Sunteți pe pagina 1din 15

Comp. Part. Mech.

DOI 10.1007/s40571-016-0102-y

A hybrid FEM-DEM approach to the simulation of fluid flow


laden with many particles
Marcus V. S. Casagrande1 · José L. D. Alves1 · Carlos E. Silva1 · Fábio T. Alves1 ·
Renato N. Elias1,2 · Alvaro L. G. A. Coutinho1,2

Received: 1 September 2015 / Revised: 7 December 2015 / Accepted: 6 January 2016


© OWZ 2016

Abstract In this work we address a contribution to the momentum between discrete and continuous phases signifi-
study of particle laden fluid flows in scales smaller than cantly affect flow dynamics as a whole [1]. Also, according
TFM (two-fluid models). The hybrid model is based on a to Topin et al. [1], examples of multiphase systems in which
Lagrangian–Eulerian approach. A Lagrangian description is their dynamics are dictated by the coexistence of granular
used for the particle system employing the discrete element phase and saturated liquid include slurry transportation in
method (DEM), while a fixed Eulerian mesh is used for the pipelines, dispersal of fuel fragments into coolant water dur-
fluid phase modeled by the finite element method (FEM). ing a (hypothetic) nuclear accident, sedimentation in water
The resulting coupled DEM-FEM model is integrated in time treatment, fluidization in catalyst reactors, and debris flow
with a subcycling scheme. The aforementioned scheme is [1]. Additionally, understanding turbidity currents (typically
applied in the simulation of a seabed current to analyze which a gravity or density current) can help explain where and how
mechanisms lead to the emergence of bedload transport and organic matter was deposited and perhaps transformed by
sediment suspension, and also quantify the effective viscos- other processes of geological scale into hydrocarbons. Flu-
ity of the seabed in comparison with the ideal no-slip wall idized beds are widely used in many industrial processes, and
condition. A simulation of a salt plume falling in a fluid col- are one of the main possible applications of a hybrid discrete
umn is performed, comparing the main characteristics of the element-finite element (DEM-FEM for short) code.
system with an experiment. Some macroscopic models obtained through experimen-
tal observations and theoretical works can reproduce simple
Keywords Discrete element method · Hybrid models · mechanical behavior [2–5]. However, the physical mech-
Particle methods · Finite element method anisms occurring at the particles’ scale are usually more
complex phenomena, such as hydrodynamic instability, col-
lapse, transport, etc. The most complete numerical approach
1 Introduction to simulate a granular flow is to solve the solid phase as
individual particles and the fluid phase at the voids between
The interest of the scientific community in the study of gran- particles as a continuum. However, no general method is effi-
ular flow has significantly increased in recent decades. This cient enough to afford the computational cost of resolving the
is due to several factors particularly by innovations in large fluid flow in the gaps between closed-spaced particles [6],
number of industrial processes, in which the transfers of therefore, Hoef et al. [7] presents different modeling strate-
gies to balance computational cost and accuracy.
B Marcus V. S. Casagrande For industrial scale studies, the domain has a dimension of
marcusscasagrande@gmail.com order of magnitude of 10 m, demanding fast results for a large
1
domain, so the indicated model is the discrete bubble model
Laboratory for Computational Methods in Engineering,
Department of Civil Engineering, COPPE/Federal University
that treats the gas bubble as discrete entities. The two-fluid
of Rio de Janeiro, Rio de Janeiro, Brazil model is an Eulerian–Eulerian model that provides a contin-
2 High Performance Computing Center, COPPE/Federal
uous description for both phases. It is widely used because
University of Rio de Janeiro, Rio de Janeiro, Brazil of the relative fast speed; however, the integration between

