Master of Science in Aerospace Engineering

2MAE203: Advanced Computational Fluid Mechanics

Analysis of flow around bluff bodies using Lattice Boltzmann

Method

Authors:
ACHARYA Sagar
BAVISHI Harshi
PHORNPIMONCHOKE Naphasthanan
 2MAE203: Advanced Computational Fluid Mechanics 24/25

Contents

1 Introduction 2

2 Lattice Boltzmann Method 3

2.1 The Lattice Boltzmann Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Spatial Discretization - Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Velocity Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Geometry and Mesh Generation in LBM . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.5 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.6 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.7 Solver setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.8 Grid Independence Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.9 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.10 Result and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 Under-Resolved DNS 14
3.1 Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Solver Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Grid and Time Independence Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Result and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.1 Drag coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.2 Vortex shedding Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4.3 Strouhal Number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4 Double Sphere 22
4.1 Result and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5 References 26

1
 2MAE203: Advanced Computational Fluid Mechanics 24/25

Abstract
The study investigates the flow characteristics around a single sphere for Reynolds Numbers 100,
300, 1000, 3000, and 10000. A pair of tandem spheres with inter-sphere distances of 2D and 2.5D
were also analyzed for Re 300. Numerical simulations were conducted using two different methods:
Star-CCM+, and the Lattice Boltzmann Method (LBM) implemented in the open-source framework
Palabos through Monopal. The results obtained from both numerical methods were compared with
available experimental data to validate their accuracy. The coefficient of drag (Cd ), and Strouhal’s
number were compared for the respective cases.

1 Introduction
The Lattice Boltzmann Method (LBM) has emerged to be a powerful tool in the computational fluid
dynamics domain. Originally derived from lattice gas models, the fundamental difference between the
two is that LBM tracks the distribution of gas particles instead of the behavior of concrete particles
which is the case of lattice gas models. The lattice gas models were notorious for generating statistical
noise. The reduction of control volume only amplified the fluctuations of density which is useful to
capture the thermal fluctuations of a real gas. However, these fluctuations create artifacts when the goal
is to simulate a macroscopic fluid [1].
The governing equations for LBM are derived from the kinetic theory of gases which do not operate
on continuum mechanics like the Navier-Stokes equation. This is a mesoscopic approach that models
the dynamics of particles on discrete lattices. Its simplicity, scalability, extensibility, and suitability for
complex geometries help us adapt this method to several engineering disciplines. The standard LBM
model is essentially a second-order accurate solver for weakly compressible Navier-Stokes equation.
Given that LBM is different from the traditional flow solvers, it is important for us to understand the
underlying physics and assess the influence of setup parameters such as the domain size, and spatial and
temporal discretization. In this study, LBM is used first to validate the canonical flow over a sphere at a
low Reynolds number (up to Re 10000) in the incompressible flow field. The coefficient of drag (Cd ) and
Strouhal number from velocity fluctuations are computed for different grid refinements. Under-resolved
DNS i.e., laminar flow solver is also employed using StarCCM+ to compare the results. The next step is
to study the interaction between two spheres in tandem arrangement. However, the extent of this study
is limited to the computational resources at hand and time constraints.

2
 2MAE203: Advanced Computational Fluid Mechanics 24/25

2 Lattice Boltzmann Method

2.1 The Lattice Boltzmann Equation

The Boltzmann equation is given by:
∂f ∂f
+ ξ · ∇f + F · = C (f )
∂t ∂ξ
where,
f is the distribution function as f(x, ξ, t) which represents the density of particles with velocity
ξ = ξx , ξy , ξz at position x and time t.
v is the particle velocity.
F represents external forces (gravity, electromagnetic forces, etc.).
C is the collision operator.
The macroscopic transport phenomena are all embedded in the collision operator C(f ). C(f )=( ∂f ∂t )collision
represents the particle distribution for each collision. The Bhatnagar-Gross-Krook (BGK) collision
operator is used for this study to avoid complexity. It considers all possible outcomes of binary collisions
and is only suitable for gas simulations and suitably produces the flow physics for incompressible low
Re flow applications. The BGK collision operator can be given by ( ∂f ∂t )collision = − τ (f − feq ). The
1

distribution function approaches feq , the equilibrium distribution which is the Maxwell-Boltzmann
distribution. τ is the relaxation time which determines how fast a system reaches equilibrium after being
perturbed.
In this study, we do not consider the external force term. The BGK-Boltzmann equation is as follows,
∂f 1
+ ξ · ∇f = − (f − feq )
∂t τ
The discretized BGK Boltzmann equation in the Lattice Boltzmann Method is given by:

fi (x, t) − fieq (x, t)

