Sei sulla pagina 1di 89

POLITECNICO DI MILANO

2 .1 . | M A R C H I O E LO G OT I P O

Corso di Laurea Magistrale in Aeronautical Engineering


Dipartimento di Scienze e Tecnologie Aerospaziali
Politecnico di Milano

Immersed Boundary Method in


OpenFOAM: Numerical Validation and
Applications to Wheel Geometries

Relatore: Prof. Paolo Schito


Correlatore: Ing. Giuseppe Porpiglia

Tesi di Laurea di:


Alessandro Ballatore, matricola 919762

Anno Accademico 2019 - 2020


Ringraziamenti

Questo lavoro di tesi suggella il mio percorso accademico al Politecnico di Milano,


un cammino impervio che con le sue asperità ha forgiato lo studente, l’ingegnere,
e, prima ancora, la persona che è in me.
Un primo ringraziamento va al Professor Paolo Schito e all’Ingegner Giuseppe
Porpiglia per avermi sapientemente guidato e istruito nello sviluppo di questo
lavoro, specialmente in un periodo così difficile e demotivante per chiunque.
Voglio ringraziare la mia famiglia, costantemente presente nel sostegno e nell’enorme
fiducia da sempre riposta in me, così come la mia ragazza, che mi ha visto diplo-
marmi al liceo, laurearmi in Triennale e ora in Magistrale, e che ha sempre provato
grande ammirazione nei miei confronti.
Infine, un sentito ringraziamento va a tutti quei miei amici presenti in questi cinque
anni, da quelli della gioventù, passando a quelli del liceo, e arrivando infine alle
nuove conoscenze universitarie. Ho apprezzato il periodo dell’università condivi-
dendo risate e confrontando idee e pensieri con voi.

Milano, ottobre 2020

iii
Sommario

L’approccio body-fitted è uno strumento ormai consolidato per la discretizzazione


del dominio nell’ambito della fluidodinamica computazionale; tuttavia, può risultare
condizionato dalla generazione di una mesh troppo pesante, che può potential-
mente rallentare la simulazione e il relativo tempo computazionale. Il metodo dei
contorni immersi (in inglese, immersed boundary method, o IBM) è una valida
alternativa che permette di utilizzare semplici griglie ortogonali Cartesiane, evi-
tando la generazione di mesh complesse sostituendo il corpo da modellare con delle
forze al contorno che ne riproducono la presenza. Queste forze sono ricostruite per
interpolazione dei valori delle variabili nell’intorno della frontiera stessa. Di con-
seguenza, tale metodo permette di risparmiare tempo nella creazione della mesh,
in quanto la griglia è relativamente semplice e veloce da creare, e conseguente-
mente anche il tempo computazionale appare minore rispetto ad un equivalente
caso body-fitted, dal momento che, per l’appunto, la mesh risulta essere più leg-
gera. Tuttavia, si mostra che, a parità di mesh tra i due metodi, l’algoritmo
implementato in OpenFOAM per il metodo dei contorni immersi è più lento del
body-fitted sulla singola cella, e può essere quindi ancora maggiormente velociz-
zato. L’obiettivo di questo lavoro è quello di provare inizialmente l’equivalenza del
metodo rispetto al classico body-fitted, tramite casi base che spaziano da simu-
lazioni laminari, stazionarie e non, fino a casi turbolenti, e quindi applicarlo ad un
caso 3D di un modello di ruota vera, andando ad investigarne applicabilità e difetti.
Si trova che il metodo IBM funziona bene con simulazioni statiche laminari, soffre
di oscillazioni di pressione spurie con simulazioni dinamiche, e le sue wall func-
tions sono in grado di ricostruire correttamente le forze in direzione tangenziale
ma riscontrano problemi per quelle in direzione normale.

Parole chiave: Metodo dei contorni immersi, Open∇FOAM, foam-extend, Aero-


dinamica della ruota

iv
Abstract

The classical body-fitted mesh approach is a consolidated tool for domain discreti-
sation within the framework of computational fluid dynamics; however, it may
suffer from the generation of a largely heavy mesh, possibly slowing down the sim-
ulation and the associated computational time. The immersed boundary method
(IBM) is a useful alternative that allows to use simple Cartesian orthogonal grids,
avoiding complex mesh generation by substituting the body to be modeled with
boundary forces that mimic its presence. These forces are reconstructed by in-
terpolation of the variables’ values in the neighbourhood of the boundary itself.
Such method hence saves time in the mesh generation step, as the grid is relatively
simple and fast to create, and consequently the computational time appears to be
lower than the one of an equivalent body-fitted case, since, as a matter of fact, the
mesh is much lighter. However, it is shown that, on equally-sized meshes between
the two methods, the algorithm implemented in OpenFOAM for the immersed
boundary method is slower than the body-fitted approach on the single cell, and
thus it can be still improved and sped up. The goal of this work is to first prove the
equivalence of IBM with respect to the body-fitted mesh approach, via benchmark
cases that range from laminar steady and unsteady simulations to turbulent ones,
and then applying the method to a true 3D wheel model, investigating applica-
bility and defects. It is found that the IBM works well with laminar stationary
cases, suffers from spurious oscillations when a dynamic simulation is performed,
and its wall functions are able to just correctly predict the drag but lack in the
computation of forces in normal direction.

Keywords: Immersed Boundary Method, Open∇FOAM, foam-extend, Wheel


aerodynamics

v
Contents

1 Introduction and state of the art[1][2][3][4][15] 7

2 Review of theory[5][6][7][8][9][10][11][12][13][14] 13
2.1 Description of the Immersed Boundary Method[5] . . . . . . . . . . 14
2.1.1 The IBM implementation in foam-extend[5] . . . . . . . . . . 15
2.2 Moving Immersed Boundary[6][13] . . . . . . . . . . . . . . . . . . . 19
2.2.1 Moving IBM surface support[6][13] . . . . . . . . . . . . . . . 19
2.2.2 Calculation of cell volumes and motion fluxes for immersed
cells[13] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 RANS equations[7] . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 2-equations models[7] . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1.1 k - model . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1.2 k -ω model . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1.3 k -ωSST model . . . . . . . . . . . . . . . . . . . . 23
2.3.2 Wall Functions[7] . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2.1 Standard body-fitted wall functions[5] . . . . . . . . 25
2.3.2.2 Immersed boundary wall functions[5] . . . . . . . . 26

3 Numerical Model[16][17][18][19] 27
3.1 Numerical Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.1 Computational domain . . . . . . . . . . . . . . . . . . . . . 27
3.1.2 Numerical schemes . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.3 Numerical solver . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Validation of the method . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Mesh generation . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . 36
3.2.3 Laminar stationary floating cylinder[16][17][18] . . . . . . . . . 37

vi
3.2.3.1 Re = 30, 40 . . . . . . . . . . . . . . . . . . . . . . 37
3.2.3.2 Re = 100, 200 . . . . . . . . . . . . . . . . . . . . . 39
3.2.4 Laminar rotating floating cylinder, Re = 200[19] . . . . . . . 43
3.2.4.1 Spurious pressure oscillations[19] . . . . . . . . . . . 44
3.2.5 Turbulent flat plate . . . . . . . . . . . . . . . . . . . . . . . 46

4 Results 51
4.1 BF reference case mesh . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2 2D cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.1 Static . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2.1.1 Grid generation . . . . . . . . . . . . . . . . . . . . 56
4.2.2 Rotating[19] . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 3D turbulent rotating case . . . . . . . . . . . . . . . . . . . . . . . 60
4.3.1 IBM mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . 66
4.3.3 Numerical schemes . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.4 Numerical solver . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

5 Conclusions and future developments 78

vii
viii
List of figures

• Figure 1.1: wheel rotation influence on aerodynamic forces

• Figure 1.2: vortex regions in the back of the wheel, either stationary (left)
of rotating (right)

• Figure 1.3: rotating contact patch influence on pressure coefficient

• Figure 1.4: rotating contact patch influence on grooved tyre pressure coeffi-
cient

• Figure 1.5: rotating wall approach vs sliding mesh approach, velocity distri-
bution

• Figure 1.6: moving reference frame approach vs sliding mesh approach on


an empty disk; strong pressure gradient in MRF

• Figure 1.7: Coincident body-fitted (left) vs IBM (right) mesh approaches

• Figure 2.1: sketch of the immersed boundary in the solid-fluid interaction

• Figure 2.2: live and dead cells

• Figure 2.3: extended stencil for interpolation

• Figure 2.4: reference frame for the Neumann boundary condition

• Figure 2.5: weight function vectors

• Figure 2.6: interpolation in the fluid cell

• Figure 2.7: moving immersed boundary cells sketch

• Figure 2.8: near-wall dimensionless velocity profiles against dimensionless


distance from the wall

• Figure 3.1: overview of the cylinder domain

• Figure 3.2: overview of the flat plate domain

1
• Figure 3.9: overall cylinder mesh

• Figure 3.10: focus on the immersed boundary cylinder mesh

• Figure 3.11: focus on the body-fitted cylinder mesh

• Figure 3.12: comparison between the two meshes

• Figure 3.13: flat plate grid and its levels of refinement

• Figure 3.3: Laminar stationary cylinder: icoFoam computational time on


single time-step

• Figure 3.4: Laminar stationary cylinder: icoFoam cumulative computational


time

• Figure 3.5: Laminar rotating cylinder: icoDyMFoam computational time on


single time-step

• Figure 3.6: Laminar rotating cylinder: icoDyMFoam cumulative computa-


tional time

• Figure 3.7: Turbulent flat plate: simpleFoam computational time on single


time-step

• Figure 3.8: Turbulent flat plate: simpleFoam cumulative computational time

• Figure 3.16: trend of the residuals of the cylinder simulations at Re = 30, 40

• Figure 3.17: drag history over time at Re = 30, 40

• Figure 3.14: recirculating region in the back wake of the cylinder at Re = 30

• Figure 3.15: recirculating region in the back wake of the cylinder at Re = 40

• Figure 3.18: vortex shedding in the wake cylinder at Re = 100

• Figure 3.19: vortex shedding in the wake cylinder at Re = 200

• Figure 3.20: trend of the residuals of the cylinder simulations at Re = 100,


200

• Figure 3.21: drag history over time at Re = 100, 200

• Figure 3.22: focus on drag oscillations at Re = 100

• Figure 3.23: focus on drag oscillations at Re = 200

2
• Figure 3.24: rotating cylinder wake visualization

• Figure 3.25: drag history over time for both body-fitted and immersed
boundary methods

