You are on page 1of 12

Multiphase modelling of hydrocyclones prediction of cut-size

M.S. Brennan
a,
*
, M. Narasimha
b
, P.N. Holtham
a
a
Julius Kruttschnitt Mineral Research Centre, The University of Queensland, Isles Road, Indooroopilly 4068, Queensland, Australia
b
R&D Division, Tata Steel, Jamshedpur, Jharkhand 831 007, India
Received 1 August 2006; accepted 6 October 2006
Available online 28 November 2006
Abstract
A comprehensive multiphase model of cyclone separators using Computational Fluid Dynamics is under development. The model is
capable of predicting velocity proles, ow splits, air core position and eciency curves in classifying hydrocyclones. The model
approach uses the Mixture model with the granular options and large eddy simulation (LES) to resolve the turbulent mixing of the par-
ticles. Multiphase simulations of Hsiehs [Hsieh, K.T., 1988. A phenomenological model of the hydrocyclone, Ph.D. thesis, University of
Utah] data show a very good prediction of the cyclone eciency curve. Whilst further model development is needed, the approach is
showing promise as a cyclone design tool.
2006 Elsevier Ltd. All rights reserved.
Keywords: Hydrocyclones; Classication; Computational uid dynamics; Large eddy simulation; Mixture model
1. Introduction
Hydrocyclones rely on the centrifugal forces that
develop under the swirling ow to eect separation of par-
ticles in slurries and particle laden ows and can classify on
density or particle size. Modelling the ow by computa-
tional uid dynamics (CFD) is a key to understanding
how cyclone separators behave. However the ow is multi-
phase, usually turbulent and an air core usually forms if the
cyclone outlets are exposed to the atmosphere. These fac-
tors inuence the separation behavior and must be
accounted for even though they complicate the CFD
modeling.
Turbulent ows in industrial processes have mostly been
modeled by CFD using RANS based turbulence models.
However previous studies (Slack and Wraith, 1997; Suasn-
abar, 2000; Cullivan et al., 2003; Delgadillo and Rajamani,
2005) have shown that the turbulence in hydro-cyclones is
too anisotropic to model with k e models and the veloc-
ities are predicted more accurately with a 2nd moment
closures.
In recent years large eddy simulation (LES) has become
available in commercial CFD packages. Whilst computa-
tionally more intensive than a RANS simulation, LES
resolves the largest turbulent scales without modeling and
has been shown to predict the mean ows in cyclone sepa-
rators with reasonable accuracy for single phase gas ows
(Slack et al., 2000; Shalaby et al., 2005) and the basic
two phase ow (water feed plus air core) in hydrocyclones
(Delgadillo and Rajamani, 2005; Brennan, 2006). In princi-
ple LES should also be able to resolve the turbulent mixing
of particles, provided that the eect of particle loading on
the ow and turbulence is accounted for in the simulation.
In this paper we report the results of multiphase CFD
studies of the Hsieh cyclone (1988) with a limestone feed,
where the turbulence has been resolved using large eddy
simulation (LES) and the limestone has been simulated
using the Mixture model (Manninen et al., 1996) with the
size distribution approximated by six phase transport equa-
tions. The tromp (cyclone eciency) curve has been con-
structed from the time averaged ow rates of each size
range out the underow and overow boundaries. The
0892-6875/$ - see front matter 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.mineng.2006.10.010
*
Corresponding author. Tel.: +61 7336 55888; fax: +61 7336 55999.
E-mail address: m.brennan@uq.edu.au (M.S. Brennan).
This article is also available online at:
www.elsevier.com/locate/mineng
Minerals Engineering 20 (2007) 395406
predicted distribution of each size in the distribution inside
the cyclone body are shown.
2. Previous work
CFD of hydrocyclones has primarily focused on model-
ing the basic ow behavior, which is understandable as
even a basic two phase water/air core simulation using a
RANS turbulence model still requires between one and
two days to reach convergence using current hardware.
As noted by Slack and Wraith (1997), Suasnabar (2000)
and Cullivan et al. (2003), the turbulence anisotropy in
hydrocyclones means that a RANS based CFD solution
of a hydrocyclone needs to use a 2nd moment closure (or
Dierential Reynold Stress Turbulence Model, DRSM)
where the transport equations for the individual Reynolds
stresses are solved. However both Delgadillo and Rajamani
(2005) and Brennan (2006) have shown that the Launder
et al. 2nd moment model (1975) under predicted the tan-
gential velocities in simulations of the Hsieh cyclone
(1988) and that large eddy simulation gave more accurate
predictions of the mean velocities.
The goal of a CFD simulation of a classifying hydrocycl-
oner is to predict classication and if this goal is to be
achieved then the interactions between the particulate
phases and the uid must be modeled. In CFD there are
two approaches to simulating particulate phases in uid sys-
tems; these are the Lagrangian and Eulerian approaches.
A Lagrangian simulation runs on top of a single
phase CFD simulation and simulates the paths of individ-
ual particles through the uid by a double integration of
the particle acceleration, which is calculated from a force
balance on the particle. The particle force balance includes
drag, buoyancy forces and other body forces and the drag
force is calculated from the local uid velocity as predicted
by the single phase CFD calculation. The particle path cal-
culation starts from an initial position, which is usually the
feed boundary condition and terminates when the particle
has left the domain via an outlet boundary condition or
reaches a stagnant zone.
Eulerian CFD techniques treat the dispersed phases as a
pseudo continuum and solve transport equations for the
phase concentrations. Flows containing dispersed solid
phases at high concentrations and where inter-particle col-
lisions result in additional stresses can be simulated using a
full Eulerian granular ow approach (Ding and Gidaspow,
1990; Gidaspow et al., 1992) where momentum and conti-
nuity equations are solved for both the dispersed phases
and the continuous phases. The full Eulerian granular ow
approach also solves transport equations for the continu-
ous phase turbulence and also solves transport equations
for the granular temperature and granular pressure which
are used to calculate eective viscous stresses in the dis-
persed phase momentum equation.
A simplied Eulerian approach is the Mixture model
(Manninen et al., 1996) where the momentum and continu-
Nomenclature
Transport equations in this paper have generally used cartesian tensor notation. Velocities, densities, viscosities,
pressure and stress tensor terms associated with the equations of motion refer to the mixture unless denoted with
the subscript k (dispersed phase k) or c (continuous uid phase)
A
k
particle cross-sectional area of phase k
C
d
drag coecient
C
s
Smagorinsky Lilly constant
e
kk
coecient of restitution of phase k
d
k
diameter of phase k
g
i
i component of gravity
h distance from wall
k turbulent kinetic energy
L
s
length scale of the sub grid scale stresses
M
k,i
i component of interphase momentum transfer
to phase k
p
k
granular pressure of phase k
S
ij
mean strain rate
t time
u
i
i component of mixture velocity
u
kc,i
i component of velocity of phase k relative to
mixture (drift velocity)
u
km,i
i component of velocity of phase k relative to
continuous phase (slip velocity)
V
g
volume of grid nite volume
V
k
particle volume of phase k
x
i
i co-ordinate
y
+
dimensionless distance from wall
a
k
volume fraction of phase k
d
ij
kronecker delta
j Von Karmanns constant = 0.41
q mixture density
s
d,ij
drift stress tensor of mixture
s
s,ij
turbulent or sub grid scale stress tensor of mix-
