A FINITE VOLUME, CARTESIAN GRID METHOD FOR...
Transcript of A FINITE VOLUME, CARTESIAN GRID METHOD FOR...
A FINITE VOLUME, CARTESIAN GRID METHOD FOR COMPUTATIONAL
AEROACOUSTICS
By
MIHAELA POPESCU
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2005
Copyright 2005
by
Mihaela Popescu
To my daughter Ileana Klein, my parents, my brother and my sister, for their love and support
iv
ACKNOWLEDGMENTS
I would like to express my deepest gratitude to my advisor Dr. Wei Shyy for his
subtle yet effective methods of encouraging all of his students. I will always value my
good fortune to have been one of them. Seldom does one encounter individuals with
intellectual caliber, scientific temperament and a spirit of humanity that he embodies. I
am equally grateful for his enduring enthusiasm and boundless patience during the
process of preparing me to be a researcher and contributor to the global scientific
community. Without his enthusiasm and commitment to excellence, my research in this
field/area could not have been accomplished.
I would like also like to express my deepest appreciation for Dr. Mark Sheplak who
assessed my strength and successfully directed my research toward acoustics. I have
benefited substantially from his guidance, experience, knowledge, and philosophy both
professionally and personally.
Similarly, I thank sincerely my committee members Dr. Lou Cattafesta, Dr. Nam
Ho Kim and Dr. Jacob Nan-Chu Chung for their support, encouragement and sharing
freely of their expertise whenever it was needed.
Being part of the Computational Fluid Dynamics (CFD) group was a source of
pride and honor. I have always felt fortune to be part of this group and am grateful
especially for their genuine friendship and support. I benefited from both collaboration
and the pleasant work environment offered by members and visitors of CFD group. They
were a real team as well as family for me. Our advisor and friend, Dr. Wei Shyy, fostered
v
close and sincere relationship between members and visitors which enhanced the
experience for all of us.
I would like to thank my cousin Catalina and Laly Chirita, my friend Rada
Munteanu, my friends from home Florina Carabet, Cosmin Carabet and Lascu Luminita
for finding the most extraordinary and creative ways of being there for me whenever I
needed them.
I would like to thank to my family members: my parents, Aurica and Conatantin
Popescu; my brothers and sisters, Pavel, Teofil, Marinela, Daniel Popescu, and Tabita
Chirita and their family, for their sustaining support and love throughout years. They
were the undisclosed source of energy and strength that I relied on during challenging
times
Finally, I would like to express my deepest love and appreciation to my daughter
Ileana Klein, who demonstrated perseverance and resilience in face of unfavorable
prognosis and the challenging circumstance of her birth.
vi
TABLE OF CONTENTS page
ACKNOWLEDGMENTS ................................................................................................. iv
LIST OF TABLES............................................................................................................. ix
LIST OF FIGURES .............................................................................................................x
ABSTRACT..................................................................................................................... xiii
CHAPTER
1 COMPUTATIONAL ASPECTS IN AEROACOUSTICS ..........................................1
Introduction...................................................................................................................1 Definition of Sound ......................................................................................................3
The Characteristics of Sound.................................................................................3 Viscous Effect in Sound Wave..............................................................................5
Classification of Aeroacoustic Problems......................................................................6 Linear Problems in Aeroacoustics.........................................................................6 Nonlinear Problems in Aeroacoustics ...................................................................9
Lightill’s acoustic analogy ...........................................................................10 Ffowcs Williams – Hawkings equation .......................................................14 The merits of Lighthill and Ffowcs Williams – Hawkings analogies..........16 Nonlinear problem: Shock wave formation .................................................17
Computational Techniques for Aeroacoustics............................................................19 Direct Numerical Simulation (DNS) ...................................................................19 Perturbation Technique .......................................................................................20 Linearized Euler Equation...................................................................................21
Computational Issues..................................................................................................21 The Numerical Approach to Reduce Dissipation and Dispersion.......................22 Complex Geometry .............................................................................................24
Scope...........................................................................................................................27
2 ASSESMENT OF DISPERSION-RELATION-PRESERVING AND SPACE-TIME CE/SE SCHEMES FOR WAVE EQUATIONS..............................................30
Introduction.................................................................................................................30 The Dispersion-Relation Preservation (DRP) Scheme...............................................32
Discretization in Space ........................................................................................32
vii
Time Discretization .............................................................................................37 The Space-Time Conservation Element and Solution Element Method ....................41
a–µ Scheme .........................................................................................................41 a-ε Scheme ..........................................................................................................46
Numerical Assessment of the DRP and Space-Time Schemes ..................................49 Short Wave: b/∆x = 3 ..........................................................................................57 Intermediate Wave: b/∆x = 6..............................................................................58 Long Wave: b/∆x = 20.........................................................................................58
Summary and Conclusions .........................................................................................62
3 FINITE VOLUME TREATMENT OF DISPERSION-RELATION-PRESERVING AND OPTIMIZED PREFACTORED COMPACT SCHEMES FOR WAVE PROPAGAION.....................................................................................63
Numerical Schemes ....................................................................................................65 DRP Scheme........................................................................................................66
Finite volume-based DRP scheme (DRP-fv) ...............................................66 Boundary treatment of the DRP scheme ......................................................69
OPC Scheme........................................................................................................70 Finite-difference-based optimized prefactored compact (OPC-fd) scheme.70 Finite volume-based OPC scheme (OPC-fv) ...............................................71 The boundary treatment of the OPC scheme ...............................................71
Time Discretization – The Low Dispersion and Dissipation Runge-Kutta (LDDRK) Method............................................................................................72
Analytical Assessment of DRP and OPC Schemes....................................................77 Operation Count ..................................................................................................77 Dispersion Characteristics ...................................................................................78 Stability................................................................................................................79
Computational Assessment of the DRP and OPC Schemes.......................................82 Test problem 1: One-Dimensional Linear Wave Equation .................................82 Test problem 2: One-Dimensional Nonlinear Wave Equation............................85 Test problem 3: One-Dimensional Nonlinear Burgers Equation ........................89 Test problem 4: Two-Dimensional Acoustic Scattering Problem.......................91
Summary and Conclusions .........................................................................................94
4 A FINITE VOLUME-BASED HIGH ORDER CARTESIAN CUT-CELL METHOD FOR COMPUTATIONAL AEROACOUSTICS.....................................98
Introduction.................................................................................................................98 Cut-Cell Procedure .....................................................................................................99 Test Cases .................................................................................................................104
Radiation from a Baffled Piston ........................................................................104 General description ....................................................................................104 Directive factor D.......................................................................................107 Pressure on the face of the piston...............................................................109 Low frequency (ka = 2) ..............................................................................110 High frequency (ka = 7.5) ..........................................................................111
viii
Reflection of a Pulse on an Oblique Wall .........................................................112 Wave Generated by a Baffled Piston and Reflected on an Oblique Wall .........115
Conclusion ................................................................................................................117
5 SUMMARY AND FUTURE WORK ......................................................................119
Assessment of DRP and Space-Time CE/SE Scheme..............................................121 Finite-Volume Treatment of Dispersion-Relation-Preserving and Optimized
Prefactored Compact Schemes ............................................................................122 Cartesian Grid, Cut-Cell Approach for Complex Boundary Treatment...................123 Future Work..............................................................................................................124
LIST OF REFERENCES.................................................................................................125
BIOGRAPHICAL SKETCH ...........................................................................................134
ix
LIST OF TABLES
Table page 2-1 The stencil coefficient for N = 3...............................................................................36
3-1 A summary of proposed CAA algorithms................................................................96
3-2 The computational cost for DRP and OPC schemes................................................97
4-1 Published cut-cell approach for different problems ...............................................118
x
LIST OF FIGURES
Figure page 1-1 Sound propagation away from a source .....................................................................3
1-2 Moving surface - Ffowcs Williams and Hawkings equation ...................................15
1-3 Schematic diagram showing the (2n + 1) point stencil on a nonuniform grid ...........................................................................................................................24
1-4 Cartesian boundary treatment of curved wall surface; b) detail around boundary...................................................................................................................27
2-1 xα∆ versus α∆x for the optimized DRP 4th order scheme, 7 point stencil, standard 6th order central scheme and 4th order central scheme ................35
2-2 Dispersive characteristics of DRP scheme...............................................................36
2-3 Scheme of the solution elements (SEs) and conservation elements (CEs) ..............43
2-4 Comparison between analytical and numerical solutions Effect of ε on the accuracy of space-time a-ε scheme ....................................................................52
2-5 The dependence of the error on ε for the space-time a-ε scheme at t = 200: a) b/∆x = 3; b) b/∆x = 6; c) b/∆x = 20.........................................................53
2-6 The dependence of the error as function of ν for short (b/∆x = 3), intermediate (b/∆x = 6) and long (b/∆x = 20) waves................................................55
2-7 Effect of ν on the accuracy of space time a-ε scheme: b/∆x = 3, t = 200 ...............56
2-8 The behavior of the error in function of the wavelength: comparison between DRP and space-time a-ε schemes ..............................................................57
2-9 Accumulation of the error in time for short wave - b/∆x = 3 ...................................59
2-10 Accumulation of the error in time for intermediate wave - b/∆x = 6 .......................60
2-11 Accumulation of the error in time for long wave - b/∆x = 20..................................61
3-1 Grid points cluster for one-dimensional problem ....................................................67
xi
3-2 Grid notation for two-dimensional problem.............................................................69
3-3 Four-six–stage optimized Runge-Kutta of order four scheme: a) dissipation error; b) phase error ...........................................................................74
3-4 Dispersive characteristics of the schemes ................................................................80
3-5 Phase speed error on a logarithmic scale .................................................................80
3-6 Errors with respect to the time step size under a fixed space ∆x, at t = 50 - linear wave equation ..............................................................................................84
3-7 Errors under a fixed CFL = 0.5, at t = 50 - linear wave equation: ...........................85
3-8 Errors with respect to the space step size under a fixed CFL = 0.5, at t = 5; nonlinear wave equation .................................................................................87
3-9 DRP–fd solution - nonlinear wave equation; t = 5; CFL = 0.5 ................................87
3-10 DRP–fv solution - nonlinear wave equation; t = 3; CFL = 0.5 ................................88
3-11 OPC-fd solution - nonlinear wave equation; t = 5; CFL = 0.5................................88
3-12 OPC-fv solution - nonlinear wave equation; t = 5; CFL = 0.5................................88
3-13 Error as a function of Pe - nonlinear Burgers equation............................................90
3-14 Numerical solution obtained by DRP schemes nonlinear Burgers quation......................................................................................................................91
3-15 Numerical solution obtained by OPC schemes nonlinear Burgers equation ....................................................................................................................91
3-16 Instantaneous pressure contours at time t = 7; two-dimensional acoustic scattering problem ....................................................................................................93
3-17 The pressure history at point A, B and C .................................................................93
4-1 Illustration of the interfacial cells and cut-and-absorption procedures ..................100
4-2 Modified cut – cell approach for CAA..................................................................101
4-3 Radiation from a baffled piston (test problem (ii)) ...............................................105
4-4 Radiation from a baffled piston: Beam patterns |D(θ)|; ka = 2 .............................108
4-5 Radiation from a baffled piston: Beam patterns |D(θ)|; ka = 7.5 ..........................108
4-6 Radiation from a baffled piston: Beam patterns |D(θ)|; ka = 12.5 ........................108
xii
4-7 Radiation from a baffled piston: Real of piston radiation impedance....................109
4-8 Radiation from a baffled piston: Numerical solution on axis (ka = 2) ..................110
4-9 Radiation from a baffled piston: Comparison between analytical and computed solutions on axis (ka = 2).......................................................................111
4-10 Radiation from a baffled piston: Contour plot of pressure (ka = 2)......................111
4-11 Radiation from a baffled piston: Comparison between analytical and computed solutions on axis (ka = 7.5) ...................................................................112
4-12 Radiation from a baffled piston: Contour plot of pressure (ka = 7.5) ...................112
4-13 Reflection of the pulse on an oblique wall (test problem (ii)): general description ..............................................................................................................113
4-14 Reflection of the pulse on an oblique wall (test problem (ii)): cell around boundary.....................................................................................................114
4-15 Reflection of the pulse on an oblique wall: history of pressure for different angle of wall ............................................................................................115
4-16 Reflection of the pulse on an oblique wall: α = 630 ...............................................115
4-17 Wave generated by a baffled piston and reflects on an oblique wall: General description of the domain .........................................................................116
4-18 Wave generated by a baffled piston and reflects on an oblique wall: α = 630; a) t = 9; b) t = 14; .....................................................................................116
xiii
Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
A FINITE VOLUME, CARTESIAN GRID METHOD FOR COMPUTATIONAL AEROACOUSTICS
By
Mihaela Popescu
August 2005
Chair: Wei Shyy Major Department: Mechanical and Aerospace Engineering
Computational Aeroacoustics (CAA) combines the disciplines from both
aeroacoustics and computational fluid dynamics and deals with the sound generation and
propagation in association with the dynamics of the fluid flow, and its interaction with
the geometry of the surrounding structures. To conduct such computations, it is essential
that the numerical techniques for acoustic problems contain low dissipation and
dispersion error for a wide range of length and time scales, can satisfy the nonlinear
conservation laws, and are capable of dealing with geometric variations.