• Figure 3.26: lift history over time for both body-fitted and immersed bound-
ary methods

• Figure 3.27: turbulent kinetic energy for the flat plate simulation

• Figure 3.28: focus on turbulent kinetic energy for the flat plate

• Figure 3.29: turbulent flat plate residuals for the -˛ model

• Figure 3.30: turbulent flat plate residuals for the -˛ωSST model

• Figure 3.31: viscous x-forces comparison between BF and IBM for the -˛
turbulence model

• Figure 3.2.5: viscous x-forces comparison between BF and IBM for the -˛ωSST
turbulence model

• Figure 3.33: semi-logarithmic plot of near-wall velocity for the -˛ model

• Figure 3.34: semi-logarithmic plot of near-wall velocity for the -˛ωSST model

• Figure 4.4: 2D slick tyre

• Figure 4.5: 2D tyre with 30 lateral grooves

• Figure 4.6: 2D tyre with 50 lateral grooves

• Figure 4.7: grid spacing between the tyre and the floor for the immersed
boundary method

• Figure 4.8: 2D 20 mm cut tyre

• Figure 4.9: deformed tyre

• Figure 4.10: drag spurious oscillations at different time-steps

• Figure 4.11: lift spurious oscillations at different time-steps

• Figure 4.1: 3D wheel model

• Figure 4.2: schematic view and actual mesh on the x-z plane

• Figure 4.3: resultant mesh of the AMI geometry

3
• Figure 4.12: Overall bounding box

• Figure 4.13: x-z plane view of the mesh

• Figure 4.14: y-z plane view of the mesh

• Figure 4.15: x-y plane view of the mesh

• Figure 4.16: x-y plane mesh view, focus on the refinement box

• Figure 4.18: x-z plane view, focus on the refinement box

• Figure 4.17: x-z plane view of the mesh with refineImmersedBoundaryMesh


utility

• Figure 4.20: y-z plane view, focus on the refinement box

• Figure 4.19: y-z plane view of the mesh with refineImmersedBoundaryMesh


utility

• Figure 4.21: x-y plane view of the mesh with refineImmersedBoundaryMesh


utility

• Figure 4.23: Drag coefficient history over time

• Figure 4.24: Lift coefficient history over time

• Figure 4.25: Ventilation moment coefficient history over time

• Figure 4.22: Velocity magnitude Courant number-related issues

• Figure 4.26: First and second order discretisation schemes

• Figure 4.27: Computational time confront between IBM and body-fitted


approach

• Figure 4.28: Velocity magnitude visualizations at 3.2 s, x-z plane

• Figure 4.29: Pressure coefficient visualisations at 3.2 s, x-z plane

• Figure 4.30: Turbulent kinetic energy contours at 3.2 s

• Figure 4.31: Surface LIC visualisations at 3.2 s, x-z plane

4
List of tables

• Table 3.2.3.1: comparison of the IBM laminar cylinder drag coefficient against
BF approach and numerical reference at Re = 30, 40

• Table 3.2.3.1: comparison of the IBM laminar cylinder drag coefficient at Re


= 40 against literature experimental results

• Table 3.2.3.2: comparison of the IBM laminar cylinder drag coefficient against
BF approach and numerical reference at Re = 100, 200

• Table 3.2.4.1: comparison of lift and drag and correspondent coefficients


between IBM and BF for the laminar rotating cylinder

• Table 4.3.5: comparison of lift, drag and moment coefficients between IBM
and body-fitted approach

• Table 4.3.5: comparison of computational times between IBM and BF ap-


proaches

5
6
1. Introduction and state of the art[1][2][3][4][15]

Aerodynamics is one of the key aspects that characterise the behaviour and perfor-
mance of a car and, in general, of a vehicle, e.g. acceleration and fuel consumption,
both mainly influenced by aerodynamic drag.
The total resistance of the flow on a car is estimated to be due for the 25% to
wheels, that hence play an important role within this matter, by means of both
their physical presence and their rotation. The flow around a wheel is quite com-
plex and results from the contributions of vehicle, wheelhouses and rims’ aero-
dynamics, that feature the overall flow topology with dense vortex regions that
depend on whether the wheel is stationary or rotating. More in detail, a horse-
shoe vortex characterises the wake region in both situations; however, the rotation
of the wheel reduces the drag and lift coefficients, as well as vorticity, as below
illustrated:

Figure 1.1: Wheel rotation influence on aerodynamic forces

Also the interaction between the tyre and the ground has a remarkable influence.
In particular, it has been shown that the pressure coefficient (Cp ) spikes beyond 1
in correspondence of the contact patch with rotating wheels. This behaviour has

7
Figure 1.2: Vortex regions in the back of the wheel, either stationary (left) or
rotating (right)

been interpreted as a consequence of tyre jetting effect where the wheel rotation
provides work to the flow. On the other hand, with laterally-grooved tyres, the Cp
has been noted to stay below 1. The explanation is to be found in the suppression of
the wheel jetting by the rain grooves. Moreover, downstream of the contact patch
a negative pressure peak has been measured with values in pressure coefficients up
to -1.5.

Figure 1.3: Rotating contact patch influence on pressure coefficient

8
Figure 1.4: Rotating contact patch influence on grooved tyre pressure coefficient

In order to target the interaction between air and vehicle, two main tools are
wind tunnel (WT) and computational fluid dynamics (CFD): while the former is
more precise but quite expensive in terms of money and time, the latter provides a
useful medium to assess the flow field in an accurate way. Nowadays, the trend is to
focus more and more on CFD when dealing with fluid dynamics, for its flexibility,
comfort and cheapness in costs and times; moreover, as all the main efforts are put
in CFD, this tool is dramatically improving its accuracy with the years passing by.
The standard fashion in flow dynamics simulation is represented by the body-
fitted mesh approach (BF), that consists of generating a grid that conforms to
the geometry of interest, and then solve the equations of conservation (see 2 for
details) along every cell resulting from the spatial discretisation of the domain.
Three different numerical approaches to assess the aerodynamics of a wheel can be
outlined, in an increasing order of accuracy: rotating wall (RW), moving reference
frame (MRF), sliding mesh (SM).
The RW approach is based on a steady simulation being featured with the addition
of a wall-tangential velocity term to reproduce the rotation; however, due to mass
conservation, the component of the velocity normal to the cell surface cannot be
described, as the solver would physically "read" it as an in/outflow through a solid
wall. This aspect is crucial when dealing with rim spooks or lateral tyre grooves,
due to their normal movement.
The MRF approach goes beyond this issue by considering a fluid region, defined by
the user, belonging to a local reference frame, rotating with respect to the global
one. In this way, the approach is able to take into account centrifugal accelerations
and Coriolis effect within the fluid itself. Nevertheless, the particular choice of the
MRF region size and interface location may possibly alter the results; this is typical

9
when dealing with rims, whose spooks are relatively large and influence strongly
the local flow field. Yet, when instead dealing with tyres, it is usually hard to
correctly define the MRF regions.
The SM is the most realistic way of approaching the wheel aerodynamics. It
consists of literally moving, sliding the mesh through an unsteady simulation,
rotating the grid every time-step. It is embedded as a rigid body motion. The
drawback is the consequent high computational cost that leads to a significant
increase in run time. Furthermore, as all the body-fitted approaches, it lacks of
efficiency in refinement, especially in complex geometries.

Figure 1.5: MRF vs SM on an empty disk, strong adverse pressure gradient in


MRF

10
Figure 1.6: Velocity distribution, RW vs SM (desired)

In this sense, the Immersed Boundary Method is a useful and efficient approach
for cheap mesh generation in Computational Fluid Dynamics: it leverages on non-
conformal boundary surfaces through a Cartesian mesh, mimicking the presence
of the body through the imposition of boundary forces, by including additional

11
terms in the governing equations so that the influence of the body on the fluid
is locally included. This is relatively simple because of the use of non-deforming
stationary grids, significantly reducing the computational effort.
There is little in literature about the application of the immersed boundary method:
the most of the matter is treated either in a theoretical way or applied to simple
benchmark cases, and almost no paper deals with a comparison to the correspon-
dent body-fitted case. All of this becomes truer if dealing with wheels and tyres,
as the author has been able to find only one paper focusing on a standalone ro-
tating tyre aerodynamics[1] ; however this study concerns the comparison with the
body-fitted stationary rotating wall velocity approach, and not with the transient
sliding mesh simulation that is, as already stated, the most realistic approach to
the aerodynamics of a wheel.
The following study is hence aimed at proving the equivalence between IBM and
"classical" approach, possibly providing a useful and alternative tool to investigate
the aerodynamics of wheels: the essay starts from the validation and verification of
the method with respect to the correspondent body-fitted mesh case and to what
can be found in literature; then, many simulations on different type of simplified
tyres, and eventually on a true 3D wheel, are performed. The simulations are run
by means of the open source "foam-extend" by Professor Hrvoje Jasak.

