This action might not be possible to undo. Are you sure you want to continue?
in
Thermo and Fluid Dynamics
Numerical Modelling of Diesel
Spray Injection, Turbulence
Interaction and Combustion
Fabian Peng K¨ arrholm
Department of Applied Mechanics
Chalmers University of Technology
G¨ oteborg, Sweden, 2008
Numerical Modelling of Diesel Spray Injection, Turbulence
Interaction and Combustion
Fabian Peng K¨ arrholm
The image on the cover of the thesis is an illustrative image of a lagrangian
spray, evaporation has been turned oﬀ to improve the visual experience.
c Fabian Peng K¨ arrholm, 2008
Doktorsavhandling vid Chalmers Tekniska H¨ ogskola
First edition
ISBN: 9789173851732
ISSN: 0346718X
Ny Serie: 2854
Department of Applied Mechanics
Chalmers University of Technology
SE412 96 G¨ oteborg
Sweden Telephone +46 31 772 10 00
Printed at Chalmers Reproservice
G¨ oteborg, Sweden 2008
I guess this is my dissertation
Homie this ... is basic, Welcome to Graduation
Good morning,
Good morning
On this day, we become legendary
 Kanye West
iii
Abstract
This thesis covers two main topics. The ﬁrst is numerical modelling of
cavitating diesel injector ﬂows, focusing on describing such ﬂows using a
singlephase cavitation model based on a barotropic equation of state together
with a homogenous equilibrium assumption. The second topic is
EulerLagrangian simulations of diesel sprays, focusing on attempts to reduce
the high grid/timestep dependencies in numerical simulations of diesel sprays.
In addition, the ability of two CFD codes to predict ﬂame liftoﬀ length and
ignition delay time, and the advection scheme’s inﬂuence on fuel distributions,
are considered.
A longterm goal was to develop a new atomization model based on calculated
ﬂows in injector nozzles, which did not have the drawback of requiring either
nonphysical parameters or information derived from speciﬁc experiments. To
validate the cavitation simulations, comparisons were made with experimental
data obtained at AVL. The experimental data (which are practically 2D)
provide information on velocity proﬁles and pressure contours. These data
were used to validate the code. However, since the code is not stable for
dieseltype pressures, no atomization model was developed.
The main part of the thesis describes how diesel sprays were simulated using
the discrete droplet model (DDM), in which the liquid is described by
Lagrangian coordinates and the vapour by an Eulerian approach. The
simulations have been used to investigate how the kε family of turbulence
models inﬂuence spray behaviour, and a simple but eﬃcient way to reduce the
dependency of the mesh resolution, by limiting the turbulence length scale in
the liquid core region, is proposed. This constraint is shown to have a positive
eﬀect on the spray behaviour, and to reduce both grid and timestep
dependencies.
In addition, the ignition delay time and ﬂame liftoﬀ lengths have been
investigated, since these two properties are believed to be important for
emissions formation. The simulations used a complex chemical mechanism
involving 83 species and 338 reactions. The eﬀects of the numerical scheme,
the turbulence model and physical parameters (like ambient temperature and
oxygen content) on these variables have also been investigated.
Keywords: Spray, Computational Fluid Dynamics, Diesel, Turbulence Model,
Cavitation, Atomization, Flame, Liftoﬀ, Ignition Delay, Numerical Scheme,
TVD, Combustion
iv
Acknowledgements
I thank Niklas Nordin for his supervision, help and guidance in OpenFOAM,
especially during the ﬁrst years of my phd studies. I am also in debt to Feng
Tao, thank you for your enthusiasm, discussions, suggestions, encouragement
and help during the latter part. I also owe a great deal to Rasmus Hemph for
the many discussions on the mysteries of OpenFOAM, CFD and autolyse.
Thanks also goes out to Ingemar Denbratt for his enthusiasm and support. I
thank AlfHugo Magnusson for providing experimental data for paper 1. I also
want to thank everyone else at the department, new and old. I have really
appreciated the lunch & coﬀee breaks as well as discussions on more
workrelated matters.
I also wish to thank some people abroad, who have helped me during my time
at Chalmers. Dr. Tommaso Lucchini who I am proud to have shared oﬃce
with, and who showed me the path to selflearning in OpenFOAM. Dr Hrvoje
Jasak for helping me with the appendix on RhieChow, the work you do for
OpenFOAM is priceless. Henry Weller for his assistance on the second paper,
and for answering to so many of my questions on cavitation and its
implementation in the code. I thank Dr. Ernst Winklhofer for sharing his
experiments on cavitating nozzles.
I thank the Combustion Engine Research Center (CERC) for providing
ﬁnancial support for this work on Diesel Sprays.
Last, I want to thank my family, Cecilia PK, Edwin P and Alicia K. Afterall,
you are the ones who makes everything worthwhile. I also thank myself for
hanging in there, and accomplishing this.
v
vi
Nomenclature
Roman Symbols
b Relaxation coeﬃcient
c Molar Concentration mol/m
3
c
l
Liquid speciﬁc heat at constant volume J/kgK
c
p,v
Fluid speciﬁc heat at constant volume J/kgK
C
1
,C
2
,C
3
,C
µ
Turbulence model constants
C
D
Drag coeﬃcient
C
d
Discharge coeﬃcient of nozzle
C
e
Constant of the D
2
law
Co Courant number
d Droplet diameter m
D Droplet Diameter m
D Mass Diﬀusion Coeﬃcient m
2
/s
E
a
Arrhenius Activation Energy J/mol
F Force on particle kgm/s
2
F
s
Spray momentum source term kgm/s
2
f Coeﬃcient in RNG turbulence model
f
heat
Correction factor for heat exchange due to mass transfer
g Gravitation acceleration vector m/s
2
h
v
Enthalpy of vapour J/kg
k Kinetic energy of turbulent ﬂuctuations m
2
/s
2
k
j
Arrhenius constant of reaction j
l
∞
Relaxation length scale m
l
t
Length scale of turbulent ﬂuctuations m
L
sgs
Maximum userset length scale m
m Mass kg
n Parameter for RosinRammler distribution
ˆ n Face Normal
N
r
Number of reactions
N
s
Number of species
Nu Nusselt number
Oh Ohnesorge number
p
0
Pressure at outlet in previous time step Pa
p
∞
Desired outlet pressure Pa
p Pressure Pa
vii
viii
p Momentum kgm/s
R
RNG
Coeﬃcient in the RNG kε turbulence model
R Universal Gas Constant J/Kmol
RR
i
Reaction Rate of species i
Re Reynolds number of ﬂuid
S Strain 1/s
S
fa
Help variable
Sc Schmidt number
Sh Sherwood number
T Temperature of ﬂuid K
t Time s
Ta Taylor number
u,U Velocity m/s
w Outgoing pressure wave velocity m/s
We Weber number
X
i
Volume fraction of scalar i
X
v,s
Mass fraction of fuel vapour at droplet surface
X
v,∞
Mass fraction of fuel vapour far away
Y
i
Mass fraction of scalar i
Greek Symbols
α Volume fraction of ﬂuid
α
k
,α
ε
Turbulence model constant
β Volume fraction of ﬂuid
β
max
Maximum spray cone angle
γ Vapour (cavitated liquid) fraction
δ Cellface distance coeﬃcient m
∆t Timestep s
ε Dissipation rate m
2
/s
3
κ
c
Thermal Conductivity W/mK
λ Limiter function for interfaceCompressionscheme
Λ Wavelength of fastest growing wave m
µ Laminar viscosity kg/ms
µ
T
Turbulent viscosity kg/ms
ν Kinematic viscosity m
2
/s
ρ Density kg/m
3
˙ ρ
s
Spray mass source term kg/m
3
s
τ
e
Relaxation time of evaporation s
τ
u
Momentum relaxation time s
φ Volume ﬂow across cell face m
3
/s
ψ Compressibility s
2
/m
2
Ω Growth rate of fastest growing wave m/s
2
Contents
1 Introduction 1
1.1 Background & Motivation . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Simple Description of the Computational Approach . . . . . . . . 4
1.3 Implementation and Code . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Nozzle Flow 9
2.1 Previous Works in the Field . . . . . . . . . . . . . . . . . . . . . 10
2.2 A Code for Cavitating Diesel Injector Flow into Air . . . . . . . 11
2.3 Boundary Conditions for Pressure . . . . . . . . . . . . . . . . . 15
2.3.1 Inlet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.2 Outlet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 A Code for Cavitating SinglePhase Flow . . . . . . . . . . . . . 16
2.4.1 Development of Injection Model . . . . . . . . . . . . . . 19
3 Spray Modelling 21
3.1 The Gas Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.1 Basic Equations . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.2 Turbulence Modelling . . . . . . . . . . . . . . . . . . . . 23
3.1.3 Turbulence/Spray Interaction . . . . . . . . . . . . . . . . 24
3.1.4 Chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Spray SubModels . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2.1 Spray Motion Equation . . . . . . . . . . . . . . . . . . . 30
ix
x CONTENTS
3.2.2 Parcel Tracking . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.3 Injection Model . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.4 Drag Model . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.5 Breakup Model . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.6 Evaporation Model . . . . . . . . . . . . . . . . . . . . . . 34
3.2.7 Heat Transfer . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.8 Particle Collisions . . . . . . . . . . . . . . . . . . . . . . 37
4 Results & Discussion 39
4.1 Simulations of a Diesel Flow in a Cavitating Model Nozzle . . . . 39
4.2 Injector Simulation Results . . . . . . . . . . . . . . . . . . . . . 43
4.2.1 Impact of Injection Model . . . . . . . . . . . . . . . . . . 44
4.3 Spray Results without Chemistry . . . . . . . . . . . . . . . . . . 46
4.3.1 Grid Dependency . . . . . . . . . . . . . . . . . . . . . . . 46
4.3.2 Time Step Dependency . . . . . . . . . . . . . . . . . . . 48
4.3.3 Comparison to Experimental Data . . . . . . . . . . . . . 50
4.4 Inﬂuence of Numerical Scheme . . . . . . . . . . . . . . . . . . . 51
4.5 Fuel Sprays with Chemistry . . . . . . . . . . . . . . . . . . . . . 53
4.5.1 Combustion Predictions Using Other Schemes . . . . . . . 57
4.6 Comment on the PaSR Model . . . . . . . . . . . . . . . . . . . . 59
5 Conclusions 61
6 Future Work 63
6.1 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7 Summary of Papers 67
A Implementation in OpenFOAM 71
A.1 RhieChow Interpolation in OpenFOAM . . . . . . . . . . . . . . 71
A.2 Formulating the Momentum Equation Matrix in OpenFOAM . . 76
B Variable Deﬁnitions and Dimensionless Numbers 79
Chapter 1
Introduction
1.1 Background & Motivation
Today, cars and heavy duty vehicles are considered to be major sources of
environmental pollution. However, when the ﬁrst automobiles appeared in the
streets of New York and London they were considered to be solutions to
another problem: horse manure. According to a rough estimate published in
Appleton’s magazine in 1908, 20 000 New Yorkers died each year from
“maladies that ﬂy in the dust, created mainly by horse manure”. Details of the
data used to obtain the estimate are unknown, but there is no doubt that
horses, and the resulting manure, constituted a major environmental
problem[1]. In the late 19th century, Nikolaus Otto in 1876, and later Rudolf
Diesel in 1897 patented their prime mover inventions. Horse manure became a
problem of the past. However, now, over 100 years later, we are facing a new
environmental threat, and this time our former saviours have become our
problem.
The Diesel engine is facing many challenges today, one of which is to reduce
soot and NO
x
generated in the process of burning the fuel. Both soot and
NO
x
are harmful to humans and animals ([2],[3]), so national governments
around the world are forcing manufacturers to decrease speciﬁc emissions from
the vehicles they produce, as shown in ﬁgure 1.1. According to an estimate
quoted in a report on the Diesel engine by P. Kageson [3], each year roughly
21 000 people over the age of thirty in France, Switzerland and Austria died
prematurely due to heart and lungdiseases caused by particulate matter from
vehicle traﬃc. It should be noted that more people are dying from traﬃc
pollution, than from traﬃc accidents in these countries.
In addition to the demands imposed by environmental legislation,
manufacturers have to meet the customers’ demands, who are often more
concerned about cupholders, extra horsepower and larger cars, than about
improving fuel eceonomy. In surveys by CNW Marketing Research [5], buyers
of cars were asked which factors were important in their decisions. These
1
2 CHAPTER 1. INTRODUCTION
1990 1995 2000 2005 2010
0
0.5
1
1.5
2
2.5
3
year
g
/
k
m
NO
x
PM
HC+NO
x
CO
Figure 1.1: Changes in legal limits for emissions from automobiles within the
EU, 19922008 [4].
surveys showed that although fuel economy increased from being important to
just 27.2% of the buyers in 1995, to 61.3% in 2005, it was important to fewer
customers than factors like interior comforts, manufacturer’s reputation, sound
system, and overall exterior styling (the reason most buyers found important).
It should be noted that the survey does not take into account how important
the factors were to the buyers, only the numbers who thought they were
important reasons for their choices. There also seems to be a discrepancy
between what consumers claim to do, and what they actually do. The same
report provides data on the average weights of US light vehicles and average
fuel consumption dating back to 1975. If most consumers really considered fuel
economy to be an important factor when choosing a car, a decline or at least
reductions in the rate of increase in vehicle size would be expected. However,
as can be seen from ﬁgure 1.2, the power of the engine per unit vehicle weight
has increased over the past 30 years, and increased even more since 2005. The
ﬁgure also shows that fuel consumption per unit distance travelled declined
sharply between 1975 and 1985, but due to the increases in power & weight,
the decrease has ﬂattened out since then. If customer awareness is increased,
or the oil price is drastically increased, perhaps the demand for large cars with
powerful engines will decline. Downsizing of the engine could then become a
viable option to meet legislative demands as well as customer needs. It should
also be mentioned that increases in safety features of modern cars have also
increased their weight.
In ﬁgure 1.3, the amounts of NO
x
and CO
2
emitted into the atmosphere
between 19902005 from vehicles in Sweden are shown. The amounts of
harmful NO
x
gases are decreasing, while those of the greenhouse gas CO
2
are
slowly increasing. CO
2
is an inevitable product of combustion when
1.1. BACKGROUND & MOTIVATION 3
1975 1980 1985 1990 1995 2000 2005 2010
0.6
0.8
1
1.2
1.4
1.6
year
f
u
e
l
c
o
n
s
u
m
p
t
i
o
n
(
l
/
1
0
k
m
)
1975 1980 1985 1990 1995 2000 2005 2010
50
60
70
80
90
100
p
o
w
e
r
/
w
e
i
g
h
t
(
w
a
t
t
/
k
g
)
Figure 1.2: Fuel consumption per unit distance travelled (solid line), and power
per unit weight of cars (dashed line), 19752007
1990 1995 2000 2005
0
50
100
150
200
year
1
0
0
0
t
o
n
n
e
s
NO
x
1990 1995 2000 2005
0
0.5
1
1.5
2
2.5
x 10
4
year
1
0
0
0
t
o
n
n
e
s
CO
2
Figure 1.3: Emissions from vehicles in Sweden, in 1000 tonnes, 19902005 [6]
carbonbased fuels are burnt, and the only way to reduce emissions of it is to
burn less. This can be done either by increasing the eﬃciency of the engine, or
using the engine less. The second option becomes more interesting as battery
power is increased. However, batteries that can power a vehicle for longhaul
highway operations are still far in the future. For the nearer future, increasing
fuel eﬃciency and engine downsizing are more likely to be viable strategies.
4 CHAPTER 1. INTRODUCTION
There are two branches of research concerning the optimization of engine
design and operators: one involving experiments with physical engines, and the
other involving attempts to model engines numerically. A symbolic
relationship between the two approaches is usually employed, since neither of
them can capture all the relevant details. The experimenter does not know
exactly what happens inside the engine, and the modeller cannot be certain
the events predicted by the model are correct without experimental
veriﬁcation. However, with a veriﬁed model, a reasonable description of the
processes taking place can be obtained. In addition, the modeller has all the
data available from them. Another advantage of modelling is that it allows
conceptual engines and combustion modes to be explored long before
prototypes are made. Simulations can also be useful for parameter studies,
where they provide a cheaper alternative than largescale engine studies.
1.2 Simple Description of the Computational
Approach
CFD is an abbreviation of Computational Fluid Dynamics, and refers to
solving a ﬁnite approximation of the NavierStokes equations numerically. The
spatial domain of interest, for example the cylinder of an engine, the shape of
a car, or the inside of a diesel injector, is divided into several small cells or
control volumes. The sum of the control volumes, called a mesh, provides a
ﬁnite approximation of the spatial domain. Solving conservation equations
using such a mesh is called an Eulerian approach, meaning that each position
in space has a numerical representation. For instance, a thermometer that
measures the outdoor temperature at the same spot every day provides an
example of an Eulerian description of temperature. Another way to describe
reality is to follow a particle, and compute its state vector, including variables
such as its position over time, conserving momentum and energy. This
approach, known as Lagrangian, is particularly appropriate when all forces
acting on an object are known. For example, the path of a rocket can be
conveniently described in Lagrangian coordinates, as all forces acting on the
rocket moves with it.
A combination of these two representations is often used in Diesel spray
simulations. The air in the combustion chamber is described using an Eulerian
framework, and the liquid spray is discretised into computational ’parcels’,
each of which is described by Lagrangian coordinates. That is, we inject part
of the spray, and track its position in time. The parcels may consist of any
number of droplets, each of which is considered identical, depending on values
set by the user. Each of these parcels is subject to the same processes as a real
diesel spray, including (inter alia): atomization, breakup, collision,
evaporation, heat transfer and turbulence. The processes mentioned, such as
heat transfer, occur as integral operations over the particle’s entire surface,
but resolution of particle surfaces is computationally prohibitive. Therefore, a
’point particle’ approximation is typically made, which requires a large
number of submodels to empirically account for integral ﬂuxes over the
1.3. IMPLEMENTATION AND CODE 5
particle surfaces. Further submodels, requiring still more empiricism, are
needed to model physical processes such as atomization. Unfortunately, most
of the submodels available today are quite unreliable, and depend on the user
to supply constants that have limited physical interpretability. These constants
can often diﬀer widely between authors, and publications, depending on the
type of spray behaviour being considered. A typical example is the breakup
constant in the KelvinHelmholzRayleighTaylor model ([7],[8],[9]), which can
vary over an order of magnitude.
1.3 Implementation and Code
The code used in the work this thesis is based upon was OpenFOAM ([10]
&[11]); an opensource code available at www.openfoam.com. It is an
objectoriented code written in C++, which makes it reasonably
straightforward to implement new models and ﬁt them into the whole code
structure. OpenFOAM is being continously developed; when I started my
doctoral studies I used Nabla FOAM 1.2, followed by OpenFOAM 1.0 up until
the recent 1.4.1 (the latest, 1.5, release has not been used). The code now
includes polyhedral mesh support, making it possible to create meshes using
any form of cells, as long as the quality of the resulting mesh is high.
Lagrangian parcels are tracked by facetoface tracking, thus no parcels are
lost when moving between cells as in Kiva3V. Further, models are
implemented to be runtimeselectable, which makes it very easy for the user
to switch between turbulence models, spray submodels, numerical schemes
etc. OpenFOAM has many other advantages
1
, but one of the most important
is the complete parallelization of the code. All solvers written in OpenFOAM
can easily be run in parallel, since the code is parallelized at such a
fundamental level, removing the need (in most cases) for the user to consider
multiple processor simulations.
OpenFOAM code was chosen both because of the high scope it oﬀers for
developing new models, and the demands it places on the user. It is more
diﬃcult to not know what you are doing in OpenFOAM, than in other codes.
One could argue that access to source code, and knowledge of what one is
doing, can easily be achieved by using an inhouse code. While this is true, I
believe that research should generally be done in a community in which
exchange of information is the norm, and an important contributor to
progress. The ability to share codes and ideas developed in the same
framework is one of OpenFOAM’s most important strengths.
1
One of the most recent additions is an automatic mesh generator.
6 CHAPTER 1. INTRODUCTION
1.4 Outline of the Thesis
The Introduction chapter presents basic information about sprays, and
motivation for the work. The idea is that this chapter should be
understandable to people without a degree in engineering. Subsequent
chapters assume a knowledge of basic university level math and physics. In the
following chapters, the project is described in the same order that the fuel
travels, starting with the nozzle ﬂow, and then on to the spray calculations,
turbulence interactions and combustion. Some parts of the modelling have
been described twice (in both papers and the thesis), but such repetition has
been kept to a minimum. Where it is included, it is included for continuity, or
because the description in the article was too concise due to size limitations.
The goals of the work have been to improve existing models for diesel spray
simulations, and to develop better models. Spray modelling today is at a stage
where there is a heavy reliance on experimental data for validation. One
example is the dependency on the computation grid. The errors in spray
simulations do not follow the common CFD rule of decreasing as the number
of cells is increased[12]. Instead, the discretization error has a minimum for a
speciﬁc grid size, due to the relationship between the liquid (Lagrangian
parcels) and the gas (Eulerian cells). The minimum has to be found on a
casetocase basis, and the results compared to experimental data to determine
which grid is most suitable. Therefore, spray simulations are more sensitive to
the grid than other CFD simulations, since an increase in the number of cells
can either increase or decrease the quality of the simulation. If some of this
grid dependency could be reduced, it would be of great value to the numerical
community.
EulerLagrangian spray simulations uses submodels to describe the processes
that aﬀect the injected liquid. One of the models that has been focused on in
this work is the injection and atomization model. The model is used to obtain
initial droplet size and velocity parameters, thus it has a major inﬂuence on
the initial breakup of the spray. Atomization refers to the primary breakup of
the liquid core of the spray, and it is separated from secondary breakup, which
describes how liquid droplets are split into smaller droplets. An atomization
model can be used to describe how the droplets break up initially, or simply
switch oﬀ breakup within a certain distance from the injector. The most
common way to inject droplets in EulerLagrangian sprays is to use a size
distribution and randomly inject droplets from the injector in directions
within the prescribed cone angle. A better injection model taking cavitation,
turbulence and the ﬂow in the nozzle into account that is only dependent on
physical parameters could increase the quality of the simulations. Therefore, a
goal of this project was to develop such a model, which does not prescribe an
injection angle and droplet distribution, but calculates it from set variables
such as the injection pressure, fuel density, length and width of the nozzle etc.
A very important parameter for diesel simulations is the ﬂame liftoﬀ length,
which is believed to have a strong eﬀects on the emissions generated during
combustion, since increasing it will cause the ﬂame to have a more premixed
character and consequently yield lower emissions of NO
x
and particulate
1.4. OUTLINE OF THE THESIS 7
matter. The liftoﬀ length phenomenon is quite complex, and requires a
welldeveloped spray model to even appear. If inappropriate schemes,
submodels and combustion models are employed the ﬂame will spread all the
way to the nozzle and no liftoﬀ length will be observed. It is also closely linked
to the ignition delay time, which is very important for engine simulations since
full control over the ignition is highly desirable. Attempts were also made to
benchmark OpenFOAM’s performance in terms of simulating these
phenomena, and compare it to that of the commonly used KIVA3V code.
The last issue addressed during the work underlying this thesis were the eﬀects
of numerical schemes and initial turbulent viscosity on the solution. It was
discovered quite early on (during the studies described in Paper 1) that the
chosen scheme and initial parameters have a large inﬂuence over the resulting
spray. Part of the reason why this was studied is the wide variety of schemes
available in OpenFOAM, in which it is very easy to switch between diﬀerent
types.
8 CHAPTER 1. INTRODUCTION
Chapter 2
Nozzle Flow
In both Diesel engines and direct injection spark ignition engines fuel is
injected into the combustion chamber by high pressure. Ideally, it should then
atomize, form small droplets, and eventually completely evaporate, yielding a
mixture that is easily combustible without forming either soot or NO
x
. In a
spark ignition engines it is also important to position the fuel cloud close to
the spark plug to promote good combustion. The atomization, i.e. the
formation of the initial droplets and their breakup, is mainly governed by the
ﬂow in the nozzle which ejects the spray ([13]). It is therefore important to
study the ﬂow inside the nozzle, to ﬁnd out how it is linked to the spray in the
cylinder. Furthermore, cavitation is believed to have a strong inﬂuence on the
ﬂow inside the nozzle, so it is also important to consider these eﬀects.
Cavitation can occur when ﬂuid of high velocity passes through a contraction
like a nozzle. A pocket of low pressure is formed in the wake of the
contraction’s edge, and if the pressure in this wake is suﬃciently low, i.e.
below the saturation pressure, the liquid will cavitate. The size of the
cavitation bubbles will depend on several factors, including injection pressure,
geometry, smoothness of the interior of the nozzle, and the properties of the
ﬂuid. For example, if the ﬂuid is pushed through the nozzle suﬃciently slow,
no cavitation will occur and ordinary pipe ﬂow will be observed.
Cavitation introduces vapour bubbles into the ﬂow and increases the
maximum velocity in the centre. The velocity is increased for two reasons
when the ﬂuid is cavitating. Firstly, if there is vapour along the walls the
liquid will not have a noslip condition boundary since the vapour will not be
stationary, thus allowing the velocity of the liquid to increase (which is
essential when there is a high pressure drop over the nozzle). Secondly, when
the ﬂuid is cavitating the liquid can not ﬁll the entire channel. Furthermore,
since the vapor bubbles formed in the nozzle are mixed with the liquid, when
they reach the combustion chamber the ligaments and droplets formed are
already partly evaporated through cavitation. The remaining liquid will
evaporate faster, and the entire atomization process will be accelerated.
9
10 CHAPTER 2. NOZZLE FLOW
2.1 Previous Works in the Field
The ﬂow inside a diesel nozzle has been studied by many researchers, both
experimentally ([14],[15],[16],[17]), and numerically ([17],[18],[19],[20]). The
results from [15] by Winklhofer et al have been used in this work to validate
the singlephase cavitation model. They have been used previously in [21]
using a Rayleigh Bubble Growth model [22]. In some studies nozzle ﬂow
calculations have been coupled to Lagrangian spray simulations ([23]) on a
casetocase basis, but few authors have performed nozzle simulations with
intention to create a new atomization model for EulerLagrangian diesel spray
simulations covering a comprehensive range of initial conditions. An exception
is Ning [24], who developed a new atomization model for the EulerLagrangian
Spray and Atomization (ELSA) approach. ELSA is a mixture of modelling
approaches in which a full Eulerian description is used for the nearnozzle ﬂow
where the spray is dense, but if the liquid volume fraction falls below a
userset value, it is transformed into a Lagrangian parcel.
An additional complication when modelling cavitation that most studies have
neglected (one exception is [25]) is the eﬀect of cavitation in multihole
injectors. As reported by Nouri et al in [26], strings of cavitation can occur
between diﬀerent nozzle holes in multihole injectors and the strings reduce
the eﬀective hole ﬂow area, which in turn increases the ﬂow velocity.
Cavitation is usually modelled using either Rayleigh bubble growth dynamics
theory in conjunction with a volume of ﬂuid method ([17],[22],[27]) or, the
Homogenous Equilibrium Model (HEM,[20],[28]), which assumes that the
cavitated mixture is perfectly mixed and the amount of cavitation can be
described by a vapour mass fraction between 0 and 1. In both approaches an
assumption about isentropic nozzle ﬂow is usually included, and the HEM
approach needs an equation of state relating pressure and density to calculate
the growth of cavitation. The Rayleigh bubble growth approach solves a
transport equation for the liquidgas interface, with a source term to model
cavitation, calculated from a simpliﬁed Rayleigh equation. There is also a
third approach, based on Gibbs free energy [29], which is very rarely used.
Previous works in the ﬁeld have modelled cavitating liquid, but not the
ambient air in the combustion chamber. This is often an essential
simpliﬁcation due to the diﬃculty of resolving three phases while maintaining
reasonable computational times. As part of this project, OpenCFD Ltd. was
contracted to develop a code handling liquid, cavitating liquid and ambient
air. The resulting code has not been validated within this project, due to
problems in its handling of diesel fuel ﬂows. However, it is possible to use it to
simulate water injections, although the simulation time extends over several
weeks. This code was originally described in [30], but since the cited
publication has limited availability it will also be described here. A simpler
code (presented in the following section) for a cavitating ﬂuid was also
developed as part of the project.
2.2. A CODE FOR CAVITATING DIESEL INJECTOR FLOW INTO AIR11
2.2 A Code for Cavitating Diesel Injector Flow
into Air
The purpose of this code is to simulate cavitating diesel fuel injected into a
chamber ﬁlled with air. Even though the code can have other uses, simulating
diesel injections was its original purpose. In this text, gas refers to the ambient
gas which the liquid is injected into, and vapour to liquid that undergoes
cavitation. To simulate the liquidgas interface, the Volume of Fluid (VOF)
method is used, in which the liquid phase is allowed to cavitate and the gas
phase can be compressed due to pressure shocks. The equations that need to
be solved are, as usual, the momentum equation and the continuity equation,
presented below:
∂ρU
∂t
+ ((U· ∇)ρU) −∇· (µ∇U) = −∇p (2.1)
with
µ = α(γµ
α,v
+ (1 −γ)µ
α,l
) +βµ
β
+µ
sgs
(2.2)
where α and β are the fuel and air volume fractions, respectively. The
relationship between these two is:
α +β = 1 (2.3)
In equation 2.2, the vapour fraction is described by γ, indicating the
proportion of the cells that are cavitated.
γ =
ρ
α,lv
− ρ
α,lSat
ρ
α,vSat
− ρ
α,lSat
(2.4)
The code uses a Large Eddy Simulation (LES) model, which is apparent in
equation 2.2 in the µ
sgs
term. LES is believed to be necessary in order to
resolve the liquid turbulence interaction with the low turbulence in the gas
phase.
In addition to the momentum equation there is the continuity equation for
each phase:
∂αρ
α
∂t
+∇· (ρ
α
αU) = 0 (2.5)
∂βρ
β
∂t
+∇· (ρ
β
βU) = 0 (2.6)
where
12 CHAPTER 2. NOZZLE FLOW
ρ = αρ
α
+βρ
β
(2.7)
and thus equations 2.5 & 2.6 can be combined to form
∂ρ
∂t
+∇· (ρU) = 0 (2.8)
Since the code is intended to account for compressibility an equation of state
(EOS) relating pressure and density is needed. There was a more detailed
rationale for the EOS used in the studies presented in Paper 2, but that code
does not use VOF and thus only one EOS is needed. For modelling diesel
injection into air two equations of state are needed, one for liquid and one for
the gas. The main reason for selecting the chosen EOS (for the liquid) is that
it obeys the liquid and vapour EOS in the limit cases of pure liquid and pure
vapour, and some form of mixture for the intermediate states. The gas EOS is
simpler, since it is assumed to not undergo any phasechanges:
ρ
liq,vap
= (1 −γ) ρ
0
l
+ (γψ
v
+ (1 −γ) ψ
l
) p
sat
+ψ(γ)
_
p −p
sat
_
(2.9)
ρ
air
= ψ
g
p (2.10)
Where ψ
v
and ψ
l
refers to the vapour and liquid compressibility, respectively.
For reference, it is related to the speed of sound by:
a =
1
√
ψ
(2.11)
The constant ρ
0
l
is deﬁned as
ρ
0
l
= ρ
l,sat
−p
sat
ψ
l
(2.12)
When the code is executed the model for ψ is chosen at runtime. Three models
have been implemented: the Wallis equation [31]:
ψ
Wallis
= (γρ
v,Sat
+ (1 −γ)ρ
l,Sat
)
_
γ
ψ
v
ρ
v,Sat
+ (1 −γ)
ψ
l
ρ
l,Sat
_
(2.13)
the Chung equation [32]:
s
fa
=
ρv,Sat
ψv
(1 −γ)
ρv,Sat
ψv
+γ
ρ
l,Sat
ψ
l
(2.14)
ψ
Chung
=
__
1 − γ
√
ψ
v
+
γs
fa
√
ψ
l
_ √
ψ
v
ψ
l
s
fa
_
2
(2.15)
2.2. A CODE FOR CAVITATING DIESEL INJECTOR FLOW INTO AIR13
and a linear model:
ψ
linear
= γψ
v
+ (1 −γ) ψ
l
(2.16)
The linear model was the one chosen for this work, since it is then consistent
with the VOFmethod, where viscosity and mass fraction are also described by
linear equations. The Wallis model represents the lower speed of sound
(compressibility) in a bubbly mixture in a more physical way, but it was found
to be quite unstable when used for high speed ﬂows. When a linear
compressibility model is used, the EOS for the liquid can be reduced to:
ρ
liq,vap
= ρ
l,sat
+ψp (2.17)
We can now return to equations 2.5 and 2.6, and form equations that are more
suitable for solving for the phase fraction and densities. We can begin by
separating the two equations into two components: one representing
incompressible ﬂow, and the other the compressibility eﬀects.
∂α
∂t
+∇· (αU) = −
α
ρ
α
Dρ
α
Dt
(2.18)
∂β
∂t
+∇· (βU) = −
β
ρ
β
Dρ
β
Dt
(2.19)
These two equations could be solved for α and β respectively with the right
hand side as the compressible “source”. However, when these equations are
discretized, they will not be massconservative, since there is no coupling
between α and β when they are solved. The way the phase fraction variables
are calculated will be described later. If the two equations above are combined
with 2.3 an equation for the divergence of U is formed:
∇· U = −
α
ρ
α
Dρ
α
Dt
−
β
ρ
β
Dρ
β
Dt
(2.20)
This equation is used with the discretized momentum equation and the
equation of state (equations 2.9 & 2.10, respectively) to form the equation for
pressure. An example of how this is done for a simpler code (icoFoam) can be
found in Appendix A.1.
Equations for U and p have now been derived, and by using the equation of
state, the density can be found. However, we also need an equation to convect
the density for the momentum equation, which has already been deﬁned in 2.5
& 2.6. They can be rearranged into:
14 CHAPTER 2. NOZZLE FLOW
∂ρ
α
∂t
+∇· (ρ
α
U) = ρ
α
β
_
1
ρ
α
Dρ
α
Dt
−
1
ρ
β
Dρ
β
Dt
_
(2.21)
∂ρ
β
∂t
+∇· (ρ
β
U) = ρ
β
α
_
1
ρ
β
Dρ
β
Dt
−
1
ρ
α
Dρ
α
Dt
_
(2.22)
These equations are used to convect ρ and illustrate that if the material
derivatives of either density is larger than the other, it will result in a
negative/positive right hand side. Further, the phase densities will remain
singlephase if either α or β is zero. The terms within the parenthesis on the
right hand side are treated explicitly, when the equations are solved for ρ
α
and
ρ
β
.
The last two variables that need to be solved for are the phasefractions. The
equation to be solved can be derived from 2.18 & 2.19, where the convection
term is evaluated and 2.20 is used to eliminate the divergence term. This
results in:
∂α
∂t
+U· ∇α = −αβ
_
1
ρ
α
Dρ
α
Dt
−
1
ρ
β
Dρ
β
Dt
_
(2.23)
∂β
∂t
+U· ∇β = −αβ
_
1
ρ
β
Dρ
β
Dt
−
1
ρ
α
Dρ
α
Dt
_
(2.24)
Unlike 2.18 and 2.19, these equations are bounded by αβ, and if care is taken
with regard to sign of the right hand side, the boundedness can be maintained
by choosing the solver to be implicit with respect to α or β, depending on the
sign of the term within the parenthesis. The code only solves equation 2.23,
and then calculates β through equation 2.3.
The last requirement is to ﬁnd a way to maintain sharp interfaces, since when
using the VOF method it is essential to ensure that the interface is not
dispersed, causing cubes to lose their shape and balls to become “wiggly”.
This is usually done by using a compressive diﬀerencing scheme such as
CICSAM [33], which unfortunately has problems to fulﬁl its nondisperse
property when both liquid and air are ﬂowing away from the interface
(divergent ﬂow) [33]. For this code a “countergradient” method was chosen.
This involves not only use of a speciﬁc scheme, but also the introduction of a
new term, which compresses the interface. The new term cannot be allowed to
aﬀect the solution in any other way than by compressing interfaces. It is also
essential for this term to reduce to zero as the grid is reﬁned, since the
distortion of interfaces and associated problems are larger for coarse grids. By
adding the term as a divergence term, the equation will still be conservative,
and the term introduced will have an αβ term, meaning that it will only be
nonzero around interfaces. Equation 2.21 is then modiﬁed by adding the third
term on the left hand side in the equation below:
∂α
∂t
+U· ∇α +∇· (U
c
αβ) = −αβ
_
1
ρ
α
Dρ
α
Dt
−
1
ρ
β
Dρ
β
Dt
_
(2.25)
2.3. BOUNDARY CONDITIONS FOR PRESSURE 15
The success of the model now depends on the choice of U
c
, which should only
compress interfaces, and thus be directed in
∇α
∇α
. The method must compress
interfaces even when no ﬂow is present, i.e. when U is low. There are many
options for ensuring this, but the one chosen for this code is:
U
c
= min(c
α
U, max(U))
∇α
∇α
(2.26)
As mentioned above, the liquid/gas interface is maintained not only by use of
a speciﬁc compressible numerical scheme but also by introducing the
mentioned term. The scheme used is implemented in OpenFOAM under
interfaceCompression, which switches between upwind and central
diﬀerencing depending on the limiter function λ.
λ = min(max[1 −max{(1 −4α
P
(1 −α
P
))
2
, (1 −4α
N
(1 −α
N
))
2
}, 0], 1) (2.27)
where subscript P refers to the current cell value and N the neighbour cell
value.
2.3 Boundary Conditions for Pressure
When modelling a high Machnumber ﬂow, the boundary conditions become
very important. If a nonreﬂective outlet condition is not used, pressure waves
from the nozzle will bounce oﬀ the outlet boundary and destroy the physical
realism of the simulation. The same is true for the inlet; using a simple ﬁxed
value for the pressure at the inlet is insuﬃcient. Therefore special care has
been applied to the boundaries.
2.3.1 Inlet
The total inlet pressure takes into account the compressibility of the mixture,
as well as the calculated local velocity, as follows
p
inlet
=
p
0
(1 +
1
2
ψ(1 −pos(φ))U
2
)
(2.28)
where: p
0
is the pressure set by the user; φ is a scalar volume ﬂux on the face
(for more information on what φ really means, see Appendix A.1); pos is a
function returning the positive part of this velocity (which is a scalar); while ψ
and U are read from the ﬁeld.
2.3.2 Outlet
The nonreﬂective pressure boundary condition implemented in OpenFOAM is
a simpliﬁcation of that proposed in [34]. Input for the model is p
∞
and l
∞
,
16 CHAPTER 2. NOZZLE FLOW
which represent the outlet pressure and relaxation length scale, respectively.
The relaxation length scale is the parameter that governs how reﬂective the
outlet will be; a low value will give a more reﬂective outlet than a high value.
The model begins by calculating the velocity of the outgoing pressure wave:
w = U· ˆ n +
_
1
ψ
(2.29)
where ˆ n is the outlet normal. The pressure wave velocity is used to calculate
the pressure wave coeﬃcient α and the relaxation coeﬃcient b:
α = w
∆t
δ
(2.30)
b =
w∆t
l
∞
(2.31)
where δ is the cellface distance coeﬃcient. The value of the pressure can now
be calculated using these properties:
p
trans
=
p
0
+bp
∞
1 +b
(2.32)
ξ =
1 +b
1 +α +b
(2.33)
where p
0
refers to the pressure at the outlet in the previous timestep. The
pressure at the outlet is not set to p
trans
, instead a mixture of p
trans
and the
pressure in the cell (p
cell
) closest to the outlet is used.
p
outlet
= ξp
trans
+ (1 −ξ) p
cell
(2.34)
This ends the description of the multiphase code for cavitating ﬂow. In the
results section some results obtained by applying it to water injection into a
chamber are presented.
2.4 A Code for Cavitating SinglePhase Flow
A simpler, more robust code was also developed as part of the previously
mentioned threephase code for cavitating nozzles. This code simulates only
cavitating liquid, and is the one used in the studies presented in Paper 2. The
implementation is publicly available under the name cavitatingFoam in
OpenFOAM 1.4.1.
A singlephase code for a cavitating ﬂuid, is usually constructed by
introducing variable density, which can be dependent on pressure and/or
2.4. A CODE FOR CAVITATING SINGLEPHASE FLOW 17
temperature. If it is only dependent on pressure, the equation of state is said
to be barotropic. Since it is implemented in a ﬁnite volume code, it is also
convenient to assume homogenous equilibrium, which means the liquid and
vapour are assumed to be always perfectly mixed in each cell. A property, γ, is
introduced to describe the amount of cavitated vapour in each cell.
As mentioned above, an equation of state is required to model cavitation. In
this case we have assumed the temperature is constant (in accordance with
reference experiments [15]). A common barotropic equation of state is then the
nonequilibrium diﬀerential equation:
Dρ
Dt
= ψ
Dp
Dt
(2.35)
This equation 2.35 can either be used directly in the continuity equation to
formulate a pressure equation, or integrated to obtain the pressure as a
function of the density. The latter approach has been applied by Schmidt et al
[28]. The former approach is problematic when it comes to consistency. The
former approach is complicated by the lack of consistency (since equation 2.35
is not an equilibrium equation of state) between the pressure and density
obtained from it with the liquid and vapour equations of state before
equilibrium is reached. Thus, until equilibrium is reached, errors from the
inconsistency accumulate, and the code will not produce accurate results when
it has reached equilibrium.
Like the equivalent threephase EOS, the equation of state for the singlephase
cavitation code should therefore obey the liquid and vapour equations of state
both in the limit cases when there is pure liquid or pure vapour, while in
intermediate cases between these states it must have some form of mixture.
The two states have the following linear equations of state:
ρ
v
= ψ
v
p (2.36)
ρ
l
= ρ
0
l
+ψ
l
p (2.37)
The property that describes the proportion of liquid in each phase is γ:
γ =
ρ −ρ
l,sat
ρ
v,sat
−ρ
l,sat
(2.38)
γ = 1 corresponds to a fully cavitated ﬂow, and γ = 0 a ﬂow without
cavitation. ρ
v,sat
is calculated from
ρ
v,sat
= ψ
v
p
sat
(2.39)
where ψ
v
is the compressibility of the vapour. These properties together form
the mixture’s equalibrium equation of state:
18 CHAPTER 2. NOZZLE FLOW
ρ = (1 −γ) ρ
0
l
+ (γψ
v
+ (1 −γ) ψ
l
) p
sat
+ψ(γ)
_
p −p
sat
_
(2.40)
The compressibility, ψ, and choice of how it is modelled. were mentioned in
section 2.2. For the singlephase code, the considerations when choosing a
model are the same, and for the simulations in Paper 2 the linear model in
equation 2.16 was chosen. For the mixture’s viscosity, a simpler version of
equation 2.2 can be used, since only one compressible phase is then present:
µ
f
= γµ
v
+ (1 −γ) µ
l
(2.41)
When a linear model is used for the compressibility, the equation of state
(2.40) can be simpliﬁed to:
ρ = (1 −γ) ρ
0
l
+ψp (2.42)
The ﬁrst term governs the liquid density when γ is low. If the ﬂuid is
cavitating, the second term becomes more dominant. The ﬁrst term contains
the property ρ
0
l
which is:
ρ
0
l
= ρ
l,sat
−p
sat
ψ
l
(2.43)
where ρ
l,sat
is the liquid density at standard conditions. The saturation density
of the vapour, used earlier to calculate γ, is important for the liquid’s
tendency to cavitate. If ρ
v,sat
is increased, ρ will not need to be so low in order
to obtain a higher γ.
The relationship between density and pressure are used in the continuity
equation to transform it from a density equation to a pressure equation. For
reference, we have the compressible continuity equation:
∂ρU
∂t
+∇· (ρUU) = −∇p + (µ
f
∇U) (2.44)
∂ρ
∂t
+∇· (ρU) = 0 (2.45)
Since this is compressible ﬂow, it is more straightforward than incompressible
ﬂow. Now, with equation 2.40 we get:
∂ψp
∂t
−
_
ρ
0
l
∂γ
∂t
+ (ψ
l
−ψ
v
) p
sat
_
∂γ
∂t
−p
sat
∂ψ
∂t
+∇· (ρU) = 0 (2.46)
2.4. A CODE FOR CAVITATING SINGLEPHASE FLOW 19
This equation can be combined with the numerical discretization of the
momentum equation to obtain the continuity equation for pressure. This
equation is solved in conjunction with the momentum equation in a PISO
loop, as normal. This code does not use any turbulence model, instead it relies
on stabilizing numerical schemes. This can be regarded as implicit LES if
small cells are used.
2.4.1 Development of Injection Model
These simulations were originally intended to provide input for the
development of a new injection model for EulerLagrangian diesel spray
simulations. The features of interest were the ways in which cavitation layer
thickness, velocity proﬁle, velocity ﬂuctuations and the discharge coeﬃcient
vary with changes in injection pressure, curvature of the inlet,
L
D
, etc.
A schematic diagram of the development of a cavitation layer in a nozzle can
be seen in ﬁgure 2.1. The size of the cavitation layer is the size of the vapour
layer close to the wall, and it ends where the liquid core begins. This
distinguishes it from the boundary layer, which is located where the freestream
velocity is reached.
δ
δ
δ
cavitation
cavitation
core
Figure 2.1: Schematic diagram of a cavitation proﬁle, and cavitation layer clas
siﬁcation
The theory behind the model is that if the cavitation layer (δ
cavitation
) is thin,
there will be more mass in the centre of the injector, and the model will return
large initial droplets. Since large droplets are assumed to originate from the
centre of the spray, this could also result in a smaller spray angle. Smaller
droplets are then formed from the cavitation layer itself, and the distance they
move from the spray axis is governed by the turbulence in the nozzle. Thus, a
thin cavitation layer can still result in a large spray angle, if the turbulence is
suﬃciently high.
This assumption and the points raised in the brief discussion on the cavitation
layer have all been mentioned before in the literature ([35],[36]). However, Huh
[35] only considers the results of turbulence, and the Max Planck Institute
(MPI) model, described in [36], only considers the eﬀect of cavitation in the
20 CHAPTER 2. NOZZLE FLOW
nozzle. The aim of the new model was to account for both cavitation and
turbulence, and to estimate their magnitude using CFD.