fi (x + vi ∆t, t + ∆t) = fi (x, t) −
τ
Here, i represents the possibilities to discretize the velocity into finite sets of velocity directions. This
will be elaborated in section 2.2 on spatial discretization in a lattice.
The equations to obtain the macroscopic variables are as follows,
for mass density, Z X
f (x, ξ, t)d3 ξ = ρ(x, t) = fi
i

for momentum density, Z X

f (x, ξ, t)ξd3 ξ = ρu(x, t) = vi fi
i

for total energy density,

|ξ|2 3
Z X1
f (x, ξ, t) d ξ = ρE(x, t) mv 2 fi
2 2 i
i

for internal energy density,

|ξ − u|2 3
Z
f (x, ξ, t) d ξ = ρe(x, t)
2
where,
m is the mass of the particle
u is the macroscopic velocity
vi is discrete particle velocity
fi is the discrete distribution function

3
 2MAE203: Advanced Computational Fluid Mechanics 24/25

2.2 Spatial Discretization - Lattices

Spatial discretization and the corresponding 3D scheme selection is important based on the flow
problem at hand. In our case, two 3D schemes are available namely, D3Q19 and D3Q27. D3Q19 is more
computationally efficient than D3Q27. The fundamental difference is the storage of velocity sets in each
lattice which affect the truncation errors generated. According to [1], D3Q27 was not considered superior
to D3Q19 as it required 40% more memory and computing power. It is recommended for turbulence
modelling while D3Q19 is suitable for laminar flows.

Figure 1: D3Q27 Lattice [1] Figure 2: D3Q19 Lattice [1]

In the current study, even though the preliminary analysis is in the laminar or lower transitional flow
regime, the D3Q27 scheme is used. This is to account for the lack of knowledge of the flow characteristics
in case of interaction of tandem spheres and to have a basis for comparison for the canonical flow study.

2.3 Velocity Sets

As already discussed in section 2.1 the discrete form of the density and momentum density are
ρ(x, t) = i fi and ρu(x, t) = i vi fi respectively. Here, vi = (vix , viy , viz ) at position x and time t.
P P
For D3Q27, i takes the value of 27. These discrete velocities together with a set of weighting coefficients
wi , form velocity sets.
In order to retrieve macroscopic N-S equations in the continuum limit, a constant cs is defined which
essentially represents the speed of sound in the lattice. cs is crucial for numerical scheme stability and it
is used to normalize the physical quantities for accurate fluid dynamics simulations. For isothermal flow,
cs determines the relation between pressure (p) and density (ρ) i.e., p = c2s ρ, which is the isothermal
2
model for the speed of sound. In all velocity sets, c2s = 13 ∆x
∆t2

Notation Velocities vi = (vix , viy , viz ) Number Length |vi | Weight wi

D3Q19 (0,0,0) 1 0 1/3
(±1,0,0), (0,±1,0), (0,0,±1) 6 1√ 1/18
(±1,±1,0), (±1,0,±1), (0,±1,±1) 12 2 1/36
D3Q27 (0,0,0) 1 0 8/27
(±1,0,0), (0,±1,0), (0,0,±1) 6 1√ 2/27
(±1,±1,0), (±1,0,±1), (0,±1,±1) 12 √2 1/54
(±1,±1,±1) 8 3 1/216

Table 1: Lattice Boltzmann velocity sets for D3Q19 and D3Q27 for ∆x = ∆t = 1 i.e., cs = √1
3
[1].

4
 2MAE203: Advanced Computational Fluid Mechanics 24/25

Recalling the BGK collision operator, we can now define the equilibrium distribution function as,

u · vi (u · vi )2 u · u
 
eq
fi (x, t) = wi ρ 1 + 2 + − (1)
cs 2c4s 2c2s

The weights wi are specific to the chosen velocity set shown in P table 1. P
The equilibrium is suchPthat the
eq
moments are the same as those of fi . Hence, we have ρ(x, t) = f
i i = i fi and ρu(x, t) = i vi fi =
P eq
i vi fi
The LBE and NSE can be linked by using the Chapman-Enskog analysis to show that the solution
to the LBE results in macroscopic flow according to the NSE. The kinematic shear viscosity can be
obtained from the relaxation time τ by,
 
∆t
2
ν = cs τ − (2)
2

If required, the stress tensor can be calculated by,

 
∆t X
σαβ ≈ − 1 − viα viβ fineq (3)
2τ
i

fineq = fi − fieq is the non-equilibrium distribution which is the deviation from equilibrium.
The two fundamental steps to update the distribution at each timestep are collision and streaming.
Figure 3 shows the collision and streaming on nodes for a 2D lattice. The complete Boltzmann BGK
equation is,
∆t
fi (x + vi ∆t, t + ∆t) = fi (x, t) − (fi (x, t) − fieq (x, t))
τ
This equation is decomposed into collision (or relaxation) and streaming (or propagation). At first,
we have collision, which is a linear algebraic operation to compute the density ρ and macroscopic velocity
u to compute the equilibrium distribution fieq and post-collision distribution fi⋆ .

