4. CFD Simulations of Single-Phase Flow

Turbulent Velocity (m/s)

1.0
0.8
0.6
0.4
0.2
0.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.73. Comparison of r.m.s. turbulent velocity, u0, for the Lightnin A315 (obtained using
Sliding Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial
profile at a vertical position 0.083 m below impeller centre.

Turbulent Velocity (m/s)

1.0
0.8
0.6
0.4
0.2
0.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.74. Comparison of r.m.s. turbulent velocity, u0, for the Lightnin A315 (obtained using
Sliding Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial
profile at a vertical position 0.035 m below impeller centre.

Turbulent Velocity (m/s)

1.0
0.8
0.6
0.4
0.2
0.0
0

0.05

0.1

0.15

0.2

0.25

Radius (m)

Figure 4.75. Comparison of r.m.s. turbulent velocity, u0, for the Lightnin A315 (obtained using Sliding
Mesh method) with experimental data of Bakker (1992) (experimental, simulation). Radial profile at a
vertical position 0.030 m above impeller centre.

134

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

5.1 Introduction
The modelling equations used for two-phase flow are reviewed in this chapter, with an
emphasis on dispersed flow of gas in a liquid. The physics of flow in such two-phase
mixtures is considerably more complex than that of a single-phase flow, and likewise,
numerical modelling methods involve greater complexity. There is a lack of agreement
amongst various published studies regarding the specification of the modelling
equations, and many aspects of multiphase flow remain subjects of on-going research.
Therefore, further clarification has been needed as to the most appropriate equations for
the CFD model.

An attempt has been made here to present the derivation of the two-fluid equations in a
systematic way. Some of the possible modelling equations and closure expressions have
been evaluated, with emphasis on issues which may be important for modelling gas
dispersion in stirred tanks. An effort has also been made to present all equations in a
standardised vector notation (following Bird et al., 1960), whereas at least three
different systems of notation can be found in the references cited here.

While this chapter is mainly concerned with dispersed bubbly flow, some consideration
must also be given to segregated two-phase flow, since in a gas-sparged stirred tank, gas
cavities are formed behind impeller blades. The extension of the modelling method to
allow for gas cavities has been considered in Chapter 7.

5.2 Approaches to modelling


There are a number of methods available for simulation of dispersed two-phase or
multiphase flows. Two-phase modelling methods have been reviewed by Lathouwers
(1999), who stated that the main problem in simulating dispersed two-phase flows is the
presence of a wide spectrum of length and time scales associated with both continuous
and dispersed phases. Modelling methods generally involve approximations and
averaging, in order to reduce a potentially vast computation to one that can be handled
within the practical limits of current computers.
135

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

In principal, it is possible to carry out a direct numerical simulation (DNS) where all
spatial and time scales are resolved including the flow around and inside the bubbles
themselves, using the basic Navier-Stokes equations without any reference to empirical
closure relations. Simulations of this type have been undertaken, although the enormous
computational demands limit application to a small number of bubbles at low Reynolds
numbers. An example of this type of simulation is the modelling of Esmaeeli and
Tryggvason (1999), who solved the Navier-Stokes equations in conjunction with a front
tracking method, and were able to simulate 8 fully deformable bubbles on a threedimensional grid of 643 cells at bubble Reynolds numbers up to ~30. Such an approach
remains too computationally demanding for simulations related to industrial
applications, where the continuous phase is typically turbulent and very large numbers
of bubbles or particles are present at high Reynolds numbers. Approximate methods are
therefore required.

At the first level of approximation, it is possible to represent the particles or bubbles


using a point-force approximation and empirical expressions for the force on each
particle. The path of such particles can be tracked by a Lagrangian method, involving
step-wise integration of a suitable equation for particle motion, while still resolving the
instantaneous turbulent flow structures of the turbulent phase, using a method such as
DNS. Examples of simulations using the point-force approximation include the
modelling reported by Maxey et al. (1994) and Spelt and Biesheuvel (1999), although
these authors adopted a kinematic simulation approach for the continuous phase (Fung
et al., 1992), where the turbulent flow field is simulated based on spectral functions with
randomised properties intended to represent homogeneous turbulence. Results of
simulations of this type have been useful for determining the average statistics of
particle or bubble motion in idealised turbulent flows. Such methods, however, are
presently limited to simple geometries and no more than a few hundred particles.

At a higher level of approximation, the Lagrangian method can be used to track pointapproximated particles or bubbles through the flow field, while calculating the
continuous phase flow field by means of Reynolds-averaged Navier-Stokes equations
(the so-called Eulerian-Lagrangian approach). In this approach (Lathouwers, 1999), the
coupling of the two phases is awkward, since the particle motion is treated as being the
instantaneous motion, while only averaged quantities are available for the continuous
136

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

phase. This results in a degree of uncertainty in the coupling of the phases. The method
has the advantage of being able to model particles or bubbles of different sizes, although
the method tends to be limited to relatively dilute flows, since modelling of a large
number of particles is computationally very demanding, and also, it is difficult to model
the particle-particle interactions in dense flows.

The most commonly applied approach to two-phase modelling is the two-fluid or


Eulerian-Eulerian approach (e.g. Pericleous & Patel, 1987; Davidson, 1990; Gosman et
al., 1992; Bakker, 1992; Mudde & Simonin, 1999). This method is the most suitable for
modelling relatively high volume fractions of the dispersed phase. In this approach, the
dispersed phase is treated in an averaged manner as an interpenetrating continuum,
characterised by a local average volume fraction and other averaged properties such as
particle size. Since both phases are averaged in a similar manner, coupling of the phases
is more straightforward. In addition, the equations are not limited in application to dilute
flows, and although a dispersed second phase is normally assumed, there is in principle
no reason why the equations should not be also suitable for describing segregated flow,
i.e. regions where there is 100% gas, provided that the physics applicable to the
situation are taken into account, e.g. appropriate modification to the terms describing
the forces between phases.

In order to specify the Eulerian two-fluid equations, some sort of averaging process is
required, meaning averaging both over the individual bubbles to obtain statistics of the
bubbly flow in a macroscopic sense, and also over the temporal fluctuations of each
phase in a turbulent flow. The motion of individual bubbles is generally taken as the
starting point for the averaging process. Due to averaging, information is lost and one
ends up with unknown terms in the equations requiring closure relations. It is possible
to postulate the macroscopic equations without reference to flow on the microscopic
scale, however systematic averaging is generally preferred for several reasons, for
example the physical meaning of the unknown correlations may provide clues as to how
to model them (Lathouwers, 1999).

137

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

5.3 Averaging procedure for the two-fluid equations


There are a number of possible approaches to deriving averaged equations for twophase flow, including time averaging, volume averaging and ensemble averaging
(Lathouwers, 1999). Ensemble averaging has been favoured by several authors (e.g.
Drew, 1983; Simonin, 1990). As outlined briefly here, this is a statistical method where
an average is calculated for a large number of possible states of the system, and this
approach provides a basis for systematic development of the equations for two-phase
flow.

Following Lathouwers (1999), the starting point for deriving averaged equations is the
instantaneous equations for conservation of mass and momentum for each phase:
i
+ i v i = 0 ,
t

(5.1)

i v i
+ i v i v i = i + i g ,
t

(5.2)

where for phase i, i is the density, vi is the velocity, i is the stress tensor, g is the
gravity vector and t is time. The phase index is taken as i = 1 for the continuous phase
and i = 2 for the dispersed phase. Note that where two vectors are written together, i.e.
v i v i , a dyadic product is intended, yielding a second order tensor (Bird et al., 1960).

These equations are supplemented by the interfacial jump conditions which relate
quantities on either side of an interface moving with velocity vint :
2

i (v i v int ) n i = 0 ,

(5.3)

i =1
2

( i v i (v i v int ) i ) n i = 0 ,

(5.4)

i =1

where it is assumed that there is no mass transfer between phases. A phase indicator i
is introduced, which is a filter determining which phase is present at a certain time and
position, i.e.

i (x, t ) = 1 if x i ,
i (x, t ) = 0 otherwise.

138

(5.5)

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

The ensemble averaging operator is given the symbol .. i , where the subscript
indicates that the averaging is with respect to phase i (Simonin, 1990). The ensemble
average of i represents the average probability of finding phase i at position x and time

t, and this is usually called the volume fraction of the phase, i.e.
i i = i .

(5.6)

The procedure for deriving the phase averaged equations is to multiply the
instantaneous equations by the phase indicator, i, and then take the ensemble average.
After applying some properties of the averaging operator, this results in the following:
i i i
+ i i v i i = 0 ,
t

(5.7)

i i v i i
+ i i v i v i i = i i i + i i g i .
t

(5.8)

The first term on the right hand side of equation 5.8 can be rearranged using the chain
rule of differentiation so the equation reads:
i i v i i
+ i i v i v i i = i i i + i g i i i i .
t

(5.9)

In this form, the last term is identified with the interfacial momentum transfer term, Mi.
By multiplying the instantaneous jump condition equations by the Dirac function and
averaging (Lathouwers, 1999), the average jump condition is obtained:
2

i =1

i =1

M i = i i i = 0 .

(5.10)

Thus, the interfacial forces are equal in magnitude and opposite in direction for the two
phases.

A phasic average for a variable f can be defined as:


Fi = f i =

i fi i i fi i
=
,
i i
i

(5.11)

or a density-weighted average can be defined as:


f
f
Fi = fi = i i i i = i i i i .
i i i
i i

(5.12)

The density-weighted average appears to be generally preferred (e.g. Simonin, 1990),


although both values are the same if phase densities are assumed constant. An upper
case letter is generally used for averaged values.

139

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

A further step for turbulent two-phase flows is to decompose the instantaneous velocity
of each phase into mean and fluctuating components. Hence:

v i = Vi + v i

(5.13)

where, following Lathouwers (1999), a double overbar is now used to indicate a


Reynolds-type decomposition. Applying the rules for both phasic averaging and
Reynolds averaging leads to the following form of the two-fluid equations:
i i
+ i i V i = 0 ,
t

(5.14)

i i V i
+ i i V i V i = i ( i + iRe ) + i i g + M i ,
t

(5.15)

where the double bar over the mean velocity will be neglected henceforth.

By comparison with the original instantaneous equations, it is seen that there are now
additional terms in the equations. Similar to the single-phase equations for turbulent
flow, the momentum conservation equation contains a Reynolds stress term iRe , which
is due to non-zero correlations of the fluctuating velocity component and is given by:

i iRe = i i v i v i

(5.16)

The other additional term is the interfacial force term, Mi. Closure expressions must be
defined for each of these terms.

The Reynolds stress exists for both phases according to the above derivation, but
generally, the stress tensor of the second phase is neglected if dispersed two-phase flow
is considered. The Reynolds stress term of the liquid phase is somewhat different in
nature to that in single-phase flow, since velocity fluctuations are not due exclusively to
turbulence in the normal sense of the word, but also occur through the disturbances on
the surrounding fluid flow caused by the presence of bubbles, which would occur even
in the absence of turbulence. Velocity fluctuations of this type are often referred to as
pseudo-turbulence (Lathouwers, 1999). Also, in formulating a turbulence model to
account for the Reynolds stress term, the model may need to take into account the
interphase transfer of momentum and energy, whereby the presence of bubbles can
modify the continuous-phase turbulence, either enhancing or attenuating the turbulence
intensity. Various approaches have been taken to modelling the two-phase Reynolds
stress term and this subject is discussed further in Section 5.8. In the present situation,
140

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

since high turbulence levels are produced by the impeller, it has been assumed that the
turbulence is dominated by the impeller-induced flow, with minimal influence due to
the bubbles. Therefore, the approach taken here has been to calculate the k- equation
for each phase, without any additional source terms. However, for the liquid phase an
additional Reynolds stress term has been added to account for bubble-induced pseudoturbulence in the liquid phase, although this is only a small correction.

In the case of the interfacial force term in the momentum conservation equation, this
can be decomposed and rewritten to expressly show terms in the pressure:

i = i pi + ( pi ,int pi ) i + M i

(5.17)

It is common to assume (Lathouwers, 1999) that the pressure in each phase, pi, and the
interfacial pressures, pi,int, are all equal, and this approach has been adopted here.
It is generally proposed (Hunt et al., 1994; Lathouwers, 1999) that the term M i can be
expressed as the linear sum of a number of forces acting between the phases which
express different effects, given by:

M i = M drag + M am + M lift + M hist + M turb ,

(5.18)

where the terms on the right hand side are (from left to right) the drag, added mass, lift,
history, and turbulent forces. The history force represents the average memory effect
and is normally taken as zero in averaged equations (Lathouwers, 1999). The most
important terms are generally the drag force and the turbulent force. The added mass
force is significant only when there is strong acceleration and hence a large difference in
velocity between the phases. The lift force is significant when there is strong vorticity or
a large velocity gradient` in the continuous phase. In terms of modelling flow in a
stirred tank, added mass and lift forces would be expected to be significant only in the
vicinity of the impeller blades. The turbulent force is an important additional force
which does not appear in the instantaneous equation but arises due to non-zero
correlations between the fluctuating variables. The turbulent force has the important
effect of spreading gas out due to interaction between bubbles and eddies.

141

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

5.4 Closure method for the interfacial force


For the detailed specification of the interfacial force term, M i , a range of mathematical
models have been proposed by various authors, often in a somewhat empirical way, by
proposing force terms as functions of the averaged fluid parameters which imitate the
form of the instantaneous forces on a particle or bubble. However, such an approach is
in conflict with formal averaging procedures, since turbulent correlation terms are
generally discarded. Also, such an approach does not lead to an expression for turbulent
dispersion, which is needed in the macroscopic description. Systematic approaches to
averaging have been pursued by a number of workers (e.g. Gosman et al., 1992; Issa &
Oliveira, 1995; Oey et al., 2003). Here, the averaging method for the interfacial force is
described following the derivation of Simonin and coworkers (Simonin & Viollet, 1989;
Simonin, 1990, 1991; Bel Fdhila & Simonin, 1992; Viollet & Simonin, 1994). This
approach is based on ensemble averaging of the forces on a single isolated bubble.
Although there remain various problems and issues with the equations developed by
Simonin and coworkers, the method of systematic averaging would appear to be an
improvement compared to other ad hoc models which have been proposed. This method
was chosen for detailed investigation because it is one of the most detailed closure
methods which has been published. A turbulent dispersion model based on the
development of Simonin and coworkers has been implemented in the CFD model, as
described in Chapter 8.

The starting point for the derivation of the interfacial forces is the equation of motion of
a single bubble. A first statement of this is merely that the force on the bubble is equal
to mass times acceleration (Newtons second law), i.e.

dv b
= Fb ,
dt

(5.19)

where is the bubble volume, g is the gas density, vb is the bubble velocity, and Fb is
the total force on the bubble. The force on a bubble depends on the details of how the
liquid flows around the surface, the pressure distribution on the surface, the extent of
flow separation and wake formation, the presence of shear fields, and whether the
surface is mobile or immobile (Clift et al., 1978). The question of how to correctly
define or calculate the force on the particle or bubble remains a subject of discussion in
the literature, and for most practical situations it is necessary to refer to empirical
142

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

correlations. Various attempts have been made to derive the equation of particle motion
by means of theoretical analysis. However, no rigorous, generalised equation has been
derived except for the case of spherical bubbles at very low Reynolds numbers (Chahed
et al., 2003). Various theoretical efforts have been made to derive expressions with a
more general range of applicability, yet it has always been necessary to impose strict
limitations on the range of applicability of the equation. For example, Auton et al.
(1988) derived an equation for the motion of spherical bubbles at high Reynolds number
in clean water. However, they assumed inviscid flow, leading to an expression for the
force consisting of terms describing pressure gradient in the liquid, added mass and lift,
but completely neglecting the drag force term. Yet at low and intermediate bubble
Reynolds numbers, the drag is a very important force which must be taken into account.