During the course of the project, several attempts were made to simulate
cavitating diesel nozzle ﬂow. Initially, the equation of state in the code
assumed that the vapour density was proportional to the pressure. The code
worked well for water injections into gas at atmospheric air pressures, but for
diesel pressures it caused the vapour density to rise above the liquid density.
This forced the use of a unphysically high saturation pressure to avoid
negative pressures in the solution and allow the vapour density to remain
below the liquid density. Attempts were made to extend this code to use
PengRobinson equation of state (EOS) [37], but the EOS was not suﬃciently
smooth to handle the large pressure gradients in the ﬂow. The compressibility
model was also changed, from the Wallis [31] model to a linear compressibility
model. The change was necessary partly to enhance stability and partly
because the same cavitation model was to be used for the multiphase code
handling diesel vapour, liquid and air, as mentioned earlier. Since a linear
combination of the vapour fraction (γ) and liquid/gas fraction (α and β) is
highly suitable for describing the mixture’s properties in such cases, the same
model was used for the singlephase code. A similar story of cavitation code
development is reported in [38]. It starts oﬀ by using what the papers refers to
as the Schmidt model [28], and ends in a modiﬁcation of that model with a
cutoﬀ for negative pressures.
The combined eﬀects of these factors caused long delays in the simulations,
and forced the author to limit the work to validating a model for cavitating
singlephase ﬂow, without being able to develop a model for Lagrangian diesel
spray simulations.
1
1
It should also be noted that it is only in special cases (mainly sprays with low evaporation,
e.g. in cold starts) that the size of the initial droplets and their maximum spreading angle
really aﬀect the combustion. This issue will be addressed in section 4.2.1
Chapter 3
Spray Modelling
In section 1.2, Lagrangian spray simulations were brieﬂy described. The
method used to introduce Lagrangian particles into an Eulerian grid is
sometimes called the discrete droplet model (DDM). This is based on the fact
that we cannot resolve the full details of the nearnozzle ﬂow using only
Eulerian cells (as yet at least). A typical nozzle has a diameter of around
200 µm, so if it is to be resolved properly a cell size of about 20 µm is needed,
but even if such small cells are only used around the injector the number of
cells required grows enormously if the calculations are run in 3D. Thus, the
spray is usually modelled in some other way. There has been a growing
number of attempts to simulate sprays using a full Eulerian approach, or by
using LES instead of the classical DDM with a kε turbulence model. These
attempts are often associated with some limitations. For instance, de Villiers
et al [39] considers primary breakup close to a noncavitating nozzle, Befrui et
al [40] uses an updated version of the same code to study primary breakup in
noncavitating gasoline direct injection, and M´enard et al [41] considers liquid
injection into incompressible air with a velocity of around 100 m/s. A spray
study using LES and an EulerLagrangian description of the gas and liquid
has also been published, by Vuorinen et al [42]. They also used lower velocities
than those of diesel sprays and assumed the surrounding air to be
incompressible, but the Reynolds number was comparable to those of diesel
sprays. However, they did not simulate the nearnozzle behaviour, and mainly
focused on studying the spray behaviour. All the papers mentioned above,
ignore cavitation, and the simulations cover only part of the relevant physical
domain.
One of the biggest problems associated with the DDM is modelling the
nearnozzle ﬂow, more speciﬁcally the conditions close to the injector. In this
region a liquid core forms from the liquid fuel being injected through the
injector. Ligaments are separated from this liquid core and form droplets that
evaporate and mix with the ambient gas. Hence, since the DDM assumes
disperse ﬂow with spherical droplets of liquid, it is not an accurate
representation of the spray in this region. An option to improve nearinjector
ﬂow is to use the previously mentioned (in section 2.1) ELSA model. Aside
21
22 CHAPTER 3. SPRAY MODELLING
from the work by [24] (mentioned in 2.1), Blokkeel et al [43] and Lebas et al
[44] have also used it to improve the primary breakup of the spray. Since the
DDM model is only used when the liquid fraction is low, representing a state
in which droplets (not ligaments) have been formed from the liquid core, the
model gives a better representation of the nearnozzle ﬂow. Another beneﬁt is
that such a modelling concept does not require a primary atomization model,
and thus there are fewer model constants to tune. However, this raises some of
the problems associated with high grid resolution to resolve the nozzle
mentioned earlier. There have also been other attempts to simulate sprays
using only Eulerian cells, in 2D ([45]). However, the EulerianLagrangian
description of the diesel spray has been the most widely used.
The large number of submodels used to simulate the various physical
processes that the parcels are subjected to before they become part of the gas
can pose further problems when using the DDM model. Each of these models
will be thoroughly described later in this chapter, but here my intention is to
mention the interactive problems they raise as a group. For instance, even if
the breakup model is fully correct, the parcels can become too small due to
the evaporation model simulating evaporation too quickly, thus nullifying the
excellence of the breakup model. Similarly, a perfect turbulence interaction
can be ruined by a dissipative numerical scheme, or an overpredicting drag
model that slows all the parcels down too much. Therefore, tuning the spray
model in a noncombusting environment to constant volume experimental
data, and then applying the results of the tuning to real engine calculations
often yields poor results. The complex interactions amongst all of the
submodels make spray simulation somewhat more of an art than a science.
3.1 The Gas Phase
3.1.1 Basic Equations
The Eulerian grid is used to discretise the conservation equations of the ﬂuid
phase. When chemical reactions and interactions between the Lagrangian and
Eulerian phases are also involved, Eq. 2.8 needs to be extended to include
liquid spray evaporation. Furthermore, the density ρ becomes more complex
since diesel combustion involves multiple species, all of which aﬀect the
density. In addition, since chemical reactions are occuring various compounds
will be formed and participate in further reactions, so the mass fraction of
each species is not conserved. However, if we consider the total density, the
only source term ( ˙ ρ
s
) is from spray evaporation.
∂ρ
∂t
+∇· (ρU) = ˙ ρ
s
(3.1)
The evaporation source term is calculated from the evaporation submodel,
which will be described later. As already mentioned, each species will be
transported, by both diﬀusion and convection, and may also be formed or
3.1. THE GAS PHASE 23
consumed by the reactions. The mass fraction for species i is denoted by Y
i
,
and its transport equation is:
∂ρY
i
∂t
+∇· (ρUY
i
) −∇((µ +µ
T
)∇Y
i
) = ˙ ρ
si
+κRR
i
(3.2)
˙ ρ
s
=
i
˙ ρ
si
(3.3)
since most fuels consist of multiple species, and when combustion occurs their
number increases dramatically. For instance, when the mechanism including 83
species was used for the calculations shown in Paper 3, the transport equation
had to be solved 83 times in each time iteration. The momentum equation also
needs to be modiﬁed due to the presence of a Lagrangian spray. Presented
here is the equation for the mean velocities using a Reynolds Average Navier
Stokes assumption:
∂ρU
∂t
+∇·
_
ρU
T
U
_
= −∇p+∇·
_
(µ +µ
T
)
1
2
(∇U+ (∇U)
T
)
_
+ρg+F
s
(3.4)
F
s
is the spray momentum source term.
3.1.2 Turbulence Modelling
Since the goal of this project was to develop new models for industrial
applications, the turbulence was modelled using kε models. Paper 1 presents
constant volume vessel simulations of diesel sprays conducted using three
diﬀerent turbulence models, which yielded results that substantially diﬀered in
terms of spray shape and penetration of liquid and vapour. For later
simulations diﬀerent turbulence models were also tested, bearing in mind the
intial results. The discrepancies between the models indicate the importance of
modelling the turbulence correctly.
The three (compressible) models used in Paper 1 were: standard kε [46],
Renormalized Group kε [47] (RNG) and LaunderSharma’s kε (LS) [48]
models. Each of the three turbulence models was developed with a diﬀerent
rationale. The RNG model and the standard kε model used for these cases
are both high Reynolds number models, meaning that they apply wall
functions for the near wall cell. The LaunderSharma turbulence model diﬀers
in this respect, since it is a low Reynolds number model, and has no wall
functions. This places higher demands on the nearwall cell size, but since all
computations in this work simulated events in a bomb, the nearwall
behaviour was not as important as the spray interactions by any of the walls
except the injectorwall.
The ﬁrst turbulence model based on k and ε was originally presented in [46]. It
is shown in equations 3.5 and 3.6 in a general form to accommodate the three
turbulence models.
24 CHAPTER 3. SPRAY MODELLING
∂
∂t
(ρk) +∇· (ρkU) −∇· [(α
k
µ
T
+µ)∇k] =
= 2µ
T
S −
2
3
ρk∇· U−ρε −D
i
(3.5)
∂
∂t
(ρε) +∇· (ρεU) −∇· [(α
ε
µ
T
+µ)∇ε] =
C
1
2µ
T
S
ε
k
−
2
3
C
1
ρε∇· U−C
2
ρ
ε
2
k
+ B
i
(3.6)
where B
i
and D
i
are listed below. Numerical constants and other variables are
listed in Paper 1.
i B
i
D
i
kε −C
3
ρε∇· U 0
RNG −2µ
T
R
RNG
Sε/k 0
LS −C
2
fρ
ε
2
k
+ 2µµ
T
∆U
2
ρ
2µ∇
√
k
2
Table 3.1: Speciﬁc terms of each turbulence model
3.1.3 Turbulence/Spray Interaction
There is no liquid spray source term in equations 3.5 & 3.6, so the turbulence
is not directly aﬀected by the presence of particles. Diﬀerent researchers have
adjusted their models in diﬀerent ways to incorporate the eﬀects of particles
on turbulence. Including a source term in the equations is the most common
way to achieve this, but not the only one. An example of a model that uses a
source term in the equations for the turbulent quantities is KIVA ([49]).
However, that term can only reduce the kinetic energy of the turbulence,
which is not consistent with the conclusions drawn from experiments by Crowe
et al [50, 51], see ﬁgure 3.1.
According to experimental investigations of gassolid and gasliquid ﬂows
summarized in 3.1, the property
d
lt
aﬀects whether particles will increase or
decrease the intensity of the turbulence. A value above 0.1 indicates that the
intensity of the turbulence will increase, and a value below the opposite. For
most diesel spray simulations, the length scale
1
l
t
is of the order of 100 µm in
cells where there are parcels, and the droplet diameter ranges from 100 µm
(depending on the size of the injector) down to about 1 µm (there may of
course be smaller droplets than this, but they will quickly evaporate). Figure
3.2 shows the distribution of the quantity
d
lt
, centered around 0.1, indicating
that some parcels will increase, and others reduce, the turbulence.
1
According to the kε turbulence model, which might not predict the turbulence entirely
correctly.
3.1. THE GAS PHASE 25
Figure 3.1: Changes in turbulent intensity as a function of length scale ratio.
Reprinted from [50] with permission from Elsevier Ltd.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
100
200
300
400
diameter/L
t
n
u
m
b
e
r
o
f
p
a
r
c
e
l
s
Figure 3.2: Distribution of parcels from ﬁve spray simulations obtained using
diﬀerent turbulence models, both combusting and noncombusting. The total
number of parcels was 1018, and L
t
was approximated to 100 µm
Aside from using new terms in the turbulence model, there are other ways to
account for the presence of droplets in a simulation. For instance, one of a list
of suggestions for improving the performance of spray simulations in a book by
Stiesch [12] is to limit the length scale of the gas phase turbulence inside the
26 CHAPTER 3. SPRAY MODELLING
spray to the jet diameter in order to reduce the grid dependency of the
calculations. Such a modiﬁcation would also cause the turbulence model to be
aﬀected by the particles. However, the suggestion raises several problems.
Notably, the jet diameter is not well deﬁned, it could refer solely to the liquid
part of the spray, or the vapour part could also be included. Similarly, the
term “inside the spray”, could refer to just the liquid core, or include the
vapour cloud.
The author chose to limit the turbulent length scale to the nozzle diameter in
cells with parcels. The spray is a cloud of parcels, and all cells that have
parcels in them must be inside this cloud. There can be cells that do not have
a parcel in them, but if we choose a high number of injected parcels there will
be few such cells. The nozzle diameter is selected rather than the jet diameter
because the nozzle diameter is the parameter that initially sets the length
scale of the turbulence. Fixing the limit prevents the turbulence length scale
from growing, but it also simpliﬁes the calculations, since calculating the jet
diameter in every timestep would be costly. Thus, the nozzle diameter is a
good ﬁrst approximation.
k governs the scale of the ﬂuctuating turbulent velocity, and ε the size of the
turbulent eddies. The limit was thus imposed on ε. The length scale of the
turbulence is deﬁned as:
l
t
= C
µ
k
3/2
ε
(3.7)
Thus, since the limit was imposed on ε, it was limited by:
C
µ
k
3/2
L
sgs
< ε (3.8)
Where L
sgs
is set to the nozzle diameter. The constraint is imposed after the
equations for k and ε are solved. If the constraint is applied before any of the
equations are solved, the constraint in equation 3.8 cannot be guaranteed.
The modiﬁcation of the turbulence model implies an assumption regarding ε;
that the equation for ε underestimates the dissipation when liquid parcels are
present. One could also think of this as overprediction of the length scale of
the spray, which is perhaps easier to imagine, and this explanation will be used
in the rationale below.
The parcels in a diesel spray problem have fairly high velocities, and the
increased momentum of the gas comes from the spray. Thus, the resulting
length scale of the turbulence should also be governed by the spray and nozzle
size, which is what this modiﬁcation does. One could argue that this only shifts
the dependency from the grid to the magnitude of the turbulence length scale
limit. While this is true, it is a much cheaper dependency computationally.
3.1. THE GAS PHASE 27
3.1.4 Chemistry
After the fuel liquid droplets have undergone breakup and evaporation, the
fuel mixes with the surrounding air and forms a combustible mixture. A real
diesel fuel consists of about 1000 diﬀerent species, reacting with the air in over
10 000 reactions. Thus, when studying combustion, either experimentally or
numerically, the complexities are generally reduced by considering the
combustion of a simpler model fuel. One such model fuel is nheptane, another
is the IDEA model fuel consisting of ndecane and αmethylnaphthalene.
Both have been used in this work: IDEA for the spray studies described in
Paper 1, and nheptane for the studies described in Papers 3 and 4.
Solving the chemistry numerically means solving a large system of reaction
equations. For each reaction
S
f
11
[X
1
] + S
f
21
[X
2
] +... − →S
r
11
[X
1
] + S
r
21
[X
2
] +... (Re. 1)
there is a corresponding reaction rate equation which determines how rapidly
the reaction is proceeding and in which direction. To formulate the reaction in
a more general manner, reaction j is written as:
Ns
i=1
S
f
ij
[X
i
]
k
f
j
k
r
j
Ns
i=1
S
r
ij
[X
i
] (Re. 2)
where S
f
and S
r
are the matrices of forward and reverse stoichiometric
coeﬃcients, respectively, k
f
j
and k
r
j
are the corresponding reaction rate
constants of reaction j, and [X
i
] is the molar concentration of species i in the
cell. The matrix of stoichiometric coeﬃcients consists of N
s
rows, with the
rows corresponding to species. The columns represent reactions, making the
matrix N
s
×N
r
. The reaction rate constant k is itself a function of the
Arrhenius constants:
k
j
= A
j
T
βj
e
−
E
a,j
RT
(3.9)
which need to be speciﬁed as part of the mechanism. It is now possible to
write the equation for the reaction rate of the basic reaction, Re. 1. The rate
of formation of species [X
1
] from reaction j is written as:
_
d[X
1
]
dt
_
j
= S
r
1j
_
k
f
j
Ns
i=1
[X
i
]
S
f
1j
−k
r
j
Ns
i=1
[X
i
]
S
r
1j
_
(3.10)
This equation is formulated for every species included in the chemical
mechanism, as well as for every reaction, resulting inn an equation system
28 CHAPTER 3. SPRAY MODELLING
consisting of N
s
×N
r
equations. As can be seen from the above equation, it is
a system of Ordinary Diﬀerential Equations (ODEs), which can be solved
coupled using an ODE solver, sequentially using a reference species technique
[52], or by an EulerImplicit method. OpenFOAM has the ability to solve the
equations using an ODE solver (an approach used in Paper 3). Some versions
of KIVA3V also oﬀer the option of solving the chemical rate reactions
coupled, one of them being the KIVA3V version also used in Paper 3. Aside
from the concentrations, it is also important to ﬁnd the right hand side of
equation 3.10, since it is used in the source term in the transport equation
(equation 3.2). The source term for species i is:
RR
i
=
W
i
ρ
Nr
j=1
(S
r
ij
−S
f
ij
) ˙ ω
j
(3.11)
˙ ω
j
= k
f
j
Ns
i=1
[X
i
]
S
f
ij
−k
r
j
Ns
i=1
[X
i
]
S
r
ij
(3.12)
As mentioned, a real mechanism for hydrocarbon fuels (including model fuels)
would be very large. While it is possible to use such mechanisms in shock tube
simulations, they are too large to use in a CFD simulation. As mentioned in
the above section, each reaction requires a term describing the amounts of
each species it forms, and each species requires a transport equation to be
solved. One example of a suﬃciently small mechanism is the reduced one used
in Paper 3, consisting of 338 reactions and 83 species. This mechanism has to
solve 83 systems of ODEs in every cell, in every timestep. Some of the
equations in the system will not be applicable, since no reaction includes all
species, but in theory the system could comprise 28054 ODEs in total.
To solve the chemical reaction equations, a stiﬀ ODE solver is needed. The one
used in the studies presented in Paper 3 is the SIBS method (SemiImplicit
Bulirsch Stoer [53]) in OpenFOAM, which is based on Richardson
extrapolation of the approximated solution. As mentioned earlier the solver
needs to solve ODEs in every timestep to determine the chemical species’
concentrations at the end of the timestep. The Richardson extrapolation of the
function y assumes that as the interval (in our case the computational timestep
∆t) is split up into increasing number of substeps, the solution will converge
to some value y
∞
. However, the solver will never apply enough substeps to
ﬁnd it. Instead, it will approximate it, depending on the solution using large
substeps. The analytical function used to approximate y
∞
is a polynom, and
the error function of the method contains only even terms of the step size.
Chalmers PaSR Model
It is necessary to use some form of treatment for the chemistry and turbulent
mixing, in the literature there are many suggestions for doing this, including:
PaSR ([54]), diﬀusion ﬂamelets [55], detailed chemical kinetics with a
perfectlystirred reactor [56], and the general ﬂame surface density model [57].
3.1. THE GAS PHASE 29
However, the Chalmers PaSR model was used to model the turbulence 
chemistry interaction in Paper 3.
The PaSR model is based on the theory that real ﬂames are much thinner
than any computational cell, so assuming that an entire cell is a perfect
reactor is a severe overestimation. Thus, the cells are divided into a reacting
part, and a nonreacting part. The reacting part is treated like a perfectly
stirred reactor, in which all present species are homogeneously mixed and
reacted. After reactions have taken place, the species are assumed to be mixed
due to turbulence for the mixing time τ
mix
, and the resulting concentration
gives the ﬁnal concentration in the entire, partially stirred, cell. The relative
sizes of the parts of the computational cell constituting the reactor and the
rest of the cell, are governed by the turbulent mixing time and the residence
time (the numerical time step, τ, in our case). The reaction rate term for
species i is then approximated as:
∂c
i
∂t
=
c
i
1
−c
i
0
τ
= κRR
i
(c
i
1
) (3.13)
Where RR
i
(c
i
1
) is the laminar chemical source term, and κ the reaction rate
multiplier, deﬁned as:
κ =
τ
c
τ
c
+τ
mix
(3.14)
where τ
mix
is the turbulence timescale mentioned above, and τ
c
the
corresponding chemical timescale. This factor appears in the species transport
equation with the chemical source term (as in equation 3.2):
∂ρY
i
∂t
+∇(ρY
i
U) −∇(µ
Eﬀ
∇(Y
i
)) =
= S
i
+κRR
i
(c
i
) (3.15)
where τ
mix
is the turbulence timescale mentioned above, it is assumed to be
determined by k and ε:
τ
mix
= C
mix
k
ε
(3.16)
with C
mix
= 0.03. The chemical timescale is determined by solving the
reaction system’s ODEs fully coupled, and ﬁnding the characteristic time for
that system. Other variations and detailed derivations of the PaSR model can
be found in [52, 54, 58, 59].
30 CHAPTER 3. SPRAY MODELLING
3.2 Spray SubModels
3.2.1 Spray Motion Equation
The motion of a Lagrangian particle, moving in an Eulerian framework, is
governed by one of the most fundamental laws of physics; Newton’s second law:
∂p
d
∂t
=
i
F
i
(3.17)
Unfortunately, this equation cannot be solved immediately, simply because the
force acting on the parcel is unknown. The full spray equation, often referred
to as the BBO equation  from Basset (1888), Boussinesq (1903) and
Oseen(1927)  includes eﬀects of added mass, pressure, Basset force
2
, Magnus
eﬀect
3
, Saﬀman force
4
, and Fax´en
5
force. They are all neglected. Most of the
terms can be neglected due to the high density ratio between the two phases,
while others like the Magnus eﬀect are neglected since rotation of the droplets
will not be very important. What remains are the forces due to drag and
gravity acting on the droplets. The latter is included for simplicity and the
former due to its physical importance. These simpliﬁcations allow us to write
the right hand side of equation 3.17 as:
∂p
d
∂t
= −ρ
g
πD
2
8
C
D
(u
d
−U) u
d
−U +ρ
d
πD
3
6
g (3.18)
This equation can be simpliﬁed further, if we assume the droplets are
spherical, and that the drag will not be aﬀected by changes in mass, then:
∂p
d
∂t
= m
d
∂u
d
∂t
= ρ
d
πD
3
6
∂u
d
∂t
(3.19)
A droplet’s mass will of course change with time (due to breakup &
evaporation), but we assume that the changes will not aﬀect the drag. This
means we can neglect the eﬀect that causes rockets to liftoﬀ and airplanes to
ﬂy, an eﬀect that is not believed to be pronounced in diesel droplets from which
some mass is evaporating. If we combine equations 3.18 and 3.19, the result is:
∂u
d
∂t
= −
3
4
ρ
g
ρ
d
1
D
C
D
(u
d
−U) u
d
−U +g (3.20)
To simplify this equation further, a momentum relaxation time is introduced:
2
This is the force that causes a motorboat to continue in its path even when the motor has
been switched oﬀ recently.
3
The force that causes a soccer ball to rotate perfectly past defenders and into the net
4
Lift force due to shear
5
force due to curvature of the ﬂow
3.2. SPRAY SUBMODELS 31
τ
u
=
8m
d
πρ
g
C
D
D
2
u
d
−u
=
4
3
ρ
d
D
ρ
g
C
D
u
d
−u
(3.21)
Equation 3.20 can now be written as:
∂u
d
∂t
= −
u
d
−U
τ
u
+g (3.22)
Even with the introduction of τ
u
, the drag cannot be approximated without a
drag coeﬃcient, which will be introduced in a later section in this chapter.
3.2.2 Parcel Tracking
In Lagrangian spray simulations, the particles representing liquid are moving
in a ﬁxed Eulerian framework as described above. Tracking them and deﬁning
the cells they are in are clearly important issues. There are two main ways of
doing this, one called the LoseFind algorithm, and one called the
FaceToFace algorithm.
The Loseﬁnd algorithm is fairly simple and can be described in four steps.
1. Update the properties of the parcel
2. Move the parcel with velocity u
d
for time ∆t
3. Find out which cell the parcel is in
4. Add momentum to the cell to which the parcel has moved
Assuming that the time step is suﬃciently small to ensure that every parcel
traverses more than one cell, this approach is acceptable. However, even for a
small time step, a parcel can still be moved across a cell without having any
momentum exchange with it. The model is used in KIVA3V, as well as some
commercial CFD codes. Due to the problems mentioned above, however,
another model is used in OpenFOAM; FaceToFace tracking. It too can be
described in four steps:
1. Move the parcel until it reaches a cell boundary or for the
entire time step ∆t if it remains in the same cell
2. If the parcel changes cell, calculate the time it took to move
out of the ﬁrst cell, and update the parcel properties
3. Add the momentumchange to the cell that the parcel has
been in
4. If the parcel still has time left to move, go back to 1.
The FaceToFace tracking in OpenFOAM includes a stability check that
begins by tracking the parcel from the centre of the cell it belongs to, rather
than from the particle’s position. This is done to ensure that particles that
might be close to the edge of the cell (or even slightly outside) are properly
tracked.
32 CHAPTER 3. SPRAY MODELLING
Parcels tracked by FaceToFace tracking cannot ’skip’ cells, which improves
the predictions of transfer of mass, momentum and energy. For a more detailed
description, see [60].
3.2.3 Injection Model
The injection model used in the Lagrangian spray simulations is a solidcone
injection model. The user supplies a drop diameter probability density
function (PDF) with parameters, in this case a RosinRammler form was
chosen. The model also requires minimum and maximum values for the droplet
size, as well as the maximum spray cone angle. The parameters for the model
can be found in table 3.2
Model parameter Numerical Value
d
min
10
−6
m
d
max
d
nozzle
d
mean
d
nozzle
n 3
β
max
20
Table 3.2: Solid cone injection model constants
The direction of a droplet to be injected into the domain is calculated by
multiplying β
max
and a random number between 0 and 1. This angle is the
angle between the set spray direction and the direction of the injected droplet.
Note that β
max
is not the halfangle of the spray, but the full spray angle. The
velocity of the injected parcel is based on the injection pressure, and the
pressure in the domain:
u
d
= C
d
¸
2∆p
ρ
i
(3.23)
where ρ
i
is the density of the injected parcel, taking the type of fuel, injection
pressure, and composition of fuel (for multicomponent fuels) into account. C
d
is the discharge coeﬃcient, a parameter that varies from nozzle to nozzle, and
is to a certain degree pressuredependent [15].
3.2.4 Drag Model
Several options for the drag coeﬃcient have been suggested in the literature
([36],[61],[62]), but the one chosen here was:
C
D
=
_
_
_
24
Re
d
_
1 +
Re
2
3
d
6
_
Re
d
< 1000
0.44 Re
d
> 1000
(3.24)
3.2. SPRAY SUBMODELS 33
OpenFOAM also oﬀers the possibility to account for changes in drag due to
oscillations of the droplet surface, i.e. the instability waves eventually resulting
in droplet breakup. If this model is used, the TAB model [9] will be used to
calculate the oscillations and the resulting drag will be added to the drag
described above. The modiﬁcation is
C
D
= C
D,model
(1 +C
D,distort
min(y
lim
, y)) (3.25)
where C
D,distort
and y
lim
are userset constants, and y is the relative deviation
from the equator of the parcel if it were spherical. y
lim
is usually set to 1.0.
The minimum function is not necessary when the TAB breakup model is used,
since a yvalue of 1.0 will result in the parcel breaking up.
3.2.5 Breakup Model
The breakup model used is the KelvinHelmholtzRayleighTaylor (KHRT)
model ([8] & [63]). This model, along with the TAB model, is one of the most
widely used models in Lagrangian spray simulations today ([64],[65],[66]). The
KHRT model was chosen for its superior performance under diesel conditions
([67]). It is also possible to use the TAB model, but often in conjunction with
some form of atomization model since the TAB model tends to break up the
droplets very rapidly. The KHRT model includes two modes of breakup: KH
breakup, accounting for unstable waves growing on the liquid jet due to the
diﬀerences in velocity between the gas and liquid; and RT breakup, accounting
for waves growing on the droplets’ surface due to acceleration normal to the
dropletgas interface. Analysis of a round liquid jet yields a fastest growing
wave with a wavelength (Λ
KH
) and a growth rate (Ω
KH
):
Λ
KH
= 9.02
r(1 + 0.45Oh
1
2
)(1 + 0.4Ta
0.7
)
(1 + 0.865We
1.67
)
0.6
(3.26)
Ω
KH
=
(0.34 + 0.38We
3
2
)
(1 + Oh)(1 + 1.4Ta
0.6
)
_
σ
ρ
d
r
3
(3.27)
The critical droplet radius r
c
, which is the radius a parcel will have after it has
been stripped oﬀ the liquid jet, is assumed to depend linearly on the
wavelength Λ
KH
and the stripping rate, or characteristic breakup time τ
KH
(which controls the breakup rate) is assumed to be a simple function of the
growth rate Ω
KH
, wavelength Λ
KH
and droplet radius, r:
r
c
= B
0
Λ
KH
(3.28)
τ
KH
=
3.726B
1
r
Ω
KH
Λ
KH
(3.29)
The continuous reduction of droplet size due to breakup can then be deﬁned
by the rate equation:
34 CHAPTER 3. SPRAY MODELLING
dr
dt
= −
r −r
c
τ
KH
(3.30)
which essentially describes how rapidly the drops will reach the stable state.
The breakup parameters B
0
and B
1
can be found in Table B.1. In addition to
these two constants, there are two more parameters to control the KH
breakup. The ﬁrst determines whether breakup will occur at all. If
We < We
limit
, the drop has reached a stable state and no breakup due to
Kelvin Helmholtz instabilities will occur. The second parameter controls the
rate of introduction of new childparcels. As the jet breaks up the parcel will
accumulate a stripped mass, m
s
, and when m
s
exceeds a certain percentage of
the average parcel mass, which is the total injected liquid mass divided by the
total number of injected parcels, a new parcel will be introduced with size r
c
,
containing the accumulated mass m
s
.
RayleighTaylor breakup, on the other hand, is governed by how rapidly
disturbances grow on the surface of the droplet. For real droplets these
disturbances originate from the droplets’ trailing edges [12]. If these
disturbances are assumed to be linear, the frequency of the fastest growing
wave, and the corresponding wavelength, is given by:
Ω
RT
=
¸
2g
t
(ρ
l
−ρ
g
)
1.5
3
√
3σ(ρ
g
+ρ
l
)
(3.31)
g
t
= (g +
du
d
dt
) ·
u
d
u
d

(3.32)
Λ
RT
= 2πC
RT
¸
3σ
g
t
(ρ
l
−ρ
g
)
(3.33)
Where C
RT
is a modelling parameter. If this wavelength of the fastest growing
wave is smaller than the diameter of the droplet diameter and the
perturbations are allowed to grow for a period of time, the droplet will be
immediately converted into a parcel with smaller droplets when the growth
time exceeds the characteristic RT time. The new size is proportional to the
wavelength Λ
RT
of the disturbances. Thus, C
RT
determines both whether or
not the droplets are going to be converted into smaller drops and how large
the new droplets are going to be. The RT mode does not add any new parcels
to the system, it only increases the statistical number of droplets in the parcel
and changes the radii of the droplets.
3.2.6 Evaporation Model
When modelling the evaporation of a droplet, the key parameter is the
lifetime, or evaporation relaxation time of the droplet. The derivation of this
time starts by evaluating the timederivate of the mass of the droplet.
3.2. SPRAY SUBMODELS 35
dm
d
dt
= ρ
d
d
dt
_
4
3
π
_
D
2
_
3
_
=
ρ
d
π
2
D
2
dD
dt
(3.34)
Further, the evaporation is assumed to follow the empirical D
2
law:
dD
2
dt
= C
e
⇒
dD
dt
=
C
e
2D
(3.35)
where C
e
is a constant. From this equation we can determine the relaxation
time by integration (since C
e
is a constant).
D
2
= C
e
t +D
2
o
(3.36)
If we want to determine the time (τ
e
) when a droplet of size D
0
has completely
evaporated, we can put D = 0 into the above equation, resulting in:
τ
e
= −
D
2
0
C
e
(3.37)
The only unknown in this equation is the constant C
e
, which will be
determined by inserting equation 3.35 into equation 3.34, the result is:
dm
d
dt
=
ρ
d
π
4
DC
e
(3.38)
However, the time derivative of the mass of the droplet can also be expressed
as ([52]):
dm
d
dt
= −πDShDρ
v
ln
_
p −p
v,∞
p −p
v,s
_
=
= −πDShDρ
v
ln
_
1 +
X
v,s
−X
v,∞
1 −X
v,s
_
(3.39)
Using these two equations for
dm
d
dt
, the value of the constant C
e
can be
expressed as:
C
e
= −4Sh
ρ
v
ρ
d
Dln
_
1 +
X
v,s
−X
v,∞
1 −X
v,s
_
(3.40)
The Sherwood number Sh is calculated using the RanzMarshall (or Fr¨ ossling)
correlation (from [51]):
Sh = 2 + 0.6Re
1/2
Sc
1/3
(3.41)
36 CHAPTER 3. SPRAY MODELLING
Equation 3.37 can now be used with known C
e
to solve equation 3.39:
dm
d
dt
= −
m
d
τ
e
(3.42)
where the relaxation time is
τ
e
=
ρ
d
D
2
6DShρ
v
ln
_
1 +
Xv,s−Xv,∞
1−Xv,s
_ (3.43)
and a similar equation for the diameter (from 3.35):
dD
dt
= −
D
3τ
e
(3.44)
A relaxation time is introduced to characterise the evolution of the particle
size. If an explicit method is used, the time step (a function of the relaxation
timescale) must not be larger than the relaxation time, or the mass will
become negative. If an implicit method is used, the method is unconditionally
stable. It is also consistent with the way the momentum and heat transfer
equations for each parcel are solved. The diﬀerence in timescales of these
processes is useful for determining the rate at which each process is proceeding.
3.2.7 Heat Transfer
The heat transfer model comes in part from [51], and is based on the
convective heat transfer of a particle with a uniform temperature. A
modiﬁcation was made to account for the latent heat phase transfer during
evaporation. The heat transfer model also uses the evaporation relaxation
time, since the evaporation of a droplet also transfers heat to the surrounding
gas.
dT
d
dt
=
πDκ
c
Nu
m
d
c
l
(T
g
−T
d
) f
heat
−
1
c
l
h
v
(T
d
)
τ
e
(3.45)
The modiﬁcation is in the factor f
heat
which is:
f
heat
=
−
cp,v ˙ m
d
πDκcNu
_
e
−
cp,v ˙ m
πDκcNu
−1
_ (3.46)
The ﬁrst part of equation 3.45 accounts for the heat transfer to the liquid, and
the second part for latent heat arising from mass transfer. Since the mass
transfer is related to the evaporation of the droplet, the relaxation time
introduced in 3.2.6 is used. The heat transfer can also be expressed with a
relaxation time:
3.2. SPRAY SUBMODELS 37
τ
h
=
m
d
c
l
πDκ
c
Nu
(3.47)
resulting in:
dT
d
dt
=
(T
g
−T
d
)
τ
h
f −
1
c
l
h
v
(T
d
)
τ
e
(3.48)
where one term describes heat transfer due to temperature diﬀerence, and the
other heat transfer due to mass of the droplet lost through evaporation.
3.2.8 Particle Collisions
OpenFOAM oﬀers two collision models: the O’Rourke model [68], and the
modiﬁcation by Nordin [52]. Other models have also been published, for
instance the collision model by Schmidt et al [69], using adaptive collision
meshing, which involves generating a separate orthogonal collision mesh with
random orientation in every time step. The mesh is reﬁned until the number of
parcels in each cell is below a set threshold.
In the simulations presented in the included papers and the results chapter the
trajectory collision model by Nordin [52] was used (unless otherwise stated).
This is based on the O’Rourke collision model, which will be brieﬂy described.
Other modiﬁcations have also been made to the O’Rourke model [70]. In the
original version, parcels collide if they are in the same cell with a probability
given by:
e
−min(N1,N2)
1
4V
cell
π(d1+d2)
2
U1−U2∆t
(3.49)
In practice, a random number between 0 and 1 is generated for each parcel. If
the number is lower than the probability, the parcels are set to collide. The
outcome of the collision depends on the Weber number, resulting in either
coalescence or grazing. Thus, the direction of the parcels has no inﬂuence on
whether they collide or not, nor does their position within the cell. If the cells
are large, parcels several droplet radii away will be able to collide, although
the probability of this will be lower, according to equation 3.49 (due to the
large cell size). At the other extreme, there will be few collisions in very small
cells, since parcels will not often occupy the same cell. In all, the O’Rourke
model has very strong grid dependency, as shown in [52, 69].
As mentioned previously, the model implemented into OpenFOAM used for
the work presented in the appended papers is a modiﬁed form of the O’Rourke
model. Instead of only colliding parcels that are in the same cell, the algorithm
calculates the paths of all parcels and collides those that will intersect within
the same timestep. Thus the parcels need to be in the same position,
regardless of the mesh. Since a simulated spray may contain several thousand
parcels, the algorithm ﬁrst calculates whether a pair of parcels is heading
38 CHAPTER 3. SPRAY MODELLING
towards each other. If they are not, they are regarded as not colliding. This
reduces the computational load of the algorithm, and is done by computing:
v
Align
=
(U
1
−U
2
) · (x
2
−x
1
)
x
2
−x
1

(3.50)
This property will give how much of the vector x
2
−x
1
can be projected onto
the vector U
1
−U
2
, if it is positive it indicates that the parcels are heading
towards one another, if it is negative they are travelling away from each other,
and if it is zero, they are on a perpendicular course. The parcels also need to
be close enough to collide since, as mentioned before, they need to reach each
other within the current timestep, this gives the next criterion:
x
2
−x
1
 −
d
1
+d
2
2
< v
Align
∆t (3.51)
On the left hand side of this equation is the distance between the droplet
surfaces, and on the right hand side the parcels’ relative displacement during
the current timestep, which must be larger. This ensures that the parcels can
reach each other within the given time (the numerical timestep). The question
that remains to resolve is whether or not the parcels reach the intersection at
the same time. If, for instance, one parcel has a high velocity and the other a
low velocity, they may miss each other. So the algorithm needs to determine
the times (designated t
α
and t
β
) at which the parcels reach the intersection,
this is done by solving:
x
1,new
= x
1
+tU
1
x
2,new
= x
2
+tU
2
(3.52)
∂x
2,new
−x
1,new

∂t
1
(t = t
α
) = 0
∂x2,new−x1,new
∂t2
(t = t
β
) = 0 (3.53)
Whether the parcels actually collide or not is governed by chance, the
probability for a collision is now given by:
P
coll
=
_
1
2
(d
1
+d
2
)
max(
1
2
(d
1
+d
2
), x
1,new
(t
α
) −x
2,new
(t
β
))
_Cspace
e
−
c
time
tα−t
β

∆t
(3.54)
where x
1,new
−x
2,new
 is the smallest distance between the two parcels during
the timestep. As this value declines, the probability of collision increases. The
two model constants C
time
and C
space
control the collision rate in time and
space, respectively.
Chapter 4
Results & Discussion
4.1 Simulations of a Diesel Flow in a
Cavitating Model Nozzle
Paper 2 describes simulations of the ﬂow inside a model nozzle using a new
code for cavitation implemented in OpenFOAM. The code, which was
described in section 2.4, was unstable and obtaining the results in this section
required extensive work. The simulations were run on a grid representing the
experimental nozzle shown in ﬁgure 4.1. The inlet pressure was set to 10 MPa,
and the outlet pressure was varied between 1.5 MPa and 5.1 MPa.
Figure 4.1: View of the experimental setup, the centre of the image contains the
1 mm model injector. Before and after the nozzle, there is a 1 cm wide channel.
The ﬁrst converged results displayed a shockwave zigzag pattern developing
through the channel. As the boundary conditions were set in a more
appropriate way, and the geometry was improved by adding more cells before
and after the channel, the following cavitation pattern was obtained:
39
40 CHAPTER 4. RESULTS & DISCUSSION
Figure 4.2: Image of the computed average vapour fraction, injection pressure
7 MPa
The pattern obtained, shown here for an injection pressure of 7 MPa, was very
smooth. It should also be noted that this pattern has a scale ranging between
0.5 and 1 to allow the reader to distinguish between the two phases, as the
case is barely cavitating. The cavitation bubble in this case was so stable, that
calculating an average made hardly any diﬀerence. However, it did match
pressure curves acquired experimentally [71]. After considering this image, a
change of convection scheme was introduced. Instead of using the limited
linear scheme, as done previously, the monotonized central scheme [72] was
used. A routine was also implemented for calculating the cavitation
probability, which the experimental data showed, instead of the mean vapour
fraction as displayed in ﬁgure 4.2. Unfortunately, the change of scheme not
only introduced instabilities in the wall cavitation bubble, it also produced
four oblong cavitation bubbles in the centre of the domain (see ﬁgure 4.3) due
to strong vortices.
These vortices could be removed by introducing a slip condition on the upper
and lower walls. The physical rationale for the slip condition was that the
boundary layer along the two walls is extremely thin, and a noslip condition
would cause the size of the layer to be overestimated, and thus produce
suﬃciently strong vortices to cavitate the ﬂuid. After the slip condition was
introduced, the results illustrated in ﬁgure 4.4 were obtained, which compare
well to experimental results (ﬁgure 4.5). A comparison of the experimental
mass ﬂow rate and the calculated mass ﬂow rate is also shown, in ﬁgure 4.6. It
is very important for the code to be able to capture the maximum mass ﬂow
as the ﬂow cavitates, since this is a fundamental property of nozzle ﬂow.
4.1. SIMULATIONS OF ADIESEL FLOWINACAVITATING MODEL NOZZLE41
Figure 4.3: Mean isosurface for a vapour fraction of 0.1, velocity ﬁeld in a plane
perpendicular to the ﬂow direction shown to visualize the vortices, with noslip
condition applied to both upper and lower walls. The image shown is taken as if
the viewer is situated in the nozzle looking at the ﬂow going downstream with
the cavitation bubbles on the left and right sides.
Figure 4.4: Image of computed cavitation probabilities in the channel section of
the model injector, injection pressure 8.5 MPa
42 CHAPTER 4. RESULTS & DISCUSSION
Figure 4.5: Image of experimental determined cavitation probabilities, injection
pressure 8.5 MPa
Figure 4.6: Modelled and experimentally determined mass ﬂows with injection
pressure of 4.9, 7 and 8.5 MPa
4.2. INJECTOR SIMULATION RESULTS 43
4.2 Injector Simulation Results
Results obtained using the cavitation singlephase code are presented in Paper
2. Since the liquidvapourgas code described in section 2.2 is not suﬃciently
stable to run for diesel conditions, it was not used in any of the published
studies. However, an example of the kind of results generated using the code is
shown in ﬁgure 4.7.
Figure 4.7: Cutplane of injector geometry, coloured according to velocity mag
nitude. Isosurface of ρ = 500kg/m
3
While the above simulation looks pretty, it requires a simulation time of over
200 days if run on a single processor (assuming perfect parallelization). The
simulation was run on 12 processors, thus reducing the time to 17 days. It
should also be noted that the mesh at 1.7 million cells is not suﬃciently ﬁne to
resolve the incylinder spray structures, but may be adequate for modelling
turbulence and cavitation inside the injector.
44 CHAPTER 4. RESULTS & DISCUSSION
4.2.1 Impact of Injection Model
As stated in the model description section, one objective for this project was
to develop a new injection model for EulerLagrangian diesel sprays. However,
since the available code for simulating a cavitating ﬂuid was not capable of
handling high pressures, no such model could be developed. Remaining of the
eﬀorts to develop an injection model is the presented code for cavitation and a
preliminary study which was made to answer questions like: How much
inﬂuence does the injector model have? Will the fueldistribution/ﬂame be
inﬂuenced by variations in the spray angle or droplet distribution and, if so,
how much? These questions were brieﬂy addressed using simulations with
several diﬀerent spray angles and droplet distributions. This was by no means
a full study, and the conclusions are tentative at best.
The numerical grid and physical data were the same as those used in Paper 3,
except that combustion was switched oﬀ for these simulations. Results of
varying the injection spray angle are shown in ﬁgure 4.8. There is a slight
change for the liquid phase, but the gas phase remains almost unchanged for a
doubled injection cone angle.
(a) 9 degrees in
jection cone an
gle
(b) 20 degrees in
jection cone an
gle
Figure 4.8: Comparison of fuel vapour distribution in a centre cutplane for a
low injection angle (9 degrees) and a large angle (20 degrees)
4.2. INJECTOR SIMULATION RESULTS 45
Another initial parameter for the spray is the initial droplet size. This
parameter was also varied, by varying the RosinRammler distribution (as
described in section 3.2.3) used. As shown by the images in ﬁgure 4.9, initial
parameters of the droplets also have quite a minor inﬂuence on the vapour
distribution. They seem to aﬀect the penetration of the fuel vapour core, but
not the overall vapour penetration. However, they strongly inﬂuence the liquid
penetration, which will have substantial eﬀects on combustion.
(a) dav. = 50 µm (b) dav. =
100 µm
(c) dav. =
200 µm
Figure 4.9: Comparison of fuel vapour distributions in a centre cutplane obtained
with three average initial diameter values
46 CHAPTER 4. RESULTS & DISCUSSION
4.3 Spray Results without Chemistry
In Paper 1, results of applying the turbulence lengthscale limiter (described in
section 3.1.3) are presented. Due to the constraints on the length of the article,
the eﬀects of the modiﬁcation on vapour penetrations on 2 and 3 dimensional
grids were mainly used to demonstrate the improvements it provided.
However, the eﬀect of the length scale limiter on liquid penetrations is also
beneﬁcial in terms of promoting grid independence, as will be shown in this
section. For details on grids, physical conditions etc., see Paper 1.
4.3.1 Grid Dependency
This section deals solely with the grid dependency of the simulations
conducted in 2D to show how the limiter decreases the dependency. The liquid
penetration obtained using the standard kε model is shown in ﬁgure 4.10
0 0.5 1 1.5 2
x 10
−3
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
l
i
q
u
i
d
(
m
)
dr
min
=0.1mm, lim on
dr=0.5mm, lim on
dr=1mm, lim on
dr
min
=0.1mm, lim off
dr=0.5mm, lim off
dr=1mm, lim off
Figure 4.10: Liquid penetration vs. time, obtained using three diﬀerent grids,
standard kε
When using the lengthscale limiter the calculations seem to exhibit grid
independency, while the calculations without the lengthscale limiter produce
diﬀerent shapes on the three grids. It can also be seen that increasing the cell
size results in more signiﬁcant deviations from the average. If we ﬁlter out the
oscillations, using a simple local ﬁlter, we can see the local average at each
point. This gives an idea of the average penetration for every timestep
sampled. The results can be seen in ﬁgure 4.11.