123
Comp. Part. Mech.

the solid and fluid phases are carried out through drag force dimensional Eulerian finite element is used to describe the
correlations. Such correlations may not be accurate, since fluid behavior.
they cannot represent the full details of interactions between
particles and particle–fluid interactions. 2.1 Particle motion
In this work we describe a contribution to the study of
particulate systems in scales smaller than two-fluid models The solid fraction of matter is represented by a set of individ-
(TFM). The hybrid model is based on a Lagrangian–Eulerian ual particles, the motion of each particle being determined
approach, that is, the unresolved discrete particle method by the integration of the Newton–Euler equations of motion,
(UDPM) according to Hoef et al. [7]. In this approach, considering contact forces between particles and external
a Lagrangian description is used for the particle system forces acting on them [12]. Historically, this approach was
employing DEM, while a fixed Eulerian mesh is used for first introduced by Cundall and Strack [13], applied to gran-
the fluid phase modeled by FEM. This technique has been ular materials and named Distinct Element Method (DEM);
successfully applied to the study of fluidized bed in catalytic however, later DEM became more usual.
reactors [8]. The momentum exchange between the fluid According to O’Sullivan [14], there are two main clas-
and solid phases is considered through the insertion of a sifications for the simulation of particle’s interaction: the
source/sink term in the momentum equation of the fluid and a hard-sphere model and the soft-sphere model. In the hard-
force-displacement law for the solid phase. Darcy’s equation sphere model the particles are modeled as rigid bodies and
is used to create a linear relationship between the fluid veloc- interact through instantaneous collisions. However, in the
ity and the pressure gradient generated by the particles, while soft-sphere model, the equations of motion of each particle
Wen & Yu experimental correlation is used to define the non- are solved numerically, requiring a contact force model. The
linear region [9]. The terms responsible for the momentum spring-damper model is the most widely used, showing a
exchange depends directly on the local porosity, which varies good compromise between accuracy and efficiency.
in time due to the spatial rearrangement of particles. Thus, in Due to the large number of possible simultaneous contacts,
each time instant it is necessary to assess the local porosity directly related to the high concentration of particles, the
throughout the domain and map it accordingly. soft-sphere model is the most suitable and, therefore, imple-
The goal of this work is to present the methodology uti- mented in this work. The equation of motion of each particle
lized to couple, in a two-directional way, a continuum fluid can be written as
solver based on the finite element method (FEM) with a
discontinuous discrete code based on the discrete element
m ẍi = f i (1)
method (DEM). Fluid motion for the incompressible and
viscous fluid is governed by Navier–Stokes equations, which J θ̈i = Mi , (2)
are solved by an appropriate FEM implementation [10]. Clo-
sure equations are used to compute drag and lift forces over where ẍi is the particles’ acceleration, m is the particles’
the particles in the DEM framework [11]. Volume averaged mass, J is the inertia moment, θ̈i is the particles’ angular
momentum sink terms are included in the fluid equations. acceleration, f i is the resultant force and Mi the resultant
The resulting coupled DEM-FEM model is integrated in time moment both acting on a particle, and the index i refers to
with a subcycling scheme. Care is taken on mapping parti- spatial direction.
cles onto a particular tetrahedral mesh arrangement under The resultant moment and force may arise from several
the assumption that porosity is constant in each tetrahedron. different phenomena, but the most important for multiphase
The remainder of this paper is organized as follows. In the systems are
next section we briefly review the governing equations for
particle and fluid motion. In Sect. 3 we discuss the numerical f i = f iG + f iB + f iC + f iD (3)
implementation. In Sect. 4 we show a validation example,
f iG = mgi (4)
a channel flow simulation, and the numerical simulation of
a salt plume. The paper ends with a summary of our main f iB = Vp ρf gi , (5)
conclusions.
where f iG is the gravity force, f iB is the buoyancy, f iC is the
contact force, f iD is the drag force (or the coupling force),
gi is the gravity acceleration, Vp is the particle’s volume,
2 Governing equations and ρf is the specific mass of the fluid. Other possible forms
of interaction between particles (e.g., adhesion, aggregation,
In this paper a three-dimensional Lagrangian approach is and disaggregation) can be naturally treated within DEM
used with the DEM to describe the particles, and a three- framework [15–19].

123
Comp. Part. Mech.

surface, modifying the effective force acting over the parti-


cle. According to Hoomans et al. [22], the drag force acting
on a particle surrounded by other particles can be modeled
as

1 3 β  p
f iD = π dp u i − vi , (11)
6 (1 − ε)
Fig. 1 Contact model between two spheres a normal direction b tan-
gential direction where dp is the particle’s diameter, β is the momentum trans-
fer coefficient, ε is the porosity, u i is the fluid’s velocity, and
p
vi is the particle’s velocity [23].
As illustrated in Fig. 1, the contact force comes from the According to Kuipers et al. [24], for porosities lower than
soft-sphere model, consisting of stiffness and damping linear 0.80, β is defined by the Ergun equation in the following
elements [14] in the normal and tangential directions of the form:
sphere contact. Additionally, for the tangential direction the
Coulomb friction law is considered. (1 − ε)2 μf ρf  p
β = 150 + 1.75 (1 − ε) u i − vi  (12)
ε 2
dp dp
f n = K n δn + Cn δ̇n (6)
   δ̇t where ρf is the fluid’s specific mass and μf is the dynamic
f t = − min |μcoulomb f n | , K t δt + Ct δ̇t   , (7)
δ̇t  viscosity.
The first term present in the Ergun equation derives from
where f n is the normal force, K n is the normal stiffness coef- Darcy’s Law using the Kozeny–Carman equation, so it is
ficient, Cn is the normal damping coefficient, δn is the normal predominantly governed by laminar flow, while the second
penetration at the sphere, δ̇n is the normal penetration rate, term is more significant in turbulent flows.
f t is the tangential force, μcoulomb is the friction coefficient, However, for porosities higher than 0.80, Ergun equation
K t is the tangential stiffness coefficient, Ct is the tangential is no longer valid, and the following correlation, related with
damping coefficient, δt is the relative tangential slip between hindered settling, is presented by Wen and Yu [25]:
two spheres in contact, and δ̇t is its rate (the relative tangen-
3 ε (1 − ε)  p
tial velocity). The relative tangential velocity produces a slip β = Cd ρf u i − vi  ε−2.65 , (13)
between particles which can be computed as 4 dp

 where Cd is the drag of an isolated particle, modeled by Rowe


δt = δ̇t dt (8) and Henwood [26] as

f iC = f n n̂ i + f t tˆi , (9) 
Rep (1 + 0.15(Rep )
24 0.687 ), Rep ≤ 1000
Cd = , (14)
0.44 Rep ≥ 1000
where n̂ i and tˆi are unit vectors, pointing in the normal and
tangential directions. Finally, the resultant moment is the
where the Reynolds number is defined as
cross product of the tangential force ( f t tˆi ) and the radial
vector to the contact point (0.5 · dp n̂ j ).  p
ερf u i − vi  dp
Rep = . (15)
μf
Mi = −0.5 · dp f t εi jk n̂ j tˆk (10)
2.2 Fluid motion
Once the particles have been mapped in the Eulerian
domain (in the present work discretized by a tetrahedral Fluid motion is governed by the Navier–Stokes equation for
mesh), the fluid velocity is interpolated at particle positions. incompressible Newtonian fluid flow, according to
The drag forces acting in each particle are computed through
empirical laws involving the relative velocity between the ∂u i ∂u i ∂p ∂ 2ui
particle and the fluid. ρf + ρf u j = − + μf + ρf gi (16)
∂t ∂x j ∂ xi ∂x j∂x j
For a single particle immersed in a fluid medium, the drag
force is well determined and is correlated with Reynolds subject to the incompressibility constraint
number [20,21], while a system composed by many particles
poses additional complexities. The presence of surrounding ∂u j
= 0, (17)
particles reduces fluid space, increasing the shear at particle ∂x j