ture
s
lij
viscous stress tensor of mixture
H
k
granular temperature of phase k
l eective molecular viscosity of the mixture
l
s
sub grid scale eddy viscosity of the mixture
Subscripts
c continuous phase c
i, j components in i and j directions
k phase k
m mixture
396 M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406
ity equations are only solved the mixture are solved and a
transport equation, which contains an equilibrium slip
velocity, is solved for the volume fraction of each phase.
The Mixture model has been modied to incorporate some
of the features of granular ow theory and these modica-
tions are discussed in the next section.
The advantage of the Lagrangian multiphase approach
is that particleparticle and particleuid interactions are
calculated dynamically for every particle present in the sys-
tem based on the instantaneous velocity of the particle. By
comparison the Eulerian approach only calculates a phase
velocity, a phase volume fraction and overall stresses asso-
ciated with the average behavior of an ensemble of phase
particles in a each nite volume in the CFD grid. Thus
the Eulerian approach innately involves averaging of some
sort, which implies more modelling. However the computa-
tional requirements of the Lagrangian approach scale
directly with the number of particles, whereas the computa-
tional requirements of the Eulerian approach only scale
with the number of resolved phases. In recent simulations
(Boivin et al., 2000), 885,000 particles have been simulated
but even with this number of particles, the particle loading
is still dilute and is miniscule compared to the number of
particles present in a small cyclone separator at reasonable
particle loadings (ca $ 10
9
particles). For higher particle
loadings the Eulerian approach is computationally more
practical.
Hsieh (1988) used the Lagrangian approach to predict
classication of limestone in a 75 mm hydrocyclone. Hsieh
used the mean velocity eld from his 2D axi-symmetric
CFD calculation of the water phase and did not incorporate
any eects due to the turbulence. The Tromp (or cyclone
eciency) curve was constructed from repeated simulations
for dierent particle sizes. Hsiehs simulations predicted the
d
50
of the 75 mm cyclone for limestone very well, but the
short circuiting of large particles was not predicted and
the cut above the d
50
was much sharper than was measured.
Hsieh (1988) attributed this shortcoming to the 2D axi-sym-
metric grid which used what was eectively a ring inlet and
Hsieh suggested that the eects of turbulence and particle/
particle interactions should be incorporated in the simula-
tion. Devulapalli (1994,1996,1997) extended Hsiehs
Lagrangian approach, and incorporated the eects of tur-
bulence by using a Lagrangian cloud model developed by
Baxter and Smith (1993), where the mean position of a
cloud of particles is calculated and where the cloud was
released from the cyclone inlet. The spread of the cloud
about the mean position is calculated by a probability dis-
tribution function (dened as a function of the turbulence)
as the cloud moves through the domain. Devulapalli (1997)
used the Lagrangian cloud approach to predict Tromp
curves for a variety of 250 mm hydrocyclones (using several
dierent vortex nders and spigots) passing limestone and
copper ore. Again the d
50
was predicted well, but again
the cut above the d
50
was sharper than was observed from
real experiments using a cyclone of the same geometry.
Devullapalli suggested that a full 3D grid of the cyclone,
with better turbulence models and other particle eects were
needed to improve these predictions.
The Lagrangian approach has been used to model mul-
tiphase ows using large eddy simulation for the continu-
ous uid phase (Boivin et al., 2000; Portela and
Oliemans, 2001; Millelli et al., 2001). These studies have
focused on simulating the damping of the turbulence by
small dense suspended particles and have used two-way
coupling where the momentum exchange with the particles
by drag is incorporated as a source term in the uid
phase momentum equation. The studies by Portela and
Oliemans (2001) and Millelli et al. (2001) have also mod-
eled the eect of the dispersed phases on the LES sub grid
scale model.
The Eulerian approach has been used to simulate med-
ium segregation in dense medium cyclones by Suasnabar
(2000). Both the full Eulerian multiphase granular ow
model (Ding and Gidaspow, 1990; Gidaspow et al., 1992)
and the simplied Mixture model (Manninen et al., 1996)
were used to model the medium and a single medium par-
ticle size was used. The Eulerian Granular ow model and
the Mixture model predicted similar velocities and medium
concentrations. The Eulerian granular ow model was felt
to be more fundamentally correct but the Mixture model
was computationally more economical.
The Mixture model has been the mainstay of the multi-
phase cyclone CFD studies at the JKMRC. In our initial
work (Brennan, 2003) the medium in a dense medium
cyclone was modeled with a single particle size. It was
found that medium segregation was over predicted com-
pared to experimental measurements of the slurry density
in a 350 mm DSM cyclone by gamma ray tomography
(Subramanian, 2000). This was attributed to the fact that
only a single medium particle size was used. It was also
found that the medium concentration was generally highest
at the wall, whereas the measurements suggested that in
apex the density reached a peak in a central zone between
the wall and the air core. In later work (Narasimha et al.,
2006) the medium distribution was simulated and the Mix-
ture model slip velocity was modied to incorporate a lift
force near the wall based on the Saman (1965, 1968)
(expression), and the mixture viscosity was modied to
use the model of Ishii and Mishima (1984) which increases
the mixture viscosity to innity as the dispersed phase
packing density was approached. These modications
improved the predictions of the density proles consider-
ably, but the lift force coecient needed to be increased
substantially over the value used by Saman to achieve this
improvement.
2.1. CFD approach
The ow in hydrocyclones is turbulent and is a multi-
phase ow. In this work the turbulence has been resolved
using large eddy simulation (LES) for the mixture and
phase segregation has been modeled using the Mixture
model (Manninen et al., 1996).
M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406 397
2.2. LES equations
The principle behind LES is that the larger turbulent
eddies are resolved as part of a dynamic integration of
the equations of motion but the eddies that are smaller
the grid are not resolved and the momentum transfer due
to these sub grid scale eddies is modeled. Thus LES
resolves the momentum transfer by the largest turbulent
eddies without any modeling. In many ows these larger
turbulent scales carry out most of the turbulent mixing of
momentum and mass. In principle this means that an
LES should give more accurate mean ow predictions than
a RANS based simulation without needing to resort to any
ad-hoc tuning of the turbulence model since the problem of
matching the predictions to experimental data becomes pri-
marily a matter of rening the grid so that enough scales
are resolved and a simple sub grid scale model is adequate.
The disadvantage of LES is that the technique requires a
ner grid than is required for a RANS simulation and
the technique is intrinsically a time integration whereas a
time averaged steady ow RANS solution can be obtained
for a statistically steady turbulent ow. LES is of interest in
multiphase ows because the largest (and resolved) turbu-
lent scales carry out most of the turbulent mixing of the
particles.
In an LES the equations of motion that are solved are
the unsteady ltered form, which for this problem are for
a variable density slurry mixture:
oq
ot