Figure 1.7: Coincident body-fitted (left) vs IBM (right) mesh approaches

12
2. Review of theory[5][6][7][8][9][10][11][12][13][14]

The flow of incompressible viscous Newtonian fluid is described by the following


dimensionless Navier-Stokes equations in a Cartesian reference frame (x, y, z ):

∂u∗ 1 2 ∗
+ ∇ · (u∗ u∗ ) = −∇p + ∇u (2.1)
∂t Re
∇ · u∗ = 0 (2.2)
where u is the vector velocity, p is the pressure normalised by the density ρ, Re =
u∞ D is the Reynolds number with ν kinematic viscosity.
ν

The momentum (i.e. Navier-Stokes) equation can be rewritten in the following


vector form:

Au = H − ∇p (2.3)
By taking the divergence of 2.1 and enforcing the incompressibility constraint 2.2,
the Poisson equation is obtained:

∇2 p = −∇ · (u∗ ∇u∗ ) (2.4)

This equation, together with 2.3, expresses the coupling between velocity and
pressure that is usually solved in a segregated manner. Each equation is solved for
the dominant variable, treated as unknown, whereas the other ones are considered
to be known; then, sequential iterations are performed to update the coefficients
of the matrices and source terms until the solution of the coupled equation is
obtained.

13
2.1 Description of the Immersed Boundary Method[5]
In the immersed boundary method, the simulation is run on a grid that does not
conform to the boundary surface, and the immersed boundary is represented by
the surface mesh itself.

Figure 2.1: Sketch of the immersed boundary in the solid-fluid interaction

The immersed boundary (IB) conditions are imposed in the bulk of the mesh and
modify the equations in the cells interacting with the immersed surface. Indeed
there are two different approaches. The first one is the continuous forcing ap-
proach, in which the forcing function concerns a band of cells and requires the
solution of the governing equation also inside the immersed surface. The second
one is the discrete forcing approach, that likewise can be classified according to
the nature of the boundary conditions. There can be an indirect imposition, in
which only the forcing term is included in the discretised equation, but the forcing
function is still spread over the band of cells. On the other hand, the imposition
can be direct, meaning that the discretised equations are modified in order to di-
rectly apply the boundary condition on a single immersed boundary cell. This last
one is the approach used in OpenFOAM and hence in this paper: the immersed
boundary dependent variable is interpolated through the neighbouring cells values
and immersed boundary condition at the corresponding point.

14
Figure 2.2: Live and dead cells

2.1.1 The IBM implementation in foam-extend[5]


In the first step ("predictor"), the intermediate velocity u∗ is predicted by solving
the above Navier-Stokes equations through the pressure and velocities values ob-
tained at the previous step. Then, a corrector loop is performed by imposing IB
conditions to correct the velocity until convergence is reached. Using a quadratic
polynomial interpolation, the boundary condition can be either a Dirichlet or a
Neumann one. For 2D cases, 5 extended cells are required for interpolating each
IB point. This interpolation prevents the relative motion between the fluid in the
IB cells touching the immersed boundary and IB points themselves (i.e. no-slip
BC). The Dirichlet Immersed Boundary Condition reads:

φp = φib +C0 (xP −xib )+C1 (yP −yib )+C2 (xP −xib )(yP −yib )+C3 (xP −xib )2 +C4 (yP −yib )2

where the unknown coefficients of the quadratic polynomial are to be computed


through weighted least square method on the extended stencil.
The Neumann Immersed Boundary condition implies a local coordinate system,
namely x’y’, in which the x’ -axis is directed as the normal to the immersed surface
at the point ib:

φp = C0 + [nib ∗ (∇φ)i b]x0P + C1 yP0 + C3 (x0P )2 + C4 (yP0 )2

15
Figure 2.3: Extended stencil for interpolation

Figure 2.4: Reference frame for the Neumann boundary condition

As regards the weighting factors, two main alternatives are considered: either an
inverse quadratic distance weight function, namely
1
ωi =
ri2

or a cosine weight function, namely


1 ri
ωi = [1 + cos(π )]
2 Srmax

16
Figure 2.5: Weight function vectors

The discretised pressure equation of a fluid cell P that is located in the neighbour-
hood of the immersed boundary is the following:

X 1 X Hp X
nf · (∇p)f Sf = nf · ( )f Sf + nfib · vfib Sfib
f αp f αp fib

and the resultant interpolated velocity is reported in figure 2.6 below.

Figure 2.6: Interpolation in the fluid cell

17
The value of the pressure at immersed boundary faces is needed to solve the mo-
mentum equation, not for the solution of the pressure equation itself, since the
other unknown (i.e. the velocity) is treated as above specified. Hence, once solved
the pressure equation, the Neumann immersed boundary condition quadratic in-
terpolation is used to compute pressure at IB cells and faces through the Poisson
equation:

∇2 pb∗ = −∇ · (u
b ∗ ∇(u
b ∗ )) (2.5)
where the " b " indicates that the immersed boundary condition is applied. The
interpolated velocity at the immersed faces must be scaled so that zero net mass
flux is imposed through the "cage" of IB faces around the immersed boundary. It
is corrected through the updated value of pressure of the above from 2.3:

b ∗∗ = A−1 H − A−1 ∇pb


u (2.6)

Finally, the velocity and pressure are updated:

b ∗∗
un+1 = u (2.7)
pn+1 = pb∗ (2.8)

18
2.2 Moving Immersed Boundary[6][13]
As already stated, the immersed boundary conditions are usually applied by means
of either forcing terms in the governing equations or velocity or momentum in-
terpolation. The near boundary points are tracked both before and after the
forcing/interpolation is applied. This indirect treatment of boundaries helps in
avoiding complexities associated with body-fitted approaches for arbitrary, mov-
ing and deforming geometries. The immersed boundary methods helps in reducing
the costly dynamic meshing, and the issues related to large deforming grids and
the solution interpolation can be bypassed. It can also help in achieving bet-
ter computational efficiency because the domain composition and load balancing
steps for the parallel process are much simpler to implement in a Cartesian mesh
framework.

2.2.1 Moving IBM surface support[6][13]


Within the Finite Volume framework, the moving immersed boundary mesh is
based on a modified integral form of the transport equation:
Z
∂φ I I Z
dV + φ[n · (u − ub )]dS + γ(n · ∇φ)dS = qv dV (2.9)
V ∂t S S V

with ub being the boundary velocity. Motion consistency implies an additional


and auxiliary "conservation of space"-like condition:
Z
∂V I
− (n · ub )dS = 0 (2.10)
V ∂t S

that can be rewritten in a discretised manner:

Vn−V0 X
− Fb = 0 (2.11)
∆t f

2.2.2 Calculation of cell volumes and motion fluxes for im-


mersed cells[13]
During motion, the immersed face sweeps a certain volume (the one in red colour
in figure 2.7) that is computed from the new immersed boundary face and motion
distance, i.e. Vb =x · sn where xb is the geometric motion distance of the body.
While the intersected area varies, swept volume must be distributed. With respect
to figure 2.7:

19
• Cell A expands from cut to full volume, hence there is no IB face in new
configuration

• Cell B is cut both at new and old configuration, hence the IB face is present

• Cell C expands from zero volume to partial volume, hence the IB face is
present

• (Cell D): i.e. the fully swept cell by the immersed boundary, it expands from
zero to full volume, hence there is no IB face

Figure 2.7: Moving immersed boundary cells sketch

20
2.3 RANS equations[7]
The Reynolds-averaged Navier-Stokes (RANS) equations are derived from the
standard Navier-Stokes equations to describe the statistical quantities that feature
turbulence velocity fields. By means of the Reynolds decomposition, the velocity
field is decomposed in the sum of a time-averaged value ū(x) plus a time-dependent
fluctuation part u0 (x, t):

ū(x, t) = ū(x) + u0 (x, t) (2.12)

whereas the continuity equation reads:

∇ · ū = 0 (2.13)
Substituting 2.12 into the standard Navier-Stokes equations and enforcing (2.13)
it yelds:

1
∇ · (ūū) + ∇ · (u0 u0 ) = − ∇p̄ + ν∇2 ū (2.14)
ρ
The second term on the left hand side is the average of the product of the fluctua-
tion parts and it is the tensor representative of the diffusive effect of the turbulent
stresses. This tensor is the responsible for the non-closure of the equations and
hence needs to be modeled, i.e. a new constitutive equation for the turbulent
stresses must be added. The most popular class of models is based on the concept
of eddy viscosity: the anisotropic part of the Reynolds’ stresses tensor is propor-
tional by means of the turbulent viscosity νt (x) defined as:

u∗ = c k (2.15)
dū
νt = u∗ lm = lm
2
(2.16)
dy
where lm is the mixing length:

lm = κy (2.17)
with κ being the von-Kàrmàn constant. There can be eddy viscosity models at 0,
1 o 2 additional differential equations, with increasing complexity.