123
Comp. Part. Mech.

Fig. 2 DEM/FEM coupling


structure

where ρf is the fluid’s specific mass, p is the pressure, μf is


the dynamic viscosity, gi is the gravity acceleration, and u i
is the velocity of the fluid in the xi direction.
The software EdgeCFD [27–29], developed in-house, is a
stabilized finite element solver for Navier–Stokes equations
of incompressible flows using edge-based data structure. It
is currently the main tool in the group to simulate fluid prob-
lems.
However, when there is a two-phase flow, assuming the
fluid is incompressible and has a constant specific mass,
according to Bouillard et al. [30], the volume averaged
Navier–Stokes equations become
 
∂ εu j ∂ε
= − (18)
∂x j ∂t
∂ε u i ∂ε u i
ρf + ρf u j
∂t ∂x j
∂p ∂ 2ui
= −ε + ε μf + ερf gi − Ffpi (19) Fig. 3 DEM subcycling diagram
∂ xi ∂x j∂x j
 p 
Ffpi = β u i − vi . (20) allows relatively large time steps without loss of accuracy
for the fluid solution. On the other hand, small time steps
The last term,Ffp , represents the momentum transfer from are mandatory in DEM solvers to capture the high-frequency
the particle to the fluid, acting as a source/sink term, which content of particle-to-particle interactions, demanding a sub-
will be further discussed in Sect. 3. cycle approach [31], i.e., the computation of various DEM
The solution of continuity and volume averaged Navier– cycles for each FEM/CFD cycle.
Stokes equations, Eqs. (18) and (19), can be found by the After DEM cycle ending, the source/sink terms Ffp are
FEM; however, the assumption of a field of local porosity evaluated for each finite element in the mesh and proceed-
in the neighborhood of each particle adds complexity to the ing to a new step of the Navier–Stokes equations integration
problem. The particular approach used in the present work for the FEM/CFD cycle. The diagram with the main DEM
to circumvent this complexity is addressed in Sect. 3. subcycle steps is presented in Fig. 3.

2.3 General aspects of the coupled implementation


3 Numerical implementation
The overall flowchart for the coupling is presented in Fig. 2.
After the solution of fluid equations (FEM/CFD cycle), the A FEM program for incompressible flow (EdgeCFD) is used
drag forces for each particle are calculated, which allows the to solve the fluid equations. This program was first designed
progress of the solution for particle motion (DEM cycle). and optimized to solve Eqs. (16) and (17) using linear tetra-
EdgeCFD, the FEM solver, is a fully implicit velocity- hedra. For the integration of particle motion equations, an
pressure solver for the time integration of incompressible explicit first-order scheme is used to update the velocities
Navier–Stokes equations. The implicit nature of the operator and the positions of the particles.

123
Comp. Part. Mech.

Fig. 4 Particle–fluid
interaction diagram

As mentioned in Sect. 2.2, porosity adds complexity to The particles mapping in element mesh is a computational
solving Eqs. (18) and (19) when compared to the standard search of the location of each particle relative to each element
Navier–Stokes solver. In the present work, it is assumed that of the fluid mesh. It creates a list that identifies the element
the variations of the porosity field in time and space are of of the fluid mesh that each particle is located. This procedure
second order, being negligible compared to the other terms can be computationally costly since particles are changing
in the equations. This assumption simplifies Eqs. (18) and positions continuously. This mapping is a requirement for
(19), which can be rewritten as the compute porosity per element task, which defines the
porosity field (ε). Since these procedures can be complex
∂u j
= 0 (21) and has a considerable impact at the total calculation time,
∂x j
they are explained in detail in Sect. 3.1.
∂u i ∂u i Once the porosity field has been determined, the next step
ρf + ρf u j
∂t ∂x j is the computation of the source/sink term. According to Eqs.
∂p ∂ 2ui Ffpi (11)–(15), the relative velocity between the fluid and the par-
=− + μf + ρf gi − . (22)
∂ xi ∂x j∂x j ε ticle is a key parameter to the computations. However, each
particle has its own velocity and the fluid velocity changes
As can be observed, in this form, these equations present a with the particle location. The adopted procedure to calcu-
slight difference from the standard Navier–Stokes equations, late the relative velocity is to interpolate, for each particle,
i.e., the additional source/sink term, Ffpi /ε. This term can be the velocity at the particle location (fluid velocity interpola-
interpreted as the total momentum per unit of volume trans- tion for each particle). Fluid velocities are interpolated within
ferred from the particle to the fluid, and can be treated as each tetrahedron with standard volume shape functions [32];
a body force in the FEM framework. Therefore, the numer- thus, the relative velocity can be computed.
ical solution procedure is the same as for the conventional After velocities have been calculated, the sequence requir-
Navier–Stokes equations, enabling a least invasive coupling. es the computation of the Reynolds number from Eq. (15),
The solution of the fluid equations requires additionally then the drag coefficient from Eq. (14), the drag force
the specification of the source term (Ffpi ) and the porosity from Eq. (11), the momentum transfer coefficient from
(ε). Thus, the overall procedure necessary to determine the Eqs. (13) and (12), and finally, the source/sink term from
source term and porosity within the particle fluid interaction Eq. (20).
subroutine is illustrated in Fig. 4. Figure 5 illustrates a fixed control volume comprising sev-
The particle–fluid interaction subroutine consists of six eral spherical particles. A complete numerical solution for
main tasks: the particles mapping in elements mesh, the com- this problem would require fluid mesh boundaries represent-
pute porosity per element, the particle loop, the source term ing particles, but it demands a very small mesh size due to
average on element, and the nodal distribution of source term. limited space between them, which needs to be updated every