A more practical equation of motion is not strictly based on theoretical developments,


but is obtained as a linear superposition of the main forces identified as acting on a
bubble (Lathouwers, 1999), where the individual forces must be specified with
reference to empirical correlations, especially for the drag force. When this equation is
further generalised to the case of a multi-particle system in the two-fluid formulation, it
is usually assumed that the effect of particle-particle interactions can be expressed in
terms of modifications to the individual force terms. Following Lathouwers (1999)
(with minor corrections), a practical equation of motion of a single bubble may be given
as:
1
dv Dv
Fb = l CD Ab vb vl (vb vl ) CA l b l CL l (vb vl ) p + g g .
2
dt Dt
(5.20)
Here, l is the liquid density, Ab is the projected area of the bubble, vl is the liquid
velocity, p is the liquid phase pressure, and is the vorticity given by:

= vl .

(5.21)

In equation 5.20, the terms on the right hand side, from left to right, are the drag force,
the added mass force, the lift force, the pressure gradient force, and the gravity force.
Associated with the first three forces are coefficients which need to be specified, namely
the drag coefficient, CD, the added mass coefficient, CA, and the lift force coefficient,

CL. An additional term sometimes included is the Basset history force, representing a
memory effect owing to the structure of the boundary layer and the wake of the bubble
143

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

(Lathouwers, 1999). This force may be important in Lagrangian tracking of individual


particles, although in deriving averaged equations for particle or bubble motion in
turbulent flow, the history force is generally ignored (e.g. Simonin & Viollet, 1989;
Lathouwers, 1999; Chahed et al., 2003).

The pressure gradient force in equation 5.20 is sometimes written in terms of the
continuous phase velocity gradient and gravity force according to:
p = l

Dvl
l g ,
Dt

(5.22)

This expression is the Euler equation (Bird et al., 1960), which is an inviscid
approximation of the Navier-Stokes equation.

In general, the coefficients for the drag, CD, added mass, CA, and lift, CL, must be
specified empirically. The values of CA and CL are generally taken as for individual
bubbles and dilute flows, although corrections have been proposed, e.g. to account for
the effect of gas volume fraction (Biesheuvel & van Wijngaarden, 1984). The value of

CD is generally determined from empirical correlations, where it may be a function of


several variables such as bubble Reynolds number, Etvs number and volume fraction,
and may also show a dependence on the level of purity of the liquid. The estimation of
the drag coefficient is reviewed in further detail in Chapter 6.

Simonin and coworkers (Simonin, 1990; Bel Fdhila & Simonin, 1992; Viollet &
Simonin, 1994) derived the averaged interfacial force starting with an equation for
bubble motion similar to equation 5.20 above, but writing the derivatives following the
bubble motion and neglecting the lift force. Applying a transformation for the pressure
gradient term according to equation 5.22, the equation for instantaneous bubble motion
can be written:

v
f 2 = 1 f D v r 1C A r + v 2 v r + 1 1 + v 2 v1 (1 2 )g .

(5.23)

The first term represents drag, where fD is given by:

fD =

3 CD
vr .
4 d

(5.24)

The second term represents added mass, the third represents the pressure gradient, and
the last term is buoyancy.
144

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

By averaging and separating out the buoyancy and the mean pressure gradient terms, the
following form for the interfacial force, M i , is obtained:

M 2 = M 1 = f 2 + (1 2 )g + 2 P 2 = 2 1 FD v r
( 2 1C A v 2 v r 2 ) + ( 2 1 v1 v 2 2 ) +

2 1C A r + V2 Vr
t

2
( 2 1 v1 v1 1 )
1
(5.25)

The first two terms on the right hand side are the average drag force and the average
added mass. The third term arises from the instantaneous added mass force where there
are non-zero correlations between the instantaneous bubble velocity and the
instantaneous relative velocity. The fourth and fifth terms were proposed by Bel Fdhila
and Simonin (1992) to represent the correlation between the instantaneous distribution
of bubbles and the instantaneous fluid pressure fluctuations. Thus, this averaging
procedure leads to additional terms which would not appear if the instantaneous
variables in the equation of motion were simply replaced by corresponding averaged
variables. These turbulent correlation terms require closure, and such closure
expressions have been proposed by Simonin and coworkers subject to various
assumptions.

The instantaneous equation of bubble motion as assumed here does not contain a lift
force term, so there is no dispersive term due to lift in the resulting averaged equation.
However, according to Chahed et al. (2003), the turbulent contribution of the lift force
may be expected to be negligible, since there is only a relatively weak correlation
between fluctuation of velocity and fluctuation of liquid vorticity. On the other hand,
there may be problems with including the mean lift force in the absence of a
corresponding term for the fluctuating lift force. This is discussed briefly in Section 5.8.

Equation 5.25 above is developed further following the approach of Simonin and
coworkers. Firstly, for the first term relating to the average drag, further consideration is
needed as to the meaning of the average coefficient FD and the average relative velocity
vr

vr

. Averaging of the instantaneous relative velocity gives:


= V2 V1 + v 2

v1

= Vr v1

(5.26)

145

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Hence, the drag force is not simply proportional to the mean velocity difference, but
there is an additional term due to the correlation between the distribution of particles
and the fluctuating velocity of the continuous phase. This additional term is called the
drifting velocity (Simonin & Viollet, 1989; Chahed et al., 2003), and represents a
mechanism of dispersion by turbulent motions of the fluid. By considering the limiting
case of particle diameter tending towards zero, Simonin (1990) proposed a model for
the drifting velocity in terms of volume fraction gradients and a turbulent dispersion
tensor, D12, where:
v d = v1

1
1
1
2
= D12 2 1 = D12
1
1 2

(5.27)

The turbulent dispersion tensor is modelled as:


t
D12 = 12
v1 v2

(5.28)

t
is the large eddy time scale, as seen by the particles. Hence, the averaged drag
where 12

force term can be separated into a term proportional to the difference between average
particle and liquid velocity, and a dispersive component. Thus, a turbulent dispersion
force is derived from the drag force, which will be referred to as Ti,1. As shown below,
other turbulent dispersion terms also arise in this derivation.

Averaging of the term fD is as follows (Bel Fdhila and Simonin, 1992):


FD = f D

3 CD
4 d

vr

(5.29)

Here, it was proposed that the average magnitude of relative velocity is given by:

vr

= Vr Vr + v r v r

Further, it was recognised that the average drag coefficient, C D

(5.30)
2

, should take into

account the variations in drag coefficient due to turbulent fluctuations along the particle
path. Therefore, the average drag coefficient may not be the same as the drag coefficient
in the absence of turbulence. Bel Fdhila and Simonin (1992) proposed that the average
drag coefficient could be calculated using a standard correlation for drag, but based on a
Reynolds number which is a function of the average relative velocity, i.e.
Re =

146

1 1d v r
1

(5.31)

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Hence, the mean drag force is a function of the mean relative velocity, but using
modified definitions of the total magnitude of relative velocity and the drag coefficient.
Problems with this approach are discussed in Chapter 6.

To achieve closure of these equations, Simonin and coworkers have developed


expressions derived through an extension to Tchens theory (Tchen, 1947; Hinze, 1959;
Gouesbet, 1984). This theory, as summarised by Lathouwers (1999), relates to estimates
of particle dispersion in terms of the Lagrangian spectral tensors of the particles and the
carrier phase. It is assumed that the particles move through homogeneous and steady
turbulent flow, and follow a simplified, linearised form of the equation of motion of a
single particle. By assuming an exponential form of the particle temporal correlation,
algebraic relationships are obtained which relate the particle fluctuating velocity and
particle-fluid covariance to the characteristics of the continuous-phase turbulence. There
are a number of additional assumptions involved, which are not really applicable to
typical sizes of bubbles, e.g. Hinze (1959) states that the particle should be smaller than
the Kolmogorov scale, and the slip velocity should be low enough for Stokes law to be
applicable. Hence, the resulting expressions using this theory may be only approximate
when applied to gas bubbles. Some caution is necessary, and it may be that better
expressions will need to be derived in the future.

The expressions developed thus far have allowed for the possibility of anisotropic
turbulence, i.e. the various correlation terms may have different values in different
directions. Simonin (1991) has developed closure relations for this more general
situation, for application with a turbulence model such as the Reynolds stress model.
However, since the modelling of turbulence in this study has been limited to the k-
turbulence model, only the isotropic case is considered further here. In this case, the
velocity correlations are assumed equal in all coordinate directions, and the velocity
correlation and diffusivity tensors simplify to diagonal tensors with three equal
components. Following from the extension to Tchens theory, the following expressions
have been obtained for the various velocity correlation tensors in isotropic turbulence
(Simonin, 1990; Bel Fdhila and Simonin, 1992; Lathouwers, 1999):

v1 v1

2
= k1I ,
3

(5.32)

147

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

v2 v2

v1 v2

v1 vr

v2 vr

v r v r

2 b2 +r
I,
= k1
3 1+r

(5.33)

2 b +r
I,
= k1
3 1 +r

(5.34)

2 b 1
I,
= k1
3 1 +r

(5.35)

2 b2 b
I,
= k1
3 1 +r

(5.36)

2 (b 1)2
= k1
I,
3 1 +r

(5.37)

where I is the identity or unit tensor. In these expressions,


b=

1+ CA

+ CA

, and r =

t
12
.
12p

(5.38)

The ratio r compares the characteristic time scales of the turbulence and of the particle.
The particle relaxation time is a measure of how quickly a particle adapts its speed to
the surrounding liquid flow, and can be calculated as:

12p =

2
+ CA
1
3 CD
vr
4 d

(5.39)
2

The characteristic time scale of the turbulence is given by:


t
=
12

1
1 + C 2

(5.40)

This equation allows for the so-called cross-trajectory effect according to Csanady
(1963) where, because of the slip velocity of the particles, the particles cut through the
turbulent eddies, and the time scale of the eddies as seen by the particles is reduced.
This effect is estimated by comparing the particle relative velocity with the fluctuation
velocity of the turbulence, hence:

r =

Vr
2

3 k1

(5.41)

The Eulerian integral time scale of the turbulence, 1, is derived from the kinematic
eddy viscosity, which is taken as a product of velocity and time scales, as follows:

148

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

T ,1
k2
2
= C 1 = v1 1 = k1 1 ,
1
1
3

(5.42)

and hence:

3
2

1 = C

k1

(5.43)

In these expressions, C = 0.09 as in the k model. The value of C is taken as 0.45 in


the direction parallel to relative particle motion and 1.8 in directions perpendicular to
particle motion. For simplicity in the CFD model, this is approximated as being
isotropic and a value of 1.0 was used.

By substituting these various expressions into equation 5.25, three terms in addition to
the mean drag force and mean added mass can be defined in the interfacial force, M i .
These are turbulent dispersion force terms, which represent dispersion mechanisms
related to fluctuations in drag, added mass, and fluid pressure gradient. Thus, the
turbulent dispersion force is found to be the sum of the following terms:

T2,1 = 2 1 FD

D12

1 2

2 , where D12 =

2 b + r

3 1 + r

1
k1
;

1 + C 2

(5.44)

2 b2 b
1C A( 2 k1 ) ;
T2, 2 =
3 1 + r

(5.45)

2 b +r
T2,3 =
3 1 +r

(5.46)

2
1( 2 k1 ) + 2 1( 2 k1 ) .
1 3

For terms T2, 2 and T2,3 , the gradient can be rewritten using the chain rule for
differentiation, i.e.
( 2 k1 ) = k1 2 + 2k1 .

(5.47)

Hence, it is found that the turbulent dispersion force consists of terms which are
proportional to the gradient of dispersed-phase volume fraction, and also contains terms
proportional to the gradient of continuous-phase turbulent kinetic energy. Comparison
with the modelling methods of other authors shows that turbulent dispersion is generally
always modelled as being proportional to volume fraction of the particles or bubbles, so
that particles will tend to spread out from regions of high concentration to regions of
low concentration. The inclusion of terms proportional to turbulent kinetic energy
149

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

represents a different effect referred to as turbulophoresis (Bel Fdhila and Simonin,


1992). Bakker (1992) also mentions this effect. Dispersion due to the gradient of
turbulent kinetic energy implies that if particles in a turbulent flow field are initially
uniformly distributed, they will tend to migrate from regions of high turbulence
intensity to regions of low turbulence intensity. However, terms in k1 have been
neglected in the CFD model.

In application of this approach to modelling turbulent dispersion, various simplifications


have been made in the published work of Simonin and coworkers. Where flows
containing solid particles were modelled, only the dispersion term due to drag, T2,1 , was
applied (Viollet & Simonin, 1994). However, when gas bubbles are modelled, it was
considered important to include the terms due to added mass and liquid phase pressure
gradient, since the bubbles differ from solids in that they have almost zero mass, and it
is the disturbance created in the surrounding fluid that is important.

This approach to closure of the two-phase equations has been applied by Simonin and
coworkers to a number of problems with reasonable success. For example, particleladen air jets were modelled using the method described (including drag force terms
only), and good agreement was found with experimental measurements of jet spreading
rate, centre-line velocity and particle dispersion (Viollet & Simonin, 1994). Bel Fdhila
and Simonin applied the method to bubbly flow in a pipe, downstream of a sudden pipe
expansion. They included turbulent dispersion terms T2,1 and T2, 2 , and a simplified
form of T2,3 (neglecting the second term which is small due to the multiplying factor

2/1). Numerical predictions compared favourably with the overall set of experimental
data, although they suggested that accuracy could be improved further by use of a nonisotropic turbulence model and improvements to the constitutive relations used to
account for the turbulent velocity correlation terms. Mudde and Simonin (1999) applied
equations very similar to those of Bel Fdhila and Simonin (1992) in numerical
simulations of the oscillatory motion of a bubble plume in a tank, where they obtained
results that agreed well with experimental measurements.

150

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Since the turbulent dispersion force terms according to Simonin and coworkers are
given as functions of the time scales of the particles and the carrier fluid, the effects of
inertia etc., this approach calculates rates of dispersion that may be more or less than
that of a fluid particle depending on relative densities and particle size.

5.5 Comparison of models for the turbulent dispersion force


The turbulent dispersion force according to Simonin and coworkers may be compared
to the dispersion force terms proposed by other authors. The turbulent dispersion force
has been included in a variety of ways. In some cases (e.g. Ranade & Deshpande, 1999),
no turbulent dispersion force has been included at all, in which case dispersion may still
occur due to numerical diffusion, although such an approach would presumably
underestimate the real dispersive forces.

In modelling of a gas plume in a liquid bath, Davidson (1990) adopted an expression for
the turbulent diffusion force following that assumed by Lee (1987), where (written in a
form as similar as possible to that of Simonin and coworkers):

T2 = 21 1 FD

Deff

2 .