In this dissertation, we first investigate two promising numerical methods for
treating convective transport: the dispersion-relation-preservation (DRP) scheme,
proposed by Tam and Webb, and the space-time a-ε method, developed by Chang.
Between them, it seems that for long waves, errors grow slower with the space-time a-ε
scheme, while for short waves, often critical for acoustics computations, errors
accumulate slower with the DRP scheme. Based on these findings, two optimized
xiv
numerical schemes, the dispersion-relation-preserving (DRP) scheme and the optimized
prefactored compact (OPC) scheme, originally developed using the finite difference
approach, are recast into the finite volume form so that nonlinear physics can be better
handled. Finally, the Cartesian grid, cut-cell method is combined with the high-order
finite-volume schemes to offer additional capabilities of handling complex geometry. The
resulting approach is assessed against several well identified test problems,
demonstrating that it can offer accurate and effective treatment to some important and
challenging aspects of acoustic problems
1
CHAPTER 1 COMPUTATIONAL ASPECTS IN AEROACOUSTICS
Introduction
Aeroacoustics deals with the sound generated by aerodynamics forces or motions
originating in a flow rather than by externally applied forces or motions of classical
acoustics. Thus, the sounds generated by vibrating violin strings and loudspeakers fall
into the category of classical acoustics, whereas the sound generated by turbulent flow or
the unsteady aerodynamic forces on propellers falls into the domain of aeroacoustics.
A main feature of aeroacoustics is the large difference between the energy levels of
the unsteady flow fluctuation and the sound. This is true even for a very loud noise. For
example, in the near-acoustic field of a supersonic jet (at about 10 jet-diameters away)
the acoustic disturbance amplitudes are about three orders of magnitude smaller than flow
disturbance (Lele, 1997). For supersonic jets one percent of the noise is from the
mechanical power of the jet. In many other cases, the efficiency can be much smaller;
thus the amplitude of disturbance is much less.
Another issue in acoustics is the difference between the scales of the unsteady flow
and sound. This phenomenon is evident in situations in which flow speed is much less
than the sound speed that characterizes the medium. This is because the time scale of the
flow and the sound must match. In low Mach number flow (M « 1) this will give an
acoustic wavelength that is M-1 times longer than the flow length scale. In this case, a
direct numerical simulation of the dynamic flow and generation of sound field will not be
possible.
2
An important challenge in computational aeroacoustics (CAA) lies in the coupling
of time and space that appears in the acoustic wave. An acoustic wave has a wavelength
λ in space, as well as a frequency ω = 2πf in time. These are coupled by the relation
λf = c, where c is the speed of sound in the medium (assumed quiescent). Thus, to
determine an acoustic wave, one must resolve both its wavelength and its frequency.
Numerical spatial and temporal approximation meets the same problem: the dissipation
must be reduced in space, like that in time. However, the CFD community has done much
work to overcome these difficulties, and high accuracy numerical schemes have been
developed. CAA schemes try to minimize dissipation and dispersion error in both space
and time and still maintain a certain order of accuracy.
A further problem lies in the effect of the grid on the solution. In CFD, grids are
often stretched to provide high resolution in regions of high gradients with lesser
resolution where gradients are less steep. However, it is very difficult to propagate an
accurate wave in this kind of grid. The dissipation and dispersion characteristics depend
upon the Courant-Friederich-Lewy (CFL) number (= c∆t/∆x in one dimension), which
can be interpreted as the distance an acoustic wave travels in one time step. If CFL is
changing because the grid has been stretched, strange phenomena can occur.
Vichnevetsky (1987) showed that if a wave is propagated over an expanding grid, the
wave could actually appear to change frequency and be reflected such that it starts
propagating back in the other direction. Anisotropy can appear because of the different
sized grid spacing along different directions. An approach to solving the problem was
suggested by Goodrich (1995), who recommended approximating the solution of the
3
governing partial differential equations rather than approximating the derivatives and
solving the resulting numerical scheme.
Boundary conditions for CAA are also problematic. Often one is interested in
solving the problem in an unbounded space. But for such computation, boundary
conditions have to be imposed; thus the computational domain is finite. The first problem
is the selection of the boundary condition to be imposed. Clearly, the proper boundary
condition depends upon what is outside the computational domain: e.g., source, reflection
boundary and free space. The second problem is in the implementation of these
conditions.
Definition of Sound
The Characteristics of Sound
A wave is the movement of a disturbance or piece of information from one point to
another in a medium (excluding electromagnetic waves, as they do not require a
medium). Sound is a wave that moves parallel to the direction of the propagation
Figure 1-1. Sound propagation away from a source
As a sound wave propagates, it disturbs the fluid from its mean state. These
disturbances are almost always small. We will consider departures from a state in which
the fluid is at rest with a uniform pressure p0 and density ρ0. When this is perturbed by a
sound wave the pressure at position xr and time t changes to p0 + p’ ( xr , t), the density
source
4
changes to ρ0 + ρ’( xr , t), and the fluid particle moves with a velocity ( ),v x tr r . The ratios
0'/p p and 0'/ρ ρ are usually much less than unity.
To illustrate the magnitude of the disturbance in an acoustic wave we will consider
pressure waves in air which can be detected by the human ear. The pressure waves are
referred to as sound. For harmonic pressure fluctuation (a pure tone), the typical range of
frequency that our ear can detected sound is
20 Hz ≤ f ≤ 20 kHz (1.1)
The range described by (1.1) is called the audible range. The maximum sensitivity
of the ear is within the frequency range of 2 kHz to 3 kHz (policeman’s whistle tone) The
acoustic pressure of intense rocket engine noise can be as much as 9 orders of magnitude
greater than the pressure of the weakest sound detectable by human ear. A logarithm
scale was necessary to be able to comprehend this large range: 9 decade in amplitude (18
decades in intensity) (Blackstock, 2000). The logarithm devised is called a level, e.g.,
sound pressure level and intensity level. Although the levels are unitless, they are
expressed in decibels (dB).
SPL = 20 log10(p’/pref) (1.2)
On this scale a fluctuation of one atmosphere in pressure corresponds to 194 dB.
The threshold of pain is between 130 and 140 corresponding to a pressure variation of
only one thousandth of an atmosphere, i.e. p’/p0 = 10-3.
At the threshold of pain, fluid particles in a 1 kHz wave vibrate with a velocity of
about 0.1 m/s, which is only about 1/104 of the speed at which the sound wave is
traveling. The displacement amplitude of the particles is thus between 10-4 and 10-5
5
meters and the wavelength is about one-third of a meter, so the displacement amplitude is
only about 10-4 of the wavelength sound (Dowling and Ffows Williams, 1983).
In the case where sounds can be approximated by small perturbations, several
effects are noteworthy. First, there is no interaction between different acoustic waves; the
sound fields add linearly. Hence, for the study of sound that comes from more than one
source, we can study the sound from each source separately, and then add the final
solutions. Second, the flow variables satisfy the linearized equation of the fluid motion,
so each flow variable is linearly related to every other variable. This leads to a great
simplification in the mathematics. Knowledge of one variable, e.g., pressure, provides a
basis for a simple evaluation of all the other variables such as the density or particle
velocity. Third, if we wish to solve the behavior of the sound numerically, we have to use
a scheme that adds a low level of numerical dissipation.
Viscous Effect in Sound Wave
The effect of Reynolds number on CAA is ambiguous. Although all sound is
ultimately dissipated into heat by viscosity, acoustics is generally considered to be
inviscid fluid phenomena. If the viscous term is considered in the standard linear
analysis, one finds its value to be
0Re 2 /cπ λρ η= (1.3)
where c is the speed of sound, λ is the wavelength, ρ0 is the density, and η is the
coefficient of viscosity. The values in air of these parameters for most practical interests
are (Blackstock, 2000)
6
0
5
331.60.331.293
1.7 10
cλρ
η −
=≅≅
≅ ⋅
(1.4)
Based on these values, Re is approximately 4.8 ×107 at the most audible frequency.
This high value of Reynolds number determines that viscous effects are negligible in a
sound field, because the pressure represents a far greater stress field than that induced by
viscosity. Hence, we will regard sound as being essentially a weak motion of an inviscid
fluid. The effect of viscosity can be taken into consideration after the sound travels about
2πλρ0/η wavelengths (i.e., after ≅ 1.5x108 m). Thus, dissipative loss becomes important
for high frequency sound propagation over long distances.
On the other hand, if one considers the generation of sound by flows rather than
propagation of sound through flows, then a significant Reynolds number effect can be
observed. For example, the most common source of sound in flows is the acceleration of
the vorticity (Powell, 1964), which is only present in viscous flow. Even there, a curious
independence is observed that is apparently due to the fact that the large-scale; hence
efficient sound generation structures in the flow change little with Reynolds.
Classification of Aeroacoustic Problems
Diverse problems of aeroacoustics can be classified on the basis of the physical
phenomena that are expected to play a central role. The main classification is between the
linear and nonlinear problem.
Linear Problems in Aeroacoustics
A major strength offered by computational approaches is the generality to deal with
linear interaction problems of aeroacoustics (where one type of physical disturbance is
7
scattered into another type of disturbance). A linearized solution of the unsteady Euler
equations subject to appropriate boundary condition can provide useful prediction of the
noise generation. The problems of linear interaction are not limited to sound generation
but also contain the “reciprocal” problem of the generation (or receptivity) of vertical
disturbance due to the incident sound (Colonius and Lele, 2004).
This category includes the classical boundary value problems of linear acoustics:
sound propagation in a uniform media in presence of reflecting surfaces, barriers,
absorbing walls, and ducts, and also propagation and scattering of sound in a prescribed
non-uniform medium. Specific examples include as: radiation from a duct opening or
engine-inlet due to a specified source or a specified duct-mode disturbance, radiation
from a specified duct-mode disturbance, and radiation from a specified source across a
finite barrier/sound wall with an absorbing treatment.
Sound propagation in a specified non-uniform time dependent medium including
refraction/scattering in steady and unsteady vortical flows, sound propagation in non-
uniform ducts including the interaction with geometrical changes, linear impedance and
mean-flow variations are considered as linear problems of scattering.
The equation of a linear acoustics wave is deduced for a homogeneous fluid
characterized by ρ = ρ0, P = p0, and 0 'u u u= +r r r . The sound waves minutely disturb the
status of the quiescent fluid
0 0
20 0 0
0
,
,
0 ', '
P p p p c
u u u c
ρ ρ δρ δρ ρ
ρ
= + <<
= + <<
= + <<r r r
(1.5)
8
To obtain the wave equation, we will start with the Euler equation and the equation
of state for an isentropic process
( )0
0t
t
u u
u u u P
ρ ρ ρ
ρ
+ ⋅∇ + ∇ ⋅ =
⎡ ⎤+ ⋅∇ ⋅ + ∇ =⎣ ⎦
r r
ur r r (1.6)
and
0p c δρ= (1.7)
Substituting Eq.(1.5) into (1.6), the following is obtained
( ) ( )0
0 0
0
0
t
t t
u u u
u u u u u u p
δρ δρ ρ δρ
ρ ρ δρ δρ
+ ⋅∇ + ∇ ⋅ + ∇ ⋅ =
+ ⋅∇ + + ⋅∇ + ∇ =
r
ur r r r r (1.8)
The first order terms are small because the perturbation is very small; hence,
second order terms or higher can be neglected. The underlined terms are second order or
higher. The result is
0
020
0
0t
t
u
u p
p c
δρ ρ
ρ
δρ
+ ∇ ⋅ =
+ ∇ =
=
r
ur (1.9)
To reduce the set to a single equation in ur , the equation of state Eq.(1.7) first is
used to eliminate p∇ from the second expression in Eq. (1.9)
( )20 0 0tu cρ δρ+ ∇ =r (1.10)
If the time derivative of this equation is subtracted from the divergence of the first
Eq. (1.9), the result is the classical wave equation
2 20 0ttu c u− ∇ =
s r (1.11)
9
Nonlinear Problems in Aeroacoustics
Nonlinear aeroacoustics is the field of acoustics devoted to very intense sound,
specifically to waves with amplitude high enough that the small-signal assumption is
violated. Retaining the nonlinear term makes the conservation equation much harder to
solve. This includes problems of nonlinear propagation, such as nonlinear steeping and
decay, focusing, viscous effects in intense sound beam (acoustics near field of high-speed
jets), and sonic boom propagation through atmospheric turbulence, sound propagation in
complex fluids and multi-phase systems (such as in lithotripsy), and internal flows of
thermo-acoustic cooling system. Also included in this category are problems of scattering
of nonlinear disturbances into sound, such as airframe noise.