∆t
fi⋆ (x, t) = fi (x, t) − (fi (x, t) − fieq (x, t))
τ
Then we stream the distribution to the neighbouring nodes. These are the two operations that are
performed in a single timestep. The distribution function is modified accordingly to obtain the required
boundary conditions.

fi (x + vi ∆t, t + ∆t) = fi⋆ (x, t)

Figure 3: Left: distribution of particles post-collision. Right: distribution of particles before collision [1]

5
 2MAE203: Advanced Computational Fluid Mechanics 24/25

2.4 Geometry and Mesh Generation in LBM

The geometry is generated in the STL format and exported in ASCII encoding. The geometry used
in the study is a single sphere of diameter D for the validation case and two tandem spheres separated
by a distance 2D from their centers.
The mesh for the Lattice Boltzmann Method is generated using a custom Grid Density Function. The
grid density function assigns a value between 0 and 1 to each node in the mesh domain, 0 is assigned to
the coarsest region of the domain and 1 is assigned to the finest region of the domain. The function has
the ability to create multiple levels of refinement within the same grid, which aids in reducing the number
of nodes away from the regions of interest in the study. In our case, we use four levels of refinement in
the mesh. The finest region is the one closest to the sphere and a few sphere diameters downstream in
its wake.

Figure 4: Domain and mesh refinement for the single sphere LBM case

The domain extends 10 sphere diameters upstream, and 20 sphere diameters downstream, laterally,
and in depth. The domain is initially divided into multiple blocks based on the num_blocks parameter
set in the input file. Each block is then further divided based on the octree concept. The minimum
level of refinment in the mesh is of level 3, meaning that the initial block is sub-divided into its octree
components recursively 3 times, while the maximum level of refinement is of level 6, with a total of 4
levels of refinement.

2.5 Stability
For the BGK collision operator, the stability criterion applied to D3Q19 and D3Q27 schemes can be
defined as, r
1 ∆x
|umax | <
3 ∆t
where, |umax | is the maximum achievable velocity magnitude. The constant 0.577 is the limiting value
for stability. This constant value is given as an input under ’ulb’ whose value is set at 0.04 in this study.
The timestep ∆t for each output is computed from the largest cell size ∆x in the domain.

6
 2MAE203: Advanced Computational Fluid Mechanics 24/25

2.6 Boundary Conditions

Boundary Condition Method/Value

x-velocity 0.15 m/s
Sphere wall no-slip
Absorbing layer width 0.02m
Outlet sponge Active

Table 2: Boundary Conditions summary

The velocity inlet boundary condition is used to specify the x-velocity of 0.15m/s. The walls of the
sphere are set to no-slip conditions.
An absorption layer of uniform width is used on the entire simulation domain to dissipate the acoustic
and vortical disturbances that reach the outer boundary, which helps in the application of boundary
condition. In addition to this, it helps with the stability and convergence of the overall simulation.
In addition to the absorption layer, an extra dissipative zone is created as a sponge zone at the outlet
of the domain, to effectively increase the kinematic viscosity of the fluid, as a result, all the vortical
structures dissipate very efficiently.
The boundary conditions in the LBM simulations use the Filippova Hänel method [2] to apply
it on the nodes. It is a type of non-equilibrium boundary condition, which improves the accuracy
by reconstructing the unknown distribution functions at the wall using a combination of equilibrium
distributions (which is based on macroscopic velocity and density) and the non-equilibrium corrections
(which are derived from the interior of the fluid). This method prevents artificial slip effects and improves
accuracy near solid boundaries. It is more accurate for cases which involve curved boundaries(which
represents the current study), moving walls and pressure-driven flows.

2.7 Solver setup

Parameter Method/Value
Density 1.225
Kinematic Viscosity Variable
Lattice D3Q27
Boundary Condition Method Filippova - Hänel [2]
Collision Operator Bhatnagar–Gross–Krook [3]
Number of Processors 24 - 36

Table 3: Solver setup

To simulate the range of Reynold’s number from 100-10000, we use kinematic viscosity as a variable
parameter, while keeping the density, velocity and the diameter of the sphere unchanged. Parallelisation
is also used to run the simulation faster on multiple processors on Rainman, the supercomputer available
for projects at ISAE. On Rainman, we use p16, p24 and bigmem partitions, based on which one is
available for computation at the time of execution.
The simulation saves .vtm files with the solution blocking information, blocks of solution are saved as
.vti files with velocity, strain-rate, and pressure fields. Due to memory limitations, only the velocity field
is saved close to the sphere on the y-plane at every iteration. The simulation records the aerodynamic
forces exerted on the spheres in a log file, which is then used to calculate the coefficient of drag, to compare
with the values from prior experimental measurements, theory and under-resolved DNS simulations
performed on StarCCM+ in the section 3.

