fluids

Article
A Comparison of Ansys Fluent and MFiX in Performing
CFD-DEM Simulations of a Spouted Bed
Filippo Marchelli * and Renzo Di Felice

Department of Civil, Chemical and Environmental Engineering (DICCA), University of Genova,

16145 Genova, Italy; [email protected]
* Correspondence: [email protected]

Abstract: The CFD-DEM methodology is a popular tool for the study of fluid–particle systems, and
there are several programs that permit using it. In this study, we employed it to simulate a pseudo-2D
spouted bed, comparing the performance of the programs Ansys Fluent and MFiX. The results
are analysed and commented on in terms of both accuracy and computational efforts. Despite the
similarity of the setup, MFiX seems to perform significantly better. The similarities and differences
between the two programs are discussed in detail, offering useful insights to researchers regarding
the selection of one over the other, depending on the application. The better suitability of the Di
Felice drag model is confirmed for the device, while it is shown that the effect of the Magnus lift force
may be more limited than was shown in a previous study.

Keywords: gas–solid fluidization; Eulerian–Lagrangian method; drag model; parallelization; Geldart-D;

fountain height







Citation: Marchelli, F.; Di Felice, R. A

Comparison of Ansys Fluent and 1. Introduction
MFiX in Performing CFD-DEM Complex numerical simulations have seen an intense growth in the last two decades,
Simulations of a Spouted Bed. Fluids thanks to the release of better-performing and widely accessible computational resources.
2021, 6, 382. https://doi.org/
Even without expensive supercomputers, researchers can provide meaningful contributions
10.3390/fluids6110382
in areas with notable open questions, such as turbulence, molecular dynamics, climate
science, etc. Among these, the behaviour of granular materials is a very popular research
Academic Editors: Federico Piscaglia
field. Despite the numerous efforts of researchers, a deep understanding of many features
and Jérôme Hélie
of these flows is still lacking [1]. Several numerical techniques have been developed
to reproduce the behaviour of granular flows at various scales, from the very detailed
Received: 28 September 2021
Accepted: 20 October 2021
but complex direct numerical simulations to lighter methods based on the continuum
Published: 25 October 2021
hypothesis [2,3]. In this framework, the combination of computational fluid dynamics
(CFD) and the discrete element method (DEM), widely known as CFD-DEM, represents a
Publisher’s Note: MDPI stays neutral
good compromise between computational complexity and accuracy [4].
with regard to jurisdictional claims in
CFD-DEM, which is a Eulerian–Lagrangian approach, simulates particles according
published maps and institutional affil- to the discrete element method, originally proposed in 1979 by Cundall and Strack [5]. The
iations. method treats each physical particle (or group of particles) as a computational particle,
with the collisions among them directly resolved. DEM was first coupled with CFD in 1992
thanks to the work of Tsuji and colleagues [6], thus allowing the simulation of multiphase
flows. CFD-DEM represents a suitable option when the number of particles is not excessive
Copyright: © 2021 by the authors.
and is generally deemed as more reliable than continuum methods, also providing more
Licensee MDPI, Basel, Switzerland.
detailed results [7–9]. One of its drawbacks is that the computational grid must be larger
This article is an open access article
than the particles, thus possibly leading to inaccuracies in predicting the behaviour of the
distributed under the terms and fluid phase [10]. Due to this requirement, the method is labelled as “unresolved”. The
conditions of the Creative Commons interaction between the fluid and solid phase(s) relies on models that estimate the drag
Attribution (CC BY) license (https:// force based on experimental or numerical techniques. Several models exist in the literature,
creativecommons.org/licenses/by/ but researchers disagree on the ideal choice, and none of them appear to be suitable in all
4.0/). operative conditions [11].

Fluids 2021, 6, 382. https://doi.org/10.3390/fluids6110382 https://www.mdpi.com/journal/fluids

Fluids 2021, 6, 382 2 of 15

CFD-DEM has become particularly common in the last decade. According to the
Scopus database (www.scopus.com, accessed on 18 October 2021), the first scientific article
that features the expression “CFD-DEM” in its title, abstract, or keywords dates back to
2003 [12]; 70 articles featured it (or a variation) in 2013, 125 in 2016, and 292 in 2020. This
growth is directly ascribable not only to the enhancement of computational resources, but
also to the release of several accessible programs. These allowed researchers to perform
such simulations without having to develop a whole in-house code, on which research
works were originally (and sometimes still are) based.
Programs can be categorised as open-source or commercial. Open-source programs
have some clear advantages, such as being free and completely editable and transparent to
the user, thus allowing more flexibility. However, the process of learning to use them at
their full potential can be cumbersome, as sometimes they do not have a GUI (graphical
user interface), require good coding skills and have limited support and tutorials [13]. On
the other hand, commercial programs are usually not free, and their code is not completely
visible or editable, but they may have more in-built features and better support resources.
Commercial programs include Rocky and EDEM, whereas open-source programs include
LIGGGHTS and Yade, among others.
This work focuses on two well-known programs that permit performing CFD-DEM
simulations: MFiX and Ansys Fluent. MFiX (Multiphase Flow with interphase eXchanges,
mfix.netl.doe.gov, accessed on 18 October 2021) is an open-source program developed by
the National Energy Technology Laboratory (NETL), part of the Department of Energy of
the USA, which allows simulating granular materials through different methodologies.
Conversely, Ansys Fluent (www.ansys.com/products/fluids/ansys-fluent, accessed on
18 October 2021) is a commercial program that allows performing a wide variety of CFD
simulations. It is part of the Ansys package, which includes numerous other programs.
Fluent can also be employed to perform the CFD calculations only, with the DEM ones
being solved by a coupled commercial program (such as the aforementioned Rocky and
EDEM). Despite their different licence, MFiX and Fluent can be considered among the most
accessible programs for someone who wants to start performing CFD-DEM simulations:
they both have a user-friendly GUI and are available for early-stage researchers either as
an open-source program (MFiX), or through the student licence (Fluent). For this reason,
they were chosen for the present analysis.
The two programs are compared for CFD-DEM simulations, focusing on the simi-
larities and differences in preparing and personalising the simulation setup and in their
accuracy and efficiency. The comparison is based on experimental data of a pseudo-2D
spouted bed [14]. Spouted beds are a particular kind of fluidised beds, particularly suitable
to be studied through CFD-DEM simulations as they typically work with relatively large
Geldart-D [15] particles. Moreover, there are some open research questions regarding
their fluid dynamic regime, despite the great number of studies focusing on these devices,
including through CFD-DEM simulations as detailed in [16].
To the best of our knowledge, there is no existing study comparing CFD-DEM sim-
ulations performed with different codes, whereas for the Eulerian–Eulerian approach,
Herzog et al. [17] compared MFiX, OpenFoam and Fluent, and Venier et al. [18] com-
pared OpenFoam and Fluent, in both cases taking fluidised beds into account. The first
study (published in 2012) concluded that MFiX and Fluent were almost equivalently ac-
curate, with OpenFoam deemed as not mature enough yet. Conversely, the second study
(published in 2019), stated that the two programs are almost equivalent in their accuracy,
pointing out that both still have notable room for improvement.