21
2.3.1 2-equations models[7]
These are the most popular and historically used models. In fact, as the turbulent
viscosity is the product of a velocity and a length, it is natural to think of two
transport equations to model it. In this way, the models obtained are complete.
The turbulent kinetic energy equation is common to all the 2-equations models:

k 3/2
" #
∂k νt ∂k ∂ui

ui = ν+ − u0i u0j − Cd (2.18)
∂xi σk ∂xi ∂xj lm
with Cd = c3 ; what distinguishes one model to another is the second modeled
variable and the correspondent transport equation.

2.3.1.1 k - model

This model, as the name suggests, takes as second quantity to model the second
additional transport equation the dissipation ratio of turbulent kinetic energy :

2
" #
∂  ∂ui ∂ νt ∂
uj = C1 u0i u0j − C2 + ν+ (2.19)
xj k xj k ∂xj σ ∂xj
where the set of coefficient is provided by Lauder & Sharma with C1 = 1.44, C2 =
1.92, σk = 1, σ = 1.3.
This model is the most popular because it is the oldest. However, it lacks of
accuracy for decelerating flows and has troubles when dealing with wall turbulence,
due to the singular term in the second right hand side term of 2.19.

2.3.1.2 k -ω model

Here, the second additional equation is modeled for ω, defined as:



ω= (2.20)
k
that can be interpreted as the frequency of the turbulence. This new variable avoid
the singular term that was present in 2.19; in fact, the equation for the transport
of ω is:
" #
∂ω ω ∂ui ∂ω νt ∂ω

uj = −Cω1 u0i u0j − C ω2 ω 2 + ν+ (2.21)
∂xj k ∂xj ∂xj σω ∂xj
It is less popular than the previous k- model, though it is more numerically robust
and accurate in wall turbulence.

22
2.3.1.3 k -ωSST model

As it has been noted, the k- model is correct in the outer scales while the k-ω
fits better within the inner wall scales. In 1996, Menter thought to combine these
two models to take the best out of them, giving rise to the k-ωSST ("Shear Stress
Transport") model. It is composed of an equation for k and one for , that however
can be evalued for ω as well:
" #
∂k ∂k ∂ νt ∂k

+ Ui = ν+ + Pk − Cµ (2.22)
∂t ∂xi ∂xi σk ∂xi

" #
∂ω ∂ω ∂ νt ∂ω ω

+ Ui = ν+ + γ Pk − βω 2 + (1 − F1 ) 2σωω ∂x
∂k ∂ω
i ∂xi
(2.23)
∂t ∂xi ∂xj σk ∂xi k

where F1 is called blending function and it can be defined as F1 = tanh(χ) with χ


being a complex function depending also on the type of implementation. The value
of the blending function represents a weight included between 0 and 1 that allows
to access, respectively, either to the k-ω model or to the k- one. The name SST
comes from details of implementation. In fact, here the turbulent viscosity νt is
2
modeled in a different way (rather than the previous and classic νt = Cµ k = Cµ ωk ):

a1 k
νt = (2.24)
max(a1 ω, ΩF2 )
This correction is aimed at making the SST model working also with adverse
pressure gradients.

2.3.2 Wall Functions[7]


Even with RANS equations, the presence of the wall gives troubles computationally
speaking, as a high grid resolution near the wall is required, that likewise needs
smaller time-steps. Wall functions model the flow near the wall and modify the wall
drag and turbulence variables. The basic idea is to move the boundary condition
away from the walls, leveraging on the fact that within this regions the velocity
wall log-law applies, as well as the pressure being uniform in normal direction and
the shear stress being almost constant at high Reynolds numbers. To distinguish
the regions near the wall, the concept of y+ needs to be introduced: it is the
dimensionless distance from the wall and it is defined as follows:

yu+
y+ = (2.25)
ν

23
q
where u+ = τρw , τw = ρν ∂u |
∂y y=0
and ν is the cinematic viscosity. As the viscous
sublayer is identified by a value of y + < 5 and the buffer layer by 5 < y + < 30,
in order to apply wall functions the first wall grid point must be placed at 30 <
y + < 300, namely the log-law region.

Figure 2.8: Near-wall velocity profile

Because of wall functions, viscous forces are slightly less accurate with the im-
mersed boundary method with respect to the body-fitted mesh approach. In fact,
in the near wall cell the velocity solution must be split into normal and tangential
components; however, wall functions act on the tangential component only.

24
2.3.2.1 Standard body-fitted wall functions[5]

Below is shown the algorithm of wall functions in a body-fitted mesh case for the
k- model:

• Collect k and near-wall distance y for near-wall cell

• Compute y ∗ through the laminar viscosity νl at the wall:


0.25

C µ ky
y∗ = (2.26)
νl

• If y ∗ indicates log-law region, calculate turbulence dissipation and generation


and account for wall shear by modifying viscosity in the near-wall cell:

νef f n · (∇u)ω
G= (2.27)
Cµ0.25 ky

Cµ0.75 k 1.5
= (2.28)
ky
ky
νω = Cµ0.25 (2.29)
νl
τwall = νω n · (∇u)ω (2.30)

In the near-wall cells, y ∗ is a function of k and u varies to match the log-law


profile according to the change in y ∗ . The introduction of νω is a stable implicit
mechanism to add momentum sink: it responds to near-wall velocity gradient
without division. It is crucial to allow k to respond to the velocity gradient (via
G) and vice-versa (via τwall )

25
2.3.2.2 Immersed boundary wall functions[5]

As regards the immersed boundary, there are some issues related to the wall func-
tions. First of all, the velocity solution in near wall cell must be decomposed into
the normal and tangential component, but wall functions act on the latter only.
Also, the near-wall point is fitted for all variables: implementing wall functions on
the near-wall IB point will not work, as the data for active k and n · (∇u)ω must
be sampled from “live” flow cells. The implementation of wall functions for the k-
in foam-extend follows:

• For each immersed boundary point, introduce the “sampling point”, 150%
further away from the wall

• At the sampling point, perform a least-square fit of fields through the inter-
polation stencil excluding other immersed boundary point

• Based on least-square fit, evaluate near-wall tangential velocity, turbulence


kinetic energy and laminar viscosity

• Calculate y ∗ based on the sampling point near-wall distance and k

• If y indicates log-law region for the sampling point, a log-law fit can be
established to the IB point, otherwise, U will be fitted quadratically, νef f = νl
and G and  are set to zero

• Since all parameters of the least square fit are known, log-law fit for the IB
point can be established:

– Modify G,  and νef f in the IB point (they are not used in actual im-
mersed boundary wall function calculation, but only as a post-processing
result)
– Log-law fit the tangential velocity; wall-normal velocity is fitted quadrat-
ically, as in low-Re flows
– Fitted log-law velocity appears in force balance for active cells and
modified near-wall velocity field.

From this, it follows that, with respect to body-fitted meshes, since the log-law
fit describes correctly the near-wall velocity profile, the drag is identical but the
smoothness of y and y ∗ is, by necessity, lower. Furthermore, since the k transport
equation is not solved in the IB cell, the value of turbulent kinetic energy follows
from local equilibrium.

26
3. Numerical Model[16][17][18][19]

In the following, the immersed boundary method is tested through a series of


benchmark cases in order to explore and understand its behaviour against the
equivalent body-fitted mesh approaches and literature too. First, a description of
the numerical environment is depicted, then simulations on a cylinder and on a
flat plate are investigated.

3.1 Numerical Setup

3.1.1 Computational domain


The validation is carried out on 2D simulations. The bounding box is 20 x 7.5 m.
With respect to the figure, the flow blows uniformly from the left inlet to the right
outlet.

Figure 3.1: Overview of the cylinder domain

27
Figure 3.2: Overview of the flat plate domain

3.1.2 Numerical schemes


As regards the discretization of the terms of the Navier-Stokes equation, for the
two-dimensional simulations it has been chosen to set as follows:

• default Euler scheme for time

• cellLimited Gauss linear 1 for the velocity gradient

• limited corrected 0.33 for Laplacian schemes

• Gauss upwind for all the divergence schemes but velocity, whose scheme is
linearUpwind

• limited corrected 0.33 for snGrad schemes

These schemes are a safe choice in terms of boundedness and convergence, espe-
cially for the upwind scheme that, however, is highly numerically dissipative and
poorly accurate.
Indeed, for the 3D wheel simulation, after having taken confidence and some ex-
perience with the immersed boundary method, it has been decided to switch to
the following discretisation schemes:

• default Euler scheme for time

• cellLimited Gauss linear 1 for the velocity gradient

• Gauss linear uncorrected for Laplacian schemes

28
• Gauss limitedLinear for all the divergence schemes but velocity, whose scheme
is linearUpwind