7
 2MAE203: Advanced Computational Fluid Mechanics 24/25

2.8 Grid Independence Study

The grid independence study is carried out by using the number of nodes across the diameter of the
sphere as a parameter. Three meshes are generated for the study and are tabulated below.

Mesh Nodes on the diameter Total nodes (in millions)

R0 20 7M
R1 40 14M
R2 52 22M

Table 4: Meshes for grid independence test

Re Mesh Fx Cd Cdexp ∆%
R0 1.265×10−6 1.1456 8.2%
100 R1 1.26×10−6 1.1641 1.08 7.8%
R2 1.19×10−6 1.098 1.7%
R0 5.45×10−7 0.5035 7.8%
1000 R1 5.69×10−7 0.5257 0.467 12.5%
R2 5.36×10−7 0.4956 6%
R0 4.65×10−7 0.4296 7.4%
10000 R1 4.95×10−7 0.4573 0.4 14%
R2 4.62×10−7 0.4268 6%

Table 5: Grid Independence Test Results

Figure 5: Variation in % error with mesh refinement

The grid independence test is carried out for the three meshes in table 4. From table 5, the coefficient
of drag obtained from the LBM is compared with those from experimental measurements [4]. The finest
mesh R2 gives the best results with reference to the drag coefficient. The maximum error is nearly 6%
compared to the experiment.
The deviation of the mesh R0 from the mesh convergence behaviour can be attributed to the very

8
 2MAE203: Advanced Computational Fluid Mechanics 24/25

low spatial discretisation of 20 nodes on the diameter of the sphere, as a result, the simulation does not
converge, and is hence ignored for the grid independence study.
The test is carried out up to 52 nodes across the diameter, following which the number of nodes
in the mesh increase drastically, as a result, the computational memory available for this project is no
longer sufficient to generate the mesh. The grid used for computation is sufficiently refined to capture
the drag force exerted on the sphere, however, it maybe insufficient to conclusively predict the flow
structures in the wake downstream required to determine the vortex shedding frequency.
For the scope of the current study, mesh R2 is used for all further simulations.

2.9 Validation

Re Cd Cdexp ∆%
100 1.098 1.08 1.7%
300 0.6744 0.66 2.2%
1000 0.4956 0.467 3.2%
3000 0.4281 0.4 7%
10000 0.4268 0.4 6%

Table 6: Coefficient of drag for a sphere (LBM)

The simulations run for a physical time of 0.75 seconds, which takes nearly 40 hours of computation
on 32 processors on Rainman.
From table 6 and figure 6, the drag obtained from the LBM computation is compared with the
measurements available from experiment[4]. The Cd from LBM closely matches the experimental values
with a maximum error of 7%.

Figure 6: Variation of coefficient of drag (Cd ) with Reynolds number

9
 2MAE203: Advanced Computational Fluid Mechanics 24/25

2.10 Result and Discussion

As the Reynolds number increases to a few hundreds, the wake of the sphere starts to develop
turbulence. As Reynolds number reaches a few thousands, the wake becomes highly turbulent [5]. LBM
is a mesoscopic method, which requires sufficient grid resolution to accurately capture the small turbulent
scales. However, the current spatial discretization in the wake of the sphere is not fine enough to entirely
capture these turbulent structures. This results in the increasing deviation of Cd from the experimental
values as the Reynolds number increases as seen in figure 5 table 6.

Figure 7: Velocity contour Re 100

Figure 8: Velocity contour Re 300

Figure 9: Velocity contour Re 1000

From figures 7, 8 and 9, we can observe a flow separation and recirculation region in the wake.

10
 2MAE203: Advanced Computational Fluid Mechanics 24/25

(a) Iteration 1800 (b) Iteration 1900

(c) Iteration 1950 (d) Iteration 2000

(e) Iteration 2050 (f) Iteration 2100

(g) Iteration 2150

Figure 10: Velocity contour Re 3000

11
 2MAE203: Advanced Computational Fluid Mechanics 24/25

(a) Iteration 1800 (b) Iteration 1900

(c) Iteration 1950 (d) Iteration 2000

(e) Iteration 2050

Figure 11: Velocity contour Re 10000

At higher Reynolds number, the flow in the wake then breaks down and transitions into eddies that
propagate downstream. This can be seen in the velocity contours for Re 3000 in figure 10 and for Re
10000 in figure 11. The vortex structure seems to diminish faster further downstream due to the coarser
mesh in the wake of the sphere. The wake structure is highly complex, and is observed to shed vortical
structures downstream at both Re=3000 and Re=10000.

