Professional Documents
Culture Documents
Powder Technology
a r t i c l e i n f o a b s t r a c t
Article history: To achieve simulating the fluid-solid systems which consist of irregular polygonal particles and walls, an im-
Received 10 June 2019 proved and extended LBM-IMB-DEM model has been developed. The method couples the Lattice Boltzmann
Received in revised form 30 July 2019 Method (LBM) for fluid flow calculation, the polygonal Discrete Element Method (DEM) for particle-particle
Accepted 7 August 2019
and particle-wall interactions computation, and the Immersed Moving Boundary (IMB) scheme for fluid-solid in-
Available online 12 August 2019
teractions calculation. The Convex Decomposition method is proposed as an extension of polygonal DEM, which
Keywords:
successfully expands the applied range to arbitrary polygons. A series of optimization algorithms and flowchart
Lattice Boltzmann method designs are adopted, including the improved progressive scanning algorithm and the proposed solid volume frac-
Discrete element method tion fast algorithm, which can effectively increase the computational efficiency. The model is comprehensively
Immersed moving boundary validated by performing several verification simulations, the results of which are consistent with the existing lit-
CFD erature and theoretical data. The simulation of the representative engineering process, clean bed filtration of ir-
Non-spherical particulate system regular airborne particles in porous media, is then accomplished as the application test. The results have proved
Filtration that the proposed LBM-IMB-DEM is a promising model which has satisfactory adaptability and computational ef-
ficiency for simulating complex fluid-solid systems, which will help complement the experimental research into
such multiphase systems.
© 2019 Published by Elsevier B.V.
1. Introduction The Lattice Boltzmann Method (LBM) has been proved successful in
simulating complex fluid flows, and coupling LBM with DEM then be-
Particles moving in fluids are commonly encountered in natural and come a popular field. Noble and Torczynski [5] proposed the Immersed
engineering applications. Most of the particles are generally of irregular Moving Boundary (IMB) method for determining the numerical bound-
shapes instead of spherical or elliptical. Chemical industrial processes ary of solid particles, which overcame the problems of earlier LBM
often involve large amounts of irregular particles interacting with fluid works where the boundary intersection handling was too coarse [6,7].
flow, moving walls (mechanical devices) and walls of complex structure IMB method utilises solid volume fraction of each LBM lattice cell cov-
(porous media), all of which make the simulating of this complex solid- ered by the particles to calculate the fluid-solid interaction force. By in-
liquid system require huge computational cost and lots of attention to troducing the solid volume fraction, the surface curve and details are
implementation details of simulation, for example dynamic mesh draw- treated smoothly when the no-slip boundary condition is satisfied.
ing and mesh quality check. Other approaches [8,9] for LBM-DEM coupling has also been suggested
The discrete element method (DEM) is widely used for dealing with on the basis of the Immersed Boundary Method (IBM). In recent year,
the behaviour of bulk solids. Coupling DEM with Computational Fluid Tan et al. [10], Owen et al. [11] and other researches [12,13] are focused
Dynamics (CFD) methods has drawn considerable interest for recent on improving these schemes.
decades. DEM has been coupled with some CFD methods successfully The LBM-DEM has been used previously in two or three dimensions,
[1] such as the solvers depending on the Navier-Stokes equations: Finite also with various particle shapes such as spherical, ellipsoidal and
Element Method (FEM) [2], Level Set Method (LSM) [3] and Particle super-quadrics, as well as particles which are made from a combination
FEM (PFEM) [4]. However, these methods have some absence of details of multiple of such shapes called glued sphere model in DEM [1,14].
at the pore scale which is a noticeable drawback for simulating particles However, this method has been limited to convex shapes so far. So
and walls with complex shapes. most of the simulation cases in above researches were based on circular
or oval particles with simplified processes. Galindo-Torres [15] pro-
⁎ Corresponding authors. posed an effective LBM-IMB-DEM model for general shapes and tested
E-mail addresses: sai.gu@surrey.ac.uk (S. Gu), gaoxin@tju.edu.cn (X. Gao). with partially simplified simulation setups. In this paper, an improved
https://doi.org/10.1016/j.powtec.2019.08.006
0032-5910/© 2019 Published by Elsevier B.V.
178 X. Li et al. / Powder Technology 356 (2019) 177–192
where δt and δ_t are the accumulated tangential sliding distance and tan-
2.1. Convex polygonal DEM
gential sliding velocity, Kt corresponds to tangential stiffness, μ is the
friction coefficient.
This study adopts the DEM for convex polygonal particles organised
A second-order backward difference formulae (BDF2) solver is used
by Matuttis and Chen [16]. The translation and rotation of particle are
to solve Newton's second law formula. The BDF2 solver provides high
determined by Newton's second law based on the acting force Fc and
precision in the position's computation and great stability for particle
torque Tc:
heaps simulation. The results of previous researches show that the
F c ¼ F ela þ F dam þ F fri þ F hyd ð1Þ BDF2 solver stays stable when the collision is resolved with five or
T c ¼ r F c þ T hyd ð2Þ
where Fela and Fdam are the elastic force and the damping force respec-
tively, Ffri is the friction, Fhyd and Thyd are the hydrodynamic force and
torque solved by IMB scheme (see Section 2.4), and r is the vector
from the particle's mass centre to the force point.
