You are on page 1of 22

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDS

Int. J. Numer. Meth. Fluids 2013; 72:528–549


Published online 16 Nov 2012 in Wiley Online Library (wileyonlinelibrary.com/journal/nmf). DOI: 10.1002/fld.3750

Dust modelling using a combined CFD and discrete


element formulation

J. E. Hilton* ,† and P. W. Cleary

CSIRO Mathematics, Informatics and Statistics, Clayton South, VIC 3169, Australia

SUMMARY
The production and dispersal of airborne dust is an important issue in both environmental and industrial
contexts. Dust pollution is a major environmental concern, and long exposure in occupational settings has
been linked with numerous respiratory health issues. Industrial dust pollution can also present a significant
explosion hazard, even in facilities with dust extraction systems. Computational models for dust generation
and dispersal have, however, generally been formulated for specific geophysical applications and restricted
to static, two-dimensional, approaches. Here, we present a method for simulating dust production from
a dynamic granular bed by using a three-dimensional coupled discrete element method and Navier–Stokes
computational model. Dust production is based on an energy formulation in which micro-scale dust particles
are assumed to overcome cohesion to macro-scale grains. This model is used over the entire range of energies
present within the system, from macro-scale collisions to aerodynamic entrainment and bombardment of
micro-scale particles. The dust concentration is modelled as a scalar density field, which is advected and
diffused through turbulence in the gas flow field. The model is tested against empirical predictions for two
test cases, a slug of granular material dropped from a set height and air flow over a granular stockpile.
Both give good agreement to the empirical relations, showing that the model can accurately predict the
production and subsequent dispersal of dust from a dynamic granular bed. Copyright © 2012 John Wiley &
Sons, Ltd.

Received 1 July 2012; Revised 27 September 2012; Accepted 11 October 2012

KEY WORDS: gas-particle flow; dust; discrete element method

1. INTRODUCTION

The dynamics of dust entrained in air is a familiar and fascinating phenomenon. Dust consists
of very small suspended particles, typically with a diameter lower than 50 m. At this size, gas
drag forces balance gravitational forces, giving the particles very long settling times. It should
be noted that the actual particle diameter does not fully describe the state of the particle when
airborne, so a more stringent limit is the particle aerodynamic diameter, defined as the diameter of
a hypothetical sphere of density 1000 kg=m3 having the same terminal settling velocity in calm
air as the particle in question, regardless of its geometric size, shape and true density [1]. At this
upper limit, particles have a reasonably short settling time: a particle with aerodynamic diameter
of 50 m has a terminal velocity of 0.07 m/s. Dust contains particles down to micron sizes,
however, where the terminal velocity is close to zero and the particles can be regarded as staying
practically indefinitely airborne.
Naturally occurring dusts include organic and mineral particles such as pollen, volcanic ash and
fine sands picked up and entrained during sandstorms. Industrial processing typically produces

*Correspondence to: J. E. Hilton, CSIRO Mathematics, Informatics and Statistics, Clayton South, VIC 3169, Australia.
† E-mail: james.hilton@csiro.au

Copyright © 2012 John Wiley & Sons, Ltd.


DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 529

mineral dusts, such as coal and cement, metallic dusts, organic dusts, such as flour and cotton, and
biological dust, such as mould and spores. Regardless of the source, any of these dust types can be
a significant contaminant and hazard to health. Long term exposure can result in occupational lung
diseases as well as systemic intoxication, and even low exposure levels have been linked to a wide
range of respiratory diseases. Fibrous dusts, such as asbestos and carbon fibre, are renowned for
their detrimental effect on heath. Dusts containing free crystalline silica cause the fatal, eponymous,
disease silicosis. Other types of dust, such as hard metal dusts, directly cause health problems
through over-exposure, which may lead to diffuse pulmonary fibrosis. Substances that have
previously been classed as innocuous are now thought to potentially accumulate in the lungs into
masses too large to be cleared. This process, known as dust overload has been suggested as a
precursor to tumour formation [2], leading to revised standards for these substances [1]. As such,
dust is a significant concern to operators, particularly in mines, plants processing or handling
granular materials and at transport hubs where dust producing material is frequently loaded or
unloaded. Dust pollution can lead to revocation of the licence to operate, regulatory difficulties
in obtaining plant expansion approvals and heavy fines.
Dust also presents a significant hazard for explosive combustion as many fine particulates,
especially organic and metallic powders, readily combust when suspended in air. Such dust
explosions are, unfortunately, relatively common and have caused many fatalities in industrial plants.
As an example of the frequency of such events, Abbasi [3] stated that 13 major explosions took
place in the USA in 2005 alone, causing two fatalities and damage to property of around $56m.
Single events can be devastating, such as an explosion caused by textile dust in China in 1987,
which claimed 58 lives. The severity of dust explosions can often be attributed to a domino effect
where an initial dust explosion mobilizes an accumulated bulk mass of dust into the air. These
suspended dust clouds can then be re-ignited creating secondary explosions, which may in turn
cause more explosions through the same effect. This is especially extreme in closed pipe networks,
as turbulence generated by the explosion increases the rate of dust production. This leads to a
strong positive feedback of the explosive pressure as it propagates through the pipe network. Such
explosions can occur even in plants with dust extraction systems. Subsequent investigation after
explosions revealed dead zones where dust could accumulate away from, or in some cases even
within, the extraction system [4]. These accumulations can turn a relatively minor primary explosion
into a series of devastating secondary explosions. Such events highlight the need for accurate
modelling of dust flow, so dead zones where dust can settle and cause a potential hazard can clearly
be identified.
The ability to accurately model dust is therefore very important for investigations of potential
contamination and hazards in both environmental and industrial scenarios. However, although
several experimental geophysical studies have been carried out, there is a surprisingly scarce amount
of theoretical work in this area. Most models to date have been formulated from a geophysical
standpoint to simulate aeolian processes such as sediment transport and soil erosion, leading to
dust formation [5]. Models directly predicting dust concentration rates include Harris et al. [6, 7],
who used a Monte–Carlo technique to generate concentration profiles. The mean wind profile
was, however, considered fixed in their approach, and the simple geometry made it unsuitable for
industrial modelling applications. Li and Guo [8] gave a two-phase continuum formulation for a
solid granular material phase coupled to a gas phase. Their model was restricted to only two-
dimensional dust formation over a flat horizontal bed. Parameters in the model were also chosen
to closely match the experimental system against which it was validated. As their method used
a continuum model for the granular phase, the grain size was assumed to follow an empirical
Rosin–Rammler distribution. Although the model showed a good match to experimental data, the
two-dimensional nature, restricted geometry and high level of parameterization make it unsuitable
for general modelling purposes.
Within this study, we present a detailed dust generation and transport model motivated by
experimental studies. The model is based on an energy formulation that includes both of the main
methods of dust generation: liberation during particle collisions and pick-up by an incident air flow.
Our results are compared with empirical expressions developed by the Verein Deutscher Ingenieure
(VDI) [9] and show good agreement over a range of possible dust production scenarios.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
530 J. E. HILTON AND P. W. CLEARY