2. Materials and Methods

2.1. Governing Equations
The simulation approach is based on the combination of CFD and DEM. The equations
of the two methodologies, including their implementation in Fluent and MFiX, have been
reported multiple times in the literature, and for the sake of conciseness they are not
Fluids 2021, 6, 382 3 of 15

fully repeated here. Interested readers are referred to previous publications [19,20] or the
official guides of the two programs [21–23]. This section discusses the main similarities
and differences in the implementation of the methods.
Regarding the CFD part, both programs employ the equations of local balance for
mass and momentum for the fluid phase. Fluent employs the so-called “model A”, whereas
MFiX allows choosing between A and B [24]. In this work, model A was employed in
both programs. The programs also offer the possibility of turning off the fluid solver
and performing pure DEM simulations. However, in that case, Fluent’s performance is
dubious [25], which is expected since its original aim is to perform fluid dynamics studies.
Coherently with its versatility, Fluent offers a wide variety of models to consider
turbulence. Conversely, MFiX only has two turbulence models in its code (the k-epsilon
and the L-scale models). Here, the k-epsilon model [26] was selected. Despite the fact it
complicates the computational setup and is sometimes neglected when simulating spouted
beds [16], turbulence can have an impact on CFD-DEM simulations [27] and it should be
included for the sake of accuracy.
Regarding the DEM part, both programs solve the Newtonian equations of motion
for all particles. While MFiX automatically solves both the translational and rotational
equations, in Fluent the rotational equation has to be explicitly activated by the user.
There are also some differences regarding the calculation of some forces that the particles
experience. These are detailed below.

2.1.1. Pressure Gradient Force

Particles can perceive an acceleration caused by the fluid phase pressure gradient.
This is usually quite small compared to other forces, especially when the fluid phase is
gaseous. In Fluent, the force Fpg is given as:
→ ρf → →
F pg = m p u p∇ u (1)
ρp

In which mp is the particle mass, ρf and ρp are the fluid and particle density, and up
and u are the particle and fluid velocity. In this equation, the pressure gradient is replaced
by the velocity gradient based on a steady-state approach of the momentum equation,
neglecting diffusive and source terms.
In MFiX, the force is instead calculated as:
→
F pg = −Vp,i ∇ P (2)

in which Vp is the particle volume and P is the fluid pressure.

2.1.2. Drag Force

Both programs offer the possibility of choosing among several models to calculate the
fluid–solid momentum exchange. Moreover, further models can be implemented through
user-defined functions (UDF). These must be written in the C language for Fluent and in
FORTRAN for MFiX. More information on implementing UDF drag models in Fluent is
provided in a previous article [11]. In this work, we considered the Gidaspow [28] and Di
Felice [29] drag models. For the Gidaspow model the force is thus calculated in Fluent:
→ µ → → 
F d = mp u − up D (3)
ρ p d2p

→ →
ρ f dp u p − u
Re p = ε f (4)
µf
Fluids 2021, 6, 382 4 of 15

 
 18 ε−f 3.65 1 + 0.15 Re p 0.687


f or ε s ≤ 0.2
D= Re (5)
 150 εε2s + 1.75 ε2p f or ε s > 0.2
f f

In MFiX, the force is calculated as:

→ β i Vp,i → → 
Fd,i = u − up (6)
εs
→ →

 150(1−ε f )µ f 1.75ρ f u − u p
+ i f ε f ≤ 0.8


ε f d2p dp
βi = → → (7)
 0.75CD u − u p ε f ρ f ε−f 2.65
i f ε f > 0.8


di
(  
24
Re p 1 + 0.15Re0.687
p i f Re p < 1000
CD = (8)
0.44 i f Re p ≥ 1000
The procedure is analogous for the Di Felice drag model, with equations for D and βi
that are as follows:
D = 0.75CD Re p ε f 1−χ (9)
→ →
β i = 0.75CD u − u p ρ f ε f 2−χ /d p (10)

2 
χ = 3.7 − 0.65 exp −0.5 1.5 − log10 Re (11)

in which CD is calculated as per Equation (8). In the above equations, µ is the fluid viscosity,
dp the particle diameter, εf and εs the fluid and solid volume fraction.
The Gidaspow drag model is actually a mere combination of the Ergun [30] and
Wen-Yu [31] models with a voidage-based switch. It is one of the most employed drag
models, despite its unphysical discontinuity at εf = 0.8 and the doubts about its accuracy for
certain applications. The Di Felice drag model was instead proposed in 1994 [29] through
physical considerations and comparison with experimental pressure drop data. It has
seen large usage in the literature, and a partial update was proposed in 2013 by Rong and
colleagues through Lattice-Boltzmann simulations [32].
Both programs employ cell values for the porosity and fluid phase velocity by default.
In MFiX, it is possible to change this and make the program interpolate these variables
to the particle’s position; we employed this interpolation in the current simulations. An
identical option is not available in Fluent, but it is possible to enable a “high-resolution
tracking” algorithm that divides the cells in tetrahedrons to achieve more realistic values
of velocity and porosity around the particle. From preliminary simulations, it appears that
this option affects the results the most when a Cartesian mesh is employed.

2.1.3. Magnus Lift Force

When a particle rotates in a fluid, it can experience a lift force caused by the pres-
sure differential along its surface. Although this force is often neglected in CFD-DEM
simulations, a previous analysis performed with Fluent [33] showed that it can have a
noteworthy influence on the particles’ motion in this spouted bed, and according to other
researchers [34] it can become even larger than the drag force for high values of Reynolds
number and porosity.
Fluent offers the possibility of including this force through three models; in the
previous analysis [33], the model by Rubinow and Keller [35] appeared to be the most
suitable for the current setup, and is thus also included in this work. Since MFiX has no
Fluids 2021, 6, 382 5 of 15

in-built option to estimate it, we included the same model by modifying its code. The force
has this formulation:
→ →
→ 1 up − u 
→ →
 1 → →

F ml = A P CRL ρ f → → up − u × ∇ × u − ωp (12)
2 1 2
2 ∇ × u − ωp

→
ω p dp
CRL = → → (13)
up − u

in which Ap is the particle surface and ω p the particle rotational velocity.

2.1.4. Contact Forces

Particle–particle and particle–wall collisions are calculated through the well-known
spring-dashpot model in both programs. The model lets particles slightly overlap during
collisions and calculates the magnitude of the collision force as a function of the overlap.
The elastic part of the collision is calculated considering a fictitious spring, while the
damping effect is based on a fictitious dashpot. The model requires the specification of a
spring constant and a restitution coefficient for each collision pair. Details and equations
on the spring dashpot models are reported in the program guides [21–23] and in the
literature [19,20], and are not repeated here for the sake of brevity. Both programs also offer
the possibility of employing the more complex Hertzian approach for the same purpose. In
fluidised beds, however, the Hertzian model notably slows down the simulation without
providing any benefit [36].
The friction force is estimated through Coulomb’s approach in both programs (see the
previous references for the complete equations). The approach provides a limit value of the
tangential force and requires the specification of the friction coefficient. In Fluent, it is also
possible to employ a friction coefficient that varies as a function of the particle–particle (or
particle–wall) relative velocity. Given the lack of experimental data, this option was not
employed in this work.
Fluent allows including the rolling friction torque, while the standard version of MFiX
does not have an option to calculate it. We decided to modify the code of MFiX to include
it, in this form [37,38]:
→
→ ωr ri r j →
Mr = − → µr Fn (14)
ω ri + r j
r

in which ω r is the particle–particle (or particle–wall) relative rotational velocity, µr the

rolling friction coefficient, ri and rj the radii of the two contacting particles, and Fn the
normal contact force.

2.2. Simulation Procedure

2.2.1. Simulated Geometry and Data
The simulations are based on the experimental work by Zhao et al. [14], which has
already been used by several authors to validate CFD-DEM simulations. The unit is a
pseudo-2D spouted bed, whose dimensions are summarised in Table 1. The comparison is
performed employing the time-averaged profile of the particles’ vertical velocity along the
central axis of the spouted bed.
Fluids 2021,
Fluids 6, x6,FOR
2021, 382 PEER REVIEW 6 of6 16
of 15

Table 1. Dimensions of the geometry.