(5.48)

This is very similar to the term T2,1 of Simonin and coworkers, and so takes into
account the drifting velocity due to particle slip, while ignoring the effects of added
mass and continuous-phase pressure gradient. However, it differs from the expression of
Simonin and coworkers by a factor (1)2, which makes only a small difference provided
that the gas-liquid mixture is dilute. Also, the term FD is calculated with a slip velocity
which is simply the difference between mean phase velocities, and the drag coefficient
is calculated using a standard correlation. A greater difference is that Davidson set the
eddy diffusivity, Deff, equal to the eddy viscosity of the liquid, i.e. T,1, whereas in the
approach of Simonin and coworkers, the calculated eddy diffusivity for gas bubbles will
generally be greater than that of the liquid.

According to some other authors, the turbulent dispersion force takes the form:

T2 = Ctd 1k1 2 .

(5.49)

151

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

This model has been proposed by Kurul and Podowski (1990) and also by Lahey et al.
(1993), who justified the form of the equation through an analogy with thermal
diffusion of molecules in the atmosphere (Lathouwers, 1999). Different values of Ctd
have been applied to fit results to experimental data, e.g. Kurul and Podowski
recommend a value of 0.1, whereas Lo (2000) used this equation with Ctd equal to 1.0.
A turbulent dispersion force of this form is available as a command language option in
the CFX4 code, where the default value for Ctd is 0.1. This form of the turbulent
dispersion force appears quite similar to the force terms of Simonin and coworkers
under conditions of homogeneous turbulence (i.e. ignoring the term in k1 ), since in
each of these terms, the force is proportional to 1k1 2 . Hence, the constant Ctd is
equivalent to a much more complicated expression in the model of Simonin and
coworkers, where the dispersion term is a function of the properties of the particles,
liquid and turbulence.

Drew (2001) has also developed equations to express the turbulent dispersion force in
the two-phase momentum conservation equations. He used a form of the instantaneous
equation of motion of a bubble which is similar to that of Simonin and coworkers, again
including added mass force and liquid pressure gradient. However, his method of
analysis was quite different. Rather than carrying out averaging of the instantaneous
equation, he developed a model for the trajectory of a bubble in response to acceleration
by a non-rotating eddy. This leads to an expression for the dispersion tensor of the
bubble, whence he derives the bubble dispersion force. Drew made no reference to the
previous work of Simonin and coworkers, however, it can be shown that the equation
for turbulent dispersion force derived by Drew has considerable similarities.

For gas bubbles, Drews proposed expression for the turbulent dispersion force is:

T2 =

1 1C A Re
2 ,
p 2 2

(5.50)

and making the assumption that the dispersed phase Reynolds stress, Re
2 , can be
approximated as:
2
Re
2 = 2 k1I ,
3
the force expression becomes:
152

(5.51)

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

T2 =

2 1
1C A k1 2 .
3 p

(5.52)

This expression is very similar to the term T2,1 of Simonin and coworkers, which can be
rearranged to make the comparison clearer. Taking into account the definition of p,
with 2/1 0, one can write:
FD =

CA

(5.53)

and including the definition of the diffusivity D12, T2,1 according to Bel Fdhila and
Simonin (1992) can be rearranged to read as follows:
2 b +r
T2,1 =
3 1+r

t
12

1C A k1 2
p

(5.54)

Hence, the force term of Drew is almost the same as the first term T2,1 of Simonin and
coworkers. The additional multiplying factor approaches 1 for small bubbles with low
relaxation time, so in this limit the equations are almost the same. The equation of Drew
t
and
also differs in that it estimates the characteristic eddy time scale as1 rather than 12

so ignores the cross-trajectory effect. Drews method does not lead to terms for
dispersion due to added mass force or liquid pressure gradient, although these forces
were certainly considered in the derivation. The simpler form of his equation may be
due to various assumptions made in the derivation, e.g. he assumed that the force on a
bubble and the bubble velocity are uncorrelated.

The turbulent dispersion force according to Davidson (1990), as given in equation 5.48,
can also be rearranged further. Davidson assumed that the effective diffusivity Deff is
equal to the kinematic turbulent viscosity, so Deff can be substituted, according to
equation 5.42, by a term in k1 and 1. The term FD can be written in terms of CA and p
according to equation 5.53. Thus, the expression of Davidson (1990) can be rewritten
as:

T2 =

2 1
11C A k1 2
3 p

(5.55)

This is nearly identical to the equation of Drew (2001), differing only by a factor of 1.

153

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

5.6 Evaluation of models for the turbulent dispersion force


The expression for turbulent dispersion force derived according to the approach of
Simonin and coworkers has been compared with those of some other authors through
some example calculations. Typical stirred tank conditions were considered, according
to the experimental set-up of Barigou and Greaves (1992, 1996), as used for validation
in this study. Two bubble sizes, with diameters 1 mm and 3 mm, were considered, and
two sets of turbulence parameters were considered. Representative values for k and in
the impeller discharge stream and in the bulk of the tank were estimated with
reference to the values reported by Wu and Patterson (1989) and Deglon et al. (1998)
for tanks stirred by a Rushton turbine. Since their results were for tanks of different
sizes and impeller speeds, k values were scaled by the square of the impeller tip speed,
and values were scaled by the power input per unit volume. For an operating speed of
180 rpm in the tank of Barigou and Greaves (1992, 1996), the estimated values of the
turbulence parameters are k = 0.9 m2/s2 and = 30 W/kg for the impeller discharge
region, and k = 0.13 m2/s2 and = 0.4 W/kg for a location in the bulk. A gas volume
fraction of 0.05 was assumed for the calculations.
For these sets of values of bubble diameter, k and , the turbulent dispersion force has
been calculated according to the approach of Simonin and coworkers with the
assumption of homogeneous turbulence, i.e. k1 = 0 . The three turbulent dispersion
terms were calculated and summed to give the overall turbulent dispersion force, from
which an equivalent value of Ctd can be obtained. In the calculation of the first turbulent
dispersion term according to Simonin and coworkers, T2,1 , (which is the term due to
fluctuations in drag force), the total magnitude of the relative velocity was calculated
according to equation 5.30, with the turbulent fluctuating component of the relative
velocity in this equation calculated as 2k1

(b 1)2
1 +r

, which follows from the isotropic

form of the tensor in equation 5.37. The drag coefficient of the 1 mm bubble was
calculated using a modified Reynolds number according to equation 5.31. However, for
the 3 mm bubble, the distorted regime would normally apply, and the drag coefficient
is a function of the Etvs number rather than the Reynolds number (as described in
Chapter 6). Therefore, the drag coefficient was not modified for the 3 mm bubble.

154

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

The turbulent dispersion force and equivalent Ctd were also calculated according to
Drew (2001), equation 5.52. Drews equation has also been shown to be representative
of the equations of Lee (1987) and Davidson (1990) (differing only by a factor of 1,
which is in the range 0.91.0 for typical stirred tank conditions considered in this
study). Calculations for the various representative conditions are summarised in
Table 5.1.

It can be seen that according to the methods of Simonin and coworkers, and
Drew (2001), Ctd is not a constant, but is dependent on the values of bubble diameter
and the turbulence parameters. The equivalent values of Ctd follow the same trend for
bubble size with each method, i.e. smaller bubbles experience a larger dispersion force.
This seems reasonable, since larger bubbles have greater inertia, and therefore the
response of the bubble is less. There is also a trend where the coefficient Ctd is larger at
the position in the bulk of the tank. The equivalent values of Ctd according to Drew
(2001) generally fall within a range which is similar to the constant values proposed in
the literature, e.g. 0.1 according to Kurul & Podowski (1990) and 1.0 according to Lo
(2000). However, the values are considerably higher using the method of Simonin and
coworkers.

The approach following Simonin and coworkers, where the drag force term is calculated
based on a magnitude of relative velocity containing an additional turbulent fluctuating
component, according to equation 5.30, is not well established. It is possible that the
approach of Simonin and coworkers does not provide a very good representation of the
average drag in turbulent flow. As discussed in Chapter 6, these expressions were
considered for the specification of the drag force in the CFD model, but they were not
adopted because they were not consistent with experimental data. Instead, a more
empirical approach was taken to calculating the mean drag. Therefore, since there is
doubt about the method of calculating magnitude of relative velocity and the drag
coefficient, the turbulent dispersion force was also calculated on the basis of simpler
definitions of relative velocity magnitude and drag coefficient, namely a CD value based
on standard correlations, and the magnitude of relative velocity expressed as the normal
terminal rise velocity in a stagnant liquid. Calculations with this modified method are
also included in Table 5.1. This approach leads to values of the equivalent Ctd which are
in quite close agreement with those of Drew (2001). Note that the term relating to drag
155

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

force in the method of Simonin and coworkers is still quite different in magnitude, but
after adding in the terms due to added mass and pressure gradient, the overall turbulent
dispersion force is found to be similar in magnitude to the one-term equation of Drew.
Since the modified calculation procedure gives values of the turbulent dispersion force
that are more consistent with those of other authors (i.e. Lee, 1987; Davidson, 1990;
Drew, 2001), and more consistent with recommended values of Ctd such as 0.1 (Kurul &
Podowski, 1990) or 1.0 (Lo, 2000), it was decided to implement this modified model in
the CFD simulations of two-phase flow.

In Chapter 8, results using this modified model of Simonin and coworkers have been
compared against CFD simulation results using the simpler expression for turbulent
dispersion force, equation 5.49, with a constant value of Ctd = 0.1. The equivalent value
of Ctd in the CFD simulations is illustrated by Figure 5.1. This plot shows the variation
in a vertical plane half way between baffles, in the results for the simulation of the tank
of Barigou and Greaves (1992, 1996) at a speed of 180 rpm. It can be seen that the
equivalent Ctd varies from about 0.1 near the impeller tip to about 3.5 in the upper part
of the tank. The equivalent turbulent dispersion coefficient varies in a similar way for
the other impeller speeds which were modelled.

A further point of note relates to the term T2,3 , describing the effect of liquid pressure
gradient, which in the derivation of Bel Fdhila and Simonin (1992) is opposite in sign
to that of the other two. This means that the term actually has an opposing effect to that
of drag and added mass and reduces the overall turbulent dispersion force. The example
calculations confirm that this term should be of opposite sign, since if it was of the same
sign as the other terms, or indeed if it was neglected, then the total turbulent dispersion
force will be far too great in comparison with the force according to the equations of
other authors.

5.7 Added mass and lift forces


The added mass and lift forces arise due to unsteady motion of dispersed particles or
bubbles relative to the surrounding continuous phase. The added mass (or virtual mass)
force, A2, results from acceleration of a certain volume of the continuous phase
surrounding the particle. The lift force, L2, results from the motion of a particle through
156

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

a rotating fluid, which leads to the generation of a force perpendicular to the vorticity
and perpendicular to the particle velocity vector (Drew & Lahey, 1987). These forces
can be given as (Drew & Lahey, 1987; Lathouwers, 1999):

DU 2 DU1
A 2 = A1 = 2 1C A

,
Dt
Dt

(5.56)

L 2 = L1 = 2 1C L (U 2 U1 ) ( U1 ) .

(5.57)

Simulations of gas-liquid flow reported in the literature vary in their approach as to


whether the added mass and lift forces have been included or not. For example, the
added mass and lift forces were not included in the modelling of gas-sparged stirred
tanks by Bakker (1992), whereas Gosman et al. (1992) included the added mass force
but neglected the lift force. Likewise, in the derivation of the averaged two-fluid
equations according to Simonin and coworkers as described earlier in this chapter (e.g.
Bel Fdhila and Simonin, 1992), the added mass was included, while the lift force was
ignored. If they had included the lift force, then their derivation of the turbulent
dispersion force should have led to an additional term due to fluctuations in lift force.
However, it is difficult to obtain a closure expression for the turbulent component of the
lift force since there are mathematical complexities due to the curl operator
(Lathouwers, 1999).

While various authors have tended to include the added mass force but neglect the lift
force, Drew and Lahey (1987) have stressed that both the lift force and added mass
should always be included together. The test that they imposed is that the net force
should always be the same regardless of the reference frame in which the flow is
viewed. It can be shown that, for example, if only the added mass is included, and the
reference frame is changed from a stationary frame to a translating or rotating frame of
reference, then additional fictitious force terms result, making the simulation results
unphysical. Therefore, they proposed that both the lift force and added mass force must
always be used together in the frame invariant form, as neither force is individually
frame invariant. The importance of this frame invariance rule may not be obvious in
simulations where the normal inertial reference frame is used. However, if the multiple
frames of reference method is used to simulate the impeller motion in a stirred tank, this

157

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

is an important consideration, since if the lift force is neglected, extra unphysical forces
arise in the rotating frame of reference.

In equations 5.55 and 5.57, the coefficients CA and CL need to be specified. For a single,
spherical particle, it has been found that the values of the added mass and lift
coefficients are both 0.5 (Auton et al., 1988). This result comes from an analysis of a
single sphere accelerating relative to an inviscid fluid undergoing a pure strain and
rotation on a scale much larger than the particle (Drew & Lahey, 1987; Hunt et al.,
1994). There are a number of underlying assumptions in the derivation, including that
the continuous phase velocity gradients are relatively small, the relative velocity is
large, and there is no boundary layer separation around the particle (Drew & Lahey,
1987; Hunt et al., 1994).

For other situations, the added mass and lift coefficients may be different, e.g. where
there are viscosity effects or where there is wake formation behind the particle. For
example, in simulations with ellipsoidal bubbles of about 2 mm diameter, Lathouwers
(1999) found it necessary to use a value of CA = 2 in order to obtain agreement with
experimental measurements. Also, unless the dispersion is highly dilute, the effects of
interactions between particles must be accounted for. For the added mass coefficient in
a swarm of bubbles, it was found by Van Wijgaarden (1976) that there is a moderate
increase in CA with gas concentration, and the effect of gas volume fraction could be
expressed as:
CA =

1
(1 + b 2 )
2

(5.58)

where b = 3.32 (Biesheuvel & van Wijngaarden, 1984). However, the increase in added
mass must only apply at fairly low volume fractions, since the mass of continuous phase
between the bubbles decreases with increasing gas fraction, and in the limit of 100%
gas, CA must go to zero. Davidson (1990) suggested multiplying the coefficient by the
liquid phase volume fraction, i.e.
C A = 1C A .

(5.59)

For the lift force, an opposite trend has been found, where the value of CL decreases
sharply with increasing gas volume fraction. According to Beyerlein et al. (1985), the
lift coefficient could be given as:
158

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

C L = 1.65 10 3 20.78 ,

(5.60)

while according to Behzadi et al. (2004),


C L = 6.51 10 4 21.2 .

(5.61)

With either equation, even with a gas fraction of only 5% or so, the value of CL
becomes very small. This is apparently because in the presence of other bubbles, the
flow around individual bubbles is influenced by the wakes of neighbouring bubbles, so
the bubble does not directly experience the rotation in the continuous phase (Behzadi et
al., 2004).

The different values for CA and CL at moderate volume fractions seem at odds with the
analysis of Drew and Lahey (1987), since if the value of CL goes to zero while there is
still a significant value of CA, then this seems equivalent to setting the lift force to zero,
which violates the principal of frame invariance. Drew and Lahey (1987) stated that the
added mass and lift coefficients were equal, and it would seem that the coefficients
should remain equal in order to conform to the rule of frame invariance. Further
clarification seems necessary.