2. DUST GENERATION AND EMISSION

2.1. Mechanism of dust generation


One of the most familiar observable types of dust production occurs when wind blows over a
horizontal granular bed, creating a plume of dust that is picked up and dispersed in the wind. The
mechanics of this phenomena were first investigated by Bagnold in his seminal work on wind-driven
granular motion [10]. Bagnold divided the motion of particles transported by a gas into one of the
three transport modes depending on their size, shown schematically in Figure 1. These three modes
are suspended particles, saltating particles and creeping particles.
Suspended particles constitute dust, which are assumed to have an infinite settling time and stay
indefinitely airborne once liberated from the bed. Saltating grains belong to the next size category.
These particles are small enough to be picked up by turbulent gas eddies at the surface of the bed
but too large to remain in suspension indefinitely. The motion of these grains therefore consists of
a series of parabolic jumps across the surface of the particle bed. When saltating particles fall back
under gravity and impact the bed, they dislodge other particles from the surface in a process known
as splash. The steady net impact of these saltating particles is known as saltation bombardment
[11]. The last transport mode consists of creeping particles, which are too heavy to lift through
aerodynamic forces and roll along the surface of the bed. Creeping particles have no effect on the
saltating or suspended layers.
The ability of particles to undergo saltation depends on the ratio of the weight of a grain to the
shear stress imparted by the gas at the granular surface, 0 . The relevant dimensionless expression
is 0 =p gds , where p is the particle density and ds the saltating particle diameter. Owen used
pthis dimensionless quantity as 0.0064 [12]. The shear,
Bagnold’s findings to give the critical value of
or friction, gas velocity is defined as u D =g , where g is the gas density. These expressions
give a minimum friction velocity for saltation, u0 , as follows:
s
p
u0 D 0.0064 gds (1)
g

The mass flux of particles in the horizontal direction is given by q (with units kg m2 s1 ), and
the vertically integrated streamwise mass flux, G, (with units kg m1 s1 ) is defined as
Z 1
GD qdy (2)
0

Through analysis G is found to scale with the friction velocity as G / u3 [10, 12]. This relation
has been verified by Bagnold [10], Owen [12] and Shao et al. [13], amongst others. Shao et al.

Figure 1. The three transport modes of particles in a gas flow: suspension, saltation and creep.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 531

found a total mass flux G from a fit to their data sets for saltating grains with diameters from 210
to 530 m, where ur , the mean horizontal wind speed, ur / u :

G D 103 e 3.23 jur j2.926 (3)

Shao et al. also considered the role of saltation in dust production. From a series of careful
experiments, they found that dust is only produced by saltation bombardment because the cohesive
forces dominate over lift forces at very small particle sizes and these cohesive forces are easily
broken by the larger impacting grains.
Their experiments consisted of granular beds in a long wind tunnel, where the granular materials
used were red sand with a particle diameters varying from 100 to 1000 m and loose kaolin
clay, with an average particle diameter of 2 m. The main experiments consisted of two different
configurations of these granular materials. The first was a layer of purely kaolin clay 3 m long
deposited 3 m from the dust collectors. The second was a configuration in which a 1 m long
layer of red sand was deposited upwind of a 2 m long layer of kaolin clay. The wind tunnel was
run at four different wind speeds ranging from 8.3 to 12.9 m/s. Saltation was found to occur at
u0  0.2 m/s for beds with particle diameters between 100 and 210 m and u0  0.4 m/s for beds
with diameters between 530 and 1000 m. This agrees reasonably well with the expression given
by Owen, Equation (1), which gives u0 D 0.15 and u0 D 0.33 m/s for the respective size classes,
using the mean values of these particle diameter ranges, 155 and 765 m.
In the first configuration, a rapid increase in dust production was found at the start of the
experiment, followed by an exponential decrease. The dust production became negligible after
around 200 s from the start of the experiment for all wind speeds. In contrast, the second
configuration gave sustained dust production over the duration of the experiment many times
higher than the initial rates found in the first configuration. For the two lowest wind speeds used,
ur D 8.3 and ur D 9.8 m/s, the dust flux was constant over the entire 9 min of experimental time.
For the higher wind speeds, the fluxes were constant until a rapid fall-off occurred at the end of
the experiment.
These findings were interpreted as showing that bombardment from the saltation layer was
responsible for the dust production, which could not be sustained from aerodynamic forces alone.
The initial dust flux in the first configuration was caused solely from a very small percentage of
very loose particles removed from the top of the deposited bed. The remainder of the kaolin clay,
below this loose layer, remained attached to the coarser particles in the bed as strong inter-particle
cohesive forces dominated over the aerodynamic forces acting on the particles.
However, when a layer of larger red sand particles was placed in front of the kaolin layer, the
saltation of the sand bombarding the kaolin gave the sustained dust production observed. The
bombardment from the sand provided enough energy to break the inter-particle cohesive forces
and liberate the kaolin particles from the bed. The rapid drop-off in the dust flux at the end of the
experiment resulted from the supply of sand being used up, confirming that bombardment was the
primary source of dust production. Prior to this, the dust flux was constant and depended only on
the relative wind speed.
On the basis of the experiment in the second configuration, Shao et al. found that the vertically
integrated streamwise dust mass flux dust was given by Gd D 103 e 3.88 jur j2.832 . Their expressions
for the saltating flux, dust flux and total flux, Equation (3), all had exponents very close to 3,
confirming Bagnold’s findings. However, their crucial finding was that the energy provided from
saltation bombardment, rather than aerodynamic lift, was responsible for creating and maintaining
the suspended dust layer.

2.2. Cohesive energy model of dust emission


Fine particles, constituting dust, are dominated by cohesive forces attaching them to the surface
of larger particles within the granular bed. If energy is imparted to these larger particles, the
dust particles can overcome their cohesive potential energy, , and escape the surface of the

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
532 J. E. HILTON AND P. W. CLEARY

larger particle. If the energy imparted is given by E, the number of dust particles liberated, N ,
can be modelled using a straightforward linear relationship [13]:
cE E
ND (4)

where cE is a dimensionless constant. To derive an expression for the saltation flux, Shao et al. used
a formula given by Owen [12] for the flux density of saltating particles, n (with units m2 s1 ), at
the surface of the bed:
 
u2
g u2 1  ut2

nD (5)
ms .juje  juji /
where ut is the bed shear velocity required for transition to saltation, ms is the mass of a saltating
grain, juje is the saltating grain ejection velocity and juji the impact velocity. This gives
 
g cG u2
G D u3 1  t (6)
g u2
where cG is a constant  1. The vertical dust flux from the surface of the bed is then
 given as
F D md nN , where md is the mass of a dust grain. Taking E D m2s juj2e  juj2i , along with
Equations (4)–(6), gives
F md cE g juje C juji
D (7)
G 2cG u
The ratio of mean relative wind velocity to friction velocity is given by Owen et al. as ur =u 
10 [12]. Using this approximation, along with cG  1, simplifies Equation (7) to
F
' 10Kg (8)
G
where K is the product of unknown quantities K D md cE = . If the binding energy is assumed to
follow a van der Waals form,  rd rp =.rd C rp /, where rd is the dust particle radius and rp is the
radius of the much larger particle to which it is cohered. As rp  rd ,  rd , and the binding energy
is independent of the larger particle radius. Hence, K can be regarded as a constant for small dust
particles, given the assumptions of Equation (4).