123
Comp. Part. Mech.

Fig. 5 Momentum transfer diagram between particle and fluid ele-


Fig. 6 Structured grid and its mapping indexes
ment. a Flow through a particulate medium, b spatial distribution of
particles, c transferred momentum from each particle inside the trian-
gular element, d equivalent transferred momentum to the element mesh
unstructured mesh, this is usually a very expensive task [35].
A structured mesh generator was specifically developed for
the purpose of particle mapping.
time step due to particle motion, exponentially increasing
computational cost.
An alternative approach, computationally feasible, is to 3.1.1 Structured mesh
consider the whole control volume as a continuous porous
medium. According to Shimizu [9] and Asgian et al. [33], The structured mesh is divided into a Cartesian grid of rectan-
for laminar fluid flow Darcy’s law can predict the pressure gular box cells, which can be identified by their respective
gradient of a porous medium, meanwhile Kozeny–Carman rows, columns, and layers (Fig. 6). Each cell has a unique
equation [34] relates the equivalent permeability of a packed combination of indexes and identifier. They have a finite
bed of solid spheres. For a wider range of Reynolds number, length in each Cartesian direction and, as such, particles can
the empirically based Ergun’s equation predicts the pres- have their positions tested against the cells [36], as described
sure gradient for turbulent flow (Eq. 12 ). Additionally, for in Eqs. (23) through (25).
lower porosities, the correlation of Wen and Yu [25] is used
(Eq. 13). i p = int[(xp − xmin )/ x ] + 1 (23)
Accordingly, the source/sink term determines, for each jp = int[(yp − ymin )/ y ] + 1 (24)
particle, the pressure drop of an equivalent porous medium
formed by several of particles with the same conditions (e.g., kp = int[(z p − z min )/ z ] + 1 (25)
relative velocity, size, distribution). Thus, each particle inside i cell = n y · n x · (kp − 1) + n x · ( jp − 1) + i p , (26)
a fluid element will return a different source/sink term, and
the final value for the fluid element is the average of the where i p , jp , kp are the row, column, and layer, respectively,
source/sink terms from every single particle inside the fluid of the particle; xp , yp , z p are the coordinates of the particle;
element (source term average on element). Figure 6 illus- xmin , ymin , z min are the minimum coordinates of the struc-
trates how the momentum is transferred from a set of particles tured mesh; x , y , z are the lengths of rows, columns
to the fluid element. After the pressure drop is calculated for and layers, respectively, in the structured mesh; i cell is the
each element, the last step is the nodal distribution of source identifier for the cell; and n x and n y are the number of rows
term. and columns, respectively, in the structured mesh. The int
function truncates its argument to an integer.
3.1 Particle mapping Since EdgeCFD was designed for tetrahedral-based
meshes, the cells are further divided into tetrahedra. To keep
Evaluation of current element porosity requires mapping all the mesh as compact as possible, the hexahedral cells are
the particles within the element at a given time step. In an subdivided into 5 tetrahedra (Fig. 7), the smallest possible

123
Comp. Part. Mech.

Fig. 7 Cell subdivision.


Central tetrahedron in red

Substituting each vertex in turn for the particle, the four


volume coordinates of the particle can be calculated, as
shown in Eqs. (28) through (30), where x p , y p , and z p are
the coordinates of the particle being mapped. If all of the
volume coordinates are positive in sign, the particle is con-
sidered inside the element.
⎡ ⎤
Fig. 8 Tree diagram of the mesh data structure, where T are the tetra- x p x j x k xl
hedra 1 ⎢ yp y j yk yl ⎥
η1 = det ⎢ ⎥
⎣ z p z j z k zl ⎦ (28)
6 · Vi jkl
Fig. 9 Tetrahedral element and
particle 1 1 1 1
⎡ ⎤
x i x p x k xl
1 ⎢ yi yp yk yl ⎥
η2 = det ⎢ ⎥
⎣ z i z p z k zl ⎦ (29)
6 · Vi jkl
1 1 1 1
⎡ ⎤
x i x j x p xl
1 ⎢ yi y j yp yl ⎥
η3 = det ⎢ ⎥
⎣ z i z j z p zl ⎦ (30)
6 · Vi jkl
number of tetrahedral per cell [32]. The resulting data struc- 1 1 1 1
⎡ ⎤
ture is shown in Fig. 8. xi x j xk xp
1 ⎢ yi y j yk yp ⎥
η4 = det ⎢ ⎥
⎣ zi z j zk zp ⎦ . (31)
3.1.2 Element identification 6 · Vi jkl
1 1 1 1
To evaluate element porosity, it must further be identified in
Given the pre-arranged element construction within the cell,
which element of the cell each particle is. The use of volume
it is possible to determine in which element the particle
coordinates allows for such evaluation [32]. The volume of
resides by testing only against the central tetrahedron, which
any tetrahedral element can be calculated through the use of
is readily mapped within the cell. If the particle is not within
Eq. (27).
the central element, one of its four volume coordinates will
⎡ ⎤ return negative, thus indicating through which facet of the
xi xj xk xl central tetrahedron the particle is outside, directly indicating
1 ⎢ yi yj yk yl ⎥
Vi jkl = det ⎢
⎣ zi
⎥, (27) the correct element. This simplifies the entire mapping to
6 zj zk zl ⎦ four direct tests per particle.
1 1 1 1 It is important to note that cell subdivision into five tetra-
hedra introduces the problem of lack of element face compat-
where x, y, and z are the coordinates of each vertex of the ibility between neighboring cells. To circumvent this issue,
tetrahedron (Fig. 9). two different mirrored configurations of cell subdivisions are