The large oscillations observed in the liquid penetration rate are due to the
liquid penetration being set to the furthest point reached by 95% of the liquid
mass in the system. If this was set to a lower value, for example 90%, the
results would be less oscillatory and possibly no ﬁltering would be needed.
4.3. SPRAY RESULTS WITHOUT CHEMISTRY 47
0 0.5 1 1.5 2
x 10
−3
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
l
i
q
u
i
d
(
m
)
dr
min
=0.1mm, lim on
dr=0.5mm, lim on
dr=1mm, lim on
dr
min
=0.1mm, lim off
dr=0.5mm, lim off
dr=1mm, lim off
Figure 4.11: Liquid penetration vs. time, obtained using three diﬀerent grids,
the standard kε model and ﬁltered penetrations
When the LaunderSharma kε model is used, the behaviour is slightly
diﬀerent, see ﬁgure 4.12. The penetration obtained with the coarsest grid
diﬀeres considerably from those obtained with the two ﬁner ones, indicating
that 3000 cells with a cell size of 1 mm may be too coarse. In Paper 1 similar
indications were shown for the vapour penetration. However, the penetration
across the three diﬀerent grids is more uniform with the limiter then without
it.
0 0.5 1 1.5 2
x 10
−3
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
l
i
q
u
i
d
(
m
)
dr
min
=0.1mm, lim on
dr=0.5mm, lim on
dr=1mm, lim on
dr
min
=0.1mm, lim off
dr=0.5mm, lim off
dr=1mm, lim off
Figure 4.12: Liquid penetration vs. time, obtained using three diﬀerent grids
and the Launder Sharma kε with no ﬁlter
48 CHAPTER 4. RESULTS & DISCUSSION
4.3.2 Time Step Dependency
The code uses a Courantbased timestep, which means the timestep changes
with the grid size and local velocity. This implies that one cannot draw
conclusions about grid independency, without also investigating the eﬀects of
diﬀerent Courant numbers. The results of a Courant number investigation are
shown in Figures 4.13 & 4.14. The limiter has a positive eﬀect on the liquid
penetration, but not as strong as its eﬀects on the vapour penetration, for
which it removed the dependency completely (Figures 4.15 & 4.16, taken from
Paper 1).
0 0.5 1 1.5 2
x 10
−3
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
l
i
q
u
i
d
(
m
)
Co=0.2, lim on
Co=0.1, lim on
Co=0.005, lim on
Co=0.2 , lim off
Co=0.1 , lim off
Co=0.005 , lim off
Figure 4.13: Liquid penetration vs. time, obtained using three diﬀerent Courant
numbers and the kε model with ﬁltering used
0 0.5 1 1.5 2
x 10
−3
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
l
i
q
u
i
d
(
m
)
Co=0.2, lim on
Co=0.1, lim on
Co=0.005, lim on
Co=0.2 , lim off
Co=0.1 , lim off
Co=0.005 , lim off
Figure 4.14: Liquid penetration vs. time, obtained using three diﬀerent Courant
numbers, the Launder Sharma kε model and ﬁltering used
4.3. SPRAY RESULTS WITHOUT CHEMISTRY 49
0 0.5 1 1.5 2 2.5 3
x 10
−3
0
0.02
0.04
0.06
0.08
0.10
0.12
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
v
a
p
o
u
r
(
m
)
limiter on
limiter off
Figure 4.15: Vapour penetration vs. time, obtained using three diﬀerent Courant
numbers and the standard kε model
0 0.5 1 1.5 2 2.5 3
x 10
−3
0
0.02
0.04
0.06
0.08
0.10
0.12
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
v
a
p
o
u
r
(
m
)
limiter on
limiter off
Figure 4.16: Vapour penetration vs. time, obtained using three diﬀerent Courant
numbers, and the LaunderSharma kε model
50 CHAPTER 4. RESULTS & DISCUSSION
4.3.3 Comparison to Experimental Data
To compare simulations using the turbulence length scale limiter to
experiments, an extensive range of injection and bomb pressures were used.
The results of that study was presented in Paper 1, and only one case will be
presented here for reference.
0 0.5 1 1.5
x 10
−3
0
0.02
0.04
0.06
0.08
0.1
0.12
time (s)
p
e
n
e
t
r
a
t
i
o
n
o
f
v
a
p
o
r
(
m
)
limit on, vap
limit off, vap
exp Vapour
Figure 4.17: Vapour penetration vs. time for a 3D case comparing experimental
to simulated data. Injection pressure, 70 MPa; bomb pressure, 4.4 MPa.
4.4. INFLUENCE OF NUMERICAL SCHEME 51
4.4 Inﬂuence of Numerical Scheme
In numerical simulations of diesel sprays, using a grid with more, smaller cells
does not necessarily increase the accuracy of the simulations. Instead, the best
grid for a speciﬁc case is found on an empirical basis. As shown in Paper 4, the
same is true for the numerical scheme, and to some degree, the initial
conditions for the turbulence. The numerical scheme referred to here is the
scheme for the advection term (for a more detailed description of how the
scheme is used when solving the momentum equations, see Appendix A.2).
The numerical schemes examined in Paper 4 were members of the total
variation diminishing (TVD) family of schemes, which were initially developed
to handle large gradients, originally in compressible high Mach number ﬂows.
However, they can also be applied with success to other cases, such as diesel
sprays where there are high gradients between the spray and the surrounding
quiescent air.
Figure 4.18: Fuel vapour distributions of three sprays, obtained using linear
interpolation (left), upwind interpolation (centre) and MC interpolation (right).
Grey scale altered (compared to the one in ﬁgure 4.19) to ensure that the fuel
distribution is visible for all cases. The LaunderSharma kε model was used for
modelling turbulence. Time, 1 ms after SOI. Initial µ
T
= 10
−5
m
2
/s.
As can be seen, the diﬀerences in vapour distribution obtained using diﬀerent
numerical schemes are quite large. Since MC interpolation uses both upwind
and linear interpolation, it mixes the properties of the two solutions using the
simpler schemes. MC yields as long penetration as upwind, while retaining
some irregularities like linear interpolation. The diﬀerence between the
schemes was found to be larger when a lower turbulence viscosity was set,
indicating that a high viscosity smoothes the solution to such a degree that no
instabilities or high gradients are present (see Figure 4.19). This is also, in a
sense, an illustration of how the TVD limiter works. When there are no
gradients, the TVD limiter will revert to a linear scheme, and the choice of
TVD limiter will not have an eﬀect on the solution.
52 CHAPTER 4. RESULTS & DISCUSSION
Figure 4.19: Fuel vapour distributions of three sprays, obtained using linear
interpolation (left), upwind interpolation (centre) and MC interpolation (right).
The LaunderSharma kε model was used for modelling turbulence. Time, 1 ms
after SOI. Initial µ
T
= 3.6 · 10
−4
m
2
/s.
4.5. FUEL SPRAYS WITH CHEMISTRY 53
4.5 Fuel Sprays with Chemistry
In Paper 3, simulations of events in the Sandia [73] constantvolumevessel
were presented. The focus was on diﬀerences between KIVA3V and
OpenFOAM predictions of ignition delay time and ﬂame liftoﬀ length, and
how well the codes could calculate these parameters compared to experimental
data. The simulations were performed on a cubic mesh with 108 mm sides for
the following conditions: ambient gas oxygen concentrations, 821 %; ambient
gas temperatures, 7501300 K; and ambient gas density, 14.8 kg/m
3
. The
injector oriﬁce diameter was 100 µm, and the injection pressure diﬀerence was
set to 1500 bar.
As can be seen from ﬁgures 4.20 through 4.23, OpenFOAM can predict the
ignition delay time and its dependence on ambient temperature and ambient
oxygen concentration fairly well. This indicates of that the initial evaporation
and chemical kinetics are “correct”. Unfortunately, the ﬂame liftoﬀ length
predictions are not as good. OpenFOAM seems to overpredict the changes in
liftoﬀ length caused by variations oxygen concentrations and ambient
temperature. KIVA3V calculations use a set atomization length, as can be
seen from ﬁgure 4.20, where the liftoﬀ length is stable above 1100 K.
OpenFOAM’s liftoﬀ length is also stable for these temperatures, but this is
because the ﬂame cannot be any closer to the wall (due to the constant
temperature condition set along the walls).
The time sequence in ﬁgure 4.24 displays how the ignition starts and ﬂame
propagates in OpenFOAM, KIVA3V and experiments. The ﬂame in
OpenFOAM and KIVA3V diﬀers in two main respects. In OpenFOAM (and
the experiments) the ﬂame ignites and then propagates upstream towards the
injector, whereas in KIVA3V the ﬂame ignites closer to the injector and then
the ﬂame liftoﬀ length propagates downstream. The second main diﬀerence is
that OpenFOAM’s ﬂame is very long and thin compared to KIVA3V’s wider,
shorter ﬂame. The long ﬂame generated by OpenFOAM penetrates much too
far, according to the experimental data. In the simulations Paper 4 is based
upon, this was found to be due to the numerical scheme and turbulence model
chosen. RNG kε is the kε turbulence model in OpenFOAM that produces the
longest ﬂames, and the limited linear scheme in OpenFOAM was found to
promote long, thin ﬂames and fuel distributions. These results, and the
quantitative results in ﬁgures 4.204.23 are extracted from Paper 3.
The computational time was much higher in OpenFOAM compared to
KIVA3V. The case with 21% oxygen took up to 70 hours on 4 processors
using OpenFOAM, about 4 times the computational time for KIVA3V. The
main reason for the high computational time is solving the chemical reaction
rate equations. In OpenFOAM they are solved fully coupled using the SIBS
ODE solver as described in section 3.1.4. If the ODE solver for the chemistry
can be optimized to reduce the computational time without losing accuracy is
an object for future work.
In addition to the two codes diﬀering in computational time, there are
numerous other diﬀerences in the solver in two codes. OpenFOAM uses a
54 CHAPTER 4. RESULTS & DISCUSSION
collocated grid arrangement, whereas KIVA3V uses a segregated grid
arrangement. OpenFOAM uses spray submodels which are the same by name,
but the way they are implemented in KIVA3V is diﬀerent. The spray
momentum coupling, chemical solver, and turbulence model are also diﬀerent.
These diﬀerences do not make the comparisons presented here and in Paper 3
irrelevant, since it was made to compare the two codes in terms of
performance and ability to match experimental data. A full comparison of the
two codes should be made starting with simulation of a simple model ﬂow,
where complexity such as compressibility, spray, spray submodels, chemistry
etc. is added stepbystep to see the eﬀects of each model.
800 1000 1200 1400
0
20
40
60
80
Ambient Temperature (K)
L
i
f
t
−
o
f
f
L
e
n
g
t
h
(
m
m
)
Measured
Kiva−3V
OpenFOAM
Figure 4.20: Comparison of OpenFOAM and KIVA3Vpredicted and measured
liftoﬀ lengths for diﬀerent ambient gas temperatures, at ambient gas oxygen
concentration and density of 21% and 14.8 kg/m
3
, respectively.
5 10 15 20 25
0
20
40
60
Ambient Oxygen Concentration (%)
L
i
f
t
−
o
f
f
L
e
n
g
t
h
(
m
m
)
Measured
KIVA−3V
OpenFOAM
Figure 4.21: Comparison of OpenFOAM and KIVA3Vpredicted and measured
liftoﬀ lengths for diﬀerent ambient gas oxygen concentrations at an ambient
gas temperature and density of 1000 K and 14.8 kg/m
3
, respectively.
4.5. FUEL SPRAYS WITH CHEMISTRY 55
5 10 15 20 25
0
0.5
1.0
1.5
2.0
Ambient Oxygen Concentration (%)
I
g
n
i
t
i
o
n
D
e
l
a
y
T
i
m
e
(
m
s
)
Measured
OpenFOAM
Figure 4.22: Comparison of OpenFOAMpredicted and measured ignition de
lay times for diﬀerent ambient gas oxygen concentrations at an ambient gas
temperature and density of 1000 K and 14.8 kg/m
3
, respectively.
600 800 1000 1200 1400
0
1
2
3
Ambient Temperature (K)
I
g
n
i
t
i
o
n
D
e
l
a
y
T
i
m
e
(
m
s
)
Measured
OpenFOAM
Figure 4.23: Comparison of OpenFOAMpredicted and measured ignition delay
times for diﬀerent ambient gas temperatures at an ambient gas oxygen concen
tration and density of 21% and 14.8 kg/m
3
, respectively.
56 CHAPTER 4. RESULTS & DISCUSSION
750 1210 1675 2140 2600
Temperature (K)
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
t = 0.5 ms t = 0.8 ms
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
t = 1.1 ms t = 1.3 ms
PSfrag replacements
Temperature
PSfrag replacements
Temperature
PSfrag replacements
Temperature
t = 1.5 ms
Figure 4.24: Time sequences of simulated temperature distributions obtained
using the OpenFOAM code (top) and KIVA3V code (middle) showing igni
tion and ﬂame development of an nheptane spray in the Sandia highpressure,
hightemperature vessel: ambient oxygen concentration, 21 %; ambient temper
ature, 1000 K; and ambient density 14.8 kg/m
3
. Corresponding, experimentally
acquired OH chemiluminescence images are shown in the bottom part of each
subﬁgure. The length of each image from simulations corresponds to 108 mm,
the experimental images are of the same scale, but extending 60 mm.
4.5. FUEL SPRAYS WITH CHEMISTRY 57
4.5.1 Combustion Predictions Using Other Schemes
In Paper 3, the limited linear scheme was used for the advection term, and the
RNG kε model for turbulence. As mentioned in the previous section, these
two choices were quite unfavourable for obtaining similar (short) ﬂames to
those observed in experiments. They did, however, provide reasonably close
ﬂame liftoﬀ lengths (which was the reason for the choices). The “best”
1
choices for advection scheme and turbulence (of those tested in the studies
presented in Paper 4) were found to be Superbee and LaunderSharma kε,
respectively. It was also found that the initial turbulent viscosity should be
slightly higher. These numerical conditions were used to examine whether
good comparisons with data in the Sandia database of experiments could be
maintained when combustion was considered in the simulation.
750 1210 1675 2140 2600
Temperature (K)
t = 0.5 ms t = 0.7 ms
t = 0.9 ms t = 1.1 ms
Figure 4.25: Comparison of fuel distributions obtained from experiments and
simulations tuned to match the experimentally determined fuel vapour distri
bution for the noncombusting case. Ambient gas temperature, 1000 K; oxygen
content, 0%; density, 14.8 kg/m
3
.
Comparing Figure 4.25 and 4.26, showing images obtained from simulations
excluding combustion and including combustion, respectively, it can be clearly
seen that while OpenFOAM predicts the vapour distribution in the displayed
part of the vessel accurately, it does not predict the combustion and (more
speciﬁcally) the ﬂame liftoﬀ length so well. Aside from the wellpredicted
vapour distribution, the chemical mechanism used has been validated against
shock tube data. These two validations reduce the number of possible reasons
for the discrepancy in the ﬂame liftoﬀ length. Unfortunately, however, the
1
In terms of matching experimental fuel vapour distribution of a noncombusting spray.
58 CHAPTER 4. RESULTS & DISCUSSION
750 1210 1675 2140 2600
Temperature (K)
t = 0.5 ms t = 0.8 ms
t = 1.1 ms t = 1.3 ms
Figure 4.26: Comparison of OH chemiluminescence images obtained from ex
periments and temperature distributions from simulations tuned to match ex
perimental fuel vapour distribution for a noncombusting case. Ambient gas
temperature, 1000 K; oxygen content, 21%; density, 14.8 kg/m
3
.
experimental data for the noncombusting case only covers the area 17 mm
downstream of the nozzle and beyond. Hence, the data for the ignition site
and early atomization zone cannot be validated. There are several possible
remaining reasons for the diﬀerence between the experimental and simulated
data; one is that the eﬀects of air entrainment and airfuel mixing might not
be correct, i.e. the PaSR model might need improvement or the parameters
used in it might be incorrect (most likely the turbulence mixing time). If our
chemistryturbulence model predicts perfect mixing in a position where it does
not occur, the ﬂame liftoﬀ length could be underpredicted. This should also
result in underprediction of ignition delay times, which is consistent with
observations. OpenFOAM, like most other codes, assumes a Lewis number of
unity for species diﬀusion, and this according to Hilbert et al [74] (page 94,
ﬁgure 17 in [75]), can lead to substantially larger errors than use of more
detailed approximations. Another set of unknowns that have already been
mentioned in part are the liquid atomization parameters, including the liquid
droplet distribution parameters, the penetration of the droplets before they
evaporate, and the initial shape/particle distribution of the fuel cloud. A good
4.6. COMMENT ON THE PASR MODEL 59
parameter to use for comparison would be the fuel liftoﬀ length for a
noncombusting case. The liquid penetration for the noncombusting case is
reasonably well predicted by the simulations, but there is no way to verify how
well it is predicted for the combusting case with currently available
experimental data.
Finally, it is possible that the temperature dependence may be incorrect in
some of the models. Unfortunately, data for only one noncombusting case
were available at the time of these simulations, so it was not possible to verify
that the vapour distribution or liquid length is correct at other temperatures
and densities.
4.6 Comment on the PaSR Model
During the course of the work reported in Paper 3, it was observed that the
PaSR model only had minor eﬀects on the solution. The model was described
in detail in section 3.1.4, but some of its properties will be repeated here for
the reader’s convenience. The parameter that will be highlighted in the ﬁgure
below is κ, which determines the combusting and quiescent proportions of a
cell, which depend in turn on local chemical and turbulent timescales. The
chemical timescale is determined by the timescales of each reaction and does
not contain any “tuning” constants. The timescale of the turbulence is
diﬀerent, and the way it should be calculated is an open question. In [54] the
mixing timescale was calculated as:
τ
mix
=
_
1
Re
T
_
1/4
k
ε
= 0.178
k
ε
(4.1)
Other versions are from [59], in which the kolmogorov timescale is used, and
[52] in which a modelling constant C
mix
is used:
τ
mix
= C
mix
k
ε
(4.2)
This is the expression used in this work. In [52] the value of C
mix
was reported
to be 0.005, while the one I used for combustion calculations was 0.03. It was
found that the simulations are almost independent of the value of C
mix
, as
long as it remains lower than 1 (in fact, major diﬀerences have not been
observed until the value reaches 1000). For the nheptane mechanism used in
this work, the chemical timescale is thus always several times larger than the
turbulent timescale. Naturally, the values of k and ε also aﬀect the result, but
for spray simulations ε has an average of the order of 10
6
m
2
/s
3
, while k is of
the order of 100 m
2
/s
3
. An image of the values of κ in a combusting spray is
shown in ﬁgure 4.27. The elevation is by temperature, to visualize the extent
of the spray. The scale should be noted, since it ranges between 0.97 and 1.0.
Obviously, the partially stirred reactor becomes a perfectly stirred reactor in
60 CHAPTER 4. RESULTS & DISCUSSION
this case. However, whether or not this is due to unrealistic chemical
timescales of the reactions involved, is not known at this time.
Figure 4.27: Display of a spray simulation (corresponding to the one shown in
ﬁgure 4.26), coloured according to κ and elevated according to temperature. κ
is the mass fraction of the cell that is involved in the chemical reactions. The
area outside of the elevated temperature area has a κ value of zero.
Chapter 5
Conclusions
A code describing how a ﬂuid cavitates when it is subjected to low pressures
due to geometric contraction was successfully used to simulate data acquired
from model nozzle experiments performed at AVL. The thesis shows that there
are many pitfalls on the way to simulations that provide good comparisons to
experimental events, as well as presenting some ﬁnal results from the study.
Examples of the degree of inﬂuence the injection model can have over the
EulerLagrangian spray simulations are also presented, together with an
example of results obtained using the threephase code (mainly for illustrative
purposes since this code has not been validated). These simulations were, as
stated in the outline of the thesis (section 1.4), intended to be used for the
development of an injection/atomization model for EulerLagrangian diesel
spray simulations. Unfortunately, due to the high complexity of the problem,
no such model could be developed. For a more detailed description of the
encountered problems when simulating cavitating diesel injections, see section
4.1.
Regarding spray grid and timestep independence, the results from Paper 1
and section 4.3 clearly show that dependence diminishes when a limiter on the
length scale of the turbulence is applied. The importance of this result cannot
be overemphasised, since shorter computational time means more
investigations can be made, or results can be obtained in a shorter time. In
short, it improves the eﬃciency of the CFD modelling.
The results also show how ﬂame liftoﬀ length and ignition delay time can be
predicted by OpenFOAM, and the eﬀects of varying the ambient oxygen
concentrations and temperatures on these parameters. OpenFOAM was found
to perform well when predicting the propagation of the ﬂame compared to the
ignition site, and in ignition delay time. However, it was found to underpredict
the ﬂame liftoﬀ length and its dependency on temperature and ambient
oxygen concentration. For a noncombusting case in the same geometry,
OpenFOAM could predict the vapour distribution very well, in addition to the
liquid and vapour penetration lengths.
61
62 CHAPTER 5. CONCLUSIONS
To further analyze the reason for the short ﬂame liftoﬀ lengths yielded by
OpenFOAM, the calculations and results of Siebers et al [76] were used. The
cited report describes a way to estimate the percentage of stoichiometric air
entrained (ξ) upstream of the liftoﬀ length. According to this estimate, the
percentage of entrained air is 29% of the air required for stoichiometric
combustion, which is consistent with the article’s relationship between
injection velocity and ξ. Furthermore, according to the report, the liquid
droplets and ﬂame should be separated in distance much more from the
injector than they are in ﬁgure 4.26. If the turbulence/chemistry interaction
model used required more air to be entrained before combustion, the ﬂame
liftoﬀ length might be increased without increasing the length of the ﬂame.
As indicated in section 4.6, the PaSR model does not suﬃce for this task.
The choice of scheme can have a large inﬂuence on the fuel vapour
distribution, and thus on the resulting ﬂame and combustion. This was
demonstrated in paper 4, and also in sections 4.4 and 4.5.1. These eﬀects are
important to be aware of, in addition to the eﬀects of numerical grid and
turbulence model. In the work this thesis was based upon, the SuperBee TVD
limiter was found to give the best results compared to experiments, but this
ﬁnding could be (like the choice of grid) problemdependent.
Chapter 6
Future Work
As shown in this thesis, diesel spray simulations using the EulerLagrangian
approach are helpful, but they are not yet predictive. Therefore, much work
needs to be done to improve them. Brieﬂy, two possible approaches could be
adopted in attempts to improve diesel spray simulations: EulerEuler type
simulations, which have only been used for primary breakup studies to date; or
more sophisticated EulerLagrangian descriptions. Following the second
approach, implies use of LES to model gas phase turbulence and better
submodels for the liquidgas interactions. Another option is the VSB2 spray
model developed by Karlsson [77], in which the Lagrangian parcels/blobs do
not interact with the grid directly, but through bubbles of varying size.
Furthermore, the Lagrangian parcels contain a distribution of blobs instead of
equally sized droplets. Details of this setup are not known at this time, but
might be published in the future.
In the nearer future, possible areas of future work are to continue eﬀorts
started in the studies published in Papers 3 and 4. As shown in the Results
section, while the models presented in Paper 4 delivered fuel vapour
distributions that approximated well to experimentally observed distributions,
the resulting ﬂames were not well predicted when chemistry was considered.
Since we can now assume that the spray and vapour distributions are well
tuned, the parts of the code that are largely responsible for the errors must be
the turbulence/chemistry interaction (PaSR model), temperature dependency
of the liquid penetration, primary breakup and initial vapour cloud formation,
or the chemical mechanism itself. However, the chemical mechanism has been
veriﬁed against shocktube experiments, and seems to produce good
comparisons for those cases. A way to rule out temperature dependency would
be to run simulations and compare them to the results of corresponding
experiments, without oxygen using diﬀerent ambient temperatures. The
validation presented in Paper 3 is only against one case, due to a lack of data
from noncombusting experiments with welldeﬁned boundary conditions to
compare to the simulations.
63
64 CHAPTER 6. FUTURE WORK
When the results presented in Paper 3 are to be improved (beyond the
improvement presented in this thesis), a large number of simulations where the
chemical reactions are considered will have to be run. It would then be
desirable to have optimized the solver for the chemical mechanism, as the
computational time occupied by the solver in OpenFOAM is currently very
high (as mentioned in section 4.5).
It would also be interesting to venture further down the path taken in Paper 2,
investigating in detail how the initial droplet parameters inﬂuence the spray
and combustion. The most important of these parameters, based on the short
preliminary study, seems to be the initial droplet size distribution. It is likely
that a good atomization and injection model would improve the ﬂame liftoﬀ
length predictions presented in Paper 3, since the liquid penetration length has
a strong inﬂuence on the ignition point and its propagation. These two
approaches are linked in the sense that they both are concerned with the
primary events of spray development. Simulating cavitating diesel ﬂow
requires a more robust model than the one used in Paper 2, and perhaps it is
not even essential to the development of a new injection model. There are
studies ([78, 79]) which links injection pressure, nozzle diameter, to spray
angle, discharge coeﬃcient etc. If these studies were used like the cavitation
simulations were intended to be used, a diesel injection model could be
developed based on experimental studies instead of numerical simulations.
Other aspects not covered in the thesis could be of interest for future work.
One such aspect is the obvious extension from constant volume vessel
simulations to engine simulations. OpenFOAM has the functionality to handle
moving meshes with valves and other topological changes [80], so using the
work described in Papers 3 and 4 as a basis for engine simulations to explore
the eﬀects of key variables in engines could be of high value to both the engine
community and the numerical community.
Aside from using kε models and LES it could also be useful to assess the
possible utility of other turbulence models such as kω type models, unsteady
RANS models or other hybrid type turbulence models. It is possible that the
turbulence plays a major role in obtaining correct air entrainment for
modelling the early spray development and ﬂame liftoﬀ lengths.
6.1 Outlook
As mentioned above, it would be helpful to have access to an experimental
database covering a larger array of noncombusting cases than the single case
from Sandia. It would also be beneﬁcial for research in numerical methods to
have more data on the initial breakup of sprays for both combusting and
noncombusting cases. A good example of a study that presented relevant
data, in this respect, is one by Ochoterena et al [81], in which both liquid and
vapour penetration lengths are presented along with ﬂame liftoﬀ lengths, cone
angles and ﬂame temperatures. Unfortunately, from a modeller’s perspective,
the fuels used are not as easy to model as, for instance, nheptane, and the
experiments were only run in combusting conditions. However, it should be
6.1. OUTLOOK 65
noted that supplying modellers with input for simulations was not one of the
stated objectives of the cited study ([81]).
In light of the increasing need for more experimental data speciﬁcally tailored
for validating numerical simulations, the author belives and strongly
recommends that more experiments tailored for numerical simulations should
be performed. Close, longterm cooperation between a modeller and an
empirical investigator with access to both spray chamber and engine equipment
would have every chance of making major advances in spray modelling.
Furthermore, the need for good spray modelling will probably increase as
engine concepts become more advanced and new fuels are introduced, to help
ensure that demands from customers and legislation are met.
Finally, we return to the starting point of the thesis; horses in early 20th
century America. During this period, commuters were facing rising prices for
hay and oats, as well as increased public fears about pollution and traﬃc jams
[1]. This situation is uncomfortably similar to current circumstances.
Hopefully, the solution we devise this time will not lead to the same problems
100 years from now.
66 CHAPTER 6. FUTURE WORK
Chapter 7
Summary of Papers
The thesis is based on the work published in the following papers
Paper I  Numerical Investigation of
Mesh/Turbulence/Spray Interaction for Diesel
Applications
This paper originated from attempts to solve the problem of grid dependence
described in section 1.4. The solution was based on a suggestion in the book
“Modeling Engine Spray and Combustion Processes” by Gunnar Stiesch [12],
p.183, that the turbulence length scale should be limited to the jet diameter
within the jet region. The rationale for this is that the jet is the driving force
for the turbulence in spray simulations. This was implemented in the
OpenFOAM code, and tested on diﬀerent sized grids in 2D and 3D, as well as
using diﬀerent timesteps. It was shown in the article that the length scale
limiter was very beneﬁcial in terms of reducing grid and timestep dependency.
The results were also compared to data acquired in experiments performed at
Chalmers on fuel injection into a highpressure, hightemperature constant
volume vessel. The multicomponent model diesel fuel IDEA was used as the
fuel. The simulations using the turbulence length scale limiter proved to
provide better matches to the experimental data than when it was turned oﬀ.
The most important ﬁnding was that the grid and time step dependency could
be reduced by introducing the length scale limiter.
The simulations and paper were planned together with Dr Niklas Nordin, who
also implemented the limiter in the code. I was responsible for the preliminary
studies that were not included in the paper, to determine limiter size, grid
optimization etc. I was also responsible for running and postprocessing all but
one of the simulations included in the paper, the remaining one was performed
by Dr Niklas Nordin. The experimental data used were acquired by AlfHugo
Magnusson at Chalmers University of Technology. The paper was written by
67
68 CHAPTER 7. SUMMARY OF PAPERS
me, together with Dr Niklas Nordin, and I presented it at the SAE Fuel and
Lubricants Conference, 2005, in Rio de Janeiro, Brazil.
Paper II  Modelling Injector Flow Including
Cavitation Eﬀects for Diesel Applications
This paper was preceded by several attempts to model diesel injector ﬂow.
The initial intention was to develop and present a new injection model,
together with validation of the code for the injector ﬂow, in the same article.
However, simulating injector ﬂow proved to be much more complex than
anticipated, so a decision was taken to focus solely on the validation in this
paper, while the injector model would be the topic of a possible later paper.
The code used in this paper was written by OpenCFD Ltd., as a byproduct of
a code to simulate diesel injector ﬂow with the ambient air as well as
cavitating diesel liquid.
I setup, ran and postprocessed the simulations, with numerical help from
Henry Weller, who developed the code used. However, as these simulations
required extensive testing, I ran months of simulations, with various numerical
setups to ensure a stable solution, before the article was written. The
empirical data used for comparison were acquired from experiments performed
at AVL. I wrote the paper with inputs from Dr Niklas Nordin and Henry
Weller. In addition, I presented it at the 5th Joint ASME/JSME Fluids
Engineering Conference, 2007, San Diego, USA.
Paper III  Three Dimensional Simulation of
Diesel Spray Ignition and Flame LiftOﬀ Using
OpenFOAM and KIVA3V CFD Codes
The paper originated from a comparison of OpenFOAM results I generated to
KIVA3V results generated by Dr Feng Tao. Thanks to the large ECN
database, it was also possible to compare the results to experimental data. We
compared the eﬀects of varying amounts of oxygen/EGR and ambient
temperatures using the two codes. There was also a limited comparison of the
eﬀects of using diﬀerent grids and injection positions (published in the
Appendix of the article). It was found that OpenFOAM produces much longer
ﬂames than KIVA3V, as well as short ﬂame liftoﬀ lengths. It is possible that
KIVA3V yields longer ﬂame liftoﬀ lengths since the atomization length is
prescribed in the code. The OpenFOAM simulations did not use an
atomization model, and thus yielded much shorter liquid penetrations.
Unfortunate use of the RNG kεturbulence model together with the limited
Linear scheme is believed to be an underlying reason for the long penetration
of the ﬂame. With the knowledge gained from the studies published in Paper
IV, I believe the results would have been better if SuperBee interpolation had
been used for the advection scheme.
69
I setup, ran and postprocessed the OpenFOAM simulations, while the
KIVA3V simulations were setup, ran and postprocessed by Dr Feng Tao. The
paper was written by me and Dr Feng Tao. I presented it at the SAE World
Congress, 2008, Detroit, USA.
Paper IV  Inﬂuence of Advection Schemes In
The Prediction of Diesel Spray and Fuel Vapour
Distributions
As mentioned regarding the previous paper, the authors were not satisﬁed by
OpenFOAM’s overprediction of ﬂame lengths. Simulations indicated that the
scheme might have a strong inﬂuence on the fuel vapour distribution.
Therefore, a more extensive study of the inﬂuence of the scheme (and eﬀects of
the initial parameters on spray development) was undertaken. It was found
that while the scheme does have a large inﬂuence, the initial turbulence
parameters can reduce or enhance its inﬂuence. If the initial viscosity of the
turbulence is low, the SuperBee scheme was shown to be the best choice for
interpolation of the advection scheme.
I setup, ran and postprocssed the simulations, and wrote the paper together
with Dr Feng Tao. It was presented at ILASS, 2008, Como Lake, Italy.
70 CHAPTER 7. SUMMARY OF PAPERS
Appendix A
Implementation in
OpenFOAM
A.1 RhieChow Interpolation in OpenFOAM
OpenFOAM is relatively new, open source code that is not fully documented.
Therefore the intention of the author here is to extend the documentation, and
at the same time share some additional knowledge that has been gained during
the course of this work. One very important aspect of CFD is RhieChow
correction, which will therefore be described here. RhieChow correction is
absolutely necessary for ﬂow simulations using a collocated grid, since it
removes oscillations in the solution, which occur if the pressure gradient does
not depend on the pressure in adjacent cells, thus allowing a “jigsaw” pattern
to develop.
Some basic information about the notation in OpenFOAM is needed to
describe how it works. Partial Diﬀerential Equations in OpenFOAM are solved
by setting up the appropriate matrix system:
fvm::operation(coefficient,U) (A.1)
This means that the solver is to solve for U, and coefficient is a coeﬃcient
which can vary in space and/or time, or be a constant. Operations like this
will be described many times throughout this section, and it is important to
understand the nomenclature. Further, this section contains two velocities: U,
which is used for the ﬂux and held constant when the momentum equations
are solved; and U
x,y,z
, which is the velocity we solve for. All equations
containing U
x,y,z
are actually three equations, referring to relationships in the
x, y, and zdirection respectively.
Rhie and Chow is usually regarded as a correction proportional to the
diﬀerence between the pressure gradient at the face and the interpolated
71
72 APPENDIX A. IMPLEMENTATION IN OPENFOAM
pressure gradient at the face. For example (from Ferziger & Peric [82]), for a
cell face, the velocity is corrected by:
u
j
= u
j
−∆
_
1
A
uj
p
_
_
∂p
∂x
j
−
_
∂p
∂x
j
_
_
(A.2)
where the overbar indicates interpolation, and ∆ is related to the mesh size.
OpenFOAM does not have such an explicit term in the equations, which
makes it diﬃcult to see how the RhieChow correction is applied. Indeed,
OpenFOAM applies a correction that some would call ’in the spirit of Rhie
and Chow’. For this description, the icoFoam code from OpenFOAM will be
used to illustrate how this correction is implemented. icoFoam is a solver for
transient incompressible laminar ﬂow of Newtonian ﬂuids. For this type of
ﬂow, the momentum equation in vector form would be:
∂U
x,y,z
∂t
+ (U· ∇)U
x,y,z
−∇· (ν∇U
x,y,z
) = −
1
ρ
∇p (A.3)
Where the above equation is solved for U
x,y,z
in each of the three directions. If
we consider the UEqn.H in icoFoam, the convection term is slightly diﬀerent.
fvVectorMatrix UEqn (A.4)
(
fvm::ddt(U)
+ fvm::div(phi, U)
− fvm::laplacian(nu, U)
);
Equation A.4 has no right hand side, and there is a ﬁeld named phi (φ)
instead of the Uterm. φ is (for incompressible ﬂows) the volume velocity ﬂux
deﬁned at the faces of each cell. It is used because OpenFOAM utilizes the
Gauss theorem, which is frequently used in applied mathematics, and deﬁnes
the transform of a volume integral into a surface integral. For an
incompressible case, equation A.3 is integrated over a volume V with face S,
and the second term can be rewritten as:
_
V
∇· (UU
x,y,z
) dV =
_
S
(UU
x,y,z
)
f
· ˆ ndS =
=
i
U
f,i
x,y,z
(U
f,i
· S
f i
) =
i
φ
i
U
f,i
x,y,z
(A.5)
A.1. RHIECHOW INTERPOLATION IN OPENFOAM 73
where
φ = U
f,i
· S
f
(A.6)
U
f,i
is the velocity that will be held constant when the equation for pressure
is solved, while U is the vector velocity that will be solved for. It is important
to note the diﬀerence in the subscript when the surface integral is introduced;
subscript f indicates that the term should be evaluated on the face. φ is
deﬁned as the scalar product of the cell face velocity and the cell face normal
(eq. A.6). The magnitude of the cell face normal is the cell face area.
We can now connect OpenFOAM’s equation to its analytical counterpart:
_
V
∇· (UU
x,y,z
) dV =⇒fvm::div(phi,U) (A.7)
In the above equation, U is turned into phi (by equation A.6). U
x,y,z
is the
velocity we solve for, which is turned into U in OpenFOAM. There is also a
diﬀerence in that phi and U
f
x,y,z
are evaluated on the surfaces of the cell, while
U is deﬁned in the cell centre. The way in which the face and cell values are
obtained from each other is discussed in section A.2.
We are now ready to return to equation A.4, which does not include the
pressure gradient (thus contributing to the separation of pressure and
velocity). The code adds the inﬂuence of p to U in another way, which is
described below.
Recalling from [82], the discretized momentum equation would be this matrix
system:
A[U
x,y,z
] = H−∇[p] (A.8)
A is the numerical equivalent of the operators in equation A.3 operating on
U
x,y,z
, and H the one containing terms without U
x,y,z
. They are created by
issuing the commands UEqn.A() and UEqn.H(), respectively. [.] denotes the
numerical approximation of the corresponding variable. This system is solved
by dividing by A, which results in:
[U] =
H
A
−
1
A
∇[p] (A.9)
Equation A.9 cannot be solved at this stage, since we have not updated the
pressure yet. If the problem of interest includes transport of a scalar property,
such as enthalpy or the mass fraction of a chemical compound, we need a
predictor for U, which is calculated using the pressure from the previous
iteration:
solve(UEqn ==  fvc::grad(p)); (A.10)
74 APPENDIX A. IMPLEMENTATION IN OPENFOAM
This is called a momentum predictor of U. So far only the momentum
equation has been used, but we also have the continuity equation, which is
used to create an equation for pressure. The ﬁrst term on the right hand side
of equation A.10 is needed for the pressure equation, and given a special
notation:
U
*
= UEqn.H()/UEqn.A(); (A.11)
The velocity from A.11 does not satisfy continuity, and lacks inﬂuence of
pressure. To create the equation for pressure, we take the divergence of
equation A.9:
∇· [U] = ∇· ([U
∗
]) −∇·
_
1
A
∇[p]
_
(A.12)
The left hand side is zero for incompressible ﬂows, since it represents the
correct velocity. Equation A.12 is then:
∇· ([U
∗
]) = ∇·
_
1
A
∇[p]
_
(A.13)
The left hand side will be treated explicitly, since we now need to ﬁnd the
pressure, and thus keep the velocity constant. In OpenFOAM, this is done by
utilizing fvc instead of fvm. The velocity U
∗
is replaced by the velocity ﬂux φ,
since the face velocities will be used to evaluate the term. The resulting
equation is then:
surfaceScalarField phi =
fvc::interpolate(U) & mesh.Sf() (A.14)
volScalarField rUA = 1.0/UEqn.A(); (A.15)
fvScalarMatrix pEqn (A.16)
(
fvm::laplacian(rUA, p) == fvc::div(phi)
);
A feature that cannot be seen in this formulation is that OpenFOAM again
uses the Gauss theorem. Thus, it is not necessary to calculate a second
derivative of p, but only a ﬁrst derivative. This derivative is needed on the cell
faces, and it is evaluated by using the cell centre values of the pressure.
When the PISO loop is ﬁnished, the velocity is corrected with the correct
pressure gradient.
1
1
Actually this is computed at the end of each PISO loop, before the boundary conditions
are corrected. The code used as an example here is icoFoam, but the method is the same in
most of the other codes available in OpenFOAM.
A.1. RHIECHOW INTERPOLATION IN OPENFOAM 75
U = rUA*fvc::grad(p); (A.17)
The gradient in equation A.17 is again evaluated using the Gauss theorem,
and therefore no gradient calculation is necessary. Instead, only the pressure
on the cell face is needed.
To summarize, there are four important points regarding the RhieChow
correction in OpenFOAM.
1. phi does not include any eﬀect of pressure when
solving the continuity equation
2. rUA does not include any eﬀect of pressure when
solving for continuity and the ﬁnal velocity corrector
3. The Laplacianterm of p uses the value of the gradient of
p on the cell face. The gradient is calculated using
neighbouring cells, not neighbouring faces
4. The gradient term of p is calculated from the cell face
values of p.
This method gives an oscillationfree velocity “with the spirit of RhieChow”,
even though there is no explicit RhieChow correction.
76 APPENDIX A. IMPLEMENTATION IN OPENFOAM
A.2 Formulating the Momentum Equation
Matrix in OpenFOAM
In Paper 4 there is a concise description of the way in which diﬀerent schemes
for diﬀerent terms can be used in OpenFOAM. This will be explained in more
detail here, from equation to matrix. We start oﬀ by simplifying equation A.3,
by assuming stationary incompressible nonviscous ﬂow:
U· ∇U
x
= ∇· (UU
x
) = −
1
ρ
∂p
∂x
(A.18)
U· ∇U
y
= ∇· (UU
y
) = −
1
ρ
∂p
∂y
(A.19)
U· ∇U
z
= ∇· (UU
z
) = −
1
ρ
∂p
∂z
(A.20)
(A.21)
where the equations are rewritten using the continuity equation. In
OpenFOAM they would be written (according to the RhieChow method
explained in the previous section) as:
fvVectorMatrix UEqn (A.22)
(
fvm::div(phi, U)
);
Rho is not included in the above equation because OpenFOAM uses p =
p
ρ
for
many incompressible codes instead of actual p, although this is only visible in
the dimension on initial p. For simplicity, we will only consider ﬂow in the
ydirection. This saves space, as we now only have one equation:
∇· (UU
y
) = −
1
ρ
∂p
∂y
(A.23)
This term is integrated over the cell volumes used for the ﬁnite volume
method, yielding:
_
V
∇· (UU
y
) dV =
_
S
(U
y
)
f,scheme
(U)
f,linear
· ˆ ndS
=
i
(U)
i
f,linear
· S
i
f
(U
y
)
i
f,scheme
= −
_
V
∂p
∂y
dV (A.24)
A.2. FORMULATINGTHE MOMENTUMEQUATIONMATRIXINOPENFOAM77
Where all the variables are evaluated on face i, belonging to the cell in
question with volume V and surface S. Two interpolations are then required to
calculate the term on the lefthand side: one for (U
y
)
i
f,scheme
, which is the
velocity to be solved for; and one for (U)
i
f,linear
, the velocity that is kept
constant as the linear equation system is solved. As indicated by the subscript,
the second velocity is usually interpolated by linear interpolation. The
faceﬂux, (phi in OpenFOAM) is either massﬂux in compressible code, or
volumeﬂux in incompressible code. Using anything other than linear for this
term could cause problems later when the divergence term is evaluated. For
instance, the faceﬂux on certain faces could be zero due to use of
upwind/downwind interpolation, and thus the TVD scheme for the divergence
term would not perform as well.
Figure A.1: Line mesh consisting of six cells.
Following the nomenclature in ﬁgure A.1, we can write equation A.24 for cell
P:
(U)
e
linear
· S
e
. ¸¸ .
φe
(U
y
)
e
scheme
+ (U)
f
linear
· S
f
. ¸¸ .
φ
f
(U
y
)
f
scheme
= S
e
p
e
−S
f
p
f
(A.25)
If OpenFOAM used a staggered grid setup, the above equation could be solved
for the face values per se, and the only interpolation needed would be the
linear interpolations for the ﬂux term and pressure. Instead, OpenFOAM
solves for the cellcentred values directly, requiring more interpolation. The
general equation for ﬁrst/second order interpolation of the cellcentred
variable U to the face is given by
U
f
= λ
f
(U
P
−U
N
) +U
N
U
e
= λ
e
(U
M
−U
P
) +U
P
(A.26)
where λ
f
and λ
e
vary between 0 and 1 depending on the chosen scheme. If we
insert this expression for U
f
and U
e
into equation A.25, we get the equations
used to compile the linear matrix system for solving the equations. For cell P
this equation is:
78 APPENDIX A. IMPLEMENTATION IN OPENFOAM
φ
e
(λ
e
(U
M
−U
P
) +U
P
) +φ
f
(λ
f
(U
P
−U
N
) +U
N
) =
φ
e
λ
e
U
M
+ (φ
e
(1 −λ
e
) +φ
f
λ
f
)U
P
+φ
f
(1 −λ
f
)U
N
=
S
e
p
e
− S
f
p
f
(A.27)
This equation is valid for all cells, although special treatment is needed for the
boundary cells. The resulting matrix formed from this equation is then:
_
_
. . . . . . . . . . . . . . .
. . . φ
e
λ
e
φ
e
(1 −λ
e
) +φ
f
λ
f
φ
f
(1 −λ
f
) . . .
. . . . . . . . . . . . . . .
_
_
_
¸
¸
¸
¸
_
. . .
U
M
U
P
U
N
. . .
_
¸
¸
¸
¸
_
=
_
_
. . .
S
e
p
e
−S
f
p
f
. . .
_
_
(A.28)
which results in a sparse matrix having elements only on the sub,super and
main diagonal. This is the matrix that is calculated when the command
fvVectorMatrix UEqn is issued in OpenFOAM. It is usually solved by
ordinary LU decomposition. Note that this equation is solved for the
cellcentered values. As mentioned in the previous section, this equation is not
solved directly, but through the procedure described there.
Appendix B
Variable Deﬁnitions and
Dimensionless Numbers
S = ∇U : dev(sym(∇U)) (B.1)
f = −0.3e
−
„
min
»
“
ρk
2
µε
”
2
,50
–«
(B.2)
R
RNG
= η
−
η
η0
+ 1
βη
3
+ 1
(B.3)
η =
_
(S)
k
ε
(B.4)
Re
d
=
ρ
d
u
rel
r
µ
l
(B.5)
We =
ρu
rel

2
r
σ
(B.6)
Oh =
√
We
l
Re
l
(B.7)
Ta = Oh
_
We
g
(B.8)
Co =
U· d
d
2
∆t
(B.9)
Sc =
ν
D
v
(B.10)
79
80APPENDIXB. VARIABLE DEFINITIONS ANDDIMENSIONLESS NUMBERS
Model parameter Numerical Value
B
0
0.61
B
1
40.0
C
τ
1.0
C
RT
0.1
m
s,limit
0.4
We
limit
6
Table B.1: KHRT model constants
Bibliography
[1] Tarr, J.A. and McShane, C., 1997. “The centrality of the horse to the
nineteenthcentury american city”. In The Making of Urban America.
