Naseem Uddin Thesis Part01

Turbulence Modeling of Complex Flows in
CFD
A thesis accepted by the Faculty of Aerospace Engineering and Geodesy

of the Universitt Stuttgart in partial fulfilment of the requirements
for the degree of Doctor of Engineering Sciences (Dr. -Ing.)
by
M. Engg. Naseem Uddin

born in Karachi, Pakistan
Committee chair: Prof. Dr.-Ing. habil. Bernhard Weigand

Committee member: Prof. Dr. rer. nat. habil. Claus-Dieter Munz
Date of defence: 28 August 2008
Institute of Aerospace Thermodynamics

Universitt Stuttgart
2008
PREFACE
This thesis was written during my work as research scientist at the Institute of
Aerospace Thermodynamics (ITLR), Universitt Stuttgart, Stuttgart, Germany. I
want to express my deep regard for my Doktorvater Prof. Dr.-Ing. habil. Bernhard
Weigand, Director of the Institute of Aerospace Thermodynamics (ITLR) for his
sincere help until the completion of this work. I have really enjoyed the freedom
for scientific inquiry under his guidance and support.
I would like to express my deep gratitude to Dr.-Ing. Sven Olaf Neumann,
ITLR, who guided me through out this work with patience.
I would like to thank Prof. Dr. rer. nat. habil. Claus-Dieter Munz, Institut
fr Aerodynamik und Gasdynamik (IAG), Universitt Stuttgart, for reviewing this
doctorate work and becoming also the co-examiner.
I would like to thank Prof. Dr. M. Schfer, Dr.-Ing. Dorete Sternel and Dipl.-
Ing. M. Kornhaas, Fachgebiet fr Numerische Berechnungsverfahren im Maschi-
nenbau (FNB), Technische Universitt, Darmstadt, Germany, for the FASTEST
code and support.
A sincere gratitude is extended to Dr. B. A. Younis (Department of Civil &
Environmental Engineering, University of California, Davis, USA), Dr.-Ing. v.
Wolfersdorf (ITLR), Dr.-Ing. M. J. Kloker (IAG), apl. Prof. Dr.-Ing. Ulrich Rist
(IAG), Dr. Michael Crawford (University of Texas, Austin, USA) for discussion
on various issues addressed in the thesis. I want to thank Dr. M. Klein, IKE,
Technische Universitt Darmstadt, for helping in understanding the procedure of
digital filter based turbulence generation technique. I also like to thank Dr. I.
Schevchuk (ITLR) for sharing the knowledge of swirling jet physics.
I want to express my gratitude to people at Hchstleistungsrechenzentrum
(HLRS), Stuttgart, who always solved the problems I faced during compilation
especially I would like to thank Dr. Peter Lammers for helping with code compi-
ii PREFACE
lation on NEC SX-8 cluster.

I would like to express especial thanks to Higher Education Commission (HEC),
Pakistan and Deutscher Akadamischer Austauch Dienst (DAAD), Germany for the
PhD Fellowship.
I want to extend my regards to my teachers Dr. Muzzaffar Mahmood and Dr.-
Ing. R. Q. Hamid, NED University of Engineering & Technology, Karachi, who
have instilled in me the interest in the fields of heat transfer and turbulence physics.
I also like to thank the Vice-Chancellor, Engr. Abul Kalam who has allowed the
study leave from the NED University of Engineering & Technology.
I want to thank all my colleagues at ITLR especially my office colleagues
Simon and Barbara. I would like to thank Simon for the translation of the abstract.
I want to thank all my friends who have helped and encouraged me through
out this work. I want to thank Dr. Arun Anand for discussions on the spectral
analysis. I want to thank my cousin Adnan for helpful comments on the draft of
this dissertation.
I do not have words to express my regard for my mother and late father, who
had taught me how to read and write. I am dedicating this work to them. I want
to thank my brother Waseem and sisters Gulnaz and Shahla who have always sup-
ported me in my endeavors.
Lastly, I am grateful to my wife Uzma and son Sheharyar. There smile is the
constant source of motivation for this work.
Naseem Uddin
ITLR
Stuttgart, August 2008
ABSTRACT
In the last two decades of the 20th century, the world had witnessed the enormous increase
in the computational resources. This has brought in the tremendous increase in human
knowledge and understanding of complex phenomenon, like turbulence and its modeling
among others. The technical development has brought new challenges in engineering de-
sign. Whether we limit ourselves to the earthly matters or indulge in space exploration, the
simulation of turbulence has become a routine task.
Turbulence is a phenomenon in nature comprising of complex eddy structures which
can greatly improve heat and mass transfer. The simplistic approach for the computation
of turbulent flows is to compute them by Reynolds Averaged Navier-Stokes (RANS) equa-
tions based models. But due to the averaging procedure, the inherent unsteadiness of the
flow is compromised. On the other hand, Direct Numerical Simulation (DNS) is the best
approach for turbulent flow computations, in which no modeling assumptions are invoked
and turbulent eddies as small as of the order of Kolmogorov scale are computed. The mid-
dle approach between these extremes is the Large Eddy Simulation (LES), in which part
of the turbulence is modeled and rest is computed. However, the computational costs and
memory requirement are still too large to take it as a general purpose engineering design
tool.
In this thesis, the effect of changes in inflow conditions on the heat transfer by an
impinging jet is investigated using LES. The Dynamic Smagorinsky model proposed by
Germano et al. has been used as a subgrid model. Results of several Large Eddy Simulations
are reported in the thesis, which are conducted with use of total 244 processors on high
performance computing clusters.
The inflow conditions explored are:
The fully developed turbulent jet

Swirling jet
Velocity field active excitation
Velocity field passive excitation
iv ABSTRACT
Temperature field excitation

An ERCOFTAC recommended benchmark test case of an orthogonally impinging round
jet at Reynolds number of 23000 is simulated first. The agreement between the experiments
and the LES gives encouragement for further investigations. Therefore in a next step, the ef-
fect of inlet velocity and temperature fields excitations of an orthogonally impinging round
jet, on the heat transfer are explored. In case of the active control of the impinging jets
inlet velocity field, it is found that the selection of excitational frequencies is important for
heat transfer enhancement. The excitation at the subharmonic frequencies of the preferred
mode is found to be a promising approach for heat transfer enhancement. A passively ex-
cited inlet velocity field is also investigated. It is found that this approach gives a better heat
transfer than an active excitation case. The novel idea of heat transfer enhancement through
jets inlet temperature field excitation is presented. The LES shows that the excitation at
the preferred mode can give surface averaged Nusselt number higher than the non-excited
jet impingement case. However the frequencies higher than the preferred mode should be
taken with care, as they might cause thermal fatigue. The effect of the addition of the swirl
to the impinging jet on the heat transfer is investigated. Also, it is found that the swirl does
not give appreciable enhancement in heat transfer for H/D=2 case.
The knowledge of flow and passive scalar flux dynamics gained through the simulation
has helped in understanding the functional relationship between different turbulence quanti-
ties and heat transfer. It is found that the assumption of a constant turbulent Prandtl number
(often used in RANS based models) is not a realistic approach. Alternative is to use scalar-
flux models, which allow the prediction of scalar-fluxes in non-isotropic turbulent state. The
knowledge gained through the LES is then used to investigate coefficients in some explicit
scalar-flux models (RANS based models). The investigation gives insight in the impinge-
ment phenomenon, which could help in the development of advanced turbulence models for
heat transfer prediction.
ZUSAMMENFASSUNG
Die 80er- und 90er-Jahren des 20. Jahrhunderts waren gekennzeichnet durch eine enorme
Steigerung des Leistungsvermgens von Computern. Diese Entwicklung fhrte dazu, dass
komplexe Phnomene, wie z.B. Turbulenz und ihre Modellierung nun sehr viel besser ver-
standen werden. Diese technische Entwicklung hlt auch neue Herausforderungen fr In-
genieure bereit. Ob man sich auf Vorgnge auf der Erde beschrnkt, oder den Weltraum
erkundet: die Simulation von Turbulenz ist zu einer Routineaufgabe geworden.
Turbulenz ist ein Naturphnomen. Sie besteht aus komplexen Wirbelstrukturen, die
den Impuls-, Wrme- und Stofftransport stark verbessern knnen. Der einfachste Weg,
turbulente Strmungen zu berechnen, ist, sie mit Hilfe der Reynolds-gemittelten Navier-
Stokes Gleichungen (RANS) zu modellieren. Allerdings wird hierbei das systemimmanente
instationre Verhalten der Strmung geglttet. Dahingegen liefert die Direkte Numerische
Simulation (DNS) die genauesten Resultate bei der Simulation turbulenter Strmungen.
Bei diesem Verfahren werden keinerlei vereinfachende Modelle verwendet und kleinste
turbulente Wirbel bis zur Kolmogorov-Grenordnung werden berechnet. Einen Mittelweg
zwischen diesen beiden Anstzen stellt die Large Eddy Simulation (LES) dar, bei der nur
ein Teil der turbulenten Strukturen aufgelst und der Rest modelliert wird. Trotz dieser
Vereinfachung sind die Kosten und der Speicheraufwand dieser Methode immer noch zu
hoch, um sie fr Ingenieursanwendungen zu einem Standardwerkzeug werden zu lassen.
In dieser Arbeit wird die LES-Methode benutzt, um die Auswirkung von Vernderun-
gen der Eintrittsbedingungen auf den Wrmebergang eines Prallstrahls zu untersuchen.
Das dynamische Smagorinsky-Modell von Germano et al. wurde als Subgrid-Modell ver-
wendet. Die Ergebnisse von mehreren LES-Rechnungen, die auf 244 Prozessorkernen auf
Hochleistungsrechnern durchgefhrt wurden, werden dargestellt.
Die verschiedenen untersuchten Eintrittsbedingungen sind:
voll ausgebildeter turbulenter Strahl
drallbehafteter Strahl
aktiv angeregtes Geschwindigkeitsfeld
passiv angeregtes Geschwindigkeitsfeld
vi ZUSAMMENFASSUNG
angeregtes Temperaturfeld
Zunchst wird ein ERCOFTAC Benchmark-Testfall eines senkrechten kreis frmigen Prall-
strahls mit der Reynoldszahl 23.000 simuliert. Die gute bereinstimmung zwischen den
Experimenten und der LES-Rechnung ist die Grundlage fr die folgenden Untersuchun-
gen. Im nchsten Schritt wird der Einfluss der Anregung von Geschwindigkeits- und Tem-
peraturfeld auf den Wrmebergang bestimmt. Fr den Fall der aktiven Anregung des
Geschwindigkeitsfeldes im Strahleintritt stellt sich heraus, dass die Wahl der Anregungs-
frequenz wichtig ist fr die Intensivierung des Wrmebergangs. Eine Anregung mit sub-
harmonischen Frequenzen der bevorzugten Mode erweist sich als vielversprechend. Ein
passiv angeregtes Geschwindigkeitsfeld wird ebenfalls untersucht. Der Wrmebergang ist
gegenber der aktiven Anregung verbessert. Eine neuartige Idee der Wrmebergangs-
steigerung durch Anregung des Geschwindigkeitsfeldes beim Strahleintritt wird prsentiert.
Die LES zeigt, dass die Anregung bei der bevorzugten Mode die flchengemittelte Nus-
seltzahl im Vergleich zum nicht angeregten Fall steigert. Allerdings ist Vorsicht geboten
bei der Auswahl von Anregungsfrequenzen, die hher sind als die bevorzugte Mode, da sie
thermische Ermdung hervorrufen knnen. Zustzlich wird die Auswirkung der Strahlver-
wirbelung auf den Wrmebergang untersucht. Darber hinaus wird festgestellt, dass die
Verwirbelung den Wrmebergang fr H/D=2 nicht nennenswert steigert.
Die durch diese Simulationen gewonnenen Kenntnisse der Strmungsmechanik und der
passiven Skalarflsse sind hilfreich fr das Verstndnis des funktionalen Zusammenhangs
zwischen verschiedenen Turbulenzgren und dem Wrmebergang. Die Annahme einer
konstanten turbulenten Prandtlzahl (wie sie oft in RANS-Modellen Anwendung findet) ist
vielfach nicht realistisch. Alternativ empfiehlt es sich, Skalarfluss-Modelle zu verwenden,
die eine Vorhersage der Skalarflsse in anisotroper Turbulenz erlauben. Das mit Hilfe der
LES-Rechnungen gewonnene Verstndnis wird benutzt, um die Koeffizienten in einigen
expliziten (RANS basierten) Skalarfluss-Modellen zu analysieren. Die Untersuchung ver-
mittelt einen Einblick in das Verhalten eines Prallstrahls, was fr die weitere Entwicklung
von Turbulenzmodellen fr die Simulation des Wrmebergangs hilfreich sein kann.
CONTENTS
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Zusammenfassung . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation: Complex Flows . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Impinging Jet: Applications . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Impinging Jet as a Complex Flow Problem . . . . . . . . . . . . . . . . . 8
1.3.1 Fluid Dynamics of an Impinging Jet . . . . . . . . . . . . . . . . . 8
1.3.2 Heat Transfer by an Impinging Jet . . . . . . . . . . . . . . . . . . 11
1.4 Research Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.1 Heat Transfer due to Natural Jet Impingement . . . . . . . . . . . . 15
1.4.2 Heat Transfer due to Excited Jet Impingement . . . . . . . . . . . . 18
1.4.3 Heat Transfer due to Swirling Jet Impingement . . . . . . . . . . . 21
1.5 Goal of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2. Turbulence: Phenomenon and Prediction . . . . . . . . . . . . . . . . . . . . . 25

2.1 Govering Equations of Fluid Motion . . . . . . . . . . . . . . . . . . . . . 25
2.2 Turbulence Scales and Spectra . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.1 Single Point Statistics . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2.2 Averaging of Turbulent Flow . . . . . . . . . . . . . . . . . . . . . 29
2.3 Turbulence Modeling and Simulation . . . . . . . . . . . . . . . . . . . . 29
2.3.1 Direct Numerical Simulation . . . . . . . . . . . . . . . . . . . . . 29
2.3.2 Large Eddy Simulation . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.3 RANS Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
viii Contents
3. Numerical Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Comparison of Discretization Approaches . . . . . . . . . . . . . . . . . . 49
3.2 FASTEST- Basic Introduction . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3 Spatial Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.1 Convective Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.2 Diffusive Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.3 Source Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4 Time Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.1 Crank-Nicolson Method . . . . . . . . . . . . . . . . . . . . . . . 55
3.5 Pressure-Velocity Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6 Solution of System of Equation . . . . . . . . . . . . . . . . . . . . . . . . 58
4. Large Eddy Simulations & Results . . . . . . . . . . . . . . . . . . . . . . . 61