Table 1. Dimensions of the geometry.
Variable Value
Variable Value
Inlet width 0.9 cm
BottomInletwidth
width 1.50.9
cmcm
Bottom width 1.5 cm
Column
Column width
width 15.2 cmcm
15.2
Column
Column depth
depth 1.51.5
cmcm
BaseBase angle
angle 60°60◦
Height
Height of the
of the inclinedpart
inclined part 11.9
11.9 cmcm
Total height 60 cm

2.2.2. Mesh
2.2.2. Mesh
The discretisation of the geometry is often a critical aspect in CFD-DEM simulations.
The one
In general, discretisation
should alwaysof thesearch
geometry
for aiscompromise
often a critical
soaspect in CFD-DEM
as to obtain simulations.
a satisfying descrip- In
general, one should always search for a compromise so as to obtain a satisfying
tion of the fluid phase without attaining unrealistic values of the porosity. Specific re- description
of thehas
search fluid phase
shown without
that a ratioattaining
betweenunrealistic
the cell and values of the
particle porosity.
sizes of about Specific research
4 is the best
choice [39], although sometimes even smaller values are successfully employed [40]. [39],
has shown that a ratio between the cell and particle sizes of about 4 is the best choice
although sometimes
The process even smaller
of creating values are successfully
the computational grid notably employed [40].
differs between Fluent and
The process of creating the computational grid notably
MFiX. For Fluent, the mesh must be generated externally. The most common differs between Fluent
option and
is to
MFiX. For Fluent, the mesh must be generated externally. The most common option is to
generate it with the program Ansys Meshing, which has a wide variety of functionalities
generate it with the program Ansys Meshing, which has a wide variety of functionalities
to create very different grids. Conversely, the MFiX mesh must be specified within the
to create very different grids. Conversely, the MFiX mesh must be specified within the
program, and only a Cartesian cut-cell algorithm [41] can be employed. Since MFiX does
program, and only a Cartesian cut-cell algorithm [41] can be employed. Since MFiX
not support importing an external mesh, and the meshes generated in MFiX are incom-
does not support importing an external mesh, and the meshes generated in MFiX are
patible with Fluent, it was not possible to employ identical grids for the simulations.
incompatible with Fluent, it was not possible to employ identical grids for the simulations.
In a previous analysis performed with Fluent [33], the best performing mesh was
In a previous analysis performed with Fluent [33], the best performing mesh was
found to be a structured one with 3 cells along the depth, 62 along the height and 20 along
found to be a structured one with 3 cells along the depth, 62 along the height and 20 along
the width (12 in the tube), with an average cell-to-particle size ratio of about 4 in the par-
the width (12 in the tube), with an average cell-to-particle size ratio of about 4 in the
allelepipedal part. Such a grid is employed in this work as well, but to make a better com-
parallelepipedal part. Such a grid is employed in this work as well, but to make a better
parison with MFiX a Cartesian grid was also created and tested. An unstructured grid
comparison with MFiX a Cartesian grid was also created and tested. An unstructured
was also
grid preliminarily
was tested but
also preliminarily wasbut
tested discarded due to its
was discarded clearly
due to itsworse
clearlyresults.
worse The MFiX
results. The
grid is instead
MFiX grid is ainstead
Cartesian grid with
a Cartesian 3 cells
grid with along
3 cells the
alongdepth, 65 along
the depth, the height
65 along and 19
the height and
along the width.
19 along The grid-to-particle
the width. size ratio
The grid-to-particle size is 4 along
ratio the width
is 4 along and height,
the width and 2.5
and height, and
along the depth.
2.5 along Significantly
the depth. finer grids
Significantly finerprevented the MFiX
grids prevented the solver
MFiX from
solverreaching con-
from reaching
vergence. The three employed grids are shown in
convergence. The three employed grids are shown in Figure 1.Figure 1.

Figure 1. The three employed meshes. Please note that the upper portion of the mesh is not shown
Figure
as it 1. Therelevant
is not three employed meshes. Please
for the comparison, andnote
thatthat the upper of
the diagonals portion of the mesh
the squares in the isMFiX
not shown
mesh are
as only
it is not relevant
aesthetic. for the comparison, and that the diagonals of the squares in the MFiX mesh are
only aesthetic.
Fluids 2021, 6, 382 7 of 15

As Figure 1 clearly shows, the Cartesian grids allow a more regular cell volume and
distribution, but do not offer a detailed description of the fluid phase in certain parts,
such as near the inlet. This is instead achieved through the structured grid, but at the
expense of having several cells whose size approaches that of particles. To avoid numerical
instabilities and unphysical porosity values, this grid can only be employed together with
an averaging procedure that is discussed below.

2.2.3. Simulation Setup

The simulations regarded a case with an initial bed height of 0.1 m, and an inlet
velocity of 26.68 m/s. The parameters employed in the simulations are summarised in
Table 2. It shall be noted that in Fluent it is possible to input only a single value of the
collision parameters per each collision pair, whereas in MFiX different values for the normal
and tangential collision parameters may selected. In this work, we employed the standard
values of 2/7 and 1/2 for the ratios between the normal and tangential values of the spring
constant and restitution coefficient.

Table 2. Parameters employed in the simulations.

Variable Value
Spring constant 1000 N/m
Restitution coefficient 0.9
Friction coefficient 0.3
Rolling friction coefficient 0.03
Air density 1.225 kg/m3
Air dynamic viscosity 1.7894·10−5 Pa·s
Particle density 2380 kg/m3
Particle diameter 2.033 mm
Total number of particles 15,990

Both Fluent and MFiX allow for reducing the number of particles by lumping several
of them in a larger “parcel”; this approach is usually known as “coarse-graining” [42,43].
As the number of involved particles is already quite low, this procedure was not employed
here. It should also be noted that using larger parcels requires enlarging the mesh as well.
In MFiX, the particle time step is automatically calculated by the program as 1/50 of
the minimum collision time. In the current setup, this procedure yielded a value of about
4.44·10−6 s. The same value was also selected in Fluent for the sake of similarity, although
previously a larger value had been successfully used [33]. For the CFD calculations, MFIX
utilises an adaptive time step. In this work, it could vary between 10−7 and 10−3 s, although
values close to the upper boundary were almost always used by the program after the first
few time steps. In Fluent, the CFD time step was instead set to 10−4 s.
The simulations were run on the 21.2.0 version of Fluent, which is incorporated in
the 2021R2 release of Ansys. For MFiX, the 21.1.4 version was employed. The simulations
were run on Windows 10 on a PC equipped with an Intel(R) Xeon(R) Gold 5218 processor
(produced by Intel, Santa Clara, CA, USA), with 16 cores.

2.2.4. Particle Data Averaging

When the size of the grid cells approaches that of particles, numerical instability or
unphysical results may appear. To avoid these, CFD-DEM may take advantage of averaging
algorithms that distribute the effect of the presence of a particle to nearby cells. When
these methods are not employed, the presence of a particle is only taken into account in the
cell in which its centre lies, which can be especially inadequate when a particle lies on the
border between two or more cells.
For this purpose, Fluent employs a so-called “node-based averaging” algorithm,
which is based on a Gaussian function. This algorithm was employed in all simulations,
since it was the only way to achieve convergence. In MFiX, two interpolation schemes can
Fluids 2021, 6, 382 8 of 15

be employed: one based on the grid size, and one based on a user-specified interpolation
width, which is recommended to be set equal to the particle size. The latter was employed
in the simulations, although an analysis of the results showed the effects of its inclusion
were rather negligible.