The basic equations that describe the nonlinear behavior are the same that describe
the general motion of a viscous, heat conducting fluid: mass conservation, momentum
conservation, entropy balance, and thermodynamic state.
The mass conservation or continuity equation is
0D uDt
ρ ρ+ ∇ ⋅ =r (1.12)
where ρ is the mass density, ur is the fluid velocity vector, and D/Dt = ∂/∂t + ur ·∇ is the
total, or material, time derivative.
The momentum equation may be written as
( )2 13B
Du P u uDt
ρ µ µ µ⎛ ⎞+ ∇ = ∇ + + ∇ ∇ ⋅⎜ ⎟⎝ ⎠
rr r (1.13)
where P is the pressure, µ is the shear viscosity, and µB is the bulk viscosity. Shear
viscosity accounts for diffusion of momentum between adjacent fluid elements having
different velocity. Bulk viscosity provides an approximation description, valid at low
10
frequency, of nonechilibrum deviation between actual local pressure and thermodynamic
pressure.
Equation of state is given by
( )
( )
,
,
P P TorP P s
ρ
ρ
=
=
(1.14)
where T is the absolute temperature, and s is the specific entropy.
A commonly used equation of state of a perfect gas is
( )00 0
exp / v
P s s cP
γρρ
⎛ ⎞= −⎡ ⎤⎜ ⎟ ⎣ ⎦
⎝ ⎠ (1.15)
where P0, ρ0, and s0 are reference values of the pressure, density, and specific entropy.
γ = cP/cv is the ratio of the specific heats at heat pressure (cp) and constant volume (cv). In
case of arbitrary fluid the equation of state is obtained by expanding of Eq.(1.14) in a
Taylor series about (ρ0, s0).
A more general description take implies relaxation, like vibrational relation of
diatomic molecules (as in air) and chemical relation in seawater (Hamilton and
Blackstock, 1997). The former occurs when the energy associated with molecular
vibration fails to keep in step with molecular translation energy associated with the
fluctuating temperature in gas.
Lightill’s acoustic analogy
The study of flow that generates acoustic waves began with Gutin’s theory (1948)
of propeller noise, which was developed in 1937. However, the theory could not be
considered until 1952, when Lighthill (1952, 1954) introduced his acoustic analogy to
11
deal with the problem of calculating acoustic radiation from a relatively small region of
turbulent flow embedded in an infinite homogeneous fluid.
It is well known that a gas (not monoatomic) allows three distinct fundamental
modes of oscillation: the vortical and entropy modes, both of which are convective, and
the acoustic mode, which is the solution of the wave equation (Chu and Kovasznay,
1958). The three modes may exist independently only if i) the oscillations are small, and
ii) the base flow is the uniform medium at rest or is in uniform motion. If the flow is
inhomogeneous, then they are coupled, and it is not easy to separate them.
To overcome these limitations, an approximation is necessary. In this section, the
acoustic approach introduced by Lighthill (1952, 1954) is presented. This approach is
used to calculate acoustic radiation from relatively small regions of turbulent flow
embedded in an infinite homogeneous fluid where the speed of the sound c0 and density
ρ0 is constant. Density and pressure fluctuations are defined by
0
0
''p p p
ρ ρ ρ= −⎧⎨ = −⎩
(1.16)
where the subscript ‘0’ is used to denote constant reference values that are usually taken
to be corresponding properties at large distance from the flow.
Lighthill’s basic idea was to reformulate the general equations of gas dynamics in
order to derive a wave equation suitable to describe sound propagation.
The continuity and momentum equation are
0jj
ijij i
j i j
ut x
u pu ut x x x
ρ ρ
σρ ρ
∂ ∂+ =
∂ ∂
⎛ ⎞ ∂∂ ∂ ∂+ = − +⎜ ⎟⎜ ⎟∂ ∂ ∂ ∂⎝ ⎠
(1.17)
12
where σij is the component of the viscous stress tensor. For a Stokestian gas it can be
expressed in terms of the velocity gradient
23
ji kij ij
j i k
uu ux x x
σ µ δ⎛ ⎞∂∂ ∂
= + −⎜ ⎟⎜ ⎟∂ ∂ ∂⎝ ⎠ (1.18)
where µ is the viscosity of the fluid. Multiplying the continuity equation by ui, adding the
result to the momentum equation, and combining terms yields
( )i i j ij ijj
u u u pt x
ρ ρ δ σ∂ ∂= − + −
∂ ∂ (1.19)
which upon adding and subtracting the term c02∂ρ/∂xi, becomes
20
iji
i i
Tu ct x x
ρ ρ ∂∂ ∂+ = −
∂ ∂ ∂ (1.20)
where
( ) ( )20 0 0ij i j ij ijT u u p p cρ δ ρ ρ σ⎡ ⎤= + − − − −⎣ ⎦ (1.21)
By differentiating the continuity equation with respect with t, and subtracting the
divergence of Eq.(1.20), Lighthill’s equation is obtained
22
2 202
' ' ij
i j
Tc
t x xρ ρ
∂∂− ∇ =
∂ ∂ ∂ (1.22)
Equation (1.22) has the same form as the wave equation that governs the acoustic
field produced by a quadrupole source ∂2Tij/∂xi∂xj in a nonmoving medium (Goldstein,
1976). Hence, there is an exact analogy between the density fluctuation that occurs in any
real flow and the small amplitude density fluctuations that would result from a
quadrupole source distribution (of strength Tij) in a fictitious (nonmoving) acoustic
medium with sound speed c0. This source will vanish in the region outside certain types
13
of turbulent flows such that Eq. (1.22) reduces to the homogeneous wave equation in
such regions.
In the case where the acoustics generated are not due to a jet with high temperature
and the flow is isentropic, the second term of Lighthill’s turbulence stress in Eq.(1.21)
can be neglected.
For a low Mach number, the third term in Eq.(1.21) can be neglected. It can be
neglected because viscosity and heat conduction cause a slow damping due to conversion
of acoustic energy into heat and have a significant effect only over large distances.
We have therefore shown that Tij is approximately equal to ρuiuj inside the flow
and is equal to zero outside this region. Assuming that the density fluctuation is
negligible within the moving fluid, then
Tij ≅ ρ0uiuj (1.23)
The Lighthill approximation has a great advantage in that it is possible to solve the
equation with a standard Green function. But, this approximation is only for isentropic,
low Mach number flows.
Lighthill’s equation could be used to study the sound generated by unsteady flows
where there are no solid boundaries (or more correctly, by flows where the effect of
boundaries can be neglected). Another limitation of this theory is that the principle of
sound generated aerodynamically as stated by Lighthill is relevant only when there is no
back reaction of the acoustic waves in flow, such as at the trailing edge or in initial shear
layer ( Hirschberg, 1994). In this case, the conversion of mechanical energy into acoustic
energy is only one way, and this is the reason why acoustics can be inferred from an
incompressible description of the flow.
14
Ffowcs Williams – Hawkings equation
Let us consider a finite volume of space containing a disturbed flow and rigid
bodies in an arbitrary motion with the surrounding fluid being at rest. The bodies and the
flow generate a sound. In this case, it is possible to replace both the flow and the surface
by an equivalent acoustic source, assuming that the whole medium is perfectly at rest.
The key assumption is that no flow-acoustics coupling occurs, i.e., the acoustic
field does not affect the flow from which the sound originates. Consequently, this
approach is not valid when some resonant conditions induce an acoustic feedback on the
flow. To represent the real medium with the flow and the obstacle in a convenient way,
Ffowcs Williams and Hawkings (1969), and Ffowcs Williams (1969,1992) defined an
equivalent medium where the rigid bodies are replaced by mathematical surfaces. In
order to preserve the kinematics of the flow and the boundary condition of no cross-flow
on the surface, a discontinuity must be imposed at the surface location by introducing
mass and momentum sources.
The mass and momentum equations are written as
( )
( ) ( ) ( )
0
' '
j sij i
i i j ij ijj j
fu u ft x x
fu u u ft x x
ρ ρ ρ δ
ρ ρ σ σ δ
∂ ∂ ∂+ =
∂ ∂ ∂
⎛ ⎞∂ ∂ ∂+ − = −⎜ ⎟⎜ ⎟∂ ∂ ∂⎝ ⎠
(1.24)
In Equation (1.24), ρ0 is the mean density, usi is the velocity field of a point on the
surface, δ is the Dirac delta function, σ’ij (= σij – (P – P0)δij) is the viscous stress tensor
(P being the static pressure with mean value P0) and f( xr ,t) = 0 defines the kinematics of
the surfaces. If the normal unit vector on the surface is nr , then the boundary condition of
no cross-flow is simply
15
su n u n⋅ = ⋅r r r r (1.25)
Figure 1-2. Moving surface - Ffowcs Williams and Hawkings equation
Following the procedure used to obtain Lighthill’s equation, we can derive the
equation for density variation ρ’ = ρ - ρ0
( ) ( )22 2
20 02 2
' ' 'ijij si
i i j i j i
T f fc f u ft x x x x x t xρ ρ σ δ ρ δ
⎛ ⎞∂ ⎛ ⎞∂ ∂ ∂ ∂ ∂ ∂− = + +⎜ ⎟ ⎜ ⎟⎜ ⎟∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂⎝ ⎠⎝ ⎠
(1.26)
The values of ρ’ and Tij are zero inside the mathematical surface, because in these
zones the fluid is considered to be at rest.
Roger (1995) showed that in the presence of flow and rigid bodies, fluctuations in
the fluid are exactly the same as those that would exist in an equivalent acoustic medium
at rest, and they are forced by three source distribution:
• a volume distribution ∂2Tij /∂xi∂xj in the outer region of the surface due to the flow
• a surface distribution ∂/∂xi (σ’ij δ(f) ∂f/∂xj ) due to the interaction of the flow with the moving bodies
• a surface distribution ∂/∂t ( ρ0 usiδ(f) ∂f/ ∂xi ) due to the kinematics of the bodies
Like in the Lighthill analogy, the analogy of Ffowcs Williams and Hawkins is
limited to flow where there is no flow acoustic coupling, or where the acoustic field does
not affect the flow from which the sound originates.
Vi(t) Fluid at rest S(t) Surface with
rigid-body motion
Ve Real flow field
sur
nr
16
The merits of Lighthill and Ffowcs Williams – Hawkings analogies
The Lighthill analogy marked an important milestone in the study of flow-
generated acoustics. His theory is able to explain a number of important characteristics of
acoustic radiation from a jet. The solution of the Lighthill analogy could be obtained
analytically using the Green function.
The approach was developed when computation capability was limited. That is
why this analogy is a big step for understanding aeroacoustic propagation. For example,
his theory gave a mathematical representation of sound and pseudosound, an
approximation of the sound emission from subsonic cold-air jet flow (which he called
self-noise, i.e., noise generated by turbulent-turbulent interaction, and shear noise, i.e.,
noise generated by turbulent mean shear), for the first time.
Lighthill’s theory represents the basis for the Ffowcs Williams and Hawkings
analogy, which accounts for sound generation along a boundary, as in a helicopter rotor,
an airplane propeller, an aircraft engine fan, a compressor, or a turbine. The solution of
the second analogy is again solved with the Green’s function. In this, it is difficult to
obtain a general solution, because the analytical solution is available only for certain
boundary shapes.
The theory of Ffowcs Williams and Hawkings appeared when the performance of
the computers increased but was still substantially less than what a computer is capable of
today. Hence, this analogy was very useful for studying sound emission from sources
such as a thin strut in a turbulent flow, propeller noise (Gutin, 1948) and sound
generation near a plate.
The advent of powerful computers has made it easier to study more complex
problems using these analogies since no analytical solutions are required.
17
Nonlinear problem: Shock wave formation
An example where a linear approximation is not suitable is subsonic flow in a pipe,
because waves do not attenuate fast with propagation. In this case a wave can propagate
as much as 102 wavelengths before friction has a significant effect. This implies that even
a nonlinearity of order 10-2 in the wave propagation can have a significant effect
(Hirschberg and Rienstra, 1994). The most spectacular nonlinear effect is the formation
of shock waves as a result of the steepening of the compression side of a wave.
