Sunteți pe pagina 1din 11

02638762/06/$30.00+0.

00
# 2006 Institution of Chemical Engineers
www.icheme.org/cherd Trans IChemE, Part A, June 2006
doi: 10.1205/cherd.05111 Chemical Engineering Research and Design, 84(A6): 495 505

CFD SIMULATIONS OF HYDROCYCLONES WITH


AN AIR CORE
Comparison Between Large Eddy Simulations and a Second
Moment Closure
M. BRENNAN
The Julius Kruttschnitt Mineral Research Centre, The University of Queensland, Queensland, Australia

C
FD simulations of the 75 mm hydrocyclone of Hsieh (1988) have been conducted
using FluentTM. The simulations used 3-dimensional body fitted grids. The simu-
lations were two phase simulations where the air core was resolved using the mixture
(Manninen et al., 1996) and VOF (Hirt and Nichols, 1981) models. Velocity predictions
from large eddy simulations (LES), using the Smagorinsky Lilly sub grid scale model
(Smagorinsky, 1963; Lilly, 1966) and RANS simulations using the differential Reynolds
stress turbulence model (Launder et al., 1975) were compared with Hsiehs experimental vel-
ocity data. The LES simulations gave very good agreement with Hsiehs data but required
very fine grids to predict the velocities correctly in the bottom of the apex. The DRSM/
RANS simulations under predicted tangential velocities, and there was little difference
between the velocity predictions using the linear (Launder, 1989) and quadratic (Speziale
et al., 1991) pressure strain models. Velocity predictions using the DRSM turbulence
model and the linear pressure strain model could be improved by adjusting the pressure
strain model constants.

Keywords: computational fluid mechanics; hydrocyclone; turbulence; second-moment


closure; large eddy simulation.

INTRODUCTION computational challenge for several reasons. The flow is


turbulent. Hydro-cyclones operate with a strong swirl
Cyclone separators are used extensively in the chemical together with a flow reversal, and a flow separation, near
and mineral processing industries to remove or classify par- the underflow. This makes the flow strained and introduces
ticles in particle laden fluid flows. Cyclones rely on the cen- anisotropy into the turbulence. Further the short residence
trifugal forces that develop under the swirling flow inside time implies that the turbulence is not at equilibrium.
the cyclone body to effect the separation and can classify Hydro-cyclones are often operated with the outlets open
on density or particle size. Cyclone separators are essen- to the atmosphere and thus develop an air core along the
tially passive devices, with a short residence time, which axis because the swirling flow generates an axial region
makes them easy to run. However, the fact that cyclones of negative gauge pressure which draws air in. The free sur-
treat particle-laden flows means that wear and its minimiz- face between the air and the water phase introduces further
ation is a major operational problem. Further the classifi- turbulence anisotropy because the turbulent stresses normal
cation behaviour is influenced by the cyclone shape, the to the free surface drop to zero as the free surface is
flow field and the flow turbulence. approached. Finally hydrocyclones treat particle laden
Modelling the hydrodynamics of hydrocyclones by com- flows, which influences the turbulence because particles,
putational fluid dynamics (CFD) is a key to understanding which are smaller than the length scale of the turbulence,
how they behave, however, the CFD is a non-trivial are known to damp the turbulence, whilst the turbulence
also influences the mixing of the particles and thus affects

the partition behaviour.
Correspondence to: Dr M. Brennan, The Julius Kruttschnitt Mineral
Research Centre, The University of Queensland, Isles Rd, Indooroopilly,
In this paper the results of two phase water/air CFD
Queensland, 4068, Australia. studies of the 75 mm cyclone of Hsieh (1988) are reported.
E-mail: m.brennan@uq.edu.au In particular velocity predictions from large eddy

495
496 BRENNAN

simulations (LES) and the differential Reynolds stress model (Manninnen et al., 1996) was used to simulate
turbulence model (DRSM) are compared Hsiehs (1988) medium segregation (Brennan, 2003), but it was apparent
LDA data. that medium segregation was over predicted.

