You are on page 1of 53

Simulation of Incompressible

Flow Around a Rotating Sphere

VIDA HEIDARPOUR-DEHKORDI

Master of Science Thesis


Stockholm, Sweden 2009
Simulation of Incompressible
Flow Around a Rotating Sphere

VIDA HEIDARPOUR-DEHKORDI

Master’s Thesis in Numerical Analysis (30 ECTS credits)


at the Scientific Computing International Master Program
Royal Institute of Technology year 2009
Supervisor at CSC was Johan Hoffman
Examiner was Michael Hanke

TRITA-CSC-E 2009:097
ISRN-KTH/CSC/E--09/097--SE
ISSN-1653-5715

Royal Institute of Technology


School of Computer Science and Communication

KTH CSC
SE-100 44 Stockholm, Sweden

URL: www.csc.kth.se
Abstract

Numerical computations of incompressible, inviscid flow at high Reynolds


number around a moving sphere, both spinning and non spinning, has been
presented. Particularly we compute forces, i.e. drag and lift, on the ball as
well as modeling the wake and separation. The trajectory of the sphere has
been plotted and the effect of rotation and lift force on the trajectory has been
studied. The G2 method with friction boundary conditions is used to compute
the turbulent flow.
We are testing both the method G2 and the incompressible Navier-Stokes solver
available for us. We show the computed results and how the solver has approx-
imated the flow around a rotating sphere. G2 method that we use to compute
the turbulent flow around a moving sphere gives a good prediction of the aero-
dynamic forces.
The problem to compute turbulent flow in high Reynolds number is the very ex-
pensive resolution of the turbulent boundary layer in high Reynolds caused by
the excessive number of mesh points required near the boundary. Even with to-
day’s parallel computers, it is not possible to perform such computations. The
variety of turbulent methods, e.g. LES, RANS, with different turbulent mod-
els, e.g. wall model, Reynold stress, have been used to solve turbulent flows.
The method we use to compute the flow around the sphere is G2 method with
friction boundary conditions. The G2 method is a stabilized Galerkin finite ele-
ment method using continues, piecewise linear trial functions in time and space
with continues piecewise linear test functions in space and piecewise constant
test functions in time. The method uses to compute approximate solutions to
the Navier-Stokes equation. Instead of resolving the turbulent boundary layer,
the method uses a friction boundary condition as a wall model.
Sammanfattning

Simulering av inkompressibel strömning runt en roterande sfär


Numeriska beräkningar av incompressibelt, icke-visköst flöde vid höga
Reynoldstal runt en boll presenteras, bade med och utan rotation. Speciellt
beräknas krafter på bollen, med modellering avseparation och vak. Trajektori-
an för bollen har plottats och effekten av rotation med avseende på lyftkraft
har studerats. En G2-metod med friktionsrandvillkor används för att beräkna
turbulent flöde.
Vi testar både metoden G2 och den för oss tillgängliga lösaren för de inkom-
pressibla Navier-Stokes ekvationer.Metoden G2 ger en bra förutsägelse av aero-
dynamiska krafter. Detta är den metod vi använder för att beräkna turbulent
strömning runt en rörlig sfär.
Problemet att beräkna turbulent strömning för höga Reynoldstal är att det
är mycket dyrt att lösa upp det turbulenta gränsskiktet. Även med dagens
parallella datorer är det inte möjligt att göra sådana beräkningar. Diverse me-
toder med turbulensmodellering har tidigare använts, t.ex. LES och RANS.
I den metod vi använder här löser vi inte upp turbulenta gränsskikt, istället
använder vi ett friktionsrandvillkor som väggmodell. Methoden G2 är ett sta-
biliserat Galerkin finita element-metoden fortsätter piecewise linjära rättegång
funktioner i tid och rum med fortsätter piecewise linjära testa funktionerna i
rymden och piecewise konstant testa funktionerna i tid. Metoden används för
att beräkna ungefärliga lösningar till Navier-Stokes ekvation. Metoden använ-
der en friktion gränsen skick som en vägg modell, i stället för att lösa turbulent
boundary layer.
Contents

Simulering av inkompressibel strömning runt en roterande sfär . . . iv

1 Introduction 1

2 Physical concepts of the model 5


2.1 Wake and separation . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Drag and drag crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Lift, Magnus effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Mathematics and Methodology 7


3.1 The Navier-Stokes equations . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 The boundary Condition . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2.1 Skin Friction Wall Model . . . . . . . . . . . . . . . . . . . . 8
3.2.2 Slip Boundary Condition . . . . . . . . . . . . . . . . . . . . 9
3.2.3 No Slip Boundary Condition . . . . . . . . . . . . . . . . . . 9
3.3 Other works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4 Some methods for turbulent flow . . . . . . . . . . . . . . . . . . . . 10
3.4.1 Discretization cG(1)cG(1) . . . . . . . . . . . . . . . . . . . . 12
3.5 Stability of G2 method . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.6 Computation of forces . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.6.1 Computation of Drag . . . . . . . . . . . . . . . . . . . . . . 14
3.6.2 Computation of Lift . . . . . . . . . . . . . . . . . . . . . . . 15
3.6.3 Computation of Gravity . . . . . . . . . . . . . . . . . . . . . 15
3.7 Rotating ball . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.8 Trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4 Computations 19
4.1 The Computational model . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1.1 Still Ball Model . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1.2 Rotating Ball Model . . . . . . . . . . . . . . . . . . . . . . . 20
4.1.3 Trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 The Computational results . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 Still Ball . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.2 Rotating Ball . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2.3 Trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5 Summary 43
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

Bibliography 45
Chapter 1

Introduction

In this work we present numerical computations of the incompressible flow around


a moving sphere in air. Both cases of rotating sphere and non-rotating sphere have
been considered. A prepossessing step will be taken by constructing a simple model
for the geometry with the use of the COMSOL multiphysics software. The geometry
will be transformed to the solver. After the completion of the solving process of the
partial differential equation which is to be performed by the solver, the results have
been presented and discussed using MATLAB, Paraview, and C++ programming.
For this simple model, we are testing the specific solver used in our project and we
will investigate the outcomes as well as how qualitatively reasonable the results are.
The model (the rotational moving sphere) which is a simple test case, is interesting
from different point of views. The particle transport in the engineering point of
view, or trajectories of spinning balls in sports. In every sports where a sport’s ball
is being used, it is very important to realize the path in which the ball is traveling
on.
Different methods, e.g. DNS (Direct Numerical Simulation), LES (Large Eddy
Simulation), RANS (Reynolds Averaged Navier-Stokes), can be used to compute the
turbulent flow. The mentioned methods shortly explained in Chapter 2. However
DNS is very expensive when it comes to solving the turbulent flow. The reason is
that DNS will fully resolve for all the physical scales of the model. In turbulent
flow, the number of mesh points needed for full resolution are approximated to be
Re3 [6] and a boundary layer around the moving object must be solved. Therefore
using DNS won’t be a good idea. Most real life experiences are at higher Re. We
will have turbulent flow in Re > 106 , and what is obvious is that it is impossible
to use DNS with the computer systems available. Therefore, usually other methods
are used which all of them use turbulence modeling. Turbulence modeling is a wide
area of research and different resolution methods have been offered to get around
the turbulent modeling problem. In this work we will just take a quick look at
some of the suggested methods and a little explanation of the method we used (G2)
and focus on the simulated results. Inspired by previous works done in [3, 6, 8]
regarding bluff bodies moving through air to predict the mean value of drag and