123
Comp. Part. Mech.

4 Numerical examples

4.1 Single particle and two-way coupling

The first problem addressed consists of a single particle in


sedimentation within a column of fluid under the effect of
gravity. The sedimentation of a particle within a viscous fluid
less dense than the particle achieves a maximum terminal
velocity, vf , determined by the balance of gravitational force
and hydrodynamic forces (drag and buoyancy).
The Stokes’ law determines the drag force of a sphere
immersed in the laminar flow of a viscous fluid [20]. It can
be used to estimate the terminal velocity of the particle.
When considering two-way coupling, the relative move-
ment between particle and fluid influences not only the
trajectory of the particle but also the flow of the fluid.
Fig. 10 “Odd” and “even” cell configurations side-by-side The current test cases employed six different particle
diameters ranging from 0.208 to 0.76 mm, with a constant
specific mass of 2650 kg/m3 . Fluid properties are specific
mass of 1000 kg/m3 and dynamic viscosity of 0.000100 Pas.
needed, an “odd” configuration and an “even” configuration Simulation time spans one second. The time step for particle
(Fig. 10), according to its identifier, as calculated by Eq. (26). motion is set at 1 × 10−5 s, while for the fluid solution time
step is set to 0.01 s, yielding 1000 particle substeps for each
CFD step.
3.2 Porosity The fluid domain is a rectangular box with square base
(44 × 44 × 80 mm). Mesh size is set at 4 mm cells, providing
The determination of the spatial distribution of porosity is 11 × 11 × 20 cells and 12100 tetrahedral elements.
essential for the evaluation of the sink/source term for the Results of the numerical experiment are shown in Table 1.
fluid and drag forces for the particles. Utilizing an adequate It compares final velocity of the particle for two-way cou-
homogenization model, the spatial distribution of discrete pling, Stokes’ law and actual experiments performed by
particles can be transferred to a continuous medium [37]. Hallermeier [21], while the Reynolds number is based on
In the present work, it is assumed that instant porosity is the experimental final velocity of the particle.
homogenous within each element [8]. The results show good agreement between experimental
Element porosity is the ratio of non-solid volume within and numerical results, presenting deviation of at most 12 %
the element over the total element volume [21]. The solid in terminal velocity. Since the Stokes’ law is based on lam-
volume inside the element is assumed as the sum of the vol- inar theory, it is valid only when Reynolds number is very
ume of each particle mapped inside the element. As such, small, much less than one, and it is expected that the error
element porosity εelk of element k is a unique propriety that of Stokes terminal velocity grows with the Reynolds num-
depends entirely on the particles located within it. Therefore, ber. Therefore, the results using Stokes’ law are consistent,
Eq. (31) can be used to evaluate the porosity of an element presenting great discrepancy at higher Reynolds, up to 3.7
k, that is, times.

π
Velk − 6 dp3i
i
4.2 Sensitivity analysis for mesh density
εelk = , (32)
Velk
For this analysis, a fixed particle diameter of 0.76 mm is taken
for varying mesh size. The fluid and time step properties are
where Velk is the element’s volume, as calculated by Equation the same as the previous experiment. Five different mesh
(27), and dpi is the diameter of each particle mapped within sizes were tested, ranging from 3D to 10D, where D is the
the element. It was assumed that each particle is completely particle’s diameter.
inside the element that contains the particle center. Therefore, Two sets of experiments were conducted with this setup.
particles intersecting multiple elements will only contribute The first set considers the particle initially centered relative
to a single element porosity. to the mesh cell as depicted in Fig. 11.

123
Comp. Part. Mech.

Table 1 Results for single


Diameter (mm) Reynolds exp. vf exp. (m/s) vf two-way (m/s) vf Stokes (m/s)
particle sedimentation
0.208 58 0.028 0.026 0.039
0.25 83 0.033 0.034 0.056
0.29 113 0.039 0.041 0.076
0.42 244 0.058 0.064 0.159
0.59 496 0.084 0.092 0.313
0.76 836 0.110 0.117 0.519

Fig. 11 Initial centered positioning of particle


Fig. 13 Convergence of particle final velocity for different mesh sizes

Fig. 12 Velocity over time for cell centered initial positioning


Fig. 14 Porosity over time for cell m centered initial positioning

Figure 12 shows the results for particle velocity over time


for the first set of experiments in which the particle is centered As the particle is centered in the mesh cell, as it settles
at the fluid cell. A discrepancy between terminal velocities down it only passes within the central tetrahedron of each
for each mesh size is evident. This can be attributed to the cell. Therefore, it is expected that the porosity remains con-
error in the evaluation of element porosity. The forces both stant throughout the simulation. It can be observed that the
the fluid and particle are subjected to are directly related larger the mesh size, the larger the porosity and porosity value
to the porosity, which, in turn, depend on the ratio of par- approaches the unity.
ticle and element volume. In the case of a single particle, The second set considers the particle shifted towards a
the theoretical porosity is unitary. The larger the mesh, the corner of the mesh cell as depicted in Fig. 15. Figure 16
closer the calculated porosity will approach the unity, which shows the results for particle velocity over time for the second
explains the observed convergence of terminal velocity with battery. Oscillations in the velocity are evident. Figure 17
increasing mesh size. Figure 13 displays a plot of the obtained elucidates the nature of such oscillation.
terminal velocity as function of relative mesh size. Figure 17 shows the variation of calculated element poros-
Figure 14 shows the variation of calculated element poros- ity for the element in which the particle is mapped over time.
ity for the element in which the particle is mapped over time. Since the tetrahedra on the edge of the cell have half the

123
Comp. Part. Mech.

Table 2 Comparison of oscillation for offset initial positioning


Cell size/particle Amplitude of Amplitude of terminal
diameter Ratio porosity (%) velocity (%)

3 6.2 9.3
4 2.5 5.0
5 1.3 2.6
7 0.46 1.0
10 0.16 0.33

Fig. 15 Initial offset positioning of particle

Fig. 16 Velocity over time for offset initial positioning

Fig. 18 Particles configuration after settling

the oscillation in terminal velocity is larger than the oscilla-


tion in porosity, indicating a possible amplification of effects.
Acceptable results were observed in cell size to particle
diameter ratios of 5 or greater, with negligible deviations in
ratios over 10. However, it must be noted that a larger mesh
size disturbs both the convergence and discretization of the
fluid, which might invalidate the overall solution.