The elastic force between polygon particle depends on the overlap-
ping geometry of the contact. Fig. 1 shows the normal case where two
convex polygons intersect and the magnitude of the elastic force is pro-
portional to the area A of overlap between the two particles:
4r1 r 2
l¼ ðr 1 ¼ kr1 k; r 2 ¼ kr2 kÞ ð4Þ Fig. 2. Calculation of the direction of elastic force based on overlap polygon: C is the mass
r1 þ r2 centre of the overlap polygon; S1 and S2 are the intersection points.
X. Li et al. / Powder Technology 356 (2019) 177–192 179
Fig. 3. The Convex Decomposition method's processing of the two-dimensional outer contour of different objects.
more time steps [16], which is the principle for determining the DEM constituent particles are computed separately, as shown in Fig. 4. The
time step in this study. BDF2 solver computes the acceleration, velocity, displacement and posi-
tion of the combined particle in the main loop.
Taking the leftmost green combined particle in Fig. 5 as example, the
2.2. Convex decomposition method acting force Fs (F1, F2, F3) on each constituent particle is determined by
the polygonal DEM introduced in Section 2.1. The calculation formulas
In order to simulate the particles and walls of arbitrary polygon for the external force Fcomb and external torque Tcomb on the green com-
shapes including convex and concave, we propose the Convex Decom- bined particle are as follows:
position method as an effective modification of the current polygonal
X
DEM. F comb ¼ Fs ð9Þ
The two-dimensional outer contour of most objects can be reduced
to a polygon, which can be further decomposed into a combination of X
T comb ¼ rcs F s ð10Þ
several small convex polygons. So, with the Convex Decomposition
method, all the concave and irregular polygonal particles' outer contour
where rcs is the direction vector pointing from Ccomb (combined
are handled as the combination of a certain number of convex DEM
particle's centre of mass) towards the centroid of each overlap polygon.
polygons attaching to each other, as presented in Fig. 3, which are called
The coordinates of Ccomb equal to the mass weighted average of the con-
combined particles.
stituent particles' centroids Cs:
For the objects with complicated outer contours, the Convex Decom-
position method can be accomplished by increasing the number of con- P
C ∙ ms
stituent polygons with reduced area. In the air filtration simulation in C comb ¼ Ps ð11Þ
ms
Section 5, a stack of large amounts of convex constituent polygons is
used to emulate the filter material with porous structure. In addition, where ms is the mass of constituent particle.
the circular or elliptical objects can be represented by the polygons It should be noted that the combined particle's mass moment of in-
with more edges, for example, a hexadecagon is used as the substitution ertia is obtained by the Parallel axis theorem:
of a circular particle in Section 4.2. For the cases where large amounts of
X
particles are involved and the requirement for the shape fidelity is rela- Icomb ¼ Is þ ms ds
2
ð12Þ
tively low, decreasing the amount of constituent polygons while in-
creasing the corresponding edge number can effectively reduce the
where Icomb is the combined particle's moment of inertia. Is and m are
computation and capture the shape characteristics at the same time.
the moment of inertia and the mass of the constituent particle respec-
The force and torque acted on the combined particle is the resultant
tively. ds is the distance between Ccomb and the mass centre of constitu-
force and torque of all constituent particles. The acting forces on
ent particle.
Theoretically, there will be various ways of the convex decompo-
sition for the same polygon. Fig. 5 illustrates two different decompo-
sition approaches for the concave polygon: (a) one quadrangle, one
pentagon and one hexagon; (b) nine triangles. If the corners of poly-
gon are divided, as shown in the red wireframes of Fig. 5 (b), DEM
collision calculations at these corners will also be divided, which
will increase the total amount of computation and calculation
error. So, the principle for the decomposition method is to introduce
where Ω is the collision operator which is determined by the dimen- Fig. 6. IMB lattice node classification and the definition of solid volume fraction ε.