Scholarly Resources Inc., pp. 105–130.
[2] World Health Organization, 2003. “Health aspects of air pollution with
particulate matter, ozone and nitrogen dioxide”. EUR/03/5042688.
[3] K˚ageson, Per, www.bilsweden.se. Dieselmotorn p˚a kort och l˚ang sikt.
[4] www.dieselnet.com/standards/eu/ld.html, 2007. Emission standards.
[5] United States Environmental Protection Agency, 2007. LightDuty
Automotive Technology and Fuel Economy Trends: 1975 through 2007.
[6] Lagerquist, M., 2007. Energil¨ aget i siﬀror 2007, Energy in Sweden 2007
Facts and Figures.
[7] Patterson, M.A., Kong, S.C, Hampson, G.J. and Reitz, Rolf D., 1994.
“Modeling the Eﬀects of Fuel Injection Characteristics on Diesel Engine
Soot and NOx Emissions”. SAE Paper 940523.
[8] Patterson, M., and Reitz, R. D., 1998. “Modeling the eﬀects of fuel spray
characteristics on diesel engine combustion and emission”. SAE Paper
980131.
[9] O’Rourke, P.J., and Amsden, A.A., 1987. “The tab method for numerical
calculation of spray droplet breakup”. SAE Paper 872089.
[10] Weller, H., Tabor, G., Jasak, H., and Fureby, C., 1998. “A Tensorial
Approach to Computational Continuum Mechanics using Object
Orientated Techniques”. Computers in Physics, 12(6), pp. 620–631.
[11] Jasak, H., Weller, H. G. and Nordin, N., 2004. “InCylinder CFD
Simulation Using a C++ ObjectOriented Toolkit”. SAE Paper
2004010110.
[12] Stiesch, G., 2003. Modeling Engine Spray and Combustion Processes.
Springer Verlag Berlin Heidelberg New York.
[13] Faeth, G. M., Hsiang, LP and Wu, PK, 1995. “Structure and Breakup
Properties of Sprays”. International Journal of Multiphase Flow, 21,
pp. 99–127.
81
82 BIBLIOGRAPHY
[14] Badock, C., Wirth, R., Fath, A., and Leipertz, A., 1999. “Investigation of
cavitation in real size diesel injection nozzles”. International Journal of
Heat and Fluid Flow, 20(5).
[15] Winklhofer, E., Kull, E., Kelz, E., and Morozov, A., 2001.
“Comprehensive Hydraulic and Flow Field Documentation in Model
Throttle Experiments Under Cavitation Conditions”. In Proceedings of
the ILASSEurope Annual Meeting, Zurich, Switzerland, pp. 574–579.
[16] Arcoumanis, C., Gavaises, M., Flora, H., Roth, H., 2001. “Visualisation of
cavitation in diesel engine injectors”. Mecanique & Industries, 2(5).
[17] Blessing, M., K¨ onig, G., Kr¨ uger, C., Michels, U., and Schwarz, V., 2003.
“Analysis of Flow and Cavitation Phenomena in Diesel Injection Nozzles
and its Eﬀects on Spray and Mixture Formation”. SAE Paper
2003011358.
[18] Schmidt, D., Rutland, C., and Corradini, M., 1997. “A numerical study of
cavitating ﬂow through various nozzle shapes”. SAE Paper 971597.
[19] Yuan, W., and Schnerr, G., 2003. “Numerical Simulation of TwoPhase
Flow in Injection Nozzles: Interaction of Cavitation and External Jet
Formation”. Journal of Fluids Engineering, 125, pp. 963–969.
[20] Mulemane, A., Subramaniyam, S., Lu, P., Han, J., Lai, M., and Poola, R.,
2004. “Comparing Cavitation in Diesel Injectors Based on Diﬀerent
Modeling Approaches”. SAE Paper 2004010027.
[21] Martynov, S., Mason, D., Heikal, M. R., Sazhin, S. S., and Gorokhovski,
M., 2006. “Modelling of Cavitation Flow in a Nozzle and its Eﬀect on
Spray Development”. In International Heat Transfer Conference.
[22] Yuan, W., Sauer, J., and G.H., S., 2001. “Modeling and Computation of
Unsteady Cavitation Flows in Injection Nozzles”. Mechanique &
Industries, 2, pp. 383–394.
[23] Masuda, R., Fuyuto, T., Nagaoka, M., Berg, E., and Tatschl, R., 2005.
“Validation of diesel fuel spray and mixture formation from nozzle
internal ﬂow calculation”. SAE Paper 2005012098.
[24] Ning, W., 2007. “Development of a nextgeneration spray and
atomization model using an eulerianlagrangian methodology”. PhD
thesis, University of Wisconsin  Madison.
[25] Roth, H., Giannadakis, E., Gavaises, M., Arcoumanis, C., Omae, K.,
Sakata, I., Nakamura, M., and Yanagihara, H., 2005. “Eﬀect of
multiinjection strategy on cavitation development in diesel injector
nozzle holes”. SAE Paper 2005011237.
[26] Nouri, J., Mitroglou, N., Yan, Y., and Arcoumanis, C., 2007. “Internal
ﬂow and cavitation in a multihole injector for gasoline directinjection
engines”. SAE Paper 2007011405.
BIBLIOGRAPHY 83
[27] Payri, R., Margot, X., and Salvador, F., 2002. “A Numerical Study of the
Inﬂuence of Diesel Nozzle Geometry on the Inner Cavitating Flow”. SAE
Paper 2002010215.
[28] Schmidt, D.P., Rutland, C.J. and Corradini, M.L., 1999. “A Fully
Compressible TwoDimensional Model of High Speed Cavitating
Nozzles”. Atomization and Sprays, 9.
[29] Vortmann, C., Schnerr, G., and Seelecke, S., 2003. “Thermodynamic
Modeling and Simulation of Cavitating Nozzle Flow”. International
Journal of Heat and Fluid Flow, 24.
[30] Weller, H. G., 2006. A New Approach to VOFbased Interface Capturing
Methods for Incompressible, Compressible and Cavitating Flow. Tech.
Rep. TR/HGW/07, OpenCFD Ltd., October.
[31] Wallis, G. B., 1969. Onedimensional Twophase Flow. McGrawHill,
New York.
[32] Chung, M.S., Park, S.B., and Lee, H.K., 2004. “Sound speed criterion
for twophase critical ﬂow”. Journal of Sound and Vibration, 276,
pp. 13–26.
[33] Ubbink, O., 1997. “Numerical prediction of two ﬂuid systems with sharp
interfaces”. PhD thesis, Imperial College.
[34] Poinsot, T.J., and Lele, S.K., 1992. “Boundary conditions for direct
simulations of compressible viscous reacting ﬂows”. Journal of
Computational Physics, 101, pp. 103–129.
[35] Huh, K.Y., and Gosman, A.D., 1990. “Atomization Mechanism of Fuel
Injection”. ILASS.
[36] Kralj, C., 1995. “Numerical Simulation of Diesel Spray Processes”. PhD
thesis, Imperial College of Science Technology and Medicine.
[37] Peng, D.Y., and Robinson, D. B., 1976. “A new twoconstant equation of
state”. Industrial & Engineering Chemistry Fundamentals, 15, pp. 59–64.
[38] Xie, W. F., Liu, T. G., and Khoo, B. C., 2006. “Application of a oneﬂuid
model for large scale homogeneous unsteady cavitation: The modiﬁed
Schmidt model”. Computers & Fluids, 35, pp. 1177–1192.
[39] de Villiers, E., Gosman, A., and Weller, H., 2004. “Large Eddy Simulation
of Primary Diesel Spray Atomization”. SAE Paper 2004010100.
[40] Befrui, B., Corbinelli, G., Robart, D., and Reckers, W., 2008. “LES
Simulation of the Internal Flow and NearField Spray Structure of an
OutwardOpening GDi Injector and Comparison with Imaging Data”.
SAE Paper 2008010137.
[41] M´enard, T., and Berlemont, A., 2007. “Interface tracking with a coupled
level set/vof/ghost ﬂuid method: Application to jet atomization”.
Proceedings of FEDSM2007, FEDSM200737377.
84 BIBLIOGRAPHY
[42] Vuorinen, V., Larmi, M., and Fuchs, L., 2008. “LargeEddy Simulation on
the Eﬀect of Droplet Size Distribution on Mixing of Passive Scalar in a
Spray”. SAE Paper 2008010933.
[43] Blokkeel, G., Barbeau, B., and Borghi, R., 2003. “A 3D Eulerian Model
to Improve the Primary Breakup of Atomizing Jet”. SAE Paper
2003010005.
[44] Lebas, R., Blokkeel, G., Beau, P.A., and Demoulin, F.X., 2005.
“Coupling Vaporization Model With the EulerianLagrangian Spray
Atomization (ELSA) Model in Diesel Engine Conditions”. SAE
2005010213.
[45] Iyer, V. and Abraham, J., 2003. “An Evaluation of a TwoFluid
EulerianLiquid EulerianGas Model for Diesel Sprays”. Journal of Fluids
Engineering, 125, pp. 660–669.
[46] Jones, W. and Launder, B., 1972. “The prediction of laminarization with
a twoequation model of turbulence”. Int. J. Heat Mass Transfer, 15,
pp. 301–314.
[47] Yakhot, V. and Orzag, S.A., 1986. “Renormalization Group Analysis of
Turbulence”. I. Basic Theory J. Sci. Comput.
[48] Launder, B. and Sharma, B., 1975. “Application of the energy dissipation
model of turbulence to the calculation of ﬂow near a spinning disc”. Lett.
Heat and Mass Transfer, 1, pp. 131–138.
[49] Amsden, A.A., O’Rourke, P.J. and Butler, T.D., 1989. “Kivaii: A
computer program for chemically reactive ﬂows with sprays”.
LA11560MS, May.
[50] Gore, R., and Crowe, C. T., 1989. “Eﬀect of particle size on modulating
turbulent intensity”. International Journal of Multiphase Flow, 15(2),
pp. 279–285.
[51] Crowe, C., Sommerfeld, M., and Tsuji, Y., 1998. Multiphase Flows with
Droplets and Particles. CRC Press LLC.
[52] Nordin, N., 2001. “Complex chemistry modeling of diesel spray
combustion”. PhD thesis, Dept. of Thermo and Fluid Dynamics,
Chalmers University of Technology, G¨ oteborg.
[53] Press, W.H., Teukolsky, S.A., Vetterling, W.T. and Flannery, B.P., 2002.
Numerical Recipes in C++: The Art of Scientiﬁc Computing, 2:nd ed.
Cambridge University Press.
[54] Karlsson, J. A. J., 1995. “Modeling AutoIgnition, Flame Propagation
and Combustion in Nonstationary Turbulent Sprays”. PhD thesis,
Chalmers University of Technology, G¨ oteborg.
[55] Venugopal, R., and Abraham, J., 2007. “A numerical investigation of
ﬂame liftoﬀ in diesel jets”. Combustion Science and Technology, 179,
pp. 2599–2618.
BIBLIOGRAPHY 85
[56] P.K. Senecal, E. P., and Richards, K., 2003. “Multidimensional Modeling
of DirectInjection Diesel Spray Liquid Length and Flame Liftoﬀ Length
using CFD and Parallel Detailed Chemistry”. SAE 2003011043.
[57] Tap, F., and Veynante, D., 2005. “Simulation of Flame Liftoﬀ on a
Diesel Jet Using a Generalized Flame Surface Density Approach”.
Proceedings of Combustion Inst.(30), pp. 919–926.
[58] Chomiak, J. and Karlsson, J.A.J., 1996. “Flame Liftoﬀ in Diesel Sprays”.
TwentySixth Symposium (International) on Combustion/The
Combustion Institute, pp. 2557–2564.
[59] Tao, F., 2003. “Numerical Modeling of Soot and NOx Formation in
NonStationary Diesel Flames with Complex Chemistry”. PhD thesis,
Chalmers University of Technology, G¨ oteborg.
[60] Macpherson, G. B., Nordin, N., and Weller, H. G., 2008. “Particle
tracking in unstructured, arbitrary polyhedral meshes for use in cfd and
molecular dynamics”. Communications in Numerical Methods in
Engineering.
[61] Schmalzing, C.O, Stapf, P., Maly, R.R., Renner, G., Stetter, H., and
Dwyer, H.A., 1999. “A holistic hydraulic and spray model  liquid and
vapor phase penetration of fuel sprays in di diesel engines”. SAE Paper
1999013549.
[62] Kimura, S., Kosaka, H., Matsui, Y.,and Himeno, R., 2004. “A numerical
simulation of turbulent mixing in transient spray by les”. SAE Paper
2004012014.
[63] Reitz, R.D., 1987. “Modeling Atomization Processes in HighPressure
Vaporizing Sprays”. Atomization and Spray Technology, 3, pp. 309–337.
[64] Park, S. W., Kim, H. J., and Lee, C. S., 2003. “Investigation of
atomization characteristics and prediction accuracy of hybrid models for
highspeed diesel fuel sprays”. SAE Paper 2003011045.
[65] Rotondi, R., Bella, G., Grimaldi, C., and Postrioti, L., 2001.
“Atomization of HighPressure Diesel Spray: Experimental Validation of a
New Breakup Model”. SAE paper 2001011070.
[66] Bianchi, G.M., and Pelloni, P., 1999. “Modeling the diesel fuel spray
breakup by using a hybrid model”. SAE Paper 1999010226.
[67] Lee, JH and Goto, S., 2000. “Comparison of spray characteristics in
butane and diesel fuels by numerical analysis”. SAE Paper 2000012941.
[68] O’Rourke, P., 1981. “Collective Drop Eﬀects on Vaporizing Liquid
Sprays”. PhD thesis, Department of Mechanical and Aerospace
Engineering, Princeton University.
[69] Schmidt, D., and Rutland, C., 2000. “A New Droplet Collision
Algorithm”. Journal of Computational Physics, 164, pp. 62–80.
86 BIBLIOGRAPHY
[70] Gustavsson, J., and Golovitchev, V., 2004. “3D Simulation of Multiple
Injections in DI Diesel Engine”. In The Sixth International Symposium
on Diagnostics and Modeling of Combustion in Internal Combustion
Engines, COMODIA.
[71] Peng K¨ arrholm, F., 2006. Numerical Modelling of Diesel Spray Injection
and Turbulence Interaction. Licentiate Thesis, Department of Applied
Mechanics, Chalmers University of Technology.
[72] van Leer, B., 1977. “Towards the ultimate conservative diﬀerence scheme
iii. upstreamcentered ﬁnitediﬀerence schemes for ideal compressible
ﬂow”. J. Comp. Phys., 23, pp. 263–75.
[73] Sandia National Laboratories, 2007. Engine combustion network
(public.ca.sandia.gov/ecn).
[74] Hilbert, R., and Th´evenin, D., 2002. “Autoignition of turbulent
nonpremixed ﬂames investigated using direct numerical simulations”.
Combustion and Flame, 128, pp. 22–37.
[75] Hilbert, R., Tap, F., ElRabii, H., and Th´evenin, D., 2004. “Impact of
detailed chemistry and transport models on turbulent combustion
simulations”. Progress in Energy and Combustion Science, 30,
pp. 61–117.
[76] Siebers, D. L., and Higgins, B. S., 2000. Eﬀects of Injector Conditions on
the Flame LiftOﬀ Length of DI Diesel Sprays. Tech. Rep. 20008249,
Sandia National Laboratories.
[77] Husberg, T., Denbratt, I., and Karlsson, A., 2008. “Analysis of advanced
multiple injection strategies in a heavyduty diesel engine using optical
measurements and cfd simulations”. SAE Paper 2008011328.
[78] Payri, F., Berm´ udez, V., Payri, R., and Salvador, F., 2004. “The inﬂuence
of cavitation on the internal ﬂow and the spray characteristics in diesel
injection nozzles”. Fuel, 83, pp. 419–431.
[79] Gavaises, M., Arcoumanis, C., Roth, H., Choi, Y., and Theodorakakos,
A., 2002. “Nozzle ﬂow and spray characteristics from vco diesel injector
nozzles”. SAE Paper 2002300002.
[80] Lucchini, T., D’Errico, G., Jasak, H., and Tukovic, Z., 2007. “Automatic
mesh motion with topological changes for engine simulation”. SAE Paper
2007010170.
[81] Ochoterena, R., Larsson, M., Andersson, S., and Denbratt, I., 2008.
“Optical studies of spray development and combustion characterization of
oxygenated and ﬁschertropsch fuels”. SAE Paper 2008011393.
[82] Ferziger, J.H. and Peric, M., 1999. Computational Methods for Fluid
Dynamics. Springer.
Paper II
MODELLING INJECTOR FLOW INCLUDING CAVITATION EFFECTS FOR DIESEL
APPLICATIONS
F. Peng K¨ arrholm
∗
Department of Applied Mechanics
Chalmers University of Technology
Sweden
Email: fabian.pengkarrholm@chalmers.se
Henry Weller
OpenCFD Ltd.
Caversham, Reading
UK
Niklas Nordin
Scania CV AB
S¨ odert ¨ alje
Sweden
ABSTRACT
In this paper, cavitation and pressure parameters measured
in a model diesel injector are compared to data acquired by nu
merical simulations using a new code developed for the Open
FOAM platform, which uses a barotropic equation of state to
gether with the homogeneous equilibrium assumption. It is a
viscid code, allowing both compressible liquid and vapour to
be modelled. The mass ﬂow and cavitation probabilities ob
tained from the simulations are compared to data obtained in
experiments performed at AVL’s laboratories, in which the ﬂow
through an almost twodimensional nozzle was examined. The
experimental data used include pressure proﬁles and cavitation
images. The model proved to be able to predict cavitation prob
abilities, mass ﬂows, and the occurrence of supercavitation in
the channel. In addition, it proved to be stable in its dependency
on physical parameters, and grid independent.
NOMENCLATURE
A ﬂow area of the nozzle
A
i
area of a face in the nozzle in ﬂow direction
Cd discharge coefﬁcient
p pressure
∆ p pressure difference
U velocity
˙ m mass ﬂow
γ vapour fraction
∗
Address all correspondence to this author.
µ
g
vapour viscosity
µ
l
liquid viscosity
µ
f
ﬂuid viscosity
ρ ﬂuid density
ρ
v
vapour density
ρ
l
liquid density
ρ
v,sat
vapour density at saturation
ρ
l,sat
liquid density at saturation
ψ
v
vapour compressibility
ψ
l
liquid compressibility
INTRODUCTION
In Diesel engines and direct injection spark ignition engines,
fuel is injected into the combustion chamber by high pressure.
Ideally, the fuel should atomise, form small droplets, and even
tually evaporate completely in order to obtain a mixture that can
be easily combusted without formation of soot or NO
x
. Impor
tant features of the atomisation, including the size of the initial
droplets and the fragmentation processes, are mainly governed
by the ﬂow in the nozzle [1] which ejects the spray. It is there
fore important to study the ﬂow inside the nozzle, to ﬁnd out how
it is linked to the ﬂow in the combustion cylinder.
When a high velocity ﬂuid passes through a contraction like
a nozzle, a pocket of low pressure is formed in the wake of its
edge. In this wake the pressure can decrease below the satura
tion pressure, and thus cause the liquid to cavitate. The size of
the cavitation bubbles thereby formed depends on several fac
tors, including (inter alia) the injection pressure, geometry, and
1 Copyright c 2007 by ASME
Proceedings of FEDSM2007
5th Joint ASME/JSME Fluids Engineering Conference
July 30August 2, 2007 San Diego, California USA
FEDSM200737518
Copyright © 2007 Chalmers University of Technology
smoothness of the interior of the nozzle, and the properties of the
ﬂuid. For example, if the ﬂuid is pushed through the nozzle suf
ﬁciently slow, there will be no cavitation and ordinary pipe ﬂow
will occur.
Cavitation introduces vapour bubbles into the ﬂow. It also
increases the maximum velocity, for two reasons. If there is
vapour along the walls, the liquid will not have a noslip bound
ary condition, since the vapour by the wall will not be station
ary. This allows a higher velocity of the liquid, which is required
when the pressure over the nozzle is high. The velocity is also
increased since the liquid can not ﬁll the entire channel. When
the vapour bubbles formed in the nozzle reach the combustion
chamber, both the bubbles and the liquid surrounding them will
be heated by the gas in the chamber. The liquid will evaporate
faster if it has already been partially evaporated by cavitation,
and thus the entire atomisation process will be accelerated.
The point where the ﬂuid is no longer cavitating is called
the reattachment point, if there is no such point the ﬂuid will be
supercavitating and the cavitation region will extend through
out the entire nozzle. Supercavitation changes the atomisation
behaviour signiﬁcantly, since the spray’s cone angle is severely
reduced when it exits the injector, resulting in a steady ﬂow with
fairly large scale disturbances on the surface [2]. Therefore,
supercavitation should be avoided, since it severely reduces the
atomisation. There are also other negative aspects of cavitation
aside from supercavitation. Notably it can erode the injector,
thus reducing its lifetime. Further, if the injection pressure is in
creased, for a cavitating injector the ﬂow rate is still constant.
This removes the option of increasing spray penetration by using
higher pressures, and thus higher atomisation. Ultimately it will
impair performance of the engine.
PREVIOUS STUDIES ON CAVITATION IN NOZZLES
Numerical studies on Diesel sprays are usually based on
EulerLagrangian modelling according to the Discrete Droplet
Model (DDM). Few attempts have been made to model the en
tire spray using only an Eulerian framework, and those that have
been made have usually been done in 2D [3] or have been lim
ited to a restricted part of the spray [4]. There are also hybrids,
like the model described in [5], in which Lagrangian parcels are
added to the spray when the liquid phase is sufﬁciently disperse.
However, the most common way of modelling sprays is still to
use a combination of Eulerian gas and Lagrangian liquid descrip
tions. Associated with the DDM model are several submodels
describing injection, atomisation, breakup, evaporation, com
bustion etc.
The injection model will be described further, as it is of
particular interest when studying cavitation. One of the most
common approaches when modelling injection is to assume the
droplets follow a distribution with some estimated parameters,
and then sample the droplet size from that distribution. This type
of model was used in [6–8]. More ambitious injection models,
which attempt to couple the nozzle ﬂow to the injected droplets,
have been presented [9–11]. However, their rationale has not
been based on CFD or experimental data, but on theoretical con
siderations and assumptions. This paper focuses on verifying a
code for a cavitating injector, thereby providing a sound founda
tion for the subsequent development of a new injection model.
Previous works related to cavitation in nozzles under diesel
conditions have focused on various aspects of the associated phe
nomena. One option, adopted by Sarre et al [12], is to use CFD
to investigate the nozzle ﬂow and couple it to the spray. Sarre
et al focused on the spray calculations, and the effect the ﬂow in
the nozzle has on the spray, rather than studying the actual nozzle
ﬂow. In other studies only the nozzle ﬂow has been simulated,
e.g. the study by Mulemane et al [13], in which two of the possi
ble modelling approaches to cavitation was explored. The cited
authors present pressure plots similar to those in this article, but
do not compare their numerical results with experimental data.
There is also a study by Yuan et al [2], in which the cavi
tating ﬂow inside the nozzle as well as the atomising spray close
to the nozzle is modelled. This approach is not common, due
to the complexities associated with the presence of three phases:
liquid, vapour formed from the cavitated liquid, and gas which
the liquid is being injected into. Diesel injection into air without
cavitation has also been simulated in [4].
Cavitation can be modelled in several ways. There are inter
face tracking methods, which have two explicit phases, and con
tinuum methods which have just one explicit phase. Continuum
methods include the RayleighPlesset bubble method [14, 15],
barotropic equation of state method [13, 16], and a third rarely
used method based on Gibbs free energy [17].
The method chosen for cavitation modelling for this work
is the barotropic equation of state approach, mainly because it is
most appropriate for the cavitation that occurs in diesel injectors,
which originates from ﬂow separation at the nozzle inlet. The
RayleighPlesset (RP) approach is therefore unsuitable, since it
is based on the speciﬁed bubble density, and assumes gas diffu
sion and bubble coalescence to be important factors [13]. How
ever, since the velocity in diesel injectors is high, diffusion and
bubble coalescence can be assumed to have negligible effects.
Further, the code we used is a simpliﬁed Volume of Fluid (VOF)
code that allows both diesel vapour and air to be modelled. Since
the Barotropic equation of state is more consistent with the VOF
framework, it was deemed to the most suitable method for this
work.
The barotropic equation of state method described in [13]
differs fromthe one used here in the sense that it assumed the two
ﬂuid phases to be incompressible. In this paper both phases are
compressible, which increases the physical validity. It also dif
fers from the one presented by Schmidt [16], in which an equa
tion of state is analytically integrated, and thus the pressure is
found as a function of density.
2 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
Experimental data in the ﬁeld are very scarce, and many
models have only been validated (if at all) by comparing sim
ulated and empirical discharge coefﬁcients, cavitation numbers
and average velocities. This is due to the difﬁculties involved in
experimentally investigating the nozzle ﬂow. Often the cavita
tion along the sides of the nozzle makes it very difﬁcult to eval
uate the ﬂow inside, since the cavitation hinders optical analysis.
However, in experiments presented in [18] a nozzle was set up in
such a way that cavitation was limited to two of the four sides of
the nozzle, thus allowing the pressure and velocities of the ﬂow
to be measured. The data acquired in these experiments are used
to validate the model presented in this article.
The main objective of this paper is to introduce the new
model, by ﬁrst describing the theory underlying the equation of
state (EOS) used, and then the momentum equations and how
they are solved. The experimental case study used to compare to
the simulations is brieﬂy described, and the numerical setup. Fi
nally, results are presented, including both pressure proﬁle com
parisons and cavitation probability images.
THEORY
To model the ﬂow in the nozzle, we need to model the cav
itation, and either resolve the turbulent scales or model them.
Turbulence is not modelled, due to the relatively weak inﬂuence
it is believed to have on the nozzle ﬂow, compared to the cavita
tion [14]. The only way turbulence is addressed is in stabilising
numerics (in contrast to DNS computations) which are not sufﬁ
ciently accurate to include all scales of the turbulence. As men
tioned above, the barotropic equation of state was the method
chosen for modelling cavitation. We also assumed homogenous
equilibrium, i.e. the liquid and vapour are assumed to always
be perfectly mixed in each cell. The parameter γ describes the
amount of cavitated vapour in each cell.
For the method chosen to model the cavitation, an equation
of state is needed. The temperature is assumed to be constant,
which is not a major limitation [18]. A common barotropic equa
tion of state is the nonequilibrium differential equation:
Dρ
Dt
=ψ
Dp
Dt
(1)
ψ refers to the compressibility, which is the inverse square
root of the speed of sound in the cavitating mixture. This equa
tion can either be used directly in the continuity equation to for
mulate a pressure equation, or integrated to obtain the pressure
as a function of the density. The latter approach was adopted by
Schmidt et al [16], as mentioned earlier. The former approach
is problematic because equation 1 is not an equilibrium equation
of state, so the pressure and density obtained are not consistent
with the liquid and vapour equations of state unless equilibrium
has been reached. Before equilibriumis attained errors will arise,
and the code will not yield corrects result when it has been at
tained.
The equation of state should therefore be consistent with the
liquid and vapour equations of state both at the limits when there
is pure liquid or pure vapour, and at intermediate states when
there is some form of mixture. The two states can be described
by the linear equation of state:
ρ
v
= ψ
v
p (2)
ρ
l
= ρ
0
l
+ψ
l
p (3)
The parameter that describes how much of the liquid is in
each phase is γ:
γ =
ρ−ρ
l,sat
ρ
v,sat
−ρ
l,sat
(4)
γ = 1 corresponds to a fully cavitated ﬂow, and γ = 0 a ﬂow
with no cavitation. ρ
v,sat
is calculated from
ρ
v,sat
=ψ
v
p
sat
(5)
ψ
v
is the compressibility of the vapour. These properties
together form the mixture’s equilibrium equation of state:
ρ = (1 −γ)ρ
0
l
+(γψ
v
+(1 −γ)ψ
l
) p
sat
+ψ(γ)
p − p
sat
(6)
The compressibility ψ was ﬁrst mentioned in equation 1, but
nothing was said of how it is chosen. It can be modelled in sev
eral ways, one very common way is the Wallis model [19]. The
model chosen here is a simple linear one:
ψ =γψ
v
+(1 −γ)ψ
l
(7)
A linear model was selected partly because the code we used
is based on a VOF code, for which linear combinations of γ are
highly suitable for describing the mixtures properties, and partly
because it gives greater stability than the Wallis model. The γ
value is also used to calculate the mixture’s viscosity, in the same
manner as the compressibility:
3 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
µ
f
=γµ
v
+(1 −γ)µ
l
(8)
When a linear model is used for the compressibility, the
equation of state (6) can be simpliﬁed into:
ρ = (1 −γ)ρ
0
l
+ψp (9)
The ﬁrst term governs the liquid’s density when γ is low. If
the ﬂuid is cavitated, the second term becomes more dominant.
The ﬁrst term contains the property ρ
0
l
which is:
ρ
0
l
=ρ
l,sat
− p
sat
ψ
l
(10)
ρ
l,sat
is the liquid density at standard conditions. The sat
uration density of the vapour, previously used to calculate γ, is
important for the liquid’s tendency to cavitate. A higher ρ
v,sat
will not require as low ρ to get a higher γ.
Unlike the model used by Schmidt et al [16], the code does
not solve the Euler equations for inviscid ﬂowto obtain the veloc
ity and density. Viscous terms are included, and the momentum
equations are solved for U, not ρ U which is common for friction
free compressible ﬂows. A brief description of the methodology
of the code is given for clarity. The iteration process starts by
solving the continuity equation (eq. 11) for ρ, which is where
rho is convected.
∂ρ
∂t
+∇· (ρU) = 0 (11)
The ρ in the divergence term ∇· (ρU) is interpolated by use
of a total variation diminishing (TVD) scheme named MUSCL.
The MUSCL scheme is second order in space, and appropriate
when there are large gradients present, like the one between liq
uid and vapor density. Several other options are easily available
in the code. ρ is also used to obtain preliminary values for γ (eq.
4) and ψ (eq. 7), and when deﬁning the momentum equations.
∂ρU
∂t
+∇· (ρUU) = −∇p +∇(µ
f
∇U) (12)
The momentum equations are used in an ordinary fashion to
get the matrices used to calculate the pressurefree velocity. The
same TVD scheme with MUSCL limiter is used for the velocity
divergence term. It was found to be very important to choose a
proper scheme for the velocity, if a too stable scheme is used the
cavitation region will be ﬂat. Following this is an ordinary PISO
loop, to solve for p and correct the velocity to achieve continuity.
For a description on how OpenFOAM achieves continuity and
solves for p, see [20]. The equation solved in the PISO loop is
again the continuity equation, but this time transformed into a
pressure equation by use of equation 6.
∂ψp
∂t
−
ρ
0
l
+(ψ
l
−ψ
v
) p
sat
∂γ
∂t
−p
sat
∂ψ
∂t
+∇· (ρU) = 0 (13)
When continuity has been reached, the remaining properties
γ, ρ and ψ are updated. ρ is now updated according to the EOS,
eq. 6. The time step is limited by both the Courant number, and
the acoustic Courant number, respectively deﬁned as:
Co = max(
U
∆x
)∆t (14)
and
Co
acoustic
= max(
1
√
ψ∆x
)∆t (15)
The Courant number was chosen to be limited to 0.125, and
the acoustic Courant number was limited to 12.5 in all of the
simulations. However, the nonacoustic Courant number is the
real limiter of the timestep, and the acoustic Courant number is
usually around 23.
The model described is implemented for the OpenFOAM
platform, under the name cavitatingFoam (in OpenFOAM 1.4).
OpenFOAM is an open source toolbox for CFD, written in C++,
available at www.openfoam.com.
STUDY
The results produced by the code are compared to data ac
quired in experiments performed at AVL’s laboratories, in which
diesel fuel was pumped through a nozzlelike structure. The
setup is described in [18], and the model nozzle is shown in ﬁg
ure 1. The experiments generated data on density/pressure as
well as images of the cavitation region, and due to the simplicity
of the geometry they are suitable for comparison to numerical
4 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
Figure 1. VIEW OF THE EXPERIMENTAL NOZZLE, 1 cm CHANNEL
BEFORE AND AFTER THE NOZZLE
Figure 2. CLOSEUP OF THE MEASURED REGION, 299 µm WIDE,
1 mm LONG
data. An image of a typical mesh can be seen in ﬁgure 2. In the
experiments the injection pressure was 10 MPa, and the outlet
pressures were varied.
Table 1. DATA ON DIESEL FUEL
property liquid vapour
density (kg/m
3
) 832 0.1361
viscosity (kg/ms) 0.0065 5.953· 10
−6
compressibility (s
2
/m
2
) 5· 10
−7
2.5· 10
−6
vaporation pressure (Pa) 5.4· 10
3
Some of the values in table 1 have large margins of error. For
example, the compressibility of diesel vapour at room tempera
ture is an estimation based on the ideal gas law. The compress
ibility of the liquid is taken from measurements of the speed of
sound in diesel fuel [21], which is around 1400 m/s. The evapo
ration pressure chosen was that of heptane, since it is commonly
used as a model fuel for diesel. The liquid density was measured
in the experiments, while the vapour density was from the equa
tion
ρ
v,sat
=ψ
v
p
sat
(16)
The gas viscosity was obtained from the Lucas viscosity
model described in [22] (again based on nheptane’s properties),
and the liquid viscosity from Petrocard’s data on their diesel fuel
products. As will be shown in the results section, the model is
not sensitive to many of these physical properties (apart from the
liquid viscosity), as long as they are of the right order.
The walls on the sides of the nozzle, have an ordinary no
slip condition for the velocity. The upper and lower walls, on
the other hand, have a slip condition for the velocity, similar to
that used in supersonic compressible ﬂow. Perpendicular to the
wall the ﬂow (or ﬂux) is zero, but the ﬂow in the tangential direc
tion of the wall is not. A noslip boundary condition would lead
to overestimates of the thickness of the boundary layer, result
ing in strong vortices in the centre of the channel, and incorrect
modelling of the ﬂow’s cavitation. Figure 3 shows these vortices
and the resulting cavitation. As mentioned earlier, viscosity is
sometimes disregarded entirely, here we have chosen to disre
gard the viscosity from two walls due to the assumed thinness of
the boundary layer on those two walls.
RESULTS
In all of the experimental cases the injection pressure was
10 MPa, and the outlet pressure was varied. The two outlet pres
sures used here for detailed comparison are 5.1 MPa, 3 MPa and
1.5 MPa. In the experiments cavitation was measured by a line
of sight method, i.e. all cavitation observed from the top of
the nozzle was recorded, and then averaged as shown in ﬁgure
4. The numerical simulations are evaluated by a similar line of
sight method, in which the domain is searched for cavitated cells.
Both the experimental and numerical values are averaged over 20
timesteps, resulting in images such as the one shown in ﬁgure 5.
The average value at time t
n
is calculated from
U
n
=
U
n−1
t
n
−t
0
+U
n
∆t
n
t
n
+∆t
n
−t
0
(17)
5 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
Figure 3. MEAN ISOSURFACE FOR VAPOR FRACTION OF 0.1, VE
LOCITY FIELD IN A PLANE PERPENDICULAR TO THE FLOW DIREC
TION SHOWN TO VISUALIZE THE VORTICES, NOSLIP CONDITION
USED ON BOTH UPPER AND LOWER WALL
t
0
is the start time for the average, and ∆t
n
is the timestep at
time n.
When comparing the different cases, the discharge coefﬁ
cient is used to measure howmuch losses there are in the injector.
This has been deﬁned as:
Cd =
˙ m
calculated
˙ m
Bernoulli
=
N
∑
i
ρ
i
U
i
A
i
2∆pρ
0
A
(18)
The simulations (ﬁgures 5 & 7) predict the experiments (ﬁg
ures 4 & 6) quite well, even though the simulations yield slightly
more cavitation than observed in the experiments. The model
manages to predict when the ﬂow is supercavitating, i.e. when
the cavitation region extends throughout the entire nozzle. This
occurs when the outlet pressure is 1.5 MPa, but not when it is
3 MPa.
The experimental data also include pressure measurements
along the channel centreline and a line 40 µm from the channel
wall. Figures 8 and 9 illustrate these results, and as can be seen
the calculations do not predict the experimental results well. It is
important to point out that the measured pressure is much higher
than the saturation pressure for the liquid, but the cavitation prob
ability is still shown to be 100 % close to the wall (4). This will
be discussed later in the Conclusions section. The pressures in
the experiments are derived from density measurements, and an
equation of state. In the pressure proﬁle plots, the channel starts
at 0 and ends at 1 mm.
The mass ﬂow is calculated for several timesteps, and av
eraged in time over the last 100 µs. The mass ﬂow behaviour is
Figure 4. IMAGE OF THE EXPERIMENTAL CAVITATION PROBABIL
ITY, INJECTION PRESSURE 7 MPa
Figure 5. IMAGE OF THE COMPUTED CAVITATION PROBABILITY, IN
JECTION PRESSURE 7 MPa
Table 2. COMPARISON OF MODELLED AND EXPERIMENTAL DATA
p
out
˙ m
experiment
˙ m
model
Std. Deviation Cd
model
5.1 MPa 6.98· 10
−3
7.53· 10
−3
2.31· 10
−5
0.93
3 MPa 8.46· 10
−3
8.54· 10
−3
1.49· 10
−4
0.88
1.5 MPa 8.46· 10
−3
8.53· 10
−3
2.25· 10
−4
0.80
predicted within 1% for the cavitating cases, and 10 % for the
non cavitating case (∆p = 5.1 MPa). The model can also pre
dict the choking of the channel which occurs when it starts to
cavitate. The discharge coefﬁcient is also listed, which shows
how much more losses there are in the strongly cavitating case.
With the mass ﬂow constant for increased pressure difference,
the discharge coefﬁcient will decrease with increasing pressure
difference.
The reason why the mass ﬂows differ, could be underesti
6 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
Figure 6. IMAGE OF THE EXPERIMENTAL CAVITATION PROBABIL
ITY, INJECTION PRESSURE 8.5 MPa
Figure 7. IMAGE OF THE COMPUTEDCAVITATION PROBABILITY, IN
JECTION PRESSURE 8.5 MPa
mation of the liquid viscosity. As mentioned when the physical
properties were shown, there are high uncertainties in the data.
Another factor could be the lack of friction in the simulations on
the two walls with slip boundary condition, which reduces the
friction more than the lack of rough walls. This simpliﬁcations
was forced due to the presence of a strong cavitating vortex as
mentioned earlier.
To save computational time, a quarter of the mesh, in this
case 90 degrees of the full mesh, is often used. However, if such
a restriction is imposed turbulence will not be correctly mod
elled, and it will not be possible to capture large eddies. It is
however, still an attractive option due to the reduction in com
putational time. The case with an outlet pressure of 3 MPa was
examined with both the full 3D mesh, consisting of 1.7 million
cells, and the quarter mesh consisting of only 508 000 cells, to
examine possible differences in the outcome. The results (Figure
10), show that the quarter mesh provides a possible approxima
tion to the three dimensional ﬂow. However, differences in mass
−5 0 5 10 15
x 10
−4
0
2
4
6
8
10
12
x 10
6
Position (m)
P
r
e
s
s
u
r
e
(
P
a
)
51 exp
51 comp
30 exp
30 comp
Figure 8. AVERAGE PRESSURE ALONG THE CENTRELINE OF THE
NOZZLE
−5 0 5 10 15
x 10
−4
0
2
4
6
8
10
12
x 10
6
Position (m)
P
r
e
s
s
u
r
e
(
P
a
)
51 exp
51 comp
30 exp
30 comp
Figure 9. AVERAGE PRESSURE ALONGTHE WALL OF THE NOZZLE
ﬂow can be seen by comparing tables 4 and 2, for the case of
500k cells and an outlet pressure of 3 MPa.
The effects of inlet rounding are often considered since it af
fects cavitation; a well rounded inlet can reduce cavitation, while
a sharp inlet gives more cavitation. Since the rounding does not
affect the total pressure drop over the nozzle, it can be seen as a
way of examining the ﬂow as if there were no cavitation at all.
The data shown in Figure 11 and table 3 were produced using a
quarter of the channel, to reduce computational time.
Rounding was not varied in the experiments reported in [18],
so the exact values can not be veriﬁed. However, the results show
that the model can predict the reduced cavitation and increased
ﬂow rate as the inlet becomes more round. The discharge coefﬁ
cient also increases, since there is less cavitation causing losses.
7 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
−5 0 5 10 15
x 10
−4
0
2
4
6
8
10
12
x 10
6
position (m)
p
r
e
s
s
u
r
e
(
P
a
)
quarter mesh, centre
3D mesh, centre
quarter mesh, wall
3D mesh, wall
Figure 10. AVERAGE PRESSURE ALONG THE CENTRELINE AND
WALL OF THE NOZZLE, QUARTER AND FULL GEOMETRY RESULTS
SHOWN, PRESSURE DIFFERENCE 7 MPa
−5 0 5 10 15
x 10
−4
0
100
200
300
400
500
600
700
800
900
1000
position (m)
d
e
n
s
i
t
y
(
k
g
/
m
3
)
r=40mm
r=50mm
r=60mm
Figure 11. DENSITY ALONG THE WALL OF THE NOZZLE, THREE
DIFFERENT INLET RADII., PRESSURE DIFFERENCE 7.0 MPa
Table 3. DISCHARGE COEFFICIENTS AND MASS FLOWS FOR DIF
FERENT VALUES OF THE INLET ROUNDING RADIUS
r
inlet
˙ m C
d
40 mm 8.60· 10
−3
0.889
50 mm 8.83· 10
−3
0.912
60 mm 9.00· 10
−3
0.930
Figure 12. CAVITATION PROBABILITY, FOUR DIFFERENT GRIDS
USED, INJECTION PRESSURE 7 MPa
0 0.5 1 1.5
x 10
−3
0
0.5
1
1.5
2
2.5
x 10
6
position (m)
p
r
e
s
s
u
r
e
(
P
a
)
300k cells
500k cells
700k cells
Figure 13. PRESSURE ALONG THE CENTRELINE OF THE NOZZLE,
WITH FOUR DIFFERENT GRIDS, INJECTION PRESSURE 7 MPa
We have chosen to display density instead of pressure, since it
provides a clearer indication of how much the ﬂow has cavitated.
An important factor to examine when trying to model a ﬂow
is the gridsize. The standard grid chosen for this study has 508
000 cells, with ca. 5 µm cells in the channel. The grid is graded
to be ﬁner close to the wall in the cavitating region, and close to
the rounded inlet.
The calculations are not sensitive to gridsize, as can be seen
in ﬁgures 1213 and table 4. The ﬁgure is zoomed in on the cav
itation region, to make it easier to see the small differences. If a
grid coarser than the one with 140 000 cells is chosen, the cavita
tion region will consist of very few cells, and the assumption of
homogenous equilibrium becomes invalid. The numerical value
for the mass ﬂow in table 4 is different from the ﬂows listed in
table 2, since the ﬁrst are from the quarter mesh, and the latter
from the full 3D grid.
As mentioned in the Study section, some of the parameters
8 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
Table 4. MASS FLOWS AND DISCHARGE COEFFICIENTS CALCU
LATED FROM THREE DIFFERENT GRIDS, QUARTER OF DOMAIN
USED
r
inlet
˙ m C
d
300k cells 8.59· 10
−3
0.89
500k cells 8.60· 10
−3
0.88
700k cells 8.57· 10
−3
0.89
0 1 2 3 4 5
x 10
−4
0
1
2
3
4
5
6
x 10
6
coordinate (m)
p
r
e
s
s
u
r
e
(
P
a
)
pSat = 5.4 10
2
, ψ
v
= 2.5 10
−5
pSat = 5.4 10
3
, ψ
v
= 2.5 10
−5
pSat = 5.4 10
3
, ψ
v
= 2.5 10
−3
pSat = 5.4 10
3
, ψ
v
= 2.5 10
−7
Figure 14. CLOSEUP OF THE PRESSURE ALONG THE WALL,
PRESSURE DIFFERENCE 7.0 MPa, WITH FOUR DIFFERENT COM
BINATIONS OF ψ
v
AND p
sat
are difﬁcult to measure experimentally, and few data are avail
able. To test the model’s sensitivity to the ψ
v
three different val
ues of this parameter were tested for the p
di f f
=7 MPa case. The
results are shown in ﬁgure 14.
CONCLUSIONS AND FUTURE WORK
A new model for calculating the ﬂow in a cavitating nozzle
has been developed. The model has been validated against data
obtained in experiments with a simple contraction type nozzle. It
has been shown that the model can predict the cavitation proba
bility in the nozzle, and the effects of variations in the pressure
differences over the nozzle. There is a slight over prediction for
the cavitation probability, which could be due to underestimation
of the liquid’s viscosity. The model is also capable of predicting
choking of the ﬂow in the channel, since the mass ﬂow does not
increase above the value it has when the ﬂow starts to cavitate.
The pressure proﬁle from the simulations are not similar in
shape or value to the experimentally derived proﬁles. They are
slightly better for the noncavitating case. However, the experi
mental data show a minimum pressure close to the wall (ﬁgure
9) at around 2 MPa, although the cavitation probability is said to
be 100 % in this region (ﬁgure 4). If the average pressure re
ally would be of the magnitude suggested by the pressure proﬁle
plots, the ﬂow would not cavitate at all. The authors have there
fore chosen not to try to match the simulated pressure curve to
the experimental ones. The experimental data were acquired by
a line of sight method, so variations in the width of the cavita
tion region at different heights in the nozzle would not have been
seen. If the cavitation region is only large close to the walls, and
smaller at greater distances from them, this could explain why
the average pressure is said to be much higher than the evapo
ration pressure of the liquid. The simulations indicate that the
cavitation region is spread out and is not interrupted by any large
noncavitating areas.