In terms of the CFD modelling method developed for this study, the added mass and lift
forces were not included, since although several attempts were made to run simulations
with these forces included, convergence could not be achieved. The reasons for this
have not been understood. However, it is likely that for most of the stirred tank, the
dominating force should be the drag force, and the added mass and lift forces will not be
large. These forces only occur where there is significant acceleration between the
phases or significant vorticity of the liquid. To illustrate this, the added mass and lift
forces have been evaluated (assuming values of 0.5 for CA and CL), based on postprocessing of the results from one of the simulations carried out using the final
modelling method, as developed in Chapter 8. Figure 5.2 shows a plot of the drag force
on the gas, calculated as the force per unit volume of gas. The drag force is generally
about the same as the buoyancy over most of the tank (with g 9800 N m 3 ), since
in most places there is no acceleration and drag and buoyancy are in equilibrium.
Higher values of the drag force occur in the impeller zone, where there is acceleration.
The values of the added mass and lift forces, as plotted in Figure 5.3, are negligible over
the bulk of the tank, since there is negligible acceleration between the phases and
159

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

vorticity in the liquid is weak, but these forces do have a significant value in the
impeller zone. However, the combined added mass and lift forces as a ratio to the drag
force do not exceed ~0.2 (see Figure 5.4). This analysis suggests that the added mass
and lift forces may not be particularly important in most parts of the stirred tank, and
neglecting these forces should not have much affect on flow patterns, gas distribution
or gas holdup. However, for more accurate prediction of the gas behaviour near the
impeller, it may be appropriate to include these forces, and this should be considered in
further modelling development.

If the added mass and lift forces are to be included, a number of issues should be
addressed. One issue is the appropriate specification of values for the added mass and
lift coefficients, and whether these need to be equal to maintain frame invariance. Also,
the validity of the equations for these forces should be considered further, since in the
various derivations of these forces, the bubble diameter is assumed to be small
compared to the scales over which the continuous phase flow field varies (Hunt et al.,
1994; Mazzitelli et al., 2003), but this is hardly true in the impeller region (which is the
main region where the forces would be significant), especially in the trailing vortices,
where there are strong velocity and pressure gradients over relatively small distances.
For example, in the tank of Barigou and Greaves (1992, 1996) as modelled in
Chapter 8, the trailing vortices have a width of about 3 mm, compared to bubble sizes of
~0.51 mm. Since the liquid velocity changes rapidly over a scale about the same as the
bubble diameter, the validity of the form of the added mass and lift forces could be
questioned.

Also, if the added mass and lift forces are to be included, the stability of numerical
method needs to be investigated. Other authors have also reported difficulties with these
forces. In the simulations of bubbly flow carried out by Lathouwers (1999), the lift
force was neglected since inclusion of this force led to physically unreasonable results.
According to Lathouwers, this was because only the mean lift force was included, while
neglecting the fluctuating part. Therefore, an additional turbulent dispersion force term
may be needed to describe the effect of lift.

160

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

5.8 Turbulence in two-phase flow


Another aspect to consider in the modelling of gas dispersion is the turbulence of the
two-phase flow. While turbulence models analogous to those for single-phase flow are
usually employed, models for two-phase turbulence may need to consider additional
physical phenomena leading to modification of the turbulence, due to the interaction
between dispersed particles or bubbles and the continuous phase. Due to these
interactions, modelling of two-phase turbulence is a complex subject.

In single-phase flow, the turbulence is generated by instabilities induced by mean flow


gradients, leading to an energy cascade where the energy of larger eddies is transferred
progressively to smaller eddies and dissipated finally as viscous heat. Compared to
single-phase flow, there are a number of additional sources of velocity fluctuations in
the liquid phase in the two-phase case (Lathouwers, 1999). Firstly, the presence of the
bubbles and their relative motion through the liquid induces velocity perturbations in
the liquid, which are generally referred to as pseudo-turbulence (since the mechanism
does not represent turbulence in the usual sense). Secondly, the presence of wakes
behind the rising bubbles induces small-scale turbulent fluctuations in the liquid.
Thirdly, velocity fluctuations may be induced through the deformation of the bubble
surface by various mechanisms (e.g. collisions with eddies). According to Lathouwers
(1999), the various mechanisms of turbulence modulation are not independent, but
rather, they interact with the mean velocity field and each other. In some cases bubbles
may reduce the overall turbulent kinetic energy of the liquid rather than increasing it
(Chahed et al., 2003). It was shown by Gore and Crowe (1989) that in pipes and jets, the
presence of particles or bubbles tended to increase the turbulence intensity for
d/Le > 0.1, and decrease the turbulence intensity for d/Le < 0.1 (where d is particle size
and Le is the integral length scale). However, in terms of the simulations of stirred tanks
in this study, it may be expected that the turbulence intensity would increase somewhat
since values of d/Le > 0.1 will prevail.
A range of mathematical models have been proposed to account for two-phase
turbulence. As in single-phase flow, a common approach to modelling turbulence in
two-phase flow is to employ the k model. However, various authors have proposed to
introduce additional source terms which attempt to account for the momentum transfer

161

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

between continuous and dispersed phases. Examples of such models include those of
Elghobashi and Abou-Arab (1983), Bel Fdhila and Simonin (1992), and Tu and
Fletcher (1994).
The conservation equations for k and in two-phase flow may be written as (CFX4
Solver Manual, 2002):


T ,i
( i i k i )
+ ( i i U i k i ) = i L,i +
t
k


T ,i
( i i i )
+ ( i i U i i ) = i L,i +
t

k i + i i i i i + S k , (5.62)

2
i + c1 i i i c 2 i i i + S

ki
ki

(5.63)

where c1, c2, k and are model constants. The terms Sk and S are additional source
terms related to the interaction between the phases. i is the production of turbulent
kinetic energy by mean velocity gradients given (for incompressible flow) by:

i = ( L,i + T ,i )U i U i + U Ti i k i U i .
3

(5.64)

The turbulent viscosity, T,i, is calculated from ki and i according to:

T , i = C i

ki2

(5.65)

and the Reynolds stress is calculated assuming the eddy viscosity hypothesis applies in
each phase, hence:

i u i u i = T ,i U i + U Ti

2
i ki I .
3

(5.66)

In principal, both phases may be regarded as turbulent, and the turbulence of each phase
could be modelled by a separate set of conservation equations. However, it is more
common to solve the conservation equations for the continuous phase k and only (i.e.,
i = 1). The Reynolds stress of the dispersed phase may then be obtained using an
algebraic expression relating it to the continuous phase Reynolds stress, as was done for
example by Simonin (1991) for the case of solid particles. However, the dispersed phase
Reynolds stress in gas-liquid flow is generally regarded as negligible (Bel Fdhila &
Simonin, 1992), and therefore the gas phase is generally treated as laminar.

162

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

For the additional source terms in equations 5.62 and 5.63, Sk and S, various options are
possible. As an illustration of a modified k model, according to Bel Fdhila and
Simonin (1992) the proposed additional source terms in the k and equations are related
to the work done by the bubbles on the liquid, as given by:
S k1 = 2 f 2 v1
S 1 = C , 3

1
k1

= 2 1 FD

2
v1 v r
2 + 1C A

+ Vd Vr

S k1 ,

(5.67)

(5.68)

where C ,3 is another constant. However, other authors have proposed quite different
forms of the two-phase equations for k and , and it is difficult to reconcile the various
approaches. There is no single standard model for two-phase turbulence available at
present. Second-order Reynolds stress turbulence models have been proposed by some
authors (e.g. Lahey, 1995), although such models are highly computationally intensive
and highly complex, containing many unknown correlations which require closure
relations.

Another approach, rather than using additional source terms in the turbulence model, is
to separate the turbulent contributions of the mean flow gradients and the particles or
bubbles. Lance and Bataille (1991) found that in decaying grid turbulence with gas
volume fractions below ~1%, the total Reynolds stress could be accounted for by the
linear addition of the grid-generated component and the component due to pseudoturbulence, although deviation from the linear behaviour was observed at higher gas
fractions. They proposed that the turbulent kinetic energy could be expressed as the sum
of components due to mean flow gradients (or shear) and bubble motion given as:

k = k SI + k BI ,

(5.69)

where the bubble induced turbulent kinetic energy is given as:


1
2
k BI = 2 C A Vr ,
2

(5.70)

and in this case the equation for kSI does not contain any additional terms.
In terms of modelling the mean flow in the tank, the purpose of the turbulence model is
to calculate the turbulent viscosity. Hence, another simpler possibility is that, rather than
modifying the turbulence model, the turbulent viscosity can be modified directly. This
163

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

was proposed by Sato and Sekoguchi (1975), who assumed that contributions due to
mean velocity gradients and bubbles can be separated, and the turbulent viscosity is
calculated as the linear sum of the velocity gradient-induced turbulent viscosity
(calculated using the usual single-phase k model) and an additional component of
viscosity due to bubble motion, which is assumed always positive. Hence:

T ,1 = 1C

k2

+ T ,b ,

(5.71)

where the additional viscosity is given by:

T ,b = 1C ,b 2 d Vr ,

(5.72)

and the constant C ,b has a value of 0.6 (Sato & Sekoguchi, 1975; Sato et al., 1981).

The importance of two-phase effects in turbulence modelling depends on the type of


flow being considered. Developments of such modelling as reported in the literature
have been oriented towards simulation of flows such as solid particles in gas jets (e.g.
Tu & Fletcher, 1994), where the inertia of the particles is comparatively high, and even
relatively small quantities of dispersed particles can significantly change the turbulence
structure of the carrier phase (Tu & Fletcher, 1994). Bel Fdhila and Simonin (1992)
modelled gas-liquid flow in a pipe, where again the bubbles could be expected to have a
significant effect on the turbulence. However, in a mechanically-stirred tank, the
turbulence is externally driven by the impeller, and under typical conditions where the
gas is well dispersed, most of the energy input to the vessel is through the impeller
rather than through the gas flow. In this case, turbulence levels tend to be relatively high
and the turbulence production is expected to be mainly due to impeller-driven liquid
flow and associated velocity gradients. The effects of gas bubbles may therefore be less
important in this situation.

Given that the effect of the gas on liquid phase turbulence is likely to be fairly small, it
was decided to model this effect using the relatively simple approach of Sato et al.
(1981), and hence, k and were calculated without extra source terms. However, as
expected, the extra turbulent viscosity term represents only a small correction. This can
be illustrated by an example calculation. For typical conditions in a stirred tank, with a
gas volume fraction of 0.1, a bubble diameter of 4.0 mm, a slip velocity of 0.23 m/s
(assumed equal to the terminal velocity), and liquid density of 1000 kg/m3, the
164

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

additional turbulent viscosity due to bubble motion is calculated to have a value of


about 0.06 Pa. This may be compared with the values of turbulent viscosity predicted by
the CFD model, where for the tank configuration and operating conditions used in the
modelling in this study (as described in Chapter 8), the turbulent viscosity was found to
be in the range of about 0.510.0 Pa. Hence, by comparison, the additional viscosity
predicted by the equation of Sato et al. (1981) is quite insignificant. This suggests that
the modification to the turbulence using additional source terms in the k equations
would also have a very limited effect on the simulation results, at least for the range of
conditions considered in this study. On the other hand, if the modelling method was
applied to tanks operating with a smaller impeller power input in comparison to the gas
flow rate, i.e. a condition closer to flooding, then the issue of modelling two-phase
turbulence may need closer consideration.
Another consideration is the lack of accuracy of the k model in stirred tank
simulations, as demonstrated in Chapter 4. Although mean velocities are well predicted
and the general pattern of spatial distribution of turbulence appears reasonable, the
individual values of k and are substantially underpredicted. Hence, in modelling twophase flow, unless other major changes to the modelling method are made, it hardly
seems worthwhile to include additional source terms in the k model which will only
make a slight difference to the results.

5.9 Conclusions
This chapter has examined the derivation and closure of the two-fluid equations. It was
shown that there are two main terms requiring closure in the averaged equation for
conservation of momentum, one being the interfacial force term, and the other being the
Reynolds stress term. The interfacial force term consists of the sum of terms for drag,
added mass, lift and turbulent dispersion forces. Following the approach of Simonin and
coworkers, a detailed model has been outlined for the turbulent dispersion force. This
has been compared with other models, including an equation with the assumption of a
constant coefficient for Ctd, and the model of Drew (2001). The added mass and lift
forces have been discussed, and some issues in their implementation were identified,
although these forces have been neglected in the further development of the modelling
in this study. For the Reynolds stress term, it seems that the use of the standard k-
165

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

model should be sufficient for the case of a gas-sparged stirred tank, at least at the
volume fractions encountered in the validation cases. A term has been included in the
modelling for the additional turbulent viscosity due to bubble slip, however it was
shown here that this term is small relative to the turbulent viscosity due to mean
velocity gradients.

166

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Table 5.1 Estimation of turbulent dispersion force for representative bubble diameters and k and values.
Table includes the methods of Bel Fdhila and Simonin (1992); Drew (2001); and a modified version of
the method of Bel Fdhila and Simonin (1992), where FD is based on stagnant fluid values.
Case No.

Bubble dia. (mm)

1.0

3.0

1.0

3.0

k (m2/s2)

0.9

0.9

0.13

0.13

(W/kg)

30

30

0.4

0.4

p (s)

0.0059

0.0117

0.0059

0.0117

1 (s)

0.0041

0.0041

0.044

0.044

Method of Bel Fdhila & Simonin (1992)

T2,1 / 2 (N/m3)

3217

2700

817

484

T2, 2 / 2 (N/m3)

1041

1341

30

60

T2,3 / 2 (N/m3)

-1325

-1525

-111

-131

T2,i / 2 (N/m3)

2933

2516

736

413

Equivalent Ctd

3.3

2.8

5.7

3.2

i =1

Method of Drew (2001) (similar to Lee, 1987; Davidson, 1990)

T2 / 2 (N/m3)

225

108

325

169

Equivalent Ctd

0.25

0.12

2.5

1.3

Method of Bel Fdhila and Simonin (1992), with modified definition of T2,1

T2,1 / 2 (N/m3)

527

269

458

222

T2, 2 / 2 (N/m3)

1041

1341

30

60

T2,3 / 2 (N/m3)

-1325

-1525

-111

-131

T2,i / 2 (N/m3)

243

85

377

151

Equivalent Ctd

0.27

0.09

2.9

1.2

i =1

167

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Figure 5.1. Equivalent turbulent dispersion coefficient, Ctd, as calculated using the modified method
according to Simonin and coworkers. Values in a vertical plane half way between baffles, using the final
modelling method for the tank of Barigou and Greaves (1992, 1996) at an impeller speed of 180 rpm
(Case 1 as described in Chapter 8).

Figure 5.2. Magnitude of the drag force per unit volume of gas (N/m3) as calculated in the CFD model of
a tank stirred by a Rushton turbine, for operating conditions corresponding to Case 1 and using the final
modelling method (as defined in Chapter 8).

168

Chapter 5. Modelling Equations for Flow in Gas-Liquid Dispersions

Figure 5.3. Magnitude of the combined added mass and lift force per unit volume of gas (N/m3) as
calculated in the CFD model of a tank stirred by a Rushton turbine, for the same modelling conditions as
in Figure 5.2.

Figure 5.4. Ratio of the combined added mass and lift force to the drag force, for the same modelling
conditions as in Figure 5.1.

