You are on page 1of 12

Powder Technology 313 (2017) 332–343

Contents lists available at ScienceDirect

Powder Technology

journal homepage:

Evaluation of drag correlations using particle resolved simulations of

spheres and ellipsoids in assembly
Long He, Danesh K. Tafti ⁎, Krishnamurthy Nagendra 1
Department of Mechanical Engineering, Virginia Tech, Blacksburg, VA, USA

a r t i c l e i n f o a b s t r a c t

Article history: Particle-resolved simulations are performed to study the momentum transfer in flow though fixed random as-
Received 7 November 2016 sembly of non-spherical particles. Ellipsoidal particles with sphericity (ψ =0.887) are investigated in a periodic
Received in revised form 26 February 2017 cubic domain to simulate an infinite assembly. The incompressible Navier-Stokes equations are solved using
Accepted 4 March 2017
the Immersed Boundary Method (IBM). Pressure and viscous force on each particle are calculated based on the
Available online 10 March 2017
resolved flow field. Flow through an assembly of spherical particles is tested, and predicted drag forces are com-
pared with previous particle resolved simulation results to validate the current framework. The assembly of el-
Particle resolved simulation lipsoidal particles is simulated for solid fraction between 0.1 and 0.35 using 191 to 669 particles, respectively,
Non-spherical particle at low to moderate Reynolds numbers (10≤Re≤200). The simulation results show that the drag force of ellipsoi-
Ellipsoid dal particles is 15% to 35% larger than equal volume spherical particles. Widely used drag force correlations are
Drag correlation evaluated based on the current simulation results. The comparisons show that for ellipsoidal particles over the
Immersed boundary method range of parameters investigated in the present study, the combination of Tenneti et al.'s correlation with Holzer's
single non-spherical particle drag model has the best performance with an average difference of 7.15%.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction drag coefficient against Reynolds number for different non-spherical

particles, they showed that as the sphericity decreased, the drag coeffi-
Gas-solid flow is fundamental to many industrial processes such as cient increased significantly. Consequently, several drag correlations
pollution control, CO2 capture, biomass gasification, chemical reactors, have been developed based on experimental data, using particle shape
sprays, pneumatic conveying, etc. Extensive experimental and numeri- and Reynolds number as parameters. Chhabra et al. [2] studied correla-
cal studies have been devoted to understand the interphase momentum tions for estimating the drag coefficient of non-spherical particles in in-
transfer in such flows during the last several decades. Most of the stud- compressible viscous fluids. 1900 experimental data points from 19
ies have focused on spherical particle shapes to simplify the challenge of independent studies were used to evaluate the accuracy of five correla-
understanding the flow and particle characteristics. However in most tions, including those proposed by Haider and Levenspiel [3], Ganser [4],
natural and industrial processes, the particle shape is seldom spherical. Chien [5], Hartman et al. [6] and Swamee and Ojha [7]. They found that
In fact, particle shape is one of the important parameters that can have a the correlation developed by Ganser, which uses the equivalent spheri-
significant impact on momentum, heat and mass transfer, which are cal diameter and the sphericity as the particle shape parameters, per-
fundamental to all processes. Although these effects of particle shape formed the best. The average error from this method was 16%, though
have already been widely recognized, only a few studies have been car- the maximum error was as high as 180%. Holzer and Sommerfeld [8] de-
ried out to quantity them. veloped a general drag correlation which depends on particle sphericity,
Most research on drag coefficient for non-spherical particle have fo- Reynolds number, and particle orientation. Comparing to 2061 experi-
cused on a single isolated particle. Yow et al. [1] collected experimental mental data points, they showed that the new correlation had a mean
data of drag coefficient for a wide range of shapes including spheres, relative deviation of 14.1%. After performing experimental measure-
cube octahedrons, octahedrons, cubes, tetrahedrons, discs, cylinders, ments of the terminal velocities of irregular particles falling in fluids,
rectangular parallelepipeds and others. By studying the variation of Dioguardi and Mele [9] proposed a drag correlation using sphericity
and circularity as shape factors. They found the new correlation is able
to predict the terminal velocity with about 11% error when the Reynolds
⁎ Corresponding author at: 213E Goodwin Hall, Dept. of Mechanical Engineering, number is known. Using a similar experimental setup, Dioguardi et al.
Virginia Tech, Blacksburg, VA 24061, USA.
E-mail address: (D.K. Tafti).
[10] proposed another shape-dependent drag coefficient and integrated
Current designation: Senior Physicist, Exa Corporation. Address: 55 Network Drive, in to the multiphase code MFIX-DEM. Bagheri and Bonadonana [11] re-
Burlington MA 01803. cently proposed a general model for prediction drag coefficient for non-
0032-5910/© 2017 Elsevier B.V. All rights reserved.
L. He et al. / Powder Technology 313 (2017) 332–343 333