12
 2MAE203: Advanced Computational Fluid Mechanics 24/25

(a) Iteration 1800 (b) Iteration 1900

(c) Iteration 1950 (d) Iteration 2000

(e) Iteration 2050

Figure 12: Strain-Rate contour Re 10000

13
 2MAE203: Advanced Computational Fluid Mechanics 24/25

3 Under-Resolved DNS
It is important to understand and compare the flow in the current study with results obtained from
commercial CFD solvers. We use starCCM+ to run simulations for both single sphere and tandem
spheres’ configuration to calculate the Cd and Strouhal’s number from velocity fluctuations in the sphere
wake. These results are then compared with experimental data and results from the LBM simulation.
An under-resolved DNS i.e., laminar flow solver is used to simulate the flow considering low-Re flow
(Remax =10000) which is significantly less than the Re for the onset of drag crisis, the turbulent flow
transition on the surface of the sphere. However, for lower Re values, the flow transition occurs in the
wake from Re 1000 [5].

3.1 Domain
The domain size for each of the cases is shown in figures 13 and 14.

Figure 13: Domain for the single sphere case

Figure 14: Domain for the two spheres in tandem arrangement

For both cases, the slip walls perpendicular to the flow are placed at 10D from the centre of the
sphere(s). A conical wake mesh refinement up to 15D with a 15o spread is done to appropriately capture
the flow physics.

14
 2MAE203: Advanced Computational Fluid Mechanics 24/25

For the current study, x is only limited to 2D and 2.5D.

3.2 Solver Setup

An implicit unsteady flow laminar flow solver with constant density is used for the single and tandem
sphere cases. The initial and boundary conditions are shown in table 7. The dynamic viscosity was
varied for each case to obtain the desired Reynolds number in the simulation.

Initial Conditions
Velocity u(x, 0) = 15 m/s
Pressure p(x, 0) = 101325 Pa
Density ρ(x, 0) = ρ0 (x) = 1.225 kg/m3
Boundary Conditions
Inlet u(x, t) = 15 m/s
Outlet Pgauge = 0
Walls Slip condition

Table 7: Initial and Boundary Conditions

It is to be noted that the simulations on StarCCM+ are performed for a velocity of 15 m/s instead
of 0.15 m/s as the latter generated numerical artifacts at the interface of the wake and the freestream
region for the timestep of 1ms for the least Re of 100 which was computed from Courant number. This
refinement mesh interface generated these artifacts which could potentially be because of the initial
and boundary conditions set up in the given simulation. Since it is an implicit unsteady solver, at
first we chose a timestep with respect to Strouhal’s number from experimental data of the sphere. We
decided to take approximately 50 samples for a single cycle and consider one value higher and one
value lower than this case. However, Strouhal’s number for obtaining the frequency is only valid for Re
above 300 where the flow oscillations exist. For the Re 100 case, we decided to take the courant number
(C = u∆t∆x ) close to 1 which gave us a value of ∆t=1ms for 0.15 m/s. This value was not fine enough to
capture the fluctuations in the sphere for higher Re. Hence, the velocity was changed to 15m/s, and the
corresponding ∆t = 10µs was chosen based on the minimum cell size. This timestep was thus a limiting
case for the simulation.
The boundary layer specifications are shown in table 8. Since we are analyzing for different Re, the
first layer thickness is bound to change. Hence, the value of the first layer thickness is chosen for the
limiting case, i.e., the highest Re 10000. 25 layers are used to make sure the ratio of the consecutive cell
size in the last prism layer does not exceed the recommended values. The cells in the wake region are
roughly 25% of the base size.

y+ 1
First Layer Thickness 2.77 ×10−5 m
Growth Rate 1.15
Number of Layers 25

Table 8: Boundary Layer Specifications

15
 2MAE203: Advanced Computational Fluid Mechanics 24/25

3.3 Grid and Time Independence Study

When sizing the domain, we used the recommended domain size and appropriate boundary layer
refinement to capture near-wall flow. This gave us a domain with approximately 4 million mesh elements.
Given the time constraint and limited memory, we decided to proceed with the timestep independence
study after consulting our professor.
Strouhals’ numbers for the sphere case are computed from experimental simulations in [4]. Table 9
shows the low-mode Strouhal’s number for the respective Re.