• uncorrected for snGrad schemes

3.1.3 Numerical solver


The solvers implied are icoIbFoam, icoDyMIbFoam and simpleIbFoam (and their
equivalent IBM algorithm) for, respectively, laminar stationary, laminar rotating
and turbulent stationary simulations. For transient simulations, i.e. icoIbFoam
and icoDyMIbFoam, the ∆t is set to 0.005 s for stationary cases and to 0.0005
s for dynamic ones, as the mesh is finer and hence requires a smaller time-step
to cope with the CFL condition Co = U∆x ∆t
< 1. Such condition states that the
domain of the original numerical PDE ("Partial Differential Equation") problem
must lie within the domain of influence of the method of discretisation (in this
case, finite difference method). This a necessary condition for stability but not a
sufficient one. Also, these two solvers rely on the PISO algorithm, which employs
more than one inner correctors (in this case, 2) to enforce the tightness of the
pressure-velocity coupling of chapter 2.
For the steady simulations, instead, five outer correctors have been set. Moreover,
relaxation factors to smooth the convergence of turbulent quantities have been
employed, setting them to 0.5.
The numerical algorithms are the preconditioned conjugate gradient (CG) for pres-
sure and the biconjugate gradient stabilized (BiCGStab) for velocity and turbu-
lent quantities; to achieve more efficiency, Cholesky and Incomplete Lower-Upper
(ILU0) preconditioners, are, respectively, applied.
Below are shown the differences between immersed boundary and body-fitted ap-
proach in terms of computational effort.
It may seem that the immersed boundary algorithm implemented in foam-extend
is still slower than the body-fitted algorithm; however, this comparison is made
on the same number of cells, and the great advantage of the immersed boundary
method is, as a matter of fact, to save the spatial discretisation size. Thus, the
algorithm looks slower in solving the flow within the single cell, but on the overall
computational time it appears to be cheaper, as it will be shown later on.

29
Figure 3.3: Laminar stationary cylinder: icoFoam computational time on single
time-step

Figure 3.4: Laminar stationary cylinder: icoFoam cumulative computational time

30
Figure 3.5: Laminar rotating cylinder: icoDyMFoam computational time on single
time-step

Figure 3.6: Laminar rotating cylinder: icoDyMFoam cumulative computational


time

31
Figure 3.7: Turbulent flat plate: simpleFoam computational time on single time-
step

Figure 3.8: Turbulent flat plate: simpleFoam cumulative computational time

32
3.2 Validation of the method
The immersed boundary method is validated on a stationary floating cylinder, at
different Reynolds numbers, and compared to equivalent body-fitted approaches
and to literature too. This problem is a useful benchmark as the dynamics of the
flow is governed by the boundary layer separation from the cylinder surface. This
means that inaccurate treatments of the immersed surface would produce spurious
and non-physical perturbations on the shear layer and, therefore, would alter the
dynamics. The flow is modeled as two-dimensional, incompressible and laminar,
by means of the icoIbFoam solver.

3.2.1 Mesh generation


In order to establish the fairest comparison possible, the chosen approach is to
use the same Cartesian mesh for both the IBM and BF cases. The basic principle
is to generate a coarse background mesh and then refining along the boundary
surface to better reconstruct the boundary itself. The overall grid is shown below.
The difference in the two mesh approaches lies in the cells within the boundary;
while on the body-fitted grid one needs to snap the internal mesh, the immersed
boundary method strictly requires the presence of the cells inside the surface, and
also needs the same level of refinement of the outside boundary in correspondence
of the internal boundary. As for the cylinder, there is a wake refinement box
measuring 12.25 x 3 m and other three levels of refinement around the circular
body; it can be seen that an equivalent level of refinement is attained between
the two cases. As regards the flat plate, instead, four simulations with increasing
mesh refinements are studied, the finest of which is reported below. The value of
y + on the first grid point is kept fixed to 46; as the cell dimension decreases, in
order to ensure the same value of y + the viscosity has been progressively reduced.
Consequently, being the inlet velocity kept fixed at 40 m/s, the Reynolds number
value increases with mesh refinement.

33
Figure 3.9: Overview of the IB cylinder mesh

Figure 3.10: IB mesh: increasing cell refinement along the boundary

34
Figure 3.11: Body-fitted mesh: obtained by snapping the IB grid

Figure 3.12: Overlapping view of the two meshes

35
Figure 3.13: Finest grid resolution for the flat plate

3.2.2 Boundary conditions


For the cylinder case, as regards the velocity, an inflow condition is applied at the
inlet (according to the Reynolds number of the specific case); zeroGradient and slip
conditions are imposed, respectively, at the outlet and on the top and bottom of
the domain. For what concerns the pressure, zeroGradient is applied at the inlet,
top and bottom of the rectangular field, while a fixed null value is imposed at the
outlet. On the immersed body, the boundary conditions are imposed as already
described. In the rotating case, a counter-clockwise rotation is applied such that
a tangential velocity equal and opposite to the freestream flow is obtained.
For the flat plate, boundary conditions for turbulent quantities need to be specified
as well. Wall functions are applied on the flat plate and on the top and bottom of
the rectangular box, as they are namely walls, whereas a zeroGradient condition
is applied at the outlet for k,  and ω, calculated for the turbulent viscosity νt .
At the inlet, turbulentIntensityKineticEnergyInlet, turbulentMixingLengthDissipa-
tionRateInlet, turbulentMixingLengthFrequencyInlet and calculated are computed
accordingly from 2.3 and imposed for, respectively, k, , ω and νt .

36
3.2.3 Laminar stationary floating cylinder[16][17][18]
3.2.3.1 Re = 30, 40

The flow presents a steady state with a recirculating region in the wake of the
cylinder, the length of which is bigger as the Reynolds number increases. The
fluid in the upper half of this region rotates clockwise, whereas in the lower half it
rotates counter clockwise.

Figure 3.14: Recirculating region at Re = 30

Figure 3.15: Recirculating region at Re = 40

37
Figure 3.16: Semi-logarithmic plots of Ux residuals for Re = 30, 40

Figure 3.17: Drag history over time, Re = 30, 40

38
Below is shown the computed drag coefficient with respect to equivalent body-
fitted cases as well as numerical and experimental literature: good agreement is
found, both in the numerical results and residuals and drag history behaviour too,
as it can be seen that the trend, up to a small shift, is the same.

Re IBM BF Numerical ref.


30 1.85 1.80 1.78
40 1.65 1.62 1.59

Re Present IBM Kawaguti Thoman and Szewczyk Fornberg


40 1.65 1.618 1.572 1.498

3.2.3.2 Re = 100, 200

Increasing the Reynolds number, while the freestream flow approaches the circular
cylinder in a uniformly steady state, in the rearwards of the cylinder it becomes
unsteady, and the phenomenon of periodic vortex shedding appears by means of
the von Kàrmàn vortex street. More in detail, the vortex stretch and begin to
oscillate in an up-down direction: the fully formed vortex is convected away on
the surrounding main flow, and a vortex subsequently generates on the other side,
then it expands before eventually dissipating and leaving the circular cylinder.
Furthermore, each couple of upper and lower vortices in the rear part of the cylin-
der is collected in an alternative regular array and generate twin rows of vortices,
known as the von Kàrmàn vortex street. The vortex shedding occurs at a dis-
crete frequency, that is a function of the Reynolds number; the Strouhal number
expresses the dimensionless frequency of the vortex shedding, being it defined as
St = uf ∞
D

39
Figure 3.18: Vortex shedding at Re = 100

Figure 3.19: Vortex shedding at Re = 200

Again, the conducted simulations show good agreement, and the trend of residual
and drag history exhibit a practically equal vortex shedding frequency.

Re Present IBM Present BF Numerical ref.


100 1.39 1.43 1.38
200 1.32 1.36 1.37

40
Figure 3.20: Semi-logarithmic plots of Ux residuals for Re = 100, 200

Figure 3.21: Drag history over time, Re = 100, 200

41
Figure 3.22: Re = 100, focus on drag oscillations

Figure 3.23: Re = 200, focus on drag oscillations

42
3.2.4 Laminar rotating floating cylinder, Re = 200[19]
In the following, a floating cylinder rotating counter-clockwise at a rotational speed
of ω = 16rad/s with a freestream Reynolds number of Re = 200 is performed.
Here, a phenomenon known as Magnus effect characterises the fluid dynamics of
the cylinder: the tangential velocity sums up to the freestream one, leading to an
acceleration in the lower side and a deceleration in the upper side, i.e. a suction
and a pressure zone, respectively. This leads to an overall downforce exerted on
the cylinder itself. The wake, from being aligned with the freestream velocity as
in chapter 3, is bent upwards by the rotational speed of the cylinder. The von
Kàrmàn street vortex can be still noted in a similar fashion.

Figure 3.24: Wake visualization

As already stated, the great advantage of the immersed boundary method is ex-
pected to emerge when dealing with moving meshes, as the body-fitted sliding
mesh approach requires the generation of a moving grid for every time-step, that
is a quite computationally expensive procedure. Thus, showing the equivalence of
the method would be beneficial, and this is exactly what this chapter aims at.

43
3.2.4.1 Spurious pressure oscillations[19]

It can be noted that the immersed boundary method suffers from spurious pres-
sure oscillations, more than the equivalent body-fitted mesh case, though the time-
averaged force values are very close. This can be related to the inner nature of
the IBM: as a point placed within the solid body becomes, due to the motion,
that of the fluid, there happens a spatial discontinuity in pressure or temporal
discontinuity in velocity. This very last discontinuity is observed[14] to decrease
with mesh refinement near the immersed boundary and with a higher compu-
tational time step, though the dependence is mostly on the grid spacing. The
dependence on time discretisation size will be shown and analysed in chapter 4.
Also, the Strouhal number, i.e. the vortex shedding frequency, is slightly higher
in the immersed boundary than the body-fitted.