PREVIOUS CFD WORK


CFD MODEL APPROACH
Boysan et al. (1982) solved the Reynolds averged
Averaged and Filtered Equations of Motion
Navier Stokes equations for a single phase gas cyclone
using an algebraic turbulence model. Subsequently, Hsieh CFD simulations of turbulent flows have traditionally
(1988) investigated hydrocyclones, with further publi- used the Reynolds averaged Navier Stokes equations
cations by co-workers from the University of Utah. (RANS), where the equations of motion are averaged
(Hsieh and Rajamani, 1991; Monredon et al., 1992; over a time scale which is long relative to the time scale
Devulapalli and Rajamani, 1994, 1996). Hsieh (1988) of the turbulent fluctuations and hence none of the turbulent
measured the velocities and turbulence parameters in a fluctuations are resolved by the simulation. The averaging
75 mm glass cyclone with a water feed and operating operation results in six additional unknowns in the Rey-
with an air core using laser Doppler anemometry. This nolds averaged form of the Navier Stokes equations
data was used to validate a two-dimensional axi-symmetric which account for the transfer of momentum by the unre-
CFD simulation which used a mixing length eddy viscosity solved turbulence and these unknowns are commonly
turbulence model where different mixing length constants called the Reynolds stresses. Model equations, typically
were used for the two components of the momentum using either a 2 equation, or a second moment turbulence
equation. Although Hsiehs CFD model required calib- model, must be supplied to calculate the Reynolds stresses
ration, it was able to simulate with reasonable accuracy and obtain a numerical solution.
the average velocity profiles inside the cyclone body and RANS turbulence models are empirical and typically are
demonstrated the effect of swirl and flow reversal on calibrated for flows where the turbulence is fully developed
turbulence behaviour. Subsequently the same approach (or at equilibrium with respect to turbulent production and
was used by Devullapalli and Rajamani (1994, 1996) who dissipation). RANS models can be recalibrated for other
modelled and measured a 250 mm Krebs D10 hydrocyclone flows but this immediately implies that RANS models
design. lack universality. Various modifications to RANS models
CFD studies of cyclones have continued since Hsiehs have been proposed to improve their universality and the
(1988) work where authors have used various forms of literature in this area is immense. A good discussion of
the k-1 model to model the turbulence (Malhotra et al., the issues associated with RANS turbulence modelling
1994; Fraser et al., 1997; He et al., 1999) but recent studies has been given by Wilcox (1998).
(Slack and Wraith, 1997; Slack et al., 2000; Suasnabar, Advances in computer hardware have meant that LES
2000; Cullivan et al., 2003) have suggest that the turbu- are becoming practical for engineering CFD problems. In
lence in hydrocyclones is too anisotropic to simulate accu- LES the larger scales of the turbulence are resolved by inte-
rately with two equation turbulence models and that at least grating the equations of motion in time and only the turbu-
a second moment (or Reynolds stress) turbulence model is lent scales which are smaller than the grid are modelled.
needed. Slack et al. (2000) also studied the flow inside The fact that an LES resolves much of the turbulence with-
single phase gas cyclones using LES on a finer grid, and out modelling implies that LES should be more accurate
LES seemed to give good predictions. Slack et al. (2000) than a RANS simulation.
in particular noted that the turbulence in gas cyclones With LES, the equations of motion are filtered and the
was developing, i.e., a flow where the production of turbu- filtering operation also results in six additional unknowns
lence exceeded the turbulent dissipation. which account for the transfer of momentum by turbulent
These more recent studies (Slack et al., 2000; Suasnabar, eddies which are smaller than the grid. These unknowns
2000; Cullivan et al., 2003) have also indicated that CFD of are known as the sub grid scale (SGS) stresses and must
hydrocyclones needs a three-dimensional grid to capture be modelled if the equations of motion are to be solved.
asymmetries in the flow. Whilst this re-introduces the same questions of accuracy
The JKMRC has completed initial CFD studies of dense and universality that apply to RANS models, in principle
medium and classifying cyclones using FluentTM (Brennan the grid can be made sufficiently small that the SGS stresses
et al., 2002; Brennan, 2003) and is using gamma ray tomo- are isotropic and can be modelled simply.
graphy to measure density profiles in dense medium The disadvantages of LES are primarily computational;
cyclones (Subramanian, 2002). The density profiles from LES requires a finer mesh than a comparable RANS simu-
this experimental work are being used to validate the lation and LES intrinsically produces an unsteady solution
CFD modelling. The initial CFD work was a 3-D simu- in time, whereas a time averaged steady state solution can
lation of a 350 mm Dutch State Mines pattern body and usually be obtained for a RANS simulation. LES has pri-
was a two phase simulation (air/water) which predicted marily been used for the single phase flows used in aerody-
the position of the air core and flow splits with reasonable namic and meteorological applications and there are still
accuracy. The work used the volume of fluid (VOF) model unanswered questions about appropriate SGS models for
(Hirt and Nichols, 1981) to predict the position of the air the multiphase flows encountered in mineral processing
core and the FluentTM implementation of the differential applications.
Reynolds stress turbulence (DRSM) model. Subsequently The Reynolds averaged and filtered equations of motion
a DRSM based multiphase approach using the mixture have a common form where tij is a symmetric tensor

