Академический Документы
Профессиональный Документы
Культура Документы
A critical review of the published literature regarding the computational fluid dynamics (CFD) modelling of single-phase turbulent flow
in stirred tank reactors is presented. In this part of review, CFD simulations of radial flow impellers (mainly disc turbine (DT)) in a fully
baffled vessel operating in a turbulent regime have been presented. Simulated results obtained with different impeller modelling approaches
(impeller boundary condition, multiple reference frame, computational snap shot and the sliding mesh approaches) and different turbulence
models (standard k − ε model, RNG k − ε model, the Reynolds stress model (RSM) and large eddy simulation) have been compared with
the in-house laser Doppler anemometry (LDA) experimental data. In addition, recently proposed modifications to the standard k − ε models
were also evaluated. The model predictions (of all the mean velocities, turbulent kinetic energy and its dissipation rate) have been compared
with the experimental measurements at various locations in the tank. A discussion is presented to highlight strengths and weaknesses of
currently used CFD models. A preliminary analysis of sensitivity of modelling assumptions in the k − ε models and RSM has been carried out
using LES database. The quantitative comparison of exact and modelled turbulence production, transport and dissipation terms has high-
lighted the reasons behind the partial success of various modifications of standard k − ε model as well as RSM. The volume integral of predicted
energy dissipation rate is compared with the energy input rate. Based on these results, suggestions have been made for the future work in this area.
Nous présentons un examen critique de la littérature concernant la modélisation de la dynamique des fluides numérique (DFN) de l’écoulement
turbulent à une phase dans les réacteurs à cuve agitée. Dans cette partie de l’examen, nous présentons les simulations de DFN de turbines à
écoulement radial (principalement des turbines à disque (TD)) dans un réservoir entièrement cloisonné effectuées dans un régime turbulent. Les
résultats des simulations obtenus grâce à différentes approches de modélisation des turbines (couche limite turbulente, méthode des référentiels
multiples, snap-shot de modélisation numérique, maillage glissant) et à différents modèles de turbulence (modèle standard k-e, modèle RNG
k-e, modèle aux tensions de Reynolds et simulation des grandes échelles) ont été comparés aux données expérimentales internes d’allocation de
Dirichlet latente (ADL). De plus, les modifications des modèles standards k-e récemment proposées ont également été évaluées. Les prédictions
du modèle (de toutes les vitesses moyennes, de l’énergie cinétique turbulente et de son taux de dissipation) ont été comparées aux données
expérimentales relevées à différents endroits de la cuve. Une discussion présente les points forts et les points faibles des modèles de DFN actuellement
utilisés. Une analyse préliminaire de la sensibilité des hypothèses de modélisation liées aux modèles k-e et RSM a été menée en utilisant la base de
données LES. La comparaison quantitative des données exactes et modélisées de la production, du transport et de la dissipation de la turbulence
a mis en évidence les raisons qui expliquent la réussite partielle de plusieurs modifications apportées au modèle standard k-e ainsi qu’au modèle
aux tensions de Reynolds. L’intégrale de volume du taux de dissipation énergétique prévue a été comparée au taux d’intrant énergétique. Sur la
base de ces résultats, d’autres études à venir dans ce domaine ont été suggérées.
Keywords: CFD, stirred vessel, radial flow impellers, impeller models, k − ε models, RSM, LES
S
tirred tank reactors, in which one or more impellers are E-mail addresses: jb.joshi@ictmumbai.edu.in, jbjoshi@gmail.com
used to generate desired flow and mixing, are amongst the Can. J. Chem. Eng. 89:23–82, 2011
most widely used reactors in chemical and allied indus- © 2011 Canadian Society for Chemical Engineering
DOI 10.1002/cjce.20446
tries. Stirred reactors offer unmatched flexibility and control over
Published online in Wiley Online Library
various transport processes occurring within the reactor. The (wileyonlinelibrary.com).
performance of a stirred reactor can be optimised by appro- This paper was meant to be part of the GLS 2010 Special Issue - PART A
priate adjustments of the reactor hardware and the operating (CJCE 20465 (Part B) shall be published soon).
1 Mixing time Peters and Smith (1969); Carreau et al. (1976); Brennan and Lehrer (1976); Khang and Levenspiel (1976); Pandit and Joshi (1983); Pandit et al.
(1984); Shiue and Wong (1984); Sano and Usui (1985); Raghava Rao and Joshi (1988); Raghava Rao et al. (1988); Takahashi et al. (1988); Saito
and Kamiwano (1989); Saito et al. (1990); Ranade et al. (1991); Rewatkar and Joshi (1991a); Rewatkar and Joshi (1991b); Carreau et al. (1992);
|
al. (2003); Patwardhan et al. (2004); Kumaresan and Joshi (2006); Kumaresan et al. (2006)
2 Power consumption Rushton et al. (1950a); Rushton et al. (1950b); Metzner and Otto (1957); Metzner et al. (1961); Godleski and Smith (1962); Bates et al. (1963);
Beckner and Smith (1966); Bourne and Butler (1969a); Novák and Rieger (1969); Nagata et al. (1970); Hall and Godfrey (1970); Fort et al.
(1971); Nienow and Miles (1971); Bertrand et al. (1980); Rewatkar and Joshi (1991a); Sarvanan et al. (1996); Karcz and Major (1998); Dohi et al.
(2004); Deshmukh and Joshi (2006)
3 Heat transfer Mizushina et al. (1966a); Mizushina et al. (1966b); Coyle et al. (1970); Edney and Edwards (1976); De Maerteleire (1978); Shamlou and Edwards
(1986); Wang and Yu (1989); Carreau et al. (1994)
4 Mass transfer Feranandes and Sharma (1967); Miller (1971); Levins and Glastonbury (1972); Kuboi et al. (1974); Boon-Long et al. (1978); Conti and Sicardi
(1982); Lai et al. (1988); Armenante and Kirwan (1989); Jadhav and Pangarkar (1991); Hiraoka et al. (1993); Barigou and Greaves (1996); Gezork
et al. (2001); Fishwick et al. (2003); Martı́n et al. (2010)
5 Scale-up procedures Rieger and Novák (1972); Khang and Levenspiel (1976); Bowen (1985); Costes and Couderc (1988); Obot (1993); Ogawa and Kuroda (1995);
Montante et al. (2003); Gimbun et al. (2009)
6 Multiphase flooding Miller (1974); Yung et al. (1979); Warmoeskerken and Smith (1985); Tanaka and Izumi (1987); Wong et al. (1987); Takahashi and Nienow (1992);
Birch and Ahmed (1997); Paglianti et al. (2000); Bombač and Žun (2006)
7 Gas–liquid dispersions Komasawa et al. (1970); Miller (1974); Joshi and Sharma (1976); Figueiredo and Calderbank (1979); Joshi and Kale (1979); Rewatkar and Joshi
(1991b); Rewatkar and Joshi (1991c); Rewatkar and Joshi (1993); Rewatkar et al. (1993); Mhetras et al. (1994); Tecante et al. (1996); Parsu Veera
et al. (2001); Joshi and Patwardhan (2002); Thatte et al. (2004); Scargiali et al. (2007a); Ford et al. (2008); Martı́n et al. (2010)
8 Gas–liquid–solid dispersion Raghava Rao and Joshi (1989); Rewatkar and Joshi (1992); Ebrahimi-Moshkabad and Winterbottom (1999); Micale et al. (2000); Nienow and
Bujalski (2002); Fishwick et al. (2003); Murthy et al. (2007b); Chen et al. (2009)
9 Liquid–liquid dispersion Sprow (1967); Chen and Middleman (1967); Brown and Pitt (1972); Weinstein and Treybal (1973); Coulaloglou and Tavlarides (1977); Brooks
|
(1979); Rounsley (1983); Davies (1985); Okufi et al. (1990); Kumar et al. (1991); Armenante et al. (1992); Ribeiro et al. (1996); Zhou and Kresta
(1998); Colenbrander (2000); Kamienski and Wójtowicz (2003); Derksen and van Den Akker (2007); Kadam et al. (2009)
10 Solid–liquid dispersion Patil et al. (1984); Raghava Rao et al. (1988); Rewatkar et al. (1989); Rewatkar and Joshi (1991a); Rewatkar and Joshi (1991b); Armenante et al.