1
Introduction

lift for a non-rotating sphere, we will use the same method to model the rotating
sphere. In particular, an adaptive General Galerkin(G2) method with skin friction
boundary condition is used. We are using the icns, which is an incompressible
Navier-Stokes solver and is part of Unicorn solver that is developed by FEniCS
project’s group, to simulate the turbulent flow around a rotating sphere. It is
relevant to mention that the G2 method was implemented to the incompressible
Navier-Stokes solver which we are using. This means choosing this solver, we are
using the G2 method automatically to compute the approximate values of interest.
There is a brief explanation in the following chapters of this project about the G2
method and skin friction boundary conditions. However, more specific information
about the method can be found in [3, 4, 6, 8].
G2 doesn’t resolve the boundary layer to compute the mean value of drag and
lift of turbulent flow. We get a good approximation of wake, separation point(s),
and mean value forces. In [8], we notice that an adaptive G2 method is based on
a posteriori error estimate which is the combination of the residual, and the dual
weight.
The FEniCS Project is a set of free software projects with the common goal to
enable automated solution of differential equations. It has many components and
one of them is the Unicorn solver which is an adaptive finite element solver for
fluid and structure mechanics that uses G2. Unicorn is based on the python/C++
interface of FEniCS called Dolfin. Dolfin provides an environment for solving partial
and ordinary differential equations1 .
The investigation of flow around a sphere can be an interesting subject in many
fields of studies such as in CFD, sports,etc. It is important in many sports to give
the sport’s ball a spin that it needs to change the straight path that it is moving
in, e.g. both in topspin and backspin in table tennis.
Also, this is an interesting test case to test the performance of the solver in use.
The computations we are doing are also important scientifically speaking because
the results of these computation can support the new resolution of d’Alambert’s
paradox, which was given by Claes Johnson and Johan Hoffman [9].
In this work we shall show that one can predict mean values of interest for
rotating sphere in small viscosity flows without resolving boundary layers, using
Adaptive G2 with slip and friction boundary conditions.
The flow separation for circular shapes depends on the Reynolds number. There-
fore, in modeling the sphere it is interesting to see how the separation point(s) and
the wake and therefore the drag will act due to changes of Re. We also can check
the accuracy of the method comparing the results from the wind tunnel experi-
ments with the results from the solver or we can study the computational cost of
the algorithm we are using, e.g. computing time and memory requirement.
In the model, skin friction (noted as β parameter) is coupled to energy dissi-
pation. Increasing the skin friction at the boundary will cause the momentum to
decrease which will result in an earlier separation and a bigger wake with an increase

1
For more information about the solver refer to http://www.fenics.org

2
Introduction

in drag. This means that friction parameter β is coupled to velocity (velocity∼ Re).
An increase in β will model a decrease in Re as well as the velocity of the sphere
while it moves in the air. The separation point(s) of circular shape’s objects such
as sphere are related closely to Re so that in our experiments by changing β, we
note a sudden change of the separation points (and wake and drag). We connect
this behavior to the drag crisis of Re ∼ 105 which corresponds to modeling the
boundary layer as it is changing from turbulent to laminar flow. This can be con-
cluded from the previous experimental results in [6]. A brief conclusion from the
mentioned work is that for a laminar boundary layer, i.e. 103 ≤ Re ≤ 105 , the
positions in which the flow will separate from the object stay almost the same with
no big change. However, as the boundary layer changes to turbulent, i.e. 105 ≤ Re,
a visible delay on the separation of the flow from the boundary will occur.
In Chapter 2, the physical explanation of the model is presented. Concepts such as
Drag crisis, Magnus effect and boundary layer theory have been presented in this
chapter. In Chapter 3, we give a summary of the history of turbulent flow and com-
putational methods as well as an explanation about the mathematical formulation
we used to solve the problem. Also, a summary about the method used and the
well-posedness and stability of the method is given. In Chapter 4, The computa-
tional model and the computational results as well as the conclusion is presented.
First, We consider the problem of a non-rotating sphere with several different fric-
tion coefficients, in different simulations. Then we will compute the flow around a
rotating ball (clock-wise and counter clock-wise) with friction boundary condition
where the skin friction is modeled to change linearly during time. Computational
models and results will be presented. At the end of this chapter the Conclusion and
future work will be presented.

3
Chapter 2

Physical concepts of the model

In this chapter, I would like to explain some of the physical concepts that we will
use in this paper.

2.1 Wake and separation


In CFD, the separation point(s) are the points in which the fluid that is moving tan-
gentially to the boundary separate from the boundary. The flow tends to separate
earlier for a laminar boundary layer with low tangential velocity near the boundary.
The track of turbulence that is left behind the moving object in the air is called
wake.

2.2 Drag and drag crisis


Drag is an aerodynamic force. Aerodynamic forces arise from two causes: (i) pres-
sure on the surface of the object, (ii) viscosity (skin friction). Drag force will be
generated when a sphere shaped object such as a sport ball is moving in air due to
(i), (ii). This force will be in the opposite direction to the moving ball and will resist
with the movement of the ball which will cause the velocity of the ball to drop. For
high Reynolds numbers, the skin friction plays a small part in determining the drag
force and instead the most determining drag force is the one due to pressure which
is coupling to separation of the flow. In high Re, one can explain the drag force by
the separation point(s). According to [7] the earlier the separation, the higher the
pressure drag, and therefore is the drag force.
While the delay in separation will cause a smaller wake behind the ball and
the drag force will drop noticeably and the ball will be permitted to travel further
than regularly expected. This sudden, high reduction of drag force is known as drag
crisis.
Drag crisis happens for 105 ≤ Re ≤ 106 . This interval happens to correspond
to when a laminar boundary layer undergoes transition to become turbulent. The
increase of momentum of the turbulent boundary layer will cause the delay in sep-

5
Physical concepts of the model

aration and therefore less drag force will be generated. One physical example of
drag crisis is the dimples of a golf ball. The effect of these dimples is to produce
turbulent boundary layers so that the ball can travel further with a moderate veloc-
ity (no need of high Re to produce turbulent boundary layer), according to Prandtl
and Wieselsberger’s experiments.

2.3 Lift, Magnus effect


When a sport ball rotates around its axis, in addition to its straight forward motion
(which will produce the drag ), another aerodynamic force called the Lift force is
also produced which is prependicular to the direction of the rotation which depends
on the angular velocity of the sphere as well as the direction of the movement. The
lift force is force that is perpendicular to the flow direction. It contrasts with the
drag force, which is parallel to the flow direction. The effect of the lift force on
the rotating object is known as Magnus effect or reverse Magnus effect depending
on the direction of rotation.The expression of the Magnus effect comes from Lord
Rayleigh, who credited Heinrich Gustav Magnus for the first explanation of the
lateral deflection of a spinning ball [7].
In most sports played by ball, these effects will play a major role in changing the
score of the game by changing the path of the ball1 . The physical example and
explanation of Magnus effect and reverse Magnus effect:
1) The traditional explanation is based on Bernoulli’s law stating that the reason
of Magnus effect is the pressure difference between two sides of the ball caused by
relative velocity vectors on each side. It claims that on the side of the ball which
is moving in the same direction of wind flow, an increase of the velocity (decrease
of pressure) will be generated and on the other side which has a velocity vector
opposite to the wind flow, a decrease in that side will be observed. Moreover, the
pressure difference will cause the ball to move toward the side that has a lower
pressure.
2) A more modern understanding of the Magnus effect is attributed to asymmet-
ric separation in turbulent boundary layers due to different relative velocities, see
e.g. [7]. In this approach the separation points are used to explain the phenomenon
and basically the cause mentioned in this approach is the horizontally unsymmet-
rical separation. This is when the delay in separation will cause turbulence in that
side as well as producing the other side to be laminar due to premature separation.
As the turbulent flow is relative to the increase of the momentum, this additional
momentum should transfer to work (here it is the generation of Magnus force).
Using this notion, it is possible to explain the topspin and backspin in tennis game.

1
An example is the free-kick of Roberto Carlos against France in 1997.

6
Chapter 3

Mathematics and Methodology

One can use the familiar Navier Stokes (N.S.) equations which were formulated
1825 − 45 to compute the fluid flow. These equations are widely used in
Computational Fluid Dynamics (CFD) for computing both laminar and turbulent
flow. We are going to use them for our computations of the sports’ ball as well.
The N.S. equations are nonlinear partial differential equations in almost every real
situation, with the exceptions of one dimensional flow, and Stokes (creeping) flow.
The nonlinearity makes most problems difficult or impossible to solve.

3.1 The Navier-Stokes equations

The most general Navier-Stokes equation end up being like (3.1), which is indicating
the conservation of momentum. The second equation is the Continuity equation and
here it indicates the conservation of mass.