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
CFD SIMULATIONS OF HYDROCYCLONES WITH AN AIR CORE 497

containing the six unknown stresses which must be equations for the shear components of the Reynolds stres-
modelled: ses. Modelling Pij has been the subject of much research
and most workers have taken the approach where Pij is
@ r @ rui the sum of slow and fast terms (Wilcox, 1996):
0 (1)
@t @xi
@ @ @uk
( rui ) (rui uj ) Pij Aij Mijkl (6)
@t @xj @xl
@ @ Both the slow pressure strain term Aij and the fast pressure
 p (tmij tij ) rgi (2)
@xi @xj strain term Mijkl are assumed to be functions of the magni-
tude of what is commonly called the Reynolds stress aniso-
The Differential Reynolds Stress Turbulence Model tropy tensor bij:
(DRSM)
u0i u0j  2=3kdij
The RANS simulations in this study have used the FluentTM bij (7)
implementation of the DRSM which is based on the Launder 2k
et al. (1975) second moment closure and solves a transport
equation for each of the six unique Reynolds stresses:
The slow pressure strain term Aij redistributes the Reynolds
tij tji ru0i u0j (3) stresses based on their magnitude and is also called the
return to isotropy term, whilst the fast pressure strain
The exact DRSM transport equations contain a number of term Mijkl is multiplied by the velocity gradients and this
unknown correlations which must be modelled. These sensitises the pressure strain model (and redistributes the
unknowns are; the dissipation tensor 1ij, the pressure strain Reynolds stresses according to) to flow strain, flow rotation,
correlation tensor Pij, and the turbulent diffusion of the Reynolds stress production and convection of the Reynolds
Reynolds stresses expressed by the tensor Cijk. The Launder stresses. Most models for Aij and Mijkl are truncated taylor
et al. (1975) model assumes that dissipation occurs only in series expansions in bij. However, the higher order expan-
the transport equations for the normal stresses and solves a sions generate a large number of additional terms contain-
transport equation for a scalar dissipation rate 1 (i.e., isotropic ing adjustable constants and much of the research work
dissipation). The Launder et al. (1975) model models Cijk seems to attend to reducing these terms to a manageable
using a generalized gradient diffusion hypothesis; however level.
the FluentTM implementation uses a simpler turbulent vis- FluentTM has a number of options for Pij and in this
cosity for stability reasons. The DRSM transport equations work the linear pressure strain (LPS) model of Launder
that are thus solved in FluentTM are et al. (1989) with the wall reflection term and the quadratic
pressure strain (QPS) model of Speziale et al. (1991) were
D used. The LPS model is linear in bij whilst the QPS is a
(ru0i u0j ) rPij r1ij  rPij quadratic expansion in bij and the QPS allows for strained
Dt
  flows where Reynolds stress redistribution would be non-
@ @ 0 0 linear. In particular the QPS does not need the LPS wall
m u u rCijk
@xk @xk i j reflection term, which introduces a non linear function
into the LPS in the vicinity of a boundary, to model the cor-
@uj @ui
Pij u0i u0k  u0j u0k rect flow behaviour in wall bounded regions.
@xk @xk (4)
m @ 0 0
rCijk t uu
s k @xk i j
2 Large Eddy Simulation (LES)
r1ij  dij r1
3 In a LES, tij in equation (2) contains the sub grid scale
s k 0:82 stresses, which in this work are modelled with the
SmagorinskyLilly (Smagorinsky, 1963; Lilly, 1966) model:
   1
D @ m @ tij  tkk dij 2mt,s Sij (8)
(r1) m t 1 3
Dt @xj se @xj
1 1 12 (5) The SmagorinskyLilly model calculates the SGS eddy vis-
C11 Pii  C21 r cosity algebraically from a length scale Ls and the mean local
2 k k
strain rate
s1 1:09, C11 1:44, C21 1:92
mt,s rLs jSj (9)
The pressure strain correlation tensor Pij acts to redistribute
the individual Reynolds stresses. In the case of the Launder Ls is normally equal to a third power of the finite volume
et al. (1975) DRSM model, which assumes isotropic dissi- size at each grid point in regions of high turbulence but the
pation, Pij also provides the only sink in the transport FluentTM implementation also makes Ls a function of the

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
498 BRENNAN

distance from the wall in wall bounded regions:

Ls min (kd, Cs Vg1=3 ) (10)

Cs is the SGS calibration constant which in this work was 0.1.

Air Core Modelling


Multiphase flows can be simulated using a full Eulerian
multiphase model, which is the most sophisticated
approach but solves a full set of transport equations for
each phase in the mixture. The mixture model (Manninnen
et al., 1996) is a less numerically intensive approach to
simulating multiphase flows where one of the phases is dis-
persed. The mixture model is derived from the full Eulerian
multiphase transport equations by making two simplifying
assumptions: (i) that the dispersed phases are moving at
their terminal slip velocity relative to the continuous fluid
phase and (ii) the interphase momentum transfer can be for-
mulated by a simple drag calculation. Assumption (i) obvi-
ates the need to solve separate momentum equations for
each phase in the system and the mixture model only
solves the equations of motion for the fluid mixture and
transport equations for the volume fractions of any
additional dispersed phases

@
ak r  (ak um ) r  (ak ukm ) 0
@t (11)
ukm uk  um

The ukm is the drift velocity of the phase k with respect to


Figure 1. Image of grid used in hydrocyclone simulations showing feed
the mixture and is calculated from the slip velocities of the port, and detail of underflow.
other dispersed phases:

X
n
ak r k continuous air phase and in the context of hydro-cyclones
ukm ukc  ulc
l1
rm (12) can treat a case where air is present both as a dispersed
phase in the feed, and as a continuous phase in the air core.
ukc uk  uc

ukc is the slip velocity of the dispersed phase k relative to Grid, Boundary Conditions and Problem Set Up
the continuous fluid phase c and is calculated from the The simulations used FluentTM V6. Initial work was
equilibrium drag assumption. done using V6.1.22 and more recent work used 6.2.15.
A further simplification of the mixture model is the VOF Three dimensional body fitted grids were used with an
(Hirt and Nichols, 1981). The VOF model is designed to accurate geometric model of the Hsieh cyclone body con-
model multiphase flows where the phases segregate totally. sisting of the feed port, main body and vortex finder. The
It solves the equations of motion for the mixture and an grids were generated using GambitTM using the Cooper
additional transport equation for each additional phase meshing facility; however the grids have been set up so
which is essentially identical to equation (11) except that that in the main body they were essentially a cylindrical
the drift velocity ukm is not calculated. O grid. The feed port used a velocity inlet boundary con-
In this work the air core has been simulated with both the dition, whilst the underflow and overflow were pressure
mixture and VOF models and the results are compared because
both models have applications in solving CFD of the multi-
phase flows encountered in mineral processing. The VOF Table 1. Grids used in simulations showing number of grid points in each
model is best suited to model CFD problems where there is coordinate. Radial is the number of points between the vortex finder and
a clear air/water free surface between a continuous air and a outer wall. Total is the total number of volume elements
continuous water phase. By comparison the primary purpose Pars 1 2 3 4
of the mixture model is to model dispersed phases and
should be suitable for modelling many mineral slurries Axial 102 112 204 224
because the slurry particles are often less than 1 mm in diam- Radial 30 40 60 80
Tangential 44 44 88 88
eter, and accelerate to their terminal slip velocity quickly. Total 2.31  105 3.14  105 18.5  105 25.1  105
However, the mixture model can be still be used to model a

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
CFD SIMULATIONS OF HYDROCYCLONES WITH AN AIR CORE 499

outlet boundary conditions. Figure 1 shows an outline of a split and velocities and hence the predictions should be
typical grid and two features of the grid should be noted. independent on how the case was first evolved.
Firstly a circular inlet port was used and this was meshed The simulations were run on Silcon Graphics Altix ser-
and merged with the main grid by projecting the face vers which are part of the Queensland Parallel Supercom-
meshes from the port inlet and a surrounding section of puting Foundation. PRESTO discretization was used for
the cyclone wall in to the outer wall face of the vortex pressure. HRIC discretization was used for the air volume
finder. Secondly the grid was graded radially near the fraction with the VOF model and QUICK was used with
underflow so that a fairly coarse grid existed in the the air volume fraction with the mixture model. All other
region expected to be occupied by the air-core. equations used QUICK. HRIC and QUICK discretization
Four grids are reported here and the grid parameters have been compared using the VOF model and predict
are shown in Table 1. The number of axial mesh points essentially the same velocities, but the HRIC option gave
is the total number including the upper body and apex. a sharper resolution of the air/water free surface.
The number of radial mesh points is the number between The simulations were time consuming. In the case of the
the wall and the vortex finder, whilst 10 points were used DRSM simulations, a velocity data set could be generated
radially inside the vortex finder. after a week of simulations. The LES on the finer grids
Grids 1 and 2, which are the coarser grids, were gener- took considerably longer even though parallel processing
ated directly using Gambit. However grid 3 and grid 4 was used.
which are the finer grids were generated using grid adap-
tation where grid 3 was generated from grid 1 and grid 4
Results
was generated from grid 2. The standard FluentTM grid
adaptation algorithm was used where the entire grid in a Hsieh (1988) reported time averaged velocities. LES is
steady case study was adapted. intrinsically a dynamic simulation and will resolve the
The y in wall bounded grid points was between 50 and large scale turbulent fluctuations plus any periodic effects,
100 for grid 1 and was between 20 and 40 for grid 4, so even when the simulation has reached a time-averaged
standard wall functions were used for the DRSM simu- steady state. To compare LES predictions with time averaged
lations. In the cases of simulations using LES, these velocity data the LES instantaneous velocities must be aver-
values of the wall bounded y imply that the log layer aged over a sufficiently long sample time. In this work each
was being resolved but the viscous sub layer and transition LES was run at time averaged steady flow and a set of instan-
region were being modelled. taneous velocity predictions were generated from the LES at
Hsieh (1988) conducted measurements on a cyclone fixed 1023 s intervals. This set of instantaneous velocity pre-
where the air core was fully developed and the flow and tur- dictions was then averaged externally. Over 200 sets of
bulence have reached a time averaged steady state. The instantaneous data from the LES were found to be sufficient
simulations reported here are for similar steady operation. to give good predictions of the mean velocities.
The following strategy was evolved because it could The DRSM simulations reached a true steady state with no
obtain, with reasonable reliability, a case study with fluctuations in the computed velocities. Hence the predicted
steady flow and a stable air core. Other approaches were velocities for the DRSM simulations presented in this work
tried but the cases invariably diverged: are also an average of between four and 10 sets of instan-
taneous velocities as calculated by the solver, generated at
1. The case was initialised with a cyclone full of water
0.1 s intervals, after the simulation reached steady flow.
with the backflow air volume fraction set to zero on
All simulations were conducted using a water feed rate of
overflow and underflow boundary conditions.
1.1165 kg s21 which is equivalent to Hsiehs (1988) series 1
2. The case was run using the steady solver and the stan-
data. Hsieh (1988) measured axial velocities at four tangen-
dard k-1 model for approximately 200 iterations.
tial positions of 08, 908, 1808 and 2708 and measured tangen-
3. The DRSM model with the LPS option was then enabled
tial velocities at 0 and 1808 where the 0 1808 plane was
and the case ran for about 25 iterations using the steady
normal to the feed port. In this work, axial velocity predic-
solver and then the unsteady solver (fixed time step) was
tions are reported in the 90 2708 plane and tangential vel-
enabled and the simulation was ran as a time integration
ocity predictions are reported in the 0 1808 plane at 60,
till a central axial core of negative pressure formed,
120, 180 and 240 below the top of the cyclone.
which led to a reversed flow on the overflow and under-
Figure 2 shows a comparison between the axial vel-
flow boundary conditions.
ocities measured by Hsieh (1988) in the 90 2708 plane
4. The back flow air volume fraction on both the overflow
and those predicted by the CFD using the DRSM turbu-
and underflow boundary conditions was then set to 1 and
lence model with the linear pressure strain correlation
the simulation was run using the unsteady solver until
and the VOF model whilst Figure 3 compares tangential
the air core was fully developed and overflow and
velocities in the 0 1808 plane for the same simulations.
underflow mass flow rates matched the feed flow rate.
The simulations are compared at 60 mm, 120 mm and
The steady case study using DRSM/LPS/VOF was saved 180 mm, for grids 1, 2 and 3.
as a base case and was used as the starting case for other The results for the axial velocities in Figure 2 show that
case studies using other model options. the DRSM turbulence model predicts the axial velocities
The methodology [steps (1) (4)] does not realistically well at 60 mm, but under predicts the axial velocities at
represent the actual dynamics of air core development, lower levels. Further the DRSM turbulence model does not
because in reality the back flow air volume fraction is 1 predict correctly the asymmetry in the measured axial velo-
from start up. However, perturbations of the final steady cities. Figure 3 shows that the DRSM model consistently
cases always saw the case stabilise back to the same flow under predicts the tangential velocities at all levels and