Lift Drag Cl Cd
IBM -1.0340 N 0.4272 N -0.2585 0.1068
BF -1.0767 N 0.4217 N -0.2691 0.1054

Figure 3.25: Drag history over time

44
Figure 3.26: Lift history over time

That said, it seems that the two methods coincide for accuracy of the results and
description of the physics (up to a small shift on the description of the vòn Kàrman
street vortex), but IBM lacks in smoothness; this source of pressure oscillations is
not that relevant in this case, but it can be an important and influencing matter
in more complicated cases, as it will be shown later on.

45
3.2.5 Turbulent flat plate
The next step is to explore and validate the behaviour of the immersed bound-
ary turbulent wall functions. In order to simplify the study, the focus is on the
turbulent flat plate in figure.

Four stationary, turbulent simulations with increasing mesh refinements are re-
ported, the finest of which is pictured in Fig. 3.13. The purpose of the wall
functions, at least in a body-fitted approach, is modeling the flow near the wall,
assuming it to be known, in order to use bigger near-wall cells and hence save
computational cost; however, in the immersed boundary method, the higher the
refinement, the better the solid boundary is interpolated and represented and thus
the truer the flow dynamics. The value of y + on the first grid point is kept fixed
to 46; as the cell dimension decreases, in order to ensure the same value of y + the
viscosity has been progressively reduced. Consequently, being the inlet velocity
kept fixed at 40 m/s, the Reynolds number value increases with mesh refinement.

As before, the starting mesh is the IBM one, then it is just snapped for the body-
fitted mesh case. In the following, according to an increasing level of refinement,
the meshes are labeled as Ref1, Ref2, Ref3, Ref4, corresponding to, respectively,
26694, 65766, 141126, 261402 cells. Notice below the numerical dissipation on
the visualizations of the turbulent kinetic energy due to the particular first order
method employed in the discretization schemes:

Figure 3.27: Visualization of turbulent kinetic energy

As regards the trend of the residuals, the k- model goes to convergence smoothly,
whereas in the k-ωSST model there is an oscillatory behaviour in the finest meshes.

46
Figure 3.28: Focus on TKE

The focus is on the wall shear stresses. As expected, the results show the k- model
incapable to correctly describe viscous forces, especially in the IBM case, whereas
the k-ωSST show good agreement in both approaches, at least in the tangential
direction (recall from chapter 2 that wall functions act only on the tangential
component of the velocity, whereas the velocity must be decomposed both in the
normal and tangential components). This is highlighted by the semi-logarithmic
plot of the module of velocity against wall-normal distance, where IBM and body-
fitted cases show a better agreement when the k-ωSST model is applied, especially
in the viscous sublayer.
The discrepancy between lift outputs in the two methods lies in the effective y +
IBM calculates the variables of interest on, and hence is claimed to be reducible
by refining on the boundary; however, for the present work, this has been scarcely
true, as it has not been possible to close the gap beyond a certain quantity.

47
Figure 3.29: Trend of the residuals, k- model

Figure 3.30: Trend of the residuals, k-ωSST model

48
Figure 3.31: Viscous x-forces comparison for the k- turbulence model

Figure 3.32: Viscous x-forces comparison for the k-ωSST turbulence model

49
Figure 3.33: Semi-log plot of velocity, k- model

Figure 3.34: Semi-log plot of velocity, k-ωSST model

50
4. Results

Finally, a more real wheel model is simulated and studied deep inside through the
immersed boundary method, modelling both the rotation and the turbulence. The
geometry is still somehow simplified, as the rim is absent. This case sums up and
makes a step forward what has been investigated so far in the paper. The wheel
model taken into account is shown in the figure below: the radius is 0.3 m and
the total width is 0.2 m, with four long grooves along its circumference, symmetric
with respect to the longitudinal plane, with a depth of 12 m and width of 10 mm,
each separated from another by 15 mm; the frontal area is hence 0.1183 m2 .

Figure 4.1: Wheel model

51
Before studying and simulating the 3D model through IBM, some 2D cases or
relevant interest are investigated. Then, in order to enforce the validity of the
3D IBM results, the comparison is made against an adaptation to the present
case of what E. Panunzio studied in his master’s thesis [20] (and hence taken as the
reference solution), in which the same wheel model is investigated via a body-fitted
mesh approach.

4.1 BF reference case mesh


For the sake of completion, here is briefly reported from[20] how the reference
body-fitted mesh, consisting of about 5 million cells, was generated. There is
a complex but efficient base mesh featured with a “O-grid” mesh within a tunnel
with a circular hole. In this way, the circular shape around the wheel was correctly
reproduced, reducing the snapping effort and improving cell quality. In order to
avoid an excessive number of cells far from the region of interest, a multi-grading
technique was applied in order to reduce the cell dimension while approaching the
geometry. However, the aspect ratio was maintained as near as possible to the
unity value to avoid high stretched cells, especially next to the wheel. Regarding
the mesh refinement in a 3D case, there are two boxes and one cylinder that include
the fluid regions of major interest, like the volume around and behind the wheel
to correctly reproduce the wake.

Figure 4.2: Schematic view and actual grid on the x-z plane

Moreover, the sliding mesh approach requires a second geometry (in this case, a
cylinder, shown in fig. 4.3) that encloses the body; it is called AMI ("Arbitrary
Mesh Interface") as it works as the interface between those cells, within the AMI,
the that will rotate with respect to surrounding others.

52
Figure 4.3: Resultant mesh of the AMI geometry

53
4.2 2D cases
In the aerodynamics of a racing vehicle, the interaction of the tyre with the ground
plays a fundamental role, as it is one of the main sources of vortices that dramati-
cally increase the drag and affect the aerodynamic performances of the rear devices
(e.g. rear wing, sidepods, floor, ...). Furthermore, when it comes to simulate the
tyre flow by means of CFD, the contact patch represents one of the biggest matter
to deal with. Because of this, in the following, several simulations on different
two-dimensional "grounded" wheel geometries are investigated: slick, 30 grooves,
50 grooves, 20 mm cut, 20 mm deformed. The Reynolds number is set to 200.
The goal of these first two-dimensional simulations is, rather than exploiting the
aerodynamic behaviour and the forces acting on a real wheel, to highlight geomet-
ric details and correspondent issues that could resemble the approaches in the 3D
simulation of an actual tyre.

Figure 4.4: Slick tyre

54
Figure 4.5: 30 grooves tyre

Figure 4.6: 50 grooves tyre

55
4.2.1 Static
4.2.1.1 Grid generation

One of the main troubles encountered in the generation of the Cartesian mesh is
that there is the need to leave at least one empty cell between the ground and the
body in order to allow the interpolation of the immersed surface, otherwise the
simulation would not start. This, however, alters the real aerodynamic behaviour
of the tyre as it allows the flow to pass beneath.

Figure 4.7: Leave at least one cell between the wall patch and the cylinder

Moreover, a higher refinement in the zones of interest is required in order to re-


construct, by interpolation, small geometric details such as the grooves.

56
4.2.2 Rotating[19]
The next step of great interest is to simulate the contact patch whilst the wheel
rotates. To thoroughly achieve this, the deformation of the tyre must be included
and simulated; unfortunately, it has been impossible in this context to predict such
behaviour.
The proposed solution is to simply lower the circular geometry with respect to the
bounding box so that a shape that resembles a squeezed tyre is included. This
approach somehow mimics what in a body-fitted case would be called overset, or
Chimera, grid: mesh regions do not need to deform to follow the motion of the
geometry, and hence it features great control of mesh characteristics locally as the
geometry moves through the domain.
This has been hereby possible only for perfectly rounded geometries, i.e. the cut
wheel could be simulated, whereas the grooved and the deformed ones could not.
This also explains why the geometry (i.e. the .ftr file) needs to remain the very
same circular one: if, instead of lowering the tyre, the geometry itself is cut, there
is no possibility of simulating the rotation, for the aforementioned matter.
Future developments need to focus on the possibility of simulating the rotation of
a deformed tyre as well; the cut model is a first simple approach to deal with the
contact patch, but it does not depict the real behaviour of the compound when
squeezed by the floor.
As noticed in the above, the moving immersed boundary method suffers from
spurious pressure oscillations; this is here enhanced by the contact patch. What is
found in [19] is here verified: the oscillations magnitude and frequency are reduced
and dumped by increasing the time-step size. This may sound counter intuitive,
as, usually, the smaller the time-step size, the higher the precision; however, while
on one hand the computational accuracy does not increase here by reducing the
temporal discretisation step, it also decreases the period of oscillations, hence
leading to a higher frequency.