oqu
i

ox
i
0 1
o
ot
qu
i

o
ox
j
qu
i
u
j

o
ox
i
p
i

o
ox
j
s
l;ij
s
d;ij
s
s;ij
qg
i
2
The viscous stress tensor in Eq. (2) is
s
l;ij
2lS
ij
3
LES requires a sub grid scale (SGS) model to calculate the
transfer of momentum by turbulent scales which are smal-
ler than the grid and which appears in Eq. (2) as the sub
grid scale stress tensor s
s,ij
. In this work the basic Smago-
rinsky Lilly SGS model (Lilly) was used, where the SGS
stresses are modelled using an SGS eddy viscosity l
s
:
s
s;ij

1
3
s
s;kk
d
ij
2l
s
S
ij
4
and the SGS eddy viscosity is calculated algebraically from
a length scale L
s
and the mean local strain rate:
l
s
qL
2
s
jSj 5
L
s
is normally equal to a third power of the nite volume
size at each grid point in regions of high turbulence but
the Fluent
TM
implementation also makes L
s
a function of
the distance from the wall h in wall bounded regions:
L
s
minjh; C
s
V
1=3
g
6
C
s
is the SGS calibration constant which in this work was
0.1.
Millelli et al. (2001) has proposed a modication to the
basic Smagorinsky Lilly SGS model which corrects the
SGS eddy viscosity in the presence of dispersed phases.
This modication has been derived from dimensional con-
siderations and is of the form:
l
s
qL
s
jSj qL
s
1 a
c
ju
c
u
k
j 7
The multiphase correction given in Eq. (7) has been tested
for the multiphase ows reported here and in Hsiehs
(1988) experimental work but was found to very small.
2.3. Mixture model equations
The Mixture model (Manninen et al., 1996) is a simpli-
ed Eulerian approach to modeling dispersed multiphase
ows, where the equations of motion are solved only for
the mixture. The mixture model assumes that the primary
phase is the continuous uid phase and solves transport
equations for the volume fractions of each distinct dis-
persed phases present in the mixture:
o
ot
a
k