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
500 BRENNAN

Figure 3. Comparison between tangential velocities measured by Hsieh


Figure 2. Comparison between axial velocities measured by Hsieh (1988, (1988, Series 1) and tangential velocities predicted by CFD using RSM
Series 1) and axial velocities predicted by CFD using RSM with LPS and with LPS and VOF for air core at 60 mm, 120 mm and 180 mm below
VOF for air core at 60 mm, 120 mm and 180 mm below the top of cyclone the top of cyclone in the 01808 plane. Shows effect of different grids.
in the 902708 plane. Shows effect of different grids.

that this under prediction of the tangential velocities is located and this will have the effect of driving the air inwards
actually worse with grids 2 and 3, which are the finer grids. under the centrifugal forces, resulting in a sharper numerical
Figure 4 shows a comparison between Hsiehs (1988) prediction of the phase boundary. This will feed back into the
measured velocities at 120 mm and DRSM predictions momentum equation through the density gradient across the
using; the LPS option with the VOF model, the QPS free surface, but the density gradient is much the same with
option with the VOF model and the LPS option with the both models and is in the same location.
mixture model. The simulations with the mixture model Figure 4 shows that the QPS model predicts marginally
treated the air as a dispersed phase with a notional particle better axial velocities than the LPS model near the air/
size of 1.0  1024 m. water phase boundary but in the main part of the flow the
The simulations with the Mixture model treated the air LPS and QPS axial velocity predictions are essentially the
as a dispersed phase with a notional particle size of same. Figure 4 also shows that tangential velocity predictions
1.0  1024 m. Figure 4 shows that the VOF and mixture from the LPS and QPS are almost identical with both models
models predict essentially the same velocities. This is to be under predicting the tangential velocities to the same extent.
expected as the only difference between the two models is The fact that the DRSM turbulence model as used here
that the mixture model calculates a drift velocity in the air under predicts the tangential velocities in Hsiehs cyclone,
phase transport equation. This will only be non-zero in the irrespective of grid refinement, implies that there is a pro-
region of the domain where the air/water free surface is blem with the DRSM turbulence model for this flow.

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
CFD SIMULATIONS OF HYDROCYCLONES WITH AN AIR CORE 501

Figure 4. Comparison between axial and tangential velocities measured by


Hsieh (1988, Series 1) and velocities predicted by CFD using RSM at
120 mm below the top of cyclone in the 01808 plane. Compares LPS
and QP, VOF and mixture models. CFD used grid 2.