(1992); Guiraud et al. (1997); Nocentini et al. (2000); Fishwick et al. (2003); Fajner et al. (2008); Unadkat et al. (2009)
11 Fractional dispersed phase Mlynek and Resnick (1972); McManamey (1979); Sembira et al. (1988); Pacek et al. (1994); Arjunwadkar et al. (1998); Kulkarni et al. (2001);
hold-up Brunazzi et al. (2004); Shewale and Pandit (2006); Sun et al. (2006); Boden et al. (2008); Kadam et al. (2009)
12 Size distribution and Magelli et al. (1990); Magelli et al. (1991); Nishikawa et al. (1994); Pacek et al. (1998); Sessiecq et al. (1999); Angeli and Hewitt (2000); Ochieng
concentration profile of and Lewis (2006); Montante et al. (2008); Fajner et al. (2008)
dispersed phase
13 Breakage and coalescence Madden and Damerell (1962); Komasawa et al. (1970); Mizoguchi et al. (1973); Brown and Pitt (1974); Molag et al. (1980); Narsimhan et al.
(1984); Das et al. (1987); Barigou and Greaves (2009); Shimizu et al. (1999); Zaccone et al. (2007)
14 Chemical reactions Komasawa et al. (1970); Robinson and Wilke (1973); Uchida et al. (1978); Leszek et al. (2004); Patwardhan et al. (2005); Milewska and Molga (2010)
15 Self-inducing impellers Joshi and Sharma (1977); Sawant and Joshi (1979); Joshi (1980); Sawant et al. (1981); Rielly et al. (1992); Forrester and Rielly (1994); Forrester et al.
(1998); Poncin et al. (2002); Murthy et al. (2007a); Scargiali et al. (2007b); Murthy et al. (2008a)
16 Flow visualisation Peters and Smith (1967); Bourne and Butler (1969b); Murakami et al. (1972); van’t Riet and Smith (1975); Günkel and Weber (1975); Hiraoka et al.
(1979); Kuriyama et al. (1982); Mochizuki and Takashima (1984); Kuboi and Nienow (1986); Yianneskis et al. (1987); Ranade and Joshi (1989);
Wu and Patterson (1989); Ranade and Joshi (1990a); Shervin et al. (1991); Mishra and Joshi (1991); Kumar et al. (1993); Mishra and Joshi (1993);
Chapple and Kresta (1994); Bakker and van den Akker (1996); Desai and Joshi (1996); Schäfer et al. (1997); Fentiman et al. (1998); Mishra et al.
(1998); Mavros et al. (2001); Alvarez-Hernández et al. (2002); Szalai et al. (2004); Nikiforaki et al. (2003); Kumaresan et al. (2005); Arratia et al.
|
25
|
for blending and heat transfer are quite different (which impeller rotation speed. However, understanding of turbulent
require more bulk flow and less shear) from those desired for flow structures close to wall and their dependency on type
gas–liquid dispersion and mass transfer (which require more of impeller can help us to design energy efficient impellers.
shear). In case of aerobic fermentation, the oxygen trans- This understanding can be developed using advanced EFD
fer rate increases with hydrodynamic shear. However, the and CFD. These tools will give transient details of veloc-
validity, growth rate and the selectivity of microorganisms ity field inside the reactor. This database can be used to
(particularly animal and plant cell culture) and the activ- estimate various properties of flow structures such as age
ity of proteins and enzymes decreases with an increase in distribution, size and shape distribution, length scale distri-
shear. Such conflicting requirements make the task of evolv- bution, energy distribution, penetration depth distribution,
ing “wish list” for the desired fluid dynamics difficult. Reactor etc. This information can be used in combination with clas-
engineer has to achieve appropriate compromise between the sical theories of heat and mass transfer such as film theory,
conflicting process requirements to achieve the best results. penetration/surface newel theory, small eddy theory, large
Using a computation model, one can switch on and off vari- eddy theory, surface divergence theory. Most of these theo-
ous processes, which otherwise is not possible while carrying ries use assumed flow field and consider only one of the above
out experiments. Such parametric experimentation can give mentioned features. Accounting only one of these properties,
useful insight into interactions between different processes that too on an averaged scale may not be appropriate (Joshi
and can help to resolve the conflicting requirements. et al., 2009). To obtain accurate predictions, the velocity
(2) Translating batch data for continuous reactors: In most of the field data need to be analysed using eddy isolation algo-
cases, laboratory and bench scale experiments required to val- rithms based on zero crossing in fluctuating velocity signals,
idate the reactor concept are carried out in a batch mode. It wavelet based methods and proper orthogonal decomposition
is then necessary to translate (or to use) the data obtained to obtain structure age, size, energy, and shape distribution
in such experiments for designing continuous reactors. The and then obtain contribution of each and every flow structure
location of feed pipes, outlets and their influence on mixing and average it over space and time. Another way is to carry
and performance needs to be understood. Computational flow out direct numerical simulation (DNS) and/or LES of stirred
models can be of great help in this regard. tank with energy equation and obtain heat transfer coeffi-
(3) Scale-down/scale-up analysis: It is essential to analyse the pos- cient for various impeller designs. The objective is impeller
sible influence of scale of reactor on its fluid dynamics and should give appropriate combination of convection and tur-
eventually the performance. It should be noted that small- bulence close to impeller and tank surface can be designed
scale reactor would invariably have higher shear and more in such a way that the surface renewal frequency can be
rapid circulation than large-scale reactor. The interfacial area improved. Such kind of efforts has already been implemented
per unit volume of reactor normally reduces as the scale of in industry in heat exchangers, where inside surface is cor-
reactor increases. Scale-up/scale-down analysis is important rugated/microfins are provided. Microfins are found to be
to plan useful laboratory and pilot plant tests. It may be almost 50 times energy efficient compared to bare smooth
often necessary to use pilot reactor configuration, which is pipes for heat transfer applications. This is due to localised
not geometrically similar to the large-scale reactor in order to breaking of flow structures close to wall due to microfin struc-
maintain the similarity of the desired process. Conventionally ture.
such analysis is carried out based on certain empirical scal-
ing rules and prior experience. Computational flow modelling This brief review of the steps in engineering of stirred reactors
can make substantial contributions to this step by providing indicate that availability of large degrees of freedom regard-
quantitative information about the fluid dynamics. ing reactor configuration, impellers can be effectively exploited
(4) Testing of new reactor concepts: More often than not, develop- to evolve better reactor technologies. This, however, requires
ment of reactor technologies relies on prior experience. New detailed knowledge and understanding of the fluid dynamics in
reactor concepts are often sidelined due to lack of resources all the regions of the stirred reactors. In line with such a broad
(experimental facilities, time, funding, etc.) to test them. view, it was thought desirable to critically examine the predic-
Experimental studies have obvious limitations regarding the tive capabilities of CFD and to identify the limitations as regards
extent of parameter space that can be studied and their sub- to the expected accuracy. Second and third sections present the
sequent extrapolation beyond the studied parameter space. review of the published models for the turbulence (single phase)
A wide variety of impellers with different shapes are used and the impeller–baffle interaction, respectively. Fourth and fifth
in practice. Different practices of impeller clearance, etc., sections provide the comprehensive survey of CFD of stirred tank
are followed for different impellers and for different appli- with radial flow impellers using RANS and LES.
cations. Computational flow models, which allow “a priori” In order to appreciate the relative merits of these models,
predictions of the flow generated in a stirred reactor of any simulations of the flow field in the stirred tank using various
configuration (impellers of any shape), with just a knowledge impeller-modelling approaches along with the turbulence models
of geometry and operating parameters, can make valuable by the commercial code as well as in-house codes were carried out.
contributions in developing new reactor technologies. An attempt has been made to compare the model predictions of
(5) Development of theories for heat and mass transfer: In case all the three mean velocity components and the turbulent kinetic
of exo/endothermic reactions in stirred vessel, heat needs to energy with the experimental data with emphasis on bulk region.
supplied/removed from the reactor. External jackets or inter- Both the radial flow impeller (part I of this review) and axial flow
nal coils can be provided for this purpose. In majority of impellers (part II of this review) have been covered.
the cases, heat transfer coefficient inside the reactor dictates In part I, second section gives details of various turbulence
the overall heat transfer coefficient. This can be increased by modelling approaches such as DNS, large eddy simulation (LES),
increasing energy dissipated close to tank wall or outside sur- RSM, standard k − ε model, and modified two-equation models.