4.3 Channel flow


Fig. 17 Porosity over time for offset initial positioning
The aim of this case is to compare the velocity profile of
the fluid near the bottom of the channel using the no-slip
volume of the central tetrahedron, a fixed particle volume boundary condition at the bottom wall and using a bed of
causes a fluctuation in calculated porosity. A larger mesh particles to represent the soil (Fig. 18).
size causes a smoother convergence, due to the smaller repre- The simulation addresses the channel flow inlet configu-
sentability of the particle in the larger volume. Alternatively, ration shown in Fig. 20. Representing the soil, a polydisperse
a different mapping scheme utilizing an uniform mesh could distribution comprising 4800 particles with diameter ranging
potentially eliminate this fluctuation. from 0.2 mm to 0.3 mm was settled in a channel of rectangular
The oscillation in porosity modifies the velocity through cross section with 12 mm height, 1.8 mm width, and 12 mm
the calculation of the force terms in a non-linear fashion. It length. For this case, no tangential forces where computed,
is important to study the impact of such fluctuations in the since the pack is almost stationary. The additional proper-
velocity to provide further insight into the approach. ties utilized in the numerical simulation are listed in Table 3.
Table 2 presents the amplitude of the oscillations in per- The entire channel was discretized with 6000 tetrahedra as
centage of average porosity and terminal velocity. In all cases, depicted in Fig. 19.

123
Comp. Part. Mech.

Table 3 Physical properties


Property Value (SI units)
attributed to the simulation of
channel flow Time step for fluid solver 5.0 × 10−3 s
Simulation time 5.0 s
Dynamic viscosity (μf ) 0.001 Pa s
Fluid specific mass (ρ f ) 1000.0 kg/m3
Fluid domain size 12.0 × 12.0 × 1.8 mm
Mesh Structured, 6000 tetrahedra with 0.6 mm side
Time step for DEM solver 10−5 s (500 subcycles)
Number of particles 4800
Particle specific mass 2170.0 kg/m3
Contact stiffness (K n ) 105 N/m
Contact damping (Cn ) 0.026 N s/m
Particle diameter (dp ) [0.20 mm, 0.30 mm]

at outlet. Each marker represents a measure realized at the


respective node of the mesh.
It can be observed from the graph that the slopes of the
profiles are slightly different. This difference is due to the
interaction of the fluid and the particle that has a different
effect than the no-slip wall condition. To assess this differ-
ence, both slopes were calculated resulting in a 1.86◦ for
the no-slip condition seabed and a 2.82◦ for the particulate
seabed. The ratio between both slopes is 152 %; therefore,
the effective viscosity of the fluid in the particulate seabed
is 50 % greater than the effective viscosity of the fluid in
the no-slip condition seabed. Furthermore, these results are
qualitative, since no experiments or mesh sensitivity has been
realized.

4.4 Salt plume falling in a fluid column

Due to the lack of physical experiments involving particles


with controlled parameters, this work presents a comparison
between a numerical simulated and a qualitative experiment
of a salt plume. This experiment consists of a batch of salt
released in a recipient filled with water. The main objective
Fig. 19 Fluid mesh utilized for simulations
of this case is to compare the numerical and experimental
features of the salt plume that develops.
The employed boundary conditions are inlet velocity of A minimal characterization of the salt batch was per-
30.0 mm/s at the left wall for the nodes above the seabed and formed over a sample of 300 salt grains, measuring grain
0.0 mm/s for the nodes bellow seabed (if applied), no-slip at sizes using digital imaging and a caliper as a reference scale.
the bottom wall, free slip at the upper wall, no penetration While the characterized sample does show slight deviation
at side walls, and outlet at the right wall (Fig. 20). At time t from a Gaussian distribution, to simplify the construction of
= 0 s the particles are frozen and the sink term is nulled so the model, one was assumed for the entire set, using com-
the fluid develops a steady-state profile. This methodology puted statistical values of the sample data. Computed values
was applied to facilitate the convergence. At t = 0.05 s the for mean and standard deviation yielded d̄p = 309.0 µm
particles were released and the sink term computed. and σ = 114.0 µm, respectively. The bar chart for the mea-
The solution was achieved after the simulation reaches the surements and the assumed Gaussian distribution used in
steady state. Fig. 21 illustrates the obtained velocity profiles the numerical model are overlaid in Fig. 22. The additional

123
Comp. Part. Mech.

Fig. 20 Channel flow inlet


boundary conditions for no-slip
wall (left) and particle bed
(right)

Fig. 22 Salt grain size distribution

Table 4 Data for the salt plume numerical experiment


Property Value (SI units)