o
ox
i
a
k
u
i

o
ox
i
a
k
u
km;i
0
u
km;i
u
k;i
u
i
8
The Mixture model can simulate a slurry containing solids
with a distribution of size or density by solving a transport
equation for the average size or average density over a sub-
range within the distribution.
Phase segregation is accounted for in the phase trans-
port equation (Eq. (8)) by the drift velocity of the phase
k relative to the mixture u
km,I
, which is calculated from
the slip velocities, u
kc,i
which are the velocities of the dis-
persed phases relative to the continuous uid phase:
u
km;i
u
kc;i

n
l1
a
k
q
k
q
u
lc;i
u
kc
u
k
u
c
9
Phase segregation results in an additional source of
momentum transfer in the mixture momentum equation
(Eq. (2)), known as the drift tensor s
d,ij
, which can calcu-
lated exactly as:
s
d;ij

n
k1
a
k
q
k
u
km;i
u
km;j
10
The premise behind the Mixture model is that the slip
velocities can be modeled algebraically without the need
to solve the dispersed phase momentum equations. Manni-
nen et al. (1996) derived a model for u
kc,i
by subtracting the
momentum equation for a dispersed phase k from the
momentum equation for the mixture and substituting a
constitutive drag law for the interphase momentum trans-
fer. The combined expression was then simplied by
assuming that the dispersed phase k accelerates to its local
terminal slip velocity very rapidly when compared to the
398 M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406
time scale of the ow. This simplication was called a local
equilibrium assumption. Manninen et al.s (1996) general
expression for the slip velocity u
kc,i
is:
a
k
3
4
C
dk
q
c
d
k
ju
kc
ju
kc;i
a
k
q
k
q g
i

o
ot
u
i
u
j
o
ox
j
u
i
_ _
a
k
o
ox
j
s
l;ij
s
d;ij
s
s;ij

o
ox
j
a
k
s
kT ;ij
M
ki
11
Eq. (11) includes forces due the various stress tensors in the
mixture and dispersed phase momentum equations. The
implementation of the Mixture model used here assumes
that the forces due to these stress terms are negligible in
their eect on the slip velocity when compared to the
hydrodynamic and gravity forces:
3
4
C
dk
q
c
d
k
ju
kc
ju
kc;i
q
k
q g
i

o
ot
u
i
u
j
o
ox
j
u
i
_ _
12
Eqs. (11) and (12) can be derived using Manninen et al.s
(1996) formalism but using the ltered LES equations for
the mixture. In this case the mixture velocities that appear
on the RHS are the instantaneous mixture velocities and
will include the resolved turbulent uctuations. Thus the
Mixture model when used in conjunction with an LES
should resolve the turbulent mixing of the dispersed phases
by the resolved scales.
2.4. Granular enhancement to the mixture model
The Mixture model used in this work is an enhanced
form of the Manninen et al. (1996) model which incorpo-
rates the granular viscosity, the granular pressure and the
drag law formulations from the kinetic theory of granular
ows. These granular terms require the granular tempera-
ture H
k
which in the Mixture model is calculated algebra-
ically from Eq. (13)
p
k
I s
k
: ru
k
cH
k
/
kl
0 13
Eq. (12) is derived by simplifying the transport equation for
H
k
as used in the full Eulerian granular model by assuming
that the convection and accumulation and diusion terms
are negligible compared to production, dissipation and
interphase transfer.
In the Mixture model, the eective molecular viscosity
of the mixture is treated as the weighted sum of the phase
viscosities (including the continuous phase viscosity):
l

n
k1
a
k
l
k
14
With granular options enabled, each dispersed phase vis-
cosity is treated as the sum of a collisional viscosity and
a kinetic viscosity (Gidaspow et al., 1992):
l
k
l
k;col
l
k;kin
l
k;col

4
5
a
k
q
k
d
k
g
0;kk
1 e
kk

H
k
p
_ _
1=2
l
k;kin

10q
k
d
k

H
k
p
p
96a
k
1 e
kk
g
0;kk
1
4
5
g
0;kk
a
k
1 e
kk

_ _
2
15
The phase granular viscosity (Eq. (15)) is a function of the
granular temperature H
k
, which is a function of shear via
Eq. (13) and thus is a non-newtonian model.
g
0,kk
is the radial distribution function and is a correc-
tion factor that modies the probability of collisions
between grains when the solid granular phase becomes
dense and is modeled by Eq. (16):
g
0;kk
1
1 a
c
1 a
c;min
_ _
1=3
_ _
1