2.3. Extension of cohesive energy model


The relation Equation (4) gives the number of dust particles, N , and hence flux of dust, F , produced
from a grain when imparted with an energy E. This relation is independent of grain diameter, so the
expression can be used for both the energy imparted by saltation bombardment, as well as energy
from inter-particle collisions within the granular bed. The total energy available for dust production,
E, is therefore divided into resolved, Er , and unresolved energies, Eu , such that E D Er C Eu .
The resolved energy components are directly measured from inter-particle collisions in the granular
model. The unresolved energies are considered to be produced by saltation bombardment, which is
not directly simulated.
The resolved dust flux Fr emitted from an area of dimension h  h m2 in t s is given by
md N
Fr D (9)
h2 t
Equation (4) can be substituted for N . However, there are generally multiple collisions for each
resolved particle per time step. If each collision is given index i, the resolved dust flux Fr is given by
md cE 1 X Ei
Fr D (10)
h2 t
i

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 533

which gives the final expression as


K X
Fr D Pi (11)
h2
i

where Pi is the power dissipated in collision i during the period t .


The unresolved flux, Fu is given by Equation (3) with Equation (8):
Fu D 102 Kge 3.23 u3r (12)
where ur is the relative gas-particle velocity. The exponent has been rounded up from 2.926 in
Equation (3) to 3, for both simplicity and to agree with Bagnold’s original finding. Despite reporting
a transition velocity in their experiments, the expression given by Shao et al., Equation (3), contains
no transitional velocity dependence. It is assumed that this expression only holds above the transi-
tion velocity, Equation (1), where the friction velocity is converted to a relative gas-particle velocity
by using ur =u  10. This gives
´ q

102 Kge 3.23 u3r ur > 10 0.0064 pg gds
Fu D (13)
0 otherwise
The constant K can be calculated using an empirical expression given by Marticorena and
Bergametti [14]:
F
D 100.1344 (14)
G
where  is the clay percentage of the material by mass. The relation, Equation (14), was derived
from a data fit by Marticorena and Bergametti motivated by the observation that environmental
dust samples are mainly composed of clay mineral aluminosilicate compounds. This relation can
be used in Equation (8) to determine K for use in Equations (13) and (11), where the only free
parameter is .

3. COMPUTATIONAL MODEL FORMULATION

The following sections detail the computational methodology behind our model. An incompressible
Navier–Stokes CFD solver is used for the gas phase, coupled to a Lagrangian soft-sphere discrete
element model for the particle phase. A LES approach is used for modelling the turbulent component
of the gas flow, as well as to calculate the turbulent diffusion of the dust. The coupling between these
methods relies on the drag relations for solid particles in a gas flow.
Dust, once generated, is assumed to have a settling time much greater than the simulation length
and is considered suspended for the duration of the simulation. The dust is modelled as an Eulerian
density field that is advected and diffused by the gas field. Saltating particles are not directly
simulated and appear via the unresolved dust flux contribution Equation (13).

3.1. Particle model


The discrete element method (DEM) predicts the behaviour of a granular system by individually
time integrating the dynamics of every grain within the system [15]. The force balance for each
particle i is given by
dvi
mi D Fci C Fdi C Fpi C mi g (15)
dt
where mi is the mass of particle and vi is the linear particle velocity. The forces on each particle
consist of contact forces with neighbouring particles, Fci , a gas-particle drag force, Fdi , a pressure
gradient body force, Fpi D rp and the gravitational force mi g. The particles are considered to be
bluff, such that gas-particle viscous forces are neglected.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
534 J. E. HILTON AND P. W. CLEARY

The contact forces on each particle, Fci , are the vector sum of the normal contact forces, Fn , and
the tangential contact forces, Ft , Fci D Fn C Ft . The forces Fn and Ft act along the normal at the
contact point and in the collision plane defined by the contact point and normal, respectively. The
magnitude of the normal force is given by

Fn D kn ıl C Cn vn (16)

where kn is a spring stiffness, chosen to ensure the overlap, ıl, between the two particles in contact
does not exceed, on average,  0.1% of the particle diameter. The parameter Cn is the normal
damping coefficient, calculated to give the required coefficient of restitution, and vn the relative
normal speed. The tangential force is incrementally calculated by
h X i
Ft D min Fn , kt vt t C Ct vt (17)

where  is the coefficient of friction, kt the tangential spring stiffness, vt the relative tangential
surface velocity, t the incremental time step and Ct is a tangential damping coefficient. The
tangential spring stiffness, kt , is set to be 0.5kn . The sum of the incremental tangential spring forces
and the damping force is limited by the Coloumb sliding limit, Fn .
The moment balance on each particle is given by
d!i
Ii D Ti C Tdi (18)
dt
where Ii is the moment of inertia in the principal frame of the particle, Ti is the inter-particle
contact torque, Tdi is the Stokesian rotational drag on the particle and !i is the spin of the particle.
The Stokesian rotation drag term, TD , incorporates the rotational drag on a spinning particle from
the surrounding gas into the model. This is given by [16]

TD D 8 r 3 !r (19)

where the relative particle spin is given by !r D 12 !g  !p , with !g the curl of the gas velocity field
and !p the spin of the particle.

3.2. Gas phase model


The constitutive equations for flow through a porous bed are derived by Anderson and Jackson [17]
and given by Kafui et al. [18] as
@.
g /
C r  .
g u/ D 0 (20)
@t

@.
g u/
C r  .
g uu/ D 
rp C r  .
/ C fd C
g g (21)
@t
where
is the local bed porosity around the particle, g the gas phase density, u the interstitial
velocity field, p the pressure, fd the body drag force from the gas phase, g the gravity vector and 
the local stress tensor. The stress tensor is given by  D D, where is the gas viscosity and D is
the rate of strain tensor:
1 
DD ru C ruT (22)
2
Two velocities can be defined for gas flow through particle beds; the interstitial velocity, u, which
is the gas velocity through the pores of the bed, and the superficial velocity, u0 , which is the average
velocity over the bed. The two velocities are related by u0 D
u. For a particle with velocity v in
a gas with interstitial velocity u, the relative interstitial velocity is ur D u  v, and the relative
superficial velocity is therefore u0r D
.u  v/.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 535

We re-formulate Equations (20) and (21) to solve for the superficial gas velocity, rather than
the interstitial gas velocity. Substituting u0 D
u into Equations (20) and (22) and making the
assumption that the gas density, g , is constant in the system gives
@

C r  u0 D 0 (23)
@t
 
@u0 1 0 0 1  
Cr  uu D 
rp C r  Œ
 0 C fd C
g (24)
@t
g
where the stress tensor  0 D D0 , with
!
0 1 u0 u0 T
D D r Cr (25)
2

Expanding the terms in Equation (24) gives


@u0 1 u0 1
C .u0  r/u0 C u0 r  D .
rp C fd C r  Œ
 / C
g (26)
@t

g
which contains the usual terms  in the Navier–Stokes equations plus a term involving the divergence
of the interstitial velocity, r  1 u0 . A modified pressure correction method can be used to solve the
constitutive equations, Equations (23) and (26). Time discretization of Equation (26) gives
 nC1

u0 nC1 D u0 n C t  rp nC1 C Sn (27)


g
where n is the current time step at time t and n C 1 is the next time step at t C t . The source term,
Sn , contains the advection, divergence, particle-gas interaction, stress and gravitational terms:
1 u0 n 1  n