face of coils. This objective can be satisfied by increasing the The various impeller–baffle interaction models are presented in
capture important phenomena in the process of transition, as 0.15 m with Reynolds number of 7275. The authors used MRF
well as turbulence itself. DNS is mathematical, and therefore, can approach for simulation impeller–baffle interaction, though they
be used to create simplified situations that are not possible in an mentioned that for DNS simulations sliding mesh (SM) technique
experimental facility, and can be used to isolate specific phenom- is necessary. Their geometry was described by a block structured
ena in the transition process. As we are aiming at a nearly exact grid of 15 blocks. The grid size also was small (1 901 824) for DNS
solution (and not “the exact”) to specific turbulent flows utilising as compared to (2 082 816) for k − ε model for the same half simu-
limited computational resources, DNS is stressed as a research lation of the vessel. The average Kolmogorov scale was 6 mm and
tool and not as a brute-force solution to engineering problems. average grid size was 0.8 mm. As the impeller region has very
The objective of DNS is not necessarily to reproduce real-life high shear as well as energy dissipation, 35% of the grid points
flows (say the flow over an airplane), but to perform controlled were located in the impeller region (which consists only 7% of
studies that provide better insight, scaling laws, and turbulent reactor volume) to resolve the trailing vortices. Although both the
models to develop. In aerodynamics, DNS is associated with a simulations (DNS and k − ε model) were in good agreement with
large-scale computationally intensive solution procedure, which the experimental data the authors found that DNS captured few
may consume hundreds to thousands of Cray super-computing more details than k − ε model. The DNS simulations showed two
resources. The earliest use of DNS began in the 1970s and with small secondary vortices behind the blades which were missed by
the growth in the computational power today, it is getting more the k − ε model, even the flow structure predicted by k − ε model
and more popular day-by-day. DNS of stirred tanks has been car- was wavier than DNS. The shape of regions with increased tur-
ried out in literature by Bartels et al. (2000); Verzicco et al. (2004); bulent kinetic energy was closer to experimental results for DNS
Sbrizzai et al. (2006); and Gillissen and van den Akker (2009). as compared to k − ε model, but in the outer flow region k − ε
These simulations are reported at very low Reynolds number rang- model gave more uniform results of turbulent kinetic energy. The
ing from 1600 to 7300. The simulations carried out by Verzicco maximum energy of turbulent kinetic energy by DNS was much
et al. (2004); and Sbrizzai et al. (2006) were in unbaffled tank higher (k/U2tip = 0.118) and closer to the experiments, whereas
whereas Bartels et al. (2000); and Gillissen and van den Akker k − ε model (k/U2tip = 0.079) underpredicted.
(2009) carried out simulations in baffled tank. These two inves- Gillissen and van den Akker (2009) simulated six bladed Rush-
tigations have been briefly reported below. ton turbine and the impeller–baffle interaction was modelled with
Bartels et al. (2000) simulated stirred tank with Rushton turbine adaptive force field technique (AFT). Their grid size was of 1
using k − ε model and DNS and compared the results of both with billion and Reynolds number of 5000. They simulated 100 000
experimental results of Schäfer et al. (1998). The simulations were time steps for a month on 128 cores. The authors focused on
performed with silicon oil as working fluid in a tank diameter of comparison between LES and DNS. They also showed
Large Eddy Simulation The Reynolds stresses (−∂ui uj /∂xj in Equation 8) arising out
In LES, large scales eddies are resolved and the small scales, which of ensemble-averaging procedure are modelled according to the
are isotropic in nature, are modelled using subgrid scale models. Boussinesq hypothesis and given by:
The major role of subgrid scale models is to provide proper dis-
sipation for the energy transferred from the large scales to small 2 ∂ui ∂uj
scales. Because of these advantages, LES technique is gradually −ui uj = kıij −t + (9)
3 ∂xj ∂xi
becoming a popular tool to investigate turbulent flows. LES stud-
ies have several valuable advantages. They are (i) much cheaper In order to close the system of equations (7)–(8), t has been for-
(ii) easier to handle than field measurements. They provide (iii) mulated in many ways in the literature in terms of zero, one, and
three-dimensional fields of flow characteristics under (iv) well- two-equation models. Some of the popular models are discussed
controlled external condition. They are (v) more accessible in the below.
sense that LES can be run in the conditions where field measure-
ments would be difficult. Moreover, they are (vi) very helpful for Standard k − ε model
conceptual understanding of detail hydrodynamic of the chemi-
Standard k − ε model is essentially a high Reynolds number model
cal process system. The motivation behind LES is the recognition
and assumes the existence of isotropic turbulence and the spectral
that the large scales of the turbulence often dominate mixing,
equilibrium. In the case of standard k − ε model, t is estimated
heat transfer, and other quantities of engineering interest, while
using turbulent kinetic energy (k) and rate of dissipation of tur-
the small scales are only of interest because of how they affect the
bulent kinetic energy (ε):
large ones. Furthermore, large scales are not universal and vary
according to the geometry and flows under consideration, while
k2
small scales exhibit a more or less universal behaviour, which t = C (10)
considerably simplifies the task of modelling the SGS stresses. ε
In an LES, the flow field is decomposed into a large-scale or
The modelling of k and ε equations result into five turbulence
resolved component and a small-scale or subgrid-scale compo-
parameters C , Cε1 , Cε2 , k and ε . These parameters have been
nent:
estimated from studies in simple flows. In the log-law region of
the boundary layer, experimentally it has been found that the
u = ū + u (4)
turbulence production and the dissipation terms are much larger
than the other terms. Turbulence in this region is considered to be
The conservation equations governing the filtered velocity
in local equilibrium. Neglecting transport and production in this
field are obtained by applying the filtering operation to the
flow, value of C = 0.09 has been obtained. For the value of Cε2 ,
Navier–Stokes equations (Equations 1 and 2). We consider
experiments have been carried out in decaying grid turbulence.
spatially uniform filters, so that filtering and differentiation com-
Turbulence is generated when free flow goes through a grid which
mutes.
generates mean flow gradients and in turn via the production
When the Navier–Stokes equations for constant density (incom-
term generates turbulence. Sufficiently far downstream, the veloc-
pressible flow) are filtered, one obtains a set of equations very
ity gradients are zero and hence the production and dissipation
similar in form to the RANS equations:
terms are zero in contrast to log-law region in the boundary layer.
∂ū The turbulent diffusion term is also negligible. The balance of
=0 (5) turbulence convection and dissipation gives Cε2 = 1.92. The con-
∂xi
stant Cε1 is obtained by looking at the dissipation rate equation
∂ū ∂(ūū) 1 ∂p̄ ∂2 ū ∂ in the log-region of boundary layer. The turbulent diffusion is not
+ =− + 2− (6) negligible in this case unlike k equation. The simplifications of ε
∂t ∂xi ∂xi ∂xi ∂xi
equation using velocity profiles in log law region and the numeri-
Various subgrid scale models are required for estimation of cal solution of these equations give Cε1 = 1.44, k and ε = 1.0 and
effect of subgrid scale stresses () such as Smagorinsky model, 1.3, respectively.
dynamic Smagorinsky model, scale similarity model, one equa- The standard k − ε model has performed satisfactorily in many
tion subgrid scale kinetic energy model, etc. These models are flows, but the applicability of this model is limited due to uncer-
discussed in Mathpati and Joshi (2007). tainties involved in the modelling of turbulence production,
turbulent transport and the assumptions in modelling dissipation
Reynolds Averaged Navier–Stokes (RANS) Approach rate equation. Further, the modelling of transport equations for
k and ε pose the difficulties to account for streamline curvature,
DNS and LES require huge computational power and are not pos- rotational strains, and the other body-force effects.