2.2.5. Result Extraction and Post Processing

Extracting and analysing the results is a key task in numerical simulation. Although
Fluent offers more in-built functionalities to do this, it is quite difficult to achieve something
beyond. For example, Fluent automatically stores several data regarding both fluid and
particles, but it is impossible to access the values of most of the forces that act on particles
(such as drag and friction). On the other hand, in MFiX the grid values of the particles’
velocity are not accessible by default, but through a modification of the code, these can
be saved (one of the most common ways to achieve it is to store them as “reaction rates”).
Moreover, most of the data extraction and visualisation cannot be performed within the
program itself, but through the command-line tool “postmfix” or through the open-source
program Paraview.
Regarding Fluent, the time-averaged data were obtained through the in-built option
“data sampling for time statistics”. The algorithm was automatically activated after 1 s of
simulation through a command. MFiX offers no in-built option to obtain time-averaged
data, but there are two ways to achieve them. One is to create a user-defined subroutine
that performs this task, while the other is to do it at the end of the simulation in Paraview,
using the saved data. The second method is clearly more immediate, but the output relies
on the frequency at which data are saved, whereas through the first method, data can be
averaged at each time step. Hence, the first method was employed.

3. Results and Discussion

The performances of the two programs are compared on the basis of the results and
computational demands. Finally, their advantages and disadvantages are summarised.

3.1. Comparison of Results

The comparison between the two programs was performed employing specifications
that were identical or as similar as possible. As already mentioned, the most outstanding
difference lies in the numerical mesh, which cannot be set as identical. Although in a
previous analysis a structured mesh appeared to be the best choice, a Cartesian mesh was
also tested to employ a more similar grid with respect to MFiX.
The first simulations regarded setups in which we did not employ any user-defined
function, to show the standard capabilities of the two programs. The drag is estimated
through the Gidaspow model, and the rolling friction and Magnus lift forces are not enabled
in MFiX. Figure 2 depicts the results visually. Qualitatively, both programs succeed in
providing a good representation of the experimental configuration of the particles. The
three typical zones of spouted beds are clearly visible: the spout, which is the central
channel in which particles travel rapidly upwards; the fountain, which is the dispersed
zone in which particles are scattered while falling back; the annulus, which is the packed
zone where particles move slowly downwards and centre-wards.
For a more quantitative assessment, Figure 3 provides the time-averaged vertical
profiles of the particles’ vertical velocity. From the plot, both programs do not provide
an accurate reproduction of the experimental data. This is not unexpected since, as per
a previous work [11], the Gidaspow drag model is rather unsuitable for this setup, even
though it was found to be the most chosen in the literature [16]. Fluent seems to moderately
underestimate the peak velocity, while it largely overestimates the fountain height (i.e.,
the highest position reached by particles). Conversely, MFiX overestimates both the peak
velocity and the fountain height, but the latter is closer to the experimental value. Another
interesting aspect is that both MFiX and Fluent with the structured mesh correctly capture
the abrupt acceleration that particles experience near the bottom of the bed. Fluent with
 using the saved data. The second method is clearly more immediate, but the output relies
Fluids 2021, 6, x FOR PEER REVIEW 9 of 16
on the frequency at which data are saved, whereas through the first method, data can be
averaged at each time step. Hence, the first method was employed.

Figure 2. Configuration
3. Results of the particles after 2 s of simulation with the Gidaspow drag model. The
and Discussion
Fluids 2021, 6, 382 9 of 15
particles are coloured according to Table 14. Copyright 2008 Elsevier.
The performances of the two programs are compared on the basis of the results and
computational demands. Finally, their advantages and disadvantages are summarised.
The first simulations regarded setups in which we did not employ any user-defined
function,
the to show
3.1.Cartesian
Comparison mesh the standard
instead
of Results showscapabilities
a more gradual of theand
twoless
programs.
accurateThe drag is estimated
acceleration, with the
through the Gidaspow
peak velocity only reachedmodel, and
near thethe rolling
end of thefriction
spout.and
ThisMagnus
can belift forces are
explained nottoena-
due the
The comparison between the two programs was performed employing specifications
bled
higher in porosity
MFiX. Figure
values 2 depicts the results
that a coarser mesh visually. Qualitatively,
generates, bothaprograms
thus yielding less intensesucceed
drag
that were identical or as similar as possible. As already mentioned, the most outstanding
in providing
effect. a good
Since the representation
Cartesian Fluent meshof the
andexperimental
the MFiX mesh configuration of the and
are quite similar, particles. The
the forces
difference lies in the numerical mesh, which cannot be set as identical. Although in a pre-
three typical zones
are calculated of spouted
in a very similarbeds
way,are
thisclearly visible:
is a strong the
hint ofspout, which is the
the differences central
between thechan-
two
vious analysis a structured mesh appeared to be the best choice, a Cartesian mesh was
nel in which
codes, particles
especially on thetravel rapidly
numerical upwards; the fountain, which is the dispersed zone
side.
also tested to employ a more similar grid with respect to MFiX.
in which particles are scattered while falling back; the annulus, which is the packed zone
where particles move slowly downwards and centre-wards.
For a more quantitative assessment, Figure 3 provides the time-averaged vertical pro-
files of the particles’ vertical velocity. From the plot, both programs do not provide an
accurate reproduction of the experimental data. This is not unexpected since, as per a pre-
vious work [11], the Gidaspow drag model is rather unsuitable for this setup, even though
it was found to be the most chosen in the literature [16]. Fluent seems to moderately un-
derestimate the peak velocity, while it largely overestimates the fountain height (i.e., the
highest position reached by particles). Conversely, MFiX overestimates both the peak ve-
locity and the fountain height, but the latter is closer to the experimental value. Another
interesting aspect is that both MFiX and Fluent with the structured mesh correctly capture
the abrupt acceleration that particles experience near the bottom of the bed. Fluent with
the Cartesian mesh instead shows a more gradual and less accurate acceleration, with the
peak velocity only reached near the end of the spout. This can be explained due to the
higher porosity values that a coarser mesh generates, thus yielding a less intense drag
effect.
Figure 2. Configuration of the Sinceafter
particles the Cartesian Fluentwith
2 s of simulation meshtheand the MFiX
Gidaspow dragmesh areThe
model. quite similar,
particles areand the forces
coloured
according to their velocity magnitude. Experimental snapshot reprinted from [14], Copyright (2008), with permission from the
are calculated in a very similar way, this is a strong hint of the differences between
Elsevier. two codes, especially on the numerical side.