Sn D  .u0 n  r/u0 n  u0 n r  C fd C r 
nC1  n C
nC1 g (28)

nC1
nC1 g
Introducing the intermediate velocity field u0
n as
 nC1
0 0n
n n
u D u C t  rp C S (29)
g
where we have used the pressure gradient from the last time step, rp n . Subtracting Equation (29)
from Equation (27) gives
 nC1
0 nC1 0

u D u C t  rıp (30)
g
where ıp is a pressure correction given by ıp D p nC1  p n . Taking the divergence of both sides of
Equation (30) allows us to formulate a Poisson equation for this pressure correction:

g

r 
nC1 rıp D r  u0  r  u0 nC1 (31)
t
From Equation (23), we can then express the divergence of the new velocity field, to first order,
as
@
1
nC1
r  u0 nC1 D  

n (32)
@t t
giving the final Poisson expression as


nC1 g 0 1  nC1 n

r
rıp D r u C

(33)
t t

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
536 J. E. HILTON AND P. W. CLEARY

3.3. Coupling methodology


The primary component of any model in which particle and gas motion is coupled is the particle
drag term. This must be chosen and calculated with care as it determines the overall dynamics of the
system. The drag force exerted on a stationary isolated spherical particle, FD0 , subject to a relative
gas velocity u0r is defined by
1
FD0 D g ju0r j2 CD A? uO0r (34)
2
where CD is a dimensionless gas drag coefficient, A? is the particle cross-sectional area perpendic-
ular to the flow and u0r is a unit vector in the relative flow direction. For multi-particle systems, the
drag force on each particle is influenced by flow and pressure effects from the surrounding particles.
From an empirical fit to experimental data, Di Felice [19] proposed the following correction to the
drag force, FD , on a spherical particle closely surrounded by neighbouring particles:

FD D FD0
 (35)

where
is the voidage fraction around the particle and is a parameter given by

.1.5  log Rep /2
D 3.7  0.65 exp  (36)
2
and the particle Reynolds number for a particle of diameter d is Rep D g ju0r jd= . The drag force,
Equation (35) with (34), becomes
1
FD D g ju  vjCD A?
.2/ .u  v/ (37)
2
For relative motion in granular beds, the particle Reynolds number can be expressed in terms of
the superficial velocity as Rep D g
ju  vjd= .
The body force from the particle on the gas, fd , is defined as
Nc Nc
1 X 1X 1
fd D FDi D g ju  vi jCD A?
.2/ .u  vi / (38)
Vc Vc 2
i D1 i D1

where Nc is the number of particles in the characteristic gas volume Vc . In terms of the superficial
velocity, substitution of u D u0 =
at time n gives the particle-gas drag body force as
Nc ˇ ˇ  
1X 1 ˇˇ u0 n ˇ
ˇ
 nC1 2 u0 n
fdn D g  vi ˇ CD A?
 vi (39)
Vc 2 ˇ
nC1
nC1
i D1

This coupling methodology, along with the reformulation of the porous gas flow constitutive
relations, has previously been applied to a diverse range of gas-particle system. These include
fluidized beds [20], pneumatic conveying [21] and discharge of fine particles from a hopper [22].
Each of these studies has validated the model against existing empirical relations for each of
these systems.

3.4. Dust modelling and turbulent diffusion


The settling time of dust particles is assumed to be much greater than the simulation time, such that
dust remains indefinitely airborne once produced. Using this assumption allows the dust density to
be modelled as a continuum Eulerian field, d , which obeys the Reynolds transport theorem over a
control volume :
Z Z Z
@  0 
d d D  d u C J d C S d (40)
@t  @ 

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 537

where J is a diffusional flow term and S is a dust source term. This results in a continuity equation
of the form
@d
D u0  rd C rJ C S (41)
@t
where the assumption that the velocity field is incompressible outside the particle bed has been used.
The molecular diffusion of dust in air can be shown to be negligibly small using the Stokes–
Einstein relation. This gives the diffusion coefficient as D D .kB T /=.6 r/, where kB is the
Boltzmann constant and T is absolute temperature. For a dust particle with radius O.50 m/
at room temperature, this gives D  O.1013 /. However, dust can be observed to readily diffuse
in a flow field, despite the negligible molecular diffusion. This is due to turbulent mixing in the
gas phase, which gives an effective diffusion rate many orders of magnitude higher than molecular
diffusion. Correctly modelling turbulence, and the effective diffusion due to turbulence, is therefore
crucially important for accurately simulating the dust field.
Turbulence is incorporated in the model by using a LES approach. This applies a spatial filter
to the Navier–Stokes equations resulting in resolved and unresolved, or sub-grid scale, stress
components. The most straightforward approach to modelling these sub-grid scale stress compo-
nents is to assume that they follow a linear relationship identical to a Newtonian viscous relation.
The effect of the sub-grid stress terms can be therefore modelled by adding an extra viscous term due
to turbulence, t , such that D g C t , where g is the molecular gas viscosity. This eddy-viscosity
term is assumed to be proportional to the filter width, , as well as a function of the magnitude of
the rate of strain tensor. The Smagorinsky model [23], which is the first model suggested of this
type, uses
p
. t /S D .Cs /2 2D W D (42)
where Cs is the Smagorinsky constant  O.101 / and the colon operator represents the scalar
product of two tensors. More sophisticated models can dynamically determine the constant [24, 25].
Here, we use the wall-adapting local eddy-viscosity method (WALE), which gives proper cubic
near-wall scaling behaviour without the need for any additional procedures [26]. The WALE model
has been shown to give very good results, conforming well to DNS simulations, for a range of
standard test flow types including cylindrical pipe flow [26] and flow past a wall-mounted cube [27].
The WALE LES eddy-viscosity term is given by
D3d
. t /W D .Cw /2 5
(43)
D5 C Dd2
where the WALE turbulent coefficient, Cw , is set to 0.325, found to give the best result over a wide
range of flow types [26]. The filter width, , is set to the grid resolution, and the tensor Dd is the
traceless symmetric part of the square of the strain rate tensor:
0 1
 0 2 " 0 T #2 
1@ u u A 1 u0 2
Dd D r C r  ı r (44)
2

3

where ı is the identity tensor.