Flow in hydrocyclones is a turbulent shear flow which


suggests that the DRSM is over-predicting the shear com-
ponents of the Reynolds stresses. This problem could be
addressed by attention to the pressure strain model. There
is little difference between the velocity predictions from
the LPS and QPS sub-models, which suggests that the
non-linearity of the QPS confers no advantage. However
the Speziale et al. (1991) QPS model is a contraction of Figure 5. Comparison between axial velocities measured by Hsieh (1988,
Series 1) and axial velocities predicted by CFD using LES and mixture
a more general form and assumes that the turbulence is model for air core at 60 mm, 120 mm and 180 mm below the top of
homogeneous. Further, the QPS reduces to the LPS when cyclone in the 902708 plane. Shows effect of different grids.
the included non-linear terms are neglected, so the under
prediction of the tangential velocities may be the result of
an incorrect assumption that applies to both pressure the radius, but in the upper apex region the velocities pre-
strain models. However under prediction of the tangential dicted by both grids are similar.
velocities could also be due inappropriate simplifying Figure 7 shows a comparison between Hsiehs average
assumptions for 1ij and Cijk. velocities at 240 mm from the top of the cyclone and
Figure 5 shows a comparison between the axial vel- those predicted by the CFD for the LES and the DRSM.
ocities measured by Hsieh (1988) in the 90 2708 plane It is difficult to interpret Hsiehs planar velocity measure-
and the ensemble averaged axial velocities predicted ments at this level with any accuracy so Hsiehs reported
using LES and the VOF model for the air core, whilst averages of all four tangential positions is used here
Figure 6 compares tangential velocities in the 0 2708 instead. The DRSM model matches the tangential average
plane. Results are shown at 60, 120 and 180 mm and are axial velocities quite well at this level. The LES with
for grids 2 and 4. The LES predictions are much closer grid 2 predicts a much sharper flow reversal in the axial
to the measured velocities than velocities predicted by the flow velocities whilst the LES with grid 4 is reasonably
DRSM turbulence model; the agreement on tangential vel- close. The LES with grid 4 gives the best prediction for
ocities is very good, but the LES still does not reproduce the tangential velocities but the LES tangential velocity
the apparent axial velocity asymmetry in Hsiehs data at predictions for grid 2 are relatively poor. This implies
180 mm. The results also show that there is some difference that grid 4 is the best grid for an LES simulation.
between the average velocities predicted by grids 2 and 4, The fact that the LES seems to predict the velocities well
with the finer grid giving a smoother velocity profile across would imply that the approach of positioning the wall

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
502 BRENNAN

Figure 7. Comparison between axial and tangential velocities measured by


Hsieh (1988, Series 1) with velocities predicted by CFD using LES and
mixutre model for air core at 240 mm below the top of cyclone. Velocities
are the average of four tangential positions. Also shown are predictions
using the RSM model with LPS and VOF.

give reasonable velocity predictions in the main part of


Figure 6. Comparison between tangential velocities measured by Hsieh the cyclone body, so adapting grid 1 (or grid 2) only
(1988, Series 1) and tangential velocities predicted by CFD using LES
and mixture model for air core at 60 mm, 120 mm and 180 mm below
below 180 mm, may be satisfactory and the resultant par-
the top of cyclone in the 0 1808 plane. Shows effect of different grids. tially adapted grid would not have the computational over-
head of grid 4.

Can the DRSM Model Be Calibrated to Give Better


bounded grid points in the log layer at y  40 100 and
Velocity Predictions?
modelling the viscous sub-layer and transition region with
a wall function does not introduce significant errors. While it is evident that LES is more accurate than the
Simulations using LES and the mixture model were also DRSM simulations, the LES are much more numerically
run with grid 2 and grid 4 but the time average velocity pre- intensive. The best LES velocity predictions are from
dictions were essentially the same as time average velocity grid 4 which contains nearly an order of magnitude more
predictions from the LES/VOF simulations, so the LES/ nodes than grid 1. In addition the LES needed a shorter
mixture simulations are not reproduced here. time step, typically 1.0  1024 s, compared to around
Adapting the grid in an LES reduces the size of the LES 1  1023 s for a DRSM simulation. Further, the LES
filter with the result that the LES resolves more of the smal- needed to run for a significant period of time and a statisti-
ler scales of turbulence. The fact that grid adaption cally sufficient amount of instantaneous velocity data
improves and more particularly flatters out the axial vel- needed to be generated, which then had to be ensemble
ocity gradients at 240 mm indicates that these smaller averaged to obtain velocity predictions. This raises the
scales contribute a significant amount of the turbulent question as to whether it might be practical to recalibrate
momentum transfer at this level and need to be resolved a DRSM simulation with the Launder et al. (1975) model
rather than modelled. Hence grids 1 and 2 are too coarse as is on a coarser grid to match either experimental results,
in this region. However LES using grids 1 and 2 still or an LES prediction at one operating point. The calibrated

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
CFD SIMULATIONS OF HYDROCYCLONES WITH AN AIR CORE 503

DRSM simulation could then be used in engineering simu- adjustable constants and this simplifies the task. The Laun-
lations at other conditions with some confidence at con- der et al. (1989) linear model is
siderably less computational cost than an LES.  
Calibration of the existing model is also a useful first step 1 0 0 2
Pij C1 r ui uj  dij k
to identifying what might be needed in terms of model k 3
modifications.  
The DRSM model used in this study implements dissipa- 1
 C2 Pij  Dij  dij (Pkk  Dkk )
tion only in the transport equations for the Reynolds normal 3
stresses, which are the diagonal components of the Reynolds  
@uj @ui
stress tensor. Apart from convection and turbulent diffusion, 0
Pij  ui uk 0 0
uj uk0 , (13)
the only sink in each transport equation for the Reynolds @xk @xk
shear stresses is the associated shear component of the  
@uk @uk
pressure strain tensor. This implies calibration could be Dij  u0i u0k u0j u0k
@xj @xi
accomplished by adjusting the pressure strain model con-
stants with the aim of damping the Reynolds shear stresses, 1
which act on the tangential velocity. k u0k u0k , C1 1:8, C2 0:6 (default)
2
Calibration of the linear pressure strain model has been
considered here because the linear model has only two The first constant C1 adjusts the slow, or return to isotropy,