sible for very large-scale industrial contactors at high Reynolds
number. In most of the cases, it is necessary to understand the
Zonal model
gross flow patterns. This can be achieved by solving the ensemble-
averaged Navier–Stokes equation. In case of flows with very strong inhomogenity, the standard tur-
Equations (1) and (2) on ensemble-averaging reduce to the bulence model parameters fail to predict the flow field. Hence the
following form: entire domain is divided into various zones and a set of turbulence
model parameters (chosen on the basis of prior experience so as
∂ ∂ui to improve upon the predictions) is assigned to each zone. Such
+ =0 (7) efforts in stirred tank are reported by Sahu et al. (1998). Thus
∂t ∂xi
effectively, it identifies the different regions of distinct physics of the near wall flows. Table 2 gives the source terms along with
flow, modifies the values of the turbulence parameters accordingly the modified eddy viscosity formulation. It contains more num-
and effectively prescribes the zonal eddy viscosity distribution. ber of turbulence model parameters (Table 3) and is applicable
The selection and distribution of turbulence parameters in stirred to the swirling flows wherein strain gradients are large where the
tank is largely arbitrary. Nere et al. (2001) carried out CFD simula- simple eddy viscosity relation of Boussinesq leads to non-realistic
tion of stirred tanks using zonal modelling for C . They have also values of the eddy viscosities.
developed a new constitute equation for eddy viscosity which gave
better agreement with experimental data for axial flow impellers. Multiscale models
Some investigators have used model parameters as a function of In contrast to the assumption in k − ε models as regards to the
geometrical and operating variables, rather than constant values. presence of equilibrium and hence the presence of a single time
For example in jet flows, model parameters are related with jet scale, these models assume the existence of two time scales Placek
half-width and centreline velocity (Launder et al., 1972; McGuirk et al. (1986). They have proposed a multiscale model for stirred
and Rodi, 1977; Rodi, 1993; Mathpati et al., 2009). tanks. It considers the existence of non-spectral equilibrium. The
energy spectrum is divided in three parts, namely production
RNG k − ε model region, a transfer region and a region in which dissipation takes
RNG k − ε model uses framework of two equations and these place. It assumes the existence of the spectral equilibrium only
equations are derived by application of renormalisation group between the transfer region and the region in which turbulence
methods to the original governing equations of momentum trans- is dissipated. Thus, it considers the presence of two time scales,
fer. This model gives the eddy viscosity relation based on the result namely one for the production range and the other for the dis-
of application of the renormalisation group theory where inertial sipation. It involves two equations: one for the transport of the
sub-range eddies are eliminated from the equations of motions turbulent kinetic energy of the large-scale vortices and the other
to yield averaged flow quantities at the integral scale of turbu- for transport of the turbulent kinetic energy of inertial sub-range
lence. This model is capable of modelling the distance between eddies in addition to the equation for the energy dissipation rate.
the production and dissipation of turbulent kinetic energy better Tables 2 and 3 give the source terms and the values of the con-
than the standard k − ε model. As opposed to the semi-empirical stants involved. Utility of this model needs to be extensively
models these are more fundamental and are capable of predicting proved for the prediction of flow pattern in stirred tanks.
Standard k − ε model (Launder and Spalding, 1974) C = 0.09, Cε1 = 1.44, Cε2 = 1.92, k = 1.0, ε = 1.3
Modified k − ε model (Ranade and Joshi, 1989) C = 0.125, Cε1 = 1.59, Cε2 = 1.62, k = 1.0, ε = 1.3
New k − ε model (Nere et al., 2001) C = (1/RUtrip )(k 2 /ε), Cε1 = 1.44, Cε2 = 1.92, k = 1.0, ε = 1.3
High Re RNG k − ε model C = 0.0845, Cε1 = 1.42, Cε2 = 1.68, k,RNG = 0.7194, ε = 1.3
Chen–Kim k − ε model C = 0.09, Cε1 = 1.15, Cε2 = 1.90, Cε3 = 0.25, k = 0.75, ε = 1.15
Optimised Chen–Kim k − ε model C = 0.09, Cε1 = 1.36, Cε2 = 1.90, Cε3 = 0.04, k = 0.75, ε = 1.15
RSM C = 0.09, Cε1 = 1.44, Cε2 = 1.92, Cs = 0.22, C1 = 1.8, C2 =
0.60, C1 = 0.5, C2 = 0.18, k = 1.0, ε = 1
LES = 0.42
Cs = 0.23, Smagorinsky constant (for homogeneous isotropic
turbulence in the inertial sub-range)
similar to that given by Pericleous and Patel (1987) (these source the system attains a satisfactory numerical convergence. Each of
terms can easily be calculated in terms of the lift and drag coef- the inner and outer simulations is carried out under steady state
ficients by using the velocity–force diagram for the blade section assumptions in its own frame of reference. As the two frames
under consideration) and the radial momentum source is derived are different, the information that is iteratively exchanged is cor-
by assuming that the interaction between the impeller and fluid rected for the relative motion and averaged over the azimuthal
results in forces perpendicular to the blade surfaces. The bound- direction. The main feature of this approach is the existence of
ary conditions on the tank wall and baffles are specified as the the overlapping region, common to the inner and outer zones,
solid surfaces. Further, the impeller blades are also modelled as which provides the iterative matching of the two solutions. The
walls. In this approach, the full tank needs to be modelled and extent of this region and the exact location of its boundaries are
not a quarter of the tank as is practiced in some other cases. largely arbitrary.
Harvey and Greaves (1982a, Rushton turbine D = 0.228 m, T = 0.456 m, flat enclosing 950 rpm Black box approach k − ε model
1982b) wall at the top, impeller blade
height = 0.022 m
Middleton et al. (1986) Rushton turbine T = 0.31 m, HL = 0.42, D = 0.15 m: N = 600, 100, 50 rpm Black box approach k − ε model
|
C/T = 0.33
Placek et al. (1986) Rushton turbine T = 0.294 m, D/T = 0.33, and 0.25 Re = 63 300 and 72 800 Black box approach kp − kT − ε multiple-scale
model
Pericleous and Patel (1987) Straight blade turbine, Source-sink approach
Rushton turbine
Ju et al. (1990) Rushton turbine Three tank geometries: T = 0.292 m, 26 × 30 × 25 and Black box approach Non-isotropic k − ε model
viscosity = 5 centiStokes, D = 76, N = 250: 23 × 30 × 25
I, D = 102, N = 150: II, D = 102, N = 200:
III
Ranade and Joshi (1990b) Rushton turbine T = 0.3 m, C/T = 1/2, D/T = 1/3, H = T Black box approach k − ε model
Kresta and Wood (1991) Rushton turbine T = 0.456 m, D/T = 0.5 and 0.33, H = T, 30 × 30 × 26 Black box approach k − ε model
C/T = 0.5, N = 950 rpm, FLUENT
Luo et al. (1993) Rushton turbine T = 0.294 m, C/T = 0.33, D/T = 0.33, 0.5, 28 × 45 × 120 (r, z, ,) MRF k − ε model
0.2
Perng and Murthy (1993) Rushton turbine 21 × 19 × 24 MRF
Dong et al. (1994) Eight straight bladed Unbaffled vessel, 100 rpm, T = 0.1 m, 39 × 11 × 23 MRF k − ε model
turbine D/T = 0.33, C/T = 0.33, H = T
Luo et al. (1994) Rushton turbine T = 0.292 m, C/T = 0.33, H = T, D/T = 0.33 Source-sink, MRF
Brucato et al. (1994) Rushton turbine 60 × 72 × 26 Inner outer approach k − ε model
Tabor et al. (1996) Rushton turbine T = 0.27 m, D = 0.093 m, C/T = 0.33 50 000, 120 000 Source-sink MRF and k − ε model
SM
|
Lee et al. (1996) Rushton Turbine T = 0.1 m, H = T, D/T = 0.33 46 016, 102 296, 138 632,
N = 2165 rpm
Re = 40 000 SM
Ciofalo et al. (1996) Rushton turbine, T = 0.6 m, D = 0.3 m, C = 0.3 m, and 26 × 46 × 16 and MRF k − ε model, second-order
straight blade turbine T = 0.19 m 20 × 30 × 16 differential stress model
Rigby et al. (1997) Rushton and flat bladed T = 0.294 m, w = 0.2D, D/T = 1/3, C/T = 1/3 SM
turbine
Ranade (1997) Rushton turbine 28 × 28 × 78, Source-sink approach k − ε model
Re = 32 000—5135 Snapshot approach
regime
Jenne and Reuss (1997) Rushton turbine T = 0.444 m, D/T = 0.33 = C/T N = 333 rpm Black box approach k − ε models, Chen–Kim