sionless relaxation time τ and the equilibrium distribution functions feq
i :
Δt eq
Ω¼− f ðx; t Þ− f i ðx; t Þ ð15Þ moving particles. As shown in Fig. 6, IMB scheme classifies the LBM lat-
τ i
tice nodes into four types. The solid volume fraction ε of lattice node
eq 3 9 3 equals to the area ratio occupied by the solid phase, which is the
f i ðx; t Þ ¼ ωi ρ 1 þ 2 ei ∙v þ 4 ðei ∙vÞ2 − 2 v∙v ði ¼ 0; …; 8Þ ð16Þ
C 2C 2C most important variable in IMB scheme. The solid volume fraction de-
termines the weighting function B directly:
4
where ω i is the weighting coefficient with fixed values ( ω0 ¼ ,
9 εðτ−0:5Þ
1 1 B¼ ð22Þ
ω1;2;3;4 ¼ , ω5;6;7;8 ¼ ), and C is the lattice speed and has the ð1−εÞ þ ðτ−0:5Þ
9 36
relationship to lattice length, h, and LBM time step (Δt):
In the IMB scheme, LBM's standard collision operator in Eq. (15) is
C ¼ h=Δt ð17Þ modified as
Δt
X
8 where an additional collision term, Ωsi , is introduced to realise the effect
ρ¼ fi ð18Þ of particle motion on fluid flow:
i¼0
eq eq
Ωsi ¼ f −i ðx; t Þ−f i ðx; t Þ þ f i ðρ; U s Þ−f −i ðρ; uÞ ð24Þ
X
8
ρv ¼ f i ei ð19Þ
i¼0 where sub-symbol −i stands for the opposite direction of i. Us is the
velocity of solid lattice node at time t + Δt which depends on the
C 2 particle translational velocity Up, the particle angular velocity ω as
P¼ pffiffiffi ρ ð20Þ
3 well as the position of the centre of mass (x c , yc), as illustrated in
Fig. 6. Us is given by
Moreover, the macroscopic fluid kinetic viscosity ν and the dimen- qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sionless relaxation time τ has the following relationship which is used
U s ¼ U p þ ω lc lc ¼ ðx−xc Þ2 þ ðy−yc Þ2 ð25Þ
to calculate the LBM time step Δt:
2
h The total hydrodynamic force Fhyd on DEM particle is considered as
ν ¼ ðτ − 0:5Þ ð21Þ
3Δt the summation of the change of momentum on fluid and solid boundary
nodes that the particle covers:
Torczynski [5] is adopted to solve the modified bounce back rule for
X. Li et al. / Powder Technology 356 (2019) 177–192 181
The torque Thyd is solved with a similar method: 3.1. LBM lattice node classification
( " !#) Scanning algorithm has been used in LBM calculation [17], and on
X X the basis of the existing algorithm, the characteristics of polygons are
T hyd ¼ Ch ðx−xc Þ Bn Ωsi ei ð27Þ
n i
utilized to reduce the computational cost of node position judgment.
Node classification in the program is built as a pre-preparation module
for solid volume fraction calculation. As introduced in Section 2.4, there
are four types of lattice nodes in the IMB scheme: solid boundary node,
interior solid node, fluid boundary node, and fluid node. For the fluid
3. Programming and algorithm optimization node and the internal solid node, the solid volume fraction ε are 0 and
1, respectively, which do not need to be calculated. Therefore, the LBM
In this section, innovative flowchart designs and optimization algo- lattice node update strategy is: Use the improved “progressive scanning
rithms are proposed to achieve the following targets: [1] To add the algorithm” to quickly determine the node classification, and then only
Convex Decomposition method to the simulation program, which is calculate ε of the fluid and solid boundary nodes with the proposed
theoretically capable of handling any irregular 2D polygonal shapes; “solid volume fraction fast algorithm” (see in Section 3.2). Fig. 7 illus-
[2] To extend the application range of the LBM-IMB-DEM model by trates the progressive scanning algorithm. The process steps are as
supporting walls with complex porous structures as well as moving follows:
walls; [3] To introduce optimization algorithms reducing the amount
1) By default, all lattice nodes are set as fluid nodes with ε being zero.
of computation, as the coupling between LBM and DEM modules has a
2) Read the bounding box of the current particle, as shown by the green
huge computational demand.
wire frame in Fig. 7 (a). Determine the nodes occupied by the
For the first two targets, we introduced an improved LBM-IMB-DEM
bounding box and extend the nodes outwards by one grid (the
computational sequence and a modified DEM module design in Appen-
brown lattice region in Fig. 7 (a)). This area is the calculation area.
dix A. And for the third target, we adopted two optimization algorithms
Number all rows and columns in the calculation area.
for LBM lattice node classification and solid volume fraction calculation
3) Determine the node classification starting from row a to row m. The
(see in Sections 3.1 and 3.2), both of which are specially designed for the
scanning of each line consists of the horizontal forward scanning and
IMB scheme and the polygonal DEM.
backward scanning:
4) Horizontal forward scanning. In the case of the row d in Fig. 7 (b), the
forward scanning starts from the node d1. Connect d1 with the parti-
cle centroid C, and judge the intersection of Cd1 with all particle
edges (red lines in Fig. 7 (b)). If there is an intersection (i.e., the
edge B1 intersects with Cd1), proceed to the node d2 and judge the in-
tersection of Cd2 with all particle edges. If there is no intersection, it
can be determined that the first “no intersection node” (i.e., d2) that
appears in a single forward scanning is the solid boundary node can-
didate, and the previous node involved (i.e., d1) is the fluid boundary
node candidate. The particle edges associated with the boundary
nodes are recorded correspondingly (i.e., both d1 and d2 are associ-
ated with B1). The forward scanning ends.