The study of the equation of a shock wave is deduced for a homentropic fluid,
namely, a fluid with constant entropy (Hamilton, 1998). The characteristic form of the
equation is obtained from mass and momentum conservation:
( ) 0dpu c ut x cρ
⎛ ⎞∂ ∂⎡ ⎤+ ± ± =⎜ ⎟⎢ ⎥∂ ∂⎣ ⎦ ⎝ ⎠∫ (1.27)
which implies that
dpJ ucρ
± = ± ∫
(1.28)
is invariant along the characteristic path C± in the (x, t) plane which are described by the
equation
C
dx u cdt ±
⎛ ⎞ = ±⎜ ⎟⎝ ⎠
(1.29)
When a C+ wave propagates into a uniform region, the C- wave emanating from the
uniform region will all carry the same information: J- = J0 is constant. Typically for a
simple wave, the characteristic C+ are simple lines in the (x, t) plane. Eq.(1.28) and
Eq.(1.29) yield
u = 1/2 (J+ + J0) (1.30)
18
and
( )012
dp J Jcρ
+= −∫ (1.31)
In other words u and ∫dp/ρc are constant along C+, the thermodynamic variable. Hence,
speed of sound can be considered constant, so that (u + c) of C+ is constant. For an ideal
gas there is the relation
21
dp ccρ γ
=−∫ (1.32)
A shock wave appears at the point where two characteristics lines of the same
family intersect. The intersection of two C+ occurs after traveling time ts:
( )1
0s
d u ct
dx
−⎛ ⎞+⎡ ⎤
= −⎜ ⎟⎢ ⎥⎜ ⎟⎣ ⎦⎝ ⎠ (1.33)
which for an ideal gas with constant γ can be expressed in terms of the space derivative
of pressure at x = 0:
( )
0
0
1s
ptpcx
γ
γ= −
∂⎛ ⎞− ⎜ ⎟∂⎝ ⎠
(1.34)
Because the variation of the speed of sound is negligible, the distance of
appearance of the shock wave is given by
( )
00
0
1s s
px c tpx
γ
γ≅ = −
∂⎛ ⎞− ⎜ ⎟∂⎝ ⎠
(1.35)
For harmonic waves
0max
ˆdp k pdx
⎛ ⎞ =⎜ ⎟⎝ ⎠
(1.36)
19
( )
00 ˆ1s
pk xp
γγ
=+
(1.37)
Hence, for an amplitude 0ˆ /p p = 10-2, the shock is expected appear within 10
acoustic waves (Goldstein 1976).
Computational Techniques for Aeroacoustics
Direct Numerical Simulation (DNS)
Direct simulation methods aim to compute both unsteady flow and the sound
generated by it. These methods must use a domain that includes the noise producing flow
region and at least a part of the near-acoustic-field. The computational mesh needs to be
selected so that both the flow and its sound can be well represented. The first issue in this
approach is that the computational cost of such direct computations is large, hence only
simple flow configurations can be studied using this direct method.
Direct simulation of the acoustic field solves the compressible Navier-Stokes
equations (or Euler equations in those cases where viscosity is not important) without
further approximation. These equations govern the total flow field including the
acoustics, so if one could solve them in a domain reaching out to the far field, then the
one can obtain the acoustic radiation emerging from the fluctuating flow. Gloerfelt at al.
(2003) did a comparison of this method with the acoustic analogy. Both methods gave
solutions that agree with the experimental data.
This approach is limited to flows where the viscosity is not important; it encounters
fundamental difficulties when the Reynolds number is high due to the range of scales
present in the flow field. The characteristic frequency of a source that radiates sound is
given by the Strouhal number, St = fL/U ≅ O(1), where L and U are the characteristic
length and velocity. This implies that the wavelength of sound produced is λ≅ L/M,
20
where M is the Mach number of the flow. On the other hand, the dissipation of turbulent
energy takes place at the Kolmogorov length scale η = L/Re3/4 (Tennekes and Lumley
1972), where Re is the Reynolds number of the flow. Thus, the ratio of the wavelength of
sound to the size of the energy dissipating eddies, λ/η≅Re3/4/M, takes high values at low
Mach numbers, even for moderate Reynolds numbers. Accordingly, the requirement for
spatial and temporal resolutions will simply be beyond computational capabilities in the
near future. Thus, although DNS has been utilized quite successfully in low Reynolds
number (Gloerfelt et al., 2003; Hu et al., 2002) (Re < 1000) , a less direct approach where
separate grids can be designed for the viscous and acoustic phenomena appears to be
preferable for high Reynolds number flows.
Perturbation Technique
A perturbation technique consists of splitting up the flow field calculation into two
parts. First, a time-independent viscous background flow is calculated, and then a
perturbation equation (that describes the sound) about this background flow is derived,
and viscous action on the perturbation is neglected. In this approach, an initial
disturbance is introduced upstream which causes an instability and causes the wave to
grow resulting in the radiation of sound. Tam and Morris (1980) developed this approach
analytically for shear layers, and Tam and Burton (1984) developed it for subsonic jets.
Hardin and Pope (1995) developed a slightly different approach to address
vorticity-dominated flow, an expansion about the time-dependent, viscous,
incompressible, subsonic flow. If the density ρ = ρ0 is constant, then the continuity and
momentum equations become a set of four equations for the three incompressible
velocity components and incompressible pressure. One solves this set of equations by
21
using a grid and numerical scheme designed for the viscous aspects of the flow for the
time dependent viscous incompressible flow. Using the solution from the previous time
step, the compressible flow equation is perturbed about the time dependent
incompressible flow. The differences between the compressible and incompressible flows
are then assumed to be inviscid and isentropic.
Linearized Euler Equation
The linearized Euler equation (LEE) comes from the modified Euler equation. The
principal issue appears when we deal with LEE with specific source term and LEE with
projected source terms (Colonius and Lele, 2004). The numerically difficulty of this
approach stems from the fact the full LEE set admit non-trivial instability wave solution
of the homogeneous equation.
Propagation of linear acoustic waves through a medium with known properties, its
refraction and scattering due to the nonuniform of the medium or the base flow, and
scattering from solid surfaces, and scattering from solid surfaces (with prescribed
boundary conditions) are problem for which computational technique are well suited.
Simulation of such phenomena can be based on linearized field equation subject to the
physical boundary conditions; refraction and scattering effects are automatically
obtained. However in such a direct approach the high frequency limit becomes
computationally demanding.
Computational Issues
Numerical methods for problems of sound generation and propagation must
overcome a host difficulties that arise because acoustic waves are very weak compared to
near-field fluctuation, and because they must propagate with little attenuation over long
distances. In practice this has dictated the use of high-order-accurate numerical methods.
22
The Numerical Approach to Reduce Dissipation and Dispersion
In computational aeroacoustics (CAA), accurate prediction of the generation of
sound is demanding due to the requirement of preserving the shape and frequency of
wave propagation and generation. Furthermore, the numerical schemes need to handle
multiple scales, including long and short waves, and nonlinear governing laws arising
from sources such as turbulence, shocks, interaction between fluid flows and elastic
structures, and complex geometries. It is well recognized (Hardin and Hussaini, 1992),
Tam (Tam and Webb, 1993; Tam et al., 1993) that in order to conduct satisfactory CAA,
numerical schemes should induce minimal dispersion and dissipation errors. In general,
higher-order schemes are more suitable for CAA than lower-order schemes since, overall,
the former are less dissipative. That is why higher-order spatial discretization schemes
have gained considerable interest in computational acoustics (Hixon, 1997; Kim et al.,
1997; Lin and Chin, 1995).
For longer wavelengths, the formal order of accuracy is sufficient to indicate the
performance of a scheme. However, for shorter waves relative to the grid size, it is
known that the leading truncation error terms are not good performance indicators (Shyy,
1985; Shyy, 1997). To handle broad band waves, the idea of optimizing the scheme
coefficients by minimizing the truncation error associated with a particular range of wave
numbers has been used over the years by many researchers, e.g., Hu et al. (1996),
Stanescu and Habashi (1998), Shu (1994), Nance et al. (1997), Wang and Sankar(1999),
Cheong and Lee (2001), Wang and Chen (2001), Ashcroft and Zang (2003), and Tang
and Baeder (1999). A successful approach is the Dispersion-Relation-Preserving (DRP)
finite difference scheme proposed by Tam (Tam and Webb, 1993; Tam et al. 1993). The
basic idea in the DRP scheme is to optimize the coefficients to satisfactorily resolve short
23
waves with respect to the computational grid, namely, waves with wavelengths of ∆x
(defined as 6-8 points per wave or PPW) or shorter. It maximizes the accuracy by
matching the wave number and frequency characteristics between the analytical and
numerical operators in the range of resolvable scales. Recently, Ashcroft and Zhang
(Ashcroft and Zang, 2003) have reported a strategy for developing optimized prefactored
compact (OPC) schemes requiring smaller stencil support than DRP. The prefactorization
strategy splits the central implicit schemes into forward and backward biased operators.
Using Fourier analysis, they have shown that it is possible to select the coefficients of the
biased operators such that their dispersion characteristics match those of the original
central compact scheme. Hixon and Turkel (1998) proved that the prefactored scheme is
equivalent to the initial compact scheme if the real components of the forward and
backward operators are equal to those at the corresponding wave number of the original
compact scheme, and the imaginary components of the forward and backward operators
are equal in magnitude and opposite in sign.
Both DRP and OPC schemes are originally designed based on the finite difference
approach. In order to satisfy the governing laws of the fluid physics, it may be
advantageous to adopt the finite volume approach (Udaykumar et al., 1997; Yang and
Ingram, 1999; Udaykumar, 1999), which ensures that fluxes estimated from different
sides of the same surface are identical, i.e., no spurious source/sink is generated due to
numerical treatment. Furthermore, a finite volume formulation can offer an easier
framework to handle the irregular geometry and moving boundaries. In this work, we
investigate a finite volume formulation (which we call DRP-fv), extending the concept
embodied in the original, finite difference-based DRP scheme (which we call DRP–fd).
24
Similarly, for the OPC scheme, we extend the basic concepts of the original, finite
difference-based OPC (OPC-fd) scheme, in a finite volume formulation, named OPC-fv.
Our ultimate goal is to extend the finite volume version of suitable schemes into a cut-
cell type of Cartesian-grid computational technique that we have developed earlier for
moving and complex boundary computations (Yang and Ingram, 1999; Udaykumar et al.,
1999; Kim and Lee, 1996; Ye et al., 1999; Lahur et al., 2000; Calhoun, 2000).
Complex Geometry
To handle problems of practical interest, a CAA scheme needs to have the
capability of handling irregular and curved geometries. It is challenging to develop
methods exhibiting desirable order of accuracy and controlling dispersion and dissipation
errors while being capable of handling complex geometries. In an attempt to address the
need for flexible grid distributions, Cheong and Lee (2001), proposed a grid-optimized
dispersion-relation-preserving scheme (GODRP). They considered the approximation of
the derivative by
( )1 n
j ij n
u a u x x xx x =−
∂≈ + ∆ ⋅ ∆
∂ ∆ ∑ (1.38)
where
( ) / 2n nx x x n−∆ = − (1.39)
( ) / ( , 1,..., 1, )i ix x x x i n n n n∆ = − ∆ = − − + − (1.40)
x0 x1 x2x-1x-2 xn-1 xnx-n x-n+1
( =x )
x0 x1 x2x-1x-2 xn-1 xnx-n x-n+1
( =x ) Figure 1-3. Schematic diagram showing the (2n + 1) point stencil on a nonuniform grid
25
The wavenumber of the scheme over a nonuniform grid is obtained by using a
Fourier transformation:
i
ni x x
jj n
i a ex
αα ∆ ⋅∆
=−
−=
∆ ∑ (1.41)
The scheme is derived by requiring that i) the real part of the scheme closely match
the analytical solution in the chosen range of wavenumbers, and ii) the imaginary part of
the scheme be as close to zero as possible. These requirements can be achieved by
minimizing the integrated error E, defined as
( )
( ) ( )
2
0
2
0
exp ln 2
r
i
e
real
e
imag
E x x d x
xx Sgn c x d x
α α α
αλ α α ασ
= ∆ − ∆ ∆
⎡ ⎤⎛ ⎞∆+ ∆ + − ⋅ − ∆ ∆⎢ ⎥⎜ ⎟
⎢ ⎥⎝ ⎠⎣ ⎦
∫
∫ (1.42)
The terms er and ei denote the upper limits of the integration intervals of the real
and imaginary parts, respectively. The term λ is the weighting factor, σ is the half-width
of a Gaussian function, and c is the speed of wave propagation in ut + cux = 0.
GODRP schemes of curvilinear grids permit an assessment of the accuracy of the
finite difference method for curvilinear meshes from the wave number point of view.
Through the grid-optimization process, high-order finite difference equations can be
solved with curvilinear grids with a guarantee of local and thus resultant global
dispersion-relation-preserving properties. Hence, the approach can be used with success
for a body-fitted grid to study the generation of sound around a body with complex
geometry.
The coefficients are obtained based on local characteristics of the grid; in other
words, for a 2n+1 – point stencil GODRP spatial discretization, the scheme implies a
26
solution of 2n+1 equations for each point in the grid. This will lead to a considerable
increase in the computational cost.