modified k − ε and RNG
k − ε models
Venneker and van den Akker Rushton turbine T = 0.286 m, D = 0.0953 30 × 28 × 50 Re = 5500 Black box approach k − ε model, Power law
(1997) (n = 0.77) model
Jaworski et al. (1997) Rushton turbine T = 0.15 m, D/T = 0.33, C/T = 0.33, H = T 20 × 28 × 40, Re = 17 000 SM k − ε and RNG k − ε model
Ng et al. (1998) Rushton turbine T = 0.1 m, H = T, D/T = 0.33, C/T = 0.33 46 016, 239 468 SM k − ε model
N = 2165 rpm
|
67 000 IO, SM
(Continued )
35
|
|
36
|
Table 4. (Continued )
Jenne and Reuss (1999) Rushton turbine T = 0.444 m, H = T, C/T = 0.5, D/T = 0.33 Black box approach k − ε model, Chen–Kim
k − ε RNG-k − ε model
and optimised
Chen–Kim k − ε model
Derksen and van den Akker Rushton turbine D/T = 0.33, C/T = 0.33 6 00 000, Re = 29 000 Adaptive force-field
(1999) technique
Bartels et al. (2000) Rushton turbine T = H = 0.152 m, D/T = 0.33, C/T = 0.33, MRF DNS and k − ε model
Re = 7275
Ranade et al. (2001) Rushton turbine 80 × 83 × 95 Snap-shot approach k − ε model
Lane et al. (2000) Rushton turbine Geometry: same as that of Luo et al. (1994) SM and MRF
|
N = 300 rpm
Montante et al. (2001) Rushton Turbine T = 0.29, H = T, D = T/3, C/T = 0:12; 0:15; 80 136, RPM = 250, SM and inner–outer Std k − ε model, RNG k − ε
0:2; 0:25; and 0.33. CFX software Re = 40 000 model, RSM
Jones et al. (2001) Paddle impeller H = T = 0.1 m, D = T/4, Case 1: C = T/2, 94 500–338 000 for 1/8th MRF k − ε model, modified k − ε
Case 2: C = T/3, FLUENT region, RPM = 100, model, RNG k − ε
Re = 3273 model, k − ω model, and
modified k − ω model
Ranade et al. (2002) Rushton turbine T = H = 0.15 m, C = T/3, D = T/3, FLUENT 269 667 for PBT (for 1/4th Computational k − ε model
region), 630 800 for snapshot
Rushton (for half region)
Kukuková et al. (2005) Rushton turbine H = T = 0.29 m, D = T/3, C = T/3, FLUENT 586 680 for Rushton MRF k − ε model
Deglon and Meyer (2006) Rushton turbine H = T = 0.15 m, D = T/3, C = T/3, FLUENT 33 000–1 900 000, MRF k − ε model
Re = 40 000
Guha et al. (2006) Rushton turbine H = T = 0.2 m, D = T/3, C = T/3, FLUENT 120–720 compartments, MRF k − ε model
150–350 RPM
Javed et al. (2006) Rushton turbine T = 0.3 and 0.15, H = T, D = T/3, C = T/3, 112 480, 138 348, SM k − ε model
FLUENT software Re = 24 000, 48 000
Ochieng et al. (2008) Rushton turbine T = 0.38, D = 0.33T, H = T, C = 0.33 T and 216 000, 436 000, and MRF and SM k − ε model, LES
0.15T 700 000, RPM: 300
Alopaeus et al. (2009) Total 17 cases. Majority T/H = 1, C = T/3, D = T/3, volume of 36 000–800 000, power MRF k − ε model, SST k − ω
with Rushton liquid = 14 dm3 , 194 dm3 and input = 1.49 W/kg to model
|
Turbine, last two 20 000 dm3 , CFX software 15.93 W/kg
with Combijet and
Phasejet
Comparison at x2 /R = 0.586
Figure 5A shows the comparison of radial velocity using standard
k − ε model and various impeller–baffle interaction models. It can
be seen that SNAP model (line 4, 28%) gave the best fit amongst
all the models. IBC (line 1, 44%), MRF (line 2, 52%), and SM (line
3, 88%) underpredicted the radial velocity. Even the axial velocity
predictions (Figure 5B) were better with SNAP model (51%), fol-
lowed by IBC (73%), MRF (79%), and SM (95%) models. None of
the models could predict tangential velocity (Figure 5C). The pre-
dictions of radial velocity with RSM as turbulence model and IBC
Figure 4. Schematic of stirred tank reactor with locations at which and MRF as impeller–baffle interaction model are given in (Figure
comparison between experiment and CFD data has been carried out.
6A). MRF (52%) gave almost similar results with RSM and stan-
dard k − ε model, while IBC (70%) gave more deviation with RSM.
illustrated with the experimental data obtained with the help of Even in case of axial velocity (Figure 6B), though both the models
in-house LDA. The values of error are reported in brackets in the under predicted, MRF (76%) with RSM gave better results than
analysis. Table 5 gives the quantitative error involved with all the standard k − ε model. IBC (81%) with RSM gave poor results as
turbulence models for mean and turbulent quantities. The error compared with standard k − ε model. Figure 6C shows the tangen-
is estimated as follows: tial velocity predictions of IBC and MRF with IBC. Both the models
(>300%) showed only qualitative agreement with experimental
Expt Value − Predicted Value results with RSM.
Error = × 100 (11)
Expt Value Figure 7A shows the prediction of turbulent kinetic energy.
Though IBC gave a peak at x1 /R = 0.1, which was not shown by
experiments, it gave satisfactory match (45%) in the rest of the
Simulation Details region. MRF (92%), SM (76%), and SNAP (65%) underpredicted.
Turbulent kinetic energy with RSM (Figure 7B) gave satisfactory
The geometrical details of the system investigated are already
results with IBC (45%), while underpredictions were observed
shown in Figure 4. The impeller speed was kept at 4.5 rps
with MRF (92%).
(Re = 45 000) to ensure the turbulent flow regime. The impeller
was placed at H/3 from the vessel bottom. All the simulations
have been carried out by using an in-house and the commercial Comparison at x2 /R = 0.24
code FLUENT. Pressure–velocity coupling has been taken care of Figure 8A shows the comparison of radial velocity using different
by using SIMPLE scheme. Grid independence studies have been impeller–baffle interaction models using standard k − ε turbu-
carried out using three different grid sizes, namely 45, 68, and lence model. It can be seen that IBC model (line 1, 835%) shows
85 K. Finally 85 K grid size has been chosen for all the simula- better agreement compared to other models (MRF: line 2, 246%,
tions. Non-uniform grids are used to maintain higher grid density SM: line 3, 1005%, SNAP: line 4). However, maximum value of
in high shear regions. velocity with experiment was observed close to x1 /R = 0.66 and
In some cases, visual observation of plots showed good agree- with IBC at x1 /R = 0.42. SM and SNAP approach do not predict
ment, whereas RMS error was high. This was because agreement negative radial velocity, whereas MRF approach shows very poor
was good over major portion; only at some locations deviations agreement with experimental data. Figure 8B shows comparison
were high making the RMS error values large. Other possible rea- of axial velocity. It can be seen that SNAP shows very good agree-
sons for relatively large deviations are (i) number of grid points ment with experimental data over the entire range. IBC predictions
for all the simulations have been decided after grid sensitivity (103%) were in good agreement up to x1 /R = 0.4, beyond which
study. In this grid sensitivity study, comparison of mean veloc- it was insignificant under prediction. SM (51%) and MRF (118%)
ity components with experimental data using IBC approach and showed under prediction over entire range and agreement was
standard k − ε model was carried out. Hence possibility of error very poor. Figure 8C shows comparison of tangential velocity. It
at this resolution for different impeller models and turbulence can be seen that no model is able to predict tangential velocity
models cannot be ruled out. (ii) The turbulence in stirred vessels neither in qualitative nor quantitative sense. The source of error
is strongly anisotropic and not homogeneous. The standard k − ε in prediction can be either from inappropriate impeller model or
model and its modifications cannot account for such anisotropy turbulence model. To check the sensitivity of turbulence model on
due to inherent assumptions involved in these models. At best, overall prediction, similar comparison of impeller models (MRF
one can use different combinations of model parameters in dif- and IBC) is presented using RSM. Figure 9A shows comparison of
Axial location Max error RMS error Max error RMS error Max error RMS error Max error RMS error
be seen that predictions from all the impeller models for turbulent
kinetic energy were in poor agreement by both the turbulence
models. IBC model (with k − ε model: 398% and with RSM:
542%) overpredicted, whereas other models severely under pre-
dicted (factor of 10).