5) Horizontal backward scanning. In the case of the row d in Fig. 7 (b),
backward scanning starts from the node d10. Similar to step [4],
judged whether or not particle edges intersect with Cd10. If inter-
sects, move the intersection judgment sequentially to node d9, d8,
and d7. When it is judged that Cd7 does not intersects with any
edge, the first “no intersection node” (i.e., d7) in the backward scan-
ning is recorded as the solid boundary node candidate, and previous
node involved (i.e., d8) is the fluid boundary node candidate. The
particle edges associated with the boundary nodes are recorded cor-
respondingly (i.e., both d7 and d8 are associated with B3). The back-
ward scanning ends.
6) After the horizontal progressive scanning from row a to row m, col-
umn 1 to column 10 are then scanned vertically with the same
method as steps [4] and [5]. Taking column 5 as an example, the ver-
tical forward scanning direction is from node a5 to m5, while the ver-
tical backward scanning starts from node m5 to a5.
7) In steps [4] to [6], all lattice nodes that have not participated in the
intersection judgment are judged as interior solid nodes.
8) Return to step [2] and continue to the next particle.
It should be noted that only candidates for boundary nodes are clas-
sified through the above steps. If the calculated ε of these candidate
nodes equal to 1 or 0, they will be classified to interior solid nodes or
fluid node. Progressive scanning algorithm can greatly reduce the com-
putation of the node classification. For example, the classification of 10
nodes in the row d only calculate the intersection of 6 sets of line
Fig. 7. The progressive scanning algorithm for LBM lattice node classification. segment.
182 X. Li et al. / Powder Technology 356 (2019) 177–192
As for judging the intersection between line segments a1a2 and b1b2, 3.2. Fast solid volume fraction calculation
this paper adopts the fast intersection judgment algorithm:
This section will introduce the “solid volume fraction fast algorithm”
which is proposed for the polygonal DEM and designed compatible with
Ints1 ¼ Vecp ða1 b1 ; a1 b2 Þ ∙ Vecp ða2 b1 ; a2 b2 Þ ð28Þ the Convex Decomposition method.
Fig. 8 shows the calculation of the solid volume fraction ε of various
boundary nodes. L represents the target node. Fig. 8 (a) to (d) are solid
Ints2 ¼ Vecp ðb1 a1 ; b1 a2 Þ ∙ Vecp ðb2 a1 ; b2 a2 Þ ð29Þ boundary nodes, while Fig. 8 (e) to (g) are fluid boundary nodes. Square
A1A2A3A4 is the node's region. The red lines indicate the edges of the par-
ticle, and the intersection points between the particle edges and the
Vecp ðx; yÞ ¼ xð1Þ ∙ yð2Þ−yð1Þ ∙ xð2Þ ð30Þ
boundary of the node's region are marked as R1, R2, …, Rn. The grey
part is the area covered by the solid phase (called the solid phase poly-
gon). The red points are the vertexes of the solid phase polygon.
where Ints1 and Ints2 are the judgment numbers. When both Ints1 and
Taking the case shown in Fig. 8 (d) as example, process steps of the
Ints2 are less than or equal to zero, line segments intersect.
solid volume fraction fast algorithm for solid boundary nodes and candi-
Using the above algorithm, intersection judgment of two line seg-
dates are as follows:
ments can be finished with 10 multiplications and 4 subtractions,
which requires small amount of computations. The fast intersection 1) Read the particle edges associated with the boundary nodes re-
judgment algorithm is packaged as an independent MATLAB® custom corded in the progressive scanning algorithm (see in Section 3.1).
function, and all the calculations involving the intersection judgment Solve the intersections R1 and R2. If no intersection occurs, L is classi-
will call this function. fied as interior solid node and ε is 1.
2) Compute the intersection of the two particle edges. If the inter- 3) Judge whether LA1, LA2, LA3, LA4 intersect with R1R3 and R2R3 in se-
section point is within the node's region, record it as R3. Record quence. If LAn does not intersect with R1R3 or R2R3, record An as the
the above R1, R2, …, R n as the vertexes of the solid phase vertex of the solid phase polygon (i.e., A2 and A3 in Fig. 8 (d)).
polygon. 4) The area of the solid phase polygon is calculated by its vertexes
R1R3R2A2A3 obtained in the steps [1] to [3]. ε equals to the area
ratio of the solid phase polygon to the node's region.
Fig. 10. Results of the flow past a rectangle polygon simulation: (a) Computed drag
coefficient vs. Reynolds number for steady flow; (b) Computed recirculation length vs. Fig. 11. The particle sedimentation simulation. (a) Simulation settings; (b) The setup of
Reynolds number. dodecagon particle with the Convex Decomposition modification.
184 X. Li et al. / Powder Technology 356 (2019) 177–192
It should be noted that the above algorithm does not support the
case where more than one particle vertexes appear in the node's region,
as shown in Fig. 8 (h). However, in the actual simulation, the lattice
spacing will be much smaller than the particle side length in order to ef-
fectively capture the shape characteristics of the particle, so the cases in
Fig. 8 (h) is very rare and currently not considered.