1.25
Particle average vertical velocity (m/s)

Fluent - structured mesh

Fluent - Cartesian mesh
1
MFiX
Experimental
0.75

0.5

0.25

0
0 0.05 0.1 0.15 0.2 0.25
y (m)
Figure 3.
3. Time-averaged
Time-averagedvertical
verticalprofiles
profilesofof
thethe
particles’ vertical
particles’ velocity
vertical when
velocity employing
when the the
employing
Gidaspow drag model.

code, as already stated, does not include rolling friction and Magnus
The MFiX code, Magnus lift,
which are
are selectable
selectableininFluent.
Fluent.Since
Sinceaaprevious
previousanalysis
analysispointed
pointed towards
towardsthethe
importance
importance of
including
of includingthethe
Magnus
Magnuslift lift
force, andand
force, for the
for sake of higher
the sake accuracy,
of higher we modified
accuracy, the MFiX
we modified the
code tocode
MFiX addto both
addvariables to the particles’
both variables force and
to the particles’ torque
force and balance. Figure 4Figure
torque balance. provides the
4 pro-
obtained time-averaged profiles of the particles’ vertical velocity. For the rolling
vides the obtained time-averaged profiles of the particles’ vertical velocity. For the rolling friction
torque, the
friction results
torque, thedo not appear
results do notto be strongly
appear affected,affected,
to be strongly and the velocity
and the values
velocityare only
values
marginally lowered. This is in line with the results of the previous analysis, although the
reduction appeared to be more marked [33]. In general, particles do not experience very
large relative rotational velocities, and thus the values yielded by Equation (13) are not
expected to have a huge impact on the particles’ trajectories.
 dicts the results of the aforementioned study, which highlighted that the inclusion of the
Magnus lift force had a strong impact on the shape of the velocity profiles. This incon-
sistency points to one of the major limits of Fluent, i.e., the lack of transparency of its code
and the impossibility to access the values of some calculated variables. The Magnus lift
Fluids 2021, 6, 382 10 of 15
force can be enabled, but the value of the force can only be inferred by its effects on the
particles’ trajectories.

1.25
Original

Particle average vertical velocity (m/s)

1 + rolling friction

+ rolling friction and

0.75 Magnus lift
Experimental

0.5

0.25

0
0 0.05 0.1 0.15 0.2 0.25
y (m)
Figure 4. Time-averaged
Time-averaged vertical
vertical profiles
profiles of
of the
the particles’
particles’ vertical
vertical velocity
velocity in
in MFiX,
MFiX,including
includingaddi-
addi-
tional forces.

With regard
Finally, to the Magnus
we repeated lift force, our
the simulations implementation
with the Di Felice of the model,
drag Rubinow–Keller model
with the aim of
also appears to have a rather negligible effect on the results. This, however,
checking whether the two programs responded analogously to such a modification. In contradicts thea
results of analysis,
previous the aforementioned study, appeared
this drag model which highlighted
to yield that the inclusion
the best results inofthethepresent
Magnuscon-
lift
force had a Figure
figuration. strong 5impact on the
provides theshape
visualofresults.
the velocity profiles. the
Qualitatively, Thisobtained
inconsistency points to
configurations
oneagain
are of theall
major limits ofThe
acceptable. Fluent, i.e., the
fountain lack of
height is transparency of its code
noticeably reduced and
in all the impossibil-
cases, becoming
ity to access the values of some
more similar to the experimental one. calculated variables. The Magnus lift force can be enabled,
but the value of the force can only be inferred by its effects on the particles’ trajectories.
Finally, we repeated the simulations with the Di Felice drag model, with the aim of
checking whether the two programs responded analogously to such a modification. In a
previous analysis, this drag model appeared to yield the best results in the present configu-
ration. Figure 5 provides the visual results. Qualitatively, the obtained configurations are
again all acceptable. The fountain height is noticeably reduced in all cases, becoming more
similar to the experimental one.
If we now observe the velocity profiles, reported in Figure 6, we notice that the one
obtained with MFiX is almost identical to the experimental one, thus clearly representing
the best choice. The profile obtained with the structured mesh in Fluent is again better
than the Gidaspow case, but the peak velocity is quite underestimated, whereas the one
obtained with the Cartesian mesh yields a good reproduction of the fountain height, but
a very poor one of the overall velocity. In general, both programs confirm that the Di
Felice model, compared to the Gidaspow one, results in a markedly lower drag force in
this configuration, thus lowering the particle velocity and fountain height. It thus is further
confirmed to be the best choice to simulate the current setup. These results also confirm
the accuracy of the computational frameworks, since the equation of the Di Felice model
generally yields lower values than the Gidaspow model [11].

3.2. Comparison of Calculation Time

Accuracy is an indispensable factor when choosing a simulation program. However,
given how CFD-DEM simulations can be demanding from a computational point of view,
the computational demand should never be overlooked. Regarding this, our observations
point largely in favour of MFiX. To perform the simulations described in the previous
section, which featured nearly identical setups and numerical parameters, MFiX required
about 60 min of calculation per 1 s of simulation, whereas Fluent needed about 1050 min.
Given that the large majority of the calculation time was employed by both programs to
perform the DEM calculations, it may be concluded that the particle-related algorithm of
MFiX is more efficient. It should be noted that, in a previous work, this same spouted bed
 0 0.05 0.1 0.15 0.2 0.25
y (m)
Figure 4. Time-averaged vertical profiles of the particles’ vertical velocity in MFiX, including addi-
tional forces.
Fluids 2021, 6, 382 11 of 15
Finally, we repeated the simulations with the Di Felice drag model, with the aim of
checking whether the two programs responded analogously to such a modification. In a
previous analysis, this drag model appeared to yield the best results in the present con-
Fluids 2021, 6, x FOR PEER REVIEW 11 of 16
was simulated
figuration. in Fluent
Figure withthe
5 provides a visual
tenfoldresults.
largerQualitatively,
DEM time step the on a less configurations
obtained performing PC,
without any
are again all significant
acceptable.variation in the
The fountain results.
height In that case,
is noticeably the required
reduced simulation
in all cases, time
becoming
was
moreabout 200tomin
similar theper 1 s.
experimental one.
Figure 5. Configuration of the particles after 2 s of simulation with the Di Felice drag model. The
particles are coloured according to their velocity magnitude. Experimental snapshot reprinted
from [14]. Copyright 2008 Elsevier.

If we now observe the velocity profiles, reported in Figure 6, we notice that the one
obtained with MFiX is almost identical to the experimental one, thus clearly representing
the best choice. The profile obtained with the structured mesh in Fluent is again better
than the Gidaspow case, but the peak velocity is quite underestimated, whereas the one
obtained with the Cartesian mesh yields a good reproduction of the fountain height, but
a very poor one of the overall velocity. In general, both programs confirm that the Di Felice
model, compared to the Gidaspow one, results in a markedly lower drag force in this
configuration, thus lowering the particle velocity and fountain height. It thus is further
confirmed to be the best choice to simulate the current setup. These results also confirm
theparticles
Figure 5. Configuration of the accuracy of the
after computational
2 s of frameworks,
simulation with the Di Felice since the equation
drag model. of theare
The particles Di coloured
Felice model
generally yields lower values than the Gidaspow model [11].
according to their velocity magnitude. Experimental snapshot reprinted from [14]. Copyright 2008 Elsevier.

1.25
Particle average vertical velocity (m/s)

Fluent - structured mesh

Fluent - Cartesian mesh
1
MFiX
Experimental
0.75

0.5

0.25

0
0 0.05 0.1 0.15 0.2 0.25
y (m)
Figure 6. Time-averaged
Time-averaged vertical
vertical profiles
profiles of
of the
the particles’
particles’vertical
verticalvelocity
velocitywhen
whenemploying
employingthe
theDi
Di
Felice drag model.