4.1 LES of Natural Jet Impingement . . . . . . . . . . . . . . . . . . . . . . . 61
4.2 LES of Jet Impingement with Active Velocity Field Excitation . . . . . . . 89
4.3 LES of Jet Impingement with Passive Velocity Field Excitation . . . . . . . 99
4.4 LES of Impinging Jet with Active Temperature Field Excitation . . . . . . 107
4.5 LES of Swirling Jet Impingement . . . . . . . . . . . . . . . . . . . . . . 114
5. An Assessment of Scalar-flux Modeling for Complex Flows . . . . . . . . . . . 125

5.1 Turbulent Heat Flux Distribution . . . . . . . . . . . . . . . . . . . . . . . 128
5.2 Calibration of Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.2.1 Daly and Harlow Model . . . . . . . . . . . . . . . . . . . . . . . 130
5.2.2 Abe and Suga Model . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.2.3 Younis et al. Model . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.3 Similarity of Mechanical and Thermal Time Scales . . . . . . . . . . . . . 135
6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Appendix 163
NOMENCLATURE
English Symbols
Symbol Units Meaning
A m2 Area
A2 , A3 Second & third anisotropy invariant tensor
AN Amplitude of excitation
ai j Reynolds stress anisotropy tensor (=2 bi j )
B Constant in law of the wall
bi j Anisotropy tensor (=ai j /2) (Lumleys definition)
C1 ,C2 ,C3 ,C4 Coefficients in Younis et al. model
Cg Germano model parameter
Ci j m2 /s3 Convection term (Reynolds Stress Model)
Ck m2 /s3 Convection of turbulent kinetic energy
Cri j N/m2 Cross terms in subgrid modeling
C1 Coefficient in Daly & Harlow model
C2 Coefficient in Abe & Suga model
cp J/kgK Specific heat at constant pressure
cs Smagorinsky constant
Coh( f ) m2 /s Coherency spectrum
D m Diameter
DTi mK/s2 Diffusion term (scalar-flux transport equation)
d m Diameter of the cylindrical insert
di j m2 /s3 Diffusion term (Reynolds Stress Model)
E m2 /s Energy
F N Body forces
F Flux
x NOMENCLATURE
English Symbols (continued)

F Flux vector in surface normal direction
f s1 Frequency
fn s1 Frequency at preferred mode
fn/2 s1 Subharmonic of preferred mode
f2n s1 Frequency twice of the preferred mode
f5n s1 Frequency five times of preferred mode
fo s1 Most amplified frequency
G Filter function
gi j Tensor (inflow turbulence generation)
H m Distance between jet outlet and target wall
h W /m2 K Convective heat transfer coefficient
I, II, III First, Second & Third invariant tensors
k m2 /s2 Turbulent kinetic energy
k m1 Wave number
Li j N/m2 Leonard terms in subgrid modeling
Mi j N/m2 Stress tensor (Germano model)
m kg/m2 Mass flux
N Number of control volumes
ni Projection of unit vector normal to surface
Nu Nusselt number
Nu1 Nusselt number at first peak
Nu2 Nusselt number at secondary peak
P N/m2 Pressure
Pi j m2 /s3 Production term (Reynolds Stress Model)
Pk m2 /s3 Production of turbulent kinetic energy
PTi mK/s2 Production term (scalar-flux transport eq.)
p Corrected pressure value
q W /m2 Heat flux
Q Q - criterion
Pr Molecular Prandtl number
Prt Turbulent Prandtl number
Prs Subgrid Prandtl number
NOMENCLATURE xi
English Symbols (continued)

R m Radius of jet at outlet
Rt Time correlation function
R` Spatial correlation function
Ri j N/m2 Unresolved Reynolds stresses
r m Radius or radial coordinate
r m Position vector
rm s Mechanical time scale
rt s Thermal/heat time scale
Re Reynolds number
Rek Turbulent Reynolds number
S Swirl number
St Strouhal number
ST Source term in energy equation
Sr Surface renewal parameter
Si j s1 Strain rate tensor
S Generalised Source term
t s Time
T K Temperature
Tad K Adiabatic temperature
To K Jet inlet temperature
Tu Turbulence intensity
U m/s Mean velocity
Ub m/s Bulk velocity
Um m/s Local maximum velocity in wall jet zone
u m/s Velocity in x- direction
ur m/s Velocity in radial direction
u m/s Velocity in azimuthal direction
uz m/s Velocity in vertical direction
u m/s Frictional velocity
V m3 Volume
v m/s Velocity in y- direction
w m/s Velocity in z- direction
x m x- Coordinate
y m y- Coordinate
Zi j N/m2 Stress (Germano model)
z m z- Coordinate
xii NOMENCLATURE
Greek Symbols
i j N/m2 Stress tensor (Germano model)
p Under-relaxation factor
m Characterstic filter-width
r,,z m Grid spacing in r, & z direction
m Boundary layer thickness
T m Thermal boundary layer thickness
m Displacement thickness
m Momentum thickness
i j Kronecker delta
i j m2 /s3 Dissipation term (Reynolds Stress Model)
m2 /s3 Dissipation of turbulent kinetic energy
T mK/s2 Destruction rate
Thermal diffusivity
e Appropriate thermal diffusivity in energy eq.
s Subgrid thermal diffusivity
t Eddy thermal diffusivity
Generic transport coefficient
e Interpolation factor
i j m2 /s3 Redistribution of pressure fluctuations
Parameter related to swirl
m Kolmogorov length scale
Index in tangential velocity expression
Von Karman constant
m/s Velocity at convective outlet
W /mK Thermal conductivity
Pa.s Dynamic or absolute viscosity
s Pa.s Subgrid viscosity
n nth time step
m2 /s Kinematic viscosity
t m2 /s Turbulent eddy momentum diffusivity
rad Azimuthal direction
Parameter based on radius
Ti mK/s2 Pressure-scalar gradient correlation
NOMENCLATURE xiii
Greek Symbols (continued)

kg/m3 Density
m Local coordinate
i j Velocity gradient tensor (antisymmetric)
i j N/m2 Stress used in Germano model
s1 Vorticity
N/m2 Stress tensor (subgrid modeling)
m2 Surface enclosing CV
Arbitrary conservation quantity
m Distance from the jet outlet
` m Integral length scale
Subscript
Symbol Meaning
<.> Ensemble averaged quantity
(.)ax Jets axial value
(.)CL Jet centerline value
(.)C Neighbouring control volumes
(.)D Based on diameter of the jet
(.) jet jets value
(.)P Central point of CV
(.)stg Stagnation value
(.) Frictional quantities
(.)w Wall values
(.)E ,(.)ea Central point of neighbouring CV on east side, face
(.)W ,(.)we Central point of neighbouring CV on west side, face
(.)N ,(.)no Central point of neighbouring CV on north side, face
(.)S ,(.)so Central point of neighbouring CV on south side, face
(.)B ,(.)bo Central point of neighbouring CV on bottom side, face
(.)T ,(.)to Central point of neighbouring CV on top side, face
xiv NOMENCLATURE
Superscript & other symbols

Symbol Meaning
(.)a Anisotropic
(.)corr Corrected value from previous iteration
(.)sgs Subgrid scale
(.)test Test filter (Germano model)
(.) Value from previous iteration
(.)0 Turbulent fluctuating quantities
(.)+ Dimensionless = (.)u/
(.) Reynolds time averaged quantity
(.)
c Large scale component which is time dependent (LES)
(.)
f Small-scale, or subgrid components (LES)
Acronyms
Acronym Name/Phrase
AIM Anisotropy Invariant Map
CDS Central Difference Scheme
CFD Computational Fluid Dynamics
CFL Courant Friedrich Levy Criterion
CV Control Volume
DNS Direct Numerical Simualtion
EARSM Explicit Algebraic Reynolds Stress Model
ERCOFTAC European Research Community on Flow, Turbulence and
Combustion
ESFM Explicit Scalar Flux Model
EVM Eddy Viscosity Model
FASTEST Flow Analysis Solving Transport Equations Simulating
Turbulence
GDH Gradient Diffusion Hypothesis
HLRS Hchstleistungsrechenzentrum
KH Kelvin Helmholtz
LDA Laser Doppler Anemometry
LES Large Eddy Simulation
LHS Left Hand Side
LU Lower-Upper decomposition
NOMENCLATURE xv
Acronyms
Acronym Name/Phrase
NGV Nozzle Guide Vane
ODE Ordinary Differential Equation
PDE Partial Differential Equation
PIV Particale Image Velocimetry
PSD Power Spectral Density
MPI Message Passing Interface
RANS Reynolds Averaged Navier-Stokes
RHS Right Hand Side
RSM Reynolds Stress Model
SIMPLE Semi-Implicit Method for Pressure-Linked Equations
SGS Subgrid Scale Model
SMC Second Moment Closure
SIP Strongly Implicit Procedure
SSG Speziale-Sarkar-Gatski
UDS Upwind Difference Scheme
1. INTRODUCTION
1.1 Motivation: Complex Flows

The increase in performance of high speed computation has changed the approaches of
engineering analysis in many fields. Benefits of this progress can be appreciated as the
solutions obtained by discretization of mass, momentum and energy equations has become
a routine task in industry and academia. Through Computational Fluid Dynamics (CFD) a
better understanding of flow physics especially turbulence has been achieved along side the
conventional analytical and experimental techniques.
CFD offers broadly the following three approaches for the analysis of a turbulent flow
problem:
Direct Numerical Simulation (DNS)
Large Eddy Simulation (LES)
Models based on Reynolds Averaged Navier-Stokes (RANS) equations
In DNS the complete flow field is simulated and it is desired to do that till the smallest
scale of the flow. This naturally demands high computational resources and can be done
for small Reynolds number flows as the number of grid points in spatial discretization is
proportional to Re9/4 , where Re is the Reynolds number of the flow [190].
In LES only the large scales are simulated and small scales are modeled, and thus in
terms of cost it is less expensive as DNS and can be applied to higher Reynolds number
flow. Both LES and DNS gives the instantaneous flow field and they must be statistically
averaged in order to get the mean flow field.
On the other side, in RANS the complete flow field is modeled so the main burden lies
on the model used. As flow in this case is statistically averaged, only the mean flow field is
predicted. Any deterministic structure in flow must be computed as part of flow. Therefore,
the models provide only the low order statistics.
There has been a constant interest in the field of CFD for improvements of turbulence
models. Turbulence is inherently a dissipative phenomenon, therefore, most of the flows
2 1. Introduction
that survive long enough to be of interest of scientist and engineers have an energy input
to turbulent motion. The shear flows (boundary layer flows, jets and wake flows etc.) are
the most common kind of flows encountered in engineering problems. Two dimensional
thin shear flows are the simple flows, in which, the mean deformation rates are mild and
turbulence has time to come to equilibrium with the mean flow. In simple flows the principal
directions of the Reynold stress tensor tend to be aligned with those of mean strain. For such
kind of flows, an eddy-viscosity representation is adequate. In real engineering problems,
the flows are three-dimensional, affected by adjacent turbulent fields, body forces, shock
waves and other perturbations. These complex shear flows dominate the engineering flows.
In complex flows the principal axes of Reynolds stress tensor are not aligned with those of
the mean deformation tensor. Because of this, the flow near the wall does not adequately
comply with universal log laws as near wall turbulence departs very far from a state of local
equilibrium. Also the flow statistics depend on more than one coordinate and possibly on
time as well. Some of the examples of complex flows are:
Shear layers on rotating bodies, such as those occur on turbomachinery rotor blades,
helicopter rotors and meteorological flows. In such flow the corriolis forces affect
directly the turbulence structures.
Shear layer developing on curved surfaces, such as boundary layer on turbine blades,
wall jets used in cooling application and curved flow in turbomachinery passages.
Three dimensional boundary layers and wakes e.g. flow over finite wings, fuselage,
marine vehicles, turbomachinery hub, and wall boundary layers.
Shear layer with lateral convergence and divergence, as in flow through annular dif-
fusers or nozzles.
Separated flows.
Swirling flows and vortices e.g. Tornadoes, hurricanes, turbomachinery annulus
flows etc.
Flow with strong compressibility effects.
Flow with buoyancy, chemical reaction etc.
These complexities affect the turbulence stresses and scales directly. The prediction meth-
ods for these complex flow are in a state of constant development and improvement. One
of the major requirements and expectation of turbulence models are that they mimic faith-
fully the flow and turbulence physics. Especially in an analysis of a problem of industrial
significance the need to model complex flows with higher accuracy and certainty is urgent.
In this thesis, the complex problem of an impinging jet with heat transfer has been sim-
ulated using LES. Impinging jet is a challenging problem for CFD modeling as it contains
many complex flow features like free shear flow, near wall flow, flow stagnation, boundary
layer acceleration and then deceleration.
1.2. Impinging Jet: Applications 3
1.2 Impinging Jet: Applications