spherical particles for Reynolds number up to 3 × 105 based on experi- (GenIDLEST). The details of the framework and methodology used in
mental data. Instead of using the sphericity, this new correlation uses GenIDLEST can be found in Tafti [22] and Tafti [23]. In this section, the
two shape factors based on the particle flatness and elongation, and di- relevant governing equations and the modified treatments of the
ameter to quantify the shape of the particles. The authors state that the governing equations under fully-developed flow conditions are pre-
new correlation has an average error of ~ 10%, which is significantly sented along with some details of the Immersed Boundary Method
lower than existing correlations. [24]. GenIDLEST with immersed boundary method has been successful-
Compared to the rich database and correlations available for drag of ly applied to oscillating cylinders [24], turbine cooling passage heat
an isolated particle, only a few studies focusing on drag coefficient of transfer [25], and fluid structure interaction problems [26].
packed non-spherical particles have been carried out. Nemec and
Levec [12] demonstrated that the original Ergun equation [13] is only 2.1. Governing equations and numerical technique
able to predict the pressure drop in flow over spherical particles, where-
as it systematically under-predicts the pressure drop in flow over non- The non-dimensional form of the Navier-Stokes equations for in-
spherical particles. They proposed that the dimensionless form of the compressible flow are as follows:
Ergun equation, developed by Niven [14] with two constant factors, Continuity:
which only vary for different particle shapes, should be used for non-
spherical particles. They conclude that with their modifications, the ∂ui
¼0 ð1Þ
Ergun equation is able to predict pressure drop in flow through a packed ∂xi
bed of certain non-spherical particles within 10% error. However, there
is no correlation to calculate the constants for arbitrary particle shapes, Momentum:
leaving that determination to experiments or particle resolved simula- !
tions. Moreover, the original Ergun equation significantly over- ∂ui ∂ u j ui ∂P 1 ∂ ui
predicts the average drag force at low solid volume fractions [15], and þ ¼− þ ð2Þ
∂t ∂x j ∂xi Reref ∂x j ∂x j
hence the modified equation applied to a dilute system of non-
spherical particles has not been validated. Machač and Dolejš [16] inves-
where the non-dimensionalizations are:
tigated the pressure drop in flow through a random fixed bed of non-
spherical particles. They suggested that a ‘bed factor’ which contains t  uref P  −P ref ρref uref Lref
xi ui
the information of the non-spherical particle surface area is needed to xi ¼  ; ui ¼  ; t ¼  ; P ¼  2 Reref ¼
Lref uref Lref ρref uref μ ref
accurately predict the pressure drop. Hua et al. [17] studied the com-
bined Ganser and Ergun correlations and modified Syamlal and O′
Brien drag models in Eulerian-Eulerian CFD framework for irregular The above governing equations are transformed to generalized coor-
shape in dense gas-solid fluidized beds. Dorai et al. [18] examined the dinates, and discretized in a conservative finite-volume formulation
local packing structure of fixed bed reactors made of cylindrical pellets using a second-order central (SOC) difference scheme on a non-
packed in cylindrical tubes. They found that the local packing structure staggered grid topology [27]. Cartesian velocities, pressure and temper-
(particle orientation) is strongly affected by the particle size distribu- ature are calculated and stored at the cell center, whereas fluxes are cal-
tion. Zhou et al. [19] studied packed ellipsoidal particles in a fluidized culated and stored at cell faces. A projection method using second order
bed using CFD-DEM simulations. The drag model developed by Holzer predictor-corrector steps is used for the time integration of the continu-
et al. [8] was used for the ellipsoids. They concluded that the accuracy ity and momentum equations. In the predictor step, an intermediate ve-
and applicability of current correlations is questionable, and a general locity field is calculated; and in the corrector step, an updated
and reliable correlation to determine fluid drag on non-spherical parti- divergence free velocity is calculated at the new time-step by solving
cles is urgently needed. Vollmari et al. [20] investigated the pressure a pressure-Poisson equation.
drop in packing of arbitrary shaped particles using CFD-DEM and exper-
iment. Based on experimental measurements, they suggested that par- 2.2. Fully developed calculations
ticle drag in DEM simulations should be calculated combining the
correlations developed by Holzer et al. [8] and De Felice [21]. They rec- The computational domain consists of a three-dimensional periodic
ommend fully resolved simulations to understand the effects of inho- box representing an unbounded particle assembly. Flow is induced
mogeneous packing on fluid flow structure introduced by the non- along the x-direction by applying a constant mean pressure gradient
spherical particle shape. to balance the form and friction losses, and the total pressure is
All these studies demonstrate that further investigation is needed to expressed in terms of the mean pressure and a fluctuating or periodic
have a better understanding of flow through an assembly of non- component as shown in Eq. (3).
spherical particles. Therefore, in this study, particle-resolved simula- *  * 
tions are carried out to investigate momentum transfer for a random as- P x ; t ¼ −β  x þ p x ; t ð3Þ
sembly of non-spherical particles. In this approach, fluid force on each
particle is calculated by resolving the flow field around each individual where β is a constant. Because the applied pressure gradient will bal-
particle. Ellipsoidal particles with aspect ratio of 2.5 are used and the ance the form and friction losses when the flow reaches a steady state,
particle-fluid boundaries are resolved using an Immersed Boundary the actual value used for β is not of any consequence to the calculated
Method (IBM). The objectives of the current study are: (1) validate forces as long as the mean Reynolds number obtained from the simula-
the current particle resolved simulation framework by comparing re- tion is the same.
sults for flow through spherical particle assemblies to existing drag cor- With the unchanged continuity equation, the momentum equation
relations; (2) simulate flow through an assembly of ellipsoids and can be written as:
extract the drag force; (3) evaluate the performance of existing drag
correlations for the non-spherical particles. ∂ui ∂ u j ui ∂p 1
∂ ui *
þ ¼− þ þ βex ð4Þ
∂t ∂x j ∂xi Reref ∂x j ∂x j
2. Numerical method
All simulations are performed using an in-house code – Generalized The β e x in Eq. 4 is the mean applied pressure gradient source term,
Incompressible Direct and Large Eddy Simulation of Turbulence which balances the mean form and friction losses in the flow direction.
334 L. He et al. / Powder Technology 313 (2017) 332–343