∂u
 
ρ + u · ∇u = −∇p + ∇ · T + f (3.1)
∂t
∂ρ
+ ∇ · (ρu) = 0. (3.2)
∂t

Here u is the flow velocity vector, ρ is the fluid density, p is the pressure, T is the
deviatoric stress tensor, and describes viscous forces, and the stress tensor is defined
by σij = −pI + T.
T in equation(3.1) has too many unknowns which makes the original N.S. not that
usefull for solving problems. Instead assumptions will be made on viscous behavior
which will eventually give us a useful equation. For incompressible, Newtonian flow
∇ · T = ν 4 u.
For incompressible, Newtonian, with unit density (ρ = 1) fluid in volume Ω the

7
Mathematics and Methodology

Navier-Stokes equations that are used altogether will be reduced to ( [4]),


∂u
+ (u · 5)u − ν 4 u + 5p = f in Ω × I, (3.3)
∂t
∇·u=0 in Ω × I,
u=0 on ∂Ω × I,
u(., 0) = u0 in Ω,
Under the assumption of high Re, the kinematic viscosity ν > 0 is constant in
(3.3). Moreover, the flow is enclosed in the volume Ω in R3 with homogeneous
Dirichlet boundary conditions with f being a given driving force, u0 initial data
and I = (0, T ) is the time interval.
In R3 , ∇ = ( ∂x∂ 1 , ∂x∂ 2 , ∂x∂ 3 ) and with the quantity ν 4 u − 5p representing the
total fluid force, and may expressed as:
ν 4 u − 5p = 5 · σ(u, p)
where σ(u, p) = (σij (u, p)) is the stress tensor where
σij (u, p) = 2νij (u) − pδij
| {z } |{z}
stress deviatoric isotropic pressure

With the Kronecker delta function δij and the strain rate tensor ij (u) = (ui,j +
uj,i )/2 where ui,j = ∂ui /∂xj , i, j changing from 1 to 3.
By definition Re = U L/ν. This along with the assumption of (3.3) being normal-
ized, gives the conclusion of Re = ν −1 . More details can be find in [3], [7].

3.2 The boundary Condition


It is not possible to resolve the turbulent boundary layer for high Re using DNS
(Direct Numerical Simulation). This is because DNS resolves all physical scales of
of the model and since for turbulent boundary layer we will need too many mesh
points, it will not be an ideal possible approach. Usually a wall model will be used
to overcome the issue.
For a wall model one may define a wall function and model the mean profile of
the turbulent flow, or use different methods in different parts of the domain such
as RANS (Reynolds Averaged Navier Stokes) near boundary, and LES (Large Eddy
Simulations) for the remaining parts of the domain.
We will use the same simple wall model as was used in [7], in which it has been
referred to it as the skin friction model.

3.2.1 Skin Friction Wall Model


Using a simple wall model such as the skin friction model gives us the chance to
avoid resolving the turbulent boundary layer fully. In this work we will use a skin
friction model to simulate the flow past a rotating sphere.

8
Mathematics and Methodology

The boundary condition model that has been used is as follow:

u · n + αnT σn = 0, (3.4)

u · τk + β −1 nT στk = 0, k = 1, 2 (3.5)

These equations are implemented weakly to the discrete equation of cG(1)cG(1) [7]
and this boundary condition was implemented to the mesh points on the boundary
of the sphere. More details on this boundary condition can be found in [5]. The
first equation is used to define the penetration of the boundary by modifying α,
which in our case is always zero since the sphere is solid and the flow will not go
through it. The second equation implements friction by the β parameter which
corresponds to slip b.c. with β = 0 and nonslip b.c. with β → ∞. β is based
on: Re, the roughness of the body of the sphere (skin friction τ ), etc. Increasing β
will cause an increase in the friction of the boundary, while decreasing β will model
reduced (normalized) friction which will cause an increase of velocity and therefore
increasing Re.
Velocity and length are both scaled to be of unit size in the computations and
using β we are able to model the velocity. Choosing β to be a function of velocity in
time will couple the value of beta to velocity and therefore to the Reynolds number.
By increasing β we model the velocity (Re number) of the sphere being decreased as
it moves in the air. Re is coupled to velocity with Re = UνL , where in the physical
scale L is the actual diameter of sphere and it is fixed in time. ν is the kinematic
viscosity of air (' 10−5 m2 /s) which is constant in time, assuming the temperature
to be constant (kinematic viscosity depends mostly on temperature).

3.2.2 Slip Boundary Condition


A slip boundary condition is when there is no friction or resistance of the boundary
of the object against the flow moving on it, another way of saying is when the profile
of velocity will not change along the boundary. It can been shown with the following
mathematical term : u · n = 0. Using β = 0 in formula (2.5) will supply the slip
boundary. By using Adaptive G2 method with slip boundary condition, one is able
to solve the Euler equation. In the solver for incompressible Navier Stokes, available
in Unicorn we declare the viscosity term in Navier Stokes to be zero. Since in the
case that boundary condition is slip (and the skin friction will vanish too), there
will be no artificial viscosity. In this case we will be solving the Euler equations.

3.2.3 No Slip Boundary Condition


No slip conditions refer to the state in which the flow will not slide on the boundary
of the object but instead the flow particles will stick to the boundary. This is when
the profile of velocity reduces as it gets closer to the boundary, and is zero at the
boundary. This simply means u = 0 at the boundary of the object.

9
Mathematics and Methodology

In our method, varying the skin friction parameter (β) means that we are using
a skin friction wall model boundary condition around the sphere. To implement
no slip boundary condition as the boundary layer approximation condition of the
model, we can simply give β = 1 in the solver.

3.3 Other works


There has been a lot of work done previously for a non-spinning sphere in which
the drag has been computed. In [6], G2 method has been used for non-rotating
ball. In [12], another method for turbulent modeling called DES has been used to
compute drag crisis. Detached Eddy Simulation (DES) is a method that acts as
RANS near the boundary and changes to LES a way from the surface.
Incompressible, viscous flow past a rotating sphere (with rotation around the
stream line axis), for low Re (50 ≤ Re ≤ 100), has been done in [2] in which a three-
dimensional spectral element/spectral method (combination of a FEM method with
high degree piecewise polynomial basis and spectral method known also as Fourier
Transform Methods) has been used. The same method was used in [1] to compute
the lift and drag of the rotating sphere in laminar flow.
In [2], the point potential vortex method has been suggested for a rotating ball
moving in an incompressible, inviscid flow as well as the numerical computation of
the vortex sheet in the a plane.
In [11], flow around a rotating baseball was computed using a combination of
types of finite difference methods in space and time. From their conclusions: (i)
it appears that the variations in the wake become larger when the speed in which
the ball spins is lower. (ii) The drag is predicted to increase as the rotating speed
decreases which explains the Magnus effect and why the topspin will land sooner
on the ground.
In most methods the boundary layer theory has been used in which a viscous flow
in a boundary layer must be solved which needs modeling and is very complicated.

3.4 Some methods for turbulent flow


It is important to predict a reasonably accurate solution to the N.S. for turbulent
flow since most fluid flows in real life are turbulence, e.g. blood flow in arteries, oil
transport in pipelines, atmosphere and ocean currents, and the flow in boat wakes,
around aircraft-wing tips and along the sport ball moving in high velocities, etc.
The variety of length scales needed to be resolved in order to give an accurate
estimation of the turbulent flow will cause to fail computing turbulent flows point-
wisely. Another way to explain it is to say that a very fine mesh is needed in order
to capture all these length scales.
One of the basic problems in CFD area is to compute the turbulent fluid flow
around a moving object. According to [6], the number of mesh points needed for
solving turbulent flow has been approximated to be Re3 in space-time, where Re =