1
2
d
k

N
k
k1
a
k
d
k
16
g
0,kk
appears in the denominator for the expression for the
kinetic viscosity l
k,kin
in Eq. (15) and has the main function
of numerically limiting the total dispersed phase concentra-
tion to less than the packing limit. This is eected by mak-
ing l
k,kin
(and hence the total mixture viscosity) approach
innity as the dispersed phase packing limit (=1 a
c,min
)
is approached. This feeds back into the slip calculation
and reduces the slip velocities of all phases to zero as the
packing limit is approached, which eliminates phase
segregation.
The pressure of the mixture is assumed to be the sum of
the continuous phase pressure and the granular pressure
associated with each phase (Syamlal et al., 1993):
P
s;t

N
k
k1
p
k
p
k
a
k
q
k
H
k
2q
k
1 e
kk
a
2
k
g
0;kk
H
k
17
The granular enhancements to the mixture model also per-
mit a number of dierent drag laws and in this work, the
Gidaspow et al. (1992) drag law has been used, with this
law the slip velocity calculation is:
3
4
C
dk
a
2:65
c
q
c
d
k
ju
kc
ju
kc;i
q
k
q g
i

o
ot
u
i
u
j
o
ox
j
u
i
_ _
a
c
P0:8
150
1 a
c
a
c
l
c
d
2
k
1:75
q
c
d
k
ju
kc
j
_ _
u
kc;i
q
k
q g
i

o
ot
u
i
u
j
o
ox
j
u
i
_ _
a
c
< 0:8
C
d

24
Re
k
1 0:15a
c
Re
k

0:687

Re
k

q
c
d
k
ju
kc
j
l
c
18
M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406 399
2.5. Grid, boundary conditions and problem set up
The simulations were conducted on 3D body tted grids
of Hsiehs (1988) 75 mm geometry, which were generated in
Gambit
TM
were unsteady LES simulations. The approach
used was to generate an initial steady water/air two phase
case on a coarse grid of 56,000 nodes using our existing
methodology (Brennan, 2006). A steady multiphase case
study using the Mixture model was then generated using
the coarse grid water/two phase case as an initial condition
and then a steady multiphase case using a ner grid of
450,000 nodes was generated by adapting the coarse grid
multiphase case using the Fluent adaptation algorithm.
Steady in this context meant that the overall material bal-
ance of each dispersed phase had reached a time average
steady state during the LES simulation over approximately
0.6 s of simulation time which is roughly the residence time
of the cyclone.
A ne grid water/air simulation was also generated by
adapting the coarse grid water/air simulation.
The Fluent grid adaptation algorithm divides every
nite volume element selected for adaptation in a 3D grid
in eight volume elements. Grid adaptation is a numerically
economical method of grid renement because the adapta-
tion algorithm uses a converged case study using a coarse
grid and avoids the process of remeshing the geometry
Axial Velocity - 60 mm
-1
-0.5
0
0.5
1
1.5
2
2.5
3
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
Radial position - m
V
e
l
o
c
i
t
y

-

m
.
s
-
1
Hsieh 0-180
Coarse
Fine
Fig. 1. Time averaged axial velocities from water/air LES simulations on coarse and ne grids at a feed rate of 1.116 kg s
1
, with Hsiehs (1988) series 1
LDA measurements on the same cyclone geometry, at 60 mm below the top of the cyclone body.
Tangential Velocity - 60 mm
0
1
2
3
4
5
6
7
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
Radial position - m
V
e
l
o
c
i
t
y

-

m
.
s
-
1
Hsieh 0-180
Fine
Coarse
Fig. 2. Time averaged tangential velocities from water/air LES simulations on coarse and ne grids at a feed rate of 1.116 kg s
1
, with Hsiehs (1988) series
1 LDA measurements on the same cyclone geometry, at 60 mm below the top of the cyclone body.
400 M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406
and generating a converged case from scratch. In this work
the entire grid was adapted.
All simulations were run using Fluent 6.2.16
TM
with
large eddy simulation with a xed time step of
1.0 10
4
s. Presto discretization was used for pressure
and quick discretization was used for the phase transport
equations. Both bounded central dierencing and the 3rd
order muscle discretization schemes were used on the
momentum equation. Simple was used for the pressure
velocity coupling. A velocity inlet boundary condition
was used for the feed where the inlet turbulence was gener-
ated using the Spectral Synthesiser option. The overow
and underow used pressure outlet boundary conditions
with an air back-ow volume fraction of 1. The feed air
volume fraction was set to zero.
The Mixture model was run using eight phase transport
equations in the multiphase simulations, where the primary
phase was water and the dispersed phases were air and six
limestone phases of density 2700 kg m
3
with sizes 7 lm,
9 lm, 13 lm, 17 lm, 25 lm and 35 lm. The volume frac-
tion of each limestone size in the feed boundary condition
was set so that the cumulative limestone size distribution
and the total limestone volume fraction matched those
used in the feed in Hsiehs experimental runs. The mixture
slip velocity was disabled for the air phase which meant in
eect that the air phase was being solved using the VOF
model with quick discretization.
The coarse grid simulations were run using the Windows
version of Fluent on a desktop PC using a P4 dual core
processor, whilst the ne grid simulations were run on
SGI Altix servers using the parallel solver with eight pro-
cessors. These ne grid simulations took approximately
167 h per second of simulation time.
3. Results
3.1. Two phase water/air results
Case studies using the coarse and ne grids and with
only two phases (air and water), were run with a water feed
rate of 1.116 kg s
1
. This is a feed rate equal to Hsiehs
(1988) series 1 single phase experimental results where the
water velocities were measured using laser doppler
anemometry.
Fig. 1 shows a comparison between the time averaged
tangential velocities as predicted by the LES from these
two phase studies and Hsiehs series 1 measurements at
60 mm below the top of the cyclone in the 0180 plane,
whilst Fig. 2 shows a comparison between the axial veloc-
ities at the same location. The results from both grids com-
pare well from a radial position of 0.01 m outwards, but
the ne grid gives better agreement with Hsiehs data
between 0.005 m (approximately the air/water free surface
core position) and 0.01 m. The third order muscle discreti-
zation scheme give better predictions of the velocities than
does bounded central dierencing on the coarse grid, but
on the ner grids both discretization schemes give similar
velocity predictions.
One test of an LES is to estimate the ratio of the turbu-
lent energy carried by the sub grid scales to the turbulent
kinetic energy of the resolved scales. This ratio can be esti-
mated by the formula in Eq. (15), where the u
0
i
are the time
averaged values of the resolved velocity uctuations:
k
sgs
k
res