The inﬂuence of the walls should also be considered, since
the walls have different boundary conditions. When normal non
slip conditions were applied, the ﬂow started to cavitate in the
central part of the channel, due to strong vortices. These cavi
tated regions were separated from the wall cavitation, and thus
did not result from the pressure drop following the contraction.
It was found that the vortices originated from the top and bottom
walls, and if the boundary condition on those walls were changed
to a slip condition, a more physical behaviour of the ﬂow was
observed. However, if the same slip condition was applied to
all walls, the cavitation region would grow larger (preliminary
calculations suggest this). It would also increase the mass ﬂow,
since the friction is reduced when the walls are made inviscid
(and, notably, many previously used cavitation codes have as
sumed completely inviscid ﬂow).
The real walls in the experiments are not perfectly smooth,
unlike the walls in the simulations. Smooth walls might inhibit
cavitation, since they provide fewer cavitation sites. However,
introduction of rough walls would require a new form of mod
elling, since the barotropic equation of state is not dependent on
the presence of cavitation sites.
To further validate the model, more simulations would be
needed, preferably with comparison to other sets of experimen
tal data. Since the goal is to simulate diesel sprays, higher injec
tion pressures should be tested as diesel injection pressures often
exceed 100 MPa.
ACKNOWLEDGMENT
We are grateful to Dr E. Winklhofer at AVL for allowing us
to use their experimental data applied in the veriﬁcation.
9 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
REFERENCES
[1] Faeth, G. M. and Hsiang, LP and Wu, PK, 1995. “Struc
ture and Breakup Properties of Sprays”. International Jour
nal of Multiphase Flow, 21, pp. 99–127.
[2] Yuan, W., and Schnerr, G., 2003. “Numerical Simulation of
TwoPhase Flow in Injection Nozzles: Interaction of Cavi
tation and External Jet Formation”. Journal of Fluids En
gineering, 125, pp. 963–969.
[3] Iyer, V. and Abraham, J., 2003. “An Evaluation of a
TwoFluid EulerianLiquid EulerianGas Model for Diesel
Sprays”. Journal of Fluids Engineering, 125, pp. 660–669.
[4] de Villiers, E., Gosman, A., and Weller, H., 2004. “Large
Eddy Simulation of Primary Diesel Spray Atomization”.
SAE Paper 2004010100.
[5] Blokkeel, G., Barbeau, B., and Borghi, R., 2003. “A 3D
Eulerian Model to Improve the Primary Breakup of Atom
izing Jet”. SAE Paper 2003010005.
[6] Peng K¨ arrholm, F., and Nordin, N., 2005. “Numeri
cal Investigation of Mesh/Turbulence/Spray Interaction for
Diesel applications”. SAE Paper 2005012115.
[7] Lippert, A., Chang, S., Are, S., and Schmidt, D., 2005.
“Mesh Independence and Adaptive Mesh Reﬁnement For
Advanced Engine Spray Simulations”. SAE Paper 2005
010207.
[8] Kim, H., and Sung, N., 2004. “Combustion and Emission
Modeling for a Direct Injection Diesel Engine”. SAE Paper
2004010104.
[9] Rotondi, R., Bella, G., Grimaldi, C., and Postrioti, L., 2001.
“Atomization of HighPressure Diesel Spray: Experimental
Validation of a New Breakup Model”. SAE paper 200101
1070.
[10] Baumgarten, C., Stegemann, J., and Merker, G., 2002. “A
New Model For Cavitation Induced Primary Breakup of
Diesel Sprays”. ILASSEurope Zaragoza.
[11] Huh, K.Y., and Gosman, A.D., 1990. “Atomization Mech
anism of Fuel Injection”. ILASS.
[12] von Kuensberg Sarre, C., Kong, S., and Reitz, R., 1999.
“Modeling the Effects of Injector Nozzle Geometry on
Diesel Sprays”. SAE Paper 1999010912.
[13] Mulemane, A., Subramaniyam, S., Lu, P., Han, J., Lai, M.,
and Poola, R., 2004. “Comparing Cavitation in Diesel In
jectors Based on Different Modeling Approaches”. SAE
Paper 2004010027.
[14] Payri, R., Margot, X., and Salvador, F., 2002. “ANumerical
Study of the Inﬂuence of Diesel Nozzle Geometry on the
Inner Cavitating Flow”. SAE Paper 2002010215.
[15] Yuan, W., Sauer, J., and G.H., S., 2001. “Modeling and
Computation of Unsteady Cavitation Flows in Injection
Nozzles”. Mechanique & Industries, 2, pp. 383–394.
[16] Schmidt, D.P., Rutland, C.J. and Corradini, M.L., 1999.
“A Fully Compressible TwoDimensional Model of High
Speed Cavitating Nozzles”. Atomization and Sprays, 9.
[17] Vortmann, C., Schnerr, G., and Seelecke, S., 2003. “Ther
modynamic Modeling and Simulation of Cavitating Nozzle
Flow”. International Journal of Heat and Fluid Flow, 24.
[18] Winklhofer, E., Kull, E., Kelz, E., and Morozov, A., Zurich,
Switzerland, 2001. “Comprehensive Hydraulic and Flow
Field Documentation in Model Throttle Experiments Un
der Cavitation Conditions”. ILASSEurope Annual Meet
ing, pp. 574–579.
[19] Wallis, G. B., 1969. Onedimensional Twophase Flow.
McGrawHill, New York.
[20] Peng K¨ arrholm, F., 2006. Numerical Modelling of Diesel
Spray Injection and Turbulence Interaction. Department of
Applied Mechanics, Chalmers University of Technology.
[21] Mustafa, E. T., and Van Gerpen, J. H., 2003. “Effect of
Temperature and Isentropic Bulk Modulus of Mixtures of
Biodiesel and Diesel Fuel”. Journal of the American Oil
Chemists’ Society, 80(11), pp. 1127–1130.
[22] Poling, B., Prausnitz, J., and O’Connell, J., 2001. The Prop
erties of Gases and Liquids, 5th ed. McGrawHill.
10 Copyright c 2007 by ASME
Copyright © 2007 Chalmers University of Technology
Numerical Modelling of Diesel Spray Injection, Turbulence Interaction and Combustion ¨ Fabian Peng Karrholm The image on the cover of the thesis is an illustrative image of a lagrangian spray, evaporation has been turned oﬀ to improve the visual experience.
¨ c Fabian Peng Karrholm, 2008 ¨ Doktorsavhandling vid Chalmers Tekniska Hogskola First edition ISBN: 9789173851732 ISSN: 0346718X Ny Serie: 2854 Department of Applied Mechanics Chalmers University of Technology SE412 96 G¨teborg o Sweden Telephone +46 31 772 10 00
Printed at Chalmers Reproservice G¨teborg, Sweden 2008 o
I guess this is my dissertation Homie this ... is basic, Welcome to Graduation Good morning, Good morning On this day, we become legendary  Kanye West
iii
Abstract
This thesis covers two main topics. The ﬁrst is numerical modelling of cavitating diesel injector ﬂows, focusing on describing such ﬂows using a singlephase cavitation model based on a barotropic equation of state together with a homogenous equilibrium assumption. The second topic is EulerLagrangian simulations of diesel sprays, focusing on attempts to reduce the high grid/timestep dependencies in numerical simulations of diesel sprays. In addition, the ability of two CFD codes to predict ﬂame liftoﬀ length and ignition delay time, and the advection scheme’s inﬂuence on fuel distributions, are considered. A longterm goal was to develop a new atomization model based on calculated ﬂows in injector nozzles, which did not have the drawback of requiring either nonphysical parameters or information derived from speciﬁc experiments. To validate the cavitation simulations, comparisons were made with experimental data obtained at AVL. The experimental data (which are practically 2D) provide information on velocity proﬁles and pressure contours. These data were used to validate the code. However, since the code is not stable for dieseltype pressures, no atomization model was developed. The main part of the thesis describes how diesel sprays were simulated using the discrete droplet model (DDM), in which the liquid is described by Lagrangian coordinates and the vapour by an Eulerian approach. The simulations have been used to investigate how the kε family of turbulence models inﬂuence spray behaviour, and a simple but eﬃcient way to reduce the dependency of the mesh resolution, by limiting the turbulence length scale in the liquid core region, is proposed. This constraint is shown to have a positive eﬀect on the spray behaviour, and to reduce both grid and timestep dependencies. In addition, the ignition delay time and ﬂame liftoﬀ lengths have been investigated, since these two properties are believed to be important for emissions formation. The simulations used a complex chemical mechanism involving 83 species and 338 reactions. The eﬀects of the numerical scheme, the turbulence model and physical parameters (like ambient temperature and oxygen content) on these variables have also been investigated. Keywords: Spray, Computational Fluid Dynamics, Diesel, Turbulence Model, Cavitation, Atomization, Flame, Liftoﬀ, Ignition Delay, Numerical Scheme, TVD, Combustion
iv
Last. Edwin P and Alicia K. I am also in debt to Feng Tao. I thank the Combustion Engine Research Center (CERC) for providing ﬁnancial support for this work on Diesel Sprays. Thanks also goes out to Ingemar Denbratt for his enthusiasm and support. you are the ones who makes everything worthwhile. thank you for your enthusiasm. the work you do for OpenFOAM is priceless. I also owe a great deal to Rasmus Hemph for the many discussions on the mysteries of OpenFOAM. Ernst Winklhofer for sharing his experiments on cavitating nozzles.Acknowledgements I thank Niklas Nordin for his supervision. Dr. I thank AlfHugo Magnusson for providing experimental data for paper 1. I also want to thank everyone else at the department. Cecilia PK. I thank Dr. and accomplishing this. I also thank myself for hanging in there. discussions. suggestions. encouragement and help during the latter part. I also wish to thank some people abroad. I want to thank my family. help and guidance in OpenFOAM. CFD and autolyse. Henry Weller for his assistance on the second paper. I have really appreciated the lunch & coﬀee breaks as well as discussions on more workrelated matters. who have helped me during my time at Chalmers. Dr Hrvoje Jasak for helping me with the appendix on RhieChow. Tommaso Lucchini who I am proud to have shared oﬃce with. new and old. especially during the ﬁrst years of my phd studies. Afterall. and for answering to so many of my questions on cavitation and its implementation in the code. and who showed me the path to selflearning in OpenFOAM. v .
vi .
C2 .C3 .v C1 .Cµ CD Cd Ce Co d D D Ea F Fs f fheat g hv k kj l∞ lt Lsgs m n n ˆ Nr Ns Nu Oh p0 p∞ p Relaxation coeﬃcient Molar Concentration Liquid speciﬁc heat at constant volume Fluid speciﬁc heat at constant volume Turbulence model constants Drag coeﬃcient Discharge coeﬃcient of nozzle Constant of the D 2 law Courant number Droplet diameter Droplet Diameter Mass Diﬀusion Coeﬃcient Arrhenius Activation Energy Force on particle Spray momentum source term Coeﬃcient in RNG turbulence model Correction factor for heat exchange due to mass transfer Gravitation acceleration vector Enthalpy of vapour Kinetic energy of turbulent ﬂuctuations Arrhenius constant of reaction j Relaxation length scale Length scale of turbulent ﬂuctuations Maximum userset length scale Mass Parameter for RosinRammler distribution Face Normal Number of reactions Number of species Nusselt number Ohnesorge number Pressure at outlet in previous time step Desired outlet pressure Pressure vii mol/m3 J/kgK J/kgK m m m2 /s J/mol kgm/s2 kgm/s2 m/s2 J/kg m2 /s2 m m m kg Pa Pa Pa .Nomenclature Roman Symbols b c cl cp.
∞ Yi Momentum Coeﬃcient in the RNG kε turbulence model Universal Gas Constant Reaction Rate of species i Reynolds number of ﬂuid Strain Help variable Schmidt number Sherwood number Temperature of ﬂuid Time Taylor number Velocity Outgoing pressure wave velocity Weber number Volume fraction of scalar i Mass fraction of fuel vapour at droplet surface Mass fraction of fuel vapour far away Mass fraction of scalar i kgm/s J/Kmol 1/s K s m/s m/s Greek Symbols α αk .viii p RRNG R RRi Re S Sfa Sc Sh T t Ta u.s Xv.αε β βmax γ δ ∆t ε κc λ Λ µ µT ν ρ ρs ˙ τe τu φ ψ Ω Volume fraction of ﬂuid Turbulence model constant Volume fraction of ﬂuid Maximum spray cone angle Vapour (cavitated liquid) fraction Cellface distance coeﬃcient Timestep Dissipation rate Thermal Conductivity Limiter function for interfaceCompressionscheme Wavelength of fastest growing wave Laminar viscosity Turbulent viscosity Kinematic viscosity Density Spray mass source term Relaxation time of evaporation Momentum relaxation time Volume ﬂow across cell face Compressibility Growth rate of fastest growing wave m s m2 /s3 W/mK m kg/ms kg/ms m2 /s kg/m3 kg/m3 s s s m3 /s s2 /m2 m/s2 .U w We Xi Xv.
. . . . A Code for Cavitating SinglePhase Flow . . . . . . . . . . . . . . . Spray SubModels . . .4 Background & Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . .1 2. . . . . . . . . . . . . . . . . Simple Description of the Computational Approach . . . .3 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . .1. . . . 3. . . . . . . . . . 3 Spray Modelling 3. . . . . . . .2 3. . . . . . .1 Development of Injection Model . . . . . . . . . . . . .1 The Gas Phase . . .2 2. . . . . . . . . .3 1. 2. Chemistry . . . . . . . . . . A Code for Cavitating Diesel Injector Flow into Air . . . . . . . . . . . . . . . . . . . . . .4 Inlet . . Turbulence Modelling . . . . . . .2. . . . . . 2.1 1. . . Boundary Conditions for Pressure . . . . . .4 3. . . . . . . . Outlet . . .1.4. . . . . . . . . . . . Implementation and Code . . . . .2 2. 1 1 4 5 6 9 10 11 15 15 15 16 19 21 22 22 23 24 27 30 30 2 Nozzle Flow 2. . . . . .2 1. . . . . .1 3. .Contents 1 Introduction 1.2 Basic Equations . . . .3. .3 Previous Works in the Field . . . . . Outline of the Thesis . . . . . . . . . . . 3. . . . . .1 Spray Motion Equation . . Turbulence/Spray Interaction . . . . . . . . . . . . . ix . . . . . . . . . . . . . . . . . . .1 2. . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . .
Time Step Dependency . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Formulating the Momentum Equation Matrix in OpenFOAM . . Drag Model . . . . . . . . . . . . . . . . . . . .3. . . . Injector Simulation Results . . . . . . . . . . . . . . .2. . . . . . . . 31 32 32 33 34 36 37 39 39 43 44 46 46 48 50 51 53 57 59 61 63 64 67 71 71 76 79 4 Results & Discussion 4. . . . . . 4. . .2. . . . . . Particle Collisions . . . . . 4.3. . . . . . . . . . . . . . . Fuel Sprays with Chemistry . . . . . . . . . . . . . . . Injection Model . . . . . . . . . .6 Comment on the PaSR Model .6 3. . . . . . . . . . . . . .2. . . . . .1 4. . . . . . . . . Spray Results without Chemistry .3 3.2. . . . . . . . . . . . Heat Transfer . . . .1 Combustion Predictions Using Other Schemes . . . . . . . .7 3. . . . . . . . . .2 Simulations of a Diesel Flow in a Cavitating Model Nozzle . . . . . . . . . . . . . . . . . . . . . . . . . .x 3. 5 Conclusions 6 Future Work 6. . . . . . . Breakup Model . . Comparison to Experimental Data . .2.2 3. A. . . . . . .5 3. . .1 RhieChow Interpolation in OpenFOAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Evaporation Model . . . . . . . . . . .2 4. .1 Outlook . .2. . .1 4. . .2.4 3. . . . . . . . . . . . . . . . . . B Variable Deﬁnitions and Dimensionless Numbers . . . . . . . .8 CONTENTS Parcel Tracking . . . .3 Grid Dependency . .5. 7 Summary of Papers A Implementation in OpenFOAM A. . . . . . . . . . . . . . . . . . . .1 4. . . . . . . . . . . . . . . .2. . . . .5 Inﬂuence of Numerical Scheme . . . . . .3 Impact of Injection Model . . 4. . . . .4 4. . . . . . . . . . . . . 4. . . .3. . . . .
we are facing a new environmental threat. one of which is to reduce soot and NOx generated in the process of burning the fuel. but there is no doubt that horses. created mainly by horse manure”. who are often more concerned about cupholders. In the late 19th century. Switzerland and Austria died prematurely due to heart and lungdiseases caused by particulate matter from vehicle traﬃc. than about improving fuel eceonomy.1 Background & Motivation Today. Details of the data used to obtain the estimate are unknown. than from traﬃc accidents in these countries. now. and the resulting manure. In addition to the demands imposed by environmental legislation. manufacturers have to meet the customers’ demands. The Diesel engine is facing many challenges today. extra horsepower and larger cars. and later Rudolf Diesel in 1897 patented their prime mover inventions. In surveys by CNW Marketing Research [5]. over 100 years later. It should be noted that more people are dying from traﬃc pollution. However. These 1 . each year roughly 21 000 people over the age of thirty in France. According to a rough estimate published in Appleton’s magazine in 1908. and this time our former saviours have become our problem. cars and heavy duty vehicles are considered to be major sources of environmental pollution. According to an estimate quoted in a report on the Diesel engine by P. so national governments around the world are forcing manufacturers to decrease speciﬁc emissions from the vehicles they produce. Nikolaus Otto in 1876. Both soot and NOx are harmful to humans and animals ([2]. as shown in ﬁgure 1. Horse manure became a problem of the past.Chapter 1 Introduction 1. 20 000 New Yorkers died each year from “maladies that ﬂy in the dust. constituted a major environmental problem[1]. buyers of cars were asked which factors were important in their decisions. when the ﬁrst automobiles appeared in the streets of New York and London they were considered to be solutions to another problem: horse manure. Kageson [3]. However.1.[3]).
sound system. the amounts of NOx and CO2 emitted into the atmosphere between 19902005 from vehicles in Sweden are shown. or the oil price is drastically increased.3.2% of the buyers in 1995. a decline or at least reductions in the rate of increase in vehicle size would be expected. However. In ﬁgure 1. and overall exterior styling (the reason most buyers found important). If customer awareness is increased. it was important to fewer customers than factors like interior comforts. to 61. The amounts of harmful NOx gases are decreasing.2 CHAPTER 1. manufacturer’s reputation.5 2 NO x PM HC+NO CO x g/km 1. The ﬁgure also shows that fuel consumption per unit distance travelled declined sharply between 1975 and 1985.5 1 0. only the numbers who thought they were important reasons for their choices. and what they actually do. as can be seen from ﬁgure 1. INTRODUCTION 3 2. There also seems to be a discrepancy between what consumers claim to do. perhaps the demand for large cars with powerful engines will decline. The same report provides data on the average weights of US light vehicles and average fuel consumption dating back to 1975. CO2 is an inevitable product of combustion when . Downsizing of the engine could then become a viable option to meet legislative demands as well as customer needs. It should be noted that the survey does not take into account how important the factors were to the buyers.3% in 2005.2. the decrease has ﬂattened out since then. It should also be mentioned that increases in safety features of modern cars have also increased their weight. 19922008 [4]. but due to the increases in power & weight. the power of the engine per unit vehicle weight has increased over the past 30 years.5 0 1990 1995 2000 year 2005 2010 Figure 1. surveys showed that although fuel economy increased from being important to just 27. If most consumers really considered fuel economy to be an important factor when choosing a car. and increased even more since 2005. while those of the greenhouse gas CO2 are slowly increasing.1: Changes in legal limits for emissions from automobiles within the EU.
3: Emissions from vehicles in Sweden.1. 19902005 [6] carbonbased fuels are burnt. However. For the nearer future. or using the engine less. batteries that can power a vehicle for longhaul highway operations are still far in the future. BACKGROUND & MOTIVATION 3 1.5 CO2 0 1990 1995 2000 2005 year 0 1990 1995 2000 2005 year Figure 1. and the only way to reduce emissions of it is to burn less.2 80 1 70 0.4 90 1.1. 19752007 x 10 4 200 2. and power per unit weight of cars (dashed line). The second option becomes more interesting as battery power is increased.2: Fuel consumption per unit distance travelled (solid line). power/weight (watt/kg) .5 150 1000 tonnes 1000 tonnes 2 1.6 1975 1980 1985 1990 1995 2000 2005 50 2010 year Figure 1. in 1000 tonnes.8 60 0.6 100 fuel consumption (l/10km) 1.5 100 1 50 NO x 0. increasing fuel eﬃciency and engine downsizing are more likely to be viable strategies. This can be done either by increasing the eﬃciency of the engine.
The parcels may consist of any number of droplets. In addition. However. and the liquid spray is discretised into computational ’parcels’. and the other involving attempts to model engines numerically. a ’point particle’ approximation is typically made. meaning that each position in space has a numerical representation. breakup. conserving momentum and energy. depending on values set by the user. collision. Another way to describe reality is to follow a particle. Solving conservation equations using such a mesh is called an Eulerian approach. since neither of them can capture all the relevant details. and compute its state vector. as all forces acting on the rocket moves with it. 1. the path of a rocket can be conveniently described in Lagrangian coordinates. for example the cylinder of an engine. and track its position in time. is particularly appropriate when all forces acting on an object are known. A symbolic relationship between the two approaches is usually employed. is divided into several small cells or control volumes. The spatial domain of interest. Each of these parcels is subject to the same processes as a real diesel spray. For example. where they provide a cheaper alternative than largescale engine studies. The experimenter does not know exactly what happens inside the engine. each of which is described by Lagrangian coordinates. The processes mentioned. which requires a large number of submodels to empirically account for integral ﬂuxes over the . a reasonable description of the processes taking place can be obtained.2 Simple Description of the Computational Approach CFD is an abbreviation of Computational Fluid Dynamics. we inject part of the spray. and the modeller cannot be certain the events predicted by the model are correct without experimental veriﬁcation. Another advantage of modelling is that it allows conceptual engines and combustion modes to be explored long before prototypes are made. such as heat transfer. and refers to solving a ﬁnite approximation of the NavierStokes equations numerically. including (inter alia): atomization. but resolution of particle surfaces is computationally prohibitive. The air in the combustion chamber is described using an Eulerian framework. including variables such as its position over time. Simulations can also be useful for parameter studies. heat transfer and turbulence. or the inside of a diesel injector. the modeller has all the data available from them. A combination of these two representations is often used in Diesel spray simulations. The sum of the control volumes. the shape of a car. a thermometer that measures the outdoor temperature at the same spot every day provides an example of an Eulerian description of temperature. known as Lagrangian. called a mesh. evaporation. provides a ﬁnite approximation of the spatial domain. Therefore. This approach. occur as integral operations over the particle’s entire surface. INTRODUCTION There are two branches of research concerning the optimization of engine design and operators: one involving experiments with physical engines. That is.4 CHAPTER 1. For instance. each of which is considered identical. with a veriﬁed model.
OpenFOAM has many other advantages1. 1 One of the most recent additions is an automatic mesh generator. thus no parcels are lost when moving between cells as in Kiva3V. which can vary over an order of magnitude. Lagrangian parcels are tracked by facetoface tracking.openfoam.2. and the demands it places on the user. most of the submodels available today are quite unreliable. IMPLEMENTATION AND CODE 5 particle surfaces. Further.3. making it possible to create meshes using any form of cells.com. The code now includes polyhedral mesh support.1 (the latest. when I started my doctoral studies I used Nabla FOAM 1. but one of the most important is the complete parallelization of the code. followed by OpenFOAM 1.3 Implementation and Code The code used in the work this thesis is based upon was OpenFOAM ([10] &[11]). are needed to model physical processes such as atomization. I believe that research should generally be done in a community in which exchange of information is the norm. which makes it very easy for the user to switch between turbulence models. than in other codes.5. removing the need (in most cases) for the user to consider multiple processor simulations. While this is true. models are implemented to be runtimeselectable. OpenFOAM code was chosen both because of the high scope it oﬀers for developing new models.0 up until the recent 1. . All solvers written in OpenFOAM can easily be run in parallel. It is an objectoriented code written in C++. 1. One could argue that access to source code. Further submodels. which makes it reasonably straightforward to implement new models and ﬁt them into the whole code structure. and depend on the user to supply constants that have limited physical interpretability.[9]). can easily be achieved by using an inhouse code. depending on the type of spray behaviour being considered. spray submodels. requiring still more empiricism. OpenFOAM is being continously developed. The ability to share codes and ideas developed in the same framework is one of OpenFOAM’s most important strengths. as long as the quality of the resulting mesh is high.1. numerical schemes etc. It is more diﬃcult to not know what you are doing in OpenFOAM. release has not been used). Unfortunately. and knowledge of what one is doing. These constants can often diﬀer widely between authors. an opensource code available at www. and publications. A typical example is the breakup constant in the KelvinHelmholzRayleighTaylor model ([7]. 1.[8].4. since the code is parallelized at such a fundamental level. and an important contributor to progress.
which describes how liquid droplets are split into smaller droplets. One example is the dependency on the computation grid. and the results compared to experimental data to determine which grid is most suitable. a goal of this project was to develop such a model. and it is separated from secondary breakup. Therefore.4 Outline of the Thesis The Introduction chapter presents basic information about sprays. but calculates it from set variables such as the injection pressure. A very important parameter for diesel simulations is the ﬂame liftoﬀ length. the project is described in the same order that the fuel travels. spray simulations are more sensitive to the grid than other CFD simulations. The most common way to inject droplets in EulerLagrangian sprays is to use a size distribution and randomly inject droplets from the injector in directions within the prescribed cone angle. Atomization refers to the primary breakup of the liquid core of the spray. Where it is included. and then on to the spray calculations. since increasing it will cause the ﬂame to have a more premixed character and consequently yield lower emissions of NOx and particulate . turbulence and the ﬂow in the nozzle into account that is only dependent on physical parameters could increase the quality of the simulations. which is believed to have a strong eﬀects on the emissions generated during combustion. and to develop better models. An atomization model can be used to describe how the droplets break up initially. since an increase in the number of cells can either increase or decrease the quality of the simulation. The errors in spray simulations do not follow the common CFD rule of decreasing as the number of cells is increased[12]. which does not prescribe an injection angle and droplet distribution. One of the models that has been focused on in this work is the injection and atomization model. If some of this grid dependency could be reduced. INTRODUCTION 1. but such repetition has been kept to a minimum. Subsequent chapters assume a knowledge of basic university level math and physics. starting with the nozzle ﬂow.6 CHAPTER 1. or because the description in the article was too concise due to size limitations. length and width of the nozzle etc. The idea is that this chapter should be understandable to people without a degree in engineering. Some parts of the modelling have been described twice (in both papers and the thesis). or simply switch oﬀ breakup within a certain distance from the injector. and motivation for the work. it would be of great value to the numerical community. A better injection model taking cavitation. fuel density. Instead. EulerLagrangian spray simulations uses submodels to describe the processes that aﬀect the injected liquid. thus it has a major inﬂuence on the initial breakup of the spray. Spray modelling today is at a stage where there is a heavy reliance on experimental data for validation. Therefore. turbulence interactions and combustion. The minimum has to be found on a casetocase basis. the discretization error has a minimum for a speciﬁc grid size. The goals of the work have been to improve existing models for diesel spray simulations. it is included for continuity. In the following chapters. due to the relationship between the liquid (Lagrangian parcels) and the gas (Eulerian cells). The model is used to obtain initial droplet size and velocity parameters.
and compare it to that of the commonly used KIVA3V code. and requires a welldeveloped spray model to even appear. The liftoﬀ length phenomenon is quite complex. Part of the reason why this was studied is the wide variety of schemes available in OpenFOAM. If inappropriate schemes. OUTLINE OF THE THESIS 7 matter. It was discovered quite early on (during the studies described in Paper 1) that the chosen scheme and initial parameters have a large inﬂuence over the resulting spray. Attempts were also made to benchmark OpenFOAM’s performance in terms of simulating these phenomena. The last issue addressed during the work underlying this thesis were the eﬀects of numerical schemes and initial turbulent viscosity on the solution.4. .1. submodels and combustion models are employed the ﬂame will spread all the way to the nozzle and no liftoﬀ length will be observed. It is also closely linked to the ignition delay time. in which it is very easy to switch between diﬀerent types. which is very important for engine simulations since full control over the ignition is highly desirable.
8 CHAPTER 1. INTRODUCTION .
Firstly. It is therefore important to study the ﬂow inside the nozzle. no cavitation will occur and ordinary pipe ﬂow will be observed. and if the pressure in this wake is suﬃciently low. The size of the cavitation bubbles will depend on several factors. The atomization. when they reach the combustion chamber the ligaments and droplets formed are already partly evaporated through cavitation. geometry. Secondly. cavitation is believed to have a strong inﬂuence on the ﬂow inside the nozzle. Ideally. it should then atomize. since the vapor bubbles formed in the nozzle are mixed with the liquid. if there is vapour along the walls the liquid will not have a noslip condition boundary since the vapour will not be stationary. and the entire atomization process will be accelerated. The velocity is increased for two reasons when the ﬂuid is cavitating. so it is also important to consider these eﬀects. and the properties of the ﬂuid. thus allowing the velocity of the liquid to increase (which is essential when there is a high pressure drop over the nozzle). below the saturation pressure. the liquid will cavitate. i. 9 . is mainly governed by the ﬂow in the nozzle which ejects the spray ([13]). including injection pressure. Cavitation introduces vapour bubbles into the ﬂow and increases the maximum velocity in the centre. i. In a spark ignition engines it is also important to position the fuel cloud close to the spark plug to promote good combustion. form small droplets. The remaining liquid will evaporate faster. For example. yielding a mixture that is easily combustible without forming either soot or NOx .e. to ﬁnd out how it is linked to the spray in the cylinder. and eventually completely evaporate. A pocket of low pressure is formed in the wake of the contraction’s edge. if the ﬂuid is pushed through the nozzle suﬃciently slow.e. Cavitation can occur when ﬂuid of high velocity passes through a contraction like a nozzle. smoothness of the interior of the nozzle. when the ﬂuid is cavitating the liquid can not ﬁll the entire channel. the formation of the initial droplets and their breakup. Furthermore. Furthermore.Chapter 2 Nozzle Flow In both Diesel engines and direct injection spark ignition engines fuel is injected into the combustion chamber by high pressure.
who developed a new atomization model for the EulerLagrangian Spray and Atomization (ELSA) approach. it is transformed into a Lagrangian parcel.[16]. but few authors have performed nozzle simulations with intention to create a new atomization model for EulerLagrangian diesel spray simulations covering a comprehensive range of initial conditions. . but not the ambient air in the combustion chamber.[22]. They have been used previously in [21] using a Rayleigh Bubble Growth model [22]. This code was originally described in [30].1 Previous Works in the Field The ﬂow inside a diesel nozzle has been studied by many researchers.[15]. which is very rarely used. ELSA is a mixture of modelling approaches in which a full Eulerian description is used for the nearnozzle ﬂow where the spray is dense. The resulting code has not been validated within this project. As reported by Nouri et al in [26]. calculated from a simpliﬁed Rayleigh equation.[17]).[27]) or.[20]. which in turn increases the ﬂow velocity. cavitating liquid and ambient air. but since the cited publication has limited availability it will also be described here. The Rayleigh bubble growth approach solves a transport equation for the liquidgas interface.[19]. which assumes that the cavitated mixture is perfectly mixed and the amount of cavitation can be described by a vapour mass fraction between 0 and 1. An exception is Ning [24]. Cavitation is usually modelled using either Rayleigh bubble growth dynamics theory in conjunction with a volume of ﬂuid method ([17]. and the HEM approach needs an equation of state relating pressure and density to calculate the growth of cavitation. However. An additional complication when modelling cavitation that most studies have neglected (one exception is [25]) is the eﬀect of cavitation in multihole injectors. both experimentally ([14]. although the simulation time extends over several weeks. was contracted to develop a code handling liquid. it is possible to use it to simulate water injections. based on Gibbs free energy [29]. NOZZLE FLOW 2. due to problems in its handling of diesel fuel ﬂows. This is often an essential simpliﬁcation due to the diﬃculty of resolving three phases while maintaining reasonable computational times. and numerically ([17]. The results from [15] by Winklhofer et al have been used in this work to validate the singlephase cavitation model.10 CHAPTER 2. OpenCFD Ltd. A simpler code (presented in the following section) for a cavitating ﬂuid was also developed as part of the project. In some studies nozzle ﬂow calculations have been coupled to Lagrangian spray simulations ([23]) on a casetocase basis. Previous works in the ﬁeld have modelled cavitating liquid.[28]). As part of this project. with a source term to model cavitation. the Homogenous Equilibrium Model (HEM. There is also a third approach. strings of cavitation can occur between diﬀerent nozzle holes in multihole injectors and the strings reduce the eﬀective hole ﬂow area.[20]). In both approaches an assumption about isentropic nozzle ﬂow is usually included.[18]. but if the liquid volume fraction falls below a userset value.
the Volume of Fluid (VOF) method is used.2) )ρU) − · (µ U) = − p (2.4) (2. and vapour to liquid that undergoes cavitation.2.vSat − ρα. A CODE FOR CAVITATING DIESEL INJECTOR FLOW INTO AIR11 2.lSat ρα.6) .2.l ) + βµβ + µsgs where α and β are the fuel and air volume fractions.5) (2. presented below: ∂ρU + ((U · ∂t with µ = α(γµα.1) The code uses a Large Eddy Simulation (LES) model. gas refers to the ambient gas which the liquid is injected into.2 A Code for Cavitating Diesel Injector Flow into Air The purpose of this code is to simulate cavitating diesel fuel injected into a chamber ﬁlled with air.3) (2. The relationship between these two is: α+β =1 In equation 2. The equations that need to be solved are.2 in the µsgs term. In addition to the momentum equation there is the continuity equation for each phase: ∂αρα + ∂t ∂βρβ + ∂t where · (ρα αU) = 0 · (ρβ βU) = 0 (2. To simulate the liquidgas interface.v + (1 − γ)µα. as usual. the vapour fraction is described by γ.lv − ρα. the momentum equation and the continuity equation. LES is believed to be necessary in order to resolve the liquid turbulence interaction with the low turbulence in the gas phase. γ= ρα. indicating the proportion of the cells that are cavitated.lSat (2. respectively. in which the liquid phase is allowed to cavitate and the gas phase can be compressed due to pressure shocks. Even though the code can have other uses. simulating diesel injections was its original purpose. In this text. which is apparent in equation 2.2.
5 & 2.15) .13) sfa = ψChung = (1 − γ) ρv. respectively.Sat + (1 − γ)ρl. one for liquid and one for the gas.8) Since the code is intended to account for compressibility an equation of state (EOS) relating pressure and density is needed. NOZZLE FLOW ρ = αρα + βρβ and thus equations 2. and some form of mixture for the intermediate states.Sat ψl √ 2 ψv ψl sfa ρ (2.9) (2.Sat ψv γsfa 1−γ √ +√ ψv ψl + γ l.6 can be combined to form ∂ρ + ∂t · (ρU) = 0 (2.7) (2. it is related to the speed of sound by: 1 a= √ ψ The constant ρ0 is deﬁned as l ρ0 = ρl. For modelling diesel injection into air two equations of state are needed.14) (2. Three models have been implemented: the Wallis equation [31]: ψv ψl + (1 − γ) ρv.vap ρair = (1 − γ) ρ0 + (γψv + (1 − γ) ψl ) psat + ψ(γ) p − psat l = ψg p (2.12) (2.11) When the code is executed the model for ψ is chosen at runtime.Sat ψv ρv. The main reason for selecting the chosen EOS (for the liquid) is that it obeys the liquid and vapour EOS in the limit cases of pure liquid and pure vapour.Sat ψWallis = (γρv.10) Where ψv and ψl refers to the vapour and liquid compressibility. There was a more detailed rationale for the EOS used in the studies presented in Paper 2. For reference. since it is assumed to not undergo any phasechanges: ρliq.Sat ρl.Sat ) γ the Chung equation [32]: (2.sat − psat ψl l (2. but that code does not use VOF and thus only one EOS is needed.12 CHAPTER 2. The gas EOS is simpler.
but it was found to be quite unstable when used for high speed ﬂows.6.19) These two equations could be solved for α and β respectively with the right hand side as the compressible “source”. they will not be massconservative.2. A CODE FOR CAVITATING DIESEL INJECTOR FLOW INTO AIR13 and a linear model: ψlinear = γψv + (1 − γ) ψl (2.10. Equations for U and p have now been derived.sat + ψp (2. and the other the compressibility eﬀects. when these equations are discretized.5 & 2. They can be rearranged into: . An example of how this is done for a simpler code (icoFoam) can be found in Appendix A. When a linear compressibility model is used.17) We can now return to equations 2.6.16) The linear model was the one chosen for this work. However. If the two equations above are combined with 2. we also need an equation to convect the density for the momentum equation.1. We can begin by separating the two equations into two components: one representing incompressible ﬂow.9 & 2. where viscosity and mass fraction are also described by linear equations.5 and 2. respectively) to form the equation for pressure. and by using the equation of state.2. However. and form equations that are more suitable for solving for the phase fraction and densities. the EOS for the liquid can be reduced to: ρliq. ∂α + ∂t ∂β + ∂t · (αU) = − α Dρα ρα Dt β Dρβ · (βU) = − ρβ Dt (2. since it is then consistent with the VOFmethod. The Wallis model represents the lower speed of sound (compressibility) in a bubbly mixture in a more physical way.18) (2. since there is no coupling between α and β when they are solved. the density can be found. which has already been deﬁned in 2. The way the phase fraction variables are calculated will be described later.vap = ρl.3 an equation for the divergence of U is formed: α Dρα β Dρβ − ρα Dt ρβ Dt ·U = − (2.20) This equation is used with the discretized momentum equation and the equation of state (equations 2.
19. these equations are bounded by αβ.25) . the boundedness can be maintained by choosing the solver to be implicit with respect to α or β. By adding the term as a divergence term.18 and 2. and the term introduced will have an αβ term.19. The last two variables that need to be solved for are the phasefractions. The equation to be solved can be derived from 2.21 is then modiﬁed by adding the third term on the left hand side in the equation below: ∂α +U· α + ∂t 1 Dρα 1 Dρβ − ρα Dt ρβ Dt · (Uc αβ) = −αβ (2.24) Unlike 2.3. The new term cannot be allowed to aﬀect the solution in any other way than by compressing interfaces. The last requirement is to ﬁnd a way to maintain sharp interfaces.22) These equations are used to convect ρ and illustrate that if the material derivatives of either density is larger than the other.18 & 2. For this code a “countergradient” method was chosen. The terms within the parenthesis on the right hand side are treated explicitly. which unfortunately has problems to fulﬁl its nondisperse property when both liquid and air are ﬂowing away from the interface (divergent ﬂow) [33].23) (2. This results in: 1 Dρα 1 Dρβ ∂α + U · α = −αβ − ∂t ρα Dt ρβ Dt 1 Dρα ∂β 1 Dρβ + U · β = −αβ − ∂t ρβ Dt ρα Dt (2. it will result in a negative/positive right hand side. since when using the VOF method it is essential to ensure that the interface is not dispersed. the phase densities will remain singlephase if either α or β is zero. meaning that it will only be nonzero around interfaces. where the convection term is evaluated and 2. and if care is taken with regard to sign of the right hand side. This involves not only use of a speciﬁc scheme. It is also essential for this term to reduce to zero as the grid is reﬁned. depending on the sign of the term within the parenthesis. Further. and then calculates β through equation 2.14 CHAPTER 2.20 is used to eliminate the divergence term.23. since the distortion of interfaces and associated problems are larger for coarse grids. causing cubes to lose their shape and balls to become “wiggly”. Equation 2. NOZZLE FLOW ∂ρα + ∂t ∂ρβ + ∂t · (ρα U) = ρα β 1 Dρα 1 Dρβ − ρα Dt ρβ Dt 1 Dρβ 1 Dρα · (ρβ U) = ρβ α − ρβ Dt ρα Dt (2. the equation will still be conservative. This is usually done by using a compressive diﬀerencing scheme such as CICSAM [33]. which compresses the interface. The code only solves equation 2.21) (2. but also the introduction of a new term. when the equations are solved for ρα and ρβ .
see Appendix A. BOUNDARY CONDITIONS FOR PRESSURE 15 The success of the model now depends on the choice of Uc . . as well as the calculated local velocity. 2. λ = min(max[1 − max{(1 − 4αP (1 − αP ))2 . which switches between upwind and central diﬀerencing depending on the limiter function λ. The same is true for the inlet.3.e.3. the liquid/gas interface is maintained not only by use of a speciﬁc compressible numerical scheme but also by introducing the mentioned term. pos is a function returning the positive part of this velocity (which is a scalar). as follows pinlet = p0 − pos(φ))U2 ) (2. The method must compress interfaces even when no ﬂow is present. 1) (2. If a nonreﬂective outlet condition is not used. when U is low.26)  α As mentioned above. Therefore special care has been applied to the boundaries. and thus be directed in  α . There are many options for ensuring this. The scheme used is implemented in OpenFOAM under interfaceCompression. while ψ and U are read from the ﬁeld. Input for the model is p∞ and l∞ . which should only α compress interfaces.2 Outlet The nonreﬂective pressure boundary condition implemented in OpenFOAM is a simpliﬁcation of that proposed in [34].2. i.28) (1 + 1 2 ψ(1 where: p0 is the pressure set by the user.3.1 Inlet The total inlet pressure takes into account the compressibility of the mixture. (1 − 4αN (1 − αN ))2 }. 0]. max(U)) (2. 2. φ is a scalar volume ﬂux on the face (for more information on what φ really means. using a simple ﬁxed value for the pressure at the inlet is insuﬃcient. the boundary conditions become very important. 2.1).27) where subscript P refers to the current cell value and N the neighbour cell value. pressure waves from the nozzle will bounce oﬀ the outlet boundary and destroy the physical realism of the simulation.3 Boundary Conditions for Pressure When modelling a high Machnumber ﬂow. but the one chosen for this code is: α Uc = min(cα U.
NOZZLE FLOW which represent the outlet pressure and relaxation length scale.30) (2. The model begins by calculating the velocity of the outgoing pressure wave: 1 ψ ˆ w = U·n+ (2.29) where n is the outlet normal. more robust code was also developed as part of the previously mentioned threephase code for cavitating nozzles.33) where p0 refers to the pressure at the outlet in the previous timestep. This code simulates only cavitating liquid. The pressure wave velocity is used to calculate ˆ the pressure wave coeﬃcient α and the relaxation coeﬃcient b: α = w b = ∆t δ w∆t l∞ (2. The relaxation length scale is the parameter that governs how reﬂective the outlet will be. A singlephase code for a cavitating ﬂuid.4. which can be dependent on pressure and/or . In the results section some results obtained by applying it to water injection into a chamber are presented. 2. poutlet = ξptrans + (1 − ξ) pcell (2. respectively. is usually constructed by introducing variable density.1. a low value will give a more reﬂective outlet than a high value.31) where δ is the cellface distance coeﬃcient. The implementation is publicly available under the name cavitatingFoam in OpenFOAM 1. and is the one used in the studies presented in Paper 2.16 CHAPTER 2.34) This ends the description of the multiphase code for cavitating ﬂow. instead a mixture of ptrans and the pressure in the cell (pcell ) closest to the outlet is used.4 A Code for Cavitating SinglePhase Flow A simpler. The value of the pressure can now be calculated using these properties: ptrans ξ = = p0 + bp∞ 1+b 1+b 1+α+b (2. The pressure at the outlet is not set to ptrans .32) (2.
Like the equivalent threephase EOS. the equation of state is said to be barotropic. is introduced to describe the amount of cavitated vapour in each cell. If it is only dependent on pressure.35 can either be used directly in the continuity equation to formulate a pressure equation. ρv. the equation of state for the singlephase cavitation code should therefore obey the liquid and vapour equations of state both in the limit cases when there is pure liquid or pure vapour. In this case we have assumed the temperature is constant (in accordance with reference experiments [15]). The latter approach has been applied by Schmidt et al [28].4. The former approach is problematic when it comes to consistency.sat = ψv psat (2. while in intermediate cases between these states it must have some form of mixture. an equation of state is required to model cavitation. As mentioned above. it is also convenient to assume homogenous equilibrium. A common barotropic equation of state is then the nonequilibrium diﬀerential equation: Dp Dρ =ψ Dt Dt (2.35 is not an equilibrium equation of state) between the pressure and density obtained from it with the liquid and vapour equations of state before equilibrium is reached.38) γ = 1 corresponds to a fully cavitated ﬂow. Since it is implemented in a ﬁnite volume code. These properties together form the mixture’s equalibrium equation of state: . A CODE FOR CAVITATING SINGLEPHASE FLOW 17 temperature.36) (2.sat − ρl.2.39) where ψv is the compressibility of the vapour. A property.sat is calculated from ρv. or integrated to obtain the pressure as a function of the density. The former approach is complicated by the lack of consistency (since equation 2. and γ = 0 a ﬂow without cavitation.sat ρv. which means the liquid and vapour are assumed to be always perfectly mixed in each cell. Thus. γ. until equilibrium is reached. The two states have the following linear equations of state: ρv ρl = ψv p = ρ0 l + ψl p (2.37) The property that describes the proportion of liquid in each phase is γ: ρ − ρl. and the code will not produce accurate results when it has reached equilibrium. errors from the inconsistency accumulate.35) This equation 2.sat γ= (2.