For the scalar transport term, Equation (40), assuming that molecular diffusion is negligible and
applying LES filtering give the diffusional flow term J D r    , where   is a sub-grid scale
diffusive flux vector. This can be expressed as a Smagorinsky-type expression [28]:
p
.  /S D .Cs /2 2D W D rd (45)
However, Equation (45) assumes that the diffusive flux vector aligns with the gradient of d ,
which may not generally be the case [28, 29]. A different approach is to directly expand the
sub-grid scale diffusive flux vector and discard high order terms. This leads to a model suggested
by Clarke et al. [30]:
.  /C D .Cc /2 ru0  rd (46)

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
538 J. E. HILTON AND P. W. CLEARY

where the constant Cc is set to 1=12 [28]. The Clarke model, while giving excellent theoretical
performance, can lead to un-physical negative diffusion and numerical instability [28]. A straightfor-
ward solution is to simply combine the Smagorinsky and Clarke models to give   ' .  /S C.  /C .
This seemingly ad hoc solution works surprisingly well, is comparatively straightforward to
implement and gives good results in cross-flow diffusion tests in comparison with other models [31].
We used this mixed approach in the model, which is implemented as
  D . t /W rd C .Cc /2 ru0  rd (47)
The turbulent viscosity . t /W is also incorporated into the Navier–Stokes stress tensor with
D g C . t /W .
The source term S is evaluated from the energy arguments given in the preceding section. The
total dust flux is F D Fu C Fr using Equations (13) and (11), respectively. These are calculated
on a per-particle basis and summed to give the total dust flux contribution. It is assumed that the
entire dust flux emitted by a particle over a time step remains within the computational CFD cell
enclosing the particle. This assumption is justified as the CFD method for the gas is limited by a
Courant condition, ensuring nothing can be advected further than one cell per time step. The flux
term, with units kg m2 s1 , can therefore be converted to a volume source term of the form @=@t
with units kg m3 s1 by dividing by the cell width, S D F= h. This gives the final form of the
source term for each particle as
1
Si D .Fr C Fu / (48)
h
where Fu and Fr are given by Equations (11) and (13), respectively. This contribution is summed
over all particles enclosed within each computational CFD grid cell. We also make the assumption
that any gas-particle relative velocity in the direction of gravity has no effect on unresolved
dust production, as only grains with a velocity component perpendicular to gravity are saltating
and contribute
p to dust generation. In our application, gravity is set in the y-direction, giving
ur D .ux  vx /2 C .u´  v´ /2 .

3.5. Computational methodology


Time integration of the coupled system is carried out in a discrete series of time steps. First, the
forces and torques on each DEM particle from contacts with neighbouring particles are calculated,
giving FcinC1 for each particle i at the next time step t nC1 . Next, the porosity of the granular bed,

nC1 , is determining for each CFD grid cell from the positions of the DEM particles. The dynamics
of the gas field are calculated from Equation (38) by using un , v n and
nC1 and Equation (28). The
pressure correction, Equation (33), is then applied and used to calculate the gas velocity field at the
next time step, unC1 , from Equation (30). Next, the forces from the gas on the DEM particles are
calculated from Equation (37) by using unC1 . These forces are added to the total force on each DEM
particle and time integrated using Equation (15) to give the particle positions and velocities, v nC1 .
The local porosity is calculated by determining the volume fraction of each particle within all grid
cells containing the particle. This is implemented by using the spherical cap formula for a sphere-
plane intersection to divide the particle volume between all incident cells. This volume fraction
calculation is also used to divide the gas body force created by each particle over the grid cells
containing the particle. The total gas body force within each cell is the sum of the force from each
particle multiplied by the fraction of the particle within that cell. We have found this approach to be
superior to a typical collocation approach, in which the gas body force from a particle is applied to
only the grid cell containing the particle centre. A collocation approach causes the gas body force
in a cell to instantaneously change as a particle crosses the boundary of the cell, which adds noise
to the resultant pressure field and can destabilize the numerical solution.
For the gas flow calculations, the constitutive relations are discretized onto a semi-staggered
Cartesian grid, which reduces the formation of pressure checker-board oscillations in the system.
Velocities and forces are defined on the cell vertices, and the porosity, pressure and local stress
tensors are defined at the cell centres. The stress tensor is calculated at each time step from the

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 539

Figure 2. Power and cumulative discrete element method (DEM) energy dissipation for a particle of 2 cm
diameter and p D 1000 kg/m3 dropped 1.0 m onto a flat plate. The instantaneous power over a DEM time
step is shown as a solid black line. The particle bounces several times, showing a series of sharp spikes in
the power due to each successive impact. The first three of these impacts are marked. The cumulative energy
dissipated in the DEM model, shown as the shaded grey background, rises in a sequence of steps. The total
energy dissipated matches the gravitational potential energy of the particle, shown as the dashed black line.

local porosity, molecular and turbulent viscosity and the gradient of the velocity. The advection
terms are calculated using upwind biassed stencils, and divergence and gradient terms are calcu-
lated using semi-staggered finite difference stencils. The non-linear Poisson term, Equation (33), is
calculated using an Symmetric Successive Overrelaxation (SSOR) pre-conditioned conjugate gra-
dient method, and CFD time steps are carried out using a second order Adams–Bashforth method.
From Equation (26), the porosity within each cell must be greater than zero, so the largest particle
must fit entirely within a cell. This restriction requires that the grid cell dimensions must therefore
be larger than the maximum particle diameter, which is common to all such unresolved gas-particle
coupling approaches.
The dust density is time integrated using Equation (41) with J D r    and   given by
Equation (47) and the source term given by Equation (48). The absorbed power is crucial to correct
simulation of dust production from the particle. This is taken to be the sum of the normal and
tangential components of the DEM power dissipation over a time step. The normal and tangential
power terms are calculated as Pn,t D Fn,t  vn,t , where Fn,t are the temporal mid-point averages
of the dissipative components of the normal and tangential force, respectively, and vn,t are the
corresponding velocity vectors.
To ensure these power measures fully capture the energy losses in our DEM algorithm, a simple
P out. Figure 2 shows the power per time step t against the
drop test with a one particle was carried
total cumulative dissipated energy Pt for a single particle of 2 cm diameter dropping from a
height of 1 m onto a flat plate. The particle initially impacts the plate at 0.4 s and subsequently
impacts at 0.8, 0.95, 1.25 s before coming to rest on the plate. The cumulative dissipated energy
by the time the particle comes to rest is identical to the initial gravitational potential energy of
the particle, showing that the measured power accurately gives the energy losses sustained by the
particle during the simulation. In the test cases given in this study, we assume that the particle
is not plastically deformed, heated, broken or produces any other lossy forms of energy (such as
acoustic energy), and all dissipated energy is available for dust production. A dissipation term in the
algorithm could, if required, easily be incorporated into the model.

4. COMPARISON WITH EXPERIMENT DUST MEASUREMENTS

Dust is primarily created by two methods. The first is active dust production during agitation of
the grains during dumping, dropping or conveying granular material. The second is passive aeolian
pick up from the surface of granular heaps, for instance, from storage stockpiles or during transport.
In the following sections, we compare the proposed numerical method to empirical rates derived
for these types of dust production. Typically, active production gives dust fluxes several orders of
magnitude higher than passive production, and these two tests ensure the model can correctly handle

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
540 J. E. HILTON AND P. W. CLEARY

the entire range of energy scales over which it should perform. Experimental studies concerning dust
production are, however, extremely scarce, although geophysical studies such as Shao et al. [13]
exist. There is also very little data for dust production in industrial contexts. Heitbrink et al. [32]
carried out extensive drop test experiments for freely falling powders, but unfortunately, these
experiments are not directly comparable with this model, as all particles in their system were
micron-size. In the following sections, we base our comparison on the VDI (German Association
of Engineers) standard 3790, part 3 VDI10, which provides empirical expressions for both active
and passive industrial dust generations [9].