10
Mathematics and Methodology

U∗ L∗ /ν and U∗ , L∗ are scaled velocity and scaled length of the object respectively
and ν is the viscosity of the fluid. Computing the turbulent flow with large number of
degrees of freedom close to the boundary layer is a challenge. Several methods have
been suggested for turbulent modeling. One of them is DNS which is a numerical
method used to solve N.S. equations without modeling. It will resolve for all physical
scales of flow and the cost will grow fast with increasing Re and it can not be used
for high Re flows. For most cases of interest (i.e. Re > 106 ), the number of degrees
of freedom in space-time that are needed to resolve all scales of the flow will be
1018 . This is not computable even with today’s massive parallel computers.
Using the laminar solver to model the turbulent flow has failed as it becomes
unstable (not converging) in time. Therefore the idea of using a time-averaged N.S.
instead of the real N.S. equations gave raise to some methods like RANS, LES,...
Thereby different methods has been proposed to get around the problem in
which most of them use turbulent modeling which will introduce a model usually
using an averaged N.S. equation. Two most common method used to get around
this problem are RANS, Reynolds Averaged Navier Stokes, and LES, Large Eddy
Simulations. In the first method, the effort is to solve a modified (averaged) Navier
Stokes equations, with some approximate solution which will usually end up adding
a term called Reynolds stress, to the modified equation. This comes down to RSM
(Reynolds Stress Modeling) which requires a complex modeling. The idea in this
method is based on Reynolds averaging in 1894.
Formulated in late 1960s, LES uses the idea to only solve for large eddies of
turbulent flow and using a subgrid model to model the smaller eddies. The idea
behind this method comes from the self similarity theory formulated on 1941 by
Andrey Kolmogorov.
In self similarity theory one assumes that large eddies of the flow are dependent
on the flow geometry, while smaller eddies are self similar and have a universal
character. Many methods have been produced by combining the above methods
together at which all will search for approximate solution to an averaged (filtered)
N.S. equation and will use turbulence modeling which are complicated. Instead in
G2 we will compute weak solutions (more specifically, CU ||hR(Û )||-weak solution,
explained in [7]) to N.S. equations. In this method we don’t resolve all physical
scales as oppose to DNS. Also, we won’t resolve turbulent boundary layer so there
is no need for turbulence modeling based on physics of unresolved scales as oppose
to other turbulent modelings . Looking for -weak solution instead of an averaged
(filtered) N.S formula we have no turbulent model (such as subgrid modeling or
Reynolds stress modeling) but more to say using G2 we will get an automatic
turbulent model. That is to say we will model the turbulent boundary layer using
a slip with friction boundary condition which is discussed in boundary condition
section. To solve the laminar boundary layer (which happens for Re < 105 ) the
method uses N.S. equation with either noslip or high friction boundary conditions.
The -weak solution is an expression used in [7] for the solution of the method
in use. That is to say the residual control will be ||R(û)||−1 <  instead of the more
common residual control ||R(û)|| < . The weak norm ||.||−1 is defined by:

11
Mathematics and Methodology

R
Ω v.wdx
||w||−1 = sup ,
v∈H01 ||v||H 1 (Ω)

||w||H 1 (Ω) = (||∇w||2 + ||w||2 )1/2


According to [7] the residual will be very small for turbulent solutions in the weak
sense while it is very large pointwisely. This is the reason that the method can
be used to compute some mean values of interest. We display the mean value of
interest with a function called M (u).
To discretize the system of equations, we are using the weighted least square
stabilized cG(1)cG(1) method. The stabilized cG(1)cG(1) method is a type of G2
method that uses continuous stabilized Galerkin finite element method in time with
piecewise linear trial functions both in time and space and linear piecewise test
functions in space and constant piecewise test functions in time.
The stabilization term in the discretized system will introduce dissipation where
the residual is large. The adaptivity is based on a posteriori error estimation of the
dual problem with respect to a function of drag. According to [7], using some
function of the solution (i.e. drag, lift, etc.) will give rather more satisfying bounds
of the error whereas the energy norm in most cases doesn’t give good bounds,
compared to what is acceptable in reality of physics.

3.4.1 Discretization cG(1)cG(1)


The method has been explained in [8]. Here is just the summary review of the
section.
We are looking for approximate solutions for velocity (Uhn ≡ Uh (tn )) and pres-
sure
(Phn ≡ Ph (tn )) which will satisfy N.S. equations with homogeneous Dirichlet bound-
ary conditions. Let 0 = t0 < t1 < ... < tN = T be a sequence of discrete time steps
where the time interval at each step is In = (tn−1 , tn ] with length kn = tn − tn−1 .
Assuming the domain to be Sn = Ω × In and let W n ⊂ H 1 (Ω) be a finite ele-
ment space of continuous piecewise linear functions on a mesh of size hn with the
functions in W n satisfying the Dirichlet boundary. H here is a Sobolev Hilbert
space. Uh (tn ) ∈ V0 n ≡ [W0 n ]3 and Ph (tn ) ∈ W n . functions in the set V0 n satis-
fying the Dirichlet boundary condition v|ΓD = 0 and functions belong to set W n
satisfy the Dirichlet boundary condition v|ΓD = w. Choosing the test functions
v ∈ V0 n ⊂ H 1 (Ω), the numerical approximation of the discretized system is defined
by:

((Uhn − Uhn−1 )kn−1 , v) + (Ūhn · ∇Ūhn , v) + (2ν(Ūhn ), (v)) − (Phn , ∇ · v) + (∇ · Ūhn , q)


+ SDδ (Ū n , P n ; v, q) = (f, v + δ1 (Ūhn · ∇v + ∇q)), ∀(v; q) ∈ V0n × W n ,

where

12
Mathematics and Methodology

SDδ (Ū n , P n ; v, q) = +(δ1 (Ûhn · ∇Ūhn + ∇Phn ), Ūhn · ∇v + ∇q) + δ2 ∇ · Ūhn , ∇ · v)

−2 − 21
with Ūhn = 12 (Uhn + Uhn−1 ), δ1 = 1
2 (kn + |U |2 h−2
n ) and δ2 = k2 h. k1 , k2 are
positive constants of unit size.

X Z
(v, w) = v · wdx,
K∈τn K
3
X
((v), (w)) = (ij (v), ij (w)).
i,j=1

3.5 Stability of G2 method


the stability estimation has been proven in [7]. The following relations holds for
f = 0;
1 1
k U (t) k2 + k h1/2 R(Û ) k20 = k U 0 k2
2 2
Where R(û) = u̇ + U · ∇u + ∇q and û = (u, q). This shows that the numerical
dissipation in G2 is proportional to the residual R(Û ). Also, it determines that
residual will not grow and therefore we have stability in the form,

2 1 2
k h1/2 R(Û ) k0 ≤ k U 0 k
2

3.6 Computation of forces


According to [3], the mean value forces effecting an object situated in air over time
interval (I) could be computed by following formula,

1
Z
N (σ(û)) = (u̇ + u · ∇u − f, Φ) − (p, ∇ · Φ) + (2ν(u), (Φ)) + (∇ · u, Θ)dt
|I| I

where û = (u, p) is the -weak solution to the N.S. equations.


Φ is a function defined in the fluid volume Ω to be a unit vector in the direction
of the force that we want to compute in Γ0 and to be zero on the rest of the boundary
Γ1 = ∂Ω/Γ0 . Γ0 is the part of the boundary that is in contact with the flow and
the remain will be Γ1 . In this model (a sphere moving in air) all the boundary will
be in contact with the flow. Therefore, it is important to know the direction of
different forces acting on the sphere in order to compute them.
This formula depends on Θ, and the distance between Φ and the boundary.
Using a finite element method cG(1)cG(1), we are able to approximate the force in
the discretized system as follow,

13
Mathematics and Methodology

N h (σ(Ûh )) = 1 R ˙ + Uh · ∇Uh − f, Φ) − (Ph , ∇ · Φ) + (2ν(Uh ), (Φ))+


|I| I (Uh

(∇ · Uh , Θ) + SDδ (Uh , Ph ; Φ, Θ)dt.


Now Φ and Θ are finite element functions. SDδ is the stabilization term. U˙h =
Uh n −Uh n−1
kn on In (In , is the nth subinterval of the total time interval I).
There are three aerodynamic forces on the ball moving in the air. In following the
computations of the forces on the ball (i.e, drag, lift, gravity,etc.) will be presented.

3.6.1 Computation of Drag


The drag force is the force that resists the movement of a solid object through air.
It is parallel and opposite to the direction in which the object is moving (more
information on drag can be fOUnd in next chapter).
To compute drag force it is only needed to define Φ = (0, 0, 1), assuming the ball
is moving in x3 direction. According to the computations in [7], with consideration
of a longer time interval we are able to compute the mean drag force more accurate.
The drag coefficient (cD ) is a dimensionless quantity which is used to quantify the
drag or resistance of an object in a fluid such as air. We are interested to compute
this coefficient. Under the assumption that Re is the same and the flow is subsonic
and steady, cD is the same even if other variables change. As a test case we have
the drag coefficient computed for sphere to be approximately 0.4 for a rough sphere
in Re = 106 . Even though Re and therefore cD can vary a lot, it is almost constant
or very small in some range of interest.
Figure 3.1., by Auchenbach (1974) [10], demonstrates the variation of cD as a
function of the Reynolds number (for different smoothness of sport’s ball).