Re Strouhal (St) Frequency (Hz) Cycle Timestep (s)

100 - - -
300 0.16 240 0.00417
1000 0.20 300 0.00333
3000 0.23 345 0.00290
10000 0.18 270 0.00370

Table 9: Flow characteristics at different Reynolds numbers

From the above table, we can get the following timesteps to conduct the time independence study.
We perform this study on our limiting case of Re 10000. The ∆t2 is obtained when we divide the time
obtained for Re 10000 by sampling 50 times in a given timestep. It is also to account for Re 3000 having
a lesser timestep value and sampling at a higher rate would ensure flow is captured well for one cycle.
Each simulation is run for a total time of 0.1s.

Timestep (in s) CdUDNS Cdexp ∆%

∆t1 1.4×10−4 s 0.3788 5.3%
∆t2 7×10−5 s 0.3859 0.4 3.5%
∆t3 1×10−5 s 0.3975 0.7%

Table 10: Coefficient of drag Cd at different timesteps for Re 10000

∆t3 is considered for this study. Due to the constraints on time and resources, we could not run the
simulation for a smaller timestep.

3.4 Result and Discussion

3.4.1 Drag coefficient

The Cd value obtained from a CFD analysis on StarCCM+ show very good agreement with experi-
mental results as shown in table 11 and figure 15. We observe the error value for Re 3000 and Re 1000
are slightly higher than that of the lower Re. This could be due to potential wake transition and the
inability of the laminar flow solver to solve for transition and turbulence. That would also explain the
higher drag value which could be attributed to an error in computing the pressure drag on the sphere.

Re Cd Cdexp ∆%
100 1.0843 1.08 0.3%
300 0.6583 0.66 -0.2%
1000 0.4668 0.467 0.0%
3000 0.4039 0.4 1.0%
10000 0.3975 0.4 0.6%

Table 11: Coefficient of drag (Cd ) for a sphere (StarCCM+)

16
 2MAE203: Advanced Computational Fluid Mechanics 24/25

Figure 15: Variation of Cd with Re

3.4.2 Vortex shedding Characteristics

The characteristics of vortex shedding from a sphere in a uniform flow were studied experimentally
and explained by [4]. The experimental data is used to validate the result from the simulation in this
section. Figure 16, shows the nature of the wake for different Reynolds numbers. In [4], regions of vortex
shedding phenomena is divided into four, identified by Reynolds number:

• Region I (300 < Re < 420): Hairpin vortex periodically shed with regularity in strength and
frequency.

• Region II ((480 < Re < 650): Hairpin-shaped vortex is irregular and the shedding point rotates
irregularly and slowly.

• Region III (800 < Re < 3000): Vortex tubes and small vortex loops are formed. The former is
formed by the vortex sheet separating from the sphere surface.

• Region IV (6×103 < Re < 3.7×105 ): the flow stabilised, completely transitioned from laminar to
turbulent in the separated shear layer.

The regions A, B, and C are intermediate regions which are called ’transition’ regions in the study
by Sakamoto [4].

Figure 16: Characteristics of sphere wake from [4]

To compare with the experimental result, fluctuating velocity downstream of the sphere, its FFT,
and vortex pattern visualised by the surface of constant q-criterion are studied. At Re = 100, within no

17
 2MAE203: Advanced Computational Fluid Mechanics 24/25

shedding region, no fluctuation of frequency is observed as shown in figure 17a. Further investigation at
q-criterion = 0.005s−2 as shown in figure 18a. It can be observed that a very weak vortex ring is formed
downstream of the sphere, which is likely to be uncaptured by the velocity probe positioned closer to the
sphere.

(a) Re = 100

(b) Re = 300

(c) Re = 1000

(d) Re = 3000

(e) Re = 10000

Figure 17: FFT of fluctuating velocity at different Re

At Re = 300, within Region I, constant fluctuation in velocity can be recognised as represented

in figure 26b. The fluctuation is periodical and constant, which is in line with what was found by [4].
The peak in the FFT represents the frequency of vortex shedding, which is found to be 211Hz. The
characteristics of the flow can also be seen in figure 18b, which further validates the simulation result
with the visualisation of the hairpin vortex at q-criterion = 30000 s−2 . Additional investigation was
done at q-criterion = 500000 s−2 to capture small-scale, high-intensity vortex structures near the sphere.