The detailed procedure used can be found in Patankar [28] and Zhang 2.3.4. Force calculation
et al. [29]. The fluid force is calculated on each surface element of the immersed
surface which is defined by an unstructured triangulated mesh. For each
surface element, a centroid and normal direction are calculated based
2.3. IBM
on the location of the vertices. Then a probe is assigned that lies along
the surface element normal from the element centroid as shown in
The immersed boundary method (IBM) used in this paper is an indi-
Fig. 2. The same tri-linear interpolation method used for IB node probes
rect forcing sharp interface approach. This method is an extension of a
(but with a different surrounding node stencil) is employed to deter-
scheme originally proposed by Gilmanov and Sotiropoulos [30], and
mine the value of the desired primitive flow variable at the probe loca-
has been modified to fit the generalized coordinate system and non-
tion. In the present simulations, pressure and velocity are calculated at
staggered grid framework of GenIDLEST. In this method, the governing
the probe location. The distance from the probe to the element (dp) is
equations are solved without modification for the computational grid in
selected such that no solid node value is used in the interpolation and
the fluid domain. Special treatment is only applied to the first layer of
the linear gradient assumption is still valid. The force on each element
the fluid nodes (fluid IB nodes) next to the immersed boundary. The
consists of pressure and viscous forces.
major steps of this method are summarized below. The details can be
found in Nagendra et al. [24]. Pressure. To estimate the pressure component of the force at each
surface element, the non-dimensional material derivative of velocity at
2.3.1. Definition of surface grid the surface is related to the normal pressure gradient as
The immersed boundary is the fluid/solid interface which is provid-
ed as input in a discretized form of an unstructured surface mesh. Each *
dU 1 dP
of the immersed boundary elements has information of surface normal, ¼− ð5Þ
dt ρd *
area, and location with respect to the fixed background fluid mesh.

by neglecting the viscous contribution to the pressure gradient. By as-

2.3.2. Node identification suming a constant acceleration in the region between the surface ele-
By employing a search-locate and interpolate algorithm, the cell ment and element probe, Eq. (5) is discretized as
types on the background mesh are determined based on their location
with respect to the immersed boundary. All the nodes inside the solid  
P p −P e
boundary are assigned as solid nodes and the rest are assigned as fluid a*n ¼ ð6Þ
nodes. Moreover, the fluid nodes that lie in the immediate vicinity of ρdp
the immersed boundary are marked as fluid IB nodes. Fig. 1 depicts
the node types on the background grid. Here a* is the surface element acceleration component in the sur-
face normal direction. In the current study, since the particles in assem-
2.3.3. Boundary treatment bly are fixed, a* ¼ 0. Thus:
In this step, modifications are made to the IB node values in order for
the fluid nodes to see the presence of the immersed boundary. For each
Pe ¼ Pp ð7Þ
IB node, a probe is assigned which lies on the surface element normal di-
rection and at one cell distance from the IB node. A tri-linear interpola-
tion method is used to determine the value of the desired primitive flow It is to note that in the current formulation, the total pressure is used
variable at the probe location from the surrounding nodes. The value in calculating the pressure force (mean + fluctuating in Eq. (3)). This
calculated at the probe is then used in the computation of the value at definition is the same as Hill et al. [31] and Tenneti et al. [32], whereas
IB node to satisfy the appropriate boundary condition. The fluid Van Der Hoef et al. [33] and Beetstra et al. [34] define the particle
governing equations are solved on the fluid nodes with the IB nodes act- force by subtracting the force component due to mean pressure
ing as the de facto boundary with the correct boundary conditions. gradient.

Fig. 1. Node type on the background fluid mesh defined by the immersed boundary method.
L. He et al. / Powder Technology 313 (2017) 332–343 335

Fig. 2. Pressure and Shear stress calculation at the immersed surface. Shear stress. To calculate the viscous forces, the velocity of the fraction. Similar methods have already been successfully used to extract
* drag laws for spherical particles in assembly by Hill et al. [31], Van Der
fluid at the probe location, U p , is used to calculate the shear stress com-
Hoef et al. [33], Beetstra et al. [34] and Tenneti et al. [32].
ponent at each surface element. The fluid velocity is decomposed into
In all calculations, Uref = U, the superficial velocity which equals the
the tangential and surface normal components. In the current study,
x-direction volumetric flowrate divided by the area of the cross-
only the tangential shear stress is used under the assumption that the
section, and Lref =Deq, the equivalent spherical diameter. Thus the Reyn-
normal stresses are insignificant. Assuming a linear tangential velocity
olds number is defined as:
profile, the non-dimensional shear stress is:
* ρUDeq
* 1 U pt Re ¼ ð13Þ
τw ¼ ð8Þ μ
Re dp
where ρ and μ are the fluid density and dynamic viscosity.
The element surface force is calculated by taking both pressure and The computational domain is a cube with side 10Deq. Results from a
shear stress into account: grid independency study are given in Table 1 for ellipsoid particles. Res-
*  olutions of 30 to 50 background grid cells per equivalent particle diam-
* *
d F e ¼ τ w −P e  n dA ð9Þ eter were tested. It is shown that the maximum difference between the
three grids is about 2% so 40 grid cells per Deq were deemed to be suffi-
* ciently fine to resolve the fluid field around each particle.
where n is the surface element normal vector.
Because of the fixed domain size, the number of particles N in the
The fluid force acting on each particle is calculated by integrating the
computation domain is a function of solid volume fraction ϕ alone:
viscous and pressure force on the surface of the particle:
* * 6ϕL3 6000ϕ
Fi ¼ ∮ Ad F e ð10Þ N¼ ¼ ð14Þ
πD3eq π

The average drag force on all particles is estimated by:

Sphericity ψ, proposed by Wadell [35] is the most widely used pa-
* 1 N * rameter to quantify the shape of non-spherical particles.
F g−s ¼ ∑ Fi ð11Þ
N i¼1
1 2
π3 6V p 3 As
ψ¼ ¼ ð15Þ
All forces are further normalized using the Stokes-Einstein relation: Ap Ap

F g−s where Ap is the surface area of the particle, Vp is the volume of the par-
F¼ ð12Þ ticle, and As is the surface area of a sphere which has the same volume as
3πμDeq U
the particle. The more spherical a particles is, the closer this factor gets
to one. The geometry parameters and number of particles for each
solid fraction are shown in Table 2. An ellipsoidal particle shape is cho-
3. Simulation setup
sen because CFD-DEM simulation methods have been developed for
such particles [19,36], but the drag correlations used in these studies
3.1. Geometry, grid and simulation parameters

In the present study, only a mono-disperse system is considered, i.e.