Tam et al. (2000) proposed to solve the problem around a boundary using ghost
points and an optimized extrapolation scheme around the complex boundary. They
considered an extrapolation scheme that is of desirable accuracy over a wavenumber
range –k ≤ α∆x ≤ k. The goal for the extrapolation is that the scheme works well for
general functions. For this purpose, it is sufficient to consider waves with unit amplitudes
over a desired band of wavenumbers
fα(x) = ei[αx + φ(α)] (1.43)
The total effect on the function f(x) will be weighted by the amplitude A(α).
The general formula for extrapolation is given by
( ) ( )1
0 00
,N
j j jj
f x x S f x x x j xη−
=
+ ∆ = = − ∆∑ (1.44)
where Sj (j = 0, 1, 2, …, N-1) are the stencil coefficients. Their values are obtained by
imposing the following: i) the difference between the left and right sides of Eq.(1.44) is
zero when the single Fourier components of Eq.(1.43) is substituted in the formula; ii)the
error is zero if the approximated function is zero. The constrained optimization problem
is handled by the Lagrange multiplier
2
1 1
0 00
1k N N
i y ijyj j
j jL e S e dy Sη λ
− −−
= =
⎛ ⎞= − + −⎜ ⎟
⎝ ⎠∑ ∑∫ (1.45)
The boundary curve is approximated by line segments joining the intersection
points of the computation mesh and the boundary. For instance, the curved surface
between A and B in Figure 1-4 is replaced by a straight line segment. G2 is a ghost point.
A ghost value is assigned to G2 as a boundary condition. The enforcement point is at E
27
and G2E is perpendicular to AB. The value of the derivatives of the function at points A
and B are found via extrapolation from the points at (1’, 2’, 3’, 4’, 5’, 6’, 7’) and (1, 2, 3,
4, 5, 6, 7), respectively.
The method is accurate only for waves with a wavelength of 8 mesh spacings or
longer. The method induces numerical instability when the wavenumber is high. Hence,
we do not expect that the approach gives accurate results for short waves or high
wavenumber.
G1
G4
G3
G2
G5
A
B
1 2 3 7654
1’
2’
3’
7’
6’
5’
4’
G1
G4
G3
G2
G5
A
B
1 2 3 7654
1’
2’
3’
7’
6’
5’
4’
G2
A
B
E
1’
2’
1 2
G2
A
B
E
1’
2’
1 2
a) b) Figure 1-4. Cartesian boundary treatment of curved wall surface; b) detail around
boundary
Scope
The present thesis has three main contributions. First we investigate two numerical
methods for treating convective transport: the dispersion-relation-preservation (DRP)
scheme, proposed by Tam and coworkers (Tam and Webb, 1993; Tam et al., 1993), and
the space-time a-ε method, developed by Chang (1995). The purpose is to examine the
characteristics of existing schemes capable of handling acoustic problems. The space-
28
time a-ε method directly controls the level of dispersion and dissipation via a free
parameter, ε, while the DRP scheme minimizes the error by matching the characteristics
of the wave. Insight into the dispersive and dissipative aspects in each scheme is gained
from analyzing the truncation error. Even though both methods are explicit in time, the
appropriate ranges of the CFL number,ν, are different between them. Different
performance characteristics are observed between the two schemes with regard to long
and short waves.
Second, the two low dispersion numerical schemes developed via optimization
procedures in the wave number domain, namely, the DRP scheme and the optimized
prefactored compact (OPC) scheme developed by Ashcroft and Zhang (2003), are
extended from their original finite difference framework to the finite volume framework.
The purpose is to extend the capability of these schemes to better handle nonlinearity and
conservation laws of the fluid motion. Linear and nonlinear wave equations, with and
without viscous dissipation, have been adopted as the test problems. By highlighting the
principal characteristics of the schemes and utilizing linear and nonlinear wave equations
with different wavelengths as test cases, the performance of these approaches is studied.
Finally, the Cartesian grid, cut-cell method is extended along with the OPC-based
finite volume scheme so that this high order scheme can treat curved geometry associated
with practical acoustic applications. The approach uses a background Cartesian grid for
the majority of the flow domain with special treatment being applied to cells which are
cut by solid bodies, thus retaining a boundary conforming capability. Surface integrals
around complex geometries are computed using flow properties at the cell surface
interpolated from cell nodes while maintaining desirable accuracy level.
29
The rest of the thesis is structured as follows. Chapter 2 presents the investigation
of the Dispersion-Relation-Preserving (DRP) scheme, and the space-time a-ε method.
The characteristics of these two schemes are emphasized using a simple wave equation
with the initial disturbance in the form of the Gaussian function
Chapter 3 presents principal characteristics and introduces two space discretization
schemes: the DRP scheme and the OPC scheme. A low dispersion and low dissipation
Runge-Kutta proposed by Hu and coworkers (1996) is employed for the time stepping
procedure, and combined with the DRP and OPC schemes. A study of the dispersive
characteristics and stability is presented for these schemes. The boundary treatment is
presented in this chapter. The DRP and OPC schemes are then extended to the finite
volume approach. Four linear and nonlinear test problems are presented to evaluate the
merits of these schemes.
Chapter 4 presents the principal characteristics for the Cartesian grid, cut-cell
approach. Additionally, we present the proposed adjustment for the acoustic approach.
Finally, several test problems are presented to demonstrate the performance of the present
approach.
Finally, in Chapter 5, summary and future perspectives are presented.
30
CHAPTER 2 ASSESMENT OF DISPERSION-RELATION-PRESERVING AND SPACE-TIME
CE/SE SCHEMES FOR WAVE EQUATIONS
Introduction
A number of numerical schemes based on various concepts have been proposed to
treat wave propagation and convective transport, including the concepts of upwinding,
flux splitting, total variation diminishing, monotonicity, non-oscillatory, higher order
differencing, and Riemann solvers. For some representative works, see van Leer (1979),
Roe (1981), Harten (1983, 1989), Osher and Chakravarthy (1983), Shyy (Shyy, 1983;
Shyy et al., 1997), Leonard (1988), Shu and Osher (1988), Hirsch (1990), Liou and
Steffen (1993), Tam (Tam and Webb, 1993; Tam et al., 1993), Chang (Chang, 1995;
Chang et al., 1999), Thakur et al. (Part I, 1996; Part II, 1996), Toro and Billet (1996), Yu
and Chang (1997), Loh et al. (1998, 1999, 2000), Wang and Moin (2000), Oran and
Boris(2001). In this chapter, we focus on two approaches: the dispersion-relation –
preservation (DRP) scheme based on an optimized high-order finite difference concept,
proposed by Tam (Tam and Webb, 1993; Tam et al., 1993), and the space-time concept,
proposed by Chang (Chang, 1995; Chang et al., 1999). These two schemes have derived
based on interesting concepts, and have been applied to compute flow problems requiring
balanced treatment for both dispersion and dissipation.
The well studied wave equation, first order linear hyperbolic equation, is:
0u uct x
∂ ∂+ =
∂ ∂, where a > 0 is a constant (2.1)
31
The exact solution of the wave equation for the initial value problem with initial
data
( ,0) ( )( , ) ( )
u x U xu x t U x ct
== −
(2.2)
A simple equation like this one has served useful purpose to guide the development
in numerical techniques.
The challenge for the numerical solution of the wave equation is to maintain the
correct sharpness of the solution without creating wiggles. In other words, the goal is to
offer satisfactory resolution, especially for the high wave number components, by
balancing numerical dissipation and dispersion. Here we use the term “balancing”
because of the finite resolution possessed by numerical computation.
The two methods optimized higher-order finite difference (DRP) method (Tam and
Webb 1993) and the space-time conservation element and solution element (CE/SE)
(Chang, 1995) address the aforementioned issues from different angles. The philosophy
of the DRP method is to maximize the accuracy by matching the wave number and
frequency characteristics between the analytical and the numerical operators in the range
of resolvable scales. The space-time CE/SE method views the flux calculations in a
joined space-time conservation element, taking into account the unified wave propagation
in a solution element. The present work intends to complement the analyses offered in the
original studies to gain further insight into issues related to dispersion, dissipation, and
resolution. The main features of both approaches will be highlighted via truncation error
analysis and numerical computation to evaluate order of accuracy, stability constraints,
and the implication in spatial and temporal resolutions of each scheme. We contrast the
32
performance of the two schemes by varying the ratio between the dominant wavelength
and numerical mesh spacing from long (~ 20 or higher) to short (~ 3).
In the following, we will first present the concepts and derivation of both schemes,
and then conduct computational evaluations. Some of details presented in Tam and
Webb’s paper (1993) and in Chang’s paper (1995) will be summarized to help discuss the
related analysis and numerical assessment conducted in the present work
The Dispersion-Relation Preservation (DRP) Scheme
In the following, we summarize the method by Tam and Webb (1993) to highlight
the salient features of DRP scheme.
Discretization in Space
As illustrations, x
∂∂
is approximated by the central difference schemes of various
orders considering uniform grids with spacing to ∆x, we give three examples.
• The second order central difference approximation is
21 1 ( )2
l l
l
u uu O xx x
+ −−∂⎛ ⎞ = + ∆⎜ ⎟∂ ∆⎝ ⎠ (2.3)
• Fourth-order approximation is
42 1 1 28 8 ( )12
l l l l
l
u u u uu O xx x
+ + − −− + − +∂⎛ ⎞ = + ∆⎜ ⎟∂ ∆⎝ ⎠ (2.4)
• Sixth-order approximation is
63 2 1 1 2 39 45 45 9 ( )60
l l l l l l
l
u u u u u uu O xx x
+ + + − − −− + − + −∂⎛ ⎞ = + ∆⎜ ⎟∂ ∆⎝ ⎠ (2.5)
The construction of the scheme (Tam and Webb, 1993)
( ) ( )1 N
jj N
u x a u x j xx x =−
∂≅ + ∆
∂ ∆ ∑ (2.6)
33
is based on two goals: (i) the behavior of the numerical solution in the resolvable wave
number range closely matches that of the exact solution, and (ii) the formal order of
accuracy of scheme spanning 2N + 1 nodes is 2(N - 1).
To facilitate the derivation, one can start with the Fourier Transform of a function -
f(x) and its inverse
( ) ( )12
i xf f x e dxααπ
∞−
−∞
= ∫% (2.7)
( ) ( ) i xf x f e dαα α∞
−∞
= ∫ % (2.8)
as well as the derivative and shift theorems
( )
( )
~
i fx
f xα α=
∂
∂% (2.9)
( ) ( )~
if x e fαλλ α+ = % (2.10)
With these tools, the relationship between the differential and discrete operators
(given in Eq.(2.6)) can be approximated as
1 Nij x
jj N
i f a e f i fx
αα α∆
=−
⎡ ⎤≅ =⎢ ⎥∆ ⎣ ⎦
∑% % % (2.11)
where
N
ij xj
j N
i a ex
αα ∆
=−
−=
∆ ∑ (2.12)
We see that α is a periodic function of α∆x with a period of 2π. Obviously the
goal is to ensure that α is as close to α as possible. To accomplish this goal the error is
minimized over a certain wavenumber range, α∆x ∈[−η; η] − the numerical dispersion is
reduced by narrowing the range of optimization (Tam and Webb, 1993)
34
2 ( )E x x d xη
η
α α α−
= ∆ − ∆ ∆∫ (2.13)
It is noted that α is real, and hence the coefficients aj must be anti-symmetric, i.e.,
a0 = 0 and a-j = -aj. (2.14)
On substituting Eq.(2.12) into Eq.(2.13) and taking
Eq.Error! Reference source not found. into account, E can be written as
( )2
12 sin
N
jj
E k a k j dkη
η =−
⎡ ⎤= − ⋅⎢ ⎥
⎣ ⎦∑∫ (2.15)
where k = α∆x
Hence, with the anti-symmetric condition, one can use the method of the least-
squares fitting to minimize E, over a wave number range, namely:
0,l
Ea
∂=
∂ l = 1, 2,…N (2.16)
Furthermore, to ensure that the scheme is accurate to O(∆x2(N-1)), additional
conditions can be made by expanding the right side of Eq.(2.6) as a Taylor series and
then equating terms. In this way we have N -1 independent equations with N unknowns:
(aj)j=1,N. Note that aj = -a-j, j = 1,…,N, and a0 = 0.