Impinging jets are used in many industrial applications because of their capability of effec-
tive heat and mass transfer. Due to the importance of impinging jet configurations some of
their application are discussed following.
Gas Turbine Component Cooling: Advanced gas turbine engines operate at high tem-
perature to improve power output and thermal efficiency. Figure 1.1 shows that the combus-
tion firing temperature increases with the advances in the material used for the gas turbines.
As the firing temperature for combustion increases the heat transfered to the turbine com-
ponents also increases. The operating temperatures are far above the permissible metal
temperatures. The typical temperature distribution experienced by the turbine vane surface
is shown in figure 1.2. Therefore for the safe operations, turbine component cooling is
mandatory. To accomplish this the complex passages inside the turbine vanes/blades are
created. Figure 1.3 shows one of such types of the complex passages inside the high pres-
sure nozzle guide vane (NGV). The blades are cooled both internally and externally by the
air extracted from compressor of the engine. Typical high pressure bleed flows are at 600
oC and used to cool the turbine immersed in gas having total temperature of 1400 oC [77].
The gas turbines component cooling by impinging jets is not possible at every location, as
the construction of an impinging jet system weakens the structural strength [77]. They are
used only at the locations where the thermal loads are excessively high like turbine guide
vanes, rotor blades, rotor disks, combuster case and combuster walls etc..
Figure 1.1: Firing temperature increases with blade material improvement [17]
4 1. Introduction
Figure 1.2: Turbine vane cross section with surface temperatures taken from [116].
Figure 1.3: Complex passages inside a gas turbine high pressure nozzle guide vane (NGV)
with impingement and film cooling [99].
Figure 1.4: Compact High Intensity Coolers with multiple jet impingement cooling con-
cept.
They are also generated indirectly by passage geometries like impingement of shear
layer from ribs (used for heat transfer enhancement in blades or vanes).
Compact High Intensity Cooler: Compact High Intensity Coolers are the kind of heat-
exchange devices, where an array of impinging jets strikes the heated surface with orifices.
The meandering flow through plates creates further impingement possibilities and this in
turn improves the heat transfer. They are used in applications where the primary criterion is
the weight of exchanger for example, air conditioning units in airplanes. Figure 1.4 shows
the heat exchanger with application of multiple impinging jets.
Electronic Cooling: The cooling of electronic components by the impingement of cold
fluid is a promising approach. The increase in computational resources has brought the
advances in the electronic components cooling approaches. The jet impingement is a well
tested approach for electronic cooling. A CPU-cooler with surface enlargement and jet
impingement is shown in figure 1.5.
Microburst: The microburst is a natural event in which mass of cold dense air falls to
the ground and spreads radially with strong hurricane-force winds. The schematic repre-
sentation of this phenomenon is depicted in figure 1.6. This impingement phenomenon is
affecting the ground structures and flying aircrafts for example, the vertical wind shear pro-
duced by microburst are considered to be the major cause of the number of fatal accidents.
The understanding of this phenomenon can help in reducing the accidents caused by them.
Vertical Take-off and Landing Aircrafts: Impinging jets are used for the generation
of the lift force in Vertical take-off and landing aircrafts. The jets are produced by directing
6 1. Introduction
Figure 1.5: CPU-cooler with surface enlargement elements. An array of impinging jets is
used for cooling (ITLR).
Figure 1.6: Atmospheric micro burst phenomenon (adapted from Encyclopedia Britannica).
the hot jet engine exhaust towards the ground. They are used in applications like, vertical
take off and landing of aircrafts, rockets, aircrafts de-icing etc.. Figure 1.7 shows the use of
VTOL.
Figure 1.7: Vertical take-off and landing (VTOL) aircraft (F35). Hot jets directed towards
the ground generate a lift force. (adapted from Lockheed Martin Inc. website)
Mass Transfer Applications: They are used in numerous industrial applications for
mass transfer like drying, mine ventilation, tunneling operation (mass transfer), paint spray-
ing, cavitation drilling, shielded arc welding, laser cutting and welding. They are also
important for safety engineering studies like accidental release of flammable and toxic ma-
terials.
The dynamics of flow and heat in an impinging jet is very complex involving jet stagna-
tion, ring vortex impact, acceleration of flow and then deceleration. In case of heat transfer
due to impinging jet a secondary peak in heat transfer distributions is observed experimen-
tally. The RANS based simulations are found to be incapable of capturing the secondary
peak. As can be seen in above mentioned applications, the arrays of impinging jets are
normally used for practical applications, however, for understanding of this phenomenon
the investigation of a single jet impingement is a starting step. Therefore, in the thesis the
focus is placed on the single jet impingement only.
8 1. Introduction
1.3 Impinging Jet as a Complex Flow Problem

1.3.1 Fluid Dynamics of an Impinging Jet
Like all other flows, impinging jet flow can be laminar or turbulent depending on the
Reynolds number (Re = Ub D/) of the jet. Different flow regimes of an impinging jet
are [185]:
The dissipated laminar jet (Re < 300)
A fully laminar jet (300 < Re < 1000)
A transition or semi-turbulent jet (1000 < Re < 3000)
A fully turbulent jet (Re > 3000)
An impinging jet structure can be divided into three different zones:
Free jet zone
Impingement zone
Radial wall jet zone
The zones are depicted schematically in figure 1.8.
Figure 1.8: Schematic representation of an jet impingement.

1.3. Impinging Jet as a Complex Flow Problem 9
(a) Free jet structure (adapted & amended from Yule [198]).
(b) Visualization of an impinging jet structure through smoke wire method (Popiel and
Trass (1981); taken from [134].)
Figure 1.9: Structure of a free and an impinging jet.

10 1. Introduction
A free jet consist of large eddies which are repetitive in structure, remains coherent for
downstream distances which are very much greater than their length scales and contribute
greatly to the properties of the turbulent flows. These structures are generally called co-
herent structures. The coherent structures in a free jet are presented schematically in figure
1.9(a). The jets shear layer near the outlet is initially dominated by the Kelvin Helmholtz
(KH) instability mechanism. The KH instabilities arise due to the fast moving jet flow and
slow moving surrounding fluid. In a circular jet the ring vortices are generated due to Kelvin
Helmholtz instabilities. The formation of a ring vortex is attributed to the most amplified
frequency of small disturbance which grows exponentially.
In literature, distinct jet modes, based on the most amplified frequency, are defined.
These modes are based on the jet outlet conditions. One is called the shear-layer mode and
the other one the jet-column mode. The initial region of jet shear layer can be considered
as parallel or two dimensional if the ratio of jet diameter to initial momentum thickness is
large. In this case, the proper scaling parameter is momentum thickness [95]. This mode
represented by the Strouhal number (scaled with momentum thickness) is called a shear
layer mode. Gutmark and Ho [73] have indicated that the shear layer mode can occur in the
range of the Strouhal numbers.
0.01 St 0.018 (1.1)
where St is defined as fo /Uo . fo is the most amplified frequency and Uo is the jets ve-
locity. This mode is important in case of jet emergence from a converging nozzle. However,
if the jet emerges in a fully developed turbulent flow state from a long pipe, the momentum
thickness is of order of the jet radius and the jet does not exhibit the shear-layer mode, in-
stead it will attain the jet column mode. The most amplified frequency in the jet column
mode scaled with diameter is called the preferred mode of the jet, defined by the range of
Strouhal numbers [95]:
0.25 StD 0.85 (1.2)
where StD is defined as fo D/Uo . According to Hussain [95], the preferred mode of an
axisymmetric jet can occur in the range of the Strouhal numbers StD = 0.3 - 0.5. Gutmark
& Ho [73] have found that the preferred mode depends upon the initial conditions. Liu and
Sullivan [128] have found that the dimensionless natural frequency of a jet decreases as it
evolves.
As the free jet develops, the ring vortices can also merge together, the phenomenon
known as vortex pairing. The pairing phenomenon can occur in impinging jets as well. It is
an intrinsic feature of the development of the free jet and occurs irregularly in natural jets
(non-excited jets) [159].
The shear layer which emanates from the nozzle exit contains the high vorticity. Large-
scale structures of a free jet will lose coherence as the distance from the outlet increases.
These structures will give rise to small scale turbulence. The free jet structure can be classi-
fied on the basis of behavior of these structures. The free jet may consist of a potential core
region [contains initial shear layer, till /D is 4 - 5], flow development region [/D 5-8]
and the fully developed region [after /D 8], where is the distance from jet outlet.
The above mentioned regions will form depending on how close the target wall is to
the jet outlet. For small jet outlet-to-target wall distances (like H/D = 2), the free jet does
not contain the flow development region and fully developed free jet region. The structures
of an impinging jet, visualized through smoke wire visualization technique, are depicted
in figure 1.9(b). Due to jet instabilities the ring vortices are generated, which strike the
target wall. As the jet approaches the wall, the static pressure increases and the momentum
decreases. Through both experiments and numerical simulation (LES) it is found that there
is a negative turbulence production rate in the stagnation region [147; 30; 78]. It means that
part of turbulence energy field goes back to the mean flow.
The jet deflects and spread radially due to the presence of the wall. The wall jet region
is comprised of very thin accelerating boundary layer. The complex flow involving strong
shear is produced. Although, the flow is now parallel to the wall, the turbulence levels in a
wall jet zone are much higher than the parallel flows.
1.3.2 Heat Transfer by an Impinging Jet

Radial distribution of the Nusselt number in an impinging jet can be defined as:
qw D
Nu = (1.3)
Tw T j
where T j is jet inlet temperature, Tw is the target wall temperature, qw is the heat flux at
the wall. D is the diameter of the jet and is thermal conductivity. The heat transfer in
an impinging jet is greatly affected by the nozzle-to-wall distance (H). For H/D less than 4,
prominent peaks appear in the radial distribution of the Nusselt number. In figure 1.10 the
radial distribution of the Nusselt number obtained from various experiments is shown. The
jet outlet-to-target wall distance (H/D) is 2.
The heat transfer is affected by large-scale coherent structures especially for small H/D
cases. zdemir and Whitelaw [150] reported that the large vortical structures cause local
augmentation of convective heat transfer coefficient. According to Fox et al. [48] it is
the competition between the vortex rings and secondary vortices induced by them on the
impinging plate that determines the wall surface temperature. Kataoka et al. [109] have
used the statistical technique with conditional sampling to study the mechanism of jet im-
pingement. They found that large vortical structures increase heat transfer. They proposed
the Surface renewal effect according to which the impingement wall experiences the large
scale structures striking the surface. Sakakibara et al. [168] used digital particle image
12 1. Introduction
Figure 1.10: Radial distribution of the Nusselt number at wall for H/D=2.
velocimetery (DPIV) and Laser induced fluorescence (LIF) to measure instantaneous and
time dependent velocity and temperature distribution. It is found that the amplification of
vorticity in stagnation region is the main source of enhancement of heat transfer in plane
impinging jets.
Maximum Stagnation Point Heat Transfer
At small nozzle-to plate spacing (H 2), the stagnation point Nusselt number varies ac-
cording to Nuo ReD0.5 [121]. Lee & Lee [121] have reported the stagnation point Nusselt
number correlation as:

Nuo = 0.443Re0.565
D (H/D)0.0384 (1.4)
This is valid for a nozzle-to-plate spacing of 2 - 10. Through experiments, it has been found
that the Nusselt number at stagnation point has its maximum value around H/D 8 [76].
A more recent investigation by Lee et al. [124] reported the maximum Nusselt number at
H/D equals to 7. According to them the physical mechanism for maximum Nu to occur
at H/D equal to 7 is that the turbulence intensity reaches roughly a maximum value in that
region. Ashforth-Frost et al. [9] reported that the maximum heat transfer rates occur at a jet
to plate spacing of 6 for unconfined jets and 6.5 for semi-confined jets. Viskanta [185] gives
the range of H/D = 5 8 for possible maximum stagnation point heat transfer. Kataoka
et al.[109] found that the Strouhal number based on characteristic time scale is maximum
at H/D 6 and then decreases rapidly for H/D > 6. According to them this suggest the
frequency of turbulent surface renewal due to large eddies becomes maximal at the optimal
nozzle-to-plate spacing of H/D 6.
Secondary Peak in the Radial Distribution of the Nusselt Number
The occurrence of secondary peaks in Nusselt number distribution for low nozzle-to-target
plate spacing in jet impingement heat transfer is a known phenomenon. It has been found
that the occurrence of secondary peaks depends not only on nozzle to plate spacing but also
on the turbulence levels in the jet [87]. Over the years investigators have proposed differ-
ent interpretations for this phenomenon, but there has been no consensus on the physical
explanation. Gardon and Akfirat [59] suggested that the peak is due to local thinning of
the boundary layer. It is also suggested that the secondary peak is due to flow transition in
the wall jet region [87]. Lytle and Webb [131] attributed the occurrence of the secondary
peak to the higher turbulence in boundary layer due to flow acceleration and intense shear
between the radially ejecting jet and stagnant ambient. Behnia et al. [14] also endorsed this
explanation. Viskanta [185] proposed that the ring shaped wall eddies are responsible for
this secondary peak which are induced by the large scale toroidal vortices hitting the plate.
It is interesting to note that even in laminar jet the peak is present. Chung and Luo [24]
have done a DNS of impinging jet of Reynolds number around 500 and 1000. For such low
Reynolds number flows the secondary peak is observed. It is expected that for such low
Reynolds number flows, the transition to turbulence is not expected to play any role. They
have attributed the secondary peak to the engulfing motion around the secondary vortex,
which reduces the thickness of the thermal boundary layer.
The relationship between the local peaks is found to be [121]:
Nu1 /Nuo = 0.696Re0.047

D (1.5)
Nu2 /Nuo = 0.400Re0.077

D (1.6)
If the jets temperature and the surrounding temperature are different, the wall is then
cooled by a jet with temperature depending on the proportion of the original jet air and the
entrained ambient air. In order to account for the entrainment of temperature mis-match
between the jet and the ambient air on the evaluation of the heat transfer coefficient, the
14 1. Introduction
adiabatic wall temperature can be used as a reference temperature:

T
k n wall
h= (1.7)
Tw Tad
where Tad is the temperature attained by a wall under adiabatic conditions i.e. qw = 0, Tw is
the target wall temperature attained in case of non-adiabatic conditions.
The temperature attained by the fluid when a moving fluid stream is brought to rest
(under adiabatic conditions) is called adiabatic wall temperature or recovery temperature.
Under adiabatic conditions, the amount of kinetic energy of the flow transformed into the
heat depends on the velociy of the jet. It is found that the thermal entrainment can be
accounted through the recovery factor defined as:
Tad T jet
recovery f actor = 2 /2c
(1.8)
U jet p
In defining the local Nusselt number, Gardon and Cobonpue [58], Goldstein et al. [66] and
Giovannini and Kim [65] employed the local adiabatic wall temperature, while others used
the jets exit temperature [131; 12].
It is found by some researchers that by using the adiabatic wall temperature in the defi-
nition of the Nusselt number the secondary peak can be removed [76]. Here, it is important
to note that in case of high speed jets with large thermal entrainment effects, (for example,
in the experiments of Gardon and Cobonpue [58] (Re = 7000 - 112000) and Goldstein et
al. [66] (Re = 61000 - 120000)) not only the jets velocities are high but also the thermal
entrainment is present, causing the large difference between the jets inlet temperature and
the adiabatic wall temperature. On the other hand, Giovannini and Kim [65] (Re = 23000)
have found that the secondary peak is present in the radial distribution of Nusselt number
even when the Nusselt number is defined on the basis of adiabatic wall temperature.
It can be summarised that in case of high thermal entrainment effects or high jet ve-
locity, the adiabatic wall temperature is different from the jets inlet temperature. In such
cases the definition of Nusselt number based on the adiabatic wall temperature actually rep-
resents a hybrid problem of wall heated by the high speed jet impacting on an adiabatic
wall plus the wall heated by jet with non-unform temperature. It does not represent the
true physics but offers only an approach towards problem simplification. Therefore, in the
authors opinion the secondary peak in the radial distribution of the Nusselt number is not
really an artifact of the way the Nusselt number is defined, but a physical phenomenon.
1.4. Research Overview 15
1.4 Research Overview