For reference. The saturation density of the vapour.sat is increased. since only one compressible phase is then present: µf = γµv + (1 − γ) µl (2. ψ. ρ will not need to be so low in order to obtain a higher γ.16 was chosen. used earlier to calculate γ. For the singlephase code. If ρv. The ﬁrst term contains the property ρ0 which is: l ρ0 = ρl. the equation of state (2. we have the compressible continuity equation: ∂ρU + ∂t · (ρU U ) = − p + (µf U ) ∂ρ + ∂t · (ρU ) = 0 (2.44) (2.41) When a linear model is used for the compressibility.sat is the liquid density at standard conditions.18 CHAPTER 2.2. For the mixture’s viscosity. NOZZLE FLOW ρ = (1 − γ) ρ0 + (γψv + (1 − γ) ψl ) psat l +ψ(γ) p − psat (2.46) .40) can be simpliﬁed to: ρ = (1 − γ) ρ0 + ψp l (2.40) The compressibility.2 can be used. is important for the liquid’s tendency to cavitate. were mentioned in section 2. and choice of how it is modelled. and for the simulations in Paper 2 the linear model in equation 2. If the ﬂuid is cavitating. with equation 2. the second term becomes more dominant. a simpler version of equation 2.42) The ﬁrst term governs the liquid density when γ is low. Now. The relationship between density and pressure are used in the continuity equation to transform it from a density equation to a pressure equation.45) Since this is compressible ﬂow.43) where ρl. the considerations when choosing a model are the same.40 we get: ∂ψp ∂γ ∂γ − ρ0 + (ψl − ψv ) psat l ∂t ∂t ∂t ∂ψ −psat + · (ρU ) = 0 ∂t (2.sat − psat ψl l (2. it is more straightforward than incompressible ﬂow.
velocity proﬁle. Smaller droplets are then formed from the cavitation layer itself.2. D . curvature of the inlet. A CODE FOR CAVITATING SINGLEPHASE FLOW 19 This equation can be combined with the numerical discretization of the momentum equation to obtain the continuity equation for pressure. 2. This assumption and the points raised in the brief discussion on the cavitation layer have all been mentioned before in the literature ([35]. and it ends where the liquid core begins. Thus. there will be more mass in the centre of the injector. Huh [35] only considers the results of turbulence. and cavitation layer classiﬁcation The theory behind the model is that if the cavitation layer (δcavitation ) is thin. and the Max Planck Institute (MPI) model. described in [36]. This equation is solved in conjunction with the momentum equation in a PISO loop. However.[36]).1 Development of Injection Model These simulations were originally intended to provide input for the development of a new injection model for EulerLagrangian diesel spray simulations.1: Schematic diagram of a cavitation proﬁle.1.4. velocity ﬂuctuations and the discharge coeﬃcient L vary with changes in injection pressure. and the model will return large initial droplets.4. The features of interest were the ways in which cavitation layer thickness. this could also result in a smaller spray angle. instead it relies on stabilizing numerical schemes. This can be regarded as implicit LES if small cells are used. A schematic diagram of the development of a cavitation layer in a nozzle can be seen in ﬁgure 2. a thin cavitation layer can still result in a large spray angle. as normal. etc. δcavitation δcore δcavitation Figure 2. The size of the cavitation layer is the size of the vapour layer close to the wall. if the turbulence is suﬃciently high. which is located where the freestream velocity is reached. Since large droplets are assumed to originate from the centre of the spray. This distinguishes it from the boundary layer. This code does not use any turbulence model. and the distance they move from the spray axis is governed by the turbulence in the nozzle. only considers the eﬀect of cavitation in the .
the same model was used for the singlephase code. The code worked well for water injections into gas at atmospheric air pressures.1 1 It should also be noted that it is only in special cases (mainly sprays with low evaporation. The aim of the new model was to account for both cavitation and turbulence.1 . liquid and air. the equation of state in the code assumed that the vapour density was proportional to the pressure. This forced the use of a unphysically high saturation pressure to avoid negative pressures in the solution and allow the vapour density to remain below the liquid density. as mentioned earlier. This issue will be addressed in section 4. but for diesel pressures it caused the vapour density to rise above the liquid density. NOZZLE FLOW nozzle. but the EOS was not suﬃciently smooth to handle the large pressure gradients in the ﬂow. The compressibility model was also changed.20 CHAPTER 2. The combined eﬀects of these factors caused long delays in the simulations. It starts oﬀ by using what the papers refers to as the Schmidt model [28]. During the course of the project. and forced the author to limit the work to validating a model for cavitating singlephase ﬂow. Initially. several attempts were made to simulate cavitating diesel nozzle ﬂow.2. e. without being able to develop a model for Lagrangian diesel spray simulations. The change was necessary partly to enhance stability and partly because the same cavitation model was to be used for the multiphase code handling diesel vapour. A similar story of cavitation code development is reported in [38].g. from the Wallis [31] model to a linear compressibility model. Attempts were made to extend this code to use PengRobinson equation of state (EOS) [37]. Since a linear combination of the vapour fraction (γ) and liquid/gas fraction (α and β) is highly suitable for describing the mixture’s properties in such cases. and ends in a modiﬁcation of that model with a cutoﬀ for negative pressures. and to estimate their magnitude using CFD. in cold starts) that the size of the initial droplets and their maximum spreading angle really aﬀect the combustion.
more speciﬁcally the conditions close to the injector.Chapter 3 Spray Modelling In section 1. All the papers mentioned above. Lagrangian spray simulations were brieﬂy described. but even if such small cells are only used around the injector the number of cells required grows enormously if the calculations are run in 3D. An option to improve nearinjector ﬂow is to use the previously mentioned (in section 2. They also used lower velocities than those of diesel sprays and assumed the surrounding air to be incompressible. This is based on the fact that we cannot resolve the full details of the nearnozzle ﬂow using only Eulerian cells (as yet at least). For instance. One of the biggest problems associated with the DDM is modelling the nearnozzle ﬂow. Thus.2. they did not simulate the nearnozzle behaviour. There has been a growing number of attempts to simulate sprays using a full Eulerian approach. Befrui et al [40] uses an updated version of the same code to study primary breakup in noncavitating gasoline direct injection. The method used to introduce Lagrangian particles into an Eulerian grid is sometimes called the discrete droplet model (DDM). so if it is to be resolved properly a cell size of about 20 µm is needed. Aside 21 . However. but the Reynolds number was comparable to those of diesel sprays. These attempts are often associated with some limitations. and M´nard et al [41] considers liquid e injection into incompressible air with a velocity of around 100 m/s. A typical nozzle has a diameter of around 200 µm.1) ELSA model. by Vuorinen et al [42]. and mainly focused on studying the spray behaviour. or by using LES instead of the classical DDM with a kε turbulence model. ignore cavitation. since the DDM assumes disperse ﬂow with spherical droplets of liquid. it is not an accurate representation of the spray in this region. the spray is usually modelled in some other way. A spray study using LES and an EulerLagrangian description of the gas and liquid has also been published. In this region a liquid core forms from the liquid fuel being injected through the injector. and the simulations cover only part of the relevant physical domain. Hence. de Villiers et al [39] considers primary breakup close to a noncavitating nozzle. Ligaments are separated from this liquid core and form droplets that evaporate and mix with the ambient gas.
a perfect turbulence interaction can be ruined by a dissipative numerical scheme. The complex interactions amongst all of the submodels make spray simulation somewhat more of an art than a science. Furthermore. tuning the spray model in a noncombusting environment to constant volume experimental data. which will be described later. SPRAY MODELLING from the work by [24] (mentioned in 2. the density ρ becomes more complex since diesel combustion involves multiple species. by both diﬀusion and convection.1) The evaporation source term is calculated from the evaporation submodel. Blokkeel et al [43] and Lebas et al [44] have also used it to improve the primary breakup of the spray. thus nullifying the excellence of the breakup model. since chemical reactions are occuring various compounds will be formed and participate in further reactions. 2. Therefore. In addition. When chemical reactions and interactions between the Lagrangian and Eulerian phases are also involved. Eq. each species will be transported. Since the DDM model is only used when the liquid fraction is low. but here my intention is to mention the interactive problems they raise as a group.8 needs to be extended to include liquid spray evaporation. the parcels can become too small due to the evaporation model simulating evaporation too quickly. However.1). For instance. so the mass fraction of each species is not conserved. However. and may also be formed or . representing a state in which droplets (not ligaments) have been formed from the liquid core. the EulerianLagrangian description of the diesel spray has been the most widely used. 3. if we consider the total density. in 2D ([45]). and thus there are fewer model constants to tune. However.1 The Gas Phase Basic Equations The Eulerian grid is used to discretise the conservation equations of the ﬂuid phase.22 CHAPTER 3. or an overpredicting drag model that slows all the parcels down too much.1. The large number of submodels used to simulate the various physical processes that the parcels are subjected to before they become part of the gas can pose further problems when using the DDM model. the model gives a better representation of the nearnozzle ﬂow. even if the breakup model is fully correct. Another beneﬁt is that such a modelling concept does not require a primary atomization model.1 3. Similarly. this raises some of the problems associated with high grid resolution to resolve the nozzle mentioned earlier. There have also been other attempts to simulate sprays using only Eulerian cells. and then applying the results of the tuning to real engine calculations often yields poor results. As already mentioned. all of which aﬀect the density. ˙ ∂ρ + ∂t · (ρU) = ρs ˙ (3. Each of these models will be thoroughly described later in this chapter. the only source term (ρs ) is from spray evaporation.
and has no wall functions. The three (compressible) models used in Paper 1 were: standard kε [46]. the transport equation had to be solved 83 times in each time iteration. Renormalized Group kε [47] (RNG) and LaunderSharma’s kε (LS) [48] models. . but since all computations in this work simulated events in a bomb. The LaunderSharma turbulence model diﬀers in this respect. For later simulations diﬀerent turbulence models were also tested. Paper 1 presents constant volume vessel simulations of diesel sprays conducted using three diﬀerent turbulence models. and when combustion occurs their number increases dramatically. bearing in mind the intial results.2 Turbulence Modelling Since the goal of this project was to develop new models for industrial applications. since it is a low Reynolds number model. This places higher demands on the nearwall cell size. when the mechanism including 83 species was used for the calculations shown in Paper 3.1.5 and 3. which yielded results that substantially diﬀered in terms of spray shape and penetration of liquid and vapour.6 in a general form to accommodate the three turbulence models. meaning that they apply wall functions for the near wall cell. The mass fraction for species i is denoted by Yi . The RNG model and the standard kε model used for these cases are both high Reynolds number models.2) (3. For instance. THE GAS PHASE consumed by the reactions.1. 3. and its transport equation is: 23 ∂ρYi + ∂t · (ρU Yi ) − ((µ + µT ) Yi ) = ρsi + κRRi ˙ ρs = ˙ i (3.4) 2 Fs is the spray momentum source term. the nearwall behaviour was not as important as the spray interactions by any of the walls except the injectorwall. The discrepancies between the models indicate the importance of modelling the turbulence correctly. The momentum equation also needs to be modiﬁed due to the presence of a Lagrangian spray. Presented here is the equation for the mean velocities using a Reynolds Average Navier Stokes assumption: ∂ρU + ∂t · ρUT U = − p+ · 1 (µ + µT ) ( U + ( U)T ) +ρg+Fs (3. Each of the three turbulence models was developed with a diﬀerent rationale.3. The ﬁrst turbulence model based on k and ε was originally presented in [46]. It is shown in equations 3.3) ρ si ˙ since most fuels consist of multiple species. the turbulence was modelled using kε models.
the property ld aﬀects whether particles will increase or t decrease the intensity of the turbulence. A value above 0. For most diesel spray simulations. indicating t that some parcels will increase. . According to experimental investigations of gassolid and gasliquid ﬂows summarized in 3.1.1 indicates that the intensity of the turbulence will increase. but not the only one. 1 According to the kε turbulence model.24 CHAPTER 3. Diﬀerent researchers have adjusted their models in diﬀerent ways to incorporate the eﬀects of particles on turbulence. i kε RNG LS Bi −C3 ρε · U −2µT RRNG Sε/k 2 2 −C2 f ρ ε + 2µµT ∆U k ρ Di 0 0 √ 2µ k2 Table 3.6. Including a source term in the equations is the most common way to achieve this.6) C1 2µT S ε 2 ε2 − C1 ρε · U − C2 ρ + Bi k 3 k where Bi and Di are listed below. so the turbulence is not directly aﬀected by the presence of particles. Figure 3.1: Speciﬁc terms of each turbulence model 3. However. and others reduce.2 shows the distribution of the quantity ld . An example of a model that uses a source term in the equations for the turbulent quantities is KIVA ([49]).1. SPRAY MODELLING ∂ (ρk) + ∂t · (ρkU) − · [(αk µT + µ) k] = · U − ρε − Di (3. and a value below the opposite. see ﬁgure 3. which is not consistent with the conclusions drawn from experiments by Crowe et al [50. centered around 0.1. that term can only reduce the kinetic energy of the turbulence. and the droplet diameter ranges from 100 µm (depending on the size of the injector) down to about 1 µm (there may of course be smaller droplets than this. 51]. Numerical constants and other variables are listed in Paper 1.5) 2 = 2µT S − ρk 3 ∂ (ρε) + ∂t · (ρεU) − · [(αε µT + µ) ε] = (3. which might not predict the turbulence entirely correctly. the length scale1 lt is of the order of 100 µm in cells where there are parcels.3 Turbulence/Spray Interaction There is no liquid spray source term in equations 3. the turbulence. but they will quickly evaporate).1.5 & 3.
2 0. The total number of parcels was 1018.9 diameter/L t 1 Figure 3.7 0. and Lt was approximated to 100 µm Aside from using new terms in the turbulence model. THE GAS PHASE 25 Figure 3. there are other ways to account for the presence of droplets in a simulation. one of a list of suggestions for improving the performance of spray simulations in a book by Stiesch [12] is to limit the length scale of the gas phase turbulence inside the .1.4 0. both combusting and noncombusting.2: Distribution of parcels from ﬁve spray simulations obtained using diﬀerent turbulence models.6 0.3. 400 number of parcels 300 200 100 0 0 0.5 0. Reprinted from [50] with permission from Elsevier Ltd.8 0.3 0. For instance.1 0.1: Changes in turbulent intensity as a function of length scale ratio.
k governs the scale of the ﬂuctuating turbulent velocity. SPRAY MODELLING spray to the jet diameter in order to reduce the grid dependency of the calculations. the nozzle diameter is a good ﬁrst approximation. which is what this modiﬁcation does. The constraint is imposed after the equations for k and ε are solved. it was limited by: Cµ k 3/2 <ε Lsgs (3. but it also simpliﬁes the calculations. If the constraint is applied before any of the equations are solved. the jet diameter is not well deﬁned. which is perhaps easier to imagine. There can be cells that do not have a parcel in them. The limit was thus imposed on ε. or the vapour part could also be included. and this explanation will be used in the rationale below. The parcels in a diesel spray problem have fairly high velocities. The length scale of the turbulence is deﬁned as: lt = C µ k 3/2 ε (3. Notably. could refer to just the liquid core. Thus. While this is true. the suggestion raises several problems. One could also think of this as overprediction of the length scale of the spray. the resulting length scale of the turbulence should also be governed by the spray and nozzle size. and all cells that have parcels in them must be inside this cloud. it could refer solely to the liquid part of the spray. the term “inside the spray”. However. that the equation for ε underestimates the dissipation when liquid parcels are present. and ε the size of the turbulent eddies. The nozzle diameter is selected rather than the jet diameter because the nozzle diameter is the parameter that initially sets the length scale of the turbulence.7) Thus. Such a modiﬁcation would also cause the turbulence model to be aﬀected by the particles.26 CHAPTER 3.8 cannot be guaranteed. . and the increased momentum of the gas comes from the spray. The modiﬁcation of the turbulence model implies an assumption regarding ε. since the limit was imposed on ε.8) Where Lsgs is set to the nozzle diameter. since calculating the jet diameter in every timestep would be costly. The spray is a cloud of parcels. Thus. it is a much cheaper dependency computationally. or include the vapour cloud. the constraint in equation 3. Fixing the limit prevents the turbulence length scale from growing. Similarly. The author chose to limit the turbulent length scale to the nozzle diameter in cells with parcels. One could argue that this only shifts the dependency from the grid to the magnitude of the turbulence length scale limit. but if we choose a high number of injected parcels there will be few such cells.
2) where S f and S r are the matrices of forward and reverse stoichiometric f r coeﬃcients. making the matrix Ns × Nr .10) This equation is formulated for every species included in the chemical mechanism.j RT (3. when studying combustion. Solving the chemistry numerically means solving a large system of reaction equations. and [Xi ] is the molar concentration of species i in the cell.1. One such model fuel is nheptane. For each reaction Sf [X1 ] + Sf [X2 ] + . reaction j is written as: Ns Sf [Xi ] ij i=1 kf j kr j Ns Sr [Xi ] ij i=1 (Re. The reaction rate constant k is itself a function of the Arrhenius constants: k j = A j T βj e− Ea.. either experimentally or numerically. The rate of formation of species [X1 ] from reaction j is written as: d[X1 ] dt Ns r = S1j j f kj i=1 f Ns i=1 r [Xi ]S1j − kj [Xi ]S1j r (3. resulting inn an equation system . 1. To formulate the reaction in a more general manner. 1) there is a corresponding reaction rate equation which determines how rapidly the reaction is proceeding and in which direction. Thus..4 Chemistry After the fuel liquid droplets have undergone breakup and evaporation. another is the IDEA model fuel consisting of ndecane and αmethylnaphthalene. the fuel mixes with the surrounding air and forms a combustible mixture. A real diesel fuel consists of about 1000 diﬀerent species. It is now possible to write the equation for the reaction rate of the basic reaction. with the rows corresponding to species. the complexities are generally reduced by considering the combustion of a simpler model fuel. The columns represent reactions.. respectively. THE GAS PHASE 27 3. − 11 11 21 21 (Re. kj and kj are the corresponding reaction rate constants of reaction j. and nheptane for the studies described in Papers 3 and 4.9) which need to be speciﬁed as part of the mechanism. Both have been used in this work: IDEA for the spray studies described in Paper 1.1.3. Re. reacting with the air in over 10 000 reactions. The matrix of stoichiometric coeﬃcients consists of Ns rows. → Sr [X1 ] + Sr [X2 ] + .. as well as for every reaction.
The analytical function used to approximate y∞ is a polynom.12) As mentioned. . it will approximate it. they are too large to use in a CFD simulation. This mechanism has to solve 83 systems of ODEs in every cell. detailed chemical kinetics with a perfectlystirred reactor [56]. SPRAY MODELLING consisting of Ns × Nr equations. OpenFOAM has the ability to solve the equations using an ODE solver (an approach used in Paper 3). Aside from the concentrations. which is based on Richardson extrapolation of the approximated solution. which can be solved coupled using an ODE solver. and the error function of the method contains only even terms of the step size. and each species requires a transport equation to be solved. the solver will never apply enough substeps to ﬁnd it. One example of a suﬃciently small mechanism is the reduced one used in Paper 3. but in theory the system could comprise 28054 ODEs in total. each reaction requires a term describing the amounts of each species it forms. As mentioned earlier the solver needs to solve ODEs in every timestep to determine the chemical species’ concentrations at the end of the timestep. it is also important to ﬁnd the right hand side of equation 3. depending on the solution using large substeps. sequentially using a reference species technique [52]. a stiﬀ ODE solver is needed.10.11) [Xi ]Sij r Ns i=1 ωj ˙ f = kj i=1 r [Xi ]Sij − kj (3. While it is possible to use such mechanisms in shock tube simulations. The Richardson extrapolation of the function y assumes that as the interval (in our case the computational timestep ∆t) is split up into increasing number of substeps. Some versions of KIVA3V also oﬀer the option of solving the chemical rate reactions coupled. in the literature there are many suggestions for doing this.2). one of them being the KIVA3V version also used in Paper 3.28 CHAPTER 3. a real mechanism for hydrocarbon fuels (including model fuels) would be very large. Chalmers PaSR Model It is necessary to use some form of treatment for the chemistry and turbulent mixing. or by an EulerImplicit method. The source term for species i is: RRi = Wi ρ Nr j=1 Ns f r (Sij − Sij )ωj ˙ f (3. it is a system of Ordinary Diﬀerential Equations (ODEs). To solve the chemical reaction equations. since it is used in the source term in the transport equation (equation 3. in every timestep. since no reaction includes all species. Some of the equations in the system will not be applicable. However. Instead. diﬀusion ﬂamelets [55]. The one used in the studies presented in Paper 3 is the SIBS method (SemiImplicit Bulirsch Stoer [53]) in OpenFOAM. the solution will converge to some value y∞ . including: PaSR ([54]). consisting of 338 reactions and 83 species. and the general ﬂame surface density model [57]. As mentioned in the above section. As can be seen from the above equation.
and ﬁnding the characteristic time for that system. The reaction rate term for species i is then approximated as: ∂ci ci − c i 0 = 1 = κRRi (ci ) 1 ∂t τ (3. This factor appears in the species transport equation with the chemical source term (as in equation 3.3.16) with Cmix = 0.2): ∂ρYi + ∂t (ρYi U) − (µEﬀ (Yi )) = (3. are governed by the turbulent mixing time and the residence time (the numerical time step. cell. in our case). THE GAS PHASE However. 59]. the species are assumed to be mixed due to turbulence for the mixing time τmix . The reacting part is treated like a perfectly stirred reactor. τ . deﬁned as: κ= τc τc + τmix (3. . and a nonreacting part.13) i Where RRi (c1 ) is the laminar chemical source term. so assuming that an entire cell is a perfect reactor is a severe overestimation. in which all present species are homogeneously mixed and reacted. it is assumed to be determined by k and ε: τmix = Cmix k ε (3. The relative sizes of the parts of the computational cell constituting the reactor and the rest of the cell. 29 The PaSR model is based on the theory that real ﬂames are much thinner than any computational cell.15) = Si + κRRi (ci ) where τmix is the turbulence timescale mentioned above. Other variations and detailed derivations of the PaSR model can be found in [52. and κ the reaction rate multiplier.1. and the resulting concentration gives the ﬁnal concentration in the entire. the cells are divided into a reacting part. the Chalmers PaSR model was used to model the turbulence chemistry interaction in Paper 3.14) where τmix is the turbulence timescale mentioned above. and τc the corresponding chemical timescale. The chemical timescale is determined by solving the reaction system’s ODEs fully coupled. 58. After reactions have taken place.03. partially stirred. 54. Thus.
often referred to as the BBO equation . and Fax´n5 force. The full spray equation. then: ∂ud πD3 ∂ud ∂pd = md = ρd ∂t ∂t 6 ∂t (3. while others like the Magnus eﬀect are neglected since rotation of the droplets will not be very important.18) This equation can be simpliﬁed further.30 CHAPTER 3. This means we can neglect the eﬀect that causes rockets to liftoﬀ and airplanes to ﬂy.2. What remains are the forces due to drag and gravity acting on the droplets.17 as: ∂pd πD2 πD3 = −ρg CD (ud − U) ud − U + ρd g ∂t 8 6 (3. simply because the force acting on the parcel is unknown. Magnus eﬀect3 . this equation cannot be solved immediately.18 and 3.17) Unfortunately.1 Spray SubModels Spray Motion Equation The motion of a Lagrangian particle. and that the drag will not be aﬀected by changes in mass. Most of the e terms can be neglected due to the high density ratio between the two phases. the result is: ∂ud 3 ρg 1 =− CD (ud − U) ud − U + g ∂t 4 ρd D (3.19. SPRAY MODELLING 3. 3 The force that causes a soccer ball to rotate perfectly past defenders and into the net 4 Lift force due to shear 5 force due to curvature of the ﬂow 2 This . These simpliﬁcations allow us to write the right hand side of equation 3. is governed by one of the most fundamental laws of physics. If we combine equations 3. They are all neglected. pressure. moving in an Eulerian framework. but we assume that the changes will not aﬀect the drag. The latter is included for simplicity and the former due to its physical importance.2 3.from Basset (1888). a momentum relaxation time is introduced: is the force that causes a motorboat to continue in its path even when the motor has been switched oﬀ recently. an eﬀect that is not believed to be pronounced in diesel droplets from which some mass is evaporating. Boussinesq (1903) and Oseen(1927) .includes eﬀects of added mass. if we assume the droplets are spherical.19) A droplet’s mass will of course change with time (due to breakup & evaporation). Saﬀman force4 .20) To simplify this equation further. Basset force2 . Newton’s second law: ∂pd = ∂t Fi i (3.
This is done to ensure that particles that might be close to the edge of the cell (or even slightly outside) are properly tracked.21) Equation 3.22) Even with the introduction of τu . and update the parcel properties Add the momentumchange to the cell that the parcel has been in If the parcel still has time left to move.2. FaceToFace tracking. There are two main ways of doing this. the particles representing liquid are moving in a ﬁxed Eulerian framework as described above. calculate the time it took to move out of the ﬁrst cell. The FaceToFace tracking in OpenFOAM includes a stability check that begins by tracking the parcel from the centre of the cell it belongs to. . SPRAY SUBMODELS 31 τu = 8md 4 ρd D = 2 u − u πρg CD D d 3 ρg CD ud − u (3. 3. rather than from the particle’s position. 3.20 can now be written as: ∂ud ud − U =− +g ∂t τu (3. Update the properties of the parcel Move the parcel with velocity ud for time ∆t Find out which cell the parcel is in Add momentum to the cell to which the parcel has moved Assuming that the time step is suﬃciently small to ensure that every parcel traverses more than one cell. Tracking them and deﬁning the cells they are in are clearly important issues. one called the LoseFind algorithm. however. Move the parcel until it reaches a cell boundary or for the entire time step ∆t if it remains in the same cell If the parcel changes cell. which will be introduced in a later section in this chapter.3. the drag cannot be approximated without a drag coeﬃcient. 1. even for a small time step. It too can be described in four steps: 1. However. this approach is acceptable. 3. go back to 1. The Loseﬁnd algorithm is fairly simple and can be described in four steps. 2. as well as some commercial CFD codes. a parcel can still be moved across a cell without having any momentum exchange with it. another model is used in OpenFOAM.2. Due to the problems mentioned above. and one called the FaceToFace algorithm.2 Parcel Tracking In Lagrangian spray simulations. 2. The model is used in KIVA3V. 4. 4.
23) where ρi is the density of the injected parcel. and composition of fuel (for multicomponent fuels) into account. a parameter that varies from nozzle to nozzle.32 CHAPTER 3.2 Model parameter dmin dmax dmean n βmax Numerical Value 10−6 m dnozzle dnozzle 3 20 Table 3. For a more detailed description.3 Injection Model The injection model used in the Lagrangian spray simulations is a solidcone injection model. This angle is the angle between the set spray direction and the direction of the injected droplet.2.2: Solid cone injection model constants The direction of a droplet to be injected into the domain is calculated by multiplying βmax and a random number between 0 and 1. in this case a RosinRammler form was chosen. The user supplies a drop diameter probability density function (PDF) with parameters.2. The model also requires minimum and maximum values for the droplet size. taking the type of fuel. and is to a certain degree pressuredependent [15]. which improves the predictions of transfer of mass.[61]. but the full spray angle.24) 0.4 Drag Model Several options for the drag coeﬃcient have been suggested in the literature ([36]. as well as the maximum spray cone angle. see [60]. The velocity of the injected parcel is based on the injection pressure. momentum and energy. and the pressure in the domain: 2∆p ρi ud = C d (3. Note that βmax is not the halfangle of the spray. 3. SPRAY MODELLING Parcels tracked by FaceToFace tracking cannot ’skip’ cells. Cd is the discharge coeﬃcient.[62]).44 . The parameters for the model can be found in table 3. 3. but the one chosen here was: 3 Red 6 2 CD = 24 Red 1+ Red < 1000 Red > 1000 (3. injection pressure.
4Ta0. The KHRT model includes two modes of breakup: KH breakup.distort min(ylim . The modiﬁcation is CD = CD. accounting for waves growing on the droplets’ surface due to acceleration normal to the dropletgas interface.25) where CD. and RT breakup. wavelength ΛKH and droplet radius. 3.7 ) (1 + 0.28) (3. the instability waves eventually resulting in droplet breakup.distort and ylim are userset constants.2. since a yvalue of 1. y)) (3.e.26) (3.34 + 0.0.865We1.38We 2 ) = (1 + Oh)(1 + 1.[66]).3.0 will result in the parcel breaking up. which is the radius a parcel will have after it has been stripped oﬀ the liquid jet. i. The minimum function is not necessary when the TAB breakup model is used. The KHRT model was chosen for its superior performance under diesel conditions ([67]). the TAB model [9] will be used to calculate the oscillations and the resulting drag will be added to the drag described above. Analysis of a round liquid jet yields a fastest growing wave with a wavelength (ΛKH ) and a growth rate (ΩKH ): r(1 + 0. is one of the most widely used models in Lagrangian spray simulations today ([64].67 )0.726B1 r τKH = ΩKH ΛKH (3.[65]. If this model is used.2.29) The continuous reduction of droplet size due to breakup can then be deﬁned by the rate equation: . SPRAY SUBMODELS 33 OpenFOAM also oﬀers the possibility to account for changes in drag due to oscillations of the droplet surface.model(1 + CD. and y is the relative deviation from the equator of the parcel if it were spherical.02 ΩKH (3.45Oh 2 )(1 + 0.5 Breakup Model The breakup model used is the KelvinHelmholtzRayleighTaylor (KHRT) model ([8] & [63]).6 ) σ ρd r 3 The critical droplet radius rc .27) (0.4Ta0. or characteristic breakup time τKH (which controls the breakup rate) is assumed to be a simple function of the growth rate ΩKH . along with the TAB model. This model. accounting for unstable waves growing on the liquid jet due to the diﬀerences in velocity between the gas and liquid. ylim is usually set to 1.6 3 1 ΛKH = 9. r: rc = B0 ΛKH 3. but often in conjunction with some form of atomization model since the TAB model tends to break up the droplets very rapidly. It is also possible to use the TAB model. is assumed to depend linearly on the wavelength ΛKH and the stripping rate.
is governed by how rapidly disturbances grow on the surface of the droplet.32) (3. which is the total injected liquid mass divided by the total number of injected parcels. For real droplets these disturbances originate from the droplets’ trailing edges [12]. The second parameter controls the rate of introduction of new childparcels. CRT determines both whether or not the droplets are going to be converted into smaller drops and how large the new droplets are going to be. a new parcel will be introduced with size rc . 3. As the jet breaks up the parcel will accumulate a stripped mass. the frequency of the fastest growing wave.34 CHAPTER 3. the droplet will be immediately converted into a parcel with smaller droplets when the growth time exceeds the characteristic RT time. SPRAY MODELLING r − rc dr =− dt τKH (3.31) (3. it only increases the statistical number of droplets in the parcel and changes the radii of the droplets. The RT mode does not add any new parcels to the system. The breakup parameters B0 and B1 can be found in Table B. the key parameter is the lifetime. the drop has reached a stable state and no breakup due to Kelvin Helmholtz instabilities will occur.33) gt = (g + ΛRT = 2πCRT 3σ gt (ρl − ρg ) Where CRT is a modelling parameter. The derivation of this time starts by evaluating the timederivate of the mass of the droplet. on the other hand.1. there are two more parameters to control the KH breakup. and the corresponding wavelength. or evaporation relaxation time of the droplet. . Thus.30) which essentially describes how rapidly the drops will reach the stable state. If W e < W elimit .5 √ 3 3σ(ρg + ρl ) dud ud )· dt ud  (3. is given by: ΩRT = 2gt (ρl − ρg )1. containing the accumulated mass ms . ms .2. The new size is proportional to the wavelength ΛRT of the disturbances. and when ms exceeds a certain percentage of the average parcel mass. The ﬁrst determines whether breakup will occur at all. In addition to these two constants. If this wavelength of the fastest growing wave is smaller than the diameter of the droplet diameter and the perturbations are allowed to grow for a period of time. If these disturbances are assumed to be linear. RayleighTaylor breakup.6 Evaporation Model When modelling the evaporation of a droplet.
39) Using these two equations for expressed as: the value of the constant Ce can be Ce = −4Sh ρv Xv. (3.6Re1/2 Sc1/3 (3.37) The only unknown in this equation is the constant Ce .∞ Dln 1 + ρd 1 − Xv. From this equation we can determine the relaxation time by integration (since Ce is a constant).∞ = p − pv. we can put D = 0 into the above equation. 2 D 2 = Ce t + Do (3.2. the evaporation is assumed to follow the empirical D 2 law: dD Ce dD2 = Ce ⇒ = dt dt 2D (3.34) Further.s (3.s − Xv.34.38) However.s − Xv.s Xv. resulting in: 2 D0 Ce τe = − (3.35) where Ce is a constant.40) The Sherwood number Sh is calculated using the RanzMarshall (or Fr¨ssling) o correlation (from [51]): Sh = 2 + 0. SPRAY SUBMODELS 35 dmd d = ρd dt dt 4 π 3 D 2 3 = ρd π 2 dD D 2 dt (3.36) If we want to determine the time (τe ) when a droplet of size D0 has completely evaporated.41) .s dmd dt . the result is: dmd ρd π = DCe dt 4 (3.35 into equation 3.∞ = −πDShDρv ln 1 + 1 − Xv.3. which will be determined by inserting equation 3. the time derivative of the mass of the droplet can also be expressed as ([52]): dmd = −πDShDρv ln dt p − pv.
46) The ﬁrst part of equation 3. since the evaporation of a droplet also transfers heat to the surrounding gas.∞ 1−Xv. πDκc Nu 1 hv (Td ) dTd = (Tg − Td ) fheat − dt md c l cl τe The modiﬁcation is in the factor fheat which is: p. or the mass will become negative. The heat transfer model also uses the evaporation relaxation time.36 CHAPTER 3.42) τe = (3. If an explicit method is used.v d − πDκc Nu (3.s −Xv. SPRAY MODELLING Equation 3. the relaxation time introduced in 3.7 Heat Transfer The heat transfer model comes in part from [51]. A modiﬁcation was made to account for the latent heat phase transfer during evaporation.s (3.35): dD D =− dt 3τe (3.45 accounts for the heat transfer to the liquid.39: md dmd =− dt τe where the relaxation time is ρd D 2 6DShρv ln 1 + Xv.v m ˙ (3. 3.6 is used. and the second part for latent heat arising from mass transfer. It is also consistent with the way the momentum and heat transfer equations for each parcel are solved.43) and a similar equation for the diameter (from 3.44) A relaxation time is introduced to characterise the evolution of the particle size.37 can now be used with known Ce to solve equation 3. If an implicit method is used.2. the method is unconditionally stable. Since the mass transfer is related to the evaporation of the droplet. The diﬀerence in timescales of these processes is useful for determining the rate at which each process is proceeding.2. the time step (a function of the relaxation timescale) must not be larger than the relaxation time.45) c m ˙ fheat = e− πDκc Nu − 1 cp. and is based on the convective heat transfer of a particle with a uniform temperature. The heat transfer can also be expressed with a relaxation time: .
If the number is lower than the probability. This is based on the O’Rourke collision model. there will be few collisions in very small cells. regardless of the mesh. parcels several droplet radii away will be able to collide. As mentioned previously. 69]. although the probability of this will be lower. using adaptive collision meshing. Other modiﬁcations have also been made to the O’Rourke model [70].N2 ) 4V1 π(d1 +d2 )2 U1 −U2 ∆t cell (3.48) where one term describes heat transfer due to temperature diﬀerence. the algorithm ﬁrst calculates whether a pair of parcels is heading . and the modiﬁcation by Nordin [52]. The mesh is reﬁned until the number of parcels in each cell is below a set threshold. In the original version. At the other extreme. parcels collide if they are in the same cell with a probability given by: e − min(N1 .3. The outcome of the collision depends on the Weber number. according to equation 3. the parcels are set to collide. In the simulations presented in the included papers and the results chapter the trajectory collision model by Nordin [52] was used (unless otherwise stated). Instead of only colliding parcels that are in the same cell. since parcels will not often occupy the same cell. SPRAY SUBMODELS md c l πDκc Nu 37 τh = resulting in: (3. In all. 3. which will be brieﬂy described. a random number between 0 and 1 is generated for each parcel. If the cells are large.2. the direction of the parcels has no inﬂuence on whether they collide or not. Thus the parcels need to be in the same position. the O’Rourke model has very strong grid dependency. as shown in [52. resulting in either coalescence or grazing.49) In practice.49 (due to the large cell size).2. the algorithm calculates the paths of all parcels and collides those that will intersect within the same timestep. Other models have also been published. Since a simulated spray may contain several thousand parcels. Thus.47) (Tg − Td ) 1 hv (Td ) dTd = f− dt τh cl τe (3. the model implemented into OpenFOAM used for the work presented in the appended papers is a modiﬁed form of the O’Rourke model. nor does their position within the cell.8 Particle Collisions OpenFOAM oﬀers two collision models: the O’Rourke model [68]. for instance the collision model by Schmidt et al [69]. which involves generating a separate orthogonal collision mesh with random orientation in every time step. and the other heat transfer due to mass of the droplet lost through evaporation.
new − x1. The two model constants Ctime and Cspace control the collision rate in time and space.38 CHAPTER 3. for instance. If they are not. if it is positive it indicates that the parcels are heading towards one another. they are on a perpendicular course. and on the right hand side the parcels’ relative displacement during the current timestep. The question that remains to resolve is whether or not the parcels reach the intersection at the same time. and if it is zero.new  (t ∂t2 (3.51) On the left hand side of this equation is the distance between the droplet surfaces.new (tα ) − x2.new = x1 + tU1 ∂x2.50) This property will give how much of the vector x2 − x1 can be projected onto the vector U1 − U2 . This reduces the computational load of the algorithm. The parcels also need to be close enough to collide since. So the algorithm needs to determine the times (designated tα and tβ ) at which the parcels reach the intersection. they may miss each other.new (tβ )) 1 2 (d1 Cspace e− ctime tα −tβ  ∆t (3.new = x2 + tU2 ∂x2.new − x2. one parcel has a high velocity and the other a low velocity.54) where x1.53) = tβ ) = 0 Whether the parcels actually collide or not is governed by chance. which must be larger.52) (3. If. This ensures that the parcels can reach each other within the given time (the numerical timestep). the probability for a collision is now given by: Pcoll = + d2 ) 1 max( 2 (d1 + d2 ). they are regarded as not colliding. this gives the next criterion: x2 − x1  − d1 + d 2 < vAlign ∆t 2 (3. the probability of collision increases. x1. as mentioned before. and is done by computing: (U1 − U2 ) · (x2 − x1 ) x2 − x1  vAlign = (3. As this value declines. this is done by solving: x1. they need to reach each other within the current timestep.new  (t = tα ) = 0 ∂t1 x2. respectively.new  is the smallest distance between the two parcels during the timestep.new −x1. . SPRAY MODELLING towards each other. if it is negative they are travelling away from each other.
The simulations were run on a grid representing the experimental nozzle shown in ﬁgure 4.Chapter 4 Results & Discussion 4. and the geometry was improved by adding more cells before and after the channel. As the boundary conditions were set in a more appropriate way.1. The ﬁrst converged results displayed a shockwave zigzag pattern developing through the channel. and the outlet pressure was varied between 1. Before and after the nozzle. The inlet pressure was set to 10 MPa.5 MPa and 5. there is a 1 cm wide channel.1 MPa.1 Simulations of a Diesel Flow in a Cavitating Model Nozzle Paper 2 describes simulations of the ﬂow inside a model nozzle using a new code for cavitation implemented in OpenFOAM.4. which was described in section 2.1: View of the experimental setup. the centre of the image contains the 1 mm model injector. the following cavitation pattern was obtained: 39 . was unstable and obtaining the results in this section required extensive work. Figure 4. The code.
2. It is very important for the code to be able to capture the maximum mass ﬂow as the ﬂow cavitates.6. as done previously. shown here for an injection pressure of 7 MPa. it also produced four oblong cavitation bubbles in the centre of the domain (see ﬁgure 4. as the case is barely cavitating.5 and 1 to allow the reader to distinguish between the two phases. Unfortunately. that calculating an average made hardly any diﬀerence. These vortices could be removed by introducing a slip condition on the upper and lower walls. in ﬁgure 4.5). RESULTS & DISCUSSION Figure 4.2: Image of the computed average vapour fraction. After the slip condition was introduced.40 CHAPTER 4. it did match pressure curves acquired experimentally [71]. injection pressure 7 MPa The pattern obtained. A routine was also implemented for calculating the cavitation probability. the change of scheme not only introduced instabilities in the wall cavitation bubble. After considering this image. was very smooth. since this is a fundamental property of nozzle ﬂow. and thus produce suﬃciently strong vortices to cavitate the ﬂuid.3) due to strong vortices. It should also be noted that this pattern has a scale ranging between 0.4 were obtained. The cavitation bubble in this case was so stable. Instead of using the limited linear scheme. . which the experimental data showed. A comparison of the experimental mass ﬂow rate and the calculated mass ﬂow rate is also shown. a change of convection scheme was introduced. instead of the mean vapour fraction as displayed in ﬁgure 4. and a noslip condition would cause the size of the layer to be overestimated. The physical rationale for the slip condition was that the boundary layer along the two walls is extremely thin. the results illustrated in ﬁgure 4. the monotonized central scheme [72] was used. which compare well to experimental results (ﬁgure 4. However.
5 MPa . with noslip condition applied to both upper and lower walls. The image shown is taken as if the viewer is situated in the nozzle looking at the ﬂow going downstream with the cavitation bubbles on the left and right sides. velocity ﬁeld in a plane perpendicular to the ﬂow direction shown to visualize the vortices.4: Image of computed cavitation probabilities in the channel section of the model injector.3: Mean isosurface for a vapour fraction of 0.4. Figure 4. injection pressure 8. SIMULATIONS OF A DIESEL FLOW IN A CAVITATING MODEL NOZZLE41 Figure 4.1.1.
RESULTS & DISCUSSION Figure 4.5 MPa Figure 4.9.6: Modelled and experimentally determined mass ﬂows with injection pressure of 4.5 MPa .42 CHAPTER 4. injection pressure 8. 7 and 8.5: Image of experimental determined cavitation probabilities.
it requires a simulation time of over 200 days if run on a single processor (assuming perfect parallelization). but may be adequate for modelling turbulence and cavitation inside the injector. Figure 4. It should also be noted that the mesh at 1. Isosurface of ρ = 500kg/m3 While the above simulation looks pretty. an example of the kind of results generated using the code is shown in ﬁgure 4.4.2.7 million cells is not suﬃciently ﬁne to resolve the incylinder spray structures. thus reducing the time to 17 days.2 is not suﬃciently stable to run for diesel conditions. coloured according to velocity magnitude. . However. The simulation was run on 12 processors.2 Injector Simulation Results Results obtained using the cavitation singlephase code are presented in Paper 2. it was not used in any of the published studies. Since the liquidvapourgas code described in section 2.7: Cutplane of injector geometry.7. INJECTOR SIMULATION RESULTS 43 4.
44 CHAPTER 4. There is a slight change for the liquid phase. since the available code for simulating a cavitating ﬂuid was not capable of handling high pressures.2. However. (a) 9 degrees in. if so. no such model could be developed. except that combustion was switched oﬀ for these simulations. This was by no means a full study. but the gas phase remains almost unchanged for a doubled injection cone angle. Results of varying the injection spray angle are shown in ﬁgure 4.8: Comparison of fuel vapour distribution in a centre cutplane for a low injection angle (9 degrees) and a large angle (20 degrees) .(b) 20 degrees injection cone an.8. and the conclusions are tentative at best.1 Impact of Injection Model As stated in the model description section. how much? These questions were brieﬂy addressed using simulations with several diﬀerent spray angles and droplet distributions.jection cone angle gle Figure 4. one objective for this project was to develop a new injection model for EulerLagrangian diesel sprays. The numerical grid and physical data were the same as those used in Paper 3. RESULTS & DISCUSSION 4. Remaining of the eﬀorts to develop an injection model is the presented code for cavitation and a preliminary study which was made to answer questions like: How much inﬂuence does the injector model have? Will the fueldistribution/ﬂame be inﬂuenced by variations in the spray angle or droplet distribution and.
= 50 µm (b) dav.3) used. However. they strongly inﬂuence the liquid penetration. (a) dav. INJECTOR SIMULATION RESULTS 45 Another initial parameter for the spray is the initial droplet size.4. initial parameters of the droplets also have quite a minor inﬂuence on the vapour distribution. but not the overall vapour penetration. They seem to aﬀect the penetration of the fuel vapour core. which will have substantial eﬀects on combustion. 100 µm = (c) dav. by varying the RosinRammler distribution (as described in section 3.2.9.9: Comparison of fuel vapour distributions in a centre cutplane obtained with three average initial diameter values . This parameter was also varied.2. As shown by the images in ﬁgure 4. 200 µm = Figure 4.
obtained using three diﬀerent grids. . lim on penetration of liquid (m) 0. physical conditions etc.05 0. time. see Paper 1. For details on grids. results of applying the turbulence lengthscale limiter (described in section 3. The liquid penetration obtained using the standard kε model is shown in ﬁgure 4.04 0.11. as will be shown in this section. the eﬀects of the modiﬁcation on vapour penetrations on 2 and 3 dimensional grids were mainly used to demonstrate the improvements it provided. It can also be seen that increasing the cell size results in more signiﬁcant deviations from the average. lim off 0 0. The large oscillations observed in the liquid penetration rate are due to the liquid penetration being set to the furthest point reached by 95% of the liquid mass in the system.10 0. we can see the local average at each point.3..3 Spray Results without Chemistry In Paper 1. lim off dr=1mm. while the calculations without the lengthscale limiter produce diﬀerent shapes on the three grids. However.46 CHAPTER 4. RESULTS & DISCUSSION 4.01 0 dr=0.02 0.3) are presented. If this was set to a lower value.06 drmin=0.03 0.10: Liquid penetration vs. the results would be less oscillatory and possibly no ﬁltering would be needed.1mm.1 Grid Dependency This section deals solely with the grid dependency of the simulations conducted in 2D to show how the limiter decreases the dependency. standard kε When using the lengthscale limiter the calculations seem to exhibit grid independency. If we ﬁlter out the oscillations. using a simple local ﬁlter. Due to the constraints on the length of the article. lim on drmin=0.1.5mm. The results can be seen in ﬁgure 4.5mm. 4. lim off dr=0. the eﬀect of the length scale limiter on liquid penetrations is also beneﬁcial in terms of promoting grid independence. lim on dr=1mm.5 x 10 −3 2 Figure 4. This gives an idea of the average penetration for every timestep sampled.5 time (s) 1 1.1mm. for example 90%.