Time step for fluid solver 10−3 s


Simulation time 1.0 s
Dynamic viscosity (μf ) 0.001 Pa s
Fluid specific mass (ρ f ) 1000.0 kg/m3
Fluid domain size 0.10 × 0.10 × 0.25 m
Fig. 21 Fluid velocity profiles at outlet Mesh Structured, 12500 tetrahedra with
0.01 m side
Time step for DEM solver 10−5 s (100 subcycles)
properties utilized in the numerical simulation are listed in
Number of particles 28224
Table 4.
Particles specific mass 2170.0 kg/m3
The salt plume was represented numerically by a set of
28224 spherical particles organized in a volume 20 × 20 × Contact stiffness (K n ) 100 N/m
40 mm volume in a fluid domain with 100 × 100 × 250 mm. Contact damping (Cn ) 8.5 × 10−4 N s/m
The salt plume is centered at the fluid column axis at a height Particles diameter (dp ) Gaussian Distribution
(d̄p = 309.0 µm; σ = 114.0 µm)
of 200 mm from the bottom (Fig. 23). The diameter of the

123
Comp. Part. Mech.

particles was generated randomly according to the adjusted


Gaussian distribution. The boundary condition for the fluid
is no-slip on every wall. For this case, no tangential forces
where computed due to the low rate of contact between par-
ticles.
Time evolution of the numerical experiment is presented
in Figs. 24 and 25. The formation of a dome shape is the main
feature observed. The dome shape formation mechanism is
due to the recirculation of the fluid. Initially, the fluid and
particles are in repose. After the particles start to move down-
ward driven by the gravity force, a drag force is generated
at the particle and, consequently, the same force acts on the
fluid through the momentum source. This momentum source
impels the fluid in the center of the domain downward that,
because of incompressibility, impels the surrounding fluid
upward, causing the circulation of the fluid. The circulation
creates a radial component of the fluid velocity that transports
the particles radially, creating the dome profile.
A sequence of snapshots for the experimental test is dis-
played in Fig. 26. A dome shape similar to the one obtained
Fig. 23 Salt plume initial setup
in the numerical simulation is observed. The set of images

Fig. 24 Evolution of the numerical “salt plume”

Fig. 25 Evolution of fluid velocity field of the numerical “salt plume”

123
Comp. Part. Mech.

Fig. 26 Evolution of the experimental “salt plume”

shows an initial small conglomerate of salt grains that gradu- diameter. In the studied cases, severe deviations are observed
ally evolves into a dome shape with a tail of grains, the same when the mesh size approaches particle diameter. These devi-
features that can be observed in the numerical simulation, ations are due to the assumption of constant porosity within
despite the salt plume not being perfectly centered. It should each finite element, which is determined by the fraction of
be considered that the numerical and experimental results particles found within it. Despite these issues, the numerical
have inherent differences in some aspects, such as the cubic experiments presented good results for mesh sizes approxi-
shape of salt grains, chemical interaction between salt and mately 10 times as large as particle diameter.
water, and initial conditions of salt grains. These deviations In all cases, several DEM subcycles were computed for
may preclude a quantitative comparison between numerical each CFD cycle. Through careful investigation, a two orders
and experimental results. of magnitude ratio, i.e., 100 DEM subcycles for each CFD
It is important to highlight that such features can only be cycle, demonstrated to be the most effective in all cases pre-
observed in a two-way coupled simulation, since the parti- sented. The exact ratio is directly dependent of the stiffness
cles’ motion due to gravity starts the fluid flow which in turn and mass of the particles and the fluid flow regime, but the
is responsible for the dome-torus shape in the particles’ tra- dynamic interaction between particles will generally require
jectory. A one-way fluid-to-particle approach would have the a much smaller time step than the fluid dynamics.
fluid remain at rest as dampened particles fell in a straight line The implemented code is capable of simulating com-
due to gravity, whereas a one-way particle-to-fluid approach plex phenomena in a macroscopic scale through a detailed
would have the fluid recirculating while particles fell in free- description of the rheology and dynamics at the particle scale.
fall due to gravity. In the case of channel flow, it was demonstrated that the
tool could be used to simulate local effects with the purpose
of better understanding differences between simplified and
more complete analyses. Finally, the case of the salt plume
5 Conclusions showed remarkable qualitative resemblance between numer-
ical and experimental data and allowed for a more thorough
In this work, a multiphase hybrid solid-fluid computational understanding of the formation of the dome feature due to
tool was developed coupling the DEM with the Finite Ele- the recirculation of fluid.
ment Method (FEM), with the objective of simulating particle
laden flows in scales smaller than two-fluid models. Acknowledgments The authors gratefully acknowledge the partial
Verification of the code is obtained through a set of support provided by Petrobras S.A., the Brazilian Oil Company, CNPq,
the National Research Counsil, and ANP, the National Petroleum
test cases, especially through the sedimentation case. In Agency.
this particular case, numerical, experimental, and analyti-
cal solutions for the terminal velocity of sedimentation of
a single particle immersed in fluid were compared. During References
laminar flow regime, the numerical solutions showed good
agreement with experimental and analytical data. During 1. Topin V, Dubois F, Monerie Y, Perales F, Wachs A (2011) Micro-
transient/turbulent flow, where the analytical results are no rheology of dense particulate flows: application to immersed
longer valid, there was good agreement between numerical avalanches. J Non-Newton Fluid Mech 166(1):63–72
2. Chhabra RP (2012) Bubbles, drops, and particles in non-Newtonian
and experimental data.
fluids. CRC press, Boca Raton
Additionally, the sedimentation simulations helped gauge 3. Peker SM, Helvaci SS (2011) Solid-liquid two phase flow. Elsevier,
the influence of the FEM mesh size relative to the particle Amsterdam