In traditional algorithms, in order to calculate the area of the solid
phase polygon, inverse trigonometric functions must be used to de-
termine the anticlockwise order of the vertexes [17]. As complex
MATLAB® functions, trigonometric functions have higher computa-
tional cost than the fast intersection judgment algorithm in
Section 3.1 (as well as several logical judgment) [18]. After replacing
the traditional algorithm with the proposed solid volume fraction
fast algorithm, the running time of the 10,000-node solid volume
fraction calculation test was shortened by about 94%. Therefore,
the solid volume fraction fast algorithm can effectively reduce calcu-
lating amount of the LBM module.
4. Model validation
Fig. 13. The free collision of DEM particles. (a) Simulation settings of one original group;
(b) The setup of particles in the corresponding comparative group.
This section will simulate the classic case—flow pasting the rectan-
gular obstacle. We calculate the drag coefficient and the recirculation
length corresponding to different Reynolds numbers, and compare the
results with the existing literature data. The LBM-IMB coupled scheme
for immersed polygonal obstacles is validated.
As shown in Fig. 9, one square polygon obstacle (with the side length
of 0.1025 m) was placed at the centre of a channel. The height and
length of the channel are 0.82 m and 3.07 m respectively, which leads
to a blockage ratio B = 1/8. Periodic boundaries were applied to the
channel's inlet and outlet. The fluid domain is formed by a 614 × 164 lat-
tice with the lattice spacing h = 0.005 m. The relaxation time τ = 0.55.
The fluid kinetic viscosity ν is taken as 1.0 × 10−4 m2/s, and the fluid
density ρ is 103 kg/m3.
Table 1
Parameters of the DEM particle free collision simulation.
Parameter Value
3
Solid density, ρs (kg/m ) 3.0 × 103
Two-dimensional Young's modulus, Y2D (N/m) 5.0 × 106
Damping coefficient, D 0 or 0.2
Tangential stiffness, Kt (N/m) 1.0 × 105
Fig. 12. The results of the particle sedimentation simulation: (a) Time histories of the
Coefficient of friction, μ 0
particle displacement in Y-axial direction; (b) Comparison of hydrodynamic force
Simulation time, ttotal (s) 60
applied to the particle.
X. Li et al. / Powder Technology 356 (2019) 177–192 185
Lr
¼ −0:065 þ 0:0554 Re ð0:5b Reb60Þ ð31Þ
D
proposed by Breuer, Bernsdorf, Zeiser and Durst [19] is also plotted. Side
length D = 0.1025 m.
The pair of symmetric vortices behind the rectangle polygon emerge
when the Reynolds number lies between approximately 1.1 to 2.4,
which leads to a small recirculation length at corresponding region, as
shown in Fig. 10 (b). Owen et al. and Breuer et al. found this interval
are 3.4–4.7 and 1.1–2.0, respectively. The Lr then shows a reasonable
agreement to the empirical expression Eq. (31). But the relationship be-
tween Lr and Reynolds number was found to be nonlinear, similar to the
results of Owen et al. which also adopted the coupled LBM-IMB method.
One possible reason for this nonlinear relationship is that the IMB
scheme is sensitive the flow parameters and the hydrodynamic
boundary's real location maybe different to the physical boundary by a
certain value according to the Reynolds number. Another possible rea-
son is the influence of inlet and outlet obstacles which partly form the
periodic domain [11].
The flow pasting a rectangle obstacle simulation gives excellent
agreement to previous researches, which supports the validity of the
adopted LBM coupled with IMB scheme for polygonal obstacles.
Table 2
Results of DEM particle free collision simulation.
Table 3
Results of the airborne particles filtration simulation.
Filter a b c d e
Fibre number 60 68 75 83 90
SVF (%) 20.3 22.2 24.5 25.9 26.9
Mean pore size (×10−3 m) 2.04 1.93 1.92 1.83 1.65
Average filtration rate (%) 45.3 66.7 70.7 74.7 93.3
Initial outlet flow velocity (×10−3 m/s) 8.22 7.7 6.63 5.56 4.62
Final outlet flow velocity (×10−3 m/s) 8.02 7.59 5.25 4.38 4.1
Wang et al.'s work. As a result, the motion data after the collision timing
(~0.2121 s), when the particle reaches the bottom, are not included in
Fig. 12 (a) and (b).
As can be seen from Fig. 12 (b), though using the same IMB scheme,
Wang et al. got a smoother time history of the hydrodynamic force. The
reason is that Wang et al.'s model used circular shape-based DEM in-
stead of polygonal DEM. An equilateral dodecagon particle is selected
in the present work as the substitution for the circular particle. The con-
tinuous change of boundary nodes' solid volume fraction ε of dodecagon
is less smooth than that of a circular one, which leads to a less stable hy-
drodynamic force. We treat this difference of hydrodynamic force as an
acceptable design error, and IMB scheme and the Convex Decomposi-
tion modification of our work is credible.