1.4.1 Heat Transfer due to Natural Jet Impingement
Jet impingement has many interesting fluid dynamics and heat transfer aspects and over the
years many researchers have investigated them. Martin (1977) [133] gives an overview of
this field. In Gas Turbine component cooling the nozzle exit area is very small and owing
to high jet velocity the Reynolds number is quite large. Martins review is concerned with
the heat and mass transfer related to industrial and drying processes for which the nozzle
diameters are large. Recent reviews by Han and Goldstein (2001) [76], Buchlin (2000) [18],
Viskanta (1993) [185], Jambunathan et al. (1992) [100] are more relevant in terms of Gas
Turbine component cooling and have discussed the number of areas significant for further
research. The jet is investigated under different conditions like cross flow, the oblique im-
pingement and impingement on curved surfaces. Also the role of turbulence, nozzle exit
shape, entrainment, jet outlet conditions and large scale eddies have been investigated.
Research in jet impingement area can be categorized on the basis of the shape of the
nozzle exit area. It is summarized as:
Single Slot Nozzle (plane jet)
Array of slot nozzles (multiple plane jets)
Single round nozzle (circular jet)
Array of round nozzles (multiple circular jets)
Experimental Investigations
The experimental investigations of impinging jets are numerous. There is vast body of
literature on this topic, covering entrainment, turbulence, cross flow, swirl, nozzle-to-plate
distance, inlet configurations, impingement angles, turbulence statistics etc.
Some of the important investigations related to a single plane or slot jet are tabulated in
table 1.1 and the investigations related to single circular jet are tabulated in table 1.2.
Numerical Investigations
The numerical simulations of an impinging jet has been reviewed by Polat et al. [158],
Zuckermann and Lior [204] and by Coussirat et al. [26]. Polat et al. have discussed the
laminar impinging jets in their review, but as most of the industrial jets are turbulent the
later reviewers have focused only on the turbulent jet impingement.
Recently the Direct Numerical simulation (DNS) of unsteady heat transfer of impinging
jet is done by Chung and Luo [23].They performed the simulation of three Reynolds num-
bers 300, 500 and 1000. They found that unsteady heat transfer characteristics are strongly
16 1. Introduction
Table 1.1: Single Plane Jet - Experiments
Researchers Year D (mm) Re H/D

Gardon and Akfirat [59] 1965 1.59 6.35 2800 22000 2 80
Striegl and Diller[179] 1984 3 15 3000 40000 6 25
Gau and Lee[54] 1992 3.5 20 2500 11000 4,8 and 12
Ashforth-Frost et al. [9] 1997 30 20000 4 and 9.2
Buchlin[18] 2000 1.3 and 3 5500 17000 8 20
Camci and Herr[21] 2002 14000 24
correlated with vortex dynamics of the jet flow. Cziesla et al. [30] have done a DNS in-
vestigation of multiple impinging jets. Tsubokura et al. [183] have done a DNS simulation
of single circular and plane jet. Hattori et al. [74] have done the DNS of plane impinging
jet. Satake et al. [169] have conducted the DNS of circular jet in order to investigate the jet
structures.
Large Eddy Simulation (LES) studies provide both statistical and structural flow infor-
mation. In terms of LES, Rizk and Menon [166] were the first to conduct a Large Eddy
Simulation of impinging jets. They only investigated the flow field. Voke, Gao and Leslie
[56] and Gao, Leslie and Voke [55], using LES, investigated the mechanisms by which ther-
mal eddies are generated and transported into impingement region in order to understand
the thermal interaction between the fluid and solid. Olsson and Fuchs [149] used LES with
two different subgrid scale (SGS) models in order to investigate the physical behavior of the
jet in the near wall region at the Reynolds number of 10000. The stress-similarity model
performed slightly better with higher resolution than the coarse grid simulations. Revst-
edt et al. [165] investigated the mixing characteristics of the Dynamic Divergence Model
(DDM) which is based on the Germano Dynamic Model and the Scale Similarity model
(SSM). It is concluded that the choice of SGS model has significant effects on the results
of the simulation. Revstedt et al. also pointed out the need of the further study for the
account of the influence of the near wall resolution on SGS models. Cizesla et al. [30; 31]
have tested the single and multiple plane impinging jets. Tsubokura et al. [183] have tested
the plane and circular jets with periodic inlet conditions. Gago et al. [52] have tested nine
types of subgrid models. They recommended the hybrid Smagorinsky model for simulation.
Hllqvist [79] have investigated an impinging jet using LES at Reynolds number of 20000
and H/D=2. In his computations, no explicit subgrid-scale model has been used. Instead,
the numerical scheme is used to account for the necessary amount of dissipation. The radial
distribution of Nusselt number is not correctly predicted. Hadziabdic [78] investigated an
Table 1.2: Single Circular Jet - Experiments
Researchers Year D (mm) Re H/D

Gardon and Cobonpue[58] 1962 2.3 9 7000 112000 0.5
Hoogendoorn [87] 1977 57 66000 2 and 10
Popiel et al.[160] 1980 13.8 1000 1900 2 20
Goldstein and Behbabhani[68] 1982 12.7 25000 124000 6, 12
Ward and Mahmood[188] 1982 32000 60000 2 12
Didden and Ho[38] 1985 38.1 19000 4
Goldstein et al.[66] 1986 12.7 61000 120000 2 12
Popiel and Boguslawski[161] 1986 4500 44000 1.2
Baughn and Shimizu[12] 1989 25 23750 2 14
Goldstein et al.[67] 1990 12.7 124000 2 12
Baughn et al.[13] 1991 26 23300 and 55000 2, 6 and 10
zdemir and Whitelaw[150] 1992 10 13000 22
Cooper et al.[28] 1993 26 101.6 23000 and 70000 2 10
Fox et al.[48] 1993 10.16 400 0.25 30
Lytle and Webb[131] 1994 7.8, 10.6 3600 28000 0.1 6
Lee et al.[120] 1995 10.85 4000 14400 2 10
Nishino et al.[147] 1996 40 10400 and 13000 5.63 and 5.86
Ashforth- Frost et al.[10] 1997 100 20000 18
Lee and Lee[121] 1999 25 5000 30000 2 10
Lee et al.[124] 2004 13.6 34 23000 2 14
Guerra et al. [63] 2005 35000 2
impinging jet case using LES at Reynolds number of 20000 and H/D=2, using unstructured
grids. The dynamic Smagorinsky model is used for computations. The predictions of the
radial distribution of Nusselt number are found to be good in the stagnation zone. However
the secondary peak is not captured.
In general the insight given by the results of any Direct Numerical Simulation (DNS)
or Large Eddy Simulation (LES) is valuable. Zuckermann and Lior have discussed that the
capability to predict the secondary peak by LES and DNS is excellent. As Reynolds num-
ber increases, the length scales involved become smaller than the macroscopic length scales.
This inhibits the simulation at higher Reynolds number. The only viable tools are the RANS
based simulations at high Reynolds numbers. Unfortunately, it is found in general that the
turbulence models available for the RANS simulation, fall short of predicting correctly the
18 1. Introduction
flow and heat transfer fields. Craft et al. [29] have tested k and Reynolds Stress turbu-
lence models. They concluded that both eddy viscosity models and basic Reynolds Stress
Models achieved very poor agreement with experiments. Many other researchers have in-
vestigated the k model and its variations. It has been found that although geometrically
simple, impinging jet involves several flow phenomenon which k model cannot handle
e.g. stagnation point heat transfer and stream line curvature. Heyerichs & Pollard [82] tested
k , k and low Reynolds number models and concluded that the wall function models
perform poorly. They recommended the k model. They also stressed the need for more
reliable hydrodynamic and heat transfer data for turbulent impinging plane jets. According
to them the data must clearly specify if the secondary Nu peak is caused by relaminarization
or boundary layer transition or impingement of vortex rings into boundary layer. Angioletti
et al. [7] have tested k , SST, RNG and RSM models. It is found that none performed
good. Park & Sung [152] have tested k f model and Merci et al. [142] has tested the
nonlinear k model. Parneix et al. [154] used 2 -f model which performed better than
k models, but it underestimates the local heat transfer coefficient at stagnation region.
Jia et al. [101] performed the numerical assessment of two low Re models and three 2 -f
models but all performed poorly as far as heat transfer is concerned. Polat et al. pointed out
that the only way to verify the validity of the results of a numerical simulation, in the case
of the turbulent impinging jets is to compare them to the results of experiments where the
boundary conditions of the flow domain are same as those of numerical study. Therefore
the choice of boundary conditions are very important in this case.
An interesting comparison of numerical computations for this case is given by Zucker-
mann & Lior. This is reproduced here in table 1.3.
1.4.2 Heat Transfer due to Excited Jet Impingement

Jet impingement is a promising approach for the heat and mass transfer enhancement in
many technological processes like electronic cooling, metal annealing, glass tempering,
drying, and gas turbine internal blade cooling. It is known that through jet excitation or
pulsation the jet behavior can be significantly altered. Pulsation affect not only the flow
structures but also the entrainment. In the case of limited mass flow rate, the possibility of
jet heat transfer enhancement is very important. In the case of excitation of the jet, not only
the frequency of excitation is important but the amplitude of the excitation is also important.
Further, if the jet is impinging, the target to wall distance (H/D) is also playing a significant
role.
The pioneering work of experimental investigation of an excited jet impingement has
been done by Nevins and Ball [146]. They investigated the heat transfer by pulsating jet
impingement. They found no significant heat transfer enhancement due to excitation. Liu
& Sullivan [128] examined the role of excitation on heat transfer at target to wall distances
Table 1.3: Comparison of Turbulence Models for Jet Impingement Case
Turbulence Model Computational Cost Nu-Prediction/ Prediction of

(Time) Anticipated % error Secondary peaks
k- Low Poor/ 15 - 60 Poor
**** * *
k- Low-medium Poor-fair/ 10- 30 Fair
**** ** **
Realizable k- Low Poor-fair/ 15 - 30 Poor-fair
**** ** **
Algebric Stress model Low Poor-fair/ 10-30 Poor
**** ** *
Reynolds Stress model Moderate-high Poor/ 25 -100 Fair
** * **
SST Low-moderate Good/ 20- 40 Fair
*** *** **
2 f Moderate Excellent/ 2-30 Excellent
*** **** ****
DNS/LES Extremely high Good-excellent Good-excellent
* **** ****
Source: Zuckerman and Lior [204]

Note: Fair indicates the incorrect prediction of location or magnitude.
* = Undesirable Model characteristics, **** = Excellent Model Characteristics
smaller than one. They found that when the jet is forced close to subharmonics of natural
frequency a stable vortex pairing can be obtained, which give rise to strong counter rotating
secondary vortices which leads than to reduction in the heat transfer. However, when the jet
is forced close to natural frequency, the intermittent vortex pairing is obtained, which gives
rise to less organized eddy-like structures which contain high turbulence and subsequently
improve heat transfer [128]. Gau et al. [53] have found that the acoustic excitation is im-
portant when the nozzle-to-plate distance is less than 4. For H/D > 8 there is no appreciable
effect of acoustic excitation on the jet. They have found that the jet excitation at inherent
frequencies (fundamental, subharmonic, second subharmonic) can enhance heat transfer.
Hwang & Cho [93] investigated the role of excitation for heat transfer enhancement. They
found that for jet excitations at the fundamental frequency of the initial vortex generation,
the potential core is small and the jets turbulence level is high. The subharmonic of the
20 1. Introduction
fundamental frequency excitation promotes vortex pairing. It is found that for H/D < 4, an
increase in heat transfer is possible in the wall jet zone, but overall this increase is small. On
the other hand, for larger H/D ratios (H/D 12 or higher) heat transfer decreases. Vejrazka
et al. [184] investigated the Reynolds number at 10000 with a small amplitude of excitation.
No significant change in the Nusselt number distribution was observed.
Sheriff & Zumbrunnen [173] found that there is no significant Nusselt number change
if the amplitude of pulsation is less than 40 % of the average inlet flow velocity. However,
at higher pulse amplitudes a reduction in Nusselt number is found. Mladin & Zumbrunnen
[137] found that the deciding factors for heat transfer enhancement are not the frequency and
amplitude independent of each other. At one particular Reynolds number and H/D ratio the
product of Strouhal number and amplitude is important. For higher values enhancement in
heat transfer is expected and for lower values reduction in heat transfer is anticipated. They
found that the pulses of small frequency and high amplitude may diminish heat transfer
in gas turbine cooling due to non-linear dynamics. Hwang & Cho [93] have found that
both frequency and amplitude affect the resulting heat transfer. Sailor et al. [167] had
investigated the role of pulsating impinging jet flows for heat transfer enhancement. The
inlet velocity with rectangular pulse form is found to enhance heat transfer. Azevedo et al.
[11] investigated heat transfer due to impinging jet. They found that the pulsations reduce
the heat transfer, which according to them is due to relatively low-magnitude, small-scale
turbulent fluctuations superimposed on the instantaneous periodic flow. Camci & Herr [21]
have investigated the oscillating planar turbulent jet with heat transfer. They found that the
oscillation in the jet, compared with stationary jet case, can improve heat transfer up to
70 %.
There are very few numerical studies devoted to study the effects of velocity field ex-
citation on heat transfer of an impinging jet. Chung et al. [23] have performed DNS simu-
lations of a laminar impinging jet at Reynolds numbers of 5300, 500, and 1000. Hofmann
[86] investigated the effects of excitation on impinging jets with heat transfer using models
based on Reynolds Averaged Navier-Stokes equations. The RANS based models are found
to be inadequate for the simulation of this complex flow. Jiang et al. [104] have used LES to
investigate the rectangular jet behavior. They found that strong vortical development is ex-
hibited when the forced frequency is almost the same as the characteristic frequency. Jiang
et al. [105] have investigated the effects of perturbations on the Nusselt number of turbu-
lent impinging jets using Direct Numerical Simulations (DNS). They found no significant
increase in heat transfer, as the near wall flow structures were unaffected by the excitation
of the jet inflow profile. The amplitude of excitation was 4%.
Apparently, in most of the experimental studies the jet issues from converging noz-
zles. In such cases, normally the issuing jet has a flat velocity profile with a thin boundary
layer near the nozzle lip [184]. In this scenario the most important jet mode is the shear
layer mode. The most amplified frequency in the jet scaled with the initial momentum
layer thickness and jet exit velocity, gives the Strouhal number for this mode (see eq. 1.1).
However, if the jet emerges in a fully developed flow state from the pipe, the momentum
thickness is of the order of the jet radius and the jet does not exhibit the shear-layer mode,
instead it will attain the preferred mode of the axisymmetric jet (see section 1.3.1).
Few researchers have investigated passive jet excitation. Herwig et al. [96] investi-
gated the effects of unsteadiness on the impinging jet heat transfer. The unsteadiness was
introduced through a circular ring introducing the periodic vortex shedding in the flow,
precessing effect and alternating flow passages. They found that the heat transfer can be
augmented only for the Karman jet nozzle in which additional unsteadiness and turbulence
is introduced without fundamentally changing the flow structure of the jet.
1.4.3 Heat Transfer due to Swirling Jet Impingement
Swirling jets are used in a wide variety of engineering applications. Because of high entrain-
ment rates, swirling jets offer good mixing characteristics, which makes them desirable in
chemical reactors, cyclone separators and mixing devices. The use of swirling jets in burn-
ers is known for flame stabilization. In the swirling jet the tangential velocity component
plays an important role. The azimuthal velocity introduces additional instabilities in the jet.
This makes the swirling jet more complex than non-swirling impinging jets. The swirling
jets are consist of complex flow features like, spiral type flow instabilities, vortex break-
down and free shear layer. A schematic representation of the swirling jet impingement is
shown in figure 1.11.
The swirl in the incompressible jet is quantified through a swirl number defined by the
following equation [72]:
1 0R r2 uax u dr
R
G
S= = (1.9)
RGax R 0R ru2ax dr
R
where G and Gax are the tangential and axial momentum flux, u and uax are the tangential
and axial velocities of the swirling jet and R is the radius of the jet at outlet.
The tangential velocity distribution in real swirling flows is intermediate between a
forced vortex flow and a free vortex flow. Swirl with constant angular velocity is called
forced vortex flow or solid body rotation. It can be described as:
u = r (1.10)
where is the angular velocity. The tangential velocity for free vortex flows is given by:
u = C/r (1.11)
22 1. Introduction
Figure 1.11: Schematic representation of a swirling jet impingement.
A real swirling flow normally has a core region behaving like a solid body rotation and a
region away from the center behaving like a free vortex. This combination is called Rank-
ine vortex. The effects of swirl in impinging jet flow on heat transfer has been investigated
experimentally by numerous researchers like [188],[148],[172]. There are two diverse opin-
ions in general about the possibility of heat transfer enhancement. Ward & Mahmood [188]
have found that heat transfer is reduced because of the swirl in the jet. Nozaki [148] has
proposed that there are two modes of swirling jets, one is called heat transfer enhancement
mode and the other one heat transfer suppression mode. The main role is played by the
dynamics of recirculation zones produced due to swirling jet impingement. Senda et al.
[172] investigated the impingement heat transfer with swirl of 0.45 and 0.22 at Re=8100.
It is found that for H/D less than and greater than one, the heat transfer at the target wall is
reduced. However near H/D=1 the heat transfer can be improved. Yan and Saniei [195] in-
vestigated experimentally the swirling jet impingement at Re=23000 and S=0.2, 0.35, 0.47
and H/D from 2 to 9. Recently, Abrantes and Azevedo [5] have conducted an experimen-
tal investigation of heat transfer due to swirling jet impingement at a Reynolds number of
21000. The swirl number of the jet was 0.5. The purpose of their study was to investigate
the effects of jet outlet-to-target wall distance on the swirling jet heat transfer. They found
that for the small H/D cases the swirl can increase heat transfer. Further, Lee et al. [123] and
Yan et al. [195] have found in separate experimental investigations, very different Nusselt
number distribution for heat transfer due to swirling jet at Reynold number of 23000.
There are very few numerical investigations of this highly complex phenomenon. Yan
1.5. Goal of the Thesis 23
and Kalvakota [196] have used a Reynolds Averaged Navier-Stokes equations (RANS)
based turbulence model for the investigation of the swirling jet impingement at Re=23000
with swirl numbers (S) 0.2, 0.35, 0.47. They found that swirl affects the heat transfer only
till r/D=2. Also, the normal velocity relaxation model (2 -f model) is found to perform
better than others. Time dependent computations can give more insight than steady state
computations in complex flow scenario like these. Therefore, the use of Direct Numerical
Simulations (DNS) or Large Eddy Simulations (LES) is justified. Hllqvist [79] has con-
ducted Large Eddy Simulations of swirling jets at two different swirl numbers (S=0.5 and
1). He found that a higher swirl number reduces the wall heat transfer. However, the LES
results were presented without experimental validation.
1.5 Goal of the Thesis