Tam (Tam and Webb, 1993) shows the relationship between xα∆ and α∆x for the
4th –order, DRP scheme and, the aforementioned 6th-order, 4th -order and 2th -order central
difference schemes over the interval 0 to π
For α∆x up to αc∆x the individual curves are nearly the same as the straight line
x xα α∆ = ∆ . Thus, the finite difference scheme can provide reasonable approximation for
wave number so α∆x becomes less than αc∆x. If we wish to resolve a short wave with
35
straightforward central difference approximations using a fixed size mesh, we need to use
a scheme with a large stencil. On the other hand, the 7-point DRP scheme has the widest
favorable range of xα∆ . Tam and Webb (1993) demonstrate that DRP can be effective in
improving the performance of a given stencil within certain wave number (see
Figure 2-1). In the following, we will only consider the DRP scheme with a symmetric
stencil and in particular the 7-point formula.
0 0.5 1 1.5 2 2.5 30
0.5
1
1.5
2
2.5
α∆x
DRP 6th order4th order
Figure 2-1. xα∆ versus α∆x for the optimized DRP 4th – order scheme, 7 point stencil,
standard 6th -order central scheme and 4th -order central scheme
To follow wave propagation, it is important to evaluate the group velocity of a
finite difference scheme. The group velocity is characterized by /d dα α , which should
be almost one in order to reproduce exact result. A way to reduce dispersion is to adjust
the range of the wave number in the optimization process. b shows /d dα α curves of the
DRP scheme, for different ranges of the optimization parameter η. Upon examining the
corresponding /d dα α curves, α∆x = 1.1 gives the best overall fit. In this case based on
the criterion 1.0 .01ddαα
− < , the optimized scheme has a bandwidth 15% wider than the
standard 6th order central difference scheme.
36
For long wave the important range of α∆x is small, and hence any value of η less
than 1.1 is reasonable. For short waves, on the other hand, dispersion can be noticeable
with any choice of η.
Table 2-1. The stencil coefficient for N = 3 η a1
a2 a3
π/2 0.7934 -0.1848 0.0254 1.1 0.7688530 -0.1650824 0.0204372 0.9 0.762145 -0.1597162 0.0190957 0.85 0.7607435 -0.1585945 0.0188153 0.7 0.7571267 -0.1557013 0.0180920 0.6 0.7551720 -0.1541376 0.0177011
0 0.2 0.4 0.6 0.8 10.99
0.992
0.994
0.996
0.998
1
1.002
1.004
1.006
1.008
1.01
α∆x
standard-2th orderstandard-4th orderstandard-6th orderDRP scheme-η=1.1
f(α∆x)=1
( )( )
d xd x
αα
∆∆
0 0.2 0.4 0.6 0.8 10.99
0.992
0.994
0.996
0.998
1
1.002
1.004
1.006
1.008
1.01
α∆x
standard-2th orderstandard-4th orderstandard-6th orderDRP scheme-η=1.1
f(α∆x)=1
( )( )
d xd x
αα
∆∆
a)
0 0.2 0.4 0.6 0.8 10.99
0.994
0.998
1.002
1.006
1.01
α∆x
η=0.7 η=0.9 η=1.1 eta=π/2
f(α∆x)=1
( )( )
d xd x
αα
∆∆
0 0.2 0.4 0.6 0.8 10.99
0.994
0.998
1.002
1.006
1.01
α∆x
η=0.7 η=0.9 η=1.1 eta=π/2
f(α∆x)=1
( )( )
d xd x
αα
∆∆
b) Figure 2-2. Dispersive characteristics of DRP scheme: a) ( ) ( )/d x d xα α∆ ∆ versus α∆x for
optimized and standard schemes. .b) ( ) ( )/d x d xα α∆ ∆ versus α∆x for the 4th order optimized scheme for different wave number range of optimization: range is from –η to η
37
Time Discretization
Following Tam and Webb (1993) suppose that u(t) is an unknown vector and the
time axis is divided into a uniform grid with time step ∆t. To advance to the next level,
we use the following 4-level finite difference approximation:
( )3
( 1) ( )
0
n jn n
jj
uu u t bt
−+
=
∂⎛ ⎞≅ + ∆ ⎜ ⎟∂⎝ ⎠∑ (2.17)
Similar to space discretization, the goal is to develop schemes which closely match
the exact solution in the frequency domain, and exhibit desired formal orders of accuracy.
In Eq.(2.17) there are four constants, namely: b0, b1, b2, b3. To determine these
constants and create a scheme of O(∆t3) accuracy, the terms in Eq.(2.17) are expanded in
a Taylor series to match exactly up to order ∆t2. This leaves one free parameter, b0. The
relationship between other coefficients and b0 are (Tam and Webb, 1993; Tam et al.,
1993):
1 0 2 0 3 053 16 233 ; 3 ;12 3 12
b b b b b b= − + = − = − + (2.18)
One can utilize the Laplace transform to determine b0. First, the Laplace transform
and its inverse of a function f(t) are related by
( ) ( ) ( ) ( )0
1 ;2
i t i tf f t e dt f t f e dω ωω ω ωπ
∞−
Γ
= =∫ ∫% % (2.19)
where Γ is a line in the upper half ω-plane parallel to the real- ω -axis above all poles and
singularities. Also, the shifting theorem for Laplace transformation is
( )~
if t e fω− ∆+ ∆ = % (2.20)
To apply the Laplace transform, we need to generalize the equation to one with a
continuous variable, namely,
38
3
0
( )( ) ( ) jj
u t j tu t t u t t bt=
∂ − ∆+ ∆ = + ∆
∂∑ (2.21)
On applying the shifting theorem to four-level scheme presented in Eq.(2.21), we
find that
~
3
0
i t i j tj
j
uue u t b et
ω ω− ∆ ∆
=
⎛ ⎞ ∂= + ∆ ⎜ ⎟ ∂⎝ ⎠
∑% % (2.22)
Hence,
( )
3
0
1i t
ij tj
j
i e
t b e
ω
ωω
− ∆
∆
=
−=
∆ ∑ (2.23)
To optimize the time stepping scheme, the error is optimized
[ ] ( ) ( ) ( )221 Re( ) 1 ImE t t t t d t
ς
ς
σ ω ω σ ω ω ω−
= ∆ − ∆ + − ∆ − ∆ ∆⎡ ⎤⎣ ⎦∫ (2.24)
i.e., 1
0
0.dEdb
=
In Eq.(2.24) σ is the weight and ζ is the non-dimensionalized frequency range
needed for the numerical scheme to match the exact solution. Substituting Eq.(2.18) and
(2.23) into Eq.(2.24) E1 becomes a function of b0 alone,
For σ = 0.36 and ζ = 0.5 the scheme becomes (Tam and Webb, 1993)
0 1 2 32.30256; 2.49100; 1.57434; 0.38589b b b b= = − = = − (2.25)
Eq.(2.23) indicates that the relationship between tω∆ and ω∆t is not one to one.
This means that spurious solutions appear. The stability will be established in function of
the real solution. It can be written in the form:
39
4 3 23 2 1 0 0i ib z b z b z b z
t tω ω⎛ ⎞+ + + + − =⎜ ⎟∆ ∆⎝ ⎠
(2.26)
where z = e iω∆t and the values of bj (j = 0, 1, 2, 3) are given by Eq.(2.26).
To maintain satisfactory temporal resolution while being stable, the imaginary part
of the solution (ω∆t) should be negative but close to zero. The interval 0 0.41tω< ∆ <
satisfies these expectations. Furthermore, Re( ) Re( )t tω ω∆ ≅ ∆ in this range. To ensure
that the damping numerical is minimized, Tam and Webb (1993) suggest the condition
4Im( ) 0.118 10tω −∆ ≤ − ⋅ (i.e., 0 0.19tω< ∆ < is adopted). This condition guarantees
numerical stability and negligible damping.
To compute stability of the scheme we take into consideration Fourier-Laplace
transformation of the wave equation (Eq.(2.1))
% % %12
initiali u ci u uω απ
∗− = − + (2.27)
where α, ω* characterize the PDE . For the long wave we can approximate wavenumber
of the scheme with wavenumber of the PDE
α α≅ (2.28)
which leads to
% % %*initialu c u kuω α= + (2.29)
Hence
* c kkω α= + (2.30)
The condition of the numerical stability is that amplification factor for time
discretization is less than 1, and hence ω*∆t <0.41. It is also noted from Figure 2-1 that
1.8xα∆ ≤ (2.31)
40
hold true. By introducing Eq. (2.31) into (2.30) and upon multiplying by ∆t it is found
that
[ ]1.8 M 1ct kk tx
ω∗∆ ≤ ⋅ + ⋅∆∆
(2.32)
where M is mach number. Therefore to ensure numerical stability it is sufficient by Eq.
(2.32) to restrict ∆t to be less than ∆tmax, where ∆tmax is given by
[ ]max
0.411.8 M 1
xtkk c
∆∆ =
⋅ + (2.33)
Therefore, for ∆t < ∆tmax the schemes are numerically stable. Consequently, the
schemes yield the following criteria for numerical stability:
0.21CFL ≤ (2.34)
Based on Eqs.(2.1), (2.6), and (2.17) one can obtain the final form of the DRP
scheme with 7-point in space and 4-point in time:
3 3
1
0 3
n n n jl l j k l k
j k
tu u c b a ux
+ −+
= =−
∆= −
∆ ∑ ∑ (2.35)
The leading truncation error of the scheme (given in Eq.(2.35)) can be evaluated
using the Taylor series expansion, yielding:
( )
3 33 3
0
4 54 3 3
4 3 5
0 3
16 4
( )32
4! 10 5!
t x j xxxxj
xxxxx
k
j kj k
u c u c j b u x
u x O xcc j j b a k
ν
ν
=
=
= =−
+ ⋅ = − + ∆
− + ∆ ∆
⎛ ⎞⎜ ⎟⎝ ⎠
⎡ ⎛ ⎞ ⎤+ + +⎢ ⎜ ⎟ ⎥⎦⎢ ⎝ ⎠⎣
∑
∑ ∑ (2.36)
Replacing in Eq.(2.36) by the numerical values of the various coefficients, the
scheme becomes:
( )3 3 4 4 51.6318 1.223 0.012814 ( )t x
xxxx xxxxx
u a u
a u x a a u x O xν ν
+ ⋅ =
− ⋅ ⋅ ⋅∆ − ⋅ + ⋅ ∆ + ∆ (2.37)
41
The following can be summarized in regard to the present DRP scheme:
• the scheme is forth order in space, and third order in time
• ν < 0.21 assures stability and reasonable accuracy
• the first term on the RHS of Eq.(2.37) is dissipative. The accuracy bound, ν < 0.21, indicates that the coefficient of the leading dissipative term is small. This observation suggests that, depending on the relative magnitude of ν and ∆x, dispersive patterns may be dominant in numerical solutions.
The Space-Time Conservation Element and Solution Element Method
The tenet in this method is to treat local and global flux conservation in a unified
space and time domain. To meet this requirement, Chang (1995) introduces solution
elements, which are subdomains in the space-time coordinates. Within each solution
element, any flux vector is then approximated in terms of some simple smooth functions.
In the last step, the computational domain is divided into conservation elements within
which flux conservation is enforced. Note that a solution element generally is not the
same as a conservation element. We summarize in the following the key concepts
adopted in Chang’s approach.
a–µ Scheme
Consider Eq.(2.1), and define F1 = -u, F2 = au, x1 = x, , and x2 = t
Applying Green theorem we obtain:
( )2 11 1 2 2
1 2 S
F F dxdy F dx F dxx xΓ
⎛ ⎞∂ ∂− = +⎜ ⎟∂ ∂⎝ ⎠
∫∫ ∫ (2.38)
where ( )1 1 2 2S
F dx F dx+∫ is contour integral on closed curve S.
Following a vector notation, Eq.(2.38) can be written as
( )1 1 2 2S
F dx F dx+∫ = ( )S V
g ds⋅∫r r (2.39)
42
where 1 2( , )g F F=r and ( )1 2,ds dx dx=
r . (dx1, dx2) is a differential vector associated with a
point (x1, x2) on the closed curve S.
Because Eq.(2.1) is valid anywhere in the definition domain, we obtain
( )
0S V
g ds⋅ =∫r r
(2.40)
where (i) S(V) is the boundary of an arbitrary space-time region V in E2 (a two-
dimensional Euclidean space) with x1 = x and x2 = t; (ii) ( )S V
g ds⋅∫r r is contour integral on
closed curve S(V); (iii) (F2, -F1) is a current density in E2. Note that g ds⋅r r is the space-
time flux of (F2, -F1) leaving the region V through the surface element.
As shown in Figure 2-3a, Ω is a set of mesh points (j, n) that is adapted to discretize
a physical domain, where 2, ij n = ± , with i = 0, 1, 2, 3, … To facilitate the construction
of the present scheme a solution element (SE) associated with (j, n) is illustrated in b.