Comparison at x2 /R = −0.24
The comparison of mean and turbulence quantities in below
impeller plane has been presented in Figures 11–13. Figure 11A–C
shows comparison of mean flow using standard k − ε model. Simi-
lar to x2 /R = 0.24, IBC (2916%) shows only qualitative agreement
with the experimental data below impeller for radial velocity Figure 8. Prediction of mean velocity components using standard k − ε
(Figure 11A). SNAP and MRF have totally different trend than model at x2 /R = 0.24. Impeller: disc turbine. (A) Radial Velocity (B) axial
experimental data, whereas SM has shown significant under pre- velocity (C) tangential velocity. Experimental, line 1: IBC, line 2: MRF,
diction (by a factor of 4). The reason behind good agreement using line 3: SM, line 4: SNAP.
IBC is it uses actual velocity field information (experimental data)
in the impeller zone, whereas in other methods, actual experi- comparison of impeller models (MRF and IBC) using RSM. From
mental data are not used for the simulation. Figure 11B shows Figure 12A, it can be seen that, a change in the turbulence model
the comparison of axial velocity. All the models show only qual- did not improve the results for MRF model, but the performance
itative match with the experimental data. SNAP approach was of IBC model with RSM (3780%) was deteriorated as compared
found to be in best agreement, followed by IBC (238%). MRF to standard k − ε model. Figure 11B shows that the effect of tur-
results were better than SM, but both the models under predict the bulence model on axial velocity (233%) is not significant and
experimental data. Figure 11C shows the comparison of tangential profiles are almost similar as in Figure 11B. The tangential velocity
velocity. IBC showed good agreement with the experimental data, (Figure 12C) predictions using IBC were in very good agree-
and other methods showed poor agreement. Figure 12 shows the ment with experimental data, whereas MRF results were in under
Comparison at x2 /R = 0.24
IBC with all the turbulence models though predicted radial veloc- Figure 14. Prediction of mean velocity components using standard k − ε
ity (Figure 17B1) trend qualitatively, the maximum was found to model at x2 /R = −0.933. Impeller: disc turbine. (A) Radial velocity (B)
axial velocity (C) tangential velocity. Experimental, line 1: IBC, line 2:
be shifted away from the vessel wall (standard k − ε model: 835%, MRF, line 3: SM, line 4: SNAP.
RSM: 279%, RNG k − ε model: 765%). RSM overpredicted radial
velocity in the central region while satisfactory agreement can be
seen result in rest of the region. While other turbulence models see that the sensitivity of the predictions to the turbulence models
yielded relatively closer agreement of radial velocity throughout. is limited to the region with x1 /R < 0.4 where RNG k − ε and STD
Axial velocity (Figure 17B2) predictions of all the simulations k − ε models resulted into larger underprediction of tangential
were practically similar (i.e., insensitive to turbulence models). velocity. All the turbulence models largely overpredicted turbulent
All the turbulence models closely predicted the axial velocity x1 /R kinetic energy (Figure 17B4) with the highest deviation as a result
up to 0.4 while underpredictions were observed in rest of the of RSM (542%) followed by the STD (398%) and RNG (159%)
radial region (0.4 < x1 /R < 0.8). All the models resulted into poor k − ε model in most of the radial region (x1 /R > 0.4). Good agree-
predictions (120–398%) of tangential velocity (Figure 17B3). RSM ment was observed for x1 /R < 0.4 with RNG k − ε model while
gave relatively closer agreement for x1 /R up to 0.4. Thus one can STD k − ε model yielded closer agreement for x1 /R up to 0.15.
captured by all the turbulence models. IBC with STD k − ε (84%) throughout the radial region. The deviation was found to be as
and RNG k − ε (70%) resulted into relatively closer comparison much as 100%. All the three turbulence models resulted into sim-
as compared to that given by RSM (114%). With each of the simu- ilar axial velocity predictions. Its predictions were opposite for
lations severe underprediction was the result. Interestingly all the x1 /R < 0.45 and it was overpredicted for rest of the region. Tan-
models were able to predict the upflow near the vessel wall. None gential velocity was severely underpredicted by all the turbulence
of the turbulence models was able to satisfactorily predict the models. RSM gave relatively closer comparison. Further STD and
tangential velocity component (Figure 17D3) with IBC. Though RNG k − ε models resulted into quantitatively similar predictions.
the standard k − ε (563%) and RNG k − ε (661%) model resulted IBC using all the turbulence models yielded satisfactory compar-
into relatively closer comparison near the wall, they failed to cap- ison. Predictions of k as a result of all the simulations were more
ture the qualitative trend of the tangential velocity in the rest of or less similar.
the radial region. Though RSM (380%) was able to qualitatively
follow the tangential velocity trend, severe underprediction (neg- Discussion for the simulations with the variants of
ligible swirling) was the result throughout the radial region. IBC the k − ε model
resulted into overall closer comparison for k (Figure 17D4) with
Figure 18 gives the comparison between the experimental flow
all the turbulence models employed. RSM (58%) overpredicted k
characteristics produced by DT and the predictions of the various
in the central region while standard (49%) and RNG (68%) k − ε
simulations with the variants of k − ε models (MOD k − ε, ZON
models underpredicted throughout the radial region. RNG k − ε
k − ε, and OPTCK k − ε models) using IBC approach.
model yielded largest underpredictions.
Comparison at x2 /R = 0.586
Comparison at x2 /R = 0.0 (impeller centre plane) All the variants of k − ε models (Figure 18A1) could capture the
Simulations with IBC using all the three turbulence models qualitative behaviour of radial velocity throughout the region.
resulted into consistent overprediction of mean radial velocity All the simulations underpredicted (∼50%) u throughout the
radial region. Again u predictions were found to be insensi- for 0.33 < x1 /R < 0.5 by all the k − ε variants. All the k − ε mod-
tive (with minor difference in magnitude) as a result of all the els resulted into similar predictions for axial velocity (Figure
turbulence models. Axial velocity (Figure 18A2) trend was pre- 18B2). MOD k − ε gave qualitatively opposite tangential velocity
dicted by all the k − ε variants. Underpredictions (57–72%) were for x1 /R <0.2 (Figure 18B3). Thus overall poor comparison was
given by all the models in the region with 0.2 < x1 /R < 0.65 and sought by all the k − ε models. Even the trend was not reproduced
x1 /R > 0.8. Relatively closer comparison by ZON k − ε model by any of the k − ε variants. It can be seen from (Figure 18B4)
was observed for x1 /R < 0.2. All the k − ε variants were suc- that the k was overpredicted more or less by all the models. Only
cessful in predicting the radial reversal and all the k − ε models MOD k − ε model was able to closely predict the experimentally
severely overpredicted the tangential velocity as much as 500% observed k for x1 /R > 0.26 while it resulted in severe underpre-
(Figure 18A3) for 0.13 < x1 /R < 0.3. MOD and OPT-CK k − ε mod- dictions in rest of the region. ZON k − ε model overpredicted k
els resulted in qualitatively opposite comparison for x1 /R < 0.4 (274%) throughout the radial coordinate. Though the OPTCK
and 0.13 < x1 /R < 0.4, respectively. All the models yielded similar k − ε model gave the closer comparison for 0.3 < x1 /R < 0.8, it
comparison for x1 /R > 0.5 with minor differences. The turbu- overpredicted the k (as much as 200%) in rest of the radial region.