Large Eddy Simulation is not a tool which can be invoked at any stage of design but it should
be used for understanding the complex flows. The LES gives very detailed information not
only about the primitive variables but informs us about the flow structures and turbulence
spectra. It can help in understanding the complex nature of turbulence and flow physics.
The information attained via Large Eddy Simulation can also be used to improve the
RANS models. In this thesis, the coefficients of a scalar-flux model (RANS based) has
been investigated on the basis of knowledge attained via Large Eddy Simulation of imping-
ing jets. In the past, some researchers had investigated the model coefficients using data
obtained from Direct Numerical Simulation [199; 2]. However, the data used was of the
problems having no separation, stagnation or streamline curvature effects. Therefore, there
is still a need to reassess models.
Further, the effects of swirl, velocity field excitation and temperature field excitation on
heat transfer is investigated. Especially interesting is the temperature field excitation of an
impinging jet which has never been done before. This gives a novel approach for cooling
and drying technology.
The goals of the thesis are :
To investigate the complex flow and heat transfer mechanism of an impinging jet.
To investigate the effects of jets inlet velocity field excitation on heat transfer by an
impinging jet.
To investigate the effects of jets inlet temperature field excitation on heat transfer by
an impinging jet.
To investigate the effects of swirl on heat transfer by an impinging jet.
To evaluate the coefficients appropriate for the functional relationship in the scalar
flux models for an impinging jet flow.
24 1. Introduction
2. TURBULENCE: PHENOMENON AND PREDICTION
2.1 Govering Equations of Fluid Motion

The set of equations describing the transport of mass, momentum & energy in a fluid are
the following.
Mass conservation (Continuity equation): The differential form of the mass conserva-
tion (Continuity) equation is:
(ui )
+ =0 (2.1)
t xi
Momentum Conservation equation: The differential form of the momentum conserva-
tion (Navier-Stokes Equation) for incompressible flows (assuming Newtonian fluid) is:

(ui ) (ui u j ) p ui u j
+ = + + + Fi (2.2)
t x j xi x j x j xi
where Fi are the body forces.
Energy Equation: The differential form of the energy conservation equation without
viscous dissipation term, is:

(c p T ) (c p ui T ) (c p T )
+ = e + ST (2.3)
t x j x j x j
where e is the appropriate molecular diffusivity and ST is the rate of creation of the prop-
erty per unit volume.
Generic form of conservation equations The above equations of mass, momentum &
energy can be written in generic differential form as:

+ div(ui ) = div( grad) + S (2.4)
t
26 2. Turbulence: Phenomenon and Prediction
The represents an arbitrary conservation quantity like velocity, temperature etc.

represents its transport coefficient. S is the sum of all source and or loss terms not included
in the convective and diffusive terms.
2.2 Turbulence Scales and Spectra

Turbulence is a three-dimensional time dependent motion in which vortex stretching causes
velocity fluctuations. Turbulence is one of the most complex phenomena in nature. Some
of the reasons for the complexity are [182; 81]:
I- Intrinsic spatio-temporal randomness: Turbulent flows are extremely sensitive to
small perturbations/disturbances due to which turbulent flows inherently have no re-
peatability. However, the statistical properties (not only some means, but almost all
statistical properties) of turbulent flows are insensitive to disturbances.
II- Intermittency: Turbulence can interact with non-turbulent fluid flow and can appear
intermittently in time at certain location.
III- Wide range of non-locally interacting scales: Turbulent flows are comprised of
wide range of scales ranging from hundreds of kilometers to parts of a millimeter.
IV- Continuous self production of vorticity and three-dimensionality: In turbulence
an important role is played by vortex stretching, elongation, spinning, breakup, co-
alescing and pairing. Turbulence is a three-dimensional phenomenon which creates
and maintains the turbulence vorticity. Note that a two-dimensional vortex cannot be
stretched, hence two-dimensional turbulence does not exist in nature.
V- Strong diffusion: A source of energy is required to maintain turbulence (gradients of
mean velocity, buoyancy, or other external forces). The energy supply mostly occur
at large scales and its dissipation occur at small scales. Turbulent fluctuations cause
the stretching and distortion of the fluid elements (containing a lump of property like
hot spots of heat, species, momentum, vorticity) until the increase in surface area and
the property gradient enables molecular effects to act efficiently. Thus the transport of
momentum, heat, species and mixing is enhanced by several orders of magnitude.
Although turbulence is a very complex flow, fortunately, it can be computed using the
purely deterministic equations (Navier-Stokes equations). However, the solution is very
sensitive to initial or boundary conditions. The perturbations or disturbances which come
from initial conditions, boundary conditions and external noise can significantly alter the
outcome. To study the turbulent flow field, the field variable are treated as random vari-
able and useful information about the behavior can be obtained by the standard statistical
analysis.
2.2. Turbulence Scales and Spectra 27
2.2.1 Single Point Statistics

A single point turbulence data recorded in time can give a nice description of this complex
phenomenon.
Correlations
0
The time history of a fluctuating function u (t) can be studied by finding the correlation
between the values at two different times. A time correlation function is defined as:
0 0
ui (t)u j (t + t)
Rt (t,t + t) = q q (2.5)
0 0
(ui (t))2 (u j (t + t))2
The corresponding time scale can be determined from the known time correlation function:
1
Z
ti j = Rt (t,t + t)d(t) (2.6)
2
The integral turbulent time scale can be interpreted as an averaged inverse rotational
frequency of the typical big eddy appearing in the spatial location x. Also, it gives the
0
measure of the time interval over which u (t) remembers its history.
For isotropic turbulence:
tii
t= (2.7)
3
Closely, connected to this time is the integral length scale (`), which represent the large
scales present in the flow field. Also, the integral length scale (`) can be interpreted as the
length scale from which point on the velocity fluctuation are predominantly uncorrelated.
The integral length scale can be computed with the help of the two-point spatial correlation
function for statistically steady (time independent) turbulence as:
0 0
`
ui (x)u j (x + x)
R (x, x + x) = q q (2.8)
0 0
(ui (x))2 (u j (x + x))2
The size of energy containing eddies depends on the geometry of a spatial domain and on
the local intensity of turbulence. This size can be related to the integral turbulent length
scale as:
1
Z
`i j = R` (x, x + x)d(x) (2.9)
2
Figure 2.1: The turbulence energy spectrum.
For isotropic turbulence:
`ii
`= (2.10)
3
Energy Spectrum
In practice usually a one-dimensional energy density spectrum is utilized for the study of
turbulent flows. It is nearly impossible to directly measure the three-dimensional energy
k
spectra E( ), it is relatively simple to measure the one dimensional energy spectrum.
A Fourier transformation of the autocorrelation function is the energy spectrum, defined
as:
Z
E( f ) = ei(2 f )t [Rt (t,t + t)]d(t) (2.11)

Compared to laminar flow, the turbulent flow comprises eddies (fluid particles moving
together) of different sizes and scales. These eddies significantly affect the overall flow
field. The large eddies have high energy and from them this energy is transfered to small
eddies. The energy of small scales can be dissipated due to viscous action. In figure 2.1
a schematic diagram of a typical energy spectrum is shown. The range between energetic
2.3. Turbulence Modeling and Simulation 29
eddies and small scale dissipative eddies is called inertial subrange. The wave number , k
which is reciprocal of the length scale, shows the size of the eddies. A representative length
scale for large eddies is called integral length scale (`) and for small scales it is called
Kolmogorov length scale ().
2.2.2 Averaging of Turbulent Flow

Time Averaging
An ensemble averaging is the most general type of Reynolds averaging which is suitable
for flows that decay in time. It is defined as:
N
1
< (xi ,t) >= lim (xi ,t) (2.12)
N N n=1
Once the flow becomes statistical stationary (statistics become independent of time) it can
be time averaged which is defined as:
Z t/2
(xi ) = lim (xi ,t)dt (2.13)
t t/2
According to the Reynolds decomposition of turbulence quantities, we have:

0
= + (2.14)
or simply:
0
= + (2.15)
2.3 Turbulence Modeling and Simulation

2.3.1 Direct Numerical Simulation
The direct measurement of turbulence by experimental techniques has tremendously helped
in our understanding of the phenomenon. However, it is not a simple matter to perform
fully satisfying measurements in turbulent flow. The measurement sensors are beset with
difficulties like resolution, reliability, precision and cost. Moreover the variables which can
be measured are very limited.
In spite of the difficulties, in the last 100 years, data has been accumulated and a large
body of knowledge is available for turbulent flow. The continuous advances in data acqui-
sition techniques is also helping in reducing the above mentioned problems.
In the early 1980s the advances in computational techniques and resources permitted
the numerical solution of Navier-Stokes Equations. The full field simulation called Direct
Numerical Simulation (DNS) has become an important source of information and insight
for turbulence. DNS helps in understanding turbulence physics and its modeling. This helps
in understanding the time evolution of the turbulence field, life histories of eddies and multi
point correlations. DNS is now applied for flows over a flat plate, pipe flow, jet and wakes.
The computational requirement for DNS increases with Reynolds number. This makes
the applicability of DNS for industrial problems limited. The restrictions come from the
fact that we want to resolve the scales having order of the Kolmogorov scale ().
1/4
3

= (2.16)