57
Figure 4.8: 20 mm cut tyre

Figure 4.9: Deformed tyre

58
Figure 4.10: Drag oscillations at different time-steps

Figure 4.11: Lift oscillations at different time-steps

59
4.3 3D turbulent rotating case

4.3.1 IBM mesh


The bounding box measures (15 x 5 x 6) m and the wheel centre is located, from
the inlet, at one third (five meters) of the total box length. Two approaches have
been faced, in order to reach a fair grid resolution around the wheel boundary.
After having generated the background mesh together to five refinement boxes,
it has been decided to further refine the cylinder either with a sixth box or just
on the boundary through the foam-extend’s refineImmersedBoundaryMesh utility;
this obviously results in two different mesh sizes, respectively of 3 and 1 millions
of cells (that, in the following, will be labeled as 3 mln and 1 mln). As the
interpolation of the variables takes place on the boundary only, the smartest choice
is to refine only on the surface of the geometry, saving as many cells as possible
within it. The 3 mln approach, in which all the cells are refined uniformly, is here
reported to underline the importance of both refining only where necessary and
also the inaccuracy of having so many fine cells within the boundary, as it will be
shown next.
The IBM y + value of the first grid cell on the wheel is set to an average of 48,
whereas in the reference body-fitted case has an average of 96 (as the height of the
near-wall cell is of 3.57 mm in IBM and about the half in the body-fitted, 1.785
mm); this is crucial, as the wall functions require a value of y + at least of 30.
However, this aspect needs a further discussion: while on the body-fitted ap-
proach the geometry is a priori defined by the user-input CAD file, in the im-
mersed boundary method the geometry of the body results from the interpolation
of boundary cells: the more refined the grid, the better the interpolation and re-
construction of the geometry details, and viceversa. This is hence a key aspect to
focus on and further develop: the trade-off between the cell dimension to correctly
reproduce the geometry of the body and the dimensionless value of near-wall cell
height to be in accordance with the wall-functions.

60
Figure 4.12: IBM overall bounding box

Figure 4.13: x-z plane mesh view

61
Figure 4.14: y-z plane mesh view

Figure 4.15: x-y plane mesh view

62
Figure 4.16: x-y plane mesh view, focus on the refinement box

Figure 4.17: x-y plane view with the refineImmersedBoundaryMesh utility

63
Figure 4.18: x-z plane view, focus on the refinement box

Figure 4.19: x-z plane view with the refineImmersedBoundaryMesh utility

64
Figure 4.20: y-z plane view, focus on the refinement box

Figure 4.21: y-z plane view with the refineImmersedBoundaryMesh utility

65
4.3.2 Boundary conditions
The freestream flow at the inlet and in the internal field is uniformly set to 12.5
m/s and the rotation of the wheel around the y axis is set accordingly such that
the tangential velocity has the same value but opposite direction. At the outlet
an inletOutlet condition is employed, ensuring more stability and preventing back
flows phenomena with respect to the zeroGradient condition instead. At the walls
a slip condition is applied.
As regards the pressure, being the flow modeled as incompressible, its value cor-
responds to the dynamic pressure, i.e. the static value is null. Accordingly, the
pressure (normalised by the density) has been set to 0 m2 /s2 uniformly at the inlet
and in the internal field. A zeroGradient condition has been applied to all the
remaining patches.
For what concerns the turbulent quantities, it has been chosen to set a low turbu-
lence level, 0.2%; the turbulent length scale is set to 0.1 m, and the other quantities
are hence computed accordingly (see section 2.3.1.3). fixedValue and zeroGradient
are set, respectively, at the inlet and the outlet. On the geometry, wall func-
tions are applied to model by default the viscous sublayer and thus sparing grid
resolution.

4.3.3 Numerical schemes


The applied numerical schemes are the ones already described in section 3.1.2.

4.3.4 Numerical solver


The simulation is run by means of pimpleDyMFoam, in order to account for an
unsteady dynamic and turbulent case. Within this solver, there is the possibility
to opt for either a fixed or an adjustable time-step. The difference is trivial: while
in the former the temporal discretisation is input by the user, in the latter the
time-step is automatically chosen by the solver to help the convergence. In both
case, the choice of the time-step size must cope with the CFL condition, or better,
with a maximum value of the Courant number either provided or desired by the
user. It looks like the obvious and best choice lies in the adjustable time-step: this
will be proven to be wrong, at least for the immersed boundary method. In fact,
a wise choice of the time-step size can both stabilise oscillations and speed up the
simulation; if ∆t is chosen too small, the simulation slows down, the oscillations
grow in magnitude and frequency, whereas, if it is chosen too big, the simulation
can loose accuracy and possibly diverge.

66
If the adjustable time-step is selected, the maximum Courant number has been set
to 10 to possibly speed up the simulation while fairly ensuring the convergence.
This choice will be shown to be right for the body-fitted case but troubling for
IBM, that needs a significantly lower CFL to obtain accurate results. The reason
may be found in figure 4.22: while on average the velocity magnitude floats around
25 m/s (as a rough sum of the freestream and tangential velocity), there are many
constant and periodic huge spurious spikes that burst the Courant number, so the
solution of keeping a low CFL works to balance this unwanted situation, that may
likewise come from some issue in the mesh interpolation. A fixed time-step of
0.0002375 s is then taken into account for IBM simulations, to obtain a Courant
number of approximately 0.83 (computed from Co = U∆x ∆t
with U = 12.5 m/s and
∆x = 3.57 mm).
Both cases deploy 2 inner correctors. It has been decided to set the residual control
to automatically guide the number of outer correctors once reached an established
convergence, i.e. a zero relative and a 10e-4 absolute tolerance for all the variables
but the pressure, whose absolute tolerance has been set to 10e-3.
The simulations have been conducted in parallel by means of the Cineca super-
computing centre on 128 cores.

Figure 4.22: Velocity magnitude Courant number-related defects

67
4.3.5 Results
The obtained results show to be in accordance with what has been found in the
previous chapters: the resultant drag is almost identical to the reference body-
fitted case, whereas the immersed boundary method lacks in accuracy in the lift
computation, for those intrinsic problems of the wall functions already above men-
tioned; consequently, the error on the ventilation moment, i.e. the moment that
acts against the rotation of the wheel, is mainly influenced by the error in lift.
Also, as mentioned earlier, the 3 mln mesh shows to be less accurate than the 1
mln one.
Below are reported the numerical results in terms of aerodynamic coefficients,
averaged over the last 1.7 seconds. They are defined as follows:

L
CL = 1
2
ρU 2 Aref

D
CD = 1
2
ρU 2 Aref
My
CM = 1
2
ρU 2 Aref R
where L, D and My are, respectively, the lift, drag and ventilation moment, ρ is
the density, set to the standard value of 1.225 kg/m3 , U is the freestream velocity
and Aref is the frontal area.

CL CD CM ∆% CL ∆% CD ∆% CM
Ref. (BF) adj. ∆t -0.1383 0.3616 0.0173 - - -
IB 1 mln adj. ∆t -0.2099 0.3742 0.0218 -51.7 +3.5 +26.0
IB 1 mln fixed ∆t -0.1893 0.3568 0.0240 -36.8 +1.3 +38.7
IB 3 mln adj. ∆t -0.2065 0.3814 0.0236 -49.3 +5.5 +36.4
IB 3 mln fixed ∆t -0.2045 0.3719 0.0243 -47.9 +2.8 +40.5

As already investigated, spurious oscillations characterize the history of IBM forces


over time. This is certainly introduced in first place by the rotation of the geometry
(but not by second order, and hence oscillatory, discretisation schemes; if first order
bounded schemes are applied, it is shown in figure 4.26 that the convergence is
reached earlier than second order schemes, but almost nothing changes in the
oscillations), but their amplitude and frequency is shown to mainly depend on
whether the time-step is fixed or variable in time. More in detail, it is shown that
IBM works better if a fixed time-step (in the present case, ∆t = 0.0002375 s) is
chosen. This confirms and places a step forward what was shown and discussed

68
Figure 4.23: Drag coefficient over time

in sections 3.2.4.1 and 4.2.2: the lower the time-step throughout the adjustable
algorithm, the higher the oscillation, and viceversa. Figure 4.23, 4.24, 4.25 picture
the 3 mln and 1 mln IBM cases’ history in time, along with the reference body-
fitted simulation’s one, proving what above stated.

69
Figure 4.24: Lift coefficient over time

70
Figure 4.25: Moment coefficient over time

71
Figure 4.26: First and second order discretisation schemes

72
Another key aspect is the computational effort. In fact, once proved the (potential)
equivalence of the methods, the time required to output the desired results is the
factor that makes one method preferable over another. The immersed boundary
method requires a mesh five times lighter than the equivalent body-fitted case
(that indeed consists of about 5 millions cells) to be both accurate and faster. As
a matter of fact, what was found in the previous chapter, i.e. that, on equally-
sized meshes, the body-fitted approach is faster than the IBM, is here confirmed.
It can be noted that the 3 mln has comparable times with respect to the reference
case when the time-step is fixed, and is even still slower if an adjustable time-
step is selected, despite being about 40% lighter. In other words, the immersed
boundary method can be computationally convenient, provided that the mesh is
wisely generated in terms of number of cells. Moreover, an adequate choice in
terms of fixed time-step can speed up the process, as shown below:

Computational time (s) ∆% t


Ref. (BF) adj. ∆t 5.81 · 104 -
IB 1 mln adj. ∆t 3.85 · 104 -33.7%
IB 1 mln fixed ∆t 2.23 · 104 -61.6%
IB 3 mln adj. ∆t 8.58 · 104 +47.7%
IB 3 mln fixed ∆t 5.78 · 104 -0.5%

Figure 4.27: Computational times comparison between IBM and BF

73
(a) Reference BF

(b) IBM

Figure 4.28: Velocity magnitude visualisations at 3.2 s, x-z plane

74
(a) Reference BF

(b) IBM

Figure 4.29: Pressure coefficient visualisations at 3.2 s, x-z plane

75
(a) Reference BF (b) IBM

Figure 4.30: Turbulent kinetic energy contours at 3.2 s

(a) Reference BF (b) IBM

Figure 4.31: Surface LIC visualisations at 3.2 s, x-z plane

The turbulent kinetic energy isolines pictured in figure 4.30 above show the differ-
ent behaviour of IBM and BF approaches in terms of turbulence. It can be seen
that in the immersed boundary simulation the top turbulent wake is thicker in the
immediate rear of the wheel. Also in the side region the flow looks more disrupted
than in the reference body-fitted case.