The stability and accuracy of the polygonal DEM module is key to the
credibility of the whole program. The instability or error of the DEM
module directly leads to changes in the kinetic energy of the system be-
fore and after the collision [16]. As a result, the section verifies the sta-
bility of DEM module by measuring the time history of kinetic energy
Fig. 15. The setup of the airborne particle filtration simulation: (a) Overall setups; of the “free collision system” by calculating the total kinetic energy of
(b) Three types of particle adopted the filtration simulation. all particles. If the kinetic energy of the system is maintained after a
large number of collisions, we assume that the DEM module has satis-
factory stability and accuracy.
well with the literature data. In the first period (0 to ~0.5 s), the particle The verification program randomly generates 8 polygonal particles
moves downwards with an acceleration, and the hydrodynamic force with similar size, and the vertex number of which is the random integer
increases correspondingly. After a distance, both the particle velocity between 3 and 8. Then the particles are separately rotated by a random
and the hydrodynamic force stay constant as the gravity force and hy- angle (0 to 2π) and placed in a square zone surrounded by four static
drodynamic force has reached a balanced state before the collision. It walls (as shown in Fig. 13 (a)). The initial velocities of the particles are
is noted that the parameters for stationary walls are unspecified in chosen randomly between 2.0 × 10−3 m/s and 6.0 × 10−3 m/s with a
Fig. 16. The two-dimensional porous media emulation with different numbers of polygons.
X. Li et al. / Powder Technology 356 (2019) 177–192 187
Fig. 17. The average flow velocity of different filters during filtration process. Four contours show the velocity distribution of Filter a and Filter e at 0.8 s and 1.6 s.
random direction angle (0 to 2π). The range of the initial velocity is se- the case without damping, while the orange curve indicates the case
lected to cover the particles' velocity in Section 5. where the damping force (damping coefficient D = 0.2) is added.
Ten random repeated simulations (called original groups) are con- For both the original group and the comparative group, in the pro-
ducted, and Fig. 13 (a) illustrates the settings of one original group. cess of about 500 collisions, system kinetic energy of cases without
Each simulation has a corresponding comparative simulation (called damping maintains almost unchanged because of the absence of any
comparative group), where 4 of the 8 polygonal particles were treated dissipative force, which is in accordance with the Law of Momentum
with the Convex Decomposition method (as shown in Fig. 13 (b)), Conservation. The kinetic energy changing rate of about 0.05% is consid-
and all the other parameter settings are consistent with the original ered a negligible calculation error. Those results show that the DEM
group. module as well as the Convex Decomposition method have satisfactory
The free collision system simulation does not consider gravity, fric- stability and accuracy.
tion or fluid drag force, so the LBM module is not activated. The time As for the cases with damping, about 45% of the initial kinetic energy
step of the DEM module is set as 6.0 × 10−6 s. The other parameter set- is dissipated in collisions, which leads to a gradually decreasing average
tings are detailed in Table 1. The same parameters are used for the par- velocity of the particles and consequent decreased collision times.
ticles and walls. The introduction of the Convex Decomposition method in the com-
There are numerical differences between the results of different parative group leads to an increase in the number of individual particles
original groups, but the trend and conclusion are the same. Fig. 14 and processed by the DEM module, which brings more calculations and will
Table 2 illustrate the result of the original group shown in Fig. 13 consequently accumulate more calculation error. As a result, the fluctu-
(a) as well as its comparative group. In Fig. 14, the blue curve represents ation of the system kinetic energy of the comparative group is slightly
Fig. 18. The average velocity of particles which are finally captured by the filter.
188 X. Li et al. / Powder Technology 356 (2019) 177–192
larger than the original group, which can be seen from Fig. 14. Also, the which involves porous media, irregular particles, random collision be-
comparative group's computation time is slightly longer than the origi- tween solid phases, and fluid-solid coupling on complex dynamic
nal group (about 16%). In general, the overall performance of the Con- boundaries. So the airborne particle filtration simulation can effectively
vex Decomposition method is stable and accurate, and it will not test the adaptability, computing capability and computational efficiency
significantly affect the computation efficiency. of the model proposed in this paper.
As shown in Fig. 15 (a), the top and bottom sides of the computa-
5. Application test tional domain (20 mm in width, 20 mm in height) are the pressure-
inlet and pressure-outlet for fluid flow individually. The length of simu-
In this section, the clean bed filtration of airborne particles is simu- lated particles and walls is set as small-scaled polygons. A group of par-
lated with the LBM-IMB-DEM coupling method. Air filtration process ticles moves downwards with the air flow, then pass through or be
is a representative complicated chemical process, the simulation of intercepted by the filtration medium. 15 sets of simulation were carried:
Fig. 19. Particle trajectories and flow velocity contour of different filtration simulation setups: (a) Filter d; (b) Filter e. Red and white lines show the trajectories of particles escaped and
been captured by the filter, respectively.
X. Li et al. / Powder Technology 356 (2019) 177–192 189
Fig. 21. The pore size distribution of Filter d with different particles after filtration.