Figure 3.1. Aerodynamic drag curves for spheres of different roughness, k/d = 0,
25, 150, 250, 500, 1250 E-5 increasing to the left

We will then compute the mean value of the drag coefficients during the time

14
Mathematics and Methodology

1 R
interval I = [0, T ] (e.g, cD = |I| I cD ) for incompressible flow by the following for-
mula [8, 7, 3],

N (σ(û))
cD ≡ 1 2 (3.6)
2 U∞ A

where the U∞ = 1 is the normalized velocity, and A is the cross-sectional area of


the object which in our study case for a sphere is Π( D 2
2 ) , with D the diameter of
the sphere. N (σ(û)) is the mean value drag force over time interval I.

3.6.2 Computation of Lift


The lift force is another force which might be acting on the moving object on the
air. that is perpendicular to the flow direction. It contrasts with the drag force,
which is the parallel to the flow direction. We will compute the mean value of the
1 R
lift coefficients during time interval I = [0, T ] (e.g, cL = |I| I cL ) for incompressible
flow by the following formula [8, 7, 3],

N (σ(û))
cL ≡ 1 2
2 U∞ A

In this case we will define N (σ(û)) to be lift force by defining Φ = (0, 1, 0) if the lift
force acts in x2 direction.

3.6.3 Computation of Gravity


Another force that we can consider is earth’s gravity which will be perpendicular to
the ground and is always downward assuming the origin of the coordinate system
to be the origin of the earth (g ' 9.8 m/s2 ).
However, in our solver the coordinate system has been transformed with the
axes defined by the direction of the movement of the sphere; Therefore, the gravity
force direction will depend on the place of the sphere in the global coordinate system
at each time. To model the gravity force, we will need to know the direction of the
gravity force at each time step which is not always downward anymore but changes
from {αi }i=0
n−1 . αn = α(tn ) is the angle between gravity and drag force (or opposite
direction velocity flow) at time step n. Note that 0 ≤ αn ≤ π.
Therefore having αn , we can easily compute the gravity force at each time step
to be,
 
0
h
G (α(tn )) = g  cos α(tn )  .
 
sin α(tn )

15
Mathematics and Methodology

It is enough though to define Φ = (0, cos(αn ), sin(αn )) in the solver, if the displace-
ment of the ball is in x2 , x3 directions and there is no displacement in the direction
of x1 (the sphere is only rotating around the x1 axis). Note that we will add the
gravity to plot the trajectory of the sphere, and since the first component is zero
we can say that the computations are in 2D. However all the other computations
which are done in the solver are in 3D.
Now the question is how to compute the αn at each time steps. Under the as-
sumption that the sphere is moving along the negative direction of x3 axis and the
positive direction of x2 axis they can be computed by,

Xn+1 (3) − Xn (3)


αn = arctan( ) tan αn ≤ 0
Xn+1 (2) − Xn (2)
π Xn (2) − Xn+1 (2)
αn = + arctan( ) tan αn > 0
2 Xn (3) − Xn+1 (3)
where Xn is the global position of the ball at time n. This will be computed by the
following explicit system of equations.
mẌ = F ,
which can be taken as two first order following equations:
mẊ = U ,
U̇ = F .
Here, m is the mass of the ball, F is the resultant of the forces on the ball. This can
be added to the solver to compute gravity force and adding it to the aerodynamic
forces in order to plot the trajectory online. However, we did not use this for our
computations. Both the gravity force and plotting the trajectory are done as an
offline operation.

3.7 Rotating ball


For an object to be able to rotate, a force is needed. The force needed for rotation
will be produced by angular velocity. We can compute the angular velocity in which
the ball is rotating clockwise around the x1 axis as follow:

uball = →

ω ×→

x, (3.7)
Where → −
ω is the angular speed (rotational speed), → −x is the distance vector from
the rotational axis (x1 ) to the surface of the ball which will differ with →

x . Assume
projecting the sphere on the p x2 -x3 plane, then we will have different isosurfaces
(circles) with different radius( (x2 )2 + (x3 )2 ). We will compute the above formula
for counter clockwise rotation (corresponding to topspin),
 
−ω


ω = 0 
 
0

16
Mathematics and Methodology

 
x1


x =  x2 
 
x3
 
0
uball =  ωx3 
 
−ωx2
The sign of → −
p computations to get |uball |max =
ω will change for backspin. In our


1/2|uair | we choose ω = uair /2R where R = (x1 )2 + (x2 )2 + (x3 )2 = rmax . rmax
is the radius of the biggest isosurface.

3.8 Trajectory
One interesting concept is to see how the rotation of the ball will effect the curvature
of its movement. Therefore, we compare the trajectory of the ball in three cases of:
(i) topspin (ii) nospin (iii) backspin. The position of the ball and where it is going
at each time step will be needed in order to plot the trajectory of the sphere. X,
the position of the ball is computable from a simple second order partial diferential
equation (this is known as Newton’s second laws of motion, or laws of acceleration).
F = mẌ where F is the resultant of the forces on the ball and m is the mass of the
ball (depends on the type of the ball we use). F contains gravity force as well as
the aerodynamic forces, i.e. drag and lift. Additionally, the initial position of the
sphere X0 and initial velocity of the sphere U0 will be given. The trajectory of the
ball can be obtained by the following discrete system,

Xn = Xn−1 + k Un−1
Un = Un−1 + (k/m) Fn−1 , n = 1, ..., N

To track the position of the ball which is needed to plot the trajectory, we could
either add it within the solver or we could do it offline. In our work, we chose the
offline routine because it was easy to implement and since we have had run our
computations and already had the forces on the ball therefore we didn’t want to
run the whole computations all over again, it seemed more efficient for us to chose
this option.
Computations of F can be done using a rotation matrix to interpolate the aero-
dynamic forces to a chosen global coordinate system where the gravity force will be
always in the opposite direction of the x2 . In this way, at each time step we have
the aerodynamic forces from the solver and we get their direction using the rotation
matrix. Another way is that we can use the rotation matrix to interpolate gravity
force at each time step and we can add that easily to the solver. This way the direc-
tion of the aerodynamic forces are fixed and the gravity force will be interpolated
accordingly to the coordinate system used by solver, at each time step. I didn’t

17
Mathematics and Methodology

use the rotation matrix. Assume that in the global coordinate system the gravity
force to be downwards at all time, the drag force to be parallel and in the opposite
direction to the velocity at each time step, and the lift force to be perpendicular
to velocity at each time step. Having the initial velocity v0 and using the velocity
which is estimated at each time step by the above formula which is dependent on the
forces at each time step, we can find out the position of the ball in our coordinate
system and plot the trajectory.

18
Chapter 4

Computations

The model we are studying is a sphere situated in a box. We used the Comsol
Multiphysics environment in order to define the geometry of the model as well as
generating the initial mesh.Comsol Multiphysics is an environment which may be
used to define the geometry, specify the physics of the model, mesh generation, solve
and post-processing the results. After generating the geometry and the initial mesh
using Comsol, we will solve the system using icns (Incompressible Navier-Stokes
solver) to solve the discrete system, and an adaptive mesh refinement algorithm to
refine the mesh which are both available in Unicorn package. The postprocessing
has been done in Paraview and MATLAB. ParaView is an open source package used
for data analysis and visualization1 . We used it to visualize the solution and used
MATLAB in order to plot the related graphs. MATLAB is a high-level language and
interactive environment that enables you to perform computationally intensive tasks
faster than with traditional programming languages such as C, C++, and Fortran2 .
It can be used for computational, visualizations, and programming purposes. It
is easy to use since problems and solutions are expressed in familiar mathematical
notation

4.1 The Computational model


4.1.1 Still Ball Model
The G2 method was used in [6] where the computations of mean drag value in time
as well as visualization of separation point(s) for a moving ball in air (changing the
velocity of the ball by taking the friction parameter β to be a function of velocity)
have been presented. The results showed that when decreasing the friction param-
eter (increasing the velocity) from β = 0.1 (correspond to the laminar boundary
layers) the drag almost remains constant until in β = 0.02 when laminar boundary
changes to turbulent a sudden drop in drag happens (drag crisis). In this work, we
1
http://www.paraview.org
2
http://www.mathworks.com/products/matlab