In order to resolve the turbulent flow field for DNS computations, different resolution
requirements are suggested in literature. According to Leonard [126], to simulate reliably
the dissipation of turbulence energy, the grid spacing must be somewhat smaller than the
length scale corresponding to the peak in dissipation spectrum and mean grid spacing cor-
responds to 3. According to Ferziger [49], the grid must not be larger than the viscously
determined . According to Moin and Mahesh [144], the smallest resolved scale is required
to be of O() not equal to .
2.3.2 Large Eddy Simulation
The rationale of LES is that the large scale structures in turbulent flow produced directly
by the instability of the mean flow (shear or buoyancy effects) should be simulated directly
as they are problem dependent, anisotropic and play a major role in transport of mass,
momentum and energy. On the other hand, small scale structures can be modeled with less
effect on the prediction quality.
The technique was originally proposed by the meteorologists Lilly (1962) and Smagorin-
sky (1963) [175]. They introduced the idea of relating the turbulent quantity, eddy viscosity,
to the large scale dynamics of the atmospheric flows. Deardorff [34; 35] tested the technique
for channel flow and planetary flows in the early 70s. The procedure is further improved
by Schumann [171] who used the idea of volume averaging over the computational cell for
the results refinement. Later, LES technique is developed at different levels and institu-
tions like Grtzbach improved the works of Schumann at Karlsruhe University, Germany;
Leonard introduced the concept of stress decomposition at Stanford University, USA; new
approaches were tested at Queens Mary College, UK and new subgrid models were devel-
oped at National Polytechnic Institute, Grenoble, France. In spite of these developments the
technique was used by few researchers mainly due to the large computational requirements.
Figure 2.2: Computer Simulation of Turbulence. The wave number corresponding to in-
tegral scales is defined as 2 /` and the wave number corresponding to Kol-
mogorov scales is defined as 2/.
In the early 90s, the computational recourses become accessible to more researchers and
LES was used for the simulation of more challenging flows.
As mentioned above, the philosophy of LES is to explicitly resolve the large-scale mo-
tion and to model the small scales. To distinguish the small scales eddies from resolved
large scales, a spatial filtering process is applied to the Navier-Stokes equations.
In case of the implicit LES approach, the only reference to the filter is through the filter
width , which is based on grid. In this thesis only implicit filtering is used.
Filtering Operation
In LES each quantity is decomposed as:
= + (2.17)
This is analogous to the Reynolds decomposition of field quantities into mean () and
0
fluctuating quantities ( ). However, here a caret () denotes the Large scale (resolved)
component which is time dependent and a tilde () denotes small-scale (unresolved) or
subgrid components. This spatial separation is made by low-pass filter:
Z Z Z
(x,t) = G(x, x0 )(x0 ,t)d 3 x0 (2.18)
where can be u, p and T etc, G is a normalized weighting function or filter. Some
commonly used filters are Top-hat filter, Gaussian filter and Sharp Fourier cut-off filter.
The Gaussian filter is smooth both in physical and Fourier space and is defined as:
r 0
0 6 6(x x )2
G(x, x ) = exp( ) (2.19)
2 2
The Cut-off filter is defined in Fourier space and eliminates all fourier coefficients
belonging to wave numbers above a particular cut-off. It is natural to use this filter in
spectral methods as it leaves more energy in the large scales field than other filters. It is
difficult to apply it in homogeneous flows. The cut-off filter is defined as:
0

(xx )
0 1 sin( )
G(x, x ) = 0 (2.20)
(x x )
The Top-hat filter or Box filter is simply an average over a rectangular region. It is
the most natural choice when finite volume or finite difference method is used. It is defined
as:
0 1/ |xx| 12
G(x, x ) = 0 |xx| > 12
(2.21)
here, is the characteristic filter width.
The box filter does not remove filter energy at the wave numbers beyond the highest
resolvable in a satisfactory manner. This energy is aliased into lower wave number compo-
nents and gives falsified amplitude to the resolvable scale waves.
The schematic representation of top hat filter in energy spectrum is shown as red line
in figure 2.3. It is interesting to see the application of the top-hat filter in the wave-number
space. When the top-hat filter is used it will not sharply divide the wave number space into
large and small scales, as in case of sharp Fourier cut off filter (dashed line).
In the thesis only the Top-hat filter is employed for Large Eddy Simulations.
Figure 2.3: Energy spectrum showing the resolved & subfilter portion in LES computation.
The sharp Fourier cut off filter is shown by dashed line. The red line shows the
top hat filter. Part of the real spectrum is lost due to numerical errors. [20]
Filtering Operations
By low pass filtering the flow is divided into resolved and subfilter-scale motions. The
subfilter-scale motion is modeled by the subgrid model. The following filter operations are
used:
Linearity
f + g = fb+ gb
[ (2.22)
Conservation of constant
Cb = C (2.23)
Commutation
cf fb
= (2.24)
x x
Size of Filter-Width
Researchers have tested different filter widths in numerical simulations, for example:
Schumanns Filter: p
3
= x1 x2 x3 (2.25)
where the x1 , x2 and x3 are mesh sizes along three directions.
Bardina Filter: Bardina et al. (1980) showed that in homogeneous turbulent flows the
better filter definition is:
s
x12 + x22 + x32
= (2.26)
3
Also in LES it is good enough to select the filter as large as the grid spacing .Therefore,
the smallest filter width would be proportional to grid spacing. All Large Eddy Simulations
in the thesis were conducted through this approach.
LES: Governing Equations
By filtering the continuity, Navier-Stokes & energy equations (2.1,2.2, 2.3) and neglecting
body forces and subgrid density fluctuations, we have:

iu j )
(ubi ) (ud pb ubi ubj
+ = + + (2.27)
t x j xi x j x j xi
In the above equations the convective terms are nonlinear and pose a problem for filtering.
The stresses in the flow field can be written as combination of resolved and un-resolved
stresses:
sgs
udi u j = u
bi ubj + i j (2.28)
Considering:
ui = ubi + uei (2.29)
one has:
iu j = u
ud i ubj + u i uej + ui ubj + ui uej (2.30)
d
b d
b d
e d
e
This can be written as:
sgs
iu j u
i j = (ud bi ubj ) = (Li j +Cri j + Ri j ) (2.31)
where
Li j = ubi ubj u
bi ubj (2.32)
d
is called Leonard -term.

and
Cri j = ub
di uej + u
d
e i ubj (2.33)
is called Cross-term
and
Ri j = ue
di uej (2.34)
is called subgrid or unresolved Reynolds stress.
Li j describes the fine-structure stresses (small-scale turbulence) which arise due to

the product of large scales, also called outscatter terms.
Cri j describes the fine structure stresses which arise due to interaction of resolved
scales & non-resolved scales.
Ri j describes the interaction between small scale eddies which causes the formation
of the large scale eddy, also called true subgrid scale term.
The above mentioned decomposition is now only mentioned for historical reasons. Pre-
viously, many researchers modeled the subgrid stresses separately. However in the recent
past it has become common to model the entire subgrid scale stress as a single unit. In the
context of LES the subgrid scale stress is in fact the large scale momentum flux caused by
the action of small or unresolved scales. The subgrid is not related to grid size as name
might incorrectly suggest, but related to the average of the small scale field.
One can write the mass and momentum equations as:
(ubi )
+ =0 (2.35)
t xi
sgs
(ubi ) (ubi ubj ) pb ubi ubj i j
+ = + + (2.36)
t x j xi x j x j xi x j
Similarly, one can write the energy equation as:
" # sgs
(c p Tb) (c p ubj Tb) Tb q j
+ = c p (2.37)
t x j x j x j x j
For the LES computations, there are broadly following approaches:
A LES computation with no subgrid model (quasi-DNS).
A LES computation with subgrid model.

A monotonically integrated LES (MILES), in which, numerical diffusion is used in
place of the subgrid-scale model.
In this thesis, a subgrid model is used.
Subgrid Models
The model used to approximate the subgrid-scale term is called subgrid-scale model. The
large scales in the flow depend on the geometry of the problem and are produced by the
mean flow, so they are inhomogeneous, anisotropic, diffusive and not universal. In contrast,
the small scales are homogeneous, isotropic, dissipative, and universal. It is easier to model
only small scales. If the trace of the tensor i j is subtracted the subgrid scale Reynolds
stress to be modeled is:
sgs 1 sgs
i j = aij + i j kk (2.38)
|{z} 3
| {z }
anisotropic isotropic
This is the subgrid scale stress term which represents the effects of unresolved small scales
on resolved large scales.
We can categorize the models used in LES as:
A- Gradient - diffusion models

B- Higher-order models
C- Multilevel models / Scale similarity or Dynamic models
D- Non-conventional models
In this thesis, a model of category C is used for computations. The model originally
based on the model of category A, therefore, the discussion here is limited to them only.
Gradient Diffusion Model

A large class of subgrid closure models are based on the gradient-diffusion hypothesis. It
can be expressed as a relation between anisotropic stress and the (large-scale )strain rate
tensor.

sgs 1 sgs 1 c
i j i j kk = 2s Scij S
i j kk (2.39)
3 3
Where s is the subgrid viscosity, which needs to be described by appropriate models. If

sgs
the subgrid scale energy 21 ue
dk uek is indicated by Es the second term in the above equation
can be expressed as:
1 sgs 2 sgs
i j kk = i j Es (2.40)
3 3
sgs
Substitute this in equation 2.39 and then in equation 2.36; by adding and subtracting 31 i j kk
we find a modified pressure with kinetic energy as:
2 sgs
Pb = pb + Es (2.41)
3
Also,
1 sgs
Pb = pb + kk (2.42)
3
Now the filtered Navier-Stokes equations can be written as:

(ubi ) (ubi ubi ) Pb ubi ubj
+ = + ( + s ) + (2.43)
t x j xi x j x j xi
Smagorinsky-Lilly Model
This model is based on the work of Smagorinsky (1963) [175] and Lilly (1966) [127], and
thus it is called Smagorinsky- Lilly or shortly Smagorinsky Model. It is a non-linear model
as subgrid viscosity is related to large-scale strain rate Sc
i j . We can express this model as:
s = c1 2 Sb (2.44)
Where q

Si j = Sb = 2Sc
i j Si j
c (2.45)
Hence, q
s = (cs )2 Sb = (cs )2 2Sc
i j Si j
c (2.46)
The cs is referred as Smagorinsky constant. The crucial issue is the choice of this constant.
Table 2.1 shows the disagreement on the optimum value of Smagorinsky constant. It can be
said that the optimum value is of the order 0.1. The limitations of the Smagorinsky model
are[139]:
1. The optimal model constant must be changed in different flows.
2. The model does not have correct limiting behavior near the wall.
3. The model does not vanish in a laminar flow and it is dissipative in laminar-turbulent
Table 2.1: Optimum value of Smagorinsky constant

Researchers Year Value Application
Lilly [127] 1966 0.23 Homogeneous-isotropic turbulence
Deardorff [36] 1970 0.17 Atmospheric boundary layer
Dardorff [34] 1970 0.1 Channel flow
Piomelli et al[155] 1989 0.1 Channel flow
Moin and Kim[138] 1981 0.065 Channel flow
Bardina at al.[19] 1983 0.2 and 0.09 Homogeneous-isotropic turbulence
Mason and Callen[143] 1986 0.2 Channel flow
transition zone.
4. The model does not account for the backscatter of energy from small scales to large
scales, which is important in the transition regime. In turbulence spectrum, on average,
the energy is transferred from large scales to small scales, however, at some locations the
reverse energy transfer is possible, as shown in figure 2.4. Piomelli et al. [156] have in-
vestigated large databases of channel flow DNS and shown that backscatter is happening at
large fraction of grid points in transitional flow especially in near wall region. The precise
amount of it depends on the flow Reynolds number. An accurate modeling of this phe-
nomenon is important for the transitional and non-equilibrium flows. However, for fully
developed channel flows it is not a crucial issue [156].
5. The model does not account for compressibility effects.
Dynamic Model
The dynamic model proposed by Germano et al. [64] is currently one of the best performing
models in LES. [15]. Germano et al. [64] proposed a procedure which allows the determina-
tion of the Smagorinsky constant dynamically. By this procedure the Smagorinsky constant
is no longer a constant, but - being more realistic - a function of space and time: Cs = Cs (x,
t). The dynamical computation of Cs is accomplished by defining a test filter with width b
which is larger than the grid filter width . The procedure is shown schematically in figure
2.5.
Applying the test filter to the filtered momentum equation, the subtest-scale stresses
test
i j can be written as:
sgs
i j = ubi ubj ud
iu j (2.47)
Figure 2.4: Schematic presentation of the energy cascade and backscatter process [15].
test bi ubj ubj ubj

ij = u (2.48)
sgs
Zi j = test
i j i j (2.49)
The last equation shows the relationship between resolved stresses Zi j and the subtest-scale
stress and Subgrid-scale stress. We can approximate the above stresses with the Smagorin-
sky model.
sgs 1 sgs
h i
sgs
i j i j kk = 2 Cg 2 Sb Sc
i j = 2Cg i j (2.50)
3
1 h i
b2 b c sgs
test test
i j i j kk = 2 Cg S Si j = 2Cg i j (2.51)
3
where Cg is square the value of Cs . Sb and Sc
i j are subgrid level and test level strain rate
tensors. For incompressible flows Skk is zero. Inserting the above equations in 2.49 relation
we have:
sgs
Zi j = 2Cg test
i j 2Cg i j (2.52)
Figure 2.5: Schematic representation of the Dynamic model [15].
sgs sgs
Approximating Cg i j Cg i j we can write:
sgs
Zi j = 2Cg (test
i j i j ) = 2Cg Mi j (2.53)
sgs
Mi j = (test
i j i j ) (2.54)
i j = 2Cg Mi j Si j
Zi j Sc (2.55)
c
The term Mi j in the above equation 2.55 can become zero. This solution is provided by
Lilly that square of error can be minimized through:
Er = Zi j 2Cg Mi j (2.56)
Er2
= 4(Zi j 2Cg Mi j )Mi j = 0 (2.57)
Cg
The model parameter to be computed is:

Zi j Mi j
Cg (x,t) = (2.58)
2Mi j Mi j
The model parameter can take negative values. A negative value indicates a negative eddy
viscosity and implies an energy transfer from small (subgrid) scales to the resolved scales
or backscatter, according to Piomelli et al. [156]. However, the negative Cg can also lead
to numerical instability. In order to avoid the instability, the present investigation uses the
relation:
Zi j Mi j
Cg (x,t) = max ,0 (2.59)
2Mi j Mi j
In this thesis, all LES computations are done using the Dynamic model proposed by
Germano et al.
Subgrid-Scale Heat Transfer Modeling

The gradient diffusion model is used for the subgrid-scale heat transfer modeling. Using
the gradient diffusion assumption for the Energy equation, we can write:
sgs Tb
qi j = s c p (2.60)
xi
Here, Tb is resolved temperature and s is the subgrid thermal diffusivity. By substitut-

ing equation 2.60 into 2.37 one has the following form of energy equation:
" #
(c p Tb) (c p ubj Tb) Tb
+ = (e + s )c p (2.61)
t x j x j x j
The subgrid thermal diffusivity s is expressed as:
s = s /Prs (2.62)
where Prs is the subgrid Prandtl number.