l
sgs
=L
s
qC
s

2
1
2
u
0
1
u
0
2
u
0
3

2
19
Fig. 3 shows contours of the ratio k
sgs
/k
res
for the coarse
grid, whilst Fig. 4 shows contours for k
sgs
/k
res
for the ne
Fig. 3. Contours of the ratio of SGS turbulent kinetic energy to the
resolved turbulent kinetic energy for the coarse grid.
Fig. 4. Contours of the ratio of SGS turbulent kinetic energy to the
resolved turbulent kinetic energy for the ne grid.
M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406 401
grid and Fig. 5 shows a plot of k
sgs
/k
res
for both grids at the
60 mm position. k
sgs
/k
res
for the coarse grid is $0.3 which is
too high which indicates that the coarse grid is too coarse
for an LES. k
sgs
/k
res
for the ne grid is more satisfactory
at $0.05, although further renement may be desirable to
reduce the ratio to below 0.01, however this would result
in a grid of 3 10
6
nodes if the entire grid was adapted.
It seems reasonable to propose that turbulent momen-
tum transfer in hydrocyclones in the region between the
air water free surface and the peak in tangential velocity
is being carried by turbulent scales which have a structure
which approaches the two component limit (where velocity
uctuations normal to the free surface are damped com-
pared to velocity uctuations in other directions) as the free
surface is approached. To properly resolve these highly
anisotropic eddies the LES grid needs to be ne in this
region at least in the radial direction, so further renement
in this region alone may be satisfactory.
3.2. Multiphase results
The multiphase simulations were run with a feed water
ow rate of 1.117 kg s
1
and a total feed limestone ow
rate of 0.0574 kg s
1
, (4.88% limestone by weight) which
SGS k to Resolved k - 60 mm
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
Radial position - m
Fine
Coarse
Fig. 5. Ratio of SGS turbulent kinetic energy to the resolved turbulent kinetic energy for ne and coarse grids at 60 mm below the top of the cyclone.
35 m Limestone flow - LES
0.000
0.001
0.002
0.003
0.004
0.005
0.006
0.007
0.008
0.009
0.010
0.011
0.012
9.9 10 10.1 10.2 10.3 10.4 10.5 10.6 10.7 10.8 10.9
Time - s
k
g
/
s
Overflow
Underflow
Fig. 6. Instantaneous mass ow rate of 35 lm limestone size out overow and underow for coarse grid multiphase LES. Feed mass ow rate of 35 lm
limestone = 6.84 10
3
kg s
1
. 3rd Order muscle on momentum equation.
402 M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406
matches Hsiehs (1988) series VII data. Coarse grid simula-
tions were run with bounded central dierencing and 3rd
order muscle on the momentum whilst ne grid simulations
were run with bounded central dierencing. The 3rd order
muscle scheme was found to give more accurate predictions
of limestone classication behavior on the coarse grid but
there seemed to be only small dierences between the two
discretization schemes on the ne grid and only the results
for bounded central dierencing on the ne grid are
reported here.
Fig. 6 shows a plot over 1 s starting at 9.9 s of the instan-
taneous overow and underow 35 lm limestone ow
rates, whilst Fig. 7 shows the instantaneous water ow
rates over the same period for the coarse grid simulations.
These plots show a number of interesting features of the
simulation. Firstly the simulation is predicting a signicant
classication of the 35 lm size to the underow (The pre-
dicted tromp or cyclone eciency curves are shown on
Fig. 8 and are discussed further below). It is evident that
the LES is predicting a relatively steady water ow rate
Water flow - LES
0.00
0.20
0.40
0.60
0.80
1.00
1.20
Time - s
k
g
/
s
Overflow
Underflow
9.9 10 10.1 10.2 10.3 10.4 10.5 10.6 10.7 10.8 10.9
Fig. 7. Instantaneous mass ow rate of water out overow and underow for coarse grid multiphase LES. Feed mass ow rate of water = 1.117 kg s
1
.
3rd Order muscle on momentum equation.
Cyclone Efficiency
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30 35 40
Particle size - m
F
r
a
c
t
i
o
n