190 X. Li et al. / Powder Technology 356 (2019) 177–192
about 9% after filtration. Because Particle A has a more complex irregular D Damping coefficient
structure, the filter with Particle A has a higher proportion of small pore ds Distance between the mass centre of combined and constitu-
size (0 to 1.2 mm) and lower proportion of mid-range size (1.2 to ent particles
1.8 mm). Particle C has a significantly higher frequency of distribution ei Discrete velocities
in mid-range size. Fc Acting force on particle
Overall, in addition to the overall properties of various porous filters, Fdam Damping force
the LBM-IMB-DEM model allows for more microscopic simulations tak- Fhyd Hydrodynamic force
ing irregular particle's shape characteristics in to consideration. The Fela Elastic force
model is capable of simulating the interaction between complex particle fi Probability distribution function
shapes and porous media structures, providing a more detailed research feq
i Equilibrium distribution function
method for the clean bed filtration process. All 15 sets of simulation h Square lattices length
show excellent computational stability. The calculation efficiency is Icomb Combined particle's moment of inertia
also satisfying, and the calculation per one million steps takes about Is Constituent particle's moment of inertia
70 min. Kt Tangential stiffness
l Length factor
6. Conclusion m∗ Reduced mass
ms Mass of constituent particle
To simulate the complex fluid-solid two-phase system involving ar- n Weighted average of direction vectors
bitrary solid phase, an improved LBM-IMB-DEM method for the two- P Macroscopic fluid pressure
dimensional fluid-solid interaction is developed in this paper. r Direction vector
The Convex Decomposition method is proposed to handle arbitrary Tc Acting torque on particle
polygonal solid phases collision. This study uses the LBM and BGK Thyd Hydrodynamic torque
scheme for the concern of computational efficiency and ease of Up Particle translational velocity
programming. IMB scheme is adopted and optimized for the character- Us Velocity of solid lattice node
istics of polygonal solid phase boundary. The simulation program in ν Macroscopic fluid kinetic viscosity
MATLAB® environment is completed with the efficiency optimized. Y2D Two-dimensional Young's modulus
The improved progressive scanning algorithm and solid volume fraction Δt LBM time step
fast algorithm can greatly reduce the computation of the lattice node δt Accumulated tangential sliding distance
classification and solid volume fraction calculation. δ_t Tangential sliding velocity
The LBM-IMB-DEM model is verified in detail. The flow pasting a ε Solid volume fraction
rectangle polygon simulation calculates the drag coefficient and the re- μ Friction coefficient.
circulation length corresponding to Reynolds number, which verifies ρ Macroscopic fluid density
the LBM and IMB modules. The particle sedimentation simulation vali- τ Dimensionless relaxation time
dates the model by measuring displacement and corresponding hydro- τDEM DEM time step
dynamic forces applied on the particle. The accuracy of the DEM module v Macroscopic fluid velocity
and the Convex Decomposition method are validated by measuring col- Ω Collision operator
lision system kinetic energy. The above simulation results match well ω Particle angular velocity
with existing literature data or theoretical data. ωi Weighting coefficient
Finally, the simulation of the airborne particle clean bed filtration is
performed to examine the adaptability, capability and efficiency of the
LBM-IMB-DEM method. The simulation illustrates that the model can Acknowledgements
efficiently simulate the interaction between complex particle shapes
and porous media structures, providing a more detailed research The authors are grateful for financial support from the National Key
method for the filtration process. Research and Development Program of China (2018YFB0604903), the
The above work proves that the LBM-IMB-DEM is a promising National Natural Science Foundation of China (Nos. 21776202 and
and efficient method for two-dimensional liquid-solid simulation in 21336007), and the Program for China Scholarship Council (No.
engineering applications. Expanding the current model to three- 201506250017).
dimensional cases would be straightforward in the future studies, how-
ever, how to deliver an efficient 3D DEM for irregular polyhedrons and Appendix A. Program flowchart design
reduce the computational cost is a challenge.
In traditional LBM-IMB-DEM methods, the program flowchart in-
Nomenclature cludes one LBM module for fluid mechanics simulation (see Supple-
A Overlap area mentary Fig. A.1 (a)), one DEM module for particle movement
B Weighting function calculation (mostly circular shape-based particles) and the IMB scheme
C Lattice speed, coupling the above LBM and DEM modules. Supplementary Fig. A.1 (b)
Ccomb Combined particle's centre of mass illustrates the computational flowchart adopted in the present study,
Cs Centroids constituent particles based on which we built the program on the MATLAB® platform.
X. Li et al. / Powder Technology 356 (2019) 177–192 191
Fig. A1. The LBM-IMB-DEM flowcharts comparison: (a) The traditional computational sequence; (b) The flowchart implementation with modified DEM module.
For the LBM module, the program follows the classical LBM calcula- used for all convex shaped particles, including the independent con-
tion process. This work introduces optimized algorithms for the compu- vex particles and the constituent particles composing the combined
tation and does not revise the process arrangement. For the DEM particle.
module, extensive modifications are introduced in the flowchart, with 2) Similar to the first sub-module, the second sub-module is used for
three DEM sub-modules designed for the DEM calculation: particle-wall contact judgment and the corresponding force, torque
calculation. It is designed for the walls which contain stationary and
1) The first sub-module is designed for the inter-particle contact judg-
moving walls.
ment as well as the corresponding force, torque calculation. It is
192 X. Li et al. / Powder Technology 356 (2019) 177–192
3) The third DEM sub-module is the BDF2 solver for Newton's second [9] X.D. Niu, C. Shu, Y.T. Chew, Y. Peng, A momentum exchange-based immersed
boundary-lattice Boltzmann method for simulating incompressible viscous flows,
law, where combined and independent particles' motion status is Phys. Lett. A 354 (2006) 173–182.
determined. Based on the results of the third sub-module, the veloc- [10] W. Tan, H. Wu, G. Zhu, Fluid-structure interaction using lattice Boltzmann method:
ity, angular velocity and position of all particles are updated. moving boundary treatment and discussion of compressible effect, Chem. Eng. Sci.