There is no fixed value for the subgrid Prandtl number. Different researchers have
used different values of subgrid Prandtl number depending on the grid quality, numerical
methodology and problem investigated. Some of the different values used by researchers as
found in scientific literature are listed in table below. As can be noted, there is no unanimous
value for the subgrid Prandtl number. Small values of Prs means that at small scales the
transfer of heat is much more intensive than that of momentum. The value of the subgrid
Prandtl number can also be dynamically adjusted. For the consideration of the backscatter
Table 2.2: Value of the Subgrid Prandtl Number (Prs )

Researchers Year Value
Deardorff [35] 1973 1
Deardorff [37] 1980 1
Antonopoulous-Domis [4] 1981 0.5
Kobayashi [106] 1985 0.85
Hunter et al. [91] 1988 0.25
Moeng and Wyngaard [145] 1988 0.4
Gao et al. [55] 1991 0.3
Metais and Lesieur [140] 1992 0.6
Bohnert and Ferziger [16] 1993 1
Hadziabdic [78] 2005 0.9
effect, the subgrid Prandtl number must be changed [70; 20]. Recently, Hadziabdic has
conducted the LES of an impinging jet with constant subgrid Prandtl number of 0.9.
In this thesis, the subgrid Prandtl number of 0.9 is used for the computations.
Near-Wall Treatment for LES
An important issue in LES computations is the near-wall treatment. There are two ap-
proaches used in LES computations. One can increase the near wall resolution or can use a
wall-model. In many cases good quality results can be obtained with out wall models. How-
ever, the issue becomes more important as the Reynolds number increases, which naturally
demand higher near wall resolution. Several wall models had been proposed for the Large
Eddy Simulations, details of which can be found in Ciofalo [25] and Breuer[20]. However
most of these models are based on the ideas taken from channel or duct flow simulations
and there performance for impinging jet flows is not investigated. In this thesis no near-wall
model is used.
2.3.3 RANS Models
The Reynolds Averaged Navier-Stokes equation based turbulence models are the most
widely used CFD tools. They are now a standard industrial simulation tool for the de-
sign and analysis of engineering systems. Around 100 years ago (1895), O. Reynolds gave
the idea of statistical time averaging. Through this technique, the Navier-Stokes equations
are averaged in time, which introduces new terms in the equations which needs to be mod-
eled. Over the years numerous investigators and researcher have proposed the turbulence-
models. These models predict the flow field accurately in many cases, however, in many
cases they also fail to predict the true flow physics. Therefore, the need to refine the models
is always present.
The time averaged Navier-Stokes equations are given as:

Ui P Ui U j 0 0
U j = + Fi + + ui u j (2.63)
x j xi x j x j xi
0 0
The ui u j term is an unknown quantity, which originates due to averaging a nonlinear
term. It is called Reynolds stresses.
Similarly the Reynolds averaged equations for a mean scalar field (neglecting source
term) is:

(T ) (U j T ) T 0 0
+ = ui T (2.64)
t x j x j x j
0 0
where is the thermal diffusivity of the fluid. The u j T term is called passive scalar-flux
and is an unknown quantity, which originates due to averaging a nonlinear term in transport
equation of total scalar field. This is analogous to Reynolds stresses.
The RANS models can be categorised as eddy viscosity based models (EVM) and non-
eddy viscosity based models, called second moment closure models. Both types of models
are tabulated in table 2.3. Here only the models which are generally present in commercial
codes are listed.
Table 2.3: Overview of Turbulence Computation- (RANS based)

Category Name Extra
PDEs
linear EVM Mixing length Model 0
One Equation Model 1
k Model 2
k Model 2
Shear Stress Transport Model (SST) 2
Normal velocity Relaxation Model 4
Second Moment Closure (SMC) Reynolds Stress Models 6
Eddy-viscosity based models: Most of the turbulence models available in commercial

codes are linear eddy viscosity (EVM) based models, which are based on the assumption
that shear stress is linearly proportional to the velocity gradients:

0 0 1 0 0 Ui U j 2 Uk
ui u j i j uk uk = t + i j (2.65)
3 x j xi 3 xk
In analogy with the eddy viscosity concept for Reynolds stresses, the passive scalar flux
is modeled through a simple gradient diffusion assumption:
0 0 t T
ui T = (2.66)
Prt xi
where t = c k2 /, k is the turbulent kinetic energy, is the dissipation rate of k and c

is dimensionless function. Prt is called turbulent Prandtl number.
Unfortunately the EVM models are drastically limited in correctly predicting flow
fields. The limitation stems mainly from the reason that turbulence is represented by a
scalar quantity i.e. kinetic energy. Moreover, it uses the eddy viscosity assumption which
rule out any turbulence anisotropy. It is assumed that the Reynolds stress tensor is instan-
taneously in equilibrium with mean rate of strain. This forces the alignment of the main
0 0
axes of ui u j and main rate of deformation Si j , which is correct in case of pure stress but
not for flows with mean vorticity. The heat transfer predictions are also affected as the level
0 0
of ui T depends crucially on the magnitude of Reynolds stresses. Therefore, in order to get
the satisfactory predictions for heat transfer in complex flows, more refined modeling of
Reynolds stresses is required.
EVM are not best suited for prediction of normal stress components. In turbulent im-
pinging jet flow computations excessive turbulence levels are generated which in turn gives
heat transfers totally wrong predictions. Several attempts have been made to correct this
problem of two-equation models for example, the 2 -f model, which corrects the EVM
weakness by defining the eddy viscosity with different time scales. In the case of high
strain rates the EVM may predict the negative normal stresses or excessively high Reynolds
shear stresses. The realizability limits implemented in EVM models constrain the excessive
stresses.
Algebraic stress models (ASM) are non-linear eddy viscosity models.
Second Moment Closure Models: The modeling of transport equations for turbulent
stresses is called Second Moment Closure (SMC) as these equations are the second moment
of momentum equations. Transport equation models for the Reynolds stress tensor and
passive scalar flux contains more physics than standard two equation models. The series of
higher order equations are generated but they too have more correlation terms and unknowns
which require approximate modeling. The experience has shown that RSM approaches do
not perform consistently better than simpler approaches like EVM in many cases including
impinging jets! The problem is that the Reynolds stresses contain information about the
directionality of the velocity fluctuations but not the orientation and elongation of turbulent
structures.
The stress field can be obtained from a system of inter coupled transport equations:

0 0 0 0
ui u j ui u j

+Uk = u0 u0 U j + u0 u0 Ui
t xk i k x j k x
k k
| {z }
Pi j
| {z }
C i j
0 0
0 0 0 1 0 0 0 0
ui u j
u u u + (p ui jk + p u j ik )
xk i j k xk (2.67)
| {z }
di
j
0 0 0 ! 0 0
p ui u j ui u j
+ + 2
x j xi xk xk
| {z } | {z }
i j i j
0 0
Dui u j
= Pi j + i j + di j i j (2.68)
Dt
where
Pi j denotes generation due to shear and force field
i j redistribution effects of pressure fluctuations
di j diffusive transport
i j viscous dissipation
Explicit Algebraic Reynolds Stress Model: is based on the simplification of RSM for
steady turbulence by assuming a local equilibrium of Reynolds stresses in the flow. By this
approach limitations of simple models are removed, like isotropy of the eddy viscosity.
Due to this simplification, the transport equation for the Reynolds stress tensor is re-
duced to a system of algebraic equations leading to so-called Implicit Algebraic Reynolds
Stress Models (IARSM). They are further reformulated by means of the invariant theory into
explicit nonlinear form yielding the so-called Explicit Algebraic Reynolds Stress Model
(EARSM). Thus the relation to the parents RSM is maintained and this in turn kept the
advantages of original RSM and keeping the production contribution closed. Also the es-
sential advantages of RSM are preserved and a cheap and accurate near wall treatment in
flow processes can be achieved.
As no EARSM type model is used, the detail discussion of these models is not included
here.
Scalar-flux Models: In analogy with transport equations for Reynolds stresses, a trans-
port equation for scalar flux vector can be written as:
0 0 0 0
Dui T T 0 Ui p T
Dt = u0i u0j 0
u jT
x j x j xi
| {z } | {z }
Ti P Ti
0 0
1 ui T
1+
Pr x j x j (2.69)
| {z }
Ti !
0 0 0 0
ui 0 0 T 0 0 0 T p
T ui + ui u j T + i j
x j x j x j
| {z }
DTi
where DTi is the molecular & turbulent diffusion, PTi is the production term, Ti is the
pressure scalar gradient correlation, Ti is the destruction rate tensor.
Explicit Algebraic Scalar Flux Models: The set of the equations is not closed and
needs relations to close it, especially important is ui T which needs to be modeled. Re-
searchers have also proposed the algebraic Scalar flux models in explicit forms called
EASFM. Some of the proposals are:
Daly and Harlow had (1970) suggested the relation:
0 0 k 0 0 T
ui T = C1 ui u j (2.70)
x j
where C1 = 0.3
Abe & Suga (2001) suggested the relation:
0 0
0 0 k 0 0 uk u j T
ui T = C2 ui uk (2.71)
k x j
where C2 = 0.6
Younis, Speziale & Clark (2005) suggested this model:
0 0 k2 T k 0 0 T k3 Ui T
ui T = C1 +C2 ui u j +C3 2
xi x j x j x j
k2

0 0 U j 0 0 Ui T
+C4 2 ui uk + u j uk (2.72)
xk xk x j
The coefficients are C1 = -0.0455, C2 = 0.373, C3 = -0.00373, C4 = -0.0235.

Some of scalar-flux models will be analyzed with the help of the LES data of an im-
pinging jet. A detailed analysis of these models is given in chapter 5.
3. NUMERICAL METHODOLOGY
In this chapter a basic introduction of the numerical code is given. In LES and DNS an
important role is played by the numerical discretization schemes. Especially in LES the
subgrid-scale model and the discretization scheme are linked. It is therefore suitable to look
briefly into various discretization approaches used in LES computations.
3.1 Comparison of Discretization Approaches

There are various numerical methods available for the solution of non-linear Partial Dif-
ferential Equations of fluid motion like for example, Spectral Methods, Finite Element
Methods, Finite Volume Methods and Finite Difference Methods. Spectral Methods are
a class of techniques used to numerically solve the partial differential equations by using
the Fast Fourier Transformation. The fact that certain physical processes can be described
by the wave interpretation have helped in development of this approach. The idea of the
presence of such basic waves can be exploited to develop the efficient algorithms for the
solutions to PDEs. The Spectral Methods take advantage of this fact by writing the solution
as its Fourier series and then substituting this series into the partial differential equations
to get a system of ordinary differential equations in the time-dependent coefficients of the
trigonometric terms in the series, and using a time-stepping method to solve those ODEs.
The Spectral Method and the Finite Element Method are closely related and built on
the same ideas; the main difference between them is that the Spectral Method approximates
the solution as linear combination of continuous functions that are generally non-zero over
the domain of solution (usually sinusoids or Chebyshev polynomials), while the Finite El-
ement Method approximates the solution as a linear combination of piecewise functions
that are non-zero on small sub-domains. Because of this, the Spectral Method takes on a
global approach while the Finite Element Method is a local approach. This is part of why
the Spectral Method work best when the solution is smooth. The Finite Element Method
produces equations for each element independently of all other elements. Only when the
equations are assembled into a global matrix are the interactions between elements taken
into account. In this method there is no restriction on the connection of the elements when
50 3. Numerical Methodology
the sides (or faces) of the elements are correctly aligned and have the same nodes for the
neighboring elements. The flexibility allows the modeling of a very complex geometry.
The Finite Difference method is based on Taylor series formulation. They offer many
advantages in regard to numerical formulation, data preparation and computation, but are
inappropriate for complex geometries.
In the Finite Volume Methods the physical law represented by PDE is not considered,
instead the conservation of specific physical quantity such as energy, mass or momentum
over a finite control volume is considered. Thus the control volume approach to derive the
discrete equations has the more physical meaning compared to the use of Taylor series. The
Finite Volume Method is based on the following steps [170]:
1. Decomposition of the problem domain into control volumes.
2. Formulation of integral balance equations for each control volume.
3. Approximation of integrals by numerical integration.
4. Approximation of function values and derivatives by interpolation with nodal values.
5. Assembling and solution of discrete algebraic system
The Finite Volume and the Finite Difference Methods generate numerical equations at
the reference point according to the values at neighboring grid points. The finite difference
method could be easily extended to multidimensional spatial domains if the chosen grid is
regular (the cells must look cuboid, in a topological sense). The grid indexing is simple
but some difficulties appear for the domain with a complex geometry. The Finite Volume
Method could also use irregular grids like the grids for the Finite Element Methods, but
keeps the simplicity of writing the equations like that for the finite difference method.
3.2 FASTEST- Basic Introduction

All computations, in this thesis, were performed with the CFD code FASTEST (Flow Anal-
ysis Solving Transport Equations Simulating Turbulence) based on a Finite Volume Nu-
merical Method for solving both three-dimensional filtered and Reynolds Averaged Navier-
Stokes equations on block structured, body fitted, nonorthogonal meshes. The code was
developed by INVENT Computing GmbH (Erlangen, Germany). A cell centered collocated
variable arrangement and cartesian vector and tensor components are used. The SIMPLE al-
gorithm is used for coupling the velocity and pressure fields. The convective transport of all
variables was discretized by a second-order central differencing scheme. Time discretiza-
tion is done by applying the second order implicit Crank Nicolson method. Parallelisation
based on domain decomposition in space using MPI (Message Passing Interface). In the
following section, the numerical methods implemented in FASTEST code are discussed
(for detailed discussion see ref. [45; 46; 98]).
3.3. Spatial Discretization 51
The main objective of Finite Volume discretization is to convert the partial differential
equations into algebraic form. The differential equations are integrated using the generic
transport equations (2.4) in the following way.
Z Z Z Z
() (ui )
dV + dV = ( )dV + S dV (3.1)
V t V xi V xi xi V
where is an arbitrary conserved quantity (like velocity, scalar quantity etc).
represents transport coefficient (like laminar or turbulent viscosity, diffusivity etc).
S is the sum of all sources (e.g. pressure gradient in momentum equation).
Using the Gauss divergence theorem, volume integrals are transformed into surface
integrals.
time Space
z }| { z }| {
Z Z Z
()
dV + ui ni d = S dV (3.2)
V t V xi V
| {z } | {z }
Flux source
where is a surface enclosing the Volume V and ni is the projection of unit vector normal
to the surface.
The computational domain is subdivided into a finite number of sufficiently small con-
trol volumes. The equation must be solved for each control volume. In FASTEST hexahe-
dral control volumes are used. The computational grid defines the control volume bound-
aries. To each control volume a computational node is assigned. The flux term contains
both the convective and diffusive fluxes.
3.3 Spatial Discretization