19
Computations

take the velocity variable fixed so that we can see how the smoothness of the ball
will effect on drag, and separation point(s) (skin friction is a function of velocity
and smoothness, etc.). An experiment has been done in [10] regarding the differ-
ent sport’s ball with different smoothness which gave different drag coefficient. In
their work they have used a radar speed gun downloading at 33 Hz to a notebook
computer in order to track the fall velocities. There are also experiments done in
wind tunnel for different ball skins. The result we show in this work are captured
using the G2 method. For modeling the flow around any sport ball through air, we
use G2 with friction boundary condition which is available in icns (Incompressible
Navier-Stokes Solver) implemented in Unicorn software. According to [7], Prandtl
and Wieselsberger showed in a classical experiment that the properties of sport
balls can effect the volume of wake behind the ball. As a result in same condi-
tions, different sport balls will produce different drags. In particular, considering
the smoothness which is coupled to skin friction of the ball to differ from smooth to
fuzzy skins and computing the drag force and separation of different ball surfaces.
As mentioned before, drag crisis will happen when the laminar boundary layer will
change to become turbulent. This will happen for 105 ≤ Re ≤ 106 . In high Re we
will have turbulent boundary layer.
Taking the skin friction (β) to be a function of velocity, one can model a moving
ball while the velocity differs. We can control the velocity by β. Since the velocity
U and the Reynolds number Re are related with Re = UνL (knowing the geometry
of the sphere will not change and assuming the kinematic viscosity to stay constant
as the ball is moving.), it is easy to model a turbulent or laminar flow by choosing
the β depending on velocity.
In the result section of this chapter screen shots of separation points and wake as
well as solutions of velocity and pressure will be presented. The separation points
will be presented using glyphs. There are plots showing how the drag force will
change relative to time and relative to mesh points (or level of refinement).

The model:
A sphere with D = 1 centered at x = (5, 5, 5) in the channel of dimension
10 × 15 × 15 oriented in x3 direction, subject to inflow velocity (0, 0, 1) has been
modeled. We are using a slip boundary condition on channel walls, Dirichlet inflow
boundary condition, and transparent outflow.

4.1.2 Rotating Ball Model

To model a spinning ball we consider two different models: (i) modeling topspin,
(ii) modeling backspin. It is easy to model both topspin and backspin, using icns.
A simple modification of the wall function which was used previously in [6] to model
the skin friction boundary, will suffice. The angular velocity of the ball uball , which
will cause the spinning, can be implemented using the following wall model. Both
the angular velocity and wind velocity are controlled by friction parameter in the

20
Computations

following formula,

(u − uball ) · n + αnT σn = 0, (4.1)


−1 T
(u − uball ) · τk + β n στk = 0, k = 1, 2. (4.2)

Using the same geometry and mesh as used in still ball we add the angular velocity
by above formula. We modify our code to model for the varying velocity and
angular velocity in a simulations. We use scaled velocity which can be controlled
by the friction parameter β. In here taking β to increase in time corresponds to
decreasing the velocity. For topspin, the ball is rotating around x1 axis clockwise
with a dimensionless angular speed (nondimensional rotation rate, or the ratio of
the surface speed and free-stream speed) α = |(uball )max |
|uair | = 1/2.
For the backspin, we use a counter clockwise direction for the rotation, every
thing else is the same.

4.1.3 Trajectory
The sphere’s trajectories are computed for different experimental problems. As-
suming the initial position X0 = (0, 0, 0), two different cases of sphere diameters (d)
have been experimented with different initial velocity in which the sphere has been
thrown. The experimental cases are as follows and the results for trajectories will
be presented at the results section;
1) d = 1 : u0 = (0, 1, −1), u0 = (0, 50, −1), u0 = (0, 300, −1), u0 = (0, 20, −1)
2) d = 0.1 :u0 = (0, 1, −1), u0 = (0, 30, −5), u0 = (0, 10, −5), u0 = (0, 50, −20),
u0 = (0, 4, −6)
The non-dimensional angular velocity is 0.5 for both topspin and back spin. In
order to draw the trajectory of the ball, one can use one of the following alternatives:
(i) Online: It is easy to improve the model of the ball by changing β dynamically
based on the actual velocity of the ball. (extending the model from the assumption
of linear increase in β). It can be done by defining and adding the gravity force in
the solver. This can be done the same way as the drag and lift forces definitions in
the solver. It will be done by giving the force direction. The only difference here is
since the mesh is not moving or say the local coordinate system is fixed on the mesh.
Drag is always in the direction of the flow and lift is perpendicular to the flow, in
the local system. Using this method, the direction of the forces will not change in
time stepping. However, the direction of the gravity force will change relative to
the flow direction during time steps. The angle between gravity and flow will be a
function of time which will vary almost by on angle of π. (|4θ(ti )| ∼ π). Varying
from acute to obtuse angle or vice versa depending on the direction in which the
ball is traveling. There is more information on how to compute the direction of
gravity force of the model in section 3.7.3.
(ii) Offline: The second alternative is to do the computations necessary to plot
the trajectory outside the solver. Using Newton’s second formula mẌ = F . Writing
a simple C++ program to solve the explicit system of equations, we can find the

21
Computations

position (X) of the sphere at each time and we can easily plot the trajectory using
MATLAB. The mentioned system of equation is a second order ODE (ordinary
differential equation) which can be solved by breaking it into two first order ODE.
At each time step we take the approximate forces on the sphere (drag and lift)
which are computed by the icns solver.
In order to plot the trajectory we used the second method.

4.2 The Computational results


Please note that except the trajectories, all the results are computed in three di-
mension.

4.2.1 Still Ball


With different skin friction and different mesh refinements we experiment that at
the beginning of the time interval, behavior of Fd is irregular but after a while the
changes of Fd during time becomes more regular and the range of oscillations will
be within an specific interval; Therefore, the mean value of Fd during time will be
constant, computable and meaningful. The following plot is an example of how the
drag force will look like at the beginning of time interval (left) and at the end of time
interval (right), for a test case with skin friction β = 0.5 in 5th mesh refinement
with 15872 number of mesh points.

Figure 4.1. β = 0.5 : Drag force (Fd ) vs. time

We know that the expected drag coefficient for a non rotating sphere moving in
the air is approximately Cd = 0.4. Considering this and the fact that Cd ' 2Fd will
give the conclusion that Fd ' 0.2. The following plot shows that as we refine the
mesh, we get closer to the expected drag force and therefore for accurate refined
meshes we expect to quantitatively predict the value of drag force, using the solver.
All the above stands for different skin frictions test cases. However, the following
shows the case where β = 0.5.

22
Computations

Figure 4.2. β = 0.5:Drag force vs. time for in different mesh refinements.

In the simulations for slip boundary condition (corresponding to β = 0), the


amount of Cd is close to the mentioned expected value. However for a sphere with
skin friction, more simulation (more refinement of the mesh) is needed to reach
to the more accurate value. The plot in the right (corresponding to β = 0.01)
shows that the function is decreasing in time. Quantitatively speaking, the value
of functions will converge to the predicted value with further mesh refinements.
The following plots show the drag coefficient for still ball in two test cases which
demonstrates Cd is reasonably close to previous studies.

Figure 4.3. cd vs. loglog number of mesh points for β = 0, β = 0.01 respectively
the left and the right.

We can get a more accurate and smoother function by adding less points at each
refinement of the mesh, if that is an interest of us. However, it was not a concern

23
Computations

for us in this work. The following plot shows how different refinements will effect
the plots. For this test we studied the sphere with skin friction of 0.01 for two cases.
One with 10% mesh refinement the other with 5% mesh refinement. The following
plot demonstrates the zoomed in statue of mean value cd vs. loglog number of mesh
points (right). The right plot shows from 6th to 12th refinement and in the left plot
we have considered the initial mesh up to the 12th refinement.