Table 1
all particles have the same geometric parameters. The computational Grid independence study for packed ellipsoidal particles at solid fraction of 0.1.
domain is a cube, with the periodic boundary condition applied along
all three directions. A pressure gradient is applied along the x-
direction to drive the fluid flow. Particles are randomly distributed in Re 30 cells per Deq 40 cells per Deq 50 cells per Deq
the computational domain, fixed at their initial positions. When the 10 3.55 3.58 3.63
flow reaches a steady state, fluid forces acting on each particle are calcu- 50 6.00 6.03 6.13
lated. By averaging the forces over all the particles in the fluid domain, 100 8.71 8.79 8.80
200 14.35 14.44 14.46
the drag force is calculated for a given Reynolds number and solid
336 L. He et al. / Powder Technology 313 (2017) 332–343

Table 2
Geometry parameters and the number of particles for each solid fraction. (Normalized
with Deq).

Sphere Ellipsoid

Geometry parameter Diameter = 1 Semi-major 0.921

Semi-minor 0.368

Sphericity ψ 1 0.887
Deq 1 1
Number of particles (N) ϕ = 10% 191 191
ϕ = 20% 382 382
ϕ = 30% 573 573
ϕ = 35% 669 669
ϕ = 40% 764 –
ϕ = 45% 860 –

are not validated. In addition, the drag coefficient for a single isolated el-
lipsoidal particle as a function of particle orientation and Reynolds num-
ber is available in the literature [37]. For 10% solid fraction 191 particles
are used which increase to 669 particles for ϕ = 35%. In contrast, past
particle resolved simulation studies on assembly of particles have
used 32 to 60 particles [31], 54 particles [33,34], 16 to 161 particles
[32] for solid fraction between ~ 0.1 to ~ 0.5. Tests were conducted to
qualify the intended randomness of the particle distribution. Three dif-
ferent random realizations were tested for ellipsoid particles at ϕ =
10% (191 particles), which showed that there was b1% difference in Fig. 3. Packing of : (a) spheres; (b) capsules; (c) cylinders; (d) mixed spheres, capsules,
the calculated mean drag. The solid fraction for ellipsoids was limited dodecahedron, cubes, cuboid, cylinders with random size.
to 35% because beyond this packing density it became difficult to gener-
ate a random distribution without particle-particle overlap. To generate
distributions with higher solid fractions it would have been necessary to developed by Tenneti et al. [32]:
give some preferential orientation to the ellipsoids and would have
been outside the scope of this investigation. F iso ðReÞ
F¼ þ F ϕ ðϕÞ þ F ϕ;Re ðϕ; ReÞ

3.2. Particle assembly 5:81ϕ 0:48ϕ3

F ϕ ðϕÞ ¼ 3
þ ð16Þ
ð1−ϕÞ ð1−ϕÞ4 !
The random particle assembly is generated using a physics simula- 0:61ϕ3
F ϕ;Re ðϕ; ReÞ ¼ ϕ3 Re 0:95 þ
tion engine SDK- PhysX by Nvidia [38]. During the assembly process, ð1−ϕÞ2
each particle is randomly spawned in the designated space and associ-
ated with a mesh to represent its geometry. If the particle spawns at a
where Fiso(Re) is the Dalla Valle [39] drag correlation for a single spher-
location which has already been occupied by another particle, an over-
ical particle.
lap is detected between the convex meshes, and the two particles are
The current predictions show very good agreement with the correla-
relocated. This process ensures that a random assembly of particles is
tion when the Reynold number is b200. At Reynolds number of 200, the
created with no overlap. Any shape created by CAD software can be
predicted drag force from the present simulation is 5%–10% higher than
imported and packed in this framework. Fig. 3 shows packing of:
the correlation. Tenneti et al. [32] have noted that their predictions,
(a) spheres; (b) capsules; (c) cylinders; (d) mixed spheres, capsules,
from which the correlation is built, consistently predict a lower drag
dodecahedron, cubes, cuboid, cylinders with random size. After the as-
at Re = 200 than that reported by other researchers.
sembly, the location and orientation of each particle is used to create
the surface mesh for IBM.
5. Results - ellipsoidal particles

4. Results - spherical particles The simulation setup for calculating the drag force of a random array
of ellipsoidal particles is the same as that of spherical particles. The sur-
To test the current model for calculating drag force, we first simulate face grid for each ellipsoidal particle is an unstructured triangular mesh
randomly packed spherical particles and compare our results with the with 2100 cells. Solid fractions ϕ of 0.1, 0.2, 0.3 and 0.35 are tested for a
correlation recently developed by Tenneti et al. [32] who also used par- Reynolds number ranging from 10 to 200. The number of particles for
ticle resolved simulation. each solid fraction is shown in Table 1. Fig. 6 shows the immersed sur-
The surface mesh of each sphere is an unstructured mesh consists of face of ellipsoids at solid fraction of 10% and 35%. Fig. 7 shows a repre-
about 2000 triangular elements. Six different solid fractions ϕ, ranging sentative velocity field in the assembly of ellipsoidal particles at
from 0.1 to 0.45, are tested. Reynolds number based on the superficial Reynolds number of 10 and solid fraction of 0.35. The velocity magni-
velocity and diameter are tested at 10, 50, 100 and 200. Drag force on tude is dependent on the local particle packing densities with larger
each particle is calculated when the fluid field reaches a steady state. fluid velocities prevailing in regions with small interstitial spaces.
Fig. 4 shows the immersed surface identified by GenIDLEST for solid The pressure force and viscous forces, normalized by the Stokes-
fractions of 10% and 45%. As the solid fraction increases, the interstitial Einstein relation, are shown in Fig. 8. The pressure force increases ap-
spaces between particles decrease, making it more challenging to ac- proximately linearly with Reynolds number for both spherical and ellip-
commodate the IBM probes in fluid regions. soidal particles but the slope is larger for ellipsoids than the slope for
Fig. 5 shows the normalized drag force of spherical particles from the spherical particles at the same solid fraction. At the same solid fraction,
current simulations. The results are compared to the drag correlations the difference between ellipsoidal and spherical particles increases as
L. He et al. / Powder Technology 313 (2017) 332–343 337