4.1. Dust production during a vertical drop


The VDI standard contains a number of empirical relations determining active dust production for
a wide range of scenarios. Of these, the most easily replicable and straightforward to model is a
vertical drop of granular material onto a flat plate. For a load of granular material of the total mass
M kg with volume V m3 , dropped from a height H m, the VDI standard gives the amount of dust
produced, q kg, as
p M 1.5 H 1.25
q D 0.85 10k15 (49)
V
where k is an empirical parameter for the dustiness of the material with k D 2 imperceptible, k D 3
low, k D 4 medium and k D 5 high.
To test the model, a cylinder of diameter 0.3 m and height 0.5 m was filled with spherical particles.
The particles were then dropped vertically H m onto a flat plate, measured from the base of the
particle bed, and the total dust produced was recorded at the end of the simulation. This set-up
is shown schematically in Figure 3. No-slip boundary conditions were imposed on boundary of
the domain for the gas flow. Particles spread radially after impacting the base plate and no longer
collided with their neighbours. For simulation efficiency, particles were therefore removed when
they reached the outer boundary, as they did not contribute any further dust to the system. The
simulations were run for 2 s, which was long enough for all particles to hit the base plate and
radially disperse so they were no longer in contact. This was also, therefore, sufficiently long to
generate the maximum possible amount of dust from the system.
A number of simulations were run at varying particle density p for identically sized particles,
where p D 500, 750, 1000, 1250, 1500 kg/m3 . From Equation (49), the amount of dust produced
from each drop depends only on the total mass of particle and is independent of particle diameter.
To ensure the model agrees with this behaviour, three particle diameters, d , were used for each p ,
d D 2, 3, 4 cm. Finally, the grid resolution dependence on dust production was also tested by using

Figure 3. Set-up for particle drop test, consisting of a cylinder filled with granular material H m above a
flat impact plate.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 541

resolutions of h D 5 cm for d D 4 cm, h D 4 cm for d D 3 cm and h D 2.5 cm for d D 2 cm.


The numbers of particles in each simulation were 4617 for d D 2 cm, 1319 for d D 3 cm and 538
for d D 4 cm. Each simulation used the same spring constant, k D 1  105 N/m and coefficient
of friction  D 0.5. Dust production in simulations using coefficients of restitution e D 0.5 and
e D 0.75 was practically identical, so the dust production is insensitive to this material parameter.
The particles and dust density from a typical solution run are shown in Figure 4, for H D 0.5 m.
The majority of the dust production occurred upon impact with the plate at t  0.4 s. The dust
was entrained in a gas vortex created by the falling particles, forming a torus around the base of the
granular column after impact and spreading radially outwards from the impact point and upwards
along the closed vertical boundaries. Gas velocity streamlines plotted in Figure 5 at 0.3, 0.4 and 0.5 s
show the formation of this vortex as the particles impact the plate. The formation and subsequent
dispersal of this toroidal dust cloud was independent of the particle size and simulation resolution.
The dust produced by the falling granular column before impact was found to be negligible in
comparison with the dust released upon impact with the plate. The toroidal spread of dust upon
impact agrees well with the qualitative descriptions given by Heitbrink et al. [32] for drops of slugs
of powder. A detailed quantitative comparison cannot be made with their results, however, as the
granular material used in their experiments consisted of only micron-sized particles.
The total dust mass produced from each drop is plotted in Figure 6 as a function of the mass of
the material dropped. Exact comparison is difficult because of the very broad qualitative dustiness
definitions used in the VDI empirical formula, so a range of results is shown for different values
of the empirical dustiness parameter, k, and the model dustiness parameter . The results generally
agree with the broad VDI dustiness definitions, with the simulation results from median dustiness
parameter of  D 10% bracketed between the high and medium dustiness values from the empirical

Figure 4. Dust generated by a falling column of particles of 2 cm diameter and p D 1000 kg/m3 , dropped
0.5 m onto a flat plate. The dust is shown as a translucent coloured field, shaded by density, superimposed
over the particles.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
542 J. E. HILTON AND P. W. CLEARY

Figure 5. Gas flow streamlines within a cross-section of a falling column of particles. Air entrainment
during the fall causes the creation of a toroidal vortex around the column. Streamlines are shaded by
velocity magnitude.

expression. Unfortunately, a more precise comparison cannot be made without a clearer definition
of the dustiness parameters, k, which is unavailable in the VDI report.
Comparison of the upper and lower plots in Figure 6 shows that the simulation results scale with
drop height in the same manner as the empirical formula for both drop heights, H . The total mass of
dust produced in the simulations therefore agrees with the empirical height dependency. The results
for each value of the dustiness parameter, , show a linear relation between the total dust mass and
the drop mass. Varying the particle diameter in the simulations gives only a minor variation from
this trend. The total mass of dust produced for each value of the dustiness parameter  is therefore
independent of particle diameter, d in agreement with the dependency of the empirical expression
Equation (49). These results also show that the numerical method was independent of the resolution
of the underlying CFD grid, as the grid resolution was also varied at each particle diameter.
The predicted dust mass was found to have a near linear dependence on the dropped mass M .
This is a different power law scaling compared with the 1.5 exponent in the empirical expression,
Equation (49). Investigation showed that the difference in the experimental and simulation trends
with drop mass was not caused by unresolved dust produced from the relative vertical velocity
between the gas and particles. The dust produced in this manner was negligible in comparison with
the dust produced during impact. A test case in which the unresolved dust flux was artificially greatly
increased was also not found to reproduce the empirical trend. An explanation of this discrepancy
may be that the empirical formula was developed for drops of material containing a high proportion
of pre-existing free fines; the mass of which scaled with the mass of coarse material. These free
fines may have added a non-linear contribution in addition to the modelled liberation of fines
during impact. Unfortunately, the details of the experiments used to generate Equation (49) are
not stated, making it unclear how to account for such factors in the computational model. Despite
this difference, the proposed computational model shows the expected scaling with H is correctly
independent of the particle size of the material being dropped and gives total dust masses closely
comparable with the broad dustiness definitions used in the VDI formula.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 543

Figure 6. Total dust mass released against the total mass of granular material dropped for a falling column of
particles. The drop heights are H D 0.5 m (top) and H D 1.0 m (bottom). The Verein Deutscher Ingenieure
empirical expression, Equation (49), is plotted as dashed lines for two values of k, k D 4, medium dustiness,
and k D 5, high dustiness. Three particle diameters are shown, d D 2, 3, 4 cm, for three values of the
dustiness parameter,  D 7.5, 10, 12.

Figure 7. Set-up for granular heap dust test. The granular material forms a stockpile with an angle of repose
of approximately 23ı . A gas pressure gradient and periodic boundary conditions in the x-direction allows
gas re-circulation. Periodic boundary conditions are also applied in the ´-direction. The gas velocity and
dust mass flow rate are measured at the downstream end of the domain, marked as the measurement region.