76
A final look at velocity and pressure visualisations in figure 4.29 underlines the
differences in the two methods. The lower back part of the wheel is the most
troubling: the reference body-fitted solution provides a low velocity stripe, absent
in the IBM, that represents a separate region of the wake. One one hand, in
Panunzio’s thesis, this was highlighted to be due to the intrinsic nature of unsteady
cases only; on the other hand, all the other simulations performed with steady
RW and MRF approaches resembled what is here found with the IBM case, that
however is unsteady. Despite being counter-intuitive, this seems to suggest that the
dynamic immersed boundary method behaves more like the steady BF approaches
rather than the transient sliding mesh. It can be related, though, to the fact that,
although it is an unsteady simulation, the grid does not physically move, as it
happens in steady BF simulations; this would explain why their flow predictions
look so similar.
Yet, as depicted in figure 4.30, the flow in the immediate neighbourhood of the cir-
cular boundary looks more attached in the IBM than the reference case. Also, the
mid-back part of the wheel is represented differently. The IBM predicts a thicker
and higher velocity/lower pressure flow pulled up by the rotation. These aspect
may be imputable to the non-perfectly rounded shape of the wheel that comes
from interpolation of the variables. This leads to the aforementioned problem of
grid refinement: the finer the mesh is, the better the geometry is represented, but
the slower the simulation and the smaller the y+ values are.

77
5. Conclusions and future developments

The present work has investigated the immersed boundary method by contrast
with the more classical approach of the body-fitted mesh, in the perspective of
tyre aerodynamics. Two-dimensional simulations were at first simulated in order to
validate and get confidence with the method, and to individually analyse aspects
and defects that then all together characterised the 3D simulation of the wheel
model. It has been mostly an exploratory job, as the literature on the immersed
boundary method is poor when dealing with wheel flows.
What has been found is that the method is extremely valid when dealing with
laminar stationary simulations, with practically equal results of the standard ap-
proaches, whereas it has difficulties in the context of a turbulent simulation, as
it has been seen that wall functions have trouble in computing forces in normal
direction.
Yet, in 2D simulations the problem of a truer contact patch modelling has been
shown. Future developments need to focus on the possibility of simulating the
rotation of a deformed tyre as well; the cut model is a first simple approach to
deal with the contact patch, but it does not depict the real behaviour of the
compound when squeezed by the floor.
Moreover, with rotating, and, more in general, moving geometries, spurious oscil-
lations constantly feature the history in time of the quantities of interest, possibly
losing accuracy in the results. IBM requires a low Courant number to obtain fair
numbers. A wise choice has been shown to fix a time-step accordingly, rather than
option for an adjustable ∆t: this stabilises the oscillations of the above and slightly
improves the numerical results too, other than being computationally faster.
Furthermore, the goal that the immersed boundary method aims at is to reduce
operative times and costs, while preserving accuracy. However, while this has
been true for the meshing phase (as the creation of Cartesian orthogonal grids is
objectively way faster than the generation of the body-fitted mesh), on the other
hand a wise trade-off of grid cells in terms of number and refinement topology
needs to be targeted in order for the IBM algorithm to achieve a computation
that is faster than the classical approach. Also, one needs to be careful in grid
refinement, as it improves geometry reconstruction but weighs on the size of the
mesh and possibly falls down the y+ values required by the wall functions, that

78
are so far the only option to simulate turbulence with IBM in the context of foam-
extend.
From this, future developments need to focus on more precise wall functions, ca-
pable of improving the prediction of lift in turbulent situations; another possible
idea could be the development of dedicated wall-functions that would model small
geometric details, e.g. the grooves, through an equivalent roughness, in order to
avoid the aforementioned problem of excessive refinement.
Also, a major hurdle is how to deal with the oscillations of the dynamic simulations’
forces, possibly reducing the dependency from the time-step size and grid spacing.
Finally, the computational algorithm may be slightly sped up, as on the single cells
is still slower than the body-fitted approach (although it is possible to attain fair
results in a faster way by carefully meshing and dosing space cells). However, this
consideration comes from the fact that the comparison has been made on a limited
number of cores and, in general, the aim of this Master’s thesis was not focused on
parallel computing performances. Hence, a pertinent food for thought to further
and better investigate the speed and efficiency of the method could be represented
by a deep study on the level of parallelisation and correspondent performance of
IBM w.r.t. BF approach, as it is reported that the former lends itself better to
high performance parallel computing than the latter.

79
Bibliography

[1] "Aerodynamic analysis of vehicle wheel-housings", Satya Prasad Mavuri, School


of Aerospace, Mechanical Manufacturing Engineering RMIT University Melbourne,
Victoria, Australia March 2009
[2] "Wheel Modelling and Cooling Flow Effects on Car Aerodynamics", Teddy
Hobeika. Department of Mechanics and Maritime Sciences Division of Vehicle En-
gineering and Autonomous Systems Chalmers University of Technology. Gothen-
burg, Sweden 2018
[3] "CFD investigation on wheel rotation modelling", Teddy Hobeika, Simone
Sebben. Road Vehicle Aerodynamics Group, Division of Vehicle Engineering and
Autonomous Systems, Department of Mechanics and Maritime Sciences, Chalmers
University of Technology, Sweden, 2018
[4] "A fluid-particle interaction method for the simulation of particle-laden fluid
problems", Ana Carolina S. Fernandes, Henrique C. Gomes, Eduardo M. B. Campello,
Paulo M. Pimenta. Conference paper, 2017
[5] Immersed Boundary Method in FOAM Theory, Implementation and Use Hrvoje
Jasak and Zeljko Tukovic Chalmers University, Gothenburg Faculty of Mechanical
Engineering and Naval Architecture, Zagreb
[6] "Moving Immersed boundary method", Shang-Gui Cai, Abdellatif Ouahsine,
Julien Favier, Yannick Hoarau. International Journal for Numerical Methods in
Fluids, Wiley, 2017, 85 (5), pp.288 - 323. 10.1002/fld.4382 . hal-01592822
[7] Turbulent Flows, Stephen B. Pope, Cornell University, 2000
[8] "Implementation of a discrete Immersed Boundary Method in OpenFOAM",
E. Constanta, C. Lia, J. Faviera, M. Meldib, P. Meligaa, E. Serrea a Aix-Marseille
Universit e, CNRS, Ecole Centrale Marseille, Laboratoire M2P2 UMR 7340, 13451,
Marseille, France Institut P, CNRS – Universit e de Poitiers – ENSMA
[9] "Simulations of self-propelled anguilliform swimming using the immersed bound-
ary method in OpenFOAM", Hui Feng a, Zhaomeng Wangb, Peter A. Toddc and
Heow Pueh Leea Department of Mechanical Engineering, National University of
Singapore, Singapore;

80
[10] Immersed Boundary Method, Development and Applications. Editors: Roy,
Somnath, De, Ashoke, Balaras, Elias (Eds.)
[11] Numerical Methods for Conservation Laws, Randall J. LeVeque, 1992
[12] Computational Methods for Fluid Dynamics, J. H. Ferziger, M. Perić, 2002
[13] Immersed Boundary Surface Method in foam-extend Hrvoje Jasak Wikki and
University of Zagreb OpenFOAM in Hydraulic Engineering Conference BAW Karl-
sruhe, 21-22/Nov/2018
[14] "Immersed Boundary Technique for Large–Eddy–Simulation", M.D. de Tullio,
R. Verzicco G. Iaccarino, 1DIMeG CEMeC, Politecnico di Bari. 2DIM, Univer-
sita‘ di Roma “Tor Vergata”, 3Center for Turbulence Research, Stanford University,
Stanford, CA, USA
[15] "Aerodymamic Simulation of a Standalone Rotating Tread Tire", Ales Ala-
jbegovic, Phoi-Tack Lew, Richard Shock, Bradley Duncan, Jamie Hoch, 2017
[16] "A fundamental study of the flow past a circular cylinder using Abaqus/CFD",
Masami Sato, and Takaya Kobayashi, Mechanical Design Analysis Corporation
[17] "Experiments on the flow past a circular cylinder at very high Reynolds num-
ber", Anatol Roshko, Guggenheim Aeronautical Laboratory, California Institute
of Technology, Pasadena, California, 1960
[18] "Steady separated flow past a circular cylinder at low Reynolds numbers",
Subhankar Sen, Sanjay Mittal, Gautam Biswas. Department of Mechanical Engi-
neering, Indian Institute of Technology Kanpur, Kanpur 208016, India
[19] "Sources of spurious force oscillations from an immersed boundary method for
moving-body problems", J. Lee, J. Kim, H. Choi, K. Yang, 2011
[20] "Aerodynamics of Rotating Wheels: Analysis of Different Modelling Ap-
proaches", E. Panunzio, 2020

81

Potrebbero piacerti anche