t
o

u
n
d
e
r

f
l
o
w
Hsieh Series VII
Coarse OF BCD
Coarse UF BCD
Coarse OF 3rd O
Coarse UF 3rd O
Fig. 8. Predicted cyclone eciency curve. Coarse grid multiphase LES. Feed mass ow rate of water = 1.117 kg s
1
. Feed Limestone mass ow
rate = 0.0574 kg s
1
. BCD Bounded central dierencing on momentum. 3rd O 3rd Order muscle on momentum. UF Fraction to underow from
underow ow rate. OF Fraction to underow by dierence from overow ow rate.
M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406 403
but there is a fairly large uctuation in the instantaneous
ow rates of the 35 lm limestone.
Figs. 8 and 9 shows the predicted cyclone eciency
curve, where the fraction of each limestone size reporting
to underow as a function of particle size is plotted.
Fig. 8 shows curves from the coarse grid with bounded cen-
tral dierencing and 3rd order muscle on momentum.
Fig. 9 shows the eciency curve for the ne (adapted grid)
with bounded central dierencing on momentum. The frac-
tion of each size to underow was calculated by taking the
time average of the instantaneous ow rates out of the
cyclone over a 1 s time interval. The open points were cal-
culated from the underow ow rates and the closed points
were calculated by dierence from the overow ow rates.
The fraction of each size to underow as calculated
directly from the underow rate and as calculated by dier-
ence from the overow rate are similar which indicates that
the simulations are at steady ow and the overall material
balance is being closed. The cyclone eciency curves on all
simulations are being predicted with quite good accuracy,
except that the simulations using bounded central dier-
encing on the momentum equation predict more short cir-
cuiting of the 25 and 35 lm size fractions to the overow.
The ner grid seems to improve the predictions at the smal-
ler size range, but has a small eect on the short circuiting,
however short circuiting has been reduced by using the 3rd
order muscle scheme on the momentum equation on the
coarse grid.
Fig. 10 shows the contours the instantaneous volume
fraction for each size fraction of limestone inside the
cyclone in elevation for the ne grid simulation with
bounded central dierencing on momentum. Note that
the contours on these graphs are plotted on a logarithmic
scale with an upper bound of 0.1 on volume fraction.
The central region on each plot, which is bright red, is
the air core. These plots show three distinct characteristics
in the distribution of limestone inside the cyclone as the
particle size increases; (i) The concentration increases at
the wall in the bottom of the apex as the particle size
increases. (ii) There is a sharp drop in the concentration
and a region of very low concentration forms around the
air core and this region become larger as the particle size
increases. (iii) A region of low concentration forms around
the vortex nder and this region of low concentration
becomes larger as the particle size increases.
Fig. 10 shows that there is distinct short circuiting of the
larger sizes to the overow, but contrary to accepted wis-
dom, the short circuiting is not due the larger particles
migrating down the outer wall of the vortex nder. Rather
for the particle sizes ranges being simulated, the larger par-
ticles simply have not had sucient time to classify in the
upper vortex nder region before they encounter the short
circuiting ow. There are also fairly distinct local varia-
tions in the concentration of the larger particle sizes which
is indicative that turbulent mixing of the particles is being
resolved as part of the simulation and this must have an
eect on the short circuiting ow as well.
Fig. 11 shows a comparison between the coarse grid
with bounded central dierencing on momentum, the
coarse grid with 3rd order muscle on momentum and the
ne grid with bounded central dierencing on momentum.
It is evident that the higher order discretization reduces
short circuiting. The ne grid appears to reduce short cir-
cuiting but this is not reected in an improvement in the
eciency. This seems to be because the smaller width of
the short circuiting ow that occurs with the ner grid is
oset by a larger concentration of limestone.
It is dicult to identify the cause of the discrepancy
between modeled and measured eciency at 35 lm, how-
ever the dierence may be due to dierences between the
Cyclone Efficiency
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30 35 40
Particle size - m
F
r
a
c
t
i
o
n