Figure 8. Comparison between axial velocities measured by Hsieh (1988, Figure 9. Comparison between tangential velocities measured by Hsieh
Series 1) with axial velocities predicted by CFD using RSM/LPS with (1988, Series 1) with tangential velocities predicted by CFD using
VOF, but with changes in C1 and C2 in the LPS model (C1 1.8, RSM/LPS with VOF, but with changes in C1 and C2 in the LPS model
C2 0.6 are default values). (C1 1.8, C2 0.6 are default).

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
504 BRENNAN

component of the pressure strain tensor which redistributes However, simulations at other feed rates and other cyclone
the stresses based on their magnitude and acts as a sink for geometries with these adjusted constants have not been
the Reynolds shear stresses. Increasing C1 will reduce the conducted.
magnitude of the Reynolds shear stresses but will also
make Reynolds normal stresses more equal to each other
(i.e., make the turbulence more isotropic), which would NOMENCLATURE
seem to be undesirable. Ak particle cross-sectional area of phase k
The second constant C2 adjusts the fast pressure strain Aij slow pressure strain tensor
bij dimensionless Reynolds stress anisotropy tensor
term which redistributes the stresses based on the magni- Cd drag coefficient
tude of production and convection. Increasing C2 will Cij convective transport tensor
reduce the production and convection of the Reynolds Cijk turbulent transport tensor
shear stresses but will also make the production of the Rey- Cs Smagorinsky Lilly constant
C1 Slow linear pressure strain model constant
nolds normal stresses more isotropic, which would also C2 Fast linear pressure strain model constant
seem to be undesirable. d distance from wall
Figures 8 and 9 show the effect of adjustments to C1 and g gravity vector
C2 on the predicted velocities. Increasing C2 to 0.9 gi i component of gravity vector
improves both the axial and tangential velocity predictions. k turbulent kinetic energy
Ls length scale of the sub grid scale stresses
However, the peak in the tangential velocities is somewhat Mp0 fluctuating component of interphase momentum transfer
broader than measured. Reducing C1 to 1 seems to have a Mijkl fast pressure strain tensor
limited effect except that it makes the tangential velocity Pij turbulent production tensor
peak at 120 mm somewhat broader. Increasing C2 much Sij mean strain rate
t time
above 0.9 resulted in a satisfactory prediction of the tangen- ukc velocity vector of phase k relative to continuous phase
tial velocities but had the effect of damping the Reynolds ukm velocity vector of phase k relative to mixture velocity
stress components in the axial direction so significantly um mixture velocity vector
that the predicted axial velocities started to fluctuate ui0 i component of velocity vector
time-wise. ui0 i component of fluctuating velocity
Vg volume of grid finite volume
It is likely that the non-linear constants in the fast Vk particle volume of phase k
pressure strain term of the quadratic pressure strain model xi i co-ordinate
could be adjusted to achieve a satisfactory velocity predic- y dimensionless distance from wall
tion as well and investigating this option would be a useful
next step. Greek Symbols
ak volume fraction of dispersed phase k
dij Kronecker delta
CONCLUSIONS 1ij dissipation tensor
Pij pressure strain correlation tensor
CFD predictions of the 75 mm hydrocyclone of Hsieh r density
(1988) have been conducted in FluentTM using 3 D body tij turbulent or sub grid scale stress tensor
fitted grids where the air core is resolved with the VOF tmij viscous stress tensor
m molecular viscosity
(Hirt and Nichols, 1981) and mixture (Manninen et al., mt eddy viscosity
1996) models and have been compared to Hsiehs mt,s sub-grid scale eddy viscosity
measured LDA data. It has been found that simulations
using the differential Reynolds stress turbulence model
under predict the axial velocities in the apex and consist-
ently under predict the tangential velocities. There is rela- REFERENCES
tively little difference between the predictions from the Brennan, M.S., Holtham, P.N., Rong, R. and Lyman, G.J., 2002, Compu-
standard LPS model (Launder et al., 1975) and the QPS tational fluid dynamic simulation of dense medium cyclones, Proceed-
model (Speziale et al., 1991). Velocity predictions using ings 9th Australian Coal Preparation Conference, Yeppoon Australia,
the VOF and mixture models are essentially the same. 1317 October 2002, Paper B3.
Brennan, M.S., 2003, Multiphase CFD simulations of dense medium and
The results are independent of the grid with grids of classifying hydrocyclones, Third International Conference on CFD in
between 2.2  105 and 2.4  106 nodes giving essentially the Minerals and Process Industries, CSIRO, Melbourne, Australia,
the same velocity predictions. 1012 December, 59 63.
LES using the basic Smagorinsky Lilly (Smagorinsky, Boysan, F., Ayers, W.H. and Swithenbank, J., 1982, Trans IChemE, 60:
222230.
1963; Lilly, 1996) sub-grid scale model give good agree- Cullivan, J.C., Williams, R.A. and Cross, C.R., 2003, Understanding the
ment with Hsiehs velocity data, however the best agree- hydrocyclone separator through computational fluid dynamics, Trans
ment is obtained using a grid of 2.4  106 nodes. IChemE, 81A: 455466.
Under prediction of the tangential velocities by the Devulapalli, B. and Rajamani, R.K., 1994, Application of LDV to the
DRSM model implies that with the both pressure strain modelling of particle size classification in industrial hydrocyclones,
ASME FED, 191: 41 48.
options, the model is over predicting the Reynolds shear Devulapalli, B. and Rajamani, R.K., 1996, A comprehensive CFD model
stresses. DRSM simulations where the constants in the for particle-size classification in industrial hydrocyclones, Hydrocy-
LPS model have been adjusted from the default values clones 96, Cambridge, UK, 83 104.
of C1 1.8 and C2 0.6 have been conducted. It has Fraser, S.M., Abdel Rasek, A.M. and Abdullah, M.Z., 1997, Compu-
tational and experimental investigations in a cyclone dust separator,
been found that increasing the fast pressure strain constant Proc Instn Mech Engrs, 211E: 247 257.
C2 to 0.9 improves the velocity predictions. These adjust- He, P., Salcudean, M. and Gartshore I.S., 1999, A numerical simulation of
ments also reduce the predicted Reynolds stresses. hydrocyclones, Trans IChemE, 77(A): 429441.