Consider a hexahedral control volume as shown in figure 3.1. The CV central point is
denoted by P has six neighboring control volumes indicated as E(east), W (west), N (north),
S (south), T (top), B (bottom), each sharing with CV a common face,(east side) ea, (west
side) we, (north side) no, (south side) so, (top side) to, (bottom side) bo, respectively. The
final goal of discretization (spatial & temporal) is to transform the equation 3.2 into an
algebraic form:
Figure 3.1: A hexahedral Control Volume with cell faces
A p p ANCV NCV = Sp (3.3)

| {z } |{z}
implicit part explicit part
The left hand side in the above equation is the implicit part where as the right hand side
in above equation is the explicit part. The subscript N CV indicates the six neighboring
control volumes. A p & ANCV denote the coefficients of dependent variable at point P.
The terms resulting from the discretization that can not be included into the implicit part of
equation 3.3 are treated explicitly and put into the source term S p on the RHS. The whole
surface of hexahedral control volume is divided into six faces that in general not necessarily
represent a plane, as four grid points defining such a face do not necessarily lie in the same
plane. But it is assumed that each face of the CV is a plane whose orientation is given by a
normal vector. The net flux through the whole CV boundaries is sum of fluxes over all the
six faces. To evaluate the surface integrals, the values of the integrand must be known on
the surface. As we are calculating the values at the center of control volume, the integrals
3.3. Spatial Discretization 53
will be approximated by a second order midpoint rule.
F Fi
Z Z
d = d (3.4)
i i
F
where i is a convective or diffusive flux vector in direction normal to the face i . The mid
point rule is used for volume integral approximation in equation 3.2 as well:
F Fp
Z
dV dV (3.5)
V
F
where p is the value of F in the CV center P and dV is the volume of CV. For the value
F p no interpolation is needed.
3.3.1 Convective Fluxes

The convective flux of the variable through the CV face represents the rate at which the
convective flux is transported into or out of the CV by the velocity of fluid. Applying the
mid point rule to convective fluxes we get:
Z
C
Fea = (ui ni )ea d (ui ni )ea dea mea ea (3.6)
ea
where mea is the mass-flux through the east face ea . As the momentum equation is a non-
linear equation, the mass flux is used from the previous iteration. The quantities on the
cell faces necessary for approximation are not available in collocated grids and thus need
to be interpolated. The choice of interpolation technique is very important. In first order
approximations the leading truncation error term includes the first order derivatives which
behaves as a numerical diffusion and damps the turbulence very strongly. For this reason
first order interpolations are not used in LES. For convective flux the interpolation used in
LES is of second order accuracy, called central differencing scheme. The ea is interpolated
linearly using the central differencing scheme:
ea E ea + P (1 ea ) (3.7)
where the interpolation factor ea is defined as:

ea P
ea = (3.8)
E P
Here, is the local coordinate.

3.3.2 Diffusive Fluxes

The approximation of the integrated diffusive fluxes by means of mid-point rule gives:
Z
D
Fea = ni d (3.9)
ea xi xi ea

The partial derivative xi
can be approximated for a Cartesian grid by using central differ-
encing scheme as:
E P
(3.10)
xi xE xP
This is a second order approximation for an equidistant grid.
3.3.3 Source Term

The source term is approximated by means of the midpoint rule. They are simply replaced
by the product of integrand value in the CV central point by the CV volume.
F FP
Z
S = dV dV (3.11)
V
F
The quantity p can be calculated as all variables are known at P. The source term in
momentum equations consist of the pressure gradient and the body forces. The body forces
are approximated as above. For the pressure gradient the Gauss Theorem is used:
Z Z Z
p
dV = pi ni d pi ni dc (3.12)
V xi c c
where, c runs through ea,we,no,so,to,bo. Using the mid-point rule it becomes:

Z
pi ni dc (pi ni )mc dc (3.13)
c c c
The value of pressure on the surface center is approximated by means of neighboring CVs.
The second order linear interpolation gives:
pea = e1 PE + e2 PP (3.14)
where the s are interpolation factors.
System of ODE: Once the time term in equation 3.2 is approximated by the mid-point
rule, the resulting system of ordinary differential equations are obtained:

(i,P )
dV = ai,P i,P + ai,c i,c + Si (3.15)
t c
3.4. Time Discretization 55
where i = 1,2,3..., N. The right hand side of this equation is obtained after rearranging the
approximated convective & diffusive fluxes and the source term. The index C ( = E, W, N,
S, T, B) denotes the neighboring -CVs, N is the number of CVs and S is the discretized
source term. Also due to conservation the coefficients are related as:
ai,P = ai,c (3.16)

c
3.4 Time Discretization

For unsteady problems the value of the flow variable (x,t) must be provided at the initial
time to . The spatial discretization of unsteady problem results into a system of ODEs. We
can write equation 3.15 as:

= L() (3.17)
t
We can simply integrate the equations as:
Z tn+1
n+1 n = L()dt (3.18)
tn
The RHS is approximated by also using the current unknown time level.
n+1 = g(n+1 , n , n1,..... ) (3.19)
The unsteady term from the transport equation can be approximated as:
Z
() ()
dV Vp (3.20)
V t t
For the finite approximation of the time derivative the Crank-Nicolson Method has been
used.
3.4.1 Crank-Nicolson Method

The Crank-Nicolson method is a two level approximation. The unknown variables at the
current time step and the known values from the previous time step are required. It is
unconditionally stable but gives oscillatory solutions if large time steps are used. However
for accuracy the Courant-Friedrich-Lewy number must be kept below one.

ui
CFL = t (3.21)
xi CV max
where maximum is taken over all CVs.
(P VP )n+1 (P VP )n
= ( An+1 n+1 n+1 n+1 n+1
Nb Nb + SP AP P )
t Nb
+(1 )( AnNb nNb + SPn AnP nP ) (3.22)
Nb
Here, t denotes the size of time step and superscript n+1 & n denotes the time step
level i.e. current and previous time steps. The equation is a combination of fully implicit
and fully explicit discretizations. If is 1, the scheme is Euler explicit and if is zero
scheme is Euler implicit. The of 0.5 is taken.
3.5 Pressure-Velocity Coupling

The determination of pressure is difficult in incompressible flows as the momentum equa-
tion contains only the pressure gradient. The continuity equation does not contain the pres-
sure term. There are many methods to determine the pressure. The pressure correction
methods solve the momentum equations sequentially for velocity field with available pres-
sure field. The pressure field is corrected in such a way that continuity is satisfied. This
procedure is repeated until divergence free velocity field is obtained.
The pressure-velocity coupling procedure used in FASTEST is derived from the SIM-
PLE algorithm (Semi-Implicit Method for Pressure-Linked Equations) of Patankar and
Spalding. It is further refined by Peric [50]. The procedure is designed for collocated
grids and schematic of the algorithm is drawn in figure 3.2. For each outer iteration n, the
mass fluxes and pressure values from the previous outer iteration n + 1 are used. The lin-
ear system of equations 3.3 for each velocity component ui can be written as follows. The
term at RHS in equation 3.3 i.e. SP term is decomposed into pressure gradient denoted as
(i pn1 )P and remaining part as Qn1
P .
n1 n1
AuP uni,P AuNb uni,Nb = Qn1
P (i p
n1
)P (3.23)
Nb
For a given pressure field p = pn1 a velocity field ui is obtained, giving ui,P :
" #
1 un1 n1
ui,P = un1 ANb ui,Nb + QP (i p )P (3.24)
AP Nb
The field ui satisfies the discretized momentum equations but not the continuity equation.
An unphysical mass source appearing in each CV is:

mP = Fea
Fwe
+ Fno
Fso + Fto Fbo

(3.25)
3.5. Pressure-Velocity Coupling 57
If the velocities on CV faces are simply interpolated between the CV central points the
well known problem of decoupling the pressure and velocity on the collocated grid leads to
pressure oscillations. In order to overcome this problem the interpolated CV face velocities
are modified so that the interpolated pressure gradients calculated directly at the CV face
are included into expressions.
u corr
i = ui + ui , p = p + pcorr (3.26)
Since both u
i and p
must satisfy momentum equation, following expression should be
satisfied. " #
1 un1 n1
ui,P = un1 ANb ui,Nb + QP (i p )P (3.27)
AP Nb
The neighboring velocity components are used. Subtract 3.27 & 3.23 to obtain the relation
between velocity and pressure correction.
ucorr
i,P = (i p
corr
)P (3.28)
For the CV face ea it follows:
ucorr
i,ea = (i p
corr
)ea (3.29)
In order to remove the mass source mP appearing due to the wrong velocity field, the
corrected mass fluxes must satisfy the equation.
mP = Fea
corr corr
Fwe corr
+ Fno corr
Fso + Ftocorr Fbo
corr
(3.30)
The substitution of expression 3.28 into above equation gives a linear equation of pressure
correction.

AP pcorr corr
P ANb pNb = mP (3.31)
Nb
The procedure in FASTEST consist of two steps. Predictor: Equation 3.23 is solved for
a given pressure and velocity field from previous outer iteration n-1 resulting in estimated
velocity field u . This estimated velocity (u ) and pressure fields (p =pn1 ) are used to
interpolate the estimated velocity values on the CV boundaries and to compute the mass
source m. Corrector: The mass source m is substituted into pressure correction equation
which is solved obtaining the pressure correction field pcorr . This pressure correction field
is added to pressure field from previous iteration.
p = p + P pcorr (3.32)
Here, the P is under relaxation factor in order to avoid instabilities due to big pressure
changes during the first iterations. The corrected pressure field p = pn is used in the next
n+1 step.
3.6 Solution of System of Equation

The solution of the system of equations through classical methods is computationally ex-
pensive. The reason is that almost all the classical iterative methods converge slowly on
finer grids. They reduce the high frequencies of the error efficiently, where as the lower
frequencies are smoothed slowly. Alternative methods are multigrid methods which are ap-
plicable to both linear and nonlinear system of equations. The linear multigrid method is
used in FASTEST which is explained as follows. The system of equations written in matrix
notation is:
A = b (3.33)
where A is the square coefficient matrix built from the coefficients of linear equations for
each CV, is a vector having values of variable in each CV and b is the vector containing
the terms on RHS of equation 3.3. The coefficient matrix is sparse i.e. most of the elements
are zero and non-zero elements lie on a small number of diagonals. Using the iterative
solution technique we can write:
g(o , A, b) (3.34)
where g() is an iterative solution technique and is approximate solution. since is not
exact solution it satisfies the equation only up to a residual re.
An = b re (3.35)
The convergence error is defined as:
en = n (3.36)
where is the converged solution. The purpose of the iteration procedure is to drive the
residual to zero and then process error also becomes zero. An iterative scheme for the linear
system, that should drive the residual to zero, can be written as:
Mn+1 = Nn + B (3.37)
Subtracting Mn from each side of equation 3.37 one can obtain:
M(n+1 n ) = B (M N)n (3.38)
or
Mn = rn (3.39)
Here n is the correction vector which is an approximation to the convergence error. For
rapid convergence the matrix M must be as close to A as possible. For this sake the strongly
3.6. Solution of System of Equation 59
implicit procedure (SIP), originally proposed by Stone is used [178]. In this method the
matrix M is chosen to be equal to the incomplete LU decomposition.
M = LU = A + N (3.40)
In the ILU decomposition, the same procedure as used for the standard LU factorization is
utilized. However, for each zero element of the original matrix A a corresponding element
of the lower triangular matrix L or the upper triangle matrix U is set to zero.
Usually, one has to define convergence criteria. The convergence criterion used is that
the normalised residual for the momentum, mass and energy equations fall below a specified
value. Iteration error arise from insufficient convergence criteria. Once the magnitude of
the error vector is less than the defined convergence criteria, iterations are stopped. For
convergence, error is brought down to at least four orders of magnitude.
Figure 3.2: Schematic representation of SIMPLE algorithm.

Naseem Uddin Thesis Part01

Uploaded by

Copyright:

Available Formats

Naseem Uddin Thesis Part01

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Naseem Uddin Thesis Part01

Uploaded by

Copyright:

Available Formats

Turbulence Modeling of Complex Flows in

A thesis accepted by the Faculty of Aerospace Engineering and Geodesy

M. Engg. Naseem Uddin

Committee chair: Prof. Dr.-Ing. habil. Bernhard Weigand

Institute of Aerospace Thermodynamics

lation on NEC SX-8 cluster.

The fully developed turbulent jet

Temperature field excitation

2. Turbulence: Phenomenon and Prediction . . . . . . . . . . . . . . . . . . . . . 25

4. Large Eddy Simulations & Results . . . . . . . . . . . . . . . . . . . . . . . 61

5. An Assessment of Scalar-flux Modeling for Complex Flows . . . . . . . . . . . 125

English Symbols (continued)

English Symbols (continued)

Greek Symbols (continued)

Superscript & other symbols

1.1 Motivation: Complex Flows

1.2 Impinging Jet: Applications

1.3 Impinging Jet as a Complex Flow Problem

Figure 1.8: Schematic representation of an jet impingement.

Figure 1.9: Structure of a free and an impinging jet.

0.01 St 0.018 (1.1)

1.3.2 Heat Transfer by an Impinging Jet

Maximum Stagnation Point Heat Transfer

number correlation as:

Secondary Peak in the Radial Distribution of the Nusselt Number

Nu1 /Nuo = 0.696Re0.047

Nu2 /Nuo = 0.400Re0.077

adiabatic wall temperature can be used as a reference temperature:

1.4 Research Overview

Table 1.1: Single Plane Jet - Experiments

Researchers Year D (mm) Re H/D

Table 1.2: Single Circular Jet - Experiments

Researchers Year D (mm) Re H/D

1.4.2 Heat Transfer due to Excited Jet Impingement

Table 1.3: Comparison of Turbulence Models for Jet Impingement Case

Turbulence Model Computational Cost Nu-Prediction/ Prediction of

Source: Zuckerman and Lior [204]

1.4.3 Heat Transfer due to Swirling Jet Impingement

Figure 1.11: Schematic representation of a swirling jet impingement.

1.5 Goal of the Thesis

2.1 Govering Equations of Fluid Motion

The represents an arbitrary conservation quantity like velocity, temperature etc.

2.2 Turbulence Scales and Spectra

2.2.1 Single Point Statistics

Figure 2.1: The turbulence energy spectrum.

For isotropic turbulence:

2.2.2 Averaging of Turbulent Flow

According to the Reynolds decomposition of turbulence quantities, we have:

2.3 Turbulence Modeling and Simulation

2.3.2 Large Eddy Simulation

LES: Governing Equations

is called Leonard -term.

Li j describes the fine-structure stresses (small-scale turbulence) which arise due to

A LES computation with subgrid model.

We can categorize the models used in LES as:

A- Gradient - diffusion models

Gradient Diffusion Model

Where s is the subgrid viscosity, which needs to be described by appropriate models. If

Table 2.1: Optimum value of Smagorinsky constant