lent kinetic energy (Figure 18A4) was underpredicted by all the
k − ε models throughout the radial co-ordinate. The deviation was
between 30% and 88%. Comparison at x2 /R = −0.24
It can be observed from Figure 18C1 that the experimental radial
velocity trend was closely followed by all the k − ε models for
Comparison at x2 /R = 0.24 x1 /R > 0.2. All the k − ε variants employed, resulted into qual-
Almost all the models underpredicted radial velocity nominally itatively opposite radial velocities near the vessel centre while
for x1 /R > 0.6, the deviation being maximum (as much as 25%) severe overpredictions for 0.2< x1 /R < 0.4. Minor deviations were
for x1 /R > 0.35 (Figure 18B1). Qualitatively opposite predictions observed for x1 /R > 0.5. Maximum deviation was due to ZON k − ε
were observed for x1 /R < 0.33 while overprediction was the result for x1 /R < 0.4. Closer comparison was obtained with the OPTCK
k − ε for x1 /R between 0.45 and 0.65 and other k − ε variants for velocity, clearly depicts the success of all the k − ε models in pre-
x1 /R > 0.65. All the k − ε models yielded good axial velocity pre- dicting the upflow near the vessel wall. ZON k − ε resulted into
dictions throughout the radial coordinate. It was found to be more the closer axial velocity comparison (40%) followed by OPT-CK
or less insensitive to the type of k − ε model (Figure 18C2). Tan- k − ε model (79%). While deviations were higher (296%) with
gential velocity predictions were largely overpredicted for x1 /R up the MOD k − ε model. Qualitatively opposite tangential velocities
to 0.3 while underpredictions (as much as 200%) was the result (Figure 18D3) were obtained as a result of all the simulations
for x1 /R > 0.3 by all the k − ε models (Figure 18C3). Predictions using all the k − ε models (except ZON k − ε model). Again MOD
of the w were more or less similar as a result of all the simu- k − ε model gave highest deviation for x1 /R < 0.4. While more
lations with k − ε models (except MOD k − ε in the region with or less similar tangential velocity predictions were the result for
x1 /R < 0.3). All the k − ε variants resulted in quantitatively similar x1 /R > 0.5 by all the k − ε models. All the variants of k − ε mod-
k predictions for x1 /R < 0.3 (Figure 18C4). Excellent comparison els severely underpredicted k (Figure 18D4) throughout the radial
was observed for x1 /R > 0.4 by ZON and OPTCK k − ε models distance. Comparison was relatively closer with ZON k − ε model
while overpredictions (100%) and underpredictions (as much as (62%). The simulations using MOD k − ε model yielded k com-
150%) were seen by NEW and MOD k − ε models, respectively. parison with maximum underpredictions throughout the radial
coordinate. Thus no k − ε variant was able to predict k near the
Comparison at x2 /R = −0.933 liquid surface.
All the variants of k − ε model predicted the radial velocity for x1 /R
up to 0.45 with minor deviations while qualitatively and quanti- Comparison at x2 /R = 0.0 (impeller centre plane)
tatively poor comparison (100–774%) was sought in rest of the All the variants of k − ε model resulted into more or less similar
radial region. MOD k − ε model though gave closer comparison predictions of radial velocity. It was underpredicted by as much as
for x1 /R < 0.4 it resulted into the largest deviation for x1 /R > 0.4 10% near the tip of impeller while the differences between the pre-
(Figure 18D1). Figure 18D2 which shows the comparison for axial dicted and the experimental data were found to decrease. Though
STD k − ε model with SM failed to indicate axial velocity (Figure seen from (Figure 20C4) that all the turbulence models result
20B2) trend and hence the downflow in near the wall (118%) into overall poor prediction of the turbulent kinetic energy when
while RNG k − ε model resulted into the wider downflow region used with SM. A peak of k could be seen with RNG k − ε model,
(94%). As can be seen from Figure 20B3, SM again failed to predict which is similar to the peak, observed in tangential velocity as
the tangential velocity with all the turbulence models employed. a result of a SM simulation with RNG k − ε model. In rest of the
It resulted almost into insignificant values of tangential veloci- region severe underpredictions (83–135%) was observed with all
ties throughout the radial region. The k-predictions (Figure 20B4) the turbulence models.
were again found to be in poor agreement (severe underpredic-
tion, 88%) with the experimental data. It was closely predicted Comparison at x2 /R = −0.933
by only RNG k − ε in a narrow region nearby to x1 /R = 0.5. SM was not able to capture even the trend of the experimental
radial velocity (Figure 20D1) using any of the turbulence models.
Comparison at x2 /R = −0.24 Again SM with all the turbulence models failed to predict the axial
SM with both the turbulence models overpredicted the radial velocity component (Figure 20D2). It showed very negligible val-
velocity (Figure 20C1) for x1 /R up to 0.55 while qualitatively ues of the axial velocity. SM with all the turbulence models failed
opposite trend was observed in rest of the region. It with all to predict tangential velocity (Figure 20D3) both qualitatively and
the turbulence models employed resulted in good comparison quantitatively. It showed almost absence of the swirling velocity.
(144–170%) of the trend in axial velocity (Figure 20C2). Large It can be seen from Figure 20D4, that the SM failed again to predict
underpredictions (1722%) were observed from the measured the turbulent kinetic energy.
tangential velocities (Figure 20C3). RNG k − ε model was able
to predict the tangential velocity closely for x1 /R < 0.4 while Comparison at x2 /R = 0.0 (impeller centre plane)
STD k − ε model resulted into severe underpredictions (3322%) SM approach resulted in satisfactory comparison of the radial
throughout the radial region (except for x1 /R < 0.15). It can be velocity using both the STD and RNG k − ε models for x1 /R > 0.5
(1) Bulk flow with respect to impeller design, internals, and ves-
while underpredictions of as much as 80% could be observed for sel configuration
x1 /R < 0.5. RNG k − ε model yielded relatively closer comparison. (a) Axial Flow
As regards to the axial velocity, STD k − ε model gave closer com- (b) Radial Flow
parison for x1 /R > 0.5 while it was underpredicted severely for (c) Tangential Flow
x1 /R < 0.45. The tangential velocity predictions were more or less (2) Gross flow parameters
similar as a result of each of the simulations. It was marginally (a) Primary flow number
underpredicted throughout. The turbulent kinetic energy was (b) Secondary flow number
severely underpredicted by STD k − ε model with the maximum (c) Power number
departure of around 400%. Though the RNG k − ε model was able (3) Quantification of bulk velocities in the near wall regions
to predict the trend of the turbulent kinetic energy deviations of (mainly for heat transfer)
as much as 80% were observed. (4) Local distribution of turbulent kinetic energy with respect to
geometric and operating variables
LARGE EDDY SIMULATION OF RADIAL FLOW (5) Quantification of local length scale distribution
(6) Trailing vortices
IMPELLERS (a) Shape with respect to impeller blade design
To meet the various process objectives, it is essential to main- (b) Quantification of associated energy
tain the desired local hydrodynamics by manipulating both the (c) Extent up to which their effect is significant
geometric variables and operating conditions. For instance, in (d) Symmetry/asymmetry with respect to disc (in case of
crystallisation the size distribution, shape, crystalline structure Rushton turbine) when impeller is located at (1/3)T
and purity of the product are affected by the dynamics of small and (1/2)T.
eddies which influence the mass transfer and degree of super- (7) Rate of turbulent kinetic energy dissipation
saturation in the immediate vicinity of precipitating and growing (a) Local distribution (crystalliser, drop and bubble size
crystals. Therefore, crystalliser design should ensure that homoge- estimation, instantaneous reaction)
neous distribution of fine length scales and local shear rate should (8) Local and global shear rate distribution with respect to geo-
be at the desired level. Similarly in the case of hollow gas inducing metric and operating conditions
impellers, which are highly recommended to handle hazardous, (9) Energy spectrum
expensive gases and when per pass conversion is low. Sufficient (a) Identification of dominant frequencies with respect to
low-pressure zone must be created for the induction of gas and geometric and operating parameters
at the same time with good dispersion characteristics to enhance (b) Quantification of instabilities (precessing, jet and cir-
mass transfer rate. Different design alternatives can be made possi- culation instabilities)
1 275 567
49 0000
490 000
761 760
615 000
480 000
750 000
763 000
6 × 106
tational efficiency and the ease of implementation on parallel
2403
1203
2403
processors. The main objectives were to make the lattice-
Boltzmann scheme based solver more general to simulate flows
in real complex geometries and to understand the detailed local
Adaptive force field technique
Inner–outer algorithm
Sliding mesh
Sliding mesh
Sliding mesh
fine mesh was found to give better predictions than the coarse
mesh. However, there were differences in the quantitative predic-
tions of the mean radial and mean axial velocities especially in the
near impeller region. The discrepancies were due to insufficient
statistical sampling time which was limited by the computing
Reynolds number
14 000, 82 000,
14 000, 82 000,
time with the fine grid and adoptive force field technique used
for the impeller modelling. Further, the quantitative comparison
350 000
350 000
has been restricted to a few flow variables (axial and radial veloc-
107 000
60 000
29 000
40 000
40 000
30 000
45 000
46 600
29 000
56 250
24 000
22 000
45 000
ity) and also at limited locations (one radial and one axial line)
7300
9720
Two year later Revstedt et al. (1998) investigated both the mean
T = 0.10, D/T = 0.33, C = T/3, H = T
H = T = 0.3, C = D = T/3
SGS model
velocity at only one axial location was given (impeller centre line),
which was very good at all the locations except 0.5 < x1 /D < 0.6.