5 time (s) 1 1. lim on penetration of liquid (m) 0. 0.05 0. lim on drmin=0.11: Liquid penetration vs. lim off min dr=0. lim off 0.03 0.1mm.04 0. see ﬁgure 4. SPRAY RESULTS WITHOUT CHEMISTRY 47 0. time. lim on dr=0.04 0.06 drmin=0.1mm.4. the standard kε model and ﬁltered penetrations When the LaunderSharma kε model is used. lim off dr=1mm.1mm.1mm. lim off dr=1mm. lim on dr=1mm. the penetration across the three diﬀerent grids is more uniform with the limiter then without it.5 x 10 −3 2 Figure 4.5mm. However. obtained using three diﬀerent grids.5 time (s) 1 1. lim off penetration of liquid (m) 0.02 0.01 0 0 dr=0.5 x 10 −3 2 Figure 4.03 0. lim on dr=1mm. obtained using three diﬀerent grids and the Launder Sharma kε with no ﬁlter . indicating that 3000 cells with a cell size of 1 mm may be too coarse. The penetration obtained with the coarsest grid diﬀeres considerably from those obtained with the two ﬁner ones.01 0 0 0.3.05 0. the behaviour is slightly diﬀerent. time.5mm. In Paper 1 similar indications were shown for the vapour penetration. lim on dr =0. lim off dr=0.06 dr min =0.12.5mm.5mm.12: Liquid penetration vs.02 0.
obtained using three diﬀerent Courant numbers and the kε model with ﬁltering used 0. without also investigating the eﬀects of diﬀerent Courant numbers.2.03 0.005 .04 0.2 Time Step Dependency The code uses a Courantbased timestep.48 CHAPTER 4. lim off Co=0. RESULTS & DISCUSSION 4. which means the timestep changes with the grid size and local velocity.14.2.1 . for which it removed the dependency completely (Figures 4.1. lim off 0 0.04 0. lim on Co=0.16.5 x 10 −3 2 Figure 4. lim on Co=0.5 time (s) 1 1. 0. lim off Co=0.06 penetration of liquid (m) 0.005 . taken from Paper 1).01 0 Co=0.1.5 x 10 −3 2 Figure 4.13: Liquid penetration vs. lim on Co=0.2 . The results of a Courant number investigation are shown in Figures 4. obtained using three diﬀerent Courant numbers.14: Liquid penetration vs. lim off Co=0. This implies that one cannot draw conclusions about grid independency. lim off Co=0. lim on Co=0. time. time.02 0.005.3. The limiter has a positive eﬀect on the liquid penetration.5 time (s) 1 1.1 . lim on Co=0. the Launder Sharma kε model and ﬁltering used .005.05 0.01 0 Co=0. lim on Co=0.15 & 4. lim off 0 0.02 0.05 0.06 penetration of liquid (m) 0.2 .13 & 4. but not as strong as its eﬀects on the vapour penetration.03 0.
5 1 time (s) limiter on limiter off 1.06 0.5 x 10 −3 limiter on limiter off 3 Figure 4.3.02 0 0 0.12 penetration of vapour (m) 0. time. SPRAY RESULTS WITHOUT CHEMISTRY 49 0.16: Vapour penetration vs.4. obtained using three diﬀerent Courant numbers and the standard kε model 0.02 0 0 0.15: Vapour penetration vs.08 0.5 1 1.08 0.04 0.06 0.10 0.5 2 2.12 penetration of vapour (m) 0.5 3 −3 x 10 Figure 4.5 time (s) 2 2.10 0.04 0. and the LaunderSharma kε model . obtained using three diﬀerent Courant numbers. time.
RESULTS & DISCUSSION 4.1 0.08 0.12 0.50 CHAPTER 4. 0. and only one case will be presented here for reference. 70 MPa.17: Vapour penetration vs. 4.4 MPa.04 0. vap exp Vapour penetration of vapor (m) 0.5 −3 Figure 4. an extensive range of injection and bomb pressures were used.02 0 0 limit on. bomb pressure.5 time (s) 1 x 10 1. vap limit off. time for a 3D case comparing experimental to simulated data.06 0. Injection pressure. The results of that study was presented in Paper 1.3. .3 Comparison to Experimental Data To compare simulations using the turbulence length scale limiter to experiments.
However. INFLUENCE OF NUMERICAL SCHEME 51 4. This is also. the best grid for a speciﬁc case is found on an empirical basis.18: Fuel vapour distributions of three sprays. Time. As shown in Paper 4. using a grid with more. Figure 4. which were initially developed to handle large gradients. The numerical schemes examined in Paper 4 were members of the total variation diminishing (TVD) family of schemes. in a sense. .19) to ensure that the fuel distribution is visible for all cases. it mixes the properties of the two solutions using the simpler schemes. they can also be applied with success to other cases. an illustration of how the TVD limiter works. see Appendix A. the diﬀerences in vapour distribution obtained using diﬀerent numerical schemes are quite large. The numerical scheme referred to here is the scheme for the advection term (for a more detailed description of how the scheme is used when solving the momentum equations.4.4. smaller cells does not necessarily increase the accuracy of the simulations. When there are no gradients.19). obtained using linear interpolation (left). originally in compressible high Mach number ﬂows. while retaining some irregularities like linear interpolation. and the choice of TVD limiter will not have an eﬀect on the solution. 1 ms after SOI. The LaunderSharma kε model was used for modelling turbulence. and to some degree. the TVD limiter will revert to a linear scheme. MC yields as long penetration as upwind. Grey scale altered (compared to the one in ﬁgure 4. Since MC interpolation uses both upwind and linear interpolation. such as diesel sprays where there are high gradients between the spray and the surrounding quiescent air. As can be seen.4 Inﬂuence of Numerical Scheme In numerical simulations of diesel sprays. Instead. Initial µT = 10−5 m2 /s. the same is true for the numerical scheme. The diﬀerence between the schemes was found to be larger when a lower turbulence viscosity was set. upwind interpolation (centre) and MC interpolation (right).2). indicating that a high viscosity smoothes the solution to such a degree that no instabilities or high gradients are present (see Figure 4. the initial conditions for the turbulence.
The LaunderSharma kε model was used for modelling turbulence.52 CHAPTER 4.6 · 10−4 m2 /s.19: Fuel vapour distributions of three sprays. . upwind interpolation (centre) and MC interpolation (right). 1 ms after SOI. Time. Initial µT = 3. obtained using linear interpolation (left). RESULTS & DISCUSSION Figure 4.
there are numerous other diﬀerences in the solver in two codes. In OpenFOAM (and the experiments) the ﬂame ignites and then propagates upstream towards the injector. In addition to the two codes diﬀering in computational time. As can be seen from ﬁgures 4. KIVA3V calculations use a set atomization length. The long ﬂame generated by OpenFOAM penetrates much too far.204. The time sequence in ﬁgure 4. according to the experimental data.20 through 4. OpenFOAM’s liftoﬀ length is also stable for these temperatures.5 Fuel Sprays with Chemistry In Paper 3. 7501300 K. Unfortunately. this was found to be due to the numerical scheme and turbulence model chosen. This indicates of that the initial evaporation and chemical kinetics are “correct”. 14. RNG kε is the kε turbulence model in OpenFOAM that produces the longest ﬂames. where the liftoﬀ length is stable above 1100 K. The ﬂame in OpenFOAM and KIVA3V diﬀers in two main respects. but this is because the ﬂame cannot be any closer to the wall (due to the constant temperature condition set along the walls).20.8 kg/m3 . and the quantitative results in ﬁgures 4. ambient gas temperatures.23 are extracted from Paper 3. as can be seen from ﬁgure 4. The main reason for the high computational time is solving the chemical reaction rate equations.4. simulations of events in the Sandia [73] constantvolumevessel were presented. If the ODE solver for the chemistry can be optimized to reduce the computational time without losing accuracy is an object for future work.1. shorter ﬂame. 821 %. In OpenFOAM they are solved fully coupled using the SIBS ODE solver as described in section 3.24 displays how the ignition starts and ﬂame propagates in OpenFOAM. about 4 times the computational time for KIVA3V. The second main diﬀerence is that OpenFOAM’s ﬂame is very long and thin compared to KIVA3V’s wider. and the limited linear scheme in OpenFOAM was found to promote long. thin ﬂames and fuel distributions. OpenFOAM can predict the ignition delay time and its dependence on ambient temperature and ambient oxygen concentration fairly well. and the injection pressure diﬀerence was set to 1500 bar.5. the ﬂame liftoﬀ length predictions are not as good. OpenFOAM uses a . In the simulations Paper 4 is based upon. whereas in KIVA3V the ﬂame ignites closer to the injector and then the ﬂame liftoﬀ length propagates downstream. The simulations were performed on a cubic mesh with 108 mm sides for the following conditions: ambient gas oxygen concentrations. KIVA3V and experiments. and ambient gas density. FUEL SPRAYS WITH CHEMISTRY 53 4. The case with 21% oxygen took up to 70 hours on 4 processors using OpenFOAM. OpenFOAM seems to overpredict the changes in liftoﬀ length caused by variations oxygen concentrations and ambient temperature. The focus was on diﬀerences between KIVA3V and OpenFOAM predictions of ignition delay time and ﬂame liftoﬀ length. These results. The injector oriﬁce diameter was 100 µm. and how well the codes could calculate these parameters compared to experimental data.4.23. The computational time was much higher in OpenFOAM compared to KIVA3V.
These diﬀerences do not make the comparisons presented here and in Paper 3 irrelevant.21: Comparison of OpenFOAM and KIVA3Vpredicted and measured liftoﬀ lengths for diﬀerent ambient gas oxygen concentrations at an ambient gas temperature and density of 1000 K and 14. .8 kg/m3 . RESULTS & DISCUSSION collocated grid arrangement. is added stepbystep to see the eﬀects of each model. and turbulence model are also diﬀerent. chemistry etc. 60 Lift−off Length (mm) Measured KIVA−3V OpenFOAM 40 20 0 5 10 15 20 25 Ambient Oxygen Concentration (%) Figure 4. whereas KIVA3V uses a segregated grid arrangement. since it was made to compare the two codes in terms of performance and ability to match experimental data. spray submodels. chemical solver.20: Comparison of OpenFOAM and KIVA3Vpredicted and measured liftoﬀ lengths for diﬀerent ambient gas temperatures. respectively.8 kg/m3 . OpenFOAM uses spray submodels which are the same by name. but the way they are implemented in KIVA3V is diﬀerent. spray. respectively.54 CHAPTER 4. at ambient gas oxygen concentration and density of 21% and 14. The spray momentum coupling. where complexity such as compressibility. A full comparison of the two codes should be made starting with simulation of a simple model ﬂow. 80 Lift−off Length (mm) 60 40 20 0 Measured Kiva−3V OpenFOAM 800 1000 1200 Ambient Temperature (K) 1400 Figure 4.
22: Comparison of OpenFOAMpredicted and measured ignition delay times for diﬀerent ambient gas oxygen concentrations at an ambient gas temperature and density of 1000 K and 14.0 1. .5.5 0 Measured OpenFOAM 5 10 15 20 25 Ambient Oxygen Concentration (%) Figure 4.5 1.4.0 0. FUEL SPRAYS WITH CHEMISTRY 55 Ignition Delay Time (ms) 2.23: Comparison of OpenFOAMpredicted and measured ignition delay times for diﬀerent ambient gas temperatures at an ambient gas oxygen concentration and density of 21% and 14.8 kg/m3 . Ignition Delay Time (ms) 3 Measured OpenFOAM 2 1 0 600 800 1000 1200 Ambient Temperature (K) 1400 Figure 4.8 kg/m3 . respectively. respectively.
5 ms PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature t = 1. Corresponding.8 ms t = 1. ambient temperature. The length of each image from simulations corresponds to 108 mm. .24: Time sequences of simulated temperature distributions obtained using the OpenFOAM code (top) and KIVA3V code (middle) showing ignition and ﬂame development of an nheptane spray in the Sandia highpressure. 21 %. RESULTS & DISCUSSION Temperature (K) PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature 750 1210 1675 2140 2600 PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature t = 0.56 CHAPTER 4.5 ms Figure 4.1 ms t = 0. the experimental images are of the same scale. but extending 60 mm. and ambient density 14. hightemperature vessel: ambient oxygen concentration.8 kg/m3 .3 ms PSfrag replacements Temperature PSfrag replacements Temperature PSfrag replacements Temperature t = 1. 1000 K. experimentally acquired OH chemiluminescence images are shown in the bottom part of each subﬁgure.
Unfortunately. These numerical conditions were used to examine whether good comparisons with data in the Sandia database of experiments could be maintained when combustion was considered in the simulation. the chemical mechanism used has been validated against shock tube data. 14. It was also found that the initial turbulent viscosity should be slightly higher. and the RNG kε model for turbulence.5.9 ms t = 1. however. Ambient gas temperature.1 ms Figure 4. 0%. respectively. Temperature (K) 750 1210 1675 2140 2600 t = 0. As mentioned in the previous section. Comparing Figure 4. Aside from the wellpredicted vapour distribution. The “best”1 choices for advection scheme and turbulence (of those tested in the studies presented in Paper 4) were found to be Superbee and LaunderSharma kε. showing images obtained from simulations excluding combustion and including combustion. 1000 K. it can be clearly seen that while OpenFOAM predicts the vapour distribution in the displayed part of the vessel accurately. provide reasonably close ﬂame liftoﬀ lengths (which was the reason for the choices).7 ms t = 0.26.25: Comparison of fuel distributions obtained from experiments and simulations tuned to match the experimentally determined fuel vapour distribution for the noncombusting case.25 and 4.5 ms t = 0.8 kg/m3 . These two validations reduce the number of possible reasons for the discrepancy in the ﬂame liftoﬀ length. oxygen content. the limited linear scheme was used for the advection term. . FUEL SPRAYS WITH CHEMISTRY 57 4. respectively.4. They did.1 Combustion Predictions Using Other Schemes In Paper 3. it does not predict the combustion and (more speciﬁcally) the ﬂame liftoﬀ length so well. however. the 1 In terms of matching experimental fuel vapour distribution of a noncombusting spray. density.5. these two choices were quite unfavourable for obtaining similar (short) ﬂames to those observed in experiments.
the penetration of the droplets before they evaporate.8 ms t = 1. and this according to Hilbert et al [74] (page 94. including the liquid droplet distribution parameters. If our chemistryturbulence model predicts perfect mixing in a position where it does not occur. Ambient gas temperature. assumes a Lewis number of unity for species diﬀusion. i.1 ms t = 1.e. RESULTS & DISCUSSION Temperature (K) 750 1210 1675 2140 2600 t = 0. 14. There are several possible remaining reasons for the diﬀerence between the experimental and simulated data.58 CHAPTER 4. Another set of unknowns that have already been mentioned in part are the liquid atomization parameters.8 kg/m3 . the data for the ignition site and early atomization zone cannot be validated.26: Comparison of OH chemiluminescence images obtained from experiments and temperature distributions from simulations tuned to match experimental fuel vapour distribution for a noncombusting case. OpenFOAM. experimental data for the noncombusting case only covers the area 17 mm downstream of the nozzle and beyond. density.5 ms t = 0. which is consistent with observations.3 ms Figure 4. ﬁgure 17 in [75]). 21%. one is that the eﬀects of air entrainment and airfuel mixing might not be correct. the PaSR model might need improvement or the parameters used in it might be incorrect (most likely the turbulence mixing time). 1000 K. the ﬂame liftoﬀ length could be underpredicted. and the initial shape/particle distribution of the fuel cloud. Hence. A good . can lead to substantially larger errors than use of more detailed approximations. oxygen content. like most other codes. This should also result in underprediction of ignition delay times.
In [54] the mixing timescale was calculated as: 1 ReT 1/4 τmix = k k = 0. An image of the values of κ in a combusting spray is shown in ﬁgure 4. as long as it remains lower than 1 (in fact. The model was described in detail in section 3.4. the values of k and ε also aﬀect the result. For the nheptane mechanism used in this work. in which the kolmogorov timescale is used.03. In [52] the value of Cmix was reported to be 0.0.4.005. it was observed that the PaSR model only had minor eﬀects on the solution. Obviously.1.1) Other versions are from [59]. while k is of the order of 100 m2 /s3 .97 and 1.178 ε ε (4. 4. The liquid penetration for the noncombusting case is reasonably well predicted by the simulations.6. but for spray simulations ε has an average of the order of 106 m2 /s3 . data for only one noncombusting case were available at the time of these simulations. Unfortunately. the chemical timescale is thus always several times larger than the turbulent timescale. which depend in turn on local chemical and turbulent timescales. The chemical timescale is determined by the timescales of each reaction and does not contain any “tuning” constants. The scale should be noted. The parameter that will be highlighted in the ﬁgure below is κ. but some of its properties will be repeated here for the reader’s convenience. since it ranges between 0. and [52] in which a modelling constant Cmix is used: k ε τmix = Cmix (4. and the way it should be calculated is an open question. so it was not possible to verify that the vapour distribution or liquid length is correct at other temperatures and densities. to visualize the extent of the spray.6 Comment on the PaSR Model During the course of the work reported in Paper 3. The timescale of the turbulence is diﬀerent. It was found that the simulations are almost independent of the value of Cmix . Finally. The elevation is by temperature. COMMENT ON THE PASR MODEL 59 parameter to use for comparison would be the fuel liftoﬀ length for a noncombusting case. Naturally. the partially stirred reactor becomes a perfectly stirred reactor in . major diﬀerences have not been observed until the value reaches 1000).27. while the one I used for combustion calculations was 0.2) This is the expression used in this work. it is possible that the temperature dependence may be incorrect in some of the models. but there is no way to verify how well it is predicted for the combusting case with currently available experimental data. which determines the combusting and quiescent proportions of a cell.
. Figure 4. However.26). The area outside of the elevated temperature area has a κ value of zero. RESULTS & DISCUSSION this case. κ is the mass fraction of the cell that is involved in the chemical reactions.60 CHAPTER 4. is not known at this time. coloured according to κ and elevated according to temperature.27: Display of a spray simulation (corresponding to the one shown in ﬁgure 4. whether or not this is due to unrealistic chemical timescales of the reactions involved.
The thesis shows that there are many pitfalls on the way to simulations that provide good comparisons to experimental events. in addition to the liquid and vapour penetration lengths. The importance of this result cannot be overemphasised.1. or results can be obtained in a shorter time. see section 4. it improves the eﬃciency of the CFD modelling. OpenFOAM was found to perform well when predicting the propagation of the ﬂame compared to the ignition site. as stated in the outline of the thesis (section 1. and in ignition delay time.3 clearly show that dependence diminishes when a limiter on the length scale of the turbulence is applied. In short. and the eﬀects of varying the ambient oxygen concentrations and temperatures on these parameters.4). Unfortunately. For a noncombusting case in the same geometry. However. no such model could be developed. due to the high complexity of the problem. as well as presenting some ﬁnal results from the study. since shorter computational time means more investigations can be made. intended to be used for the development of an injection/atomization model for EulerLagrangian diesel spray simulations. These simulations were.Chapter 5 Conclusions A code describing how a ﬂuid cavitates when it is subjected to low pressures due to geometric contraction was successfully used to simulate data acquired from model nozzle experiments performed at AVL. it was found to underpredict the ﬂame liftoﬀ length and its dependency on temperature and ambient oxygen concentration. The results also show how ﬂame liftoﬀ length and ignition delay time can be predicted by OpenFOAM. 61 . together with an example of results obtained using the threephase code (mainly for illustrative purposes since this code has not been validated). Regarding spray grid and timestep independence. OpenFOAM could predict the vapour distribution very well. Examples of the degree of inﬂuence the injection model can have over the EulerLagrangian spray simulations are also presented. For a more detailed description of the encountered problems when simulating cavitating diesel injections. the results from Paper 1 and section 4.
in addition to the eﬀects of numerical grid and turbulence model. the calculations and results of Siebers et al [76] were used. As indicated in section 4.4 and 4. In the work this thesis was based upon. which is consistent with the article’s relationship between injection velocity and ξ. and also in sections 4. the SuperBee TVD limiter was found to give the best results compared to experiments. .1. This was demonstrated in paper 4.6. Furthermore. according to the report. but this ﬁnding could be (like the choice of grid) problemdependent. The choice of scheme can have a large inﬂuence on the fuel vapour distribution.26. the PaSR model does not suﬃce for this task. CONCLUSIONS To further analyze the reason for the short ﬂame liftoﬀ lengths yielded by OpenFOAM. These eﬀects are important to be aware of. the liquid droplets and ﬂame should be separated in distance much more from the injector than they are in ﬁgure 4. The cited report describes a way to estimate the percentage of stoichiometric air entrained (ξ) upstream of the liftoﬀ length.5. the percentage of entrained air is 29% of the air required for stoichiometric combustion. According to this estimate.62 CHAPTER 5. the ﬂame liftoﬀ length might be increased without increasing the length of the ﬂame. and thus on the resulting ﬂame and combustion. If the turbulence/chemistry interaction model used required more air to be entrained before combustion.
the chemical mechanism has been veriﬁed against shocktube experiments. implies use of LES to model gas phase turbulence and better submodels for the liquidgas interactions. the parts of the code that are largely responsible for the errors must be the turbulence/chemistry interaction (PaSR model). Details of this setup are not known at this time. Following the second approach. 63 . while the models presented in Paper 4 delivered fuel vapour distributions that approximated well to experimentally observed distributions. temperature dependency of the liquid penetration. but might be published in the future. much work needs to be done to improve them. possible areas of future work are to continue eﬀorts started in the studies published in Papers 3 and 4. but they are not yet predictive. In the nearer future. primary breakup and initial vapour cloud formation. The validation presented in Paper 3 is only against one case. two possible approaches could be adopted in attempts to improve diesel spray simulations: EulerEuler type simulations. A way to rule out temperature dependency would be to run simulations and compare them to the results of corresponding experiments. Furthermore. Therefore. but through bubbles of varying size.Chapter 6 Future Work As shown in this thesis. in which the Lagrangian parcels/blobs do not interact with the grid directly. Another option is the VSB2 spray model developed by Karlsson [77]. Brieﬂy. which have only been used for primary breakup studies to date. and seems to produce good comparisons for those cases. Since we can now assume that the spray and vapour distributions are well tuned. the Lagrangian parcels contain a distribution of blobs instead of equally sized droplets. due to a lack of data from noncombusting experiments with welldeﬁned boundary conditions to compare to the simulations. As shown in the Results section. However. diesel spray simulations using the EulerLagrangian approach are helpful. or the chemical mechanism itself. the resulting ﬂames were not well predicted when chemistry was considered. without oxygen using diﬀerent ambient temperatures. or more sophisticated EulerLagrangian descriptions.
so using the work described in Papers 3 and 4 as a basis for engine simulations to explore the eﬀects of key variables in engines could be of high value to both the engine community and the numerical community. a diesel injection model could be developed based on experimental studies instead of numerical simulations.5). It is possible that the turbulence plays a major role in obtaining correct air entrainment for modelling the early spray development and ﬂame liftoﬀ lengths. The most important of these parameters. discharge coeﬃcient etc. It would also be interesting to venture further down the path taken in Paper 2. to spray angle.1 Outlook As mentioned above. in which both liquid and vapour penetration lengths are presented along with ﬂame liftoﬀ lengths. in this respect. Aside from using kε models and LES it could also be useful to assess the possible utility of other turbulence models such as kω type models. Other aspects not covered in the thesis could be of interest for future work. A good example of a study that presented relevant data. One such aspect is the obvious extension from constant volume vessel simulations to engine simulations. a large number of simulations where the chemical reactions are considered will have to be run. Simulating cavitating diesel ﬂow requires a more robust model than the one used in Paper 2. it would be helpful to have access to an experimental database covering a larger array of noncombusting cases than the single case from Sandia. based on the short preliminary study. since the liquid penetration length has a strong inﬂuence on the ignition point and its propagation. These two approaches are linked in the sense that they both are concerned with the primary events of spray development. However. seems to be the initial droplet size distribution. If these studies were used like the cavitation simulations were intended to be used. the fuels used are not as easy to model as. and perhaps it is not even essential to the development of a new injection model. unsteady RANS models or other hybrid type turbulence models. it should be . and the experiments were only run in combusting conditions. It would also be beneﬁcial for research in numerical methods to have more data on the initial breakup of sprays for both combusting and noncombusting cases. It would then be desirable to have optimized the solver for the chemical mechanism. from a modeller’s perspective. is one by Ochoterena et al [81].64 CHAPTER 6. for instance. 6. nheptane. cone angles and ﬂame temperatures. It is likely that a good atomization and injection model would improve the ﬂame liftoﬀ length predictions presented in Paper 3. 79]) which links injection pressure. investigating in detail how the initial droplet parameters inﬂuence the spray and combustion. There are studies ([78. nozzle diameter. OpenFOAM has the functionality to handle moving meshes with valves and other topological changes [80]. FUTURE WORK When the results presented in Paper 3 are to be improved (beyond the improvement presented in this thesis). as the computational time occupied by the solver in OpenFOAM is currently very high (as mentioned in section 4. Unfortunately.
OUTLOOK 65 noted that supplying modellers with input for simulations was not one of the stated objectives of the cited study ([81]). Hopefully. commuters were facing rising prices for hay and oats. In light of the increasing need for more experimental data speciﬁcally tailored for validating numerical simulations.6. . Close.1. to help ensure that demands from customers and legislation are met. Furthermore. horses in early 20th century America. longterm cooperation between a modeller and an empirical investigator with access to both spray chamber and engine equipment would have every chance of making major advances in spray modelling. we return to the starting point of the thesis. Finally. This situation is uncomfortably similar to current circumstances. the author belives and strongly recommends that more experiments tailored for numerical simulations should be performed. the solution we devise this time will not lead to the same problems 100 years from now. the need for good spray modelling will probably increase as engine concepts become more advanced and new fuels are introduced. as well as increased public fears about pollution and traﬃc jams [1]. During this period.
66 CHAPTER 6. FUTURE WORK .
Numerical Investigation of Mesh/Turbulence/Spray Interaction for Diesel Applications This paper originated from attempts to solve the problem of grid dependence described in section 1.183. that the turbulence length scale should be limited to the jet diameter within the jet region. This was implemented in the OpenFOAM code. who also implemented the limiter in the code. The multicomponent model diesel fuel IDEA was used as the fuel. The simulations using the turbulence length scale limiter proved to provide better matches to the experimental data than when it was turned oﬀ.4. I was responsible for the preliminary studies that were not included in the paper. as well as using diﬀerent timesteps. The solution was based on a suggestion in the book “Modeling Engine Spray and Combustion Processes” by Gunnar Stiesch [12].Chapter 7 Summary of Papers The thesis is based on the work published in the following papers Paper I . to determine limiter size. grid optimization etc. hightemperature constant volume vessel. The paper was written by 67 . and tested on diﬀerent sized grids in 2D and 3D. The most important ﬁnding was that the grid and time step dependency could be reduced by introducing the length scale limiter. The rationale for this is that the jet is the driving force for the turbulence in spray simulations. The results were also compared to data acquired in experiments performed at Chalmers on fuel injection into a highpressure. the remaining one was performed by Dr Niklas Nordin. It was shown in the article that the length scale limiter was very beneﬁcial in terms of reducing grid and timestep dependency. I was also responsible for running and postprocessing all but one of the simulations included in the paper. p. The simulations and paper were planned together with Dr Niklas Nordin. The experimental data used were acquired by AlfHugo Magnusson at Chalmers University of Technology.
as these simulations required extensive testing. as a byproduct of a code to simulate diesel injector ﬂow with the ambient air as well as cavitating diesel liquid. USA. However. 2005. I wrote the paper with inputs from Dr Niklas Nordin and Henry Weller. We compared the eﬀects of varying amounts of oxygen/EGR and ambient temperatures using the two codes.. before the article was written. With the knowledge gained from the studies published in Paper IV. The OpenFOAM simulations did not use an atomization model. who developed the code used. In addition. I believe the results would have been better if SuperBee interpolation had been used for the advection scheme. It is possible that KIVA3V yields longer ﬂame liftoﬀ lengths since the atomization length is prescribed in the code. It was found that OpenFOAM produces much longer ﬂames than KIVA3V. while the injector model would be the topic of a possible later paper. San Diego. in Rio de Janeiro. However. There was also a limited comparison of the eﬀects of using diﬀerent grids and injection positions (published in the Appendix of the article). so a decision was taken to focus solely on the validation in this paper. SUMMARY OF PAPERS me. with various numerical setups to ensure a stable solution. . and thus yielded much shorter liquid penetrations. as well as short ﬂame liftoﬀ lengths. The empirical data used for comparison were acquired from experiments performed at AVL. I ran months of simulations. I presented it at the 5th Joint ASME/JSME Fluids Engineering Conference. The initial intention was to develop and present a new injection model. ran and postprocessed the simulations. simulating injector ﬂow proved to be much more complex than anticipated. Brazil. I setup. Paper III .Modelling Injector Flow Including Cavitation Eﬀects for Diesel Applications This paper was preceded by several attempts to model diesel injector ﬂow.68 CHAPTER 7. Paper II . 2007. Unfortunate use of the RNG kεturbulence model together with the limited Linear scheme is believed to be an underlying reason for the long penetration of the ﬂame. in the same article. it was also possible to compare the results to experimental data. and I presented it at the SAE Fuel and Lubricants Conference. Thanks to the large ECN database. together with validation of the code for the injector ﬂow. together with Dr Niklas Nordin. with numerical help from Henry Weller.Three Dimensional Simulation of Diesel Spray Ignition and Flame LiftOﬀ Using OpenFOAM and KIVA3V CFD Codes The paper originated from a comparison of OpenFOAM results I generated to KIVA3V results generated by Dr Feng Tao. The code used in this paper was written by OpenCFD Ltd.
Simulations indicated that the scheme might have a strong inﬂuence on the fuel vapour distribution. Italy.69 I setup. ran and postprocessed by Dr Feng Tao. It was presented at ILASS. . ran and postprocssed the simulations. ran and postprocessed the OpenFOAM simulations. Therefore. while the KIVA3V simulations were setup. The paper was written by me and Dr Feng Tao. I setup. and wrote the paper together with Dr Feng Tao. If the initial viscosity of the turbulence is low. a more extensive study of the inﬂuence of the scheme (and eﬀects of the initial parameters on spray development) was undertaken. I presented it at the SAE World Congress. 2008. Como Lake. the initial turbulence parameters can reduce or enhance its inﬂuence.Inﬂuence of Advection Schemes In The Prediction of Diesel Spray and Fuel Vapour Distributions As mentioned regarding the previous paper. 2008. USA. Detroit. It was found that while the scheme does have a large inﬂuence. Paper IV . the authors were not satisﬁed by OpenFOAM’s overprediction of ﬂame lengths. the SuperBee scheme was shown to be the best choice for interpolation of the advection scheme.
70 CHAPTER 7. SUMMARY OF PAPERS .
and Ux. Further. Partial Diﬀerential Equations in OpenFOAM are solved by setting up the appropriate matrix system: fvm::operation(coefficient.Appendix A Implementation in OpenFOAM A. RhieChow correction is absolutely necessary for ﬂow simulations using a collocated grid.y.z are actually three equations.y. and at the same time share some additional knowledge that has been gained during the course of this work. which is the velocity we solve for.U) (A. and coefficient is a coeﬃcient which can vary in space and/or time. this section contains two velocities: U. which occur if the pressure gradient does not depend on the pressure in adjacent cells. which is used for the ﬂux and held constant when the momentum equations are solved. One very important aspect of CFD is RhieChow correction. thus allowing a “jigsaw” pattern to develop.1 RhieChow Interpolation in OpenFOAM OpenFOAM is relatively new. referring to relationships in the x. which will therefore be described here. All equations containing Ux. Therefore the intention of the author here is to extend the documentation. open source code that is not fully documented. Some basic information about the notation in OpenFOAM is needed to describe how it works. since it removes oscillations in the solution. Operations like this will be described many times throughout this section. or be a constant.1) This means that the solver is to solve for U. Rhie and Chow is usually regarded as a correction proportional to the diﬀerence between the pressure gradient at the face and the interpolated 71 . and zdirection respectively. y. and it is important to understand the nomenclature.z .
i Ux.z in each of the three directions. U) − fvm::laplacian(nu. and ∆ is related to the mesh size.z) dV f.y.H in icoFoam.z + (U · )Ux. for a cell face. the icoFoam code from OpenFOAM will be used to illustrate how this correction is implemented. U) ). fvVectorMatrix UEqn ( fvm::ddt(U) + fvm::div(phi.z(Uf.4) Equation A.y.z )f · n dS = ˆ f. (A.3 is integrated over a volume V with face S. For this description. OpenFOAM does not have such an explicit term in the equations. and there is a ﬁeld named phi (φ) instead of the Uterm. If we consider the UEqn. For an incompressible case. It is used because OpenFOAM utilizes the Gauss theorem.4 has no right hand side.i · Sf i ) = S (UUx. which makes it diﬃcult to see how the RhieChow correction is applied.z = i = i (A.5) . equation A.y. the velocity is corrected by: 1 u Ap j ∂p − ∂xj ∂p ∂xj uj = u j − ∆ (A. IMPLEMENTATION IN OPENFOAM pressure gradient at the face. Indeed.y.y. For example (from Ferziger & Peric [82]).y. and the second term can be rewritten as: V · (UUx.z − ∂t · (ν Ux.3) Where the above equation is solved for Ux. φ is (for incompressible ﬂows) the volume velocity ﬂux deﬁned at the faces of each cell.z) = − 1 p ρ (A. and deﬁnes the transform of a volume integral into a surface integral.2) where the overbar indicates interpolation.72 APPENDIX A. icoFoam is a solver for transient incompressible laminar ﬂow of Newtonian ﬂuids. the momentum equation in vector form would be: ∂Ux.y. For this type of ﬂow. which is frequently used in applied mathematics. OpenFOAM applies a correction that some would call ’in the spirit of Rhie and Chow’.i φi Ux. the convection term is slightly diﬀerent.y.
U) (A. We can now connect OpenFOAM’s equation to its analytical counterpart: · (UUx.9) Equation A.10) . we need a predictor for U.2.A() and UEqn. The magnitude of the cell face normal is the cell face area. φ is deﬁned as the scalar product of the cell face velocity and the cell face normal (eq.4. respectively.i · Sf 73 (A. which is turned into U in OpenFOAM.7) V In the above equation.y.z .6) Uf.y. The way in which the face and cell values are obtained from each other is discussed in section A.y.fvc::grad(p)). A. and H the one containing terms without Ux.z.9 cannot be solved at this stage. which is described below.z is the velocity we solve for. since we have not updated the pressure yet.6). [.y. while U is deﬁned in the cell centre. which results in: [U] = H 1 − [p] A A (A. They are created by issuing the commands UEqn. which does not include the pressure gradient (thus contributing to the separation of pressure and velocity). There is also a f diﬀerence in that phi and Ux. If the problem of interest includes transport of a scalar property. the discretized momentum equation would be this matrix system: A[Ux. such as enthalpy or the mass fraction of a chemical compound.y.] denotes the numerical approximation of the corresponding variable. This system is solved by dividing by A. while U is the vector velocity that will be solved for.y.1.6). It is important to note the diﬀerence in the subscript when the surface integral is introduced. (A.A.i is the velocity that will be held constant when the equation for pressure is solved. The code adds the inﬂuence of p to U in another way.H().8) A is the numerical equivalent of the operators in equation A. which is calculated using the pressure from the previous iteration: solve(UEqn == .3 operating on Ux. Recalling from [82].z) dV =⇒ fvm::div(phi.z are evaluated on the surfaces of the cell. U is turned into phi (by equation A. We are now ready to return to equation A.z ] = H − [p] (A. RHIECHOW INTERPOLATION IN OPENFOAM where φ = Uf. subscript f indicates that the term should be evaluated on the face. Ux.
When the PISO loop is ﬁnished. This derivative is needed on the cell faces.0/UEqn. since we now need to ﬁnd the pressure.16) A feature that cannot be seen in this formulation is that OpenFOAM again uses the Gauss theorem.13) The left hand side will be treated explicitly. (A. Thus. The code used as an example here is icoFoam. To create the equation for pressure. and lacks inﬂuence of pressure. before the boundary conditions are corrected.Sf() volScalarField rUA = 1. IMPLEMENTATION IN OPENFOAM This is called a momentum predictor of U. this is done by utilizing fvc instead of fvm.A(). In OpenFOAM. since it represents the correct velocity. the velocity is corrected with the correct pressure gradient. we take the divergence of equation A. since the face velocities will be used to evaluate the term. So far only the momentum equation has been used. but only a ﬁrst derivative. and it is evaluated by using the cell centre values of the pressure.A(). it is not necessary to calculate a second derivative of p. The resulting equation is then: surfaceScalarField phi = fvc::interpolate(U) & mesh. which is used to create an equation for pressure. fvScalarMatrix pEqn ( fvm::laplacian(rUA.11) The velocity from A. . The velocity U∗ is replaced by the velocity ﬂux φ.12) The left hand side is zero for incompressible ﬂows. and given a special notation: U* = UEqn.74 APPENDIX A. Equation A. The ﬁrst term on the right hand side of equation A.11 does not satisfy continuity.H()/UEqn.15) (A. but the method is the same in most of the other codes available in OpenFOAM.12 is then: · ([U∗ ]) = 1 [p] A · (A. and thus keep the velocity constant.9: · ([U∗ ]) − 1 [p] A · [U] = · (A.10 is needed for the pressure equation. (A.1 1 Actually this is computed at the end of each PISO loop. p) == fvc::div(phi) ).14) (A. but we also have the continuity equation.
not neighbouring faces The gradient term of p is calculated from the cell face values of p. 3.A. This method gives an oscillationfree velocity “with the spirit of RhieChow”. 4. and therefore no gradient calculation is necessary. 2.17) The gradient in equation A. there are four important points regarding the RhieChow correction in OpenFOAM. (A. even though there is no explicit RhieChow correction.1. Instead. 1. RHIECHOW INTERPOLATION IN OPENFOAM 75 U = rUA*fvc::grad(p). The gradient is calculated using neighbouring cells. To summarize.17 is again evaluated using the Gauss theorem. only the pressure on the cell face is needed. . phi does not include any eﬀect of pressure when solving the continuity equation rUA does not include any eﬀect of pressure when solving for continuity and the ﬁnal velocity corrector The Laplacianterm of p uses the value of the gradient of p on the cell face.
3. although this is only visible in the dimension on initial p.18) (A. yielding: V · (UUy ) dV = = i i S (Uy )f.23) This term is integrated over the cell volumes used for the ﬁnite volume method.linear · Si (Uy )f.scheme = − f ∂p dV ∂y (A.24) .22) Rho is not included in the above equation because OpenFOAM uses p = p for ρ many incompressible codes instead of actual p. For simplicity.21) where the equations are rewritten using the continuity equation. This saves space. This will be explained in more detail here. we will only consider ﬂow in the ydirection.scheme (U)f.19) (A. IMPLEMENTATION IN OPENFOAM A.76 APPENDIX A.2 Formulating the Momentum Equation Matrix in OpenFOAM In Paper 4 there is a concise description of the way in which diﬀerent schemes for diﬀerent terms can be used in OpenFOAM. U) ).linear · n dS ˆ i V (U)f. as we now only have one equation: 1 ∂p ρ ∂y · (UUy ) = − (A.20) (A. In OpenFOAM they would be written (according to the RhieChow method explained in the previous section) as: fvVectorMatrix UEqn ( fvm::div(phi. (A. from equation to matrix. We start oﬀ by simplifying equation A. by assuming stationary incompressible nonviscous ﬂow: U · Ux = U · Uy = U · Uz = · (UUx ) = − 1 ∂p ρ ∂x 1 ∂p · (UUy ) = − ρ ∂y 1 ∂p · (UUz ) = − ρ ∂z (A.
Two interpolations are then required to i calculate the term on the lefthand side: one for (Uy )f. Instead. we can write equation A.24 for cell P: (U)linear · Se (Uy )scheme + (U)linear · Sf (Uy )scheme = Se pe − Sf pf φe φf e e f f (A. the second velocity is usually interpolated by linear interpolation. and thus the TVD scheme for the divergence term would not perform as well.1: Line mesh consisting of six cells. the faceﬂux on certain faces could be zero due to use of upwind/downwind interpolation.26) where λf and λe vary between 0 and 1 depending on the chosen scheme. Following the nomenclature in ﬁgure A. and one for (U)f. and the only interpolation needed would be the linear interpolations for the ﬂux term and pressure. we get the equations used to compile the linear matrix system for solving the equations. FORMULATING THE MOMENTUM EQUATION MATRIX IN OPENFOAM77 Where all the variables are evaluated on face i. If we insert this expression for Uf and Ue into equation A. i Figure A. which is the velocity to be solved for. As indicated by the subscript. OpenFOAM solves for the cellcentred values directly. The faceﬂux. Using anything other than linear for this term could cause problems later when the divergence term is evaluated.scheme .25) If OpenFOAM used a staggered grid setup. the velocity that is kept constant as the linear equation system is solved.2. The general equation for ﬁrst/second order interpolation of the cellcentred variable U to the face is given by Uf = λf (UP − UN ) + UN Ue = λe (UM − UP ) + UP (A.A.linear . the above equation could be solved for the face values per se. For cell P this equation is: .1. requiring more interpolation. belonging to the cell in question with volume V and surface S.25. or volumeﬂux in incompressible code. (phi in OpenFOAM) is either massﬂux in compressible code. For instance.
IMPLEMENTATION IN OPENFOAM φe λe UM + (φe (1 − λe ) + φf λf )UP + φf (1 − λf )UN = Se pe − Sf pf φe (λe (UM − UP ) + UP ) + φf (λf (UP − UN ) + UN ) = (A. ..28) which results in a sparse matrix having elements only on the sub.. but through the procedure described there. .. . The resulting matrix formed from this equation is then: .27) This equation is valid for all cells.. . . As mentioned in the previous section. φf (1 − λf ) ....diagonal.and main..78 APPENDIX A.... .. φ e λe ... . . (A... This is the matrix that is calculated when the command fvVectorMatrix UEqn is issued in OpenFOAM. . . . Note that this equation is solved for the cellcentered values. although special treatment is needed for the boundary cells. ... ..super. UM UP UN ........ this equation is not solved directly. = Se pe − Sf pf .. φe (1 − λe ) + φf λf . . .. It is usually solved by ordinary LU decomposition.
10) f = −0.Appendix B Variable Deﬁnitions and Dimensionless Numbers S= U : dev(sym( U)) –« „ »“ 2 ”2 .1) (B.9) (B.8) (B.3e RRNG = +1 +1 k η = (S) ε ρd urel r Red = µl ρurel 2 r We = σ √ W el Oh = Rel Ta = Oh Weg U·d Co = d2 ∆t ν Sc = Dv 79 .5) (B.2) (B.7) (B.4) (B.3) (B.50 − min ρk µε η − η0 η 3 βη (B.6) (B.
0 0. VARIABLE DEFINITIONS AND DIMENSIONLESS NUMBERS Model parameter B0 B1 Cτ CRT ms.61 40.1 0.0 1.80APPENDIX B.1: KHRT model constants .limit W elimit Numerical Value 0.4 6 Table B.
105–130. H. Energy in Sweden 2007 a Facts and Figures. M. [12] Stiesch. pp. SAE Paper 980131. LightDuty Automotive Technology and Fuel Economy Trends: 1975 through 2007. N.. Scholarly Resources Inc. ozone and nitrogen dioxide”.. G. 21.. Rolf D. www. pp. Jasak.. Energil¨get i siﬀror 2007.. 2007. Tabor. 2003.com/standards/eu/ld. In The Making of Urban America.. 1997.. A.bilsweden. M. G. “A Tensorial Approach to Computational Continuum Mechanics using Object Orientated Techniques”. [10] Weller. Computers in Physics. P. Springer Verlag Berlin Heidelberg New York. R. “Structure and Breakup Properties of Sprays”. SAE Paper 940523.dieselnet. “Modeling the Eﬀects of Fuel Injection Characteristics on Diesel Engine Soot and NOx Emissions”.. 12(6).se. “InCylinder CFD Simulation Using a C++ ObjectOriented Toolkit”. 2004. H.C. and Reitz. Dieselmotorn p˚ kort och l˚ sikt.A. Hampson. [7] Patterson. [5] United States Environmental Protection Agency. [9] O’Rourke. Weller.... 2003.A. 1994. “Modeling the eﬀects of fuel spray characteristics on diesel engine combustion and emission”. G. [2] World Health Organization. and Amsden.. LP and Wu. [3] K˚ ageson. and McShane. M. 2007. Kong. G. 2007. and Fureby.. 1995. 1998.. [13] Faeth.. 99–127. D. J. [6] Lagerquist. C. [11] Jasak. pp. M. [8] Patterson. and Reitz. “The centrality of the horse to the nineteenthcentury american city”. Emission standards. 620–631. H. and Nordin. “Health aspects of air pollution with particulate matter. H. C. Per. Hsiang.J. 81 . G. S. 1998. Modeling Engine Spray and Combustion Processes.Bibliography [1] Tarr.J.. a ang [4] www. 1987.. SAE Paper 2004010110. PK. SAE Paper 872089. “The tab method for numerical calculation of spray droplet breakup”. EUR/03/5042688. International Journal of Multiphase Flow.html.A.