Fig. 4. Immersed surface of spherical particles at :(1) solid fraction of 10%;(2) solid fraction of 45%.

the Reynolds number increases. This trend has also been reported for fraction is also plotted in Fig. 9. Correlations developed by Wen and
single isolated particles, the influence of particle shape on drag in- Yu [40] and De Felice [21] have the form of:
creases as the Reynolds number increases [4]. At the same Reynolds
number, the difference between ellipsoidal and spherical particles in- F ðRe; ϕÞ ¼ F iso ðReÞð1−ϕÞ−α ð19Þ
creases as the solid fraction increases. The viscous force varies as a
sublinear power of the mean flow Reynolds number. The difference in
where Fiso(Re) is the drag correlation for isolated single particle, and the
viscous force between ellipsoid and sphere also increases as the Reyn-
neighboring effects are accounted by the term (1− ϕ)−α. α is set to 4.7
olds number and void fraction increase: at solid fraction of 0.1, the vis- 2
cous force shows little difference while at solid fraction of 0.35, the in Wen and Yu's correlation, and 4:7−0:65 exp½− ð1:5−logReÞ 2  in De
force on the ellipsoids is 15% to 20% larger than spheres. Felice's correlation. In the comparisons, the correlation developed by
In the following section, we compare simulation results to different Turton and Levenspiel [41] and Dalla Valle [39] for single sphere are
existing drag correlations in the literature. The comparison is divided used for F(Re, 0). Syamlal and O′Brien [42] developed their drag law
into three categories based on the parameters required by the correla- based on the terminal velocity correlation from Richardson and Zaki
tion: (I) only equivalent spherical diameter (Deq) and solid fraction [43]. Finally, Tenneti et al.’s correlation is based on particle resolved
(ϕ) are used; (II) particle sphericity ψ is used along with Deq and ϕ. In simulations.
these two comparisons, solid fraction is the only parameter needed For all the solid fractions tested in the current work, the drag for el-
from the assembly. (III) The final comparison is at the scale of the parti- lipsoidal particles is larger than the corresponding force for an assembly
cle, in which each particle's orientation is used along with ψ, ϕ and Deq to of spherical particles from IBM. This force difference increases as the
calculate the drag force. This method can only be applied to numerical Reynolds number increase. At solid fraction of 0.1, the current simula-
simulations where the detailed orientation of each particle is available. tion results for ellipsoids agree with Syamlal and O′Brien's correlation
and which gives values much larger than other correlations. At solid
fraction of 0.2 to 0.35, the Wen and Yu correlation with Dalla Valle ex-
5.1. Category I: Comparing to spherical particle packing correlations hibits the closest agreement with the current simulation results for el-
lipsoids while other correlations under-predict by 15% to 50%. We can
First we compare IBM results with widely used correlations using conclude that no correlations works consistently at all solid fraction
the equal volume diameter as the characteristic length, shown in range for ellipsoidal particles. It is noteworthy that even for spherical
Fig. 9. In many two-phase flow research studies with non-spherical par- particles, there is considerable deviation between the correlations as
ticles, the assumption that the particle shape is spherical is used. This Re increases, with the combination of Wen-Yu with Turton and
comparison shows the error introduced by this assumption. The drag Levenspiel correlations and the correlation of Tenneti et al. giving the
force of spherical particle at the same Reynolds number and solid best agreement with current results.

Fig. 5. Normalized drag force from IBM compared to drag correlation developed by Tenneti et al. [25] using particle resolved simulations.
338 L. He et al. / Powder Technology 313 (2017) 332–343

Fig. 6. Immersed surfaces of ellipsoidal particles at (1) 10% solid fraction (2) 35% solid fraction.

5.2. Category II: comparing to non-spherical packing correlations using the particle flatness, elongation, equivalent diameter and
particle-to-fluid density ratio as parameters, is also tested here. The
In this section, we test the performance of combined correlations drag force predicted by the combined correlations and current simu-
by substituting the F(Re, 0) in Wen and Yu (hereinafter referred to as lations are plotted in Fig. 10. For solid fractions of 0.1 and 0.2, all the
‘W&Y’) and Di Felice's correlations with single non-spherical particle correlations under-predict the drag force by 5% to 35% for ellipsoids.
drag models. These type of combined correlations have been used in At solid fraction of 0.3 and 0.35, W&Y with H&L capture the trend
CFD-DEM simulations for predicting drag in non-spherical particle while still under predicting by 10% to 27%. W&Y with B&B shows a
fluidized beds [19,36]. For single non-spherical particles, correla- slight improvement over the W&Y and H&L correlation. However,
tions derived by Harder and Levenspiel (hereinafter referred to as this combination still under predicts the drag force at low solid frac-
‘H&L’) [3] and Chien [5] which use the sphericity (ψ) to describe tion. This could be because of the low Reynolds number in this study
the particle shape are tested. The drag correlation developed by and the fact that the particle-fluid density ratio term is set to infinity
Bagheri and Bonadonna (hereinafter referred to as ‘B&B’) [11], in the correlation because the particles are fixed to their initial

Fig. 7. Representative velocity field (fluid flow in x-direction) at Reynolds number of 10 and solid fraction of 0.35.
L. He et al. / Powder Technology 313 (2017) 332–343 339

Fig. 8. Normalized pressure force and viscous force.