184 (2018) 273–284.
[11] D. Owen, C. Leonardi, Y. Feng, An efficient framework for fluid–structure interaction
It should be noted that essentially all walls are regarded as regular using the lattice Boltzmann method and immersed moving boundaries, Int. J.
particles in the contact judgment and the force calculation in DEM, but Numer. Methods Eng. 87 (2011) 66–95.
[12] H. Zhang, F.X. Trias, A. Oliva, D. Yang, Y. Tan, S. Shu, Y. Sheng, PIBM: particulate im-
the motion is restricted. So, the velocity, rotation and position updating mersed boundary method for fluid–particle interaction problems, Powder Technol.
of walls are governed by pre-set rules instead of Newton's second law. 272 (2015) 1–13.
[13] Y. Mino, A. Hasegawa, H. Shinto, H. Matsuyama, Lattice-Boltzmann flow simulation
of an oil-in-water emulsion through a coalescing filter: effects of filter structure,
References
Chem. Eng. Sci. 177 (2018) 210–217.
[14] C. Cheng, S.A. Galindo-Torres, X. Zhang, P. Zhang, A. Scheuermann, L. Li, An im-
[1] W. Zhong, A. Yu, X. Liu, Z. Tong, H. Zhang, DEM/CFD-DEM modelling of non-spher-
proved immersed moving boundary for the coupled discrete element lattice
ical particulate systems: theoretical developments and applications, Powder
Boltzmann method, Comput. Fluids 177 (2018) 12–19.
Technol. 302 (2016) 108–152.
[15] S.A. Galindo-Torres, A coupled discrete element lattice Boltzmann method for the
[2] J. Feng, H.H. Hu, D.D. Joseph, Direct simulation of initial value problems for the mo-
simulation of fluid–solid interaction with particles of general shapes, Comput.
tion of solid bodies in a Newtonian fluid part 1. Sedimentation, J. Fluid Mech. 261
Methods Appl. Mech. Eng. 265 (2013) 107–119.
(1994) 95–134.
[16] H.-G. Matuttis, J. Chen, Understanding the Discrete Element Method: Simulation of
[3] G. Tryggvason, B. Bunner, A. Esmaeeli, D. Juric, N. Al-Rawahi, W. Tauber, J. Han, S.
Non-spherical Particles for Granular and Multi-Body Systems, John Wiley & Sons,
Nas, Y.-J. Jan, A front-tracking method for the computations of multiphase flow, J.
2014.
Comput. Phys. 169 (2001) 708–759.
[17] A.A. Mohamad, Lattice Boltzmann Method: Fundamentals and Engineering Applica-
[4] S.R. Idelsohn, E. Oñate, F.D. Pin, The particle finite element method: a powerful tool
tions With Computer Codes, Springer Science & Business Media, 2011.
to solve incompressible flows with free-surfaces and breaking waves, Int. J. Numer.
[18] S. Attaway, Matlab: A Practical Introduction to Programming and Problem Solving,
Methods Eng. 61 (2004) 964–989.
Butterworth-Heinemann, 2013.
[5] D. Noble, J. Torczynski, A Lattice-Boltzmann method for partially saturated compu-
[19] M. Breuer, J. Bernsdorf, T. Zeiser, F. Durst, Accurate computations of the laminar flow
tational cells, Int. J. Modern Phys. C 9 (08) (1998) 1189–1201.
past a square cylinder based on two different methods: lattice-Boltzmann and
[6] A.J. Ladd, Numerical simulations of particulate suspensions via a discretized
finite-volume, Int. J. Heat Fluid Flow 21 (2000) 186–196.
Boltzmann equation. Part 1. Theoretical foundation, J. Fluid Mech. 271 (1994)
[20] M. Wang, Y.T. Feng, C.Y. Wang, Numerical investigation of initiation and propaga-
285–309.
tion of hydraulic fracture using the coupled bonded particle–lattice Boltzmann
[7] A.J. Ladd, Numerical simulations of particulate suspensions via a discretized
method, Comput. Struct. 181 (2017) 32–40.
Boltzmann equation. Part 2. Numerical results, J. Fluid Mech. 271 (1994) 311–339.
[8] K. Zhou, J. Hou, Q. Sun, L. Guo, S. Bing, Q. Du, C. Yao, An efficient LBM-DEM simula-
tion method for suspensions of deformable preformed particle gels, Chem. Eng. Sci.
167 (2017) 288–296.