Figure 4.4. β = 0.01:cd vs. loglog number of mesh points

In the following plot the mean value of drag coefficient against different values of
skin frictions has exhibited. We considered this experiment for the mesh with 15872
nodes in the 5th iteration of mesh refinement. It shows that for 0 ≤ β ≤ 0.05 we
have a dramatic change of drag coefficient in compare with the larger skin friction
and as the boundary converges to be no slip.

24
Computations

Figure 4.5. cd vs. β at 5th refinement

screen shots of solutions with three different skin frictions has been shown in
the following plots. For β = 0 (zero friction at the boundary) the delayed in the
separation can be seen. Also, for β = 0.01, β = 0.5 the separation points and wake
are presented.

Figure 4.6. β = 0 : Wake (left), separation points (right) in mesh with 20116 nodes

25
Computations

Figure 4.7. β = 0.01 : Wake (left), separation points (right) in mesh with 66082
nodes

Figure 4.8. β = 0.5 : Wake (left), separation points (right) in mesh with 66082
nodes

Here is just three screen shoots of how the solution looks like in different time
steps in one simulation.

Figure 4.9. β = 0.5 : Velocity in three samples in time

In the following the solutions of last time sample has been shown. The first
figure belongs to β = 0.5 with 15872 mesh points and the second one belongs to
β = 0.01 with 66082 mesh points. In the solutions it is apparent that the velocity
and pressure of the flow are horizontally symmetric. There is no lift force. pressure
is low at the separation points while velocity is high. The flow in second figure is
more mature in a sense that the mesh is finer. In the third figure that belongs to

26
Computations

β = 0 and 20116 mesh points, the substantial drag is noticeable. The pressure and
velocity are not vertically symmetric and therefore the drag won’t be zero. Instead
we are dealing with substantial drag which will be more obvious in case of further
refinements.

Figure 4.10. β = 0.5 : Pressure (left) and velocity (right)

Figure 4.11. β = 0.01 : Pressure (left) and velocity (right)

Figure 4.12. β = 0 : Pressure (left) and velocity (right), (substantial drag)

27
Computations

The following pictures demonstrates the velocity using streamlines. In left, the
Euler solution (β = 0) has been presented while the N.S solution with small friction
(β = 0.005) and turbulent fluid is shown in right.

Figure 4.13. Euler solution (left) and N.S solution (right)

4.2.2 Rotating Ball


Topspin :

Plots of how the drag force acts in time for a Coarser mesh with 12544 nodes and a
finer mesh with 103918 nodes has been presented. The drag force varies in a small,
constant range in time which gives us the ability to compute a reasonable value for
the drag force and at some point in time the mean value of Fd will act regularly.
In figure(4.15), the zoomed in plots are demonstrated in which it is noticeable that
the length of range changes of Fd is very small. The plots of Fd against time has
been presented below,for the simulation of rotating sphere, for two different meshes
(after one and four refinements).

Figure 4.14. Drag force (Fd ) in time for the coarse mesh (left) and the finer mesh
(right).

28
Computations

Figure 4.15. Drag force at the end of time interval for the coarse mesh (left) and
the finer mesh (right)

29
Computations

When a sphere rotates in addition to the drag force that we disscussed before,
another force named lift force (Fl ) will produce. This requires us to investigate this
force and to visualize the data received using the computational method in hand.
The following plots shows the lift force in time for two different meshes with 12544
and 103918 nodes (respectively the coarse mesh in the first, the finer mesh in fourth
refinements). Investigating the results it has shown that the lift force will act in the
regular, certain way after a time interval has passed. Therefore the same thing that
we mentioned for the drag force stands for the lift force as well and we are able to
have a meaningful mean value of the lift force in time.

Figure 4.16. Lift force in time for the coarse mesh (left) and the finer mesh (right)

Figure 4.17. Lift force at the end of time interval for the coarse mesh (left) and the
finer mesh (right)

Plots of drag force and lift force has been presented in the following for different
mesh refinements. The drag force will become less oscillatory as the mesh is refined
whereas the lift force becomes more oscillatory with mesh refinement. However with
lift forces we observe that even though for higher mesh refinements we have more

30
Computations

oscillations about the force axis, we still remain in a finite range which gives the
ability to easily take the mean value of Fl as a meaningful, computable value.

Figure 4.18. Drag force (Fd ) vs. time in different mesh refinements (left) and lift
forces (Fl ) vs. time as we refine the mesh (right)

#ref # nodes c¯l c¯d f̄l f¯d


itr00 12544 -0.4109 0.8796 -0.1614 0.3482
itr01 14481 -0.4264 0.8699 -0.1675 0.3416
itr02 16834 -0.4517 0.8561 -0.1774 0.3362
itr03 19617 -0.4091 0.8346 -0.1607 0.3278
itr04 22755 -0.4023 0.7964 -0.1580 0.3127
itr05 26407 -0.3898 0.7742 -0.1531 0.3040
itr06 30607 -0.4008 0.7893 -0.1574 0.3100
higher 103918 -0.4517 0.8561 -0.1835 0.2954
Table 4.1. Computational results of mean value lift coefficient (cl ), drag coefficient
(cd ),lift force (fl ), and drag force (fd ) for different mesh refitments

In the following plots the behavior of the drag and lift force with respect to log10
number of mesh points has been presented. Although it is not possible to rule a
accurate prediction on how the function will act with further mesh refinement with
the few experiments, it is obvious that the drag force is decreasing intensely while
the lift force seems to act more monotonous with the mesh refinement.

31
Computations

Figure 4.19. Drag force vs. log10 of mesh points (left) and lift force vs. log10 of
mesh points (right)

Here some results from topspin generated by the solver is demonstrated using
Paraview as visualizational tool. The results belong to rotational case and with
103918 number of mesh points. The first two screen shots exhibitad the wake and
separation points. We can see that for the rotational sphere the separation points
are not summetric and as shown in the screen shots, we have a delay separation in
left in case of topspin. In the second pair of screen shots, we can see the results for
pressure and velocity. The flow is unsymmetrical and lift force is produced because
of the delay in separation. The low pressure in left will cause the sphere to move
left.

Figure 4.20. Wake (left) and separation (right) in x2 x3 -plan

32
Computations

Figure 4.21. Pressure solution (left), velocity solution (right)

The following figures show nice results of turbulence formed behind the moving
sphere. The shows the turbulent and wake using streamlines and glyph as tools
available for users in Paraview.

Figure 4.22. Separation and wake in left (with glyphs) and turbulence in right
(with streamlines), in x2 x3 -plan

33
Computations

Figure 4.23. Screen shots of turbulence generated behind the sphere

The following pictures are the velocity solutions for rotating sphere, solved for
a mesh with 98766 number of mesh points. Three samples of velocity solution has
been demonstrated in screen shots while the flow is developing in time.

Figure 4.24. Velocity solutions in different tme steps

For a mesh with 98766 number of mesh points, velocity and pressure solutions
as well as dual solutions underlying the adaptive mesh refinement algorithm using
a posteriori error estimation of the force on the sphere has been viewed below.

34
Computations

Figure 4.25. Solutions of pressure (upper left) and velocity (upper right) and Dual
solutions for pressure (lower left) and velocity (lower right).

In the following pictures, we are able to see the mesh refinement for two different
meshes. The blue edges corresponding to edges from the initial mesh (with 2404
mesh points)and the white edges are the edges added in mesh refinement procedure
(mesh with 3460 nodes).The left figure is in x1 x3 -plan and the right figure is in
x1 x2 -plan.

Figure 4.26. Mesh refinement

35
Computations

Backspin :

The following solutions are extracted from solving the problem assuming the sphere
to rotate in the opposite direction of the topspin case. The results are produced by
implementing the model and boundary conditions to the mesh with 103918 nodes.
In the first pair of screen shots, the wake and separation ponts are presented in
x2 x3 -plan. It is noticable that there is a delay in the separation in right side. The
second pair of screen shots, show the pressure and velocity. The unsymmetrical flow
will produce a lift force. The low pressure in right will cause the sphere to move
right.

Figure 4.27. Wake (left) and separation (right)

Figure 4.28. Pressure solution (left), velocity solution (right)

# nodes c¯l c¯d f̄l f¯d


103918 -0.4623 0.7508 -0.0713 0.2949
Table 4.2. Computational results of mean value lift coefficient(cl ), drag
coefficient(cd ),lift force(fl ), and drag force(fd ) for different mesh refitments

36
Computations

4.2.3 Trajectory

(i)Experiments for a sphere with diameter = 1

Trajectory of sphere with the initial velocity U0 = (0, 1, −1).