For any (x, y) ∈SE(j, n), u(x, t), is approximated by u∗(x, t; j, n):
*( , ; , ) ( ) ( ) ( ) ( )n n n nj x j j t ju x t j n u u x x u t t= + − + − (2.41)
with nju , ( )n
x ju , and ( )n
t ju are constants in SE(j,n), ( ), n
jx t are coordinates of the mesh
index (j,n), and
( ) ( )* *
* ( , ; , ) ( , ; , )( , ; , ) , ,n nnj j x tj j
u x t j n u x t j nu x t j n u u ux t
∂ ∂= = =
∂ ∂ (2.42)
43
Figure 2-3. Scheme of the solution elements (SEs) and conservation elements (CEs): a) The index positions of SEs and CEs; b) SE(j,n); c) CE-(j,n); d) CE+(j,n); e) CE+(j-1/2, n+1/2); f) CE_(j+1/2, +1/2);
n+1 n+1/2 n n-1/2 n-1
j-3/2 j-1 j-1/2 j j+1/2 j+1 j+3/2 a)
/ 2x∆ / 2x∆
/ 2t∆
/ 2t∆
(j,n)
b)
c) d)
( -1 2 , -1 2)SE j n⊂
(j,n) A
B(j-1/2, n-1/2)
( , )SE j n⊂
( 1 2, -1 2)SE j n⊂ +
(j,n) A
(j+1/2, n-1/2)
(j,n) (j,n)
f) e)
( , )SE j n
( 1 2 , 1 2)SE j n⊂ + + ( -1 2, 1 2)SE j n⊂ +
44
Eq.(2.41) has the form of the first-order Taylor series expansion. Furthermore, if
one takes into account that ( ) ( )n nt xj j
u c u= − , because u∗(x, t; j, n) satisfies wave equation,
then Eq.(2.41) becomes
*( , ; , ) ( ) [( ) ( )]n n nj x j ju x t j n u u x x c t t= + − − − (2.43)
In each SE(j, n), ( , ) ( ( , ), ( , ))g x t u x t cu x t= −r is approximated by
* * *( , ;, , ) ( ( , ; , ), ( , ; , ))g x t j n u x t j n cu x t j n= −r (2.44)
In order to develop appropriate approximation for the flux, one divides the physical
domain into nonoverlapping rectangular elements, referred to as conservation elements
(CEs). Specifically, CE receive sign “-“ or “+” in function of the slope of the line that
connects the vertex from Ω of a CE. If the slope is negative, CE receives the positive
sign; otherwise the sign is negative
In Figure 2-3c the line that unifies the vertices founded in Ω is positive, and CE
index Ω is (j,n). The final notation is CE-(j, n). The surface of CE belongs to two
different SE, SE(j, n) and SE(j-1/2,n-1/2). To specify that a part of surface is in a certain
SE in Figure 2-3c, that part is around by a certain type of line. Figure 2-3d, Figure 2-3e,
and Figure 2-3f illustrate three other corresponding cases relating the conservation
element to the solution element.
The approximation of Eq.(2.40) is
( )( )( ),
, 0S CE j n
F j n g ds±
± = ⋅ =∫r r (2.45)
for all (j, n)∈Ω. ( )( ),S CE j n
g ds±
⋅∫r r is contour integral over closed path S(CE±(j,n)) and
represents the total flux leaving the boundary of any conservation element is zero. The
45
flux at any interface separating two neighboring CEs is calculated using the information
from a single SE. As an example, the interface AC depicted in Figure 2-3c) and Figure 2-
3d) is a subset of SE(j, n). Thus the flux at this interface is calculated using the
information associated with SE(j, n).
We integrate along the entire boundary S(CE±) of CE± such that CE± is on the left
as we advance in the direction of integration (S(CE±) is traversed counterclockwise).
With the above preliminaries, it follows from (39) that:
( )( ) ( )( ) ( )( ) ( ) ( )1 22 2 1 21
1 222 1 2
2 14 , 1 1n n n nx x j jj j
F j n u u u uxxν
ν ν − −± ±±
⎡ ⎤= ± − + − + −⎣ ⎦ ∆∆
m
(2.46)
where, again, c tx
ν ∆=
∆ is CFL number.
With the aid of Eqs.(2.45) and (2.46), nju and ( )n
x ju can be solved in terms of 1 2
1 2nju −−
and ( ) 1 2
1 2
nx j
u −
±, if (1 - ν2) ≠ 0, i.e.,
( ) ( ) ( ), 1 2, 1 2 1 2, 1 2q j n Q q j n Q q j n+ −= − − + + − (2.47)
where
( )( )
( , )/ 4
nj
nx j
uq j n
x u
⎛ ⎞⎜ ⎟=⎜ ⎟∆⎝ ⎠
for all (j, n) ∈ Ω (2.48)
( )
( )( )
22 1 11 11 1;2 21 1 1 1
Q Qν νν ν
ν ν+ −
⎛ ⎞− − −⎛ ⎞+ −⎛ ⎞ ⎛ ⎞⎜ ⎟= =⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟− − −⎝ ⎠ ⎝ ⎠ + − +⎝ ⎠ ⎝ ⎠ (2.49)
We have employed the Taylor series expansion to evaluate the leading truncation
errors of Eq.(2.46), yielding the following two expressions for F+ and F- respectively:
46
22 4
22 4
( )48 16 16 48
( )48 16 16 48
t x xxx
t x xxx
c c c cu cu u x O x
c c c cu cu u x O x
ν νν
ν νν
⎧ ⎛ ⎞+ = − + + + ⋅∆ + ∆⎪ ⎜ ⎟
⎪ ⎝ ⎠⎨
⎛ ⎞⎪ + = − + − + ⋅∆ + ∆⎜ ⎟⎪ ⎝ ⎠⎩
(2.50)
From our derivation indicated by Eqs.(2.49), and (2.50) we can establish the
following observation for the space-time CE/SE scheme:
• The method is second order in space and time;
• The method is stable if ν2 < 1
• The dispersion strength of the scheme increases when ν (or time step) is reduced because ν appears in the denominator in the leading truncation error term in Eq.(2.50). Hence a small value of ν may not be desirable.
• The third order term O(∆x3) is zero in Eq(2.50), indicating that not only the ∂2u/∂x2 term but also the ∂4u/∂x4 terms are non-present. This is the second reason, in addition to previous observation, why the present scheme can be highly dispersive.
a-ε Scheme
In order to help to reduce the dispersive nature of the present scheme, Chang
(1995) introduces an “artificial viscosity” into original CE/SE scheme. Instead of
F±(j,n) = 0, the following modification is made:
( )( )
( )221
( , )4
nx j
xF j n du
ε ν±
− ∆= ± (2.51)
where ε is an independent parameter of the numerical variables, and
( ) ( ) ( ) ( ) ( )1/ 2 1/ 2 1/ 2 1/ 211/ 2 1/ 22 1/ 2 1/ 2
/n n n n nx x x j jj j j
du u u u u x− − − −+ −+ −
⎡ ⎤= + − − ∆⎣ ⎦ (2.52)
Because the magnitude of the added terms in the scheme is controlled by ε, the
numerical dissipation is controlled by ε. However, because F±(j,n)≠0 if ε ≠ 0, strictly
specking CE+(j,n) and CE-(j,n) are no longer conservation elements in the a-ε scheme. On
the other hand, although the net flux entering the interface separating CE+(j,n) and
47
CE-(j,n) is not zero, but the sum of F+(j,n) and F-(j,n) is zero. Hence the total flux leaving
CE(j,n) vanishes. As a result, CE(j,n) is a conservation element in the a-ε scheme. These
are seemingly subtle but relevant feature of the present scheme.
Using Eqs.(2.47), (2.51), (2.52) one obtains the solution of the a-ε scheme as:
( ) ( ) ( ), 1 2, 1 2 1 2, 1 2q j n M q j n M q j n+ −= − − + + − (2.53)
where
( ) ( ) ( )221 12 2
1 11 1;
1 2 1 1 2 1M M
ν νν νε ε ν ε ε ν
+ −
⎛ ⎞− − −⎛ ⎞+ −= = ⎜ ⎟⎜ ⎟ ⎜ ⎟− − + − − +⎝ ⎠ ⎝ ⎠
(2.54)
To assess the leading truncation errors in the a-ε scheme a modified equation is
derived by substituting Taylor series expansion into Eq.(46) for ( )1212
nx j
u −
±,
1212
nju −±
The resulting equations are:
22
3
22
3
12 4 4 12 3 3 2
6 6 2
12 4 4 12 3 3 2
6 6 2
t x xxx
xxxx
t x xxx
xxxx
c c c c c c xu cu u
c c xu HOT
c c c c c c xu cu u
c c xu HOT
ν ν ε εν ν
νεν
ν ν ε εν ν
νεν
⎧ ⎛ ⎞ ∆⎛ ⎞+ =− + + + − −⎪ ⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠⎪
⎪ ∆⎛ ⎞⎛ ⎞⎪ + − +⎜ ⎟⎜ ⎟⎪ ⎝ ⎠⎝ ⎠⎨
⎛ ⎞ ∆⎛ ⎞⎪ + = − + − + − +⎜ ⎟⎜ ⎟⎪ ⎝ ⎠⎝ ⎠⎪⎪ ∆⎛ ⎞⎛ ⎞+ − +⎪ ⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠⎩
(2.55)
From Eq.(2.55), we can see that the a-ε scheme introduce dissipative term uxxxx and
reduce the strength of the dispersive uxxx. The net outcome term is that the formal order of
accuracy is maintained while the scheme becomes less dispersive. Again, it is undesirable
to employ a small ν, which is different from the DRP scheme discussed previously. As
48
already mentioned, in the a-ε scheme, both dispersive and dissipative aspects are affected
by ε.
In Refs. (Chang, 1995; Chang et al., 1999), insight is offered in regard to the choice
of ε and ν, based on the von Neumann stability analysis of the a-ε scheme, without
detailed information of the truncation error. The analysis reported in these references
suggests that: (i)ν should be large; (ii) ε shouldn’t be close to 0 or 1; (iii) the spurious
solution can be effectively suppressed for ε = 0.5 in the case of long-wavelength.
Reference (Loh et al., 1998) suggests that for aeroacoustic computation, it is essential to
choose a large CFL number and a small ε.
Further insight can be gained based on the present truncation error analysis. If we
relate ε and ν ( CFL number) by ε = ν + ε1, then, in Eqs. (2.55):
• Coefficient of uxxx in first equation of the system Eq. (2.55) is equal to:
2
11 ( 1)
3 4cν ν ε
ν⎡ ⎤+ −⎛ ⎞ −⎜ ⎟ ⎢ ⎥⎝ ⎠ ⎣ ⎦
(2.56)
• Coefficient of uxxx in second equation of system Eq. (2.55) is equal to:
( )2
1
6 113 4
cν νν ε
ν
⎡ ⎤− +−⎛ ⎞ ⎢ ⎥−⎜ ⎟⎝ ⎠ ⎢ ⎥⎣ ⎦
(2.57)
The above equations indicate that if ν is close to 1, and ε and ν are close to each
other, i.e., if ε1 is small, then the leading dispersion error shown in Eq.(2.55) is small. On
the other hand, to maintain adequate numerical dissipation, it is helpful to let ε vary in the
same manner as ν. These observations as well as the analysis in Chang (1995) indicate
that the value of ν and ε should be coordinated. Further evaluation will be made based on
numerical computations, to help establish a more explicit guideline.
49
Numerical Assessment of the DRP and Space-Time Schemes
To assess the individual and relative merits of the DRP and space-time schemes the
following simple test problem is adopted.
0.u ut x
∂ ∂+ =
∂ ∂ (2.58)
The initial condition imposed at time t = 0 are:
2
2
exp ln 2
2ln 2 exp ln 2x
xub
x xub b
⎧ ⎡ ⎤⎛ ⎞= −⎪ ⎢ ⎥⎜ ⎟⎝ ⎠⎢ ⎥⎪ ⎣ ⎦
⎨⎡ ⎤⎪ ⎛ ⎞ ⎛ ⎞= − −⎢ ⎥⎜ ⎟ ⎜ ⎟⎪ ⎝ ⎠ ⎝ ⎠⎢ ⎥⎣ ⎦⎩
(2.59)
which is a Gaussian profile.
The exact solution is:
2
2
exp ln 2
2ln 2 exp ln 2x
x tUb
x t x tUb b
⎧ ⎡ ⎤−⎛ ⎞= −⎪ ⎢ ⎥⎜ ⎟⎝ ⎠⎢ ⎥⎪ ⎣ ⎦
⎨⎡ ⎤− −⎪ ⎛ ⎞ ⎛ ⎞= − −⎢ ⎥⎜ ⎟ ⎜ ⎟⎪ ⎝ ⎠ ⎝ ⎠⎢ ⎥⎣ ⎦⎩
(2.60)
In this study we evaluate the performance of the schemes in short, intermediate,
and long waves relative to the grid spacing, which is assured by the value b/∆x. In all
cases ∆x =1, and c = 1, so that ∆t = ν.