positions. Similar to using the spherical particle drag correlation, the used along with De Felice's correlation:
non-spherical particle drag correlation when combined with the Di
Felice and W&Y correlation, do not give a consistently accurate esti- F ðRe; ψ∥ ; ψ; ψ⊥ Þ !
mation of drag force for the assembly of ellipsoidal particles. 8 1 16 1 3 1 0:2 1
¼ pffiffiffiffiffi þ pffiffiffiffi þ pffiffiffiffiffiffi 3 þ 0:42  100:4ð−logψÞ
Re ψ∥ Re ψ Re ψ 4 ψ⊥
5.3. Category III: comparing to non-spherical packing correlations with ori- Re
entation data (Deq , ψ and orientation) 24

In CFD-DEM studies of packed non-spherical particles, the single In this equation, ψ, ψ⊥ and ψ∥ are the sphericity, crosswise sphericity,
non-spherical drag correlation developed by Holzer et al. [8] has been and lengthwise sphericity. The crosswise sphericity (ψ⊥) is the ratio

Fig. 9. Normalized drag force of spherical and ellipsoidal particles compared to widely used correlations for spherical particles in assembly.
340 L. He et al. / Powder Technology 313 (2017) 332–343

Fig. 10. Normalized drag force of ellipsoidal particles comparing to modified correlations.

between the cross-sectional area of the equal volume sphere and the (The original drag coefficient correlation is modified based on the
projected cross-sectional area of the particle in the flow direction, and normalization by Stoke's drag used in the present paper.) This correla-
the lengthwise sphericity (ψ∥) is the ratio between the cross-sectional tion is also tested here.
area of the equal volume sphere and the difference between half the By substituting the above single non-spherical particle correlation
surface area and the mean projected longitudinal cross-sectional area into De Felice's correlation and Tenneti et al.’s correlation, the drag
of the particle. In the current simulations, the orientation of the ellipsoi- force of each particle in assembly can be estimated. This type of com-
dal particle can be fully described using the angle (α) which is the angle bined correlations have been used in CFD-DEM simulations for
between the fluid flow direction and the particle's semi-major axis as predicting drag in non-spherical particle beds [19,20,36].
shown in Fig. 11. Each particle's ψ⊥ and ψ∥ can be calculated based on Figs. 12 and 13 show the normalized drag force on each particle vs the
this angle. For the elliposid particle shape tested in the current study, particle orientation α at solid fraction of 0.1 and 0.35. The scatter markers
Zastawny et al. [37] proposed the following correlation for a single iso- in the figures are the drag force calculated on each individual particle
lated particle: from one representative simulation at a given solid fraction and Reynolds
number. The open squares are averaged force values from the simulation
  2 over an interval of π/36 . The solid line is the drag force predicted using
F ðRe; α Þ ¼ F α¼0
 þ F α¼π=2 −  F α¼0  ðα Þ
 Sin the combined De Felice and Holzer correlation (hereinafter referred to
5:1 15:52 Re
Fα¼0 ¼ þ as ‘F&H correlation’), the dash-dot line is the combined Tenneti et al.
Re0:48 Re1:05 24  ð21Þ
 and Holzer correlation (hereinafter referred to as ‘T&H correlation’),
24:68 3:19 Re
Fα¼π=2 ¼ þ and the dash line is the combined Tenneti et al. and Zastawny correlation
Re0:98 Re0:21 24
(hereinafter referred to as ‘T&Z correlation’). The simulation results show
that the particle drag can have significant variance even at the same ori-
entation. However, in general at a given Reynolds number and solid frac-
tion, the mean particle drag tends to increase as the angle α increases.
Using the combined correlations, each particle's drag force can be
calculated based on its orientation, α. Table 3 shows the normalized
mean drag force from the current IBM simulations comparing to F&H,
T&H and T&Z correlations. The correlation predictions are based on
the orientation data of each individual particle from the simulations.
The F&H correlation predicts significantly lower drag force at Reynolds
number of 10. At Reynolds number of 50, 100 and 200 and solid fraction
of 0.1, the F&H correlation has good agreement with the IBM simulation
Fig. 11. Particle orientation. results. At other solid fractions and Reynolds number, the F&H
L. He et al. / Powder Technology 313 (2017) 332–343 341

Fig. 12. Normalized drag force on each particle vs the particle orientation α at solid fraction of 0.1.

correlation has 5% to 24% difference with current results. Other than the fraction range from 0.1 to 0.35 and at low to moderate Reynolds num-
15% under prediction at Reynolds number of 10 and solid fraction of bers (10 ≤ Re ≤ 200). Widely used drag force correlations are evaluated
0.35, the T&H correlation agrees very well with the IBM simulation re- based on the current simulation results.
sults. The T&Z correlation has relatively good agreement at solid fraction The following conclusions can be drawn from this study:
of 0.1 while consistently under predicting the drag value by 9% to 22% at
• Drag force predicted by current simulations for assembly of spherical
higher solid fractions. Over all, among the three correlations tested here,
particles shows good agreement with previous study. The framework
T&H with 7.15% average difference gives the closest prediction com-
developed in this study is reliable to investigate the drag force for as-
pared to current results, and therefore, should be used for ellipsoidal
semblies of non-spherical particles.
particles in assembly if the particle shape is similar to the shape studied
• Drag force of ellipsoidal particle assemblies is 15% to 35% larger than
equal volume spherical particles at the same solid fraction and Reyn-
olds number.
6. Summary • The performance of available drag correlations applied to non-
spherical particles is examined. Correlations based only on particle
In this paper, particle-resolved simulations are performed to study equal volume sphere diameter (Deq) under predict the drag force.
the momentum transfer for non-spherical particles in assembly. A ran- Combined correlations using particle shape factors and particle
dom arrangement of ellipsoidal particles with sphericity (ψ = 0.887) is equal volume sphere diameter (Deq) also under predict the drag
used in the assembly and the particle-fluid boundaries are resolved force: the correlations give 2% to 35% smaller drag forces than the cur-
using the immersed boundary method. The ellipsoidal particles are rent particle resolved simulation results.
packed using a physics engine which provides random particle configu- • Combined correlations taking the particle orientation into account
rations with no overlap between particles. Incompressible Navier- show better agreement. The combined Tenneti et al. and Holzer corre-
Stokes equations are solved in the fluid with no slip boundary condition lation, with 7.15% average error, has the closest prediction capability
applied at the fluid-particle boundary. The fluid field is a periodic cubic with current simulation results.
domain. The pressure and viscous force on each particle are calculated
based on the resolved flow field. Flow through an assembly of spherical Nomenclature
particles is tested, and calculated drag forces are compared with previ- A area
ous particle resolved simulation results to validate the current frame- d distance
work. Then packed ellipsoidal particles are simulated in a solid Deq equal volume sphere diameter
342 L. He et al. / Powder Technology 313 (2017) 332–343