They have also obtained a fairly good prediction for the turbulent
kinetic energy in the impeller centre plane and a good agreement
between the experimental and predicted value of pumping num-
Derksen and van den Akker (1999)
fluid entrainment due to the trailing vortex) pair was well simu-
lated. Prediction of radial and tangential velocity with the angle
Murthy and Joshi (2008)
from LES supported the blade frequency and the existence of Kol-
Eggels (1996)
Figure 27. Turbulence transport term in TKE equation in stirred tank with disc turbine.
reviewed by Kresta and Wood (1993); and Sahu et al. (1999). One where u is a characteristic velocity scale, L is a characteristic
of the simplest ways of estimating the turbulent energy dissipa- length scale, and A is a constant of proportionality. Different meth-
tion rates is to assume that the eddy transfers all its energy within ods reported in the previous literature use different methods for
one lifetime. This results in the following estimate of ε: estimation of u and L.
In the published literature, the characteristic velocity scale has
u 3 been taken as the
√ RMS velocity or the square root of the turbulent
ε=A (14) kinetic energy ( k). The characteristic length scale has also been
L
defined in a number of ways: (i) as a simplest approximation, the methods yield different values of L and hence different estimates
characteristic length scale has been taken to be proportional to of ε.
the impeller diameter. The proportionality constant has been cho- The energy dissipation rate can also be estimated from Equa-
sen arbitrarily to be between 0.1 and 1 (Kresta and Wood, 1993). tions (11) and (12) by assuming Taylor’s frozen turbulence
Moreover, these authors have not considered the possible varia- hypothesis (Kresta and Wood, 1993). Further, if the turbulence
tion of the characteristic length scale throughout the vessel. The is isotropic, then, ε can be estimated using the following relation-
assumption of constant L is totally unrealistic; (ii) the length scale ship:
has been estimated using the auto-correlation function (Wu and
Patterson, 1989). The length scale can be estimated by first calcu- uRMS
2
lating the integral time scale () using auto-correlation function ε = 15 (17)
2
as:
∞ Here, the
is the Taylor microscale, which can be estimated
r= r()d (15) from the curvature of the auto-correlation function.
The energy associated with the fluctuating velocity component
0
represents the energy associated with the turbulent motion. The
Here the integration may be carried out for a long time or up Fourier transform of the velocity fluctuations give the energy con-
to the first zero crossing. Commonly the integration is carried out tributions of the fluctuations at different frequencies. This can
up to the first zero crossing. then be represented as a plot of energy associated with a par-
Anandha Rao and Brodkey (1972) have reported that the oscil- ticular frequency versus the frequency. This plot is commonly
lations in the auto-correlation function were due to the periodicity called the power spectrum or the energy spectrum (Mujumdar et
of the flow especially near the impeller blades. Thus, if the inte- al., 1970). The spectra are commonly represented in terms of the
gration is carried out for a long time, the negative values of the wave number. The wave number can be related to the frequency
auto-correlation function may actually reduce the value of the as:
integral time scale.
The Taylor microscale can be obtained from the curvature of 2f
k1 = (18)
the auto-correlation function as: V
d2 r()
−2
=0= 2 (16) If the turbulence is assumed to be isotropic, the energy dis-
d 2
t sipation rate can be estimated from the one-dimensional energy
spectrum using the following relationship:
The above time scales can be converted to the length scale by
multiplication with the mean velocity or the RMS velocity. Thus
the two ways of obtaining the time scale and the two ways of ∞
getting the length scale from the time scale result into four possible ε = 15 k12 E(k1 )dk1 (19)
ways of obtaining an estimate of L. It is quite likely that these 0
2 H R
Utip
3
P= εx1 dx1 dx2 dx3 (23)
R
0 0 0
Impeller Impeller model Turbulence model Grid size inimpeller swept region Grid size Np ( based) Np (ε based)
Previous work of LDV. The Reynolds stresses were also estimated from the
Drbohlav et al. (1978) have investigated the turbulence charac- relationship:
teristics produced by a Rushton turbine in a 1.0 m diameter tank.
2 ∂ui ∂uj
The Reynolds normal stresses and shear stresses were measured u i u j = kıij + T + (25)
with the help of a pitot tube. The measured values of the Reynolds 3 ∂xj ∂xi
stresses and the mean velocity gradients were used for the calcu-
lation of the eddy viscosity. It was observed that the eddy viscosity The values of eddy diffusivity were estimated from the relation:
values reduce with an increase in the radial distance away from
the impeller tip. No attempts were made to obtain the distribution k2
T = C (26)
of the eddy viscosity in the rest of the vessel. ε
Mahoust et al. (1988) have measured the turbulence charac-
teristics of stirred vessels equipped with a Rushton turbine. The The values of k were measured and that of ε were taken as 15
vessel diameter was 0.2 m and the measurements were made with times the average energy dissipation rate. It was observed that
the help of LDV. They have made simultaneous measurements of the predicted Reynolds stresses showed the same trend as that
the radial and tangential velocity components, which enabled the observed experimentally. However, their magnitude was as much
calculation of the Reynolds shear stresses. It was observed that, as 50% higher or lower in the impeller discharge stream.
in the impeller discharge stream and close to the impeller tip, the Derksen and van den Akker (1999) performed three-
Reynolds shear stress u i u j was about three times smaller than dimensional, angle-resolved LDV measurements of the turbulent
the normal stresses. Whereas, near the wall, it was found to be flow field (Re = 2.9 × 104 ) in the vicinity of a Rushton turbine in
about 10 times smaller than the normal stresses. a baffled mixing tank. They have presented results on the average
Mahoust et al. (1989) have established energy balance by flow field, as well as on the complete set of Reynolds stresses. The
computing the power transmitted from the torque calculations. It anisotropy of the turbulence has been characterised by the invari-
was observed that the power number based on torque was about ants of the anisotropy tensor. The trailing vortex structure, which
4.5, whereas, the experimentally observed power number was is characteristic for the flow induced by a Rushton turbine, is
close to 5. demonstrated to be associated with strong, anisotropic turbulent
Kemoun et al. (1994) have investigated the turbulent stresses activity.
produced by standard Rushton turbines. The experiments were Galletti et al. (2004) used 3-D LDV to measure Reynolds stresses
carried out in a 0.2 m diameter stirred vessel and the measure- and triple products in stirred tank with DT. They presented the pro-
ments were made with the help of a LDV. Their measurements files of ensemble-averaged normal and shear stresses throughout
have shown that the three components of the turbulent stresses the entire vessel. They observed that near the tank bottom and
as well as the shear stresses are not similar in shape and are not vessel axis, the three normal stresses differed significantly. In this
equal in magnitude. This indicates that the turbulence is highly region, radial-tangential shear stress was highest and the other
anisotropic. Also a comparison of the turbulent shear stresses and two being negligible.
the normal stresses shows that the former are smaller by as much
as five times than the latter. Comparison of Reynolds stresses
Derksen et al. (1997) have measured the Reynolds stresses in Extensive comparison of predicted Reynolds stresses by all the
the stirred vessel equipped with a Rushton turbine. The mea- simulations with the experimental data of Kumar (1997) was
surements were made in a 0.3 m diameter tank with the help carried out. Figure 30 shows the comparison for the predicted
2 = (Ci −C0 )2 number, blade width, blade height, disc thickness (in case
n−1 of DT), have been found to have a profound impact on
i=0
the overall flow pattern, power number of impeller and the
The value of the variance is representative of the extent mixing time. It has been reported that an increase in the
of the overall mixing and should be decided as per the blade width can be more energy efficient as compared to
desired extent of mixing. This criterion has scarcely been the increase in impeller diameter. Further, the number and
used in the past investigations. This may be due to the dimensions of the impeller blades (thickness, in particular)
fact that it requires a number of measurements of the can be of great help in the optimisation of the mixing pro-
concentration at number of points in order to eliminate cesses. Further, in view of this a great deal of care should be
the effect on variance. With the advent/development of taken while comparing mixing performance of same and/or