In the following figure the trajectory for the topspin has landed sooner with
compare to the other two. The reason is the backward lift force which was produced
due to the spin of the sphere has added a component vector in the direction of the
gravity force; Therefore, we have an extra component force which will force the
sphere to land sooner. With the same reason we can explain why the trajectory of
backspin landed later. The vertical component of the lift force is opposite to the
direction of gravity force and therefore resists to reach to the ground sooner. For
the still ball the forces are the ball are gravity and velocity. The non-rotating ball
will move in the direction of the horizontal velocity component and az the velocity
decreases, the sphere can move forward until the gravity force is larger than the
vertical component of velocity. This is when the sphere will go downwards.

Figure 4.29. U0 = (0, 1, −1), D = 1

Trajectory of sphere with the initial velocity U0 = (0, 50, −1).


In the following figure both the initial angle and speed in which the sphere has
thrown increased compare to the previous figure. The trajectory for the topspin has
landed later. This is because the initial velocity is high the aerodynamic forces are
more important with respect to gravity and the sphere will travel further in the air.

37
Computations

Figure 4.30. U0 = (0, 50, −1), D = 1

Trajectory of sphere with the initial velocity U0 = (0, 300, −10).


In the following figure the initial angle has not changed much with compare to
the previous figure, just a little decrease. However the initial speed has increased
a lot. The same behavior will happen with compare to the previous figure. The
difference here is that due to the higher initial velocity, gravity is even less important
than the previous figure.

Figure 4.31. U0 = (0, 300, −10), D = 1

38
Computations

Trajectory of sphere with the initial velocity U0 = (0, 20, −1).


The same happens in the following figure, as in previous figures.

Figure 4.32. U0 = (0, 20, −1), D = 1

(ii)Experiments for a sphere with diameter = 0.1

Trajectory of sphere with the initial velocity U0 = (0, 1, −1).


In this figure the sphere initial velocity is the same as the experiment in figure
(4.31). The trajectory shows that the landing point is farther from the initial
position when using a sphere with smaller diameter in compare to the trajectory of
the sphere with a larger diameter. Also, the plot has shown that the lift force with
the chosen initial conditions will not have a large effect on the sphere. As seen in
the plot the three trajectories travel very close together until they hit the ground
in close positions together; This is shown in right plot.

39
Computations

Figure 4.33. U0 = (0, 1, −1), D = 0.1

Trajectory of sphere with the initial velocity U0 = (0, 30, −5).


The folowing figure is the same as the previuos one. The difference is that the
area of the sphere is smaller which will cause the aerodynamic forces to be less
important with respect to gravity and therefore it lands earlier.

Figure 4.34. U0 = (0, 30, −5), D = 0.1

Trajectory of sphere with the initial velocity U0 = (0, 10, −5).


In the following figure the interesting part is what happens at the top of the
hill. At the top of the hill the vertical component of the lift force is upwards and is
the most effective component here (since the horizontal component is almost zero).
Comparing to the non-rotating sphere trajectory, we can conclude that with the
initial conditions of the problem, the vertical component is larger than the gravity

40
Computations

force; Therefore, it will cause a lift at the top of the trajectory. This will make the
sphere to travel higher in the air. The same happens in the case of the top spin.
The two forces, gravity and the vertical component of the lift force, will force the
sphere to travel downward.

Figure 4.35. U0 = (0, 10, −5), D = 0.1

Trajectory of sphere with the initial velocity U0 = (0, 50, −20).


The trajectory of the backspin will go rightwards as the cause of resultants of
forces of gravity, velocity, and the lift force. The gravity force will act weaker as
there is a lift force upwards which will cause the sphere to travel further upwards.
For the trajectory of the topspin, we think that probably there has been an error
in "offline" approach since the topspin has moved backwards.

41
Computations

Figure 4.36. U0 = (0, 50, −20), D = 0.1

Trajectory of sphere with the initial velocity U0 = (0, 4, −6).


This trajectory is an example in which the trajectories in all three different
cases will travel very close to each other. However, in the right plot we can see
that they landed on the ground at different places and the lift force has effected the
trajectories. With initial conditions in this case the effect of the force seems to be
lesser than other problems (initial conditions).

Figure 4.37. U0 = (0, 4, −6), D = 0.1

42
Chapter 5

Summary

5.1 Conclusion
Using the G2 method, we were able to simulate the drag force of the non-rotating
sphere traveling in air. Also, we computed the drag and lift forces on the rotating
sphere. We have computed the separation, wake, flow velocity, pressure and tur-
bulent flow that were formed behind the moving sphere. Using the offline solver
then we were able to plot the trajectories of some test cases with different initial
conditions. The computations of trajectories appear qualitatively correct.
We noticed that the finer the mesh was, the computed drag and lift had more
oscillations in their graphs. In finer meshes, the drag coefficient decreases while the
lift coefficient increases
qas the mesh refines.
The larger |u0 | = u0x + u0y + u0z (u0 being the initial velocity) was, the larger
the effects of aerodynamic forces (i.e. drag and lift) are compare to weight (W =
mg).
Depending on the angle in which the sphere has thrown, it will differ whether
the trajectory of topspin will land sooner than the trajectory of backspin or vice
versa. This is depending on the effect and direction of the lift force on the global
coordinate system.

5.2 Future work


In future work, we would like to apply gravity to the NSESolver in unicorn. This
enables us to compute the trajectory in the solver more accurately. Also, we like to
model β as a function of velocity and even roughness of the sphere instead of a linear
time dependent function. The beta chosen here was to simplify the problem and
we can compute the flow around a sphere in a more realistic situation by choosing
β as an accurate function of velocity. Another change that we can make is to use
the solver to compute the flow for a specific sport’s ball with specific properties.

43
Bibliography

[1] K. Horigan C. J. Pregnalato, M . C. Thompson. Flow transition in the wake


of a rotating sphere. Technical report, Advanced Computing Center, RIKEN
and NEC Informatec Systems, Ltd and Communication Media Research, NEC
Corporation, 2001.

[2] K. Horigan C. J. Pregnalato, M. C. Thompson. Flow patterns in the wake


of a streamwise-rotating sphere. Technical report, Advanced Computing Cen-
ter, RIKEN and NEC Informatec Systems, Ltd and Communication Media
Research, NEC Corporation, 2002.

[3] J. Hoffman. Computation of mean drag for bluff body problems using adaptive
dns/les. SIAM J. Sci. Comput., 27(1):184–1207, 2005.

[4] J. Hoffman. Efficient computation of mean drag for subcritical flow past a
circular cylinder using general galerkin g2, accepted for int. J. Numer. Meth.
Fluids,, 2005.

[5] J. Hoffman. Computation of turbulent flow past bluff bodies using adaptive
general galerkin methods: Drag crisis and turbulent euler solutions. Computa-
tional Mechanics, 38, 2006.

[6] J. Hoffman. Simulating drag crisis for sphere using skin friction boundary con-
ditions. European Conference on Computational Fluid Dynamics, ECCOMAS
CFD, May 2006.

[7] J. Hoffman and C. Johnson. Computational Turbulent Incompressible Flow,


volume 4. Springer, 2006.

[8] J. Hoffman and C. Johnson. A new approach to computational turbulence


modeling. Comput. Methods Appl. Mech. Engrg., 195:2865–2880, 2006.

[9] J. Hoffman and C. Johnson. Finally : Resolution of d’alembert’s paradox. J.


Math. Fluid Mech., May 2008.

[10] Acousto-Scan John Dunlop. Free flight aerodynamics of sports balls. 2003.

45
Summary

[11] Shun Doi Ryutaro Himeno, Hideki Matsumoto. A numerical analysis of flows
around a rotating baseball. Technical Report 30, Department of Mechanical
Engineering, Monash University, Australia, 2000.

[12] P. Sagaut. Large Eddy Simulation for Incompressible Flows. Springer-Verlag,


Berlin, Heidelberg, New York, 2001.

46
TRITA-CSC-E 2009: 097
ISRN-KTH/CSC/E--09/097--SE
ISSN-1653-5715

www.kth.se

You might also like