CFD simulations
3.2. Comparison can beTime
of Calculation notably sped up by employing parallel computing, which
in some cases is mandatory
Accuracy is an indispensableto achieve
factorthe
whendesired results
choosing within reasonable
a simulation program.time scales.
However,
Both programs allow employing this functionality, and especially for MFiX
given how CFD-DEM simulations can be demanding from a computational point of view, several details
about
the its implementation
computational demandare available
should neverinbethe literature [44,45].
overlooked. It needs
Regarding to be
this, our noted that
observations
parallel computation was not supported in MFiX for Windows before
point largely in favour of MFiX. To perform the simulations described in the previous of version 21.3,
with Windows users having to employ a virtual machine to perform such calculations.
section, which featured nearly identical setups and numerical parameters, MFiX required
However, any number of cores may be used. Conversely, Fluent can easily be run in
about 60 min of calculation per 1 s of simulation, whereas Fluent needed about 1050 min.
parallel on Windows, but the student licence currently does not allow employing more
Given that the large majority of the calculation time was employed by both programs to
than 4 cores. Figure 7 compares the normalised computation time per 1 s of simulation
perform the DEM calculations, it may be concluded that the particle-related algorithm of
that was needed in Fluent and MFiX when employing different numbers of cores. The
MFiX is more efficient. It should be noted that, in a previous work, this same spouted bed
normalised computation time is calculated by dividing the calculation time by the one
was simulated in Fluent with a tenfold larger DEM time step on a less performing PC,
required by a standard simulation in serial mode, with either Fluent or MFiX. From the
without any significant variation in the results. In that case, the required simulation time
plot, it is clear that the parallelisation algorithm of Fluent is more efficient, enabling a
was about 200 min per 1 s.
greater reduction in the computational time when employing the same number of cores.
CFD simulations can be notably sped up by employing parallel computing, which in
To achieve the same relative timesaving with MFiX, roughly double the number of cores is
some cases
required. is mandatory
Moreover, it cantoalso
achieve thethat
be seen desired results within
an excessive reasonable
number time not
of cores does scales. Both
provide
programs allow employing this functionality, and especially for MFiX several details
about its implementation are available in the literature [44,45]. It needs to be noted that
parallel computation was not supported in MFiX for Windows before of version 21.3, with
 needed in Fluent and MFiX when employing different numbers of cores. The normalised
computation time is calculated by dividing the calculation time by the one required by a
standard simulation in serial mode, with either Fluent or MFiX. From the plot, it is clear
that the parallelisation algorithm of Fluent is more efficient, enabling a greater reduction
Fluids 2021, 6, 382 12 of 15
in the computational time when employing the same number of cores. To achieve the
same relative timesaving with MFiX, roughly double the number of cores is required.
Moreover, it can also be seen that an excessive number of cores does not provide any
advantage: mostmost
any advantage: likely, for such
likely, a relatively
for such small-scale
a relatively problem,
small-scale the the
problem, communication
communicationbe-
comes too heavy to bring any advantage. In absolute terms, however, MFiX is
becomes too heavy to bring any advantage. In absolute terms, however, MFiX is still the still the
winner, reaching aa minimum
winner, reaching minimum of of 36
36 min/s
min/s against
against the
the 562
562 of
of Fluent.
Fluent.

100%
MFiX
Normalised calculation time
90% Fluent

80%

70%

60%

50%
0 2 4 6 8 10 12 14 16 18
Number of employed cores
Normalised calculation
Figure 7. Normalised calculation time upon employing a different number of cores.

CFD-DEM simulations
CFD-DEM simulations are
are also
also strongly
strongly affected
affected byby the
the number
number of of particles.
particles. To To
achieve a rough idea of how the two programs respond when this variable
achieve a rough idea of how the two programs respond when this variable changes, we changes, we
ran identical
ran identical simulations
simulations with
with halved
halved and
and doubled
doubled amounts
amounts of of particles. Figure 88 depicts
particles. Figure depicts
the results in a normalised fashion. The two programs appear to respond
the results in a normalised fashion. The two programs appear to respond quite similarlyquite similarly
to the
to the number
number of of particle
particle variation,
variation, with
with Fluent
Fluent being
being slightly
slightly more
more efficient. If this
efficient. If this trend
trend
continued for larger numbers of particles, at a certain point the two
Fluids 2021, 6, x FOR PEER REVIEW continued for larger numbers of particles, at a certain point the two programs wouldprograms would
13 show
of 16
showsimilar
very very similar performances,
performances, butcourse
but of of course
thisthis is not
is not somethingthat
something thatcan
canbe
be confirmed
confirmed
through the
through the current
current investigation.
investigation.
200%

175% MFiX
Normalised calculation time

150% Fluent

125%

100%

75%

50%

25%

0%
0 1 2 3
Normalised number of particles
Figure
Figure8.8.Normalised
Normalisedcalculation
calculationtime
timeupon
uponvarying
varyingthe
thetotal
totalnumber
number of
of particles.
particles.

3.3.Summary
3.3. SummaryofofAdvantages
AdvantagesandandDisadvantages
Disadvantages
The previous
The previous results
results have
haveshown
shownthat
thatboth
bothprograms
programs have advantages
have advantagesand disadvan-
and disad-
tages. For the specific case of this spouted bed, MFiX clearly appears as the superior
vantages. For the specific case of this spouted bed, MFiX clearly appears as the superiorchoice,
providing better results within a shorter timescale, but for a different application, other
choice, providing better results within a shorter timescale, but for a different application,
considerations may apply. In any case, both programs are powerful and versatile tools
other considerations may apply. In any case, both programs are powerful and versatile
that can be employed to perform CFD-DEM simulations. Table 3 sums up a point-by-point
tools that can be employed to perform CFD-DEM simulations. Table 3 sums up a point-
comparison of the two programs.
by-point comparison of the two programs.

Table 3. Point-by-point comparison of Ansys Fluent and MFiX applied to CFD-DEM.

Ansys Fluent MFiX

Commercial program, a free student licence with
Cost Open-source program, entirely free.
limitations is available.
Fluids 2021, 6, 382 13 of 15

Table 3. Point-by-point comparison of Ansys Fluent and MFiX applied to CFD-DEM.

Ansys Fluent MFiX

Commercial program, a free student licence with
Cost Open-source program, entirely free.
limitations is available.
GUI Both programs are equipped with a user-friendly GUI.
The geometry can be created in MFiX or imported,
Ansys programs provide numerous options to but only Cartesian cut-cell meshes can be
Geometry and mesh
generate geometry and mesh. employed. This may be a limiting factor for
some geometries.
Available for spherical particles, including
Available for spherical particles, including
CFD-DEM coarse-graining. Lift forces and rolling friction
coarse-graining. It is also possible to include lift
methodology torque are not included in the standard version of
forces, rolling friction torque and other forces.
the code.
Code visibility The code is not accessible by the user. The code is completely transparent to the user.
User-defined functions (written in C) allow some
The code (written in Fortran) is entirely editable by
Personalisation level of personalisation, but some variables (such
the user.
as contact forces) cannot be modified.
The standard options are more limited, but
Numerous and flexible options are available.
Results visualisation anything can be accessed by editing the code.
However, some variables (such as the drag force)
and analysis Paraview is often needed to analyse and visualise
cannot be accessed.
the results.
For the present application, it is about 17.5 times
For the present application, it is about 17.5 times
smaller than with Fluent. However, it seems to be
CPU cost larger than with MFiX. However, it seems to be less
more sensitive to increases of the number
sensitive to increases in the number of particles.
of particles.
With the same number of cores, the relative
With the same number of cores, the relative
Parallelisation speed-up is larger than with MFiX. The student
speed-up is smaller than with Fluent.
licence does not allow more than 4 cores.
The user and theory guide are very detailed, and Several tutorials on the CFD-DEM methodology
Available material the interface is clear, but the material related to the are available, but learning how to modify the code
CFD-DEM methodology is scarce. can require some effort.
The program has many options and can be The program is specifically aimed at situations
Other applications employed for a variety of applications in involving granular materials. It allows employing
different fields. other related methodologies (TFM, MP-PIC, DEM).