Fig. 13. Normalized drag force on each particle vs the particle orientation α at solid fraction of 0.35.

F normalized force Greek

L length
N number of particles α angle describing ellipsoidal particle orientation
n surface normal β mean pressure gradient constant
P, p pressure ρ density
Re Reynolds number μ viscosity
t time ϕ solid fraction
u streamwise interstitial velocity ψ sphericity
U superficial flow velocity τ shear stress

Table 3 Subscripts
Normalized mean drag force from current simulation compare to F&H, T&H and T&Z
e surface element
Re Φ IBM F&H % diff T&H % diff T&Z % diff p probe
10 10% 3.58 2.78 −22.51% 3.65 1.99% 3.49 −2.72%
50 5.82 5.66 −2.80% 6.26 7.47% 5.92 1.67% Superscripts
100 8.46 8.76 −3.58% 9.14 8.14% 8.29 −1.96%
200 14.10 14.45 2.53% 14.60 3.61% 12.32 −12.62% ∗
dimensional value
10 20% 6.87 4.39 −35.10% 6.57 −4.34% 6.30 −8.37%
50 11.13 8.95 −19.66% 10.43 −6.28% 9.88 −11.29%
100 15.81 13.66 −13.57% 14.74 −6.75% 13.41 −15.16% Miscellaneous
200 24.97 22.02 −11.78% 22.89 −8.30% 19.47 −22.01%
10 30% 13.14 7.46 −43.18% 11.98 −8.81% 11.57 −11.89% x , y, z Cartesian coordinates
50 20.20 15.38 −23.83% 18.38 −9.02% 17.56 −13.06%
100 27.58 23.29 −15.56% 25.59 −7.21% 23.63 −14.31%
200 42.82 36.77 −14.11% 39.34 −8.13% 34.27 −19.97% Acknowledgements
10 35% 19.38 10.85 −44.01% 16.41 −15.32% 15.99 −17.46%
50 26.83 21.42 −20.16% 25.02 −6.75% 24.20 −9.83% Danesh Tafti would like to acknowledge support from the National
100 36.59 34.39 −6.01% 34.81 −4.86% 32.66 −10.75% Energy Technology Laboratory (NETL) as ORISE Faculty. The authors ac-
200 57.76 60.34 4.46% 53.52 −7.34% 47.64 −17.53%
knowledge Advanced Research Computing at Virginia Tech for
L. He et al. / Powder Technology 313 (2017) 332–343 343