123
Comp. Part. Mech.

4. Phillips RJ, Armstrong RC, Brown RA, Graham AL, Abbott JR 21. Hallermeier RJ (1981) Terminal settling velocity of commonly
(1992) constitutive equation for concentrated suspensions that occurring sand grains. Sedimentology 28(6):859–865
accounts for shear-induced particle migration. Phys Fluids A 22. Hoomans BPB, Kuipers JAM, Van Swaaij WPM (2000) Granular
4(1):30–40 dynamics simulation of segregation phenomena in bubbling gas-
5. Zhang DZ, Prosperetti A (1997) Momentum and energy equations fluidised beds. Powder Technol 109(1):41–48
for disperse two-phase flows and their closure for dilute suspen- 23. Anderson TB, Jackson R (1967) Fluid mechanical description
sions. Int J Multiph Flow 23(3):425–453 of fluidized beds. Equations of motion. Ind Eng Chem Fundam
6. Wu S, Yuan L (2015) A hybrid FD-DEM solver for rigid particles 6(4):527–539
in viscous fluid. Comput Fluids 118:159–166 24. Kuipers JAM, Van Duin KJ, Van Beckum FPH, Van Swaaij WPM
7. Van der Hoef MA, Annaland MS, Deen NG, Kuipers JAM (2008) (1992) A numerical model of gas-fluidized beds. Chem Eng Sci
Numerical simulation of dense gas-solid fluidized beds: a multi- 47(8):1913–1924
scale modeling strategy. Annu Rev Fluid Mech 40:47–70 25. Wen CY, Yu YH (1966) Mechanics of fluidization. Chem Eng Prog
8. Hoomans BPB, Kuipers JAM, Briels WJ, Van Swaaij WPM (1996) Symp Ser 62:100–111
Discrete particle simulation of bubble and slug formation in a two- 26. Rowe PN, Henwood GA (1961) Drag forces in a hydraulic model
dimensional gas-fluidised bed: a hard-sphere approach. Chem Eng of a fluidized bed-part I. Trans Inst Chem Eng 39:43–54
Sci 51(1):99–118 27. Elias RN, Coutinho ALGA (2007) Stabilized edge-based finite ele-
9. Shimizu Y (2004) Fluid coupling in PFC2D and PFC3D. Numerical ment simulation of free-surface flows. Int J Numer Methods Fluids
modeling in micromechanics via particle methods. In: Proceedings 54((6–8)):965–993
of the 2nd international PFC symposium, Kyoto, pp 281–287 28. Guerra GM, Zio S, Camata JJ, Rochinha FA, Elias RN, Paraizo
10. Elias RN, Martins MAD, Coutinho ALGA (2005) Parallel edge- PLB, Coutinho ALGA (2013) Numerical simulation of particle-
based inexact newton solution of steady incompressible 3D navier- laden flows by the residual-based variational multiscale method.
stokes equations., Euro-Par 2005 parallel processingSpringer, Int J Numer Methods Fluids 73(8):729–749
Berlin, pp 1237–1245 29. Lins EF, Elias RN, Rochinha FA, Coutinho ALGA (2010)
11. Cho SH, Choi HG, Yoo JY (2005) Direct numerical simulation of Residual-based variational multiscale simulation of free surface
fluid flow laden with many particles. Int J Multiph Flow 31(4):435– flows. Comput Mech 46(4):545–557
451 30. Bouillard JX, Lyczkowski RW, Gidaspow D (1989) Porosity dis-
12. Radjaï F, Dubois F (2011) Discrete-element modeling of granular tribution in a fluidized bed with an immersed obstacle. AlIChE J
materials. Wiley-ISTE, New York 35(6):908–922
13. Cundall PA, Strack ODL (1979) A discrete numerical model for 31. Belytschko T, Yen HJ, Mullen R (1979) Mixed methods for time
granular assemblies. Geotechnique (Thomas Telford) 29(1):47–65 integration. Comput Methods Appl Mech Eng 17:259–275
14. O’Sullivan C (2011) Particulate discrete element modelling: a geo- 32. Zienkiewicz OC, Taylor RL, Zhu JZ (2013) The finite ele-
mechanics perspective. Taylor & Francis, New York ment method: its basis and fundamentals, 7th edn. Butterworth-
15. Kobayashi T, Kawaguchi T, Tanaka T, Tsuji Y (2002) DEM analysis Heinemann, Saint Louis
on flow pattern of Geldart’s group A particles in fluidized bed. In: 33. Asgian MI, Cundall PA, Brady BH (1995) Mechanical stability
Proceedings of the world congress on particle technology, pp 21–25 of propped hydraulic fractures—a numerical study. J Pet Technol
16. Li J, Kuipers JAM (2002) Effect of pressure on gas-solid flow (Society of Petroleum Engineers) 47(3):203–208
behavior in dense gas-fluidized beds: a discrete particle simulation 34. Bear J (2012) Hydraulics of groundwater. Courier Dover Publica-
study. Powder Technol 127(2):173–184 tions, New York
17. Li J, Kuipers JAM (2003) Gas-particle interactions in dense gas- 35. Lohner R (1995) Robust, vectorized search algorithms for interpo-
fluidized beds. Chem Eng Sci 58(3):711–718 lation on unstructured grids. J Comput Phys 118(2):380–387
18. Moon SJ, Kevrekidis IG, Sundaresan S (2006) Particle simulation 36. Munjiza AA (2004) The combined finite-discrete element method.
of vibrated gas-fluidized beds of cohesive fine powders. Ind Eng Wiley, Chichester
Chem Res 45(21):6966–6977 37. Zhu HP, Zhou ZY, Yang RY, Yu AB (2007) Discrete particle simu-
19. Ye M, Van der Hoef MA, Kuipers JAM (2005) The effects of par- lation of particulate systems: theoretical developments. Chem Eng
ticle and gas properties on the fluidization of Geldart A particles. Sci 62(13):3378–3396
Chem Eng Sci 60(16):4567–4580
20. Batchelor GK (2000) An introduction to fluid dynamics. Cam-
bridge University Press, Cambridge

123

S-ar putea să vă placă și