169

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

6.1 Introduction
This chapter discusses investigations regarding the drag coefficient for the gas bubbles.
Work on developing the CFD model has shown that the proper specification of the drag
coefficient is very important. Although averaging of the two-phase equations leads to a
number of terms describing the interphase force, the most important term will generally
be the drag force term, since in the absence of large accelerations (which are limited to
the impeller zone only), the balance between drag and buoyancy determines the slip
velocity of the gas, which in turn is the most important factor for determining gas
holdup and distribution.

Drag coefficients are usually determined from empirical correlations, such as the
correlation method recommended by Ishii and Zuber (1979). The value of the drag
coefficient from such a correlation is usually identified with that in the instantaneous,
Lagrangian equation for bubble motion. However, the CFD modelling method is carried
out on the basis of averaged transport equations for turbulent flow, as described in
Chapter 5, and the drag coefficient therefore represents an averaged value, which is not
necessarily the same as the instantaneous value, or a value in steady flow without
turbulence.

The usual published correlations for bubble drag coefficient, such as that according to
Ishii and Zuber (1979), and others recommended by Clift et al. (1978), have been
determined based on measurements of bubbles rising through a stagnant liquid, where
the only turbulence is that generated by the bubbles themselves. In a liquid with forced,
external turbulence such as a stirred tank, the bubble interacts with eddies equal to or
greater than the bubble diameter, which have sufficient energy to alter the path of the
bubble. Therefore, the bubble experiences continual accelerations and decelerations, and
the mean value of the drag coefficient is an average over these fluctuations. As
described further on in this chapter, there is evidence that the mean drag coefficient in
turbulent flow is not the same as that for bubble rise in a stagnant liquid. The mean drag

171

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

coefficient reflects a combination of effects including a variable instantaneous drag


coefficient, and the effects of instantaneous virtual mass and lift forces.

In most papers published to date, dealing with CFD simulations of gas-liquid flow in
stirred tanks, a drag correlation for the bubbles has been adopted which is based on rise
through a stagnant liquid. This has often been done without any comment at all as to
the validity of such a correlation, or acknowledgement that there may be some
modification due to the large scale turbulence generated by the impeller. An exception
was the modelling developed by Bakker (1990), who proposed a correlation to account
for the effect of turbulence. In the modelling work presented in this thesis, the
formulation of the drag coefficient was found to be very important for predicting the
correct gas distribution and gas holdup. Applying standard drag correlations was found
to give holdups of only about one third to one half of the experimentally-determined
values.

6.2 Drag coefficient in stagnant flow


Before proceeding to discuss the issue of drag coefficient in turbulent flow, the
estimation of the bubble drag coefficient is firstly reviewed briefly for the conditions
under which most data have been obtained and correlated, namely stagnant flow, where
single or multiple bubbles are released into a liquid which is otherwise at rest. For this
situation, a considerable body of literature exists pertaining to the estimation of the drag
coefficient, as reviewed, for example, by Clift et al. (1978). Although in some
circumstances drag coefficients have been derived analytically, in general the drag
coefficient must be determined though an empirical correlation. The drag coefficient
depends on several factors, including the shape of the bubble and the bubble Reynolds
number. In dense bubbly flows, the bubble-bubble interaction or swarm effect also
needs to be taken into account.

For the case of single solid spherical particles in a stagnant Newtonian liquid, a
standard drag curve has been obtained, where drag coefficient is a function of particle
Reynolds number only, and various correlations have been proposed to account for the
shape of this curve (Clift et al., 1978). For bubbles and droplets, the drag coefficient

172

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

differs from this standard curve due to the deformability of the bubble and the nonrigid interface, which may allow internal circulation within the bubble.

With regards to shape, bubbles rising in a stagnant fluid only remain spherical at small
diameters or low bubble Reynolds numbers, whereas larger bubbles (e.g. >2 mm for air
bubbles in water) tend to deform into an ellipsoidal shape, in the so-called ellipsoidal
or distorted regime, with the shorter axis being in the direction of motion. In this case,
the drag coefficient is found to depend on the Etvs number, Eo (as described below).
The shape of the bubble can also be correlated as a function of the Etvs number and
the Morton number (Clift et al., 1978). In this regime, bubbles in stagnant liquid tend
not to rise in a straight line, but to follow helical paths, possibly with some wobbling as
well, and the drag coefficient is defined in terms of the average vertical motion rather
than the motion following the actual bubble path. At yet larger sizes, bubbles tend
towards a spherical cap shape.

The drag coefficient of a bubble is also found to be strongly affected by the level of
surfactants or contaminants in the liquid. This leads to discrepancies in the measured
drag coefficient depending on the level of purity obtained. The effect of surfactants is to
cause the bubble surface to become partially immobilised, thus reducing the internal
circulation and making the bubble more like a solid particle. A marked difference in the
rise velocity of pure and contaminated bubbles is seen, especially in the range 1 < d <
10 mm, where the rise velocity in a pure system can be twice the value in
contaminated water (Clift et al., 1978). However, in terms of practical systems,
especially stirred tanks for chemical reactions, pure liquids are generally irrelevant,
since the liquid contains some reagent or substrate for carrying out a reaction.
Therefore, the drag coefficient for contaminated systems is the one of interest.
Furthermore, in developing simulations in this study, it has been assumed that the data
for comparison were obtained in a contaminated system. With the experiments of
Barigou and Greaves (1992, 1996) using air-water dispersions, even though the water
was treated by ion exchange, it has been assumed that the water still contains a
significant level of impurities.

For bubbles in a contaminated, stagnant liquid, a number of correlations are available to


estimate the drag coefficient. Clift et al. (1978) presented a generalised correlation by
173

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

which the terminal velocity can be calculated explicitly as a function of Reynolds,


Etvs and Morton numbers; however, the correlation is limited to dilute systems. In
practical bubbly flows, a further factor to consider is the interaction between bubbles, or
the bubble swarm effect. An empirical correlation method which takes into account
multiple particle effects has been proposed by Ishii and Zuber (1979). Different
equations apply depending on the bubble Reynolds number, Reb. For Reb < 1, the
Stokes regime applies. The undistorted regime applies for bubbles up to about 2 mm.
The distorted regime applies for diameters ~210 mm and the churn turbulent regime
for bubbles >10mm. They give the drag coefficient as follows:

Stokes regime:
CD =

U d
24
, where Re b = 1 T .
Re b
m

(6.1)

Undistorted (viscous) regime:


CD =

24
(1 + 0.1Re b0.75 ) .
Re b

(6.2)

Distorted (ellipsoidal) regime:


2
g d 2
0.5
CD =
Eo (1 d ) , where Eo =
and c >>d .
3

(6.3)

Churn turbulent flow regime:


CD =

8
(1 d )2 .
3

(6.4)

In these equations, a modified Reynolds number, Reb, is used, where the mixture
viscosity is given as:

m = c 1 d
d ,max

2.5 d , mas ( d + 0.4 c ) / ( d + c )

(6.5)

For gas bubbles, the maximum dispersed fraction, d ,max , is taken as 1, and the
dispersed phase viscosity, d, is taken as zero, so this simplifies to:

m = c (1 d )1

(6.6)

According to these equations, the drag coefficient is the same as that for a spherical
solid particle up to the point where the bubble becomes deformable, with CD decreasing
as Reb increases. Then, in the ellipsoidal regime, CD again increases, until the spherical
cap regime is reached, where CD becomes constant (at constant gas fraction). The drag
174

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

coefficient as a function of Reynolds number is plotted in Figure 6.1, for the case of
dilute flow.

The corresponding terminal velocities for individual bubbles are plotted as a function of
equivalent volume diameter in Figure 6.2. It is seen that for smaller bubbles, where the
viscous regime applies, the rise velocity increases approximately linearly with diameter.
Then, in the ellipsoidal regime, the bubble rise velocity becomes constant
(UT 0.23 m/s for 2.3 mm < d < 10 mm). At larger diameters, with the onset of the
spherical cap regime, the rise velocity increases again.

These correlating equations are based on a range of data mainly obtained in columns or
pipes. Hence the various regimes, as well as referring to ranges of bubble sizes, also
refer to ranges of volume fractions, since greater gas volume fraction tends to encourage
coalescence and growth of larger bubbles. For example, the churn turbulent regime is
expected to apply for voidages greater than about 0.3. In terms of the effects of multiple
particles or bubbles, it can be noted in these equations that for lower bubble sizes and
void fractions in the Stokes, viscous and distorted regimes, the effect of increasing
volume fraction is to increase the drag, similar to hindered settling of solids. However,
at higher volume fractions in the churn turbulent regime, overlapping wakes and
boundary layers between neighbouring bubbles lead to a reduction in drag, and hence,
the bubble swarm velocity is higher than the terminal velocity of an individual bubble.

The bubble swarm velocity (at volume fractions greater than about 0.3) has been
measured by other authors, e.g. Schlueter and Raebiger (1998), who found that bubbles
in a swarm had a rise velocity up to 40% higher than a single bubble. They described
the effect in terms of bubbles disturbing each other and following a more vertical and
less helical path. They did not, however, propose a generalised correlation for the effect.
Other authors (e.g. Lo et al., 1999) have proposed (similar to the expression of Ishii and
Zuber for the churn turbulent regime) that the effect of gas volume fraction on drag
coefficient can be expressed by a simple power law:

C D = C D , (1 d ) .
n

(6.7)

While n = 2 according to Ishii and Zuber (1979), Lo et al. (1999) proposed a value

n = 4.
175

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

In correlations such as the one proposed by Ishii and Zuber (1979), the liquid may be
described as stagnant, since turbulence in the liquid is only due to the bubble motion.
However, the drag coefficient may be substantially modified by external turbulence, as
discussed in greater detail in the remainder of this chapter. The correction to the drag
coefficient to account for turbulent flow has been taken as a multiplying factor on an
underlying drag coefficient, which is still assumed to follow the equations proposed by
Ishii and Zuber (1979). This assumption is discussed further in Section 6.5.

6.3 Previous studies of drag in turbulent flow


For several decades the possibility that liquid-phase turbulence may modify the drag
coefficient on particles and bubbles has been recognised and investigated, but data have
been quite limited due to experimental difficulties. Most data relate to solid particles,
whereas only a limited amount of experimental data for gas bubbles has been published,
although there are also results for numerical simulations which can provide information
about the effects on gas bubbles.

The drag coefficient in nearly all published studies has been based on an equation for
drag force of the form:

FD =

3 CD
U 2 U1 (U 2 U1 ) ,
4 d

(6.8)

where C D represents the effective drag coefficient in turbulent flow. Studies may refer
to the drag coefficient itself, or the particle slip velocity (where the particle could be
solid, gas or an immiscible liquid droplet). Such data should be equivalent, since under
steady conditions (no acceleration of the mean flow), these are related through equating
buoyancy force to drag force, leading to the simple relationship:

CD =

4 gd 2 1 1
,
3 1
U S2

(6.9)

where US = U2 U1. Data are often presented as the ratio US/UT, where US is the
measured slip velocity and UT is the terminal velocity in stagnant liquid. According to
the above, where the ratio C D C D , 0 is required rather than US/UT , where CD,0 is the
value in stagnant liquid, this can be obtained according to:
176

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

CD
1
=
C D ,0 U
S
UT

(6.10)

Brucato et al. (1998b) reviewed experimental efforts to determine turbulent drag


coefficient in various systems. As described in their paper, the earliest studies date back
to the 1960s (Torobin & Gauvin, 1961; Clamen & Gauvin, 1969), where both
reductions and increases in drag were measured depending on conditions. However,
particle sizes in these early studies were rather large and turbulence intensities were
relatively low, so that the data do not cover the typical range of conditions of practical
applications in the process industries.

The earliest study of drag coefficient and particle slip velocities in a stirred tank was
carried out by Schwartzberg and Treybal (1968), who found that particle settling
velocities were 3050% of those in stagnant conditions. Another study of particles in a
stirred tank (Nienow & Bartlett, 1974) found settling velocities in the range 3070% of
those in stagnant conditions.

Laser doppler velocimetry was applied by Nouri and Whitelaw (1992) to attempt to
measure particle settling velocities in a stirred tank. Reductions in settling velocity were
observed in some cases. However, as pointed out by Brucato et al. (1998b), the method
is affected by a number of uncertainties: in particular, the slip velocity is obtained as the
small difference between the particle and liquid velocities, which are both typically
much larger than the slip velocity, so that the measured slip velocity is subject to a large
error.

Most of the more recent work has avoided direct measurement of particle velocities, due
to the difficulties in obtaining accurate measurements. In the approach of Magelli et al.
(1990), solids concentration profiles were measured in a tall vessel fitted with multiple
impellers. The axial profiles of solids concentration were interpreted by means of a
sedimentation-dispersion model. The work was extended to include a wider range of
experimental conditions by Pinelli et al. (1996). In these studies it was found that
settling velocities were in the range 25100% of that expected in stagnant flow. Results

177

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

were correlated with the parameter K/d, where K is the Kolmogorov scale of
turbulence, although there was quite a high degree of scatter around the correlation line.

Brucato et al. (1998b) developed an alternative experimental setup based on a


Taylor-Couette type vessel, where an inner cylinder rotates within a fixed outer cylinder
and the annular space between them defines the vessel volume. This system was chosen
since high turbulence levels can be generated in the gap while having small axial
velocities, making the system suited to measuring particle slip velocities. The
experimental technique involved introducing a small quantity of particles at the top of
the vessel, and measuring the concentration as a function of time at two lower levels,
using a non-intrusive laser attenuation method. The resulting data could be analysed by
a residence time model to obtain settling velocity data. Glass beads and silica particles
in various size ranges (average diameters ranging ~65460 m) were tested, and settling
velocities were generally found to be reduced, to as low as ~15% of the stagnant value
for the largest glass beads (425500 m). The authors found that the data were quite
consistent with the data of Magelli et al., and a trend was observed by plotting US/UT
versus K/d, as illustrated in Figure 6.3. Thus, similar results were obtained despite large
differences in the overall flow pattern in a stirred vessel compared to a Taylor-Couette
vessel. A correlation for particle settling velocity was proposed by Brucato et al.
(1998b), according to which (in rearranged form for clarity):
d
CD
= 1 + 8.76 10 4
C D ,0
K

(6.11)

According to Brucato et al. (1998b), the mechanism by which the drag coefficient is
increased is not completely clear. Possible mechanisms may include acceleration
effects, where the particle experiences on-going accelerations, and in conjunction with
these accelerating motions, there is a non-linear relationship between drag and particle
velocity, and additional forces due to the virtual mass.

Another study which is somewhat related to this topic is that of Tunstall and Houghton
(1968), who investigated the motion of various solid beads in a sinusoidally oscillating
column of water. Experimental measurements revealed substantial retardation, with the
settling velocity reduced to as low as 40% of the value in stagnant water for 1.59 mm
178

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

glass spheres. Numerical simulations were also carried out (Houghton, 1968), based on
a one-dimensional model of particle motion. Similarly, these simulations demonstrated
a varying reduction in settling velocity, depending on the velocity and amplitude of the
liquid oscillations. Of course, an essential difference in this work was that the liquid
oscillations consisted of a single, regular velocity fluctuation, whereas turbulence is
characterised by random, three-dimensional velocity fluctuations with a spectrum of
length scales. Hence, it is difficult to make much use of this work.