18
 2MAE203: Advanced Computational Fluid Mechanics 24/25

(a) Re = 100; Q-criterion = 0.005 s−2

(b) Re = 300 (top: top view, Q-criterion = 500000 s−2 ; middle: top
view; Q-criterion = 30000 s−2 , bottom: side view, Q-criterion =
30000 s−2

(c) Re = 1000; Q-criterion = 30000 s−2

(d) Re = 3000; Q-criterion = 30000 s−2

(e) Re = 10000; Q-criterion = 30000 s−2

Figure 18: Patterns of vortex shedding at different Re

19
 2MAE203: Advanced Computational Fluid Mechanics 24/25

(a) Re = 100

(b) Re = 300

(c) Re = 1000

(d) Re = 3000

(e) Re = 10000

Figure 19: Velocity magnitude field at different Re

20
 2MAE203: Advanced Computational Fluid Mechanics 24/25

At Re = 1000, in Region III, the pattern of vortex shedding is no longer regular. As shown in figure
17c, the simulation can no longer capture the vortex shedding frequency through power spectrum analysis.
This could be explained by the fact that the vortices have begun to transition to a turbulent regime,
making the fluctuation less discernable. Nevertheless, vortex characteristics can still be visualised in
figure 18c, where vortex tubes and small vortex loops can be seen.
According to experimental results, the transition region between Region III and Region IV is where
the laminar to turbulent transition of vortex sheets separating from the sphere occurs. This further
explains the lack of evident peak in figure 17d at Re = 3000. Lastly, at Re = 10000, within Region
IV, two peaks are observed as shown in figure 17e at 106 Hz and 317 Hz. The latter matches with the
experimental results where the low-mode St is expected to be approximately 0.2.
Furthermore, the velocity magnitude field is investigated. The results are in line with previous
observations, as shown in figure 19. At Re = 100, the wake is steady and symmetric. At Re = 300, the
beginning of vortex shedding is visible and a wave pattern can be seen. At higher Re, vortex shedding is
fully developed and transitions to a turbulent wake, as characterised by random disorganised patterns.
Compared to the result using LBM, the result from under-resolved DNS simulation lacks vortex
details at higher Re as turbulent features are smoothed out. The structure of the turbulence is captured
more clearly using LBM simulation.

3.4.3 Strouhal Number

According to [4], effects of Re on St can be classified into the four regions mentioned in the previous
section. However, there is a distinct change in its trend at Re = 1000. At Re = 300 to 1000, St represents
the frequency of hairpin-shaped vortex shedding. Meanwhile, at Re = 1000 to 10000, there exists two
distinct St: low-mode St representing the alternate fluctuation of the wave motion and high-mode St
representing the frequency of the vortex tube shedding.
The values of St in this study is extracted from the peak frequency observed in the FFT as illustrated
in figure 17. At Re = 100, there is no oscillation in the flow, hence the absence of St. St at Re = 300
is retrieved form the peak at frequency of 211 Hz. For Re = 1000 to 10000, the velocity fluctuation is
irregular and there is no discernable peaks in the FFT. However, one of the peaks in the FFT may be
considered or representing low-mode St, as the values are deemed close to the experimental value, as
shown in table 12. The disregarded peaks at a lower frequency do not appear to follow a clear pattern,
and their significance remains uncertain for reasons that are currently unknown.

Re St Stexp ∆%
100 - - -
300 0.14 0.16 -12%
1000 0.26 0.20 13%
3000 0.26 0.23 -2%
10000 0.21 0.18 17%

Table 12: Coefficient of drag for a sphere (StarCCM+)

Additionally, the wave pattern can also be recognised from figure 19c to 19e. This would match
the low-mode St extracted as shown in the table as it represents the progressive wake motion. All of
them exhibit approximately 3 periods of wake before the vortex breaks down into small vortex loops.
On the other hand, high-mode St were not captured in the FFT at any Re. There could be several
factors contributing to this phenomenon. One of which is insufficient grid resolution, since high-mode St
corresponds to small-scale vortex structures near the shear layer. Another factor could be numerical
dissipation which affects high-frequency structures such as small vortices at a relatively higher magnitude.

21
 2MAE203: Advanced Computational Fluid Mechanics 24/25

4 Double Sphere
In the study done by L. Prahl [6], the interaction between two tandem spheres has been analyzed in
steady and pulsating flows at Re 300. Results from this work have been taken to compare the double
sphere simulation. However, due to the time constraint and limited computational resources, analysis
for LBM is only performed for x=2D case. The velocity fluctuations between two spheres and in the
downstream wake of the sphere are exported and a fast Fourier transform is performed.