4. Conclusions
A pseudo-2D spouted bed was simulated through the CFD-DEM approach, com-
paring the performance of the commercial program Ansys Fluent and the open-source
program MFiX. Both programs can provide acceptable qualitative predictions when em-
ploying standard settings. If the Di Felice drag model is applied, MFiX yields better results
and provides a very good quantitative reproduction of the experimental particle velocity
profile. Moreover, despite employing similar mesh and time steps and the same number of
particles, MFiX is about 17 times faster. However, Fluent seems to respond slightly more
efficiently to an increase in the particle number and appears to have better parallelisation
functionalities. The Magnus lift force is also implemented in MFiX, and the results do not
change significantly, which is very different from what was observed in Fluent.

Author Contributions: Conceptualization, data curation, formal analysis, methodology, software,

visualization, writing—original draft: F.M.; funding acquisition, supervision, writing—review and
editing: R.D.F. All authors have read and agreed to the published version of the manuscript.
 efficiently to an increase in the particle number and appears to have better para
functionalities. The Magnus lift force is also implemented in MFiX, and the res
change significantly, which is very different from what was observed in Fluen

Fluids 2021, 6, 382 Author Contributions: Conceptualization, data curation, formal analysis,
14 ofmethodolog
15
visualization, writing—original draft: F.M.; funding acquisition, supervision, writing—
editing: R.D.F. All authors have read and agreed to the published version of the manus
Funding: Filippo Marchelli’s research grant was co-funded by the Liguria Region und
Funding: Filippo Marchelli’s research grant
gramma Operativo wasRegione
Por FSE co-funded by the
Liguria Liguria (code
2014-2020 Region under the Pro-
RLOF18ASSRIC/42/1).
gramma Operativo Por FSE Regione Liguria 2014–2020 (code RLOF18ASSRIC/42/1).

Institutional ReviewInstitutional Review

Board Statement: Board
Not Statement: Not applicable.
applicable.
Informed Not
Informed Consent Statement: Consent Statement: Not applicable.
applicable.
Data Availability
Data Availability Statement: Statement:
The experimental The
data experimental
employed for thedata employed
validation are for the validation
published and are pu
available at doi.org/10.1016/j.powtec.2007.11.044 (last accessed on 18 October 2021, published by 2021, p
available at doi.org/10.1016/j.powtec.2007.11.044 (last accessed on 18 October
Elsevier, Amsterdam,Elsevier, Amsterdam,
Netherlands). The dataNetherlands).
produced by The data produced
the simulations are by the simulations
available are availabl
in the article
cle and can be provided by the corresponding
and can be provided by the corresponding author in higher detail. author in higher detail.

Conflicts of Interest: Conflicts

The authors of declare
Interest:
noThe authors
conflict declare
of interest. nofunders
The conflicthad
of no
interest.
role inThe funders had no
the design
design of the study; in the collection, analyses, or interpretation of data;
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, in the
or writing o
script, or in the
in the decision to publish the results. decision to publish the results.

References References
1. Kennedy, D. What 1. Don’tKennedy,
We Know?D. What Don’t
Science We309,
2005, Know? Science 2005, 309, 75–75.
75. [CrossRef]
2. 2.
Van der Hoef, M.A.; van
Ye, der
M.; Hoef, M.A.;
van Sint Ye, M.; van
Annaland, M.;Sint Annaland,
Andrews, A.T.;M.; Andrews, A.T.;
Sundaresan, Sundaresan,
S.; Kuipers, J.A.M. S.;Multiscale
Kuipers, J.A.M. Multiscale
Modeling of Mod
Fluidized
Gas-Fluidized Beds; Elsevier: Beds; Elsevier:
Amsterdam, The Amsterdam,
Netherlands,The Netherlands,
2006; Volume 31, 2006;
ISBN Volume
0120085313.31; ISBN 0120085313.
3. 3. Multiscale
Ngo, S.I.; Lim, Y.-I. Ngo, S.I.;Eulerian
Lim, Y.-I.CFD
Multiscale Eulerian
of Chemical CFD ofAChemical
Processes: Review. Processes: A Review.
ChemEngineering 2020,ChemEngineering
4, 23. [CrossRef]2020, 4, 23.
4. 4.
Kieckhefen, P.; Pietsch, Kieckhefen, P.; Pietsch,
S.; Dosta, M.; Heinrich,S.; S.
Dosta, M.; Heinrich,
Possibilities S. Possibilities
and Limits and Limits
of Computational of Computational
Fluid Dynamics–Discrete FluidElement
Dynamics–Discr
Method Simulations in Process Engineering: A Review of Recent Advancements
Method Simulations in Process Engineering: A Review of Recent Advancements and Future Trends. Annu. Rev. Chem. Biomol. and Future Trends. Annu. Rev. C
Eng. 2020, 11, 397–422. Eng. 2020, 11, 397–422.
[CrossRef]
5. 5.
Cundall, P.A.; Strack, Cundall,
O.D.L. AP.A.; Strack,
discrete O.D.L. model
numerical A discrete numerical
for granular model for Géotechnique
assemblies. granular assemblies.
1979, 29, Géotechnique 1979, 29, 47–65.
47–65. [CrossRef]
6. Tsuji, Y.; Tanaka,6.T.; Ishida,
Tsuji, T.
Y.;Lagrangian
Tanaka, T.;numerical
Ishida, T.simulation
Lagrangian numerical
of plug flow ofsimulation
cohesionlessof particles
plug flow in of cohesionless
a horizontal particles
pipe. Powder in a hori
Powder
Technol. 1992, 71, 239–250. Technol. 1992, 71, 239–250.
[CrossRef]
7. 7.
Moliner, C.; Marchelli, Moliner, C.; Marchelli,
F.; Spanachi, F.; Spanachi,A.;
N.; Martinez-Felipe, N.;Bosio,
Martinez-Felipe,
B.; Arato, E.A.; Bosio,
CFD B.; Arato,
simulation of E. CFD simulation
a spouted of a spouted bed: C
bed: Comparison
between the Discrete between
Element the Discrete
Method (DEM)Element
and the Method (DEM)
Two Fluid and the
Model TwoChem.
(TFM). FluidEng.
Model (TFM).
J. 2019, Chem.
377, Eng.[CrossRef]
120466. J. 2019, 377, 120466.
8. 8.
Ostermeier, P.; DeYoung, Ostermeier, P.; DeYoung,
S.; Vandersickel, S.; Vandersickel,
A.; Gleis, A.;H.
S.; Spliethoff, Gleis, S.; Spliethoff,investigation
Comprehensive H. Comprehensive investigation
and comparison and comparis
of TFM,
DenseDPM and CFD-DEM for dense fluidized beds. Chem.
DenseDPM and CFD-DEM for dense fluidized beds. Chem. Eng. Sci. 2019, 196, 291–309. [CrossRef] Eng. Sci. 2019, 196, 291–309.
9. Lungu, M.; Siame, J.; Mukosha, L. Comparison of CFD-DEM and TFM approaches for the simulation of the small scale challenge
problem 1. Powder Technol. 2021, 378, 85–103. [CrossRef]
10. Liu, D.; Bu, C.; Chen, X. Development and test of CFD–DEM model for complex geometry: A coupling algorithm for Fluent and
DEM. Comput. Chem. Eng. 2013, 58, 260–268. [CrossRef]
11. Marchelli, F.; Hou, Q.; Bosio, B.; Arato, E.; Yu, A. Comparison of different drag models in CFD-DEM simulations of spouted beds.
Powder Technol. 2020, 360, 1253–1270. [CrossRef]
12. Li, J.; Campbell, G.M.; Mujumdar, A.S. Discrete Modeling and Suggested Measurement of Heat Transfer in Gas–Solids Flows.
Dry. Technol. 2003, 21, 979–994. [CrossRef]
13. Albano, A.; le Guillou, E.; Danzé, A.; Moulitsas, I.; Sahputra, I.H.; Rahmat, A.; Duque-Daza, C.A.; Shang, X.; Ching Ng, K.;
Ariane, M.; et al. How to Modify LAMMPS: From the Prospective of a Particle Method Researcher. ChemEngineering 2021, 5, 30.
[CrossRef]
14. Zhao, X.-L.L.; Li, S.-Q.Q.; Liu, G.-Q.Q.; Yao, Q.; Marshall, J.-S.S. DEM simulation of the particle dynamics in two-dimensional
spouted beds. Powder Technol. 2008, 184, 205–213. [CrossRef]
15. Geldart, D. Types of gas fluidization. Powder Technol. 1973, 7, 285–292. [CrossRef]
16. Moliner, C.; Marchelli, F.; Bosio, B.; Arato, E. Modelling of spouted and spout-fluid beds: Key for their successful scale up.
Energies 2017, 10, 1729. [CrossRef]
17. Herzog, N.; Schreiber, M.; Egbers, C.; Krautz, H.J. A comparative study of different CFD-codes for numerical simulation of
gas–solid fluidized bed hydrodynamics. Comput. Chem. Eng. 2012, 39, 41–46. [CrossRef]
18. Venier, C.M.; Urrutia, A.R.; Capossio, J.P.; Baeyens, J.; Mazza, G. Comparing ANSYS Fluent ®and OpenFOAM ®simulations of
Geldart A, B and D bubbling fluidized bed hydrodynamics. Int. J. Numer. Methods Heat Fluid Flow 2019, 30, 93–118. [CrossRef]
19. Marchelli, F.; Moliner, C.; Curti, M.; Bosio, B.; Arato, E. CFD-DEM simulations of a continuous square-based spouted bed and
evaluation of the solids residence time distribution. Powder Technol. 2020, 366, 840–858. [CrossRef]
Fluids 2021, 6, 382 15 of 15