t
o

u
n
d
e
r
f
l
o
w
Hsieh Series VII
Fine OF BCD
Fine UF BCD
Fig. 9. Predicted cyclone eciency curve. Fine grid multiphase LES. Feed mass ow rate of water = 1.117 kg s
1
. Feed Limestone mass ow
rate = 0.0574 kg s
1
. BCD Bounded central dierencing on momentum. 3rd O 3rd Order muscle on momentum. UF Fraction to underow from
underow ow rate. OF Fraction to underow by dierence from overow ow rate.
404 M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406
modeled geometry and the real geometry or possibly
because the simulations modeled only six size ranges in
the feed size distribution. Alternatively there may have
some measurement error.
4. Conclusions
CFD simulations of the Hsieh cyclone (1988) using large
eddy simulation and the mixture model with the granular
options in Fluent
TM
have been used to predict classication
and the distribution of limestone inside the cyclone. The
cyclone eciency curve is predicted with good accuracy
even on coarse grids, but rening the grid and using a
higher order discretization scheme improves the predic-
tions by reducing the simulated short circuiting of the lar-
ger sized limestone fractions.
The simulated distribution of the limestone particles
suggests that short circuiting of the 35 lm occurs not by
migration down the outer wall of the vortex nder but
rather because this size has not had sucient time to clas-
sify before some portion is entrained in the up ow through
the vortex nder.
Fig. 10. Contours of instantaneous limestone (plus air) volume fraction inside cyclone in xz plane. The central region in red is the air core. Fine grid
simulation, large eddy simulation with bounded central dierencing on momentum. Note that contours are plotted on a log scale. (For interpretation of
the references in colour in this gure legend, the reader is referred to the web version of this article.)
Fig. 11. Instantaneous distribution of 35 lm limestone volume fraction inside cyclone in xz plane. BCD Bounded central dierencing, 3O 3rd Order
muscle.
M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406 405
References
Baxter, L.L., Smith, P.J., 1993. Turbulent dispersion of particles. Energy
Fuels, 852859.
Boivin, M., Simonin, O., Squires, K.D., 2000. On the prediction of gas
solid ows with two-way coupling using large eddy simulation. Phys.
Fluids 12, 20802090.
Brennan, M.S., 2003. Multiphase CFD simulations of dense medium and
classifying hydrocyclones. In: Third International Conference on CFD
in the Minerals and Process Industries, CSIRO, Melbourne, Australia,
1012 December, 5963.
Brennan, M.S., 2006. CFD simulations of hydrocyclones with an air core
comparison between large eddy simulations and a second moment
closure. Trans IChemE (Chemical Engineering Research and Design)
84A, 495505.
Cullivan, J.C., Williams, R.A., Cross, C.R., 2003. Understanding the
hydrocyclone separator through computational uid dynamics.
Transac. ChemE. 81A, 455466.
Delgadillo, J.A., Rajamani, R.K., 2005. A comparative study of three
turbulence-closure models for the hydrocyclone problem. Int. J.
Mineral Process. 77, 217230.
Devulapalli, B., Rajamani, R.K., 1994. Application of LDV to the
modelling of particle size classication in industrial hydrocyclones.
ASME FED 191, 4148.
Devulapalli, B., Rajamani, R.K., 1996. A comprehensive CFD model for
particle size classication in industrial hydrocyclones. Hydrocyclones
96, Cambridge UK, 83104.
Devulapalli, B., 1997. Hydrodynamic modelling of solidliquid ows in
large scale hydrocyclones, Ph.D. thesis, University of Utah.
Ding, J., Gidaspow, D., 1990. A bubbling uidization model using kinetic
theory of granular ow. AIChE J. 36, 523528.
Gidaspow, D., Bezburuah, R., Ding, J. 1992. Hydrodynamics of circu-
lating uidized beds, kinetic theory approach. In: Fluidization VII,
Proceedings of the 7th Engineering Foundation Conference on
Fluidization, 7582.
Hsieh, K.T., 1988. A phenomenological model of the hydrocyclone,
Ph.D. thesis, University of Utah.
Ishii, M., Mishima, K., 1984. 2-Fluid model and hydrodynamic consti-
tutive relations. Nucl. Eng. Des. 82, 107126.
Manninen, M., Taivassalo V., Kallio, S., 1996. On the Mixture model for
multiphase ow, Valtion Teknillinen Tutkimuskeskus, Espoo,
Finland.
Millelli, M., Smith, B.L., Lakehal, D., 2001. Large-eddy simulation of
turbulent shear ows laden with bubbles. In: Guerts, B.J. (Ed.), Direct
and Large Eddy Simulation IV. Kluwer Academic Publishing,
Dordrecht, pp. 61470.
Narasimha, M., Brennan, M.S., Holtham, P.N., 2006. Numerical Simu-
lation of magnetite segregation in a dense medium cyclone. Minerals
Engineering 19, 10341047.
Portela, L.M., Oliemans, R.V.A., 2001. Direct and large eddy simulation
of particle-laden ows using the point particle approach. In: Guerts,
B.J. (Ed.), Direct and Large Eddy Simulation IV. Kluwer Academic
Publishing, Dordrecht, pp. 53460.
Saman, P.G., 1965. The lift on a small sphere in a slow shear ow. J.
Fluid Mech. 22 (2), 385400. Also Corrigendum (1968), Journal of
Fluid Mechanics, 31, 624.
Slack, M.D., Prasad, R.O., Bakker, A., Boysan, F., 2000. Advances in
cyclone modeling using unstructured grids. Trans. IChemE. 78 (A),
10981104.
Slack, M.D., Wraith, A.E., 1997. Modelling the velocity distribution in a
hydrocyclone. In: 4th International Colloquium on Process Simula-
tion, 1113 June, Espoo Finland, 6583.
Shalaby, H., Pachler, K., Wozniak K., Wozniak, G., 2005. Comparative
study of the continuous phase ow in a cyclone separator using
dierent turbulence models. In: International Journal for Numerical
Methods In Fluids 48, 11751197.
Suasnabar, D.J., 2000. Dense Medium Cyclone Performance, Enhance-
ments via computational modeling of the physical process, Ph.D.
thesis, University of New South Wales.
Syamlal, M., Rogers, W., OBrien, T.J., 1993. MFIX Documentation:
Volume 1, Theory Guide. National Technical Information Service,
Springeld, VA, DOE/METC-9411004, NTIS/DE9400087.
406 M.S. Brennan et al. / Minerals Engineering 20 (2007) 395406

You might also like