R. and Poola. S. Sakata. 20(5). PhD thesis.. International Journal of Heat and Fluid Flow.. “Visualisation of cavitation in diesel engine injectors”. SAE Paper 2004010027.. “Development of a nextgeneration spray and atomization model using an eulerianlagrangian methodology”.. W.. Berg. K. Subramaniyam. SAE Paper 2007011405.H. E. E. D. Sauer. Nakamura. In International Heat Transfer Conference. Mecanique & Industries. “Comparing Cavitation in Diesel Injectors Based on Diﬀerent Modeling Approaches”. S... 2007. 383–394.. Omae. [23] Masuda. and Tatschl. and Arcoumanis. E. 1999. 2001. In Proceedings of the ILASSEurope Annual Meeting. 2003. [16] Arcoumanis.. [21] Martynov.. E.. Roth.. 2001. Mitroglou. SAE Paper 2003011358. Flora... Arcoumanis. D.. R. [20] Mulemane. J... and Gorokhovski. and Yanagihara. 125... J. [17] Blessing. M. N. “Numerical Simulation of TwoPhase Flow in Injection Nozzles: Interaction of Cavitation and External Jet Formation”.. R. Kelz. “Eﬀect of multiinjection strategy on cavitation development in diesel injector nozzle holes”. J.. W. 2003. “Internal ﬂow and cavitation in a multihole injector for gasoline directinjection engines”. 2005... and Schwarz. R. . Lai. [24] Ning. C. A. Wirth. “A numerical study of cavitating ﬂow through various nozzle shapes”.. and Corradini. Kull.. M.. Fath..82 BIBLIOGRAPHY [14] Badock.. Mason.... 963–969. 2006. [26] Nouri.. R. C.. 1997. 2(5). Gavaises. I. H. C. C. and Morozov. 2004. [22] Yuan. “Comprehensive Hydraulic and Flow Field Documentation in Model Throttle Experiments Under Cavitation Conditions”. Yan. S. [25] Roth. 2007. Switzerland. S. M.. Lu. pp. P. Gavaises. A. SAE Paper 2005011237.. M. Michels...... T. Journal of Fluids Engineering. A. G. Mechanique & Industries. pp. University of Wisconsin .. Sazhin.Madison. A.. [18] Schmidt. pp.. C. SAE Paper 971597. “Validation of diesel fuel spray and mixture formation from nozzle internal ﬂow calculation”.. SAE Paper 2005012098. [15] Winklhofer. H.. 574–579. Kr¨ger. 2001. M.. “Modeling and Computation of Unsteady Cavitation Flows in Injection Nozzles”. E. 2. K¨nig. [19] Yuan. o u “Analysis of Flow and Cavitation Phenomena in Diesel Injection Nozzles and its Eﬀects on Spray and Mixture Formation”.. Y.. Han. Nagaoka. and Schnerr. S. C. M... H.. Giannadakis. V... “Investigation of cavitation in real size diesel injection nozzles”. Rutland. Zurich. G. M. H. M. U.. “Modelling of Cavitation Flow in a Nozzle and its Eﬀect on Spray Development”. and Leipertz. Fuyuto. and G. Heikal. M. W.. 2005.
H. C. M. H.. Corbinelli. [31] Wallis. pp. R. [29] Vortmann. B. .P. 1990. [40] Befrui. SAE Paper 2004010100. W.. SAE Paper 2002010215.. October.. Rutland.. pp.. “Application of a oneﬂuid model for large scale homogeneous unsteady cavitation: The modiﬁed Schmidt model”.. [35] Huh. C. 1177–1192.Y.D. 1976. Imperial College. PhD thesis. S. Industrial & Engineering Chemistry Fundamentals.. 2008. FEDSM200737377. “Numerical prediction of two ﬂuid systems with sharp interfaces”.. K. [30] Weller. [37] Peng.. Tech.. 9.S. 103–129. “A Numerical Study of the Inﬂuence of Diesel Nozzle Geometry on the Inner Cavitating Flow”. and Robinson.J. 1992. B. 1995. H. S. “Large Eddy Simulation of Primary Diesel Spray Atomization”. G. A.. Schnerr. G. B. and Seelecke. Proceedings of FEDSM2007. 2003. S.. [38] Xie. D. A. [32] Chung. 1969. D. 2006. [41] M´nard. “Boundary conditions for direct simulations of compressible viscous reacting ﬂows”. New York. F. A New Approach to VOFbased Interface Capturing Methods for Incompressible. B. G. pp. D.. “Numerical Simulation of Diesel Spray Processes”.. C. O. X. [34] Poinsot. ILASS. 35... OpenCFD Ltd. TR/HGW/07. Computers & Fluids. G... T. 24. Imperial College of Science Technology and Medicine. “A Fully Compressible TwoDimensional Model of High Speed Cavitating Nozzles”. Park. Margot. International Journal of Heat and Fluid Flow. 1997. C. [39] de Villiers. T. D. A. G. pp. and Corradini... 2004.K... “Thermodynamic Modeling and Simulation of Cavitating Nozzle Flow”. 276. 13–26. and Lele. [36] Kralj. Journal of Computational Physics. “Sound speed criterion for twophase critical ﬂow”.K...L. Journal of Sound and Vibration. [28] Schmidt. W. M. and Lee. McGrawHill. Liu. 59–64. F. and Gosman. Rep. “A new twoconstant equation of state”.J. and Khoo. SAE Paper 2008010137. “LES Simulation of the Internal Flow and NearField Spray Structure of an OutwardOpening GDi Injector and Comparison with Imaging Data”. [33] Ubbink. 2004.. Onedimensional Twophase Flow... PhD thesis. T. Compressible and Cavitating Flow.. “Atomization Mechanism of Fuel Injection”. 1999.. Gosman. and Weller.B. and Reckers. 101.Y. Robart. Atomization and Sprays.. and Salvador. 2007.BIBLIOGRAPHY 83 [27] Payri.. “Interface tracking with a coupled e level set/vof/ghost ﬂuid method: Application to jet atomization”. 2002.. and Berlemont. 2006.. 15. E.
“Modeling AutoIgnition. G¨teborg. R.... O’Rourke. 2003. T. 2008. 2002. [44] Lebas.D. . “A 3D Eulerian Model to Improve the Primary Breakup of Atomizing Jet”. 1986. o [55] Venugopal. Dept. L. 301–314. I... P. Sci. CRC Press LLC... B. V.. C.. 15.. S. and Crowe. and Demoulin.. J. and Orzag. [54] Karlsson. 660–669. 1972. W.A. 2001. A. and Launder. and Fuchs. “Eﬀect of particle size on modulating turbulent intensity”. Numerical Recipes in C++: The Art of Scientiﬁc Computing. “An Evaluation of a TwoFluid EulerianLiquid EulerianGas Model for Diesel Sprays”.. J. Larmi... 179. 2005. “A numerical investigation of ﬂame liftoﬀ in diesel jets”.. Sommerfeld. PhD thesis. Barbeau. [52] Nordin. 1975. G¨teborg.. and Tsuji. LA11560MS.T. [43] Blokkeel. A. B. [51] Crowe. Heat and Mass Transfer. 15(2). M.P. SAE Paper 2003010005.. and Borghi. SAE Paper 2008010933. [46] Jones.. [45] Iyer.. 1998.A. F. Heat Mass Transfer. “The prediction of laminarization with a twoequation model of turbulence”. [48] Launder. S. pp.... 2003. pp. 1989. and Butler. pp. 1995. 279–285. J.84 BIBLIOGRAPHY [42] Vuorinen. G. Lett. “Renormalization Group Analysis of Turbulence”. and Abraham. B. Basic Theory J. of Thermo and Fluid Dynamics. “Coupling Vaporization Model With the EulerianLagrangian Spray Atomization (ELSA) Model in Diesel Engine Conditions”.X.H. “LargeEddy Simulation on the Eﬀect of Droplet Size Distribution on Mixing of Passive Scalar in a Spray”. M. J. Y. P. Combustion Science and Technology. Blokkeel. [47] Yakhot. C. Chalmers University of Technology. R. Comput. T. Int. Teukolsky. pp. [49] Amsden. 125. “Complex chemistry modeling of diesel spray combustion”.. V. o [53] Press.. Chalmers University of Technology. and Flannery. Vetterling. “Kivaii: A computer program for chemically reactive ﬂows with sprays”. V. G. J.A. W. Cambridge University Press. 2007. International Journal of Multiphase Flow. 131–138. Journal of Fluids Engineering. SAE 2005010213.. Flame Propagation and Combustion in Nonstationary Turbulent Sprays”. Multiphase Flows with Droplets and Particles.. “Application of the energy dissipation model of turbulence to the calculation of ﬂow near a spinning disc”. W.. Beau. 1. and Abraham. pp. 2599–2618. 2:nd ed.J.. May. B.A. PhD thesis. and Sharma. B. R. N. 1989. R. [50] Gore.
.. H. SAE paper 2001011070. “Modeling the diesel fuel spray breakup by using a hybrid model”. [61] Schmalzing.. S. 1999. and Dwyer.. 2001. 1999.. and Veynante.. 2003.. 2003. Kosaka. R. [62] Kimura.D. S...R.. Renner.BIBLIOGRAPHY 85 [56] P. PhD thesis. “Investigation of atomization characteristics and prediction accuracy of hybrid models for highspeed diesel fuel sprays”. SAE 2003011043.. [60] Macpherson..M. PhD thesis. G. [59] Tao. 1996. C. J. N. “Comparison of spray characteristics in butane and diesel fuels by numerical analysis”. [63] Reitz. 164. [69] Schmidt.. S. 309–337... H. Maly. E. C..J. 2000. TwentySixth Symposium (International) on Combustion/The Combustion Institute. 62–80.and Himeno. SAE Paper 2000012941. [64] Park. G. “Numerical Modeling of Soot and NOx Formation in NonStationary Diesel Flames with Complex Chemistry”..A. and Lee. Proceedings of Combustion Inst.. G¨teborg. Matsui... Kim. S. R.. Atomization and Spray Technology. SAE Paper 1999010226. 1987. and Rutland. and Karlsson. [57] Tap. pp. Princeton University.. JH and Goto. Journal of Computational Physics. “A holistic hydraulic and spray model .A. G. pp. Grimaldi. L. 2004. H. “Flame Liftoﬀ in Diesel Sprays”. J.. F. SAE Paper 2004012014. [58] Chomiak. C.. J. “A New Droplet Collision Algorithm”. G. [68] O’Rourke.K.. P. Bella. [65] Rotondi. 2003. “Multidimensional Modeling of DirectInjection Diesel Spray Liquid Length and Flame Liftoﬀ Length using CFD and Parallel Detailed Chemistry”. 919–926.. 2557–2564. Department of Mechanical and Aerospace Engineering.. pp. F. D. and Richards.. arbitrary polyhedral meshes for use in cfd and molecular dynamics”. and Weller. P. 2005. Stapf. “Particle tracking in unstructured. H. P. and Pelloni. “A numerical simulation of turbulent mixing in transient spray by les”. “Atomization of HighPressure Diesel Spray: Experimental Validation of a New Breakup Model”. W. o Chalmers University of Technology. “Collective Drop Eﬀects on Vaporizing Liquid Sprays”..(30). “Modeling Atomization Processes in HighPressure Vaporizing Sprays”.. R. Y. [67] Lee. pp. C. 1981. Stetter.. P. Communications in Numerical Methods in Engineering. 3. D. H.liquid and vapor phase penetration of fuel sprays in di diesel engines”. SAE Paper 2003011045. 2000.O. B. R.. SAE Paper 1999013549. 2008. “Simulation of Flame Liftoﬀ on a Diesel Jet Using a Generalized Flame Surface Density Approach”. and Postrioti. Senecal. K. Nordin. . G. [66] Bianchi.
86 BIBLIOGRAPHY [70] Gustavsson. [74] Hilbert... and Higgins. 20008249. 61–117.... 2007. SAE Paper 2008011328. 2004. ElRabii. SAE Paper 2002300002.. Choi. F. J. [77] Husberg. Phys. Numerical Modelling of Diesel Spray Injection a and Turbulence Interaction. “Impact of detailed chemistry and transport models on turbulent combustion simulations”.. D’Errico. SAE Paper 2008011393.. D. Rep. J. and Th´venin. F. Andersson. “Autoignition of turbulent e nonpremixed ﬂames investigated using direct numerical simulations”. Springer. “3D Simulation of Multiple Injections in DI Diesel Engine”. pp. pp. and Denbratt. V. [79] Gavaises. Eﬀects of Injector Conditions on the Flame LiftOﬀ Length of DI Diesel Sprays. 2000.H. [82] Ferziger. H. “Automatic mesh motion with topological changes for engine simulation”. Chalmers University of Technology.. . Licentiate Thesis. B. 1977. D. Roth. I. D.... 128. 22–37..gov/ecn).. “Optical studies of spray development and combustion characterization of oxygenated and ﬁschertropsch fuels”. T. Combustion and Flame. J. 2002. S. [78] Payri. 2007.. upstreamcentered ﬁnitediﬀerence schemes for ideal compressible ﬂow”.. V. R.. R. “Analysis of advanced multiple injection strategies in a heavyduty diesel engine using optical measurements and cfd simulations”. and Peric. M. 263–75.. Department of Applied Mechanics. B. [71] Peng K¨rrholm. G. [81] Ochoterena.. Berm´dez. and Th´venin. and Theodorakakos. e [75] Hilbert... Larsson.. H. 83. 30.. In The Sixth International Symposium on Diagnostics and Modeling of Combustion in Internal Combustion Engines. Fuel.ca.. A. and Karlsson.. 2008. F.. Jasak. M. [72] van Leer. S. Sandia National Laboratories. [73] Sandia National Laboratories. 2006. Computational Methods for Fluid Dynamics. Payri. 2004. R. and Salvador.. Z. Engine combustion network (public.. Tap.. and Tukovic. COMODIA. Comp. 23. Progress in Energy and Combustion Science. Tech. T. Arcoumanis.. M. A. [76] Siebers. 2008. Denbratt. 419–431. SAE Paper 2007010170. “Towards the ultimate conservative diﬀerence scheme iii. I. “Nozzle ﬂow and spray characteristics from vco diesel injector nozzles”.sandia. R. Y. 2004. and Golovitchev.. L.. “The inﬂuence u of cavitation on the internal ﬂow and the spray characteristics in diesel injection nozzles”. 1999. pp.. F. C. [80] Lucchini. pp. H. 2002..
Paper II .
.
When a high velocity ﬂuid passes through a contraction like a nozzle. and the occurrence of supercavitation in the channel. In addition. fuel is injected into the combustion chamber by high pressure.se Henry Weller OpenCFD Ltd. a pocket of low pressure is formed in the wake of its edge. Ideally. Reading UK Niklas Nordin Scania CV AB ¨ ¨ Sodertalje Sweden ABSTRACT In this paper. which uses a barotropic equation of state together with the homogeneous equilibrium assumption. are mainly governed by the ﬂow in the nozzle [1] which ejects the spray. .sat liquid density at saturation ψv vapour compressibility ψl liquid compressibility ∗ Address INTRODUCTION In Diesel engines and direct injection spark ignition engines. cavitation and pressure parameters measured in a model diesel injector are compared to data acquired by numerical simulations using a new code developed for the OpenFOAM platform. The mass ﬂow and cavitation probabilities obtained from the simulations are compared to data obtained in experiments performed at AVL’s laboratories. Caversham.Proceedings of FEDSM2007 5th Joint ASME/JSME Fluids Engineering Conference July 30August 2. and eventually evaporate completely in order to obtain a mixture that can be easily combusted without formation of soot or NOx . The size of the cavitation bubbles thereby formed depends on several factors.sat vapour density at saturation ρl. including (inter alia) the injection pressure. In this wake the pressure can decrease below the saturation pressure. It is therefore important to study the ﬂow inside the nozzle.pengkarrholm@chalmers. California USA FEDSM200737518 MODELLING INJECTOR FLOW INCLUDING CAVITATION EFFECTS FOR DIESEL APPLICATIONS ¨ F. mass ﬂows. Peng Karrholm∗ Department of Applied Mechanics Chalmers University of Technology Sweden Email: fabian. NOMENCLATURE A ﬂow area of the nozzle Ai area of a face in the nozzle in ﬂow direction Cd discharge coefﬁcient p pressure ∆ p pressure difference U velocity m mass ﬂow ˙ γ vapour fraction µg vapour viscosity µl liquid viscosity µ f ﬂuid viscosity ρ ﬂuid density ρv vapour density ρl liquid density ρv. and grid independent. to ﬁnd out how it is linked to the ﬂow in the combustion cylinder. Important features of the atomisation. The experimental data used include pressure proﬁles and cavitation images. 2007 San Diego. in which the ﬂow through an almost twodimensional nozzle was examined. and thus cause the liquid to cavitate. The model proved to be able to predict cavitation probabilities. the fuel should atomise. including the size of the initial droplets and the fragmentation processes. it proved to be stable in its dependency on physical parameters. geometry. allowing both compressible liquid and vapour to be modelled. form small droplets. and 1 Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology all correspondence to this author. It is a viscid code.
and thus the pressure is found as a function of density. There are also hybrids. Associated with the DDM model are several submodels describing injection. in which two of the possible modelling approaches to cavitation was explored. for two reasons. This type 2 of model was used in [6–8]. the most common way of modelling sprays is still to use a combination of Eulerian gas and Lagrangian liquid descriptions. evaporation. rather than studying the actual nozzle ﬂow. is to use CFD to investigate the nozzle ﬂow and couple it to the spray. and a third rarely used method based on Gibbs free energy [17]. adopted by Sarre et al [12]. Continuum methods include the RayleighPlesset bubble method [14. if the injection pressure is increased. both the bubbles and the liquid surrounding them will be heated by the gas in the chamber. since the velocity in diesel injectors is high. If there is vapour along the walls. the code we used is a simpliﬁed Volume of Fluid (VOF) code that allows both diesel vapour and air to be modelled. There are interface tracking methods. However. diffusion and bubble coalescence can be assumed to have negligible effects. and thus the entire atomisation process will be accelerated. This paper focuses on verifying a code for a cavitating injector. if there is no such point the ﬂuid will be supercavitating and the cavitation region will extend throughout the entire nozzle. There is also a study by Yuan et al [2]. which have two explicit phases. thus reducing its lifetime. It also increases the maximum velocity. When the vapour bubbles formed in the nozzle reach the combustion chamber. Cavitation can be modelled in several ways. In this paper both phases are compressible. Notably it can erode the injector. atomisation. This removes the option of increasing spray penetration by using higher pressures. it was deemed to the most suitable method for this work. Further. For example. but on theoretical considerations and assumptions. due to the complexities associated with the presence of three phases: liquid. there will be no cavitation and ordinary pipe ﬂow will occur. as it is of particular interest when studying cavitation. and assumes gas diffusion and bubble coalescence to be important factors [13]. combustion etc. in which Lagrangian parcels are added to the spray when the liquid phase is sufﬁciently disperse. since it severely reduces the atomisation. The cited authors present pressure plots similar to those in this article. since the vapour by the wall will not be stationary. Sarre et al focused on the spray calculations. One of the most common approaches when modelling injection is to assume the droplets follow a distribution with some estimated parameters. supercavitation should be avoided.smoothness of the interior of the nozzle. However. The injection model will be described further. It also differs from the one presented by Schmidt [16]. like the model described in [5]. which is required when the pressure over the nozzle is high. and the effect the ﬂow in the nozzle has on the spray. since the spray’s cone angle is severely reduced when it exits the injector. e. and gas which the liquid is being injected into. their rationale has not been based on CFD or experimental data. 15]. The liquid will evaporate faster if it has already been partially evaporated by cavitation. However. for a cavitating injector the ﬂow rate is still constant. since it is based on the speciﬁed bubble density. The velocity is also increased since the liquid can not ﬁll the entire channel. Previous works related to cavitation in nozzles under diesel conditions have focused on various aspects of the associated phenomena. which attempt to couple the nozzle ﬂow to the injected droplets. the liquid will not have a noslip boundary condition. barotropic equation of state method [13. The barotropic equation of state method described in [13] differs from the one used here in the sense that it assumed the two ﬂuid phases to be incompressible. Therefore. Diesel injection into air without cavitation has also been simulated in [4]. The RayleighPlesset (RP) approach is therefore unsuitable. which originates from ﬂow separation at the nozzle inlet. vapour formed from the cavitated liquid. and then sample the droplet size from that distribution.g. Ultimately it will impair performance of the engine. mainly because it is most appropriate for the cavitation that occurs in diesel injectors. if the ﬂuid is pushed through the nozzle sufﬁciently slow. in which an equation of state is analytically integrated. and continuum methods which have just one explicit phase. and the properties of the ﬂuid. resulting in a steady ﬂow with fairly large scale disturbances on the surface [2]. the study by Mulemane et al [13]. The point where the ﬂuid is no longer cavitating is called the reattachment point. and thus higher atomisation. Few attempts have been made to model the entire spray using only an Eulerian framework. thereby providing a sound foundation for the subsequent development of a new injection model. in which the cavitating ﬂow inside the nozzle as well as the atomising spray close to the nozzle is modelled. The method chosen for cavitation modelling for this work is the barotropic equation of state approach. 16]. More ambitious injection models. Cavitation introduces vapour bubbles into the ﬂow. have been presented [9–11]. Supercavitation changes the atomisation behaviour signiﬁcantly. In other studies only the nozzle ﬂow has been simulated. but do not compare their numerical results with experimental data. breakup. and those that have been made have usually been done in 2D [3] or have been limited to a restricted part of the spray [4]. Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology . This allows a higher velocity of the liquid. Since the Barotropic equation of state is more consistent with the VOF framework. which increases the physical validity. PREVIOUS STUDIES ON CAVITATION IN NOZZLES Numerical studies on Diesel sprays are usually based on EulerLagrangian modelling according to the Discrete Droplet Model (DDM). There are also other negative aspects of cavitation aside from supercavitation. One option. This approach is not common. Further.
Experimental data in the ﬁeld are very scarce, and many models have only been validated (if at all) by comparing simulated and empirical discharge coefﬁcients, cavitation numbers and average velocities. This is due to the difﬁculties involved in experimentally investigating the nozzle ﬂow. Often the cavitation along the sides of the nozzle makes it very difﬁcult to evaluate the ﬂow inside, since the cavitation hinders optical analysis. However, in experiments presented in [18] a nozzle was set up in such a way that cavitation was limited to two of the four sides of the nozzle, thus allowing the pressure and velocities of the ﬂow to be measured. The data acquired in these experiments are used to validate the model presented in this article. The main objective of this paper is to introduce the new model, by ﬁrst describing the theory underlying the equation of state (EOS) used, and then the momentum equations and how they are solved. The experimental case study used to compare to the simulations is brieﬂy described, and the numerical setup. Finally, results are presented, including both pressure proﬁle comparisons and cavitation probability images.
has been reached. Before equilibrium is attained errors will arise, and the code will not yield corrects result when it has been attained. The equation of state should therefore be consistent with the liquid and vapour equations of state both at the limits when there is pure liquid or pure vapour, and at intermediate states when there is some form of mixture. The two states can be described by the linear equation of state:
ρ v = ψv p ρl = ρ 0 + ψl p l
(2) (3)
The parameter that describes how much of the liquid is in each phase is γ: ρ − ρl,sat ρv,sat − ρl,sat
γ=
(4)
THEORY To model the ﬂow in the nozzle, we need to model the cavitation, and either resolve the turbulent scales or model them. Turbulence is not modelled, due to the relatively weak inﬂuence it is believed to have on the nozzle ﬂow, compared to the cavitation [14]. The only way turbulence is addressed is in stabilising numerics (in contrast to DNS computations) which are not sufﬁciently accurate to include all scales of the turbulence. As mentioned above, the barotropic equation of state was the method chosen for modelling cavitation. We also assumed homogenous equilibrium, i.e. the liquid and vapour are assumed to always be perfectly mixed in each cell. The parameter γ describes the amount of cavitated vapour in each cell. For the method chosen to model the cavitation, an equation of state is needed. The temperature is assumed to be constant, which is not a major limitation [18]. A common barotropic equation of state is the nonequilibrium differential equation:
γ = 1 corresponds to a fully cavitated ﬂow, and γ = 0 a ﬂow with no cavitation. ρv,sat is calculated from
ρv,sat = ψv psat
(5)
ψv is the compressibility of the vapour. These properties together form the mixture’s equilibrium equation of state:
ρ = (1 − γ)ρ0 + (γψv + (1 − γ)ψl ) psat l +ψ(γ) p − psat
(6)
Dp Dρ =ψ Dt Dt
(1)
The compressibility ψ was ﬁrst mentioned in equation 1, but nothing was said of how it is chosen. It can be modelled in several ways, one very common way is the Wallis model [19]. The model chosen here is a simple linear one:
ψ refers to the compressibility, which is the inverse square root of the speed of sound in the cavitating mixture. This equation can either be used directly in the continuity equation to formulate a pressure equation, or integrated to obtain the pressure as a function of the density. The latter approach was adopted by Schmidt et al [16], as mentioned earlier. The former approach is problematic because equation 1 is not an equilibrium equation of state, so the pressure and density obtained are not consistent with the liquid and vapour equations of state unless equilibrium 3
ψ = γψv + (1 − γ)ψl
(7)
A linear model was selected partly because the code we used is based on a VOF code, for which linear combinations of γ are highly suitable for describing the mixtures properties, and partly because it gives greater stability than the Wallis model. The γ value is also used to calculate the mixture’s viscosity, in the same manner as the compressibility: Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology
µ f = γµv + (1 − γ)µl
(8)
When a linear model is used for the compressibility, the equation of state (6) can be simpliﬁed into:
same TVD scheme with MUSCL limiter is used for the velocity divergence term. It was found to be very important to choose a proper scheme for the velocity, if a too stable scheme is used the cavitation region will be ﬂat. Following this is an ordinary PISO loop, to solve for p and correct the velocity to achieve continuity. For a description on how OpenFOAM achieves continuity and solves for p, see [20]. The equation solved in the PISO loop is again the continuity equation, but this time transformed into a pressure equation by use of equation 6.
ρ = (1 − γ)ρ0 + ψp l
(9) ∂γ ∂ψp − ρ0 + (ψl − ψv ) psat l ∂t ∂t ∂ψ + ∇ · (ρU) = 0 −psat ∂t
The ﬁrst term governs the liquid’s density when γ is low. If the ﬂuid is cavitated, the second term becomes more dominant. The ﬁrst term contains the property ρ0 which is: l
(13)
0 ρl = ρl,sat − psat ψl
(10)
ρl,sat is the liquid density at standard conditions. The saturation density of the vapour, previously used to calculate γ, is important for the liquid’s tendency to cavitate. A higher ρv,sat will not require as low ρ to get a higher γ. Unlike the model used by Schmidt et al [16], the code does not solve the Euler equations for inviscid ﬂow to obtain the velocity and density. Viscous terms are included, and the momentum equations are solved for U, not ρ U which is common for frictionfree compressible ﬂows. A brief description of the methodology of the code is given for clarity. The iteration process starts by solving the continuity equation (eq. 11) for ρ, which is where rho is convected.
When continuity has been reached, the remaining properties γ, ρ and ψ are updated. ρ is now updated according to the EOS, eq. 6. The time step is limited by both the Courant number, and the acoustic Courant number, respectively deﬁned as:
Co = max( and
U )∆t ∆x
(14)
1 Coacoustic = max( √ )∆t ψ∆x
(15)
∂ρ + ∇ · (ρU) = 0 ∂t
(11)
The ρ in the divergence term ∇ · (ρU) is interpolated by use of a total variation diminishing (TVD) scheme named MUSCL. The MUSCL scheme is second order in space, and appropriate when there are large gradients present, like the one between liquid and vapor density. Several other options are easily available in the code. ρ is also used to obtain preliminary values for γ (eq. 4) and ψ (eq. 7), and when deﬁning the momentum equations.
The Courant number was chosen to be limited to 0.125, and the acoustic Courant number was limited to 12.5 in all of the simulations. However, the nonacoustic Courant number is the real limiter of the timestep, and the acoustic Courant number is usually around 23. The model described is implemented for the OpenFOAM platform, under the name cavitatingFoam (in OpenFOAM 1.4). OpenFOAM is an open source toolbox for CFD, written in C++, available at www.openfoam.com.
∂ρU + ∇ · (ρUU) = −∇p + ∇ (µ f ∇U) ∂t
(12)
The momentum equations are used in an ordinary fashion to get the matrices used to calculate the pressurefree velocity. The 4
STUDY The results produced by the code are compared to data acquired in experiments performed at AVL’s laboratories, in which diesel fuel was pumped through a nozzlelike structure. The setup is described in [18], and the model nozzle is shown in ﬁgure 1. The experiments generated data on density/pressure as well as images of the cavitation region, and due to the simplicity of the geometry they are suitable for comparison to numerical Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology
Some of the values in table 1 have large margins of error. For example, the compressibility of diesel vapour at room temperature is an estimation based on the ideal gas law. The compressibility of the liquid is taken from measurements of the speed of sound in diesel fuel [21], which is around 1400 m/s. The evaporation pressure chosen was that of heptane, since it is commonly used as a model fuel for diesel. The liquid density was measured in the experiments, while the vapour density was from the equation
ρv,sat = ψv psat
(16)
Figure 1. VIEW OF THE EXPERIMENTAL NOZZLE, BEFORE AND AFTER THE NOZZLE
1 cm CHANNEL
The gas viscosity was obtained from the Lucas viscosity model described in [22] (again based on nheptane’s properties), and the liquid viscosity from Petrocard’s data on their diesel fuel products. As will be shown in the results section, the model is not sensitive to many of these physical properties (apart from the liquid viscosity), as long as they are of the right order. The walls on the sides of the nozzle, have an ordinary noslip condition for the velocity. The upper and lower walls, on the other hand, have a slip condition for the velocity, similar to that used in supersonic compressible ﬂow. Perpendicular to the wall the ﬂow (or ﬂux) is zero, but the ﬂow in the tangential direction of the wall is not. A noslip boundary condition would lead to overestimates of the thickness of the boundary layer, resulting in strong vortices in the centre of the channel, and incorrect modelling of the ﬂow’s cavitation. Figure 3 shows these vortices and the resulting cavitation. As mentioned earlier, viscosity is sometimes disregarded entirely, here we have chosen to disregard the viscosity from two walls due to the assumed thinness of the boundary layer on those two walls.
Figure 2. CLOSEUP OF THE MEASURED REGION,
299 µm WIDE, RESULTS In all of the experimental cases the injection pressure was 10 MPa, and the outlet pressure was varied. The two outlet pressures used here for detailed comparison are 5.1 MPa, 3 MPa and 1.5 MPa. In the experiments cavitation was measured by a line of sight method, i.e. all cavitation observed from the top of the nozzle was recorded, and then averaged as shown in ﬁgure 4. The numerical simulations are evaluated by a similar line of sight method, in which the domain is searched for cavitated cells. Both the experimental and numerical values are averaged over 20 timesteps, resulting in images such as the one shown in ﬁgure 5. The average value at time t n is calculated from
1 mm LONG data. An image of a typical mesh can be seen in ﬁgure 2. In the experiments the injection pressure was 10 MPa, and the outlet pressures were varied.
Table 1.
DATA ON DIESEL FUEL
property density (kg/m3) viscosity (kg/ms) compressibility (s2 /m2 ) vaporation pressure (Pa)
liquid 832 0.0065 5 · 10−7
vapour 0.1361 5.953 ·10−6 5.4 ·103 2.5 ·10−6
Un = 5
U n−1 t n − t 0 + U n∆t n t n + ∆t n − t 0
(17)
Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology
i. could be underestiCopyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology . This has been deﬁned as: Cd = mcalculated ˙ = mBernoulli ˙ ∑ ρiUi Ai i N 2∆pρ0A (18) Figure 5. MEAN ISOSURFACE FOR VAPOR FRACTION OF 0. and 10 % for the non cavitating case (∆p = 5.1 MPa 3 MPa 1. IN JECTION PRESSURE 7 MPa Table 2.1 MPa).46 ·10−3 mmodel ˙ 8. and ∆t n is the timestep at time n. INJECTION PRESSURE 7 MPa LOCITY FIELD IN A PLANE PERPENDICULAR TO THE FLOW DIRECTION SHOWN TO VISUALIZE THE VORTICES. The mass ﬂow behaviour is 6 pout 5. IMAGE OF THE EXPERIMENTAL CAVITATION PROBABILITY. The reason why the mass ﬂows differ. when the cavitation region extends throughout the entire nozzle. VE Figure 4.31 ·10−5 2. Figures 8 and 9 illustrate these results.Figure 3.e.98 ·10−3 8. Deviation 2. which shows how much more losses there are in the strongly cavitating case.5 MPa mexperiment ˙ 6. The mass ﬂow is calculated for several timesteps.1.5 MPa. COMPARISON OF MODELLED AND EXPERIMENTAL DATA The simulations (ﬁgures 5 & 7) predict the experiments (ﬁgures 4 & 6) quite well. This occurs when the outlet pressure is 1.80 1.25 ·10−4 Cdmodel 0.54 ·10−3 7. The experimental data also include pressure measurements along the channel centreline and a line 40 µm from the channel wall.49 ·10−4 predicted within 1% for the cavitating cases.46 ·10−3 8.53 ·10−3 Std. The discharge coefﬁcient is also listed. The model manages to predict when the ﬂow is supercavitating.53 ·10−3 8. It is important to point out that the measured pressure is much higher than the saturation pressure for the liquid. but the cavitation probability is still shown to be 100 % close to the wall (4). When comparing the different cases. With the mass ﬂow constant for increased pressure difference. The model can also predict the choking of the channel which occurs when it starts to cavitate. the discharge coefﬁcient will decrease with increasing pressure difference. and an equation of state. and as can be seen the calculations do not predict the experimental results well. and averaged in time over the last 100 µs.93 0. even though the simulations yield slightly more cavitation than observed in the experiments. the channel starts at 0 and ends at 1 mm. NOSLIP CONDITION USED ON BOTH UPPER AND LOWER WALL t 0 is the start time for the average. In the pressure proﬁle plots. The pressures in the experiments are derived from density measurements. IMAGE OF THE COMPUTED CAVITATION PROBABILITY. the discharge coefﬁcient is used to measure how much losses there are in the injector.88 0. This will be discussed later in the Conclusions section. but not when it is 3 MPa.
is often used. and the quarter mesh consisting of only 508 000 cells. However. The case with an outlet pressure of 3 MPa was examined with both the full 3D mesh. which reduces the friction more than the lack of rough walls. Another factor could be the lack of friction in the simulations on the two walls with slip boundary condition. consisting of 1. while a sharp inlet gives more cavitation. since there is less cavitation causing losses. if such a restriction is imposed turbulence will not be correctly modelled. still an attractive option due to the reduction in computational time. in this case 90 degrees of the full mesh. However. However. INJECTION PRESSURE 8. Rounding was not varied in the experiments reported in [18].5 MPa 0 −5 0 5 Position (m) 10 x 10 15 −4 mation of the liquid viscosity.12 10 8 6 4 2 x 10 6 51 exp 51 comp 30 exp 30 comp Pressure (Pa) 0 −5 0 5 Position (m) 10 x 10 15 −4 Figure 6. INJECTION PRESSURE 8. Since the rounding does not affect the total pressure drop over the nozzle. Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology . It is however. for the case of 500k cells and an outlet pressure of 3 MPa. differences in mass 7 Figure 9. IMAGE OF THE COMPUTED CAVITATION PROBABILITY. The effects of inlet rounding are often considered since it affects cavitation. This simpliﬁcations was forced due to the presence of a strong cavitating vortex as mentioned earlier. a quarter of the mesh. so the exact values can not be veriﬁed. The data shown in Figure 11 and table 3 were produced using a quarter of the channel. AVERAGE PRESSURE ALONG THE CENTRELINE OF THE NOZZLE 6 12 10 Pressure (Pa) 8 6 4 2 x 10 51 exp 51 comp 30 exp 30 comp Figure 7. there are high uncertainties in the data. it can be seen as a way of examining the ﬂow as if there were no cavitation at all. The results (Figure 10). To save computational time. The discharge coefﬁcient also increases.7 million cells.5 MPa Figure 8. to reduce computational time. a well rounded inlet can reduce cavitation. to examine possible differences in the outcome. As mentioned when the physical properties were shown. show that the quarter mesh provides a possible approximation to the three dimensional ﬂow. IMAGE OF THE EXPERIMENTAL CAVITATION PROBABIL ITY. and it will not be possible to capture large eddies. AVERAGE PRESSURE ALONG THE WALL OF THE NOZZLE ﬂow can be seen by comparing tables 4 and 2. the results show that the model can predict the reduced cavitation and increased ﬂow rate as the inlet becomes more round.
CAVITATION PROBABILITY. as can be seen in ﬁgures 1213 and table 4. the cavitation region will consist of very few cells. WITH FOUR DIFFERENT GRIDS.12 x 10 6 10 quarter mesh. The ﬁgure is zoomed in on the cavitation region. As mentioned in the Study section. with ca.0 MPa Table 3.. INJECTION PRESSURE 7 MPa 15 x 10 −4 Figure 11. since the ﬁrst are from the quarter mesh.912 0.5 position (m) 1 x 10 1. FOUR DIFFERENT GRIDS USED. since it provides a clearer indication of how much the ﬂow has cavitated. An important factor to examine when trying to model a ﬂow is the gridsize. The standard grid chosen for this study has 508 000 cells. some of the parameters 8 Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology . and close to the rounded inlet. DISCHARGE COEFFICIENTS AND MASS FLOWS FOR DIFFERENT VALUES OF THE INLET ROUNDING RADIUS rinlet 40 mm 50 mm 60 mm m ˙ 8. PRESSURE DIFFERENCE 7. The numerical value for the mass ﬂow in table 4 is different from the ﬂows listed in table 2. to make it easier to see the small differences.5 x 10 6 Figure 10.00 ·10−3 8. wall 3D mesh.5 0 0 0. 5 µm cells in the channel.5 1 600 500 400 300 200 100 0 −5 0 5 position (m) 10 r=40mm r=50mm r=60mm 0. The grid is graded to be ﬁner close to the wall in the cavitating region.889 0. AVERAGE PRESSURE ALONG THE CENTRELINE AND WALL OF THE NOZZLE.5 −3 Figure 13.60 ·10−3 9. centre 3D mesh. and the latter from the full 3D grid. The calculations are not sensitive to gridsize.930 We have chosen to display density instead of pressure. PRESSURE ALONG THE CENTRELINE OF THE NOZZLE. QUARTER AND FULL GEOMETRY RESULTS SHOWN. If a grid coarser than the one with 140 000 cells is chosen. THREE DIFFERENT INLET RADII. PRESSURE DIFFERENCE 7 MPa 2 1000 900 800 700 density (kg/m ) 3 300k cells 500k cells 700k cells pressure (Pa) 1. wall 8 pressure (Pa) 6 4 2 0 −5 0 5 position (m) 10 x 10 15 −4 Figure 12.83 ·10 −3 Cd 0. DENSITY ALONG THE WALL OF THE NOZZLE. centre quarter mesh. and the assumption of homogenous equilibrium becomes invalid. INJECTION PRESSURE 7 MPa 2.
unlike the walls in the simulations. Smooth walls might inhibit cavitation. To further validate the model.Table 4. CLOSEUP OF THE PRESSURE ALONG THE WALL. PRESSURE DIFFERENCE 7. which could be due to underestimation of the liquid’s viscosity. and smaller at greater distances from them. It has been shown that the model can predict the cavitation probability in the nozzle. If the cavitation region is only large close to the walls. many previously used cavitation codes have assumed completely inviscid ﬂow).5 10 v 2 3 3 −5 −5 −3 pSat = 5.88 0. higher injection pressures should be tested as diesel injection pressures often exceed 100 MPa.89 8. However. although the cavitation probability is said to be 100 % in this region (ﬁgure 4).0 MPa. MASS FLOWS AND DISCHARGE COEFFICIENTS CALCU LATED FROM THREE DIFFERENT GRIDS. and thus did not result from the pressure drop following the contraction. ψ = 2. Since the goal is to simulate diesel sprays. If the average pressure really would be of the magnitude suggested by the pressure proﬁle plots. and the effects of variations in the pressure differences over the nozzle.5 10 pSat = 5. The real walls in the experiments are not perfectly smooth. The experimental data were acquired by a line of sight method. since the friction is reduced when the walls are made inviscid (and. There is a slight over prediction for the cavitation probability. The inﬂuence of the walls should also be considered. When normal nonslip conditions were applied. since the walls have different boundary conditions. since the mass ﬂow does not increase above the value it has when the ﬂow starts to cavitate. WITH FOUR DIFFERENT COMBINATIONS OF ψv AND psat are difﬁcult to measure experimentally. It would also increase the mass ﬂow. The simulations indicate that the cavitation region is spread out and is not interrupted by any large noncavitating areas.4 10 . the ﬂow would not cavitate at all. ψ = 2.60 ·10−3 6 5 4 3 2 1 x 10 6 pSat = 5. more simulations would be needed. since they provide fewer cavitation sites. The model is also capable of predicting choking of the ﬂow in the channel. CONCLUSIONS AND FUTURE WORK A new model for calculating the ﬂow in a cavitating nozzle has been developed.59 ·10−3 Cd 0. However.57 ·10−3 8. pressure (Pa) ACKNOWLEDGMENT We are grateful to Dr E.4 10 .89 0.5 10 v v pSat = 5. These cavitated regions were separated from the wall cavitation. if the same slip condition was applied to all walls. ψ = 2. preferably with comparison to other sets of experimental data. The pressure proﬁle from the simulations are not similar in shape or value to the experimentally derived proﬁles. notably. introduction of rough walls would require a new form of modelling. Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology . However. It was found that the vortices originated from the top and bottom walls.4 103. To test the model’s sensitivity to the ψv three different values of this parameter were tested for the pdi f f = 7 MPa case. a more physical behaviour of the ﬂow was observed. and few data are available. since the barotropic equation of state is not dependent on the presence of cavitation sites. The results are shown in ﬁgure 14. They are slightly better for the noncavitating case.4 10 . due to strong vortices. the experi9 mental data show a minimum pressure close to the wall (ﬁgure 9) at around 2 MPa. QUARTER OF DOMAIN USED rinlet 300k cells 500k cells 700k cells m ˙ 8.5 10−7 0 0 1 2 3 4 x 10 5 −4 coordinate (m) Figure 14. ψv = 2. The authors have therefore chosen not to try to match the simulated pressure curve to the experimental ones. this could explain why the average pressure is said to be much higher than the evaporation pressure of the liquid. and if the boundary condition on those walls were changed to a slip condition. the cavitation region would grow larger (preliminary calculations suggest this). the ﬂow started to cavitate in the central part of the channel. The model has been validated against data obtained in experiments with a simple contraction type nozzle. so variations in the width of the cavitation region at different heights in the nozzle would not have been seen. Winklhofer at AVL for allowing us to use their experimental data applied in the veriﬁcation.
Barbeau. “Structure and Breakup Properties of Sprays”. “Combustion and Emission Modeling for a Direct Injection Diesel Engine”. B... [16] Schmidt.P. S.L. J. Han. S. G.. “Atomization of HighPressure Diesel Spray: Experimental Validation of a New Breakup Model”. P. B. and Salvador. ILASSEurope Zaragoza. and Gosman. and Merker. Journal of Fluids Engineering. SAE Paper 2003010005.. SAE Paper 2004010100. 2003. 24. Journal of Fluids Engineering.. and Schnerr. R. 5th ed. and Van Gerpen. 80(11). Grimaldi. The Properties of Gases and Liquids.D. “Modeling the Effects of Injector Nozzle Geometry on Diesel Sprays”. X. and Corradini. G. [3] Iyer. N. [9] Rotondi. E.. A. J. “Atomization Mechanism of Fuel Injection”.. D. 2004. Switzerland. 2001. R.. “A Fully Compressible TwoDimensional Model of High Speed Cavitating Nozzles”. pp. F. J... “Comprehensive Hydraulic and Flow Field Documentation in Model Throttle Experiments Under Cavitation Conditions”. 2003. [2] Yuan. pp. B.. SAE Paper 2005010207. [22] Poling. G. [7] Lippert. 2. SAE paper 2001011070. and Poola. pp. pp. C.. S.. [8] Kim.. Lai.J. “A 3D Eulerian Model to Improve the Primary Breakup of Atomizing Jet”. M. Schnerr. H. Atomization and Sprays. [4] de Villiers. and Schmidt... [14] Payri. McGrawHill. J. “A Numerical Study of the Inﬂuence of Diesel Nozzle Geometry on the Inner Cavitating Flow”. S. Bella.. [21] Mustafa. Sauer. 2001. N. 383–394. pp..H.. Onedimensional Twophase Flow. and Weller. [15] Yuan. T. 1990. ILASS. S. “Numerical Simulation of TwoPhase Flow in Injection Nozzles: Interaction of Cavitation and External Jet Formation”. Subramaniyam.. and Seelecke.. 1127–1130.. A. Prausnitz. 125... “Large Eddy Simulation of Primary Diesel Spray Atomization”. pp. [18] Winklhofer. C. 99–127. E. 2003. Kull. W. 2001. 1999. A. H. and Morozov. [19] Wallis. H. 21.REFERENCES [1] Faeth. [11] Huh. R. Copyright c 2007 by ASME Copyright © 2007 Chalmers University of Technology . Kelz. G.. SAE Paper 2002010215. J. E. Department of Applied Mechanics. “An Evaluation of a TwoFluid EulerianLiquid EulerianGas Model for Diesel Sprays”. F. and Nordin. 1969..... Zurich. C. McGrawHill. and Borghi.. and Reitz. and G. J. Lu.. and Postrioti. International Journal of Heat and Fluid Flow. W.. [5] Blokkeel.. G. 2003. D. SAE Paper 1999010912. A. A.. 2006. “Modeling and Computation of Unsteady Cavitation Flows in Injection Nozzles”. [12] von Kuensberg Sarre. 125. 2002. [20] Peng K¨ rrholm.. E.. 2004. Gosman. LP and Wu. L. Are. Chang. 2005. PK. SAE Paper 2004010104.. C. M. 1999. 2001.. C. S. 574–579.. Stegemann. “Effect of Temperature and Isentropic Bulk Modulus of Mixtures of Biodiesel and Diesel Fuel”. M.. New York. Journal of the American Oil Chemists’ Society.. SAE Paper 2005012115. Margot.. 1995.. E. K. 2002.... R. F. G.. [10] Baumgarten... and Sung. SAE Paper 2004010027. Mechanique & Industries. “Numeria cal Investigation of Mesh/Turbulence/Spray Interaction for Diesel applications”.. V. and O’Connell.Y.. Kong. J.. 2003. R. 2005. Chalmers University of Technology. Numerical Modelling of Diesel a Spray Injection and Turbulence Interaction. “Mesh Independence and Adaptive Mesh Reﬁnement For Advanced Engine Spray Simulations”. “Comparing Cavitation in Diesel Injectors Based on Different Modeling Approaches”. [13] Mulemane. 2004.. “Thermodynamic Modeling and Simulation of Cavitating Nozzle Flow”. G. and Hsiang. and Abraham.. ILASSEurope Annual Meeting.... 10 [17] Vortmann. “A New Model For Cavitation Induced Primary Breakup of Diesel Sprays”. 660–669. 9. International Journal of Multiphase Flow. [6] Peng K¨ rrholm. 963–969.. Rutland.