4.2. Dust pick-up from a granular heap


The ability of the model to predict passive dust production was tested by modelling air flow
over a static granular heap, shown schematically in Figure 7. The computational domain used
was 0.5 m  0.1 m  0.1 m in the x-, y- and ´-directions, respectively. The particles and gas
were subject to periodic boundary conditions in the x- and ´-directions. A size distribution of

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
544 J. E. HILTON AND P. W. CLEARY

particles of between 3.0 to 3.5 mm was used, with density p D 2700 kg/m3 , coefficient of friction
 D 0.3 and coefficients of restitution e D 0.5. Particles were created in a vertical line spanning the
´-direction at a height of 0.075 m and allowed to fall freely under gravity onto the solid base of the
domain, building the heap. The final heap was approximately 0.25 m in base length with a height
of approximately 0.053 m at the apex of the heap, composed of 26 750 particles with an angle of
repose of ˛ D 23ı . The mean diameter of the saltating particles, ds , in Equation (13) was taken
to be 200 m, in accordance with the stated mean diameter for the saltating sand grains used by
Shao et al. [13]
To drive the gas flow, a gas pressure gradient was imposed by splitting the total pressure into a
fluctuating and steady part, p D pf C ps . From Equation (24), this gives

   
@u0 1 0 0 1  0
 rps
Cr  uu D 
rpf C r  Œ
 C fd C
g  (50)
@t
g g

such that the gradient of ps gives a driving pressure gradient and can straightforwardly be
implemented as a body force that can be combined with gravity. A number of pressure gradients
were used, from rps D 10 N/m3 to 55 N/m3 in 5 N/m3 increments, and each was run until the
gas flow reached an average statistically stationary velocity. This pressure boundary condition was
chosen in preference to imposing a boundary velocity from an analytic profile, such as a Blasius
profile, as it was unclear whether such a profile would accurately represent the true boundary
layer over a granular surface. Instead, the imposed pressure gradient allowed the velocity profile
to naturally develop. A further advantage is that an imposed profile would be smooth, whereas a
periodic boundary with re-circulating flow gives a turbulent inflow profile, which better matches the
expected inflow conditions for this application. The dust density was set to zero at the upwind face
of the domain and measured at the downwind face, marked on Figure 7 as the measurement region.
The measurement region, of area Am ,Pspanned the entire domain in the y- and ´-directions.
The dust flux over this region is qm D i di jvi j=Nm , where di is the dust density within the
computational grid cell i and jvi j the gas velocity within the cell. The number of cells spanning
the measurement region is given by Nm D Am = h2 . The mean bed gas speed, vN m , was also
measured by taking the average over this region. The vy and v´ velocity components were  103
smallerPthan vx , so only the vx component was used for the calculation of @m=@t and vN m , giving
vN m D i vxi =Nm .
The dust density for a range of pressure gradients at t D 10 s is shown in Figure 8 where the
dustiness parameter in Equation (14) is  D 17.5%. The flow initially underwent a short transient
period of around 5 s duration, during which the granular stockpile slightly shifted under the applied
pressure from the gas flow. After this initial period, the gas flow settled to a statistically stationary
state, with turbulent wake fluctuations around a stationary average flow velocity. The scalar dust
field was drawn in a long turbulent plume from the apex of the heap; the magnitude of which can be
seen to scale with the applied pressure gradient. Figure 9 shows the gas flow velocity field over the
heap for a steam-wise pressure gradient of 50 N/m3 at intervals of 2 s up to t D 6 s. The gas initially
flowed smoothly over the heap, creating a vortex on the downwind slope. The flow rapidly became
turbulent, with vortices formed on the downwind slope detaching into a turbulent downstream wake.
The turbulent wake advected through the periodic boundary onto the front of the heap, creating a
turbulent inflow profile. The majority of gas was forced up and over the heap, which formed a high
speed stream, with the remainder percolating at a reduced flow rate through the granular bed.
Both the dust flux through the measurement region, qm , and the mean bed gas speed, vN m , are
plotted against time in Figure 10 for pressure gradients of rps D 40 and 50 N/m3 . Both dust flow
rates show a spikes  1 s, with no corresponding velocity peak. These spikes are caused by the bed
re-adjusting under the applied gas pressure. After this brief re-adjustment period, the bed settled and
the gas flow and dust production fluctuated about mean values. These mean velocity and dust mass
flow rates are plotted on Figure 10 with dashed lines, where the mean flow has been evaluated from
5 to 30 s. A starting point of 5 s was chosen to ensure the transient behaviour during the start of the
simulation was not included in the statistically stationary mean values.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 545

Figure 8. Dust generated from an imposed gas flow over a granular heap, shown at t D 10 s for a range of
static pressure gradients. The dust is shown as a translucent coloured field, shaded by density, superimposed
over the particles. The dustiness parameter for these results is  D 17.5%.

An empirical expression given by the VDI for the vertical dust pickup rate from a dry granular
stockpile is [9]
 2 1.6
6 0.1vN m
F D 1.39  10 1 (51)
dN  tan ˛

where dN is the mean saltator diameter and  the saltator density. The expression is stated with the
assumption of a dry bed, removing a parameter of the order of unity from the original expression.
This expression was determined by the VDI using a logarithmic fit to dust production data for a
range of particle shapes, diameters and densities. The range of grain diameters was 150 m to
1 mm and the range of densities was 2600 to 4900 kg/m3 . The expression has been confirmed using
field measurements at iron ore shipment sites [9], with a stated error of approximately ˙30%.
The empirical expression, Equation (51), is compared with the simulation results in Figure 11,
where the stated error in the empirical expression, ˙30%, has also been plotted. Both expressions
were converted to a mass flow rate for comparison, as the vertical mass flow rate from the bed should
be equal to the mass flow exiting the domain during steady state. The empirical expression was
converted by multiplication by the upwind area incident to the flow, Ae D 0.1  0.125= cos.˛/ '
0.0136 m2 , and the simulation results were converted by multiplication by the dimension of the
measurement area, Am D 0.12 m2 . The resulting mass flow rates are plotted against the mean bed
gas speed, vN m in Figure 11. A mean saltator diameter of dN D 200 m was used in these simulations,
where the saltator density was assumed to that of the bulk density  D 2700 kg/m3 of the coarse
particles. The angle of repose was ˛ D 23ı . The saltator transition velocities also closely match,
with Equation (1) giving a saltation velocity of 1.68 m/s compared with an empirical transition
velocity of 1.51 m/s from Equation (51).

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
546 J. E. HILTON AND P. W. CLEARY

Figure 9. Gas velocity field within a cross-section of a granular heap, shown at t D 0.1, 2, 4 and 6 s. Velocity
vectors are shaded by velocity magnitude. A turbulent wake rapidly forms downstream from the heap.

Figure 10. The x-component of the gas velocity and the dust mass flow rate measured at the downstream
end of the domain over 30 s of simulation time. The mean gas velocity and mass flow rates, marked with
dashed lines, are taken from 5 to 30 s, after an initial transient phase. The dustiness parameter for these
results is  D 17.5%.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 547

Figure 11. Measured dust mass flow rate plotted against an empirical formula given by the Verein Deutscher
Ingenieure, Equation (51). The measured mass flow rate is shown with error bars equal to one standard
deviation. A mean saltator diameter of dN D 200 m and density  D 2700 kg/m3 were used, with a
dustiness parameter of  D 17.5%.

Simulation results using a dustiness parameter of  D 17.5% show an excellent match to the
empirical expression. Although the dust generation coefficient, , and the mean saltator diameter,
dN , will vary depending on the application, the close match between the VDI empirical expression
and the simulation results gives good confidence that the model provides realistic and accurate
results for passive dust generation scenarios.