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505
CFD SIMULATIONS OF HYDROCYCLONES WITH AN AIR CORE 505

Hirt, C.W. and Nichols, B.D., 1981, Volume of fluid (VOF) method for the Smagorinsky, J., 1963, General circulation experiments with the primitive
dynamics of free boundaries, Journal of Computational Physics, 39: equation. I. The basic experiment, Monthly Weather Review, 91:
201 225. 99164.
Hsieh, K.T., 1988, A phenomenological model of the hydrocyclone, PhD Speziale, C.G., Sarkar, S. and Gatski, T.B., 1991, Modeling the pressure
Thesis, University of Utah. strain correlation of turbulence, J Fluid Mech, 227: 245272.
Hsieh, K.T. and Rajamani, R.K., 1991, Mathematical model of the hydro- Suasnabar, D.J., 2000, Dense medium cyclone performance, enhancements
cyclone based on the physics of fluid flow, AIChE J, 37: 735. via computational modeling of the physical process, PhD Thesis, Uni-
Launder, B.E., 1989, Second-moment closure: present . . . and future?, versity of New South Wales.
International Journal of Heat and Fluid Flow, 10: 282 300. Subramanian, V.S., 2002, Measuring medium segregation in the dense
Launder, B.E., Reece, G.J. and Rodi, W., 1975, Progress in the develop- medium cyclone using gamma ray tomography, PhD Thesis, JKMRC,
ment of a Reynolds-stress turbulence closure, J Fluid Mech, 68: University of Queensland.
537 566. Wilcox, D.C., 1998, Turbulence Modelling for CFD (DCW Industries, La
Lilly, D.K., 1966, On the application of the eddy viscosity concept in the Canada California).
inertial subrange of turbulence, NCAR manuscript 123.
Malhotra, A., Branion R.M.R. and Hauptmann, E.G., 1994, Modeling the
flow in a hydrocyclone, The Canadian Journal of Chemical Engineer- ACKNOWLEDGEMENTS
ing, 72: 953 960.
Manninen, M., Taivassalo, V. and Kallio, S., 1996, On the Mixture Model The computational resources that were used in this work are part of the
for Multiphase Flow (Valtion Teknillinen Tutkimuskeskus, Espoo, Queensland Parallel Supercomputing Foundation. The author would also
Finland). like to thank Prof. Klaus Bremhorst, Mechanical Engineering UQ and
Monredon, T.C., Hsieh, K.T. and Rajamani, R.K., 1992, Fluid flow of the Dr Peter Holtham JKMRC UQ for their assistance in preparing this paper.
hydrocyclones: An investigation of the device dimensions, International
Journal of Mineral Processing, 35: 65 83.
Slack, M.D., Prasad, R.O., Bakker, A. and Boysan, F., 2000, Advances in NOTES
cyclone modeling using unstructured grids, Trans IChemE, 78(A):
10981104. Fluent and Gambit are trademarks and software of Fluent Inc.
Slack, M.D. and Wraith, A.E., 1997, Modelling the velocity distribution in
a hydrocyclone, 4th International Colloquium on Process Simulation, The Manuscript was received 9 May 2005 and accepted for publication
1113 June, Espoo, Finland, 6583. after revision 10 March 2006.

Trans IChemE, Part A, Chemical Engineering Research and Design, 2006, 84(A6): 495 505

S-ar putea să vă placă și