For gas bubbles in a turbulent liquid, no experimental data were available until very
recently. However, a correlation was proposed by Bakker (1992) for the effect of
turbulence on drag coefficient, and this was apparently developed intuitively without
any data. According to Bakker, the bubble slip velocity should be reduced when a
bubble is moving in a turbulent flow field due to increased momentum transport around
the bubble. Similarly to the idea in the paper of Brucato et al. (1998b), he suggested that
the extent of the effect should depend on the ratio between the bubble size and the
turbulent length scale, so that the effect would be zero where the bubble size equals the
Kolmogorov length scale, and increases as the bubble size exceeds the Kolmogorov
length scale. He proposed that the effect of turbulence could be accounted for by using a
modified Reynolds number,
Re b =

1 U s d
,
1 + C* 1k12 / 1

(6.12)

where the viscosity is taken as the sum of the liquid laminar viscosity and a term
proportional to the liquid turbulent viscosity. The drag coefficient is obtained by
inserting this modified (and smaller) Reynolds number in a correlation for the drag on
spherical particles under stagnant conditions (ignoring the ellipsoidal regime). Although
Bakker recognised that the effect of turbulence should depend on the ratio of bubble
size to turbulence length scale, he preferred to simplify the situation by proposing a
constant value for C* , which was probably necessary due to a lack of any data for
different bubble diameters. Bakker suggested that value of this constant should be less
than C in the equation for turbulent viscosity, since not all turbulent eddies of all length
scales will affect the momentum transport around the bubbles. By tuning his
computational model to match simulation results with experimental measurements of
holdup, a value of 0.02 was found for C* .
179

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

There are a small number of published papers in which the question of the effect of a
turbulent flow field on bubble drag has been addressed through a theoretical or
computational approach, rather than through experiment in the laboratory. Maxey et al.
(1994) carried out numerical simulations to investigate the effect of turbulence on the
rise velocity of small bubbles (less than about 0.5 mm). The numerical method involved
the application of a random, statistically stationary forcing function to a grid of
96 96 96 cells, to generate a flow field with homogeneous, isotropic turbulence,
where the Kolmogorov length scale of turbulence and bubble diameters were similar in
magnitude. Particle tracking was applied for bubbles released at random positions into
the flow, and averaging was applied to obtain the statistics of the bubble motion. The
instantaneous drag coefficient was assumed to be given as CD = 48/Re, and an equation
of motion was assumed for the bubbles, which included terms for drag and added mass
forces but excluded the lift force. They found that the slip velocity reduced as the ratio

p/K was increased, where K is the Kolmogorov time scale, with a reduction of 40% in
slip velocity at p/K equal to 1.0.

Spelt and Biesheuvel (1997) investigated the motion of gas bubbles about 1.0 mm
diameter in homogeneous, isotropic turbulence, by carrying out numerical simulations
to determine the statistical properties of the average bubble motion. In considering how
the average bubble motion could be related to the properties of the bubble and the
characteristic parameters of the turbulence, they quoted Hunt et al. (1994), according to
whom the bubble motion through a turbulent flow is characterised by three
dimensionless groups:
uo
L11
T
,
, L
U T bU T b

(6.13)

where u0 is the r.m.s. turbulent velocity of the liquid, L11 is the integral length scale of
turbulence, b is the bubble relaxation time, and TL is the large eddy time scale.
However, according to the analysis of Spelt and Biesheuvel (1997), TL b was not
considered as an independent group, but rather it was taken to be a function of the other
two parameters (since in their simulation, the value of TL (L11 / u 0 ) is approximately

180

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

constant). Spelt and Biesheuvel proposed that the relevant dimensionless groups for
characterising bubble motion are:

uo
,
UT

* =
* =

(6.14)

L11
,
bU T

bU T

(6.15)

(6.16)

where they introduce an additional dimensionless group defined in terms of the Taylor
microscale, T, so as to allow for comparison of different turbulent energy spectrum
functions. However, for a given turbulent energy spectrum equation, a relationship is
fixed between * and *, so data can then be considered to be a function of just two
dimensionless groups, and *.

The average rise velocity of bubbles was determined by carrying out numerical
simulations, in which bubbles were tracked as they moved through an isotropic
turbulent flow field generated by forcing functions (in similar fashion to Maxey et al.,
1994). These forcing functions provided a distribution of random eddies, whose
statistics correspond to specified energy spectra with defined values of turbulent
macroscale, Taylor microscale, and turbulent kinetic energy. Two energy spectra were
considered, being the Kraichnan and von Karman-Pao spectra. The Kraichnan function
is considered to be representative of low-Reynolds number turbulence behind a grid,
while the von Karman-Pao spectrum is representative of high Reynolds number
turbulence.

Statistical properties were obtained by time-averaging over each bubble trajectory, and
ensemble-averaging over a large number of bubble tracks. An equation of motion of the
bubbles was specified which takes into account drag, virtual mass and lift. The rise
velocity in stagnant water was quoted as 25 cm/s for a 1.0 mm bubble, and the
(instantaneous) drag coefficient was assumed to follow the linear law for a bubble in
uncontaminated water, which is CD = 48/Re. Simulations were carried out over a range
of values of the turbulence parameters, and it was found that the velocity of bubble rise
could be markedly reduced, to as low as 50% of that expected in stagnant liquid.
181

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Results were analysed in terms of the dimensionless groups proposed, and were also
shown to be in agreement with approximate analytical solutions at low values of
(where is defined in equation 6.14). The results show that US decreases with
increasing turbulence intensity, , and a stronger effect is observed at smaller *. These
trends show that the rise velocity is reduced to a greater extent with increasing turbulent
kinetic energy, and at constant turbulent kinetic energy the effect increases with smaller
integral scale, which may be because eddies of a similar length scale as the bubble occur
more frequently.

Spelt and Biesheuvel suggested a mechanism for the reduction in the bubble rise
velocity. According to their analysis, when a bubble interacts with a turbulent eddy, the
effect of the lift force is such that the bubble tends to move towards regions where the
difference between the velocity of the bubble and that of the fluid is largest. Since
bubbles move upwards on average, they tend to move to regions of maximum
downward liquid flow. Due to this uneven sampling of the liquid flow, biased toward
regions of downward fluctuating velocity, the average rise velocity of the bubbles is
reduced.
Besides an approximate analytical solution for low values of , Spelt and Biesheuvel
did not offer any general correlation for US/UT, as is required for input to a CFD model.
In the same university department where simulations were carried out as described in
the paper of Spelt and Biesheuvel (1997), experimental measurements were carried out
for the bubble slip velocity in turbulent flow (Poorte & Biesheuvel, 2002). This work
was aimed at providing a check on the previous theoretical work of Spelt and
Biesheuvel, and although there is a fair amount of scatter in the results, these
experimental results provide fairly convincing confirmation of the previous numerical
results.

The apparatus of Poorte and Biesheuvel consisted of a recirculating water tunnel fitted
with a glass-walled vertical test section, in which turbulence was generated by means of
a mechanically active grid. This grid was constructed of a set of bi-plane rods fitted with
182

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

agitator wings, which was programmed to flap in a random manner to generate


approximately isotropic turbulence. Bubbles were tracked by a light scattering
technique to determine the ensemble-averaged statistics of bubble motion. Eight data
points were obtained for bubble diameters 0.68 mm and 1.14 mm over a range of
turbulence conditions. Using this method, reductions of up to 35% in bubble rise
velocity were measured. It was found that except for one data point, measured mean rise
velocities differed from the theoretical predictions of Spelt and Biesheuvel by less than
9% of UT.
For correlating their data, they defined another length scale of turbulence, L:

kE (k )dk ,
L=
E (k )dk
0

(6.17)

where E(k) is the Eulerian energy spectrum function of the turbulence. According to
these authors, at low values of , an approximate analytical solution can be obtained as:

US
3
1
= 1 2
,
UT
4
L*

(6.18)

where
L* =

bU T
L

By plotting their data as US/UT versus

(6.19)

L , good agreement with this theoretical

equation was obtained at low , however the function did not extrapolate well to higher

values.
In another study (Miettinen et al., 2002), experimental measurements were reported for
bubble slip velocity in a small laboratory tank stirred by a four-bladed open paddle.
Axial slip velocities of air bubbles in tap water were obtained at a number of positions
as the difference between liquid and gas velocities, as measured by particle image
velocimetry (PIV). The slip velocities were plotted as a function of bubble size, as
shown in Figure 6.4. It can be seen that the magnitudes of the slip velocities are
substantially reduced, e.g. for a 2 mm bubble, the highest measured slip velocity was
about 0.11 m/s compared to about 0.23 m/s in stagnant liquid. Also, it was found that
the slip velocities decreased with increasing intensity of turbulence, with the lowest

183

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

velocities being measured in the impeller discharge stream (position D). This finding
is consistent with that of Spelt and Biesheuvel.

6.4 Development of a correlation for use in CFD simulations


A review of the literature reveals the need for a correlation for drag coefficient in the
CFD model which takes into account the effects of turbulence. The correlations
proposed by Brucato et al. (1998b) and Bakker (1992) have been considered, as
discussed below, but were not found to be suitable. Since no correlation has been found
in the literature which was suitable for this purpose, a new correlation has been
developed. This was based on available data in the literature, with the expectation that
the form of correlating equation would be a function of liquid and gas properties, and
other variables available in the CFD model, such as bubble size, k and .

According to the derivation of the averaged conservation equations following Simonin


and coworkers (Simonin & Viollet, 1989; Bel Fdhila & Simonin, 1992), as outlined in
Chapter 5, the mean drag force on the gas phase should have the form:
CD
3
F2 = 2 1
d
4

vr

(U 2 U1 ) .

(6.20)

This equation was first considered, together with the closure relation proposed by Bel
Fdhila and Simonin (1992) for mean total magnitude of slip velocity, v r

, which

has components for the mean slip and the fluctuating slip (equations 5.30 and 5.37).
Some preliminary calculations of the drag coefficient, C D

, and relative velocity,

(U2-U1), were carried out following this approach, with the drag coefficient in the
viscous regime calculated from a standard correlation based on a Reynolds number as
a function of

vr

. Analysis of this approach indicated that average slip velocities

were generally reduced to some extent, although with high values of k an increase in
slip velocity was found. In general, the predicted values of C D

were not in good

agreement with data from the literature. Therefore, another correlating method was
needed for the drag coefficient. Also, equation 6.20 is very inconvenient for fitting of
data, since as well as the unknown parameter, C D

184

, there is an additional unknown

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

term, being the mean total magnitude of slip velocity,

vr

. For practicality,

therefore, the mean total magnitude of slip velocity has been assumed to be based
simply on the difference between the average phase velocities, |U2 - U1|, and a
correlation has been developed for a mean effective C D , based on the simpler form of
equation 6.8.

Reviewing the literature has disclosed a range of data related to drag in turbulent flow
conditions. Nearly all experimental data are for solid, heavy particles (mostly silica or
glass), while the information for gas bubbles is mainly based on numerical simulations,
with a small number of experimental measurements reported more recently. This data
can provide the basis for the correlating equations required in CFD modelling.
Nevertheless the data remain incomplete in terms of the range of particle/bubble
properties and turbulence parameters covered, and there is a need for further
measurements.

Two correlations for drag in turbulent conditions have appeared in the literature, one
(Brucato et al., 1998b) based on experiments and being for solid particles (and not
necessarily suitable for gas bubbles), the other (Bakker, 1992) being for gas bubbles,
but limited in its formulation and not based on actual data. Both these equations have
been tested in the CFD modelling development in this study.

The correlation according to Brucato et al. (1998b) was considered, since although that
equation was based on measurements of solid particles, bubbles might follow similar
trends if the mechanism by which the drag coefficient is modified involves a similar
mechanism. However, the correlation of Brucato and coworkers was not found to be
suitable, since the predicted effect on the drag coefficient was found to be much too
large for the conditions applicable to gas bubbles in a stirred tank. This can be
illustrated by reference to the operating conditions of the stirred tank, as used in the
development of the two-phase model, described in detail in Chapter 8. In the simulation
defined in Chapter 8 as Case 4, the average bubble size, d, is about 4 mm and the
average turbulent energy dissipation rate is about 2.7 W/kg, which implies an average
Kolmogorov length scale, K, of 25 m. With these values, the drag coefficient in
turbulent flow is 3589 times higher than the value for rise in a stagnant liquid, and the
185

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

predicted slip velocity is only 1.7% of the stagnant-flow terminal velocity. If local
values of turbulent energy dissipation rate are used rather than averages over the tank,
even stronger effects are predicted. In either case, the reduction in slip velocity is much
too large. A simulation of the tank stirred by a Rushton turbine was attempted using
equation 6.11 for the drag coefficient, but it was found that gas became trapped and
accumulated excessively in the lower recirculation loop of the tank due to unrealistically
low slip velocities.

It might be possible to adopt the form of the correlation of Brucato and coworkers, but
3

. This would require


with a reduced value of the coefficient before the term d

fitting of an equation of this form to available data for bubbles. However, a different
approach was pursued, as described below, where the inertial response of particles to
turbulence is taken into account, rather than just the ratio of length scales of the particles
and the turbulence. The inertial response of particles depends not only on the particle
size, but on factors such as the density difference between the particle and the liquid and
particle/liquid density ratio. However, the form of the correlation according to Brucato
and coworkers does not take particle or liquid density into account.

The approach of Bakker (1992) was also considered in the CFD modelling, where the
drag coefficient is calculated with a modified Reynolds number according to equation
6.12. However, the effect on drag was again too strong for the values of k and found in
the stirred tank, with operating conditions as used in the modelling development. Again,
gas became trapped and accumulated excessively in the lower recirculation loop of the
tank. The value of the fitted constant could be changed, however, Bakker suggested that
the value of C* was probably not constant, but depends on the characteristics of the
bubbles and the turbulence. Therefore, another modelling equation would be necessary
for C* . It was preferred to follow another approach since it is unclear why the effect of
turbulence should be expressed in terms of an additional turbulent viscosity.

Given that the correlations of Brucato et al. (1998b) and Bakker (1992) were found to
be unsatisfactory, it was decided to make use of available literature data to attempt to
develop another correlation, possibly taking into account data for both solid particles
186

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

and gas bubbles, and aiming to have a more theoretically sound basis compared to the
correlations of Brucato et al. (1998b) and Bakker (1992). By including data for gas and
solids, such a correlation should provide a higher level of confidence. It might be
expected that a correlation for solid and gas particles might take a similar form if
densities of particles were taken into account, at least where approximately spherical
bubbles are compared with spherical solid particles.

Representative data have been selected from the studies of Brucato et al. (1998b), Spelt
and Biesheuvel (1997), and Poorte and Biesheuvel (2002). With the data of Brucato et
al., the analysis has been limited to glass spherical particles with size ranges 212
250 m micron and 425500 m (they obtained similar results for irregular silica
particles, but these are not included here). With the data of Spelt and Biesheuvel, the
analysis has been limited to the results obtained using the von Karman-Pao spectrum
with fixed integral scale. These data are the most relevant to a stirred tank, since high
Reynolds number turbulence is expected.

It would also have been interesting to consider the results of Miettinen et al. (2002) for
the slip velocity of bubbles in a tank stirred by an open paddle. However, these results
were not available at the time that this part of the study was carried out. Also, the
interpretation of their results would require further information regarding the values of k
and at different positions, specifically for a tank stirred by an open paddle rather than
a Rushton turbine.