4.1 Result and Discussion

Tables 13, 14, and 15 show the respective results from experimental and numerical simulations from
StarCCM+ and LBM. According to MIT Open Courseware [5], for Re of a few hundred, the wake starts
to transition where the shearing or strain rate is high.

x Cdupstream ∆Cd St Cddownstream ∆Cd St

2 0.64 – – 0.194 – –
2.5 0.621 0.0025 0.1 0.2, 0.3 0.267 0.0094 0.1 0.2, 0.3

Table 13: Cd for double sphere case in [6]

x Cdupstream ∆Cd St Cddownstream ∆Cd St Cdtotal

LBM
2 0.629 – – 0.189 – – 0.821
2.5 0.610 0.00129 0.112 0.224, 0.333 0.264 0.0189 0.112 0.224, 0.333 -

Table 14: StarCCM+ and LBM Cd for double sphere case

% error % error % error % error % error

x % error St % error St
Cdupstream ∆Cd Cddownstream ∆Cd Cdtotal
LBM
2 -1.79 % – – 0.194 – – -7.5%
2.5 -1.77% -100% 12% 12%, 11% -1.1% 101% 0.1 0.2, 0.3 -

Table 15: StarCCM+ and LBM Cd and St for double sphere case. The value is bold is for the dominant
frequency

Figure 20: Vortex structures from two spheres at x=2D from StarCCM+

22
 2MAE203: Advanced Computational Fluid Mechanics 24/25

Figure 21: Velocity Contour: Wake from two spheres at x=2D from StarCCM+

Figure 22: Velocity Contour: Wake from two spheres at x=2D from LBM

When we observe the flow profiles in figures 21 and 25, there is a separation region and consequent
reattachment on the downstream sphere. The downstream sphere is subjected to minimal flow velocity
as it is present in the wake of the upstream sphere. This results in a lower Cd on the downstream sphere
as shown in tables 14 and 13. This can also be seen in figure 23 in the strain-rate contour. The upstream
flow is highly strained on the first sphere giving us a higher drag value. Hence, the effective Cd on the
downstream sphere is almost 70% less that of a single sphere. It is to be noted that the discretization
along the sphere diameter for LBM simulation is not fine enough to characterize the flow around the
sphere as well as in the wake thus leading to higher error values. The asymmetry in the velocity contour
can be due to the inability of the laminar flow solver to characterize weakly turbulent structures in the
wake, leading to almost negligible oscillating values (of the order 10−5 )of Cd for the case where x=2D.
Steady plane symmetric vortex structures are formed in the wake as shown in 20

23
 2MAE203: Advanced Computational Fluid Mechanics 24/25

Figure 23: Strain Rate from two spheres at x=2D from LBM

For the x=2.5D case, we performed an FFT on the velocity fluctuations between the tandem spheres
and in the downstream flow at 1D distance from the downstream sphere. The Cd values show very good
agreement with the reference values in [6]. The fluctuations in Cd highly deviate from the reference case
as noted in 15. The use of a purely laminar flow solver could lead to these deviations. A transition
model could probably help capture the flow better in the wake. The Strouhal number obtained from the
spectral analysis in figures 26a and 26b gave approximately a 12% error from the reference case. Period
vortical structures are formed as observed in figure 24.

Figure 24: Patterns of vortex shedding from two spheres

Figure 25: Patterns of vortex shedding from two spheres at x=2.5D

24
 2MAE203: Advanced Computational Fluid Mechanics 24/25

(a) Velocity fluctuations from probe between two spheres

(b) Downstream velocity fluctuations

Figure 26: FFT of the fluctuating velocity of two spheres at Re = 300

25
 2MAE203: Advanced Computational Fluid Mechanics 24/25

5 References
[1] Timm Krüger et al. The Lattice Boltzmann Method: Principles and Practice. Graduate Texts in
Physics. Cham: Springer, 2017. isbn: 978-3-319-83103-9. doi: 10.1007/978-3-319-44649-3.
[2] O. Filippova and D. Hänel. “Lattice-Boltzmann simulation of gas-particle flow in filters”. In: Com-
puters Fluids 26.7 (1997), pp. 697–712. issn: 0045-7930. doi: https : / / doi . org / 10 . 1016 /
S0045 - 7930(97 ) 00009 - 1. url: https : / / www . sciencedirect . com / science / article / pii /
S0045793097000091.
[3] M. Krook P. L. Bhatnagar E. P. Gross. “A Model for Collision Processes in Gases. I. Small
Amplitude Processes in Charged and Neutral One-Component Systems”. In: Physical Review 94.1
(1954), pp. 511–525.
[4] Haniu Sakamoto and H Haniu. “A study on vortex shedding from spheres in a uniform flow”. In:
(1990).
[5] Massachusetts Institute of Technology. Chapter 3, Flow past a sphere II - MIT OpenCourseWare.
2025.
[6] L Prahl, Asim Jadoon, and Johan Revstedt. “Interaction between two spheres placed in tandem
arrangement in steady and pulsating flow”. In: International Journal of Multiphase Flow 35.10
(2009), pp. 963–969.

26