20. Garg, R.; Galvin, J.; Li, T.; Pannala, S. Open-source MFIX-DEM software for gas-solids flows: Part I-Verification studies. Powder
Technol. 2012, 220, 122–137. [CrossRef]
21. ANSYS ANSYS Fluent User Guide; ANSYS: Canonsburg, PA, USA, 2019.
22. ANSYS ANSYS Fluent Theory Guide; ANSYS: Canonsburg, PA, USA, 2019.
23. Syamlal, M.; Rogers, W.; O‘Brien, T.J. MFIX Documentation Theory Guide; Department of Energy: Oak Ridge, TN, USA, 1993;
Volume 1004.
24. Chu, K.W.; Kuang, S.B.; Zhou, Z.Y.; Yu, A.B. Model A vs. Model B in the modelling of particle-fluid flow. Powder Technol. 2018,
329, 47–54. [CrossRef]
25. Marchelli, F.; Di Felice, R. A Discrete Element Method Study of Solids Stress in Cylindrical Columns Using MFiX. Processes 2020,
9, 60. [CrossRef]
26. Launder, B.E.; Spalding, D.B. The numerical computation of turbulent flows. Comput. Methods Appl. Mech. Eng. 1974, 3, 269–289.
[CrossRef]
27. Weaver, D.S.; Mišković, S. A Study of RANS Turbulence Models in Fully Turbulent Jets: A Perspective for CFD-DEM Simulations.
Fluids 2021, 6, 271. [CrossRef]
28. Gidaspow, D. Multiphase Flow and Fluidization: Continuum and Kinetic Theory Descriptions; Academic Press: Cambridge, MA,
USA, 1994.
29. Di Felice, R. The voidage function for fluid-particle interaction systems. Int. J. Multiph. Flow 1994, 20, 153–159. [CrossRef]
30. Ergun, S. Fluid flow through packed columns. Chem. Eng. Prog. 1952, 48, 89–94.
31. Wen, C.Y.; Yu, Y.H. Mechanics of Fluidization. Chem. Eng. Prog. Symp. Ser. 1966, 162, 100–111.
32. Rong, L.W.; Dong, K.J.; Yu, A.B. Lattice-Boltzmann simulation of fluid flow through packed beds of uniform spheres: Effect of
porosity. Chem. Eng. Sci. 2013, 99, 44–58. [CrossRef]
33. Marchelli, F.; Moliner, C.; Bosio, B.; Arato, E. A CFD-DEM sensitivity analysis: The case of a pseudo-2D spouted bed. Powder
Technol. 2019, 353, 409–425. [CrossRef]
34. Liu, R.-J.; Xiao, R.; Ye, M.; Liu, Z. Analysis of particle rotation in fluidized bed by use of discrete particle model. Adv. Powder
Technol. 2018, 29, 1655–1663. [CrossRef]
35. Rubinow, S.I.; Keller, J.B. The transverse force on a spinning sphere moving in a viscous fluid. J. Fluid Mech. 1961, 11, 447.
[CrossRef]
36. Marchelli, F.; Di Felice, R. On the influence of contact models on friction forces in discrete element method simulations. Chem.
Eng. Trans. 2021, 86, 811–816.
37. Ai, J.; Chen, J.F.; Rotter, J.M.; Ooi, J.Y. Assessment of rolling resistance models in discrete element simulations. Powder Technol.
2011, 206, 269–282. [CrossRef]
38. Zhou, Y.C.; Wright, B.D.; Yang, R.Y.; Xu, B.H.; Yu, A.B. Rolling friction in the dynamic simulation of sandpile formation. Phys. A
Stat. Mech. Its Appl. 1999, 269, 536–553. [CrossRef]
39. Boyce, C.M.; Holland, D.J.; Scott, S.A.; Dennis, J.S. Limitations on Fluid Grid Sizing for Using Volume-Averaged Fluid Equations
in Discrete Element Models of Fluidized Beds. Ind. Eng. Chem. Res. 2015, 54, 10684–10697. [CrossRef]
40. Hamidouche, Z.; Dufresne, Y.; Pierson, J.-L.; Brahem, R.; Lartigue, G.; Moureau, V. DEM/CFD Simulations of a Pseudo-2D
Fluidized Bed: Comparison with Experiments. Fluids 2019, 4, 51. [CrossRef]
41. Berger, M. Cut Cells: Meshes and Solvers. In Handbook of Numerical Analysis; Elsevier: Amsterdam, The Netherlands, 2017;
Volume 18, pp. 1–22.
42. Chu, K.W.; Chen, J.; Wang, B.; Yu, A.B.; Vince, A.; Barnett, G.D.; Barnett, P.J. Understand solids loading effects in a dense medium
cyclone: Effect of particle size by a CFD-DEM method. Powder Technol. 2017, 320, 594–609. [CrossRef]
43. Di Renzo, A.; Napolitano, E.; Di Maio, F. Coarse-Grain DEM Modelling in Fluidized Bed Simulation: A Review. Processes 2021,
9, 279. [CrossRef]
44. Gopalakrishnan, P.; Tafti, D. Development of parallel DEM for the open source code MFIX. Powder Technol. 2013, 235, 33–41.
[CrossRef]
45. Liu, H.; Tafti, D.K.; Li, T. Hybrid parallelism in MFIX CFD-DEM using OpenMP. Powder Technol. 2014, 259, 22–29. [CrossRef]