For the range of data considered, the reduction in rise velocity needs to be accounted for
in terms of the properties of the particles/bubbles and the turbulence. With the
assumption of a fixed turbulence spectrum, two parameters of the turbulence should be
sufficient to characterise the turbulent eddies interacting with a bubble. These
parameters could be, for example, a velocity scale and a length scale. One parameter
could be the turbulent kinetic energy or the turbulent fluctuating velocity, u0. Another
parameter is required to specify the spread of the energy over the spectrum of
turbulence wavelengths. This parameter could be the integral length scale for
example, where the integral length scale is larger, the turbulent energy is spread over a
wider range of wavelengths, so that for a given eddy size, the frequency is lower. It
should be equivalent to specify a characteristic length scale such as the integral scale,
187

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

L11, or to specify the turbulent energy dissipation rate, , since there should be a
relationship between k, , and L11 , such that, given any two variables, the third can be
calculated. Wu and Patterson (1989) gave the equation:
3

L11 = A

(6.21)

where A is constant of order unity.

The dimensionless groups proposed by Spelt and Biesheuvel (1997) were considered in
the analysis, since these take into account u0 and L11. The equation of Brucato et al.
(1998b) is too simplistic, since it only compares the length scales of turbulence and
particle, without considering the energy or frequency of the eddies or the inertia of the
particle. Poorte and Biesheuvel suggested another length scale for correlation, as
defined in equation 6.17, however this intermediate length scale was not adopted as it is
not readily determined in terms of the parameters k and only, which are all that are
available in the CFD model.

Hence, following Spelt and Biesheuvel, it is proposed that the ratio of slip velocity in a
turbulent flow field to that in stagnant flow, US/UT, is a function of and *. The form
of this function is not immediately clear, however, to simplify the situation further, it
was proposed that and * might be combined in some way as a single variable.
It was observed that in general, the reduction in slip velocity increases in proportion to
and in inverse proportion to *. Also, if the two dimensionless groups are combined
simply as /* then one can write:
u U
u

= 0 b T = 0 b
*
U T L11
L11

(6.22)

In isotropic turbulence one can write:

L11 = TL u 0 ,

(6.23)

so substituting for L11, the expression reduces to:

b
=
.
* TL

(6.24)

Hence, combining and * in this manner reduces the parameters to the ratio of
characteristic time scales, which is a physically meaningful dimensionless group, often
188

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

called the Stokes number, being a measure of how quickly the particle adapts its speed
to an interacting eddy. This parameter has been used in explaining the effect of
turbulence on particle dispersion (Pozorski & Minier, 1998). This dimensionless group
is preferable to a ratio of length scales as used by Magelli et al. (1990) and Brucato et al.
(1998b), since inertia is taken into account, i.e. two particles of different densities have
the same length scale, but the particle with a greater density difference compared to the
liquid has a greater relaxation time, so a greater modification by turbulence is predicted.
There is no apparent fundamental reason why and * should combine in this way, but
for the available data, a good fit was obtained, and it is preferable to keep the correlation
as simple as possible, at least until further data become available. In addition, various
combinations ()n(*)m were tested, but this was the only one for which the gas bubble
data and solids data were found to collapse onto single continuous curves, with the
curves for gas and solids also falling close to each other or overlapping.

It can also be noted that, by combining the two dimensionless groups in this way, one
gets the third dimensionless group which was proposed by Hunt et al. (1994), but
discarded by Spelt and Biesheuvel. Thus, the approach taken here implies an
assumption that there is only one independent dimensionless group governing the effect
of turbulence, not three as proposed by Hunt et al.

In the correlation of Brucato et al. (1998b), the turbulence is characterised by the


Kolmogorov length scale, which can be calculated according to:

3
K =

0.25

(6.25)

For calculating the Stokes number, the turbulence is characterised by TL, which can be
calculated according to (Bel Fdhila & Simonin, 1992):

k
TL = 0.135 .

(6.26)

Hence, it is seen that the correlation of Brucato et al. describes the turbulence in terms
of only, while using TL provides a two-parameter characterisation of the turbulence, in
terms of both k and .

189

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

The relaxation time of a bubble or particle can be calculated according to (Bel Fdhila &
Simonin, 1992):

p =

2
+ CA
1
3 CD
UT
4 d

(6.27)

where CD is the drag coefficient in stagnant liquid and CA is the added mass coefficient.
This parameter also contains more information than using the bubble diameter alone,
since the densities of particle and liquid are taken into account, as is the bubble shape,
implicit through the assumed bubble regime and associated drag coefficient. For
bubbles, assuming 2 0 and CA = 0.5, the expression for relaxation time simplifies to:

b =

UT
,
2g

(6.28)

as given by Spelt and Biesheuvel (1997).

Further interpretation of the data of Brucato and coworkers was required, to recalculate
the data as values of p and TL. The data presented in their paper consist of settling
velocities of particles of different sizes over a range of cylinder speeds, with the
corresponding power input, where the settling speed reduces as cylinder speed
increases. From this data, the particle relaxation times are readily calculated. However,
the integral times of turbulence were not reported. Values of TL are calculated here by
first estimating the average turbulent integral length scale, L11. This was assumed to be
equal to the gap width in the apparatus, which is 8 mm. Using this value of L11, and a
mean value of based on the power input, values of the average turbulent kinetic energy
were obtained using equation 6.21, with a value of 0.85 for A, following Wu and
Patterson (1989). Mean values of k and were then used to calculate TL in accordance
with equation 6.26.
Hence, the data of Brucato and coworkers could be re-interpreted as sets of values of ,

*, and /* (which is the same as the Stokes number, p/TL). These data and the
calculation procedure, for glass beads with average sizes 220 m and 460 m, are
summarised in Tables 6.1 and 6.2.

190

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

The numerically-generated data of Maxey et al. (1994) are summarised in Table 6.3.
They presented data for US/UT as a function of p/TK (ratio of particle relaxation time to
Kolmogorov time scale), however since they report that the simulations had a constant
ratio of 16 between the Kolmogorov and integral scales, these data can be readily
recalculated as US/UT versus p/TL.

The data for bubbles as generated by the numerical methods of Spelt and Biesheuvel are
summarised in Table 6.4 (being the data obtained for a fixed integral length scale and
using the von Karman-Pao spectrum). The experimental data of Poorte and Biesheuvel
are summarised in Table 6.5.

The ratio of turbulent slip velocity to stagnant fluid terminal velocity, US/UT is plotted
versus the ratio p/TL (equivalent to /*) in Figure 6.5. This figure includes the data of
Brucato et al. (1998b), Maxey et al. (1994), Spelt and Biesheuvel (1997), and Poorte
and Biesheuvel (2002). In all cases, the data from each source are clearly correlated
with /*. It is remarkable that within a certain degree of scatter, all the data fall on
approximately the same line, with the exception of the data of Maxey et al. This is
despite the data originating from distinctly different methods (experimental and
numerical) and being either for heavy solid particles or gas bubbles.

The data of Maxey et al. (1994) fall on a lower, but parallel curve. The effect of
turbulence seems very strong in these simulations, and it is difficult to reconcile this
data completely with the other data. Analysis of their results shows that even for a
bubble of the same length scale as the Kolmogorov length scale, a reduction in rise
velocity of about 25% is predicted, whereas the results of other authors indicate that the
rise velocity of a bubble of this size should not be affected by turbulence. The
simulation results of Maxey et al. (1994) might be questioned since they did not include
the lift force in their calculations. Also, there is no experimental confirmation of their
results, and given these uncertainties, their data have not been considered further.

For the remaining data, a relatively simple equation that provides a fit through all data
points is:

191

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

US
= exp 2.1
UT
TL

(6.29)

An equation that provides a good fit specifically for the data of Spelt and Biesheuvel is:

p
US

= exp 1.7
UT

TL

(6.30)

The data of Poorte and Biesheuvel are shown in Figure 6.6 for greater clarity and are
compared with the proposed correlating equation for bubbles, equation 6.30. It can be
seen that the data only extends to small values of p/TL, and there is a fair degree of
scatter in the data. Nevertheless, these data points are consistent with equation 6.30.

There are a number of uncertainties in interpreting the data for the Taylor-Couette
apparatus; for example, the estimated values of k and may only be approximate. The
estimates of k may be in error because equation 6.21 applies strictly only for fully
developed, homogeneous isotropic turbulence, while at lower speeds in the TaylorCouette apparatus, turbulence may not be fully developed. Also, single average values
of k and are assumed, while turbulence is not strictly uniform in the apparatus, but
instead shows a pattern reflecting the Taylor vortices (see Parker & Merati, 1996). Also,
the particles may not sample all parts of the flow field equally, but may experience
some centrifuging, away from the centre of the Taylor vortices and away from the inner
wall. In Figure 6.5, some of the data for solids fall on a line higher than that for the gas,
and one explanation for this could be that average effective k values have been
overestimated (and therefore TL is overestimated).
Uncertainties also exist in the applicability of the results of the numerical modelling of
Spelt and Biesheuvel, due to the various assumptions applied, such as homogeneous,
isotropic turbulence, a linear instantaneous drag equation, and the assumed values of
added mass and lift coefficients. However, the experimental results of Poorte and
Biesheuvel (2002) provide a fair degree of confirmation of the numerical results.

For the CFD model, a curve specifically fitting the data of Spelt and Biesheuvel (1997)
has been preferred, rather than a correlation including data for both bubbles and solid

192

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

particles, since Spelt and Biesheuvels data refers specifically to gas bubbles, and
experimental confirmation of Spelt and Biesheuvels numerical results has been carried
out, at least for low values of u0/UT. The fact that the data for solids almost coincide,
provides for additional confirmation of the form of the equation. As discussed, the
differences between the data for solids and bubbles may reflect a number of
uncertainties. These differences could also point to a real difference between solids and
bubbles, or may reflect a deficiency in the correlation method.

6.5 Additional considerations for the CFD model


Further consideration needs to be given as to how the correlating equations, equations
6.29 and 6.30, extrapolate to higher values of p/TL. The available experimental and
computational data were limited to values of p/TL less than one, but in a stirred tank, it
is possible to have larger values, and this actually occurs in the chosen simulation
conditions for the tank stirred by a Rushton turbine, in the region of the impeller
discharge stream, where turbulence intensities are high and integral length scales are
relatively low.

Considering the limit of small bubbles with very small relaxation times, it would be
expected that the effects of turbulence will become negligible, since the bubbles can
follow all fluctuations in the liquid flow field and there is no acceleration between the
phases. Therefore, small bubbles should remain at their normal terminal velocity, and as
found in the various data in the literature, in the limit p/TL 0, the ratio US/UT is 1.
Then, as p/TL increases, the effects of acceleration become evident, due to the
increasing response time of the bubbles to turbulent velocity fluctuations, and it is seen
that US/UT decreases. But in the limit of p/TL becoming very large, this implies either
that the bubble is very large (a high relaxation time), or the characteristic time scale of
the turbulence is much lower than that of the bubbles. In the first case, the inertia of the
bubble should be so great that eddies cannot deflect it from its path, while in the second
case, the frequency of the eddies is so high that the bubble does not have time to
respond. Hence, in the limit as p/TL becomes large, the ratio US/UT should return to 1.0.

193

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Having discerned the limits for US/UT as p/TL becomes very small or large, it would be
useful to be able to predict the shape of the curve at values of US/UT larger than those
for which experimental data are available.

To explore this matter further, one can consider the trends in the particle fluctuating
velocity and the fluctuating relative velocity between phases. The mean square values of
these fluctuating velocities can be estimated in terms of the fluctuating velocity of the
continuous phase according to (Pozorski & Minier, 1998):
1

v1 2 ,

(6.31)

p

T
L
v r 2 = v1 2 ,
p

1 +
T
L

(6.32)

v 2 2 =

p
1 +
T
L

where v1 is the r.m.s. fluctuating velocity of the liquid in one dimension, so that with
the assumption of isotropic turbulence, one can write:
v1 2 = 2 k .
3

(6.33)

These expressions are similar to those of Simonin and coworkers, as outlined in


Chapter 5, but with b = 0 in equations 5.33 and 5.37, and ignoring the cross trajectory
effect. These expressions show the expected limiting behaviour. The magnitude of the
fluctuating particle velocity, expressed as the root mean square value,

v2 2 , becomes

equal to that of the continuous phase as p/TL 0, because the particle readily follows
all fluctuations of the liquid. On the other hand, as p/TL , the fluctuation of the
particle velocity goes to zero, either because p is very large, so that the particle has no
response to turbulence, or TL 0, so there is no time for the particle to accelerate. The
fluctuating relative velocity,

vr 2 , shows the opposite limiting behaviour. The

magnitude of the fluctuating relative velocity goes to zero in the limit of p/TL 0,
since the particle follows all liquid fluctuations. Whereas, as p/TL , the fluctuating
relative velocity is equal to the liquid fluctuating velocity, since it sees the complete
continuous phase fluctuation, having no fluctuation of its own.

194

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

These expressions have typically been used in modelling of turbulent particle dispersion
(e.g. Pozorski and Minier, 1998), and in such a case the main interest is in calculating
the displacement caused by eddy interactions, leading to estimates of the dispersion
coefficient. Here, these expressions may assist in obtaining an estimate of the extent of
acceleration of the particle, since under accelerating conditions, particle slip may be
reduced due to factors such as changes in the instantaneous drag coefficient and the
action of added mass and lift forces. Where acceleration effects are at a maximum, it
might be expected that the reduction in slip velocity should be the greatest. The
maximum effect may be expected to occur at some intermediate value of p/TL, where
the particle does not follow all the liquid fluctuations, but there is sufficient interaction
for acceleration during each encounter with an eddy.

The approach taken here is to propose that the reduction in slip velocity is proportional
to the energy input due to acceleration of a particle by an eddy. This can be calculated
as the work done, W, in acceleration of a particle during some eddy lifetime te, which is
either the product of force and displacement, or the product of force and particle
velocity integrated over the time of the eddy, i.e.:
W =

te
0

F2 v2 dt =

te
0

dt

( 2 + C A 1 ) dv 2 v2 dt .

(6.34)

Since the details of how the velocity and acceleration of the particle vary as a function
of time are not known, and since there is a range of eddy lifetimes, the average work has
been estimated by a simplified approach. The work is instead estimated as the product
of a characteristic force (mass times acceleration) and a characteristic displacement
during a typical particle-eddy collision. The displacement is taken as the product of a
characteristic velocity and a typical timescale, Tm. Hence:
W ( 2 + C A 1 )

dv2 , m
dt

v2 , m Tm .

(6.35)

where v2 , m is a mean, characteristic velocity. Furthermore, the acceleration of the


particle in response to an eddy, relative to the liquid velocity, is estimated as the
characteristic change in relative velocity during the interaction time, which is:
dv2 , m
dt

vr , m
Tm

(6.36)

195

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

It is assumed further that the force and the displacement in this simplified picture are
colinear and both positive. These velocities, v2 , m and vr , m , have then been estimated
as the root mean square of the particle fluctuating and relative velocities, as given in
equations 6.31 and 6.32. Thus, the work done is taken as being proportional to the
product of

v2 2 and vr 2 .