providing computational resources and technical support that have [22] D.K. Tafti, GenIDLEST - a scalable parallel computational tool for simulating complex
turbulent flows, Proc. ASME Fluids Eng. Div., ASME-IMECE, vol. 256, 2001.
contributed to the results reported within this paper. URL: http:// [23] D.K. Tafti, Time-accurate techniques for turbulent heat transfer analysis in complex geometries, advances in computational fluid dynamics and heat transfer, in: R.
Amano, B. Sunden (Eds.), Dev. Heat Transf, WIT Press, 2011.
[24] K. Nagendra, D.K. Tafti, K. Viswanath, A new approach for conjugate heat transfer
References problems using immersed boundary method for curvilinear grid based solvers, J.
Comput. Phys. 267 (2014) 225–246,
[1] H.N. Yow, M.J. Pitt, A.D. Salman, Drag correlations for particles of regular shape, Adv.
[25] L. He, D. Tafti, Evaluating the immersed boundary method in a ribbed duct for the
Powder Technol. 16 (2005) 363–372,
internal cooling of turbine blades, ASME Turbo Expo 2015 Turbine Tech. Conf.
[2] R.P. Chhabra, L. Agarwal, N.K. Sinha, Drag on non-spherical particles: an evaluation
Expo, American Society of Mechanical Engineers, Montreal, Quebec, Canada, 2015.
of available methods, Powder Technol. 101 (1999) 288–295,
[26] L. He, K. Joshi, D.K. Tafti, Study of fluid structure interaction using sharp interface im-
mersed boundary method, ASME 2016 Fluids Eng. Div. Summer Meet, American So-
[3] A. Haider, O. Levenspiel, Drag coefficient and terminal velocity of spherical and non-
ciety of Mechanical Engineers, Washington, DC, 2016.
spherical particles, Powder Technol. 58 (1989) 63–70,
[27] D.K. Tafti, GenIDLEST - a scalable parallel computational tool for simulating complex
turbulent flows, ASME-PUBLICATIONS-FED 256, American Society of Mechanical En-
[4] G.H. Ganser, A rational approach to drag prediction of spherical and nonspherical
gineers 2001, pp. 347–356.
particles, Powder Technol. 77 (1993) 143–152,
[28] S. Patankar, Fully developed flow and heat transfer in ducts having streamwise-
periodic variations of cross-sectional area, Trans. ASME 99 (1977) 180–186 http://
[5] S.-F. Chien, Settling velocity of irregularly shaped particles, SPE Drill. Complet. 9 (De-
cember 1994) 281–289,
(accessed October 20, 2014).
[6] M. Hartman, O. Trnka, K. Svoboda, Free settling of nonspherical particles, Ind. Eng.
[29] L. Zhang, D. Tafti, Computations of flow and heat transfer in parallel-plate fin heat
Chem. Res. (1994) 1979–1983,
exchangers on the CM-5: effects of flow unsteadiness and three-dimensionality,
[7] P. Swamee, C. Ojha, Drag coefficient and fall velocity of nonspherical particles, J.
Int. J. Heat Mass Transf. 40 (1997)
Hydraul. Eng. 117 (1991) 660–667.
pii/S0017931096002074 (accessed October 21, 2014).
[8] A. Hölzer, M. Sommerfeld, New simple correlation formula for the drag coefficient of
[30] A. Gilmanov, F. Sotiropoulos, A hybrid Cartesian/immersed boundary method for
non-spherical particles, Powder Technol. 184 (2008) 361–365,
simulating flows with 3D, geometrically complex, moving bodies, J. Comput. Phys.
207 (2005) 457–492,
[9] F. Dioguardi, D. Mele, A new shape dependent drag correlation formula for non-
[31] R.J. Hill, D.L. Koch, A.J.C. Ladd, The first effects of fluid inertia on flows in ordered and
spherical rough particles. Experiments and results, Powder Technol. 277 (2015)
random arrays of spheres, J. Fluid Mech. 448 (2001) 243–278,
[10] F. Dioguardi, P. Dellino, D. Mele, Integration of a new shape-dependent particle–
[32] S. Tenneti, R. Garg, S. Subramaniam, Drag law for monodisperse gas-solid systems
fluid drag coefficient law in the multiphase Eulerian–Lagrangian code MFIX-DEM,
using particle-resolved direct numerical simulation of flow past fixed assemblies
Powder Technol. 260 (2014) 68–77,
of spheres, Int. J. Multiphase Flow 37 (2011) 1072–1092,
[11] G. Bagheri, C. Bonadonna, On the drag of freely falling non-spherical particles, Pow-
[33] M.a. van der Hoef, R. Beetstra, J.A.M. Kuipers, Lattice-Boltzmann simulations of low-
der Technol. 301 (2016) 526–544,
Reynolds-number flow past mono- and bidisperse arrays of spheres: results for the
[12] D. Nemec, J. Levec, Flow through packed bed reactors: 1. Single-phase flow, Chem.
permeability and drag force, J. Fluid Mech. 528 (2005) 233–254,
Eng. Sci. 60 (2005) 6947–6957,
[13] S. Ergun, Fluid flow through packed columns, Chem. Eng. Prog. 48 (1952) 89–94.
[34] R. Beetstra, M.A. Van Der Hoef, J.A.M. Kuipers, Drag force of intermediate Reynolds
[14] R.K. Niven, Physical insight into the Ergun and Wen and Yu equations for fluid flow
number flow past mono- and bidisperse arrays of spheres, AICHE J. 53 (2007)
in packed and fluidised beds, Chem. Eng. Sci. 57 (2002) 527–534,
[35] H. Wadell, Volume, shape, and roundness of quartz particles, J. Geol. 95 (2015)
[15] D.L. Koch, R.J. Hill, Inertial effects in suspension and porous -media flows, Annu. Rev.
Fluid Mech. (2001) 619–647,
[36] J. Gan, Z. Zhou, A. Yu, CFD-DEM modeling of gas fluidization of fine ellipsoidal par-
[16] I. Machač, V. Dolejš, Flow of generalized newtonian liquids through fixed beds of
ticles, AICHE J. 62 (2016) 62–77,
nonspherical particles, Chem. Eng. Sci. 36 (1981) 1679–1686,
[37] M. Zastawny, G. Mallouppas, F. Zhao, B. van Wachem, Derivation of drag and lift
force and torque coefficients for non-spherical particles in flows, Int. J. Multiphase
[17] L. Hua, H. Zhao, J. Li, J. Wang, Q. Zhu, Eulerian–Eulerian simulation of irregular par-
Flow 39 (2012) 227–239,
ticles in dense gas–solid fluidized beds, Powder Technol. 284 (2015) 299–311,
[38] Nividia, NVIDIA PhysX SDK 3.3.4 documentation,
gameworks/content/gameworkslibrary/physx/guide/Manual/Index.html .
[18] F. Dorai, M. Rolland, A. Wachs, M. Marcoux, E. Climent, Packing fixed bed reactors
[39] J.M. DallaValle, Micromeritics: The Technology of Fine Particles, Pitman Pub.Corp,
with cylinders: influence of particle length distribution, 20th Int. Congr. Chem. Pro-
New York, 1948.
cess Eng. CHISA 2012, pp. 1335–1345,
[40] C.Y. Wen, Y.H. Yu, A generalized method for predicting the minimum fluidization
velocity, AICHE J. 12 (1966) 610–612,
[19] Z.Y. Zhou, D. Pinson, R.P. Zou, A.B. Yu, Discrete particle simulation of gas fluidization
[41] R. Turton, O. Levenspiel, A short note on the drag correlation for spheres, Powder
of ellipsoidal particles, Chem. Eng. Sci. 66 (2011) 6128–6145,
Technol. 47 (1986) 83–86,
[42] M. Syamlal, T.J. O'Brien, A Generalized Drag Correlation for Multiparticle Systems,
[20] K. Vollmari, T. Oschmann, S. Wirtz, H. Kruggel-Emden, Pressure drop investigations
Morgantown, 1987.
in packings of arbitrary shaped particles, Powder Technol. 271 (2015) 109–124,
[43] J.F. Richardson, W.N. Zaki, Sedimentation and fluidisation: part I, Trans. Inst. Chem.
Eng. 32 (1954) S82–S100,
[21] R. Di Felice, The voidage function for fluid-particle interaction systems, Int. J. Multi-
phase Flow 20 (1994) 153–159,