5. CONCLUSIONS

Microscopic dust particles typically cohere to the surface of larger particles in granular media.
Imparting energy to these larger particles can break these cohesive bonds, forming suspended dust
in the gas around the granular material. In a dynamic granular bed, the energies imparted onto the
largest particles are typically dominated by inter-particle collisions with similarly sized particles.
However, for quasi-static granular systems with an imposed gas flow the cumulative energy from
collisions with much smaller particles can become significant. An imposed gas flow causes saltation,
where small particles are picked up from the surface of the bed by turbulent gas eddies, fall back
under gravity and impact the surface of the bed. At a steady gas flow rate, this process, known
as saltation bombardment, results in a net imparted collisional energy over the surface of the bed,
causing a steady rate of dust production.
Using this energy-based formulation, we have presented a computational methodology for
determining dust production and distribution in granular systems on the basis of a coupled
DEM–CFD approach. The DEM was used to model the large scale granular dynamics, and the CFD
method was used to model both the dust, as a continuum scalar density field, and the turbulent
gas flow. The methods were coupled through inter-particle drag relations. Dust production was
determined from the energy imparted to each DEM particle. This energy was split into inter-particle
collisional energies, resolved by the DEM simulation, and unresolved energies from the microscopic
bombardment of saltating particles. The unresolved component of the energies was calculated using
an empirical expression on the basis of the speed of the gas flow relative to the bed. The total
imparted energy was used to determine a dust flux within each CFD cell, which acted as a source
term in a dust density advection–diffusion equation. Turbulent diffusion of the dust density field was
crucial to the model and was incorporated using a LES WALE formulation, with a mixed approach
for turbulent scalar diffusion.
Measurements from simulations compared well with empirical expressions for both active and
passive dust productions. Active dust production was measured using a drop test, which gave good
qualitative agreement with experimental tests using powders. The dust mass was found to scale with
height and depended only on the total mass dropped, in agreement with an empirical expression.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
548 J. E. HILTON AND P. W. CLEARY

Passive dust production was tested using air flow over a granular heap. This gave excellent results
in comparison with a second empirical expression given by the VDI. The model therefore provides
accurate prediction of dust production and diffusion over a wide range of temporal scales, from
sudden impact events to long-term aeolian processes. It is particularly well suited to industrial
operations where dust is a major concern to the health of workers, as well as for monitoring and
predicting dust pollution in large-scale environmental systems.

REFERENCES
1. World Health Organization. Hazard prevention and control in the work environment: airborne dust, 1999.
WHO/SDE/OEH/99.14.
2. Morrow PE. Dust overloading of the lungs: update and appraisal. Toxicology and Applied Pharmacology 1992;
113:1–12.
3. Abbasi T, Abbasi SA. Dust explosions – cases, causes, consequences, and control. Journal of Hazardous Materials
2007; 140:744.
4. Frank WL. Dust explosion prevention and the critical importance of housekeeping. Process Safety Progress 2004;
23:175–184.
5. Gregory JM, Wilson GR, Singh UB, Darwish MN. TEAM: integrated, process-based wind-erosion model.
Environmental Modelling and Software 2004; 19:205–215.
6. Harris AR, Davidson CI. Particle resuspension in turbulent flow: a stochastic model of individual soil grains. Aerosol
Science and Technology 2008; 42:613–628.
7. Harris AR, Davidson CI. A Monte-Carlo model for soil particle resuspension including saltation and turbulent
fluctuations. Aerosol Science and Technology 2009; 43:161–173.
8. Li Y, Guo Y. Numerical simulation of Aeolian dusty sand transport in a marginal desert region at the early
entrainment stage. Geomorphology 2008; 100:335–344.
9. Verein Deutscher Ingenieure. Emissions of gases odours and dusts from diffusive sources, 2010. VDI 3790 part 3.
10. Bagnold RA. The Physics of Blown Sand and Desert Dunes. Methuen: London, 1941.
11. Anderson RS. Wind modification and bed response during saltation of sand in air. Acta Mechanica Supplement 1991;
1:21–51.
12. Owen PR. Saltation of uniform grains in air. Journal of Fluid Mechanics 1964; 20:225–242.
13. Shao Y, Raupach MR, Findlater PA. Effect of saltation bombardment on the entrainment of dust by wind. Journal of
Geophysical Research 1993; 98:12719–12726.
14. Marticorena B, Bergametti G. Modeling the atmospheric dust cycle: 1. Design of a soil-derived dust emission scheme.
Journal of Geophysical Research 1995; 100:16415–16430.
15. Cleary PW. Large scale industrial DEM modelling. Engineering Computations 2004; 21(2–4):169–204.
16. Delaney GW, Hilton JE, Cleary PW. Defining random loose packing for nonspherical grains. Physical Review E
2011; 83:051305.
17. Anderson TB, Jackson R. A fluid mechanical description of fluidised beds. Industrial and Engineering Chemistry
Fundamentals 1967; 6:527–539.
18. Kafui DK, Thornton C, Adams MJ. Discrete particle-continuum fluid modelling of gas-solid fluidised beds. Chemical
Engineering Science 2002; 57(13):2395–2410.
19. Di Felice R. The voidage function for fluid-particle interaction systems. International Journal of Multiphase Flow
1994; 20:153–159.
20. Hilton JE, Mason LR, Cleary PW. Dynamics of gas-solid fluidised beds with non-spherical particle geometry.
Chemical Engineering Science 2010; 65:1584–1596.
21. Hilton JE, Cleary PW. The influence of particle shape on flow modes in pneumatic conveying. Chemical Engineering
Science 2011; 66:231–240.
22. Hilton JE, Cleary PW. Granular flow during hopper discharge. Physical Review E 2011; 84:1307–1307.
23. Smagorinsky J. General circulation experiments with the primitive equations. Monthly Weather Review 1963;
91:99–164.
24. Germano M, Piomelli U, Moin P, Cabot W. A dynamic subgrid-scale eddy viscosity model. Physics of Fluids A
1991; 3:1760–1765.
25. Lilly DK. A proposed modification of the Germano subgrid-scale closure method. Physics of Fluids A 1992;
4:633–636.
26. Nicoud F, Ducross F. Subgrid-scale stress modelling based on the square of the velocity gradient tensor. Flow
Turbulence and Combustion 1999; 62:183–200.
27. Weickert M, Teike G, Schmidt O, Sommerfeld M. Investigation of the LES WALE turbulence model within the
Lattice Boltzmann framework. Computers and Mathematics with Applications 2010; 59:2200–2214.
28. Chumakov SG. A priori study of subgrid-scale flux of a passive scalar in isotropic homogeneous turbulence. Physical
Review E 2008; 78:036313.
29. Corrsin S. Limitation of gradient transport models in random walks and in turbulence. Advances in Geophysics 1974;
18A:25–60.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld
DUST MODELLING USING A COMBINED CFD AND DISCRETE ELEMENT FORMULATION 549

30. Clark RA, Ferziger JH, Reynolds WC. Evaluation of subgrid models using an accurately simulated turbulent flow.
Journal of Fluid Mechanics 1979; 91:1–16.
31. Sun OS, Su LK. Experimental assessment of scalar mixing models for large-eddy simulation. AIAA Paper 2550,
2004.
32. Heitbrink WA, Baron PA, Willeke K. An investigation of dust generation by free falling powders. American Industrial
Hygiene Association Journal 1992; 53:617–624.

Copyright © 2012 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids 2013; 72:528–549
DOI: 10.1002/fld

You might also like