Substituting for the fluctuating particle velocity and the fluctuating relative velocity, the
work done, or energy input, due to particle acceleration, can be estimated as:

p
W =2

( 2 + 1C A )k

TL

p
1 +
T
L

(6.37)

The fluctuating velocities and the energy input (at constant k) due to eddy interaction
are plotted versus p/TL in Fig 6.7. The work function rises rapidly to a maximum at

p/TL = 1, and then for higher values of p/TL it tapers off slowly. If the reduction in slip
velocity due to turbulence is proportional to this energy input, then this suggests that the
curve for US/UT as a function of p/TL has a minimum value at p/TL = 1, where the
energy input is a maximum, and thereafter the effect on slip velocity should diminish,
with US/UT = 1 for p/TL . Experimental values obtained at values of p/TL close to
1 suggest that the curve has a minimum value of about 0.15 at p/TL= 1. Since the
energy input tapers off at higherp/TL, this suggests that there is no point at which US/UT

0, and hence particles or bubbles are not expected to be trapped completely by


turbulence.

The shape of this curve suggests that there are significant effects on slip velocity at
quite high values of p/TL, but this needs further examination in terms of the physical
situation. Large values of p/TL can be achieved either through very large or inert
particles, or through ever decreasing values of TL. Limiting the discussion to practical
particles of interest (i.e. bubbles with diameters about 15 mm), a large p/TL ratio
implies a small eddy lifetime, and the meaning of very small TL values should be
considered. Since the large eddy time scale can be given as:

196

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

TL =

L11
,
u0

(6.38)

then decreasing values of TL means either decreasing values of the integral length scale,
or increasing values of u0. However if the integral length scale decreases to the extent
that d >> L11, then there will be no turbulent eddies of sufficient size or energy to
deflect the bubble from its path. If decreasing TL is achieved this way, then the curve for
the work done will return more rapidly to zero than suggested in Figure 6.7.

If the bubble size remains within the range of eddy sizes, it is still possible to achieve
high values of p/TL by making u0 very large at constant integral length scale. This
means that the fluctuating velocities are large, but the interaction time is very short. The
total displacement of the bubble will be the product of fluctuating velocity and eddy
time, but this displacement should be compared with the non-zero physical dimensions
of the bubble (which is often treated as if it were a point). The ratio of the displacement
caused by interaction and the bubble radius can be written:
X =

v 2 TL d
d

2=

1
1+

p
TL

v1TL
2
1 L11
1 =
1.
d
p d
3
2
1+
TL

(6.39)

As p/TL becomes large at constant integral length scale, it is seen that a point is
inevitably reached where the displacement during interaction becomes small compared
to the bubble diameter or radius. Under such conditions, it seems unlikely that the
turbulence would affect slip velocity, since there is little interaction with the structure of
the turbulent field, and little distance (or time) over which the bubble can change its
drag coefficient. Hence as X 0, US/UT 1.

As seen from above, the point at which X becomes small depends on L11/d as well as

p/TL. The situation is further illustrated for just one case, where a bubble of diameter 2
mm is rising through turbulence with L11 = 8 mm (this is typical of the conditions in the
discharge stream of the Rushton turbine under the conditions of the CFD simulations).
Under these conditions it can be calculated that X falls to zero at p/TL approximately
equal to 8.

197

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

To demonstrate the effect of the real physical size of the bubble, it is proposed to
multiply the work done by a modifying factor as a function of X. The form of this
function is not known, but for illustration a function with the required properties is:
Fdisp = 1 exp( X ) .

(6.40)

The modified work done is given by:


W = FdispW .

(6.41)

The modified work function, W , is plotted as a function of p/TL in Figure 6.8.

Based on an assumption that the reduction in slip velocity depends in a linear fashion on
the work done in accelerating the bubble, another function can be plotted of the form:
Us
= 1 CW .
Ut

(6.42)

where C is some constant. This curve can be made to fit roughly with the literature data
at low p/TL values for a value C = 3.6. The continuation of this curve predicts,
hypothetically, the shape of the curve at higher p/TL values. This is illustrated as the
example function in Figure 6.8. It should be pointed out that since X is a function of
L11/d, this is not a universal curve but one of a series of curves for different values of
L11/d.

Since this analysis gives a rather complicated form of the function for US/UT , which in
any case remains hypothetical, it has been preferred to fit the results to as simple a curve
as possible, while extrapolating to higher values of p/TL with a curve that is in
agreement with the above analysis.

The above analysis suggests a curve which flattens out with a minimum value at US/UT
0.15 for values of p/TL around 1.0, while at higher values of p/TL the curve
eventually returns to US/UT = 1.0. No information has been found regarding the effects
of turbulence at p/TL >1.0, so the shape of the full curve is not known, and it is not
known how large p/TL becomes before the value of US/UT returns to 1.0. Nevertheless,
a correlation has been proposed for the modelling work which assumes this shape of
curve. An equation which was tested in the CFD modelling development, and found to
work fairly well, is given by:

198

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

p
US
= 1 1.4
UT
TL

0.7

exp 0.6
TL

(6.43)

This equation gives a return to 1.0 for US/UT at about p/TL = 10. The equation is shown
in Figures 6.9 and 6.10.

In the CFD model, it is the drag coefficient rather than the slip velocity which is
specified in the code, so C D C D , 0 is calculated from US/UT according to equation 6.10,
and the drag coefficient is specified by multiplying the drag coefficient for bubble rise
in a stagnant liquid by this factor.

It should be noted that the correlation has been based on data for spherical bubbles, up
to about 1.0 mm diameter. However, in the stirred tank, bubble sizes are larger than this
in nearly all parts of the tank, ranging up to ~5.0 mm equivalent spherical diameter. For
stagnant flow, this puts the bubbles mostly in the distorted regime, where an
ellipsoidal shape is assumed. Due to lack of any other data, it is assumed that the drag
coefficient can be corrected in the same way, as a multiplying factor to the value
calculated by the distorted regime equation. However, the application of the distorted
regime correlation is somewhat uncertain. The assumption of an ellipsoidal shape
should be questioned, since the effect of constant bombardment by eddies would be to
alter the shape from ellipsoidal, possibly making the bubble more spherical due to the
random approach of eddies from different directions. This effect has been observed
experimentally (Machon et al., 1997). Also, the bubble drag coefficients in this regime
are based on a regular, helical upward path, but the path of the bubbles will be altered
by the induced random motions of the bubbles due to turbulence. Therefore, the
distorted regime correlation of Ishii and Zuber (1979) may not be strictly applicable.
Further investigation is needed.

199

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

3.2
Spherical cap regime

2.8

Drag coefficient

2.4
Ellipsoidal regime for
bubbles

2
1.6
1.2
0.8

Standard drag curve

0.4
0
10

100

1000

10000

Reynolds number
Figure 6.1 Drag coefficient as a function of Reynolds number for air bubbles in water according to the
correlation of Ishii and Zuber (1979).

Terminal velocity (m/s)

0.3
0.25
0.2
0.15
0.1
0.05
0
0

12

16

20

Diameter (mm)
Figure 6.2 Terminal velocity of a single bubble in stagnant liquid as a function of bubble equivalent
diameter, for air bubbles in water according to the correlation of Ishii and Zuber (1979).

201

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Figure 6.3. Ratio of measured slip velocity and terminal velocity in stagnant fluid, as a function of the
ratio of Kolmogorov scale and particle diameter (Brucato et al., 1998b). Data for solid particles as
obtained by Magelli et al. (1990) (small dots) and Brucato et al. (1998b) (larger symbols).

Figure 6.4. Local axial slip velocities of bubbles in air-water system, as measured with PIV by Miettinen
et al. (2002) in a 13.8 litre tank stirred by a paddle, for a gassing rate of 0.25 l/min and impeller speed of
400 rpm.
position C; position D; position E; position F (as defined in Miettinen et al., 2002).

202

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

1.2

Brucato et al 460
micron
Brucato et al 220
micron

S & B mu = 1.25

Us/Ut

0.8

S & B mu = 5.0
Maxey et al

0.6

Data of P & B
Correlation for
S&B; Brucato

0.4

Correlation for gas


bubbles
0.2

0
0.000

0.200

0.400

0.600

0.800

1.000

1.200

tp/tL

Figure 6.5. Data for US/UT plotted against p/TL (=/*) for solid particles and gas bubbles, with
correlation equations.

1.2

Us/Ut

0.8

Data of Poorte & Biesheuvel


Correlation for S&B data

0.6

0.4

0.2

0
0

0.01

0.02

0.03

0.04

0.05

0.06

beta/mu

Figure 6.6. Comparison of the data of Poorte and Biesheuvel (2003) with the correlation according to
equation 6.30.

203

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

1
0.9
0.8
0.7
0.6
Fluctuating particle velocity
Fluctuating relative velocity
Product of fluctuations

0.5
0.4
0.3
0.2
0.1
0
0

10

tp/tL

Figure 6.7. R.m.s. fluctuating particle velocity, r.m.s. fluctuating relative velocity, and the work done due
to acceleration by turbulent eddies, expressed as W/(2/3(2+1CA)k).

1
0.9
0.8
0.7
0.6
Example function for Us/Ut
Modified work

0.5
0.4
0.3
0.2
0.1
0
0

10

tp/tL
Figure 6.8. Modified work function, expressed as W/(2/3(2+1CA)k), and example of curve for slip
velocity, where the reduction in slip is assumed proportional to the modified work function.

204

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

1.2

Brucato et al 460
micron
1

Brucato et al 220
micron

Us/Ut

0.8

S & B mu = 1.25

S & B mu = 5.0

0.6

Maxey et al
0.4

Data of P & B
0.2

Extended
correlation
0
0.0

1.0

2.0

3.0

4.0

5.0

6.0

7.0

8.0

9.0

10.0

tp/tL

Figure 6.9. Extended correlation proposed for higher values of p/TL according to equation 6.43, with
data from the literature as in Figure 6.5. Simulations in Chapter 8 indicate that region of interest extends
to p/TL 5.

1.2

Brucato et al 460
micron
1

Brucato et al 220
micron

Us/Ut

0.8

S & B mu = 1.25

S & B mu = 5.0

0.6

Maxey et al
0.4

Data of P & B
0.2

Extended
correlation
0
0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

tp/tL

Figure 6.10. Extended correlation proposed for higher values of p/TL according to equation 6.43. Scale
on horizontal axis adjusted to show fit with data more clearly.

205

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Table 6.1. Data of Brucato et al. for 215250 m glass beads and interpretation for development of
correlation. Integral length scale, L11, is estimated to be 8.0 mm, and * is 54.2.

Speed

Power (W)

US (cm/s)

US/UT

(rpm)

u0

/* or

(W/kg)

(cm/s)

p/TL

2.69

1.000

0.00

0.00

0.000

50

0.01

2.69

1.000

0.00

0.00

0.000

100

0.05

2.69

1.000

0.00

0.00

0.000

150

0.12

2.69

1.000

0.00

4.53

0.031

200

0.25

2.55

0.948

0.03

9.75

0.067

250

0.44

2.44

0.907

0.10

13.89

0.095

300

0.7

2.24

0.833

0.19

17.20

0.118

350

1.03

2.14

0.796

0.30

20.11

0.138

400

1.44

2.15

0.799

0.44

22.84

0.157

450

1.93

2.08

0.773

0.61

25.42

0.174

500

2.52

1.98

0.736

0.82

27.97

0.192

550

3.22

1.88

0.699

1.06

30.48

0.209

600

4.07

1.91

0.710

1.35

33.07

0.227

650

5.05

1.66

0.617

1.69

35.62

0.244

700

6.17

1.56

0.580

2.08

38.15

0.262

750

7.44

1.48

0.550

2.51

40.66

0.279

800

8.85

1.37

0.509

3.00

43.12

0.296

206

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Table 6.2. Data of Brucato et al. for 425500 m glass beads and interpretation for development of
correlation. Integral length scale, L11, is estimated to be 8.0 mm, and * is 8.63.

Speed

Power (W)

US (cm/s)

US/UT

(rpm)

u0

/* or

(W/kg)

(cm/s)

p/TL

6.74

1.0

0.00

0.00

0.000

50

0.01

6.74

1.0

0.00

0.00

0.000

100

0.05

6.58

0.976

0.00

0.00

0.000

150

0.12

4.7

0.697

0.00

4.53

0.078

200

0.25

4.89

0.726

0.03

9.75

0.168

250

0.44

4.74

0.703

0.10

13.89

0.239

300

0.7

3.73

0.553

0.19

17.20

0.296

350

1.03

3.08

0.457

0.30

20.11

0.346

400

1.44

2.75

0.408

0.44

22.84

0.393

450

1.93

2.32

0.344

0.61

25.42

0.437

500

2.52

1.98

0.294

0.82

27.97

0.481

550

3.22

2.09

0.310

1.06

30.48

0.524

600

4.07

2.03

0.301

1.35

33.07

0.569

650

5.05

1.88

0.279

1.69

35.62

0.612

700

6.17

1.48

0.220

2.08

38.15

0.656

750

7.44

1.32

0.196

2.51

40.66

0.699

800

8.85

1.02

0.151

3.00

43.12

0.741

207

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Table 6.3 Data of Maxey et al. (1994) for bubbles (around 0.5 mm diameter) in isotropic turbulence.

p/K

p/TL

(or/*)

208

uo
UT

US/UT

0.2

0.0125

7.38

590.4

0.8

0.4

0.025

3.69

147.6

0.73

0.6

0.0375

2.46

65.6

0.726

0.8

0.05

1.85

36.9

0.7

0.0625

1.48

23.6

0.58

1.2

0.075

1.23

16.4

0.59

1.4

0.0875

1.05

12.0

0.57

1.6

0.1

0.92

9.2

0.55

1.8

0.1125

0.82

7.3

0.53

Chapter 6. The Mean Drag Coefficient in Turbulent Flow

Table 6.4. Data of Spelt and Biesheuvel (1997) as applied in development of correlation: results of
numerical simulations of 1.0 mm air bubbles in water (UT = 25 cm/s) (data obtained with von KarmanPao spectrum and two different constant values of integral length scale).

* = 1.25

* = 5.0

US/UT

/* or

uo
UT

US/UT

p/TL

/* or

uo
UT

p/TL

0.9

0.12

0.096

0.96

0.12

0.024

0.7

0.19

0.152

0.875

0.19

0.038

0.6

0.26

0.208

0.79

0.26

0.052

0.55

0.33

0.264

0.75

0.33

0.066

0.5

0.42

0.336

0.71

0.42

0.084

0.42

0.55

0.44

0.68

0.55

0.11

0.4

0.66

0.528

0.63

0.66

0.132

0.33

0.82

0.656

0.62

0.82

0.164

0.3

0.8

Table 6.5. Data of Poorte and Biesheuvel (2003) obtained in water tunnel with mechanically active grid.

Case No.

Bubble radius
(mm)

uo
UT

/* or

US/UT

p/TL

E1

0.57

0.066

20.6

0.0032

1.051

D1

0.57

0.115

18.8

0.0061

0.974

B2

0.34

0.443

43.0

0.0103

0.79

A1

0.57

0.123

11.9

0.0103

1.054

C1

0.57

0.161

13.3

0.0121

0.935

F2

0.34

0.428

31.2

0.0137

0.65

B1

0.57

0.200

7.0

0.0286

0.923

F1

0.57

0.212

5.7

0.0372

0.852

209

Das könnte Ihnen auch gefallen