For the optimized DRP scheme, we adopt the one with 4th level in time and 7-point
in space, and 4th order formal accuracy in space and 3rd order formal accuracy in time:
( )3
1
0
n n n jl l j l
j
u u b Fν+ −
=
= + ∑ (2.61)
where tx
ν ∆=
∆ (because c = 1)
50
( ) ( )3
3
k kl j l j
j
F a u +=−
= −∑ (2.62)
The values of the of bj are given in Eq.(24), and the value aj are as chosen to
minimize difference between ddαα
and 1 (see Table 2-1).
For the space-time a-ε scheme, the following formulas result:
( )
( )
( )
( )
1 122 2
1 12 2
1 122 2
1 12 2
1 12 21 12 2
1 12 21 12 2
1 1 (1 ) ( )2 4
1 (1 ) ( )4
1( ) 1 (2 1 ) ( )2 4
1 (2 1 ) ( )4
n nnj xj j
n n
xj j
n nnx j xj j
n n
xj j
xu u u
xu u
xu u u
xu u
ν ν
ν ν
ε ε ν
ε ε ν
− −
− −
− −
+ +
− −
− −
− −
+ +
⎧ ⎡ ∆= + + −⎪ ⎢
⎣⎪⎪ ⎤∆⎪ + − − − ⎥⎪⎪ ⎦⎨
⎡ ∆⎪ = − + − +⎢⎪⎣⎪
⎪ ⎤∆+ − + − −⎪ ⎥
⎪ ⎦⎩
(2.63)
Before evaluating the DRP and space-time a-ε method, we further address the
influence of the parameters: ε and ν for the space-time a-ε scheme, and ν for the DRP
scheme.
An effort is made in this study to construct a simple guidance appropriate for long
as well as short waves. Evidences based on the test problems, aided by the truncation
error analysis, indicate that in order to reduce numerical dispersion and to maintain
satisfactory resolution, for short wave ( e.g., b/∆x = 3), ν and ε are preferable to be close
to each other (see Figure 2-4 and Figure 2-5). For long and intermediate waves there are
virtually no need to introduce much numerical dissipation, hence ε can be reduced close
to zero. On the other hand, mismatched ε and ν can significantly reduce the accuracy of
the scheme. To demonstrate this fact, Figure 2-4a and Figure 2-4b show two solutions for
51
the intermediate wave, all with ν = 0.5, respectively for ν = 0.9. Changing ε from 0.5 to
0.99 causes significant numerical dissipation. On the other hand, ε = 0 can be a very
acceptable choice in for long wave. Figure 2-5 compiles long and intermediate wave
solution with different value of ν and ε. Taking into consideration the previous
observations, it is decided that ν = ε = 0.99 is a good choice, and is used in the results
presented for the space-time scheme. In more complicated computations involving
coupled system with nonlinear effects, such a choice may not be stable. However, for a
simple wave equation, this coordinated choice is beneficial.
For the present DRP scheme a value of ν less than 0.21 guarantees numerical
stability and negligible numerical damping. If we decrease the value of ν much further,
then, as indicated in Eq.(2.37), the numerical damping, as indicated by the leading
dissipation term uxxxx, may be less than adequate due to the small value of ν. For these
reasons we decided to use ν = 0.1 for longer waves and ν = 0.2 for short wave (b/∆x = 3).
To summarize, the DRP scheme is higher formal order than the space-time a-ε
scheme. The DRP scheme prefers smaller ν (or ∆t), indicating that it is more expensive to
compute than the space-time a-ε scheme.
To evaluate the solution accuracy, we define an error vector as:
[ ]1,...,T
NE E E=r
(2.64)
U(xi) is the exact solution at the point xi, and ui is the numerical solution at the
point xi.
The error norm is adopted as norm one
52
1
N
ii
EE
N==∑
(2.65)
where
( ) , 1i i iE U x u i N= − ≤ ≤ (2.66)
which will be used to help to measure the order of accuracy in actual computations.
x
u(x,
t)
160 180 200 220
-0.2
0
0.2
0.4
0.6
0.8
1numerical solution =0.10E-01analytical solution
ε
x
u(x,
t)
160 180 200 220
-0.2
0
0.2
0.4
0.6
0.8
1numerical solution =0.10E+00analytical solution
ε
x
u(x,
t)
160 180 200 220
-0.2
0
0.2
0.4
0.6
0.8
1numerical solution =0.50E+00analytical solution
ε
x
u(x,
t)
160 180 200 220
-0.2
0
0.2
0.4
0.6
0.8
1
numerical solution =0.70E+00analytical solution
ε
x
u(x,
t)
160 180 200 220
-0.2
0
0.2
0.4
0.6
0.8
1
numerical solution =0.90E+00analytical solution
ε
x
u(x,
t)
160 180 200 220
-0.2
0
0.2
0.4
0.6
0.8
1
numerical solution =0.99E+00analytical solution
ε
a)
x
u(x,
t)
20 30 40 50 60
0
0.2
0.4
0.6
0.8
1
numerical solution =0.50E+00analytical solution
ε
x
u(x,
t)
20 30 40 50 60
0
0.2
0.4
0.6
0.8
1
numerical solution =0.99E+00analytical solution
ε
b) Figure 2-4. Comparison between analytical and numerical solutions – Effect of ε on the
accuracy of space-time a-ε scheme: a) ν = 0.9, b/∆x = 3, t = 200; b) b/∆x = 6, t = 200, ν = 0.5
53
ε0.25 0.5 0.75 1
1.80E-03
2.20E-03
2.60E-03
ν=0.99error
ε0.25 0.5 0.75 1
2.5E-02
3.5E-02
4.5E-02
ν=0.5error
ε0.25 0.5 0.75 1
4.0E-02
6.0E-02
ν=0.1error
a)
ε0.25 0.5 0.75 1
2.0E-02
4.0E-02
6.0E-02
8.0E-02
ν=0.1error
ε0.25 0.5 0.75 1
1.5E-02
2.5E-02
3.5E-02
4.5E-02
5.5E-02 ν=0.5error
ε0.25 0.5 0.75 1
4.65E-04
4.75E-04
4.85E-04
4.95E-04
ν=0.99error
b)
ε0.25 0.5 0.75 1
2.0E-02
4.0E-02
6.0E-02
8.0E-02
1.0E-01
1.2E-01
ν=0.1error
ε0.25 0.5 0.75 1
5.0E-03
1.0E-02
1.5E-02
2.0E-02
2.5E-02ν=0.5error
ε0.25 0.5 0.75 1
4.06E-05
4.07E-05
4.08E-05
4.09E-05
ν=0.99error
e
erro
r
0.25 0.5 0.751.53E-03
1.55E-03
1.57E-03
e
erro
r
0.2 0.4
2.04E-032.06E-032.08E-032.10E-032.12E-03
c) Figure 2-5. The dependence of the error on ε for the space-time a-ε scheme at t = 200:
a) b/∆x = 3; b) b/∆x = 6; c) b/∆x = 20
54
In the following we will present the results based on three categories:
• long wave (b/∆x = 20)
• intermediate wave (b/∆x = 6)
• short wave (b/∆x = 3)
They are defined according to the ratio b/∆x, where b is a parameter that
characterizes the wavelength.
Figure 2-6 offer an overview of the two schemes for long, intermediate and short
wave computation at two different time instants: t/∆t = 200 and 2000. It is noted that
since ∆x is fixed in each plot, varying ν is the same as varying ∆t. First, we analyze errors
for DRP scheme applied to both intermediate and long waves. The following are some of
the main observations: (i) the characteristics are similar for short and long times; (ii) the
error increases with ν: (iii) a slower increase with respect to ν, for ν little.
For short wave, DRP scheme behaves differently as time progresses. As previously
discussed in the context of truncation error analysis, Eq.(30), a small value of ν can cause
insufficient numerical damping. Hence, it is advisable to use a somewhat larger value of
ν, but still within the stability bound of ν ≤ 0.21.
The space–time a-ε scheme has exhibited unusual error variations as a function of ν
(or ∆t): the error changes very little with ν, and then suddenly decreases rapidly for ν
close to 1 (see Figure 2-6 and Figure 2-7). This trend holds for all resolution. This type
of behavior is unusual at first glance, but can be explained from the viewpoint of
truncation error analysis. Specifically, with ν = ε, both uxxx and uxxxx terms are small as ν
approach 1, making the scheme sharply improves the apparent order of accuracy.
Next, we present the error as a function of spatial resolution (b/∆x).
55
ν0.1 0.2
1.3E-03
1.3E-03
1.4E-03
1.4E-03
1.5E-03
1.5E-03
1.6E-03
1.6E-03
DRP scheme
t=200b/ x=3.∆
error
1
2
0.011
ν 0.1 0.2
1.0E-04
2.0E-04
3.0E-04
4.0E-045.0E-04
DRP scheme
t=200b/ x=20.∆
error
1
0.9
0.41
ν0.1 0.2
1.0E-04
2.0E-04
3.0E-04
4.0E-045.0E-04
0.9
1
DRP scheme
t=200b/ x=6.∆
error
0.31
a1)
ν0.1 0.2
7.5E-03
7.6E-03
7.6E-03
7.6E-03
7.7E-03
DRP scheme
t=2000b/ x=3.∆
error
ν0.1 0.2
6.0E-04
7.0E-04
8.0E-04
9.0E-04
1.0E-030.4
1
DRP scheme
t=2000b/ x=6.∆
error
0.11
ν 0.1 0.2
5.0E-04
6.0E-04
7.0E-04
8.0E-04
9.0E-04
DRP scheme
t=2000b/ x=20.∆
error
1
0.4
0.11
a2)
ν
erro
r
10-2 10-1 100
5.0E-05
5.0E-04
9.5E-041.4E-031.9E-03
Space - time CE SE scheme
b/ x=20.t=200
∆
ν=ε
ν
erro
r
10-2 10-1 1005.0E-04
5.0E-03
9.5E-031.4E-021.9E-02
Space - time CE SE scheme
b/ x=6.t=200
∆
ν=ε
ν
erro
r
10-2 10-1 100
2.5E-03
1.3E-02
2.3E-02
3.3E-02
b/ x=3.t=200
∆
Space - time CE SE scheme
ν=ε
b1)
ν0.1 0.2
7.5E-03
7.6E-03
7.6E-03
7.6E-03
7.7E-03
DRP scheme
t=2000b/ x=3.∆
error
ν0.1 0.2
6.0E-04
7.0E-04
8.0E-04
9.0E-04
1.0E-030.4
1
DRP scheme
t=2000b/ x=6.∆
error
0.11
ν 0.1 0.2
5.0E-04
6.0E-04
7.0E-04
8.0E-04
9.0E-04
DRP scheme
t=2000b/ x=20.∆
error
1
0.4
0.11
b2)
Figure 2-6. The dependence of the error as function of ν for short (b/∆x = 3), intermediate (b/∆x = 6) and long (b/∆x = 20) waves for: a) DRP scheme; b) Space time a-ε scheme, at two different time instants t = 200 and t = 2000. In all cases ∆t = ν
56
x
u(x,
t)
180 190 200 210 220
0
0.2
0.4
0.6
0.8
1
numerical solution =0.99E+00analytical solution
ν=ε=0.99
ε
x
u(x,
t)
180 190 200 210 220
0
0.2
0.4
0.6
0.8
1numerical solution =0.90E+00analytical solution
ν=ε=0.9
ε
Figure 2-7. Effect of ν on the accuracy of space time a-ε scheme: b/∆x = 3, t = 200
Figure 2-8 depicts the dependence of error for b/∆x between 3 and 300. The graphs
show that the space-time a-ε method is second order accuracy in space. In contrast with
the DRP scheme does not exhibit consistent trends for varying spatial resolutions: (i) for
short and intermediate waves the order accuracy of this scheme is four; (ii) for long wave,
the error tends to become almost independent of space (the slope is close to 0.1).
Comparing the performance of these two schemes, we can deduce that the DRP scheme
gives a better solution for b/∆x less than 10 (for a short and intermediate wave). For long
wave, the space-time CE/SE scheme gives a better order of accuracy; nevertheless both
schemes are satisfactory in practical terms. Finally, we compare these two methods with
respect to their performance with sufficient accumulation of the time steps.
57
b/ x
erro
r
101 10210-5
10-4
10-3
10-2
∆
1
Space - Time schemea −ε
2
ν=ε=0.1t=200
b/ x
erro
r
101 10210-5
10-4
10-3
10-2
∆
DRP scheme
1
4
0.011
ν=0.02t=200
b/ x
erro
r
101 102
10-6
10-5
10-4
10-3
∆
2
1
Space - Time schemea −ε
ν=ε=0.99t=200
b/ x
erro
r
101 102
10-6
10-5
10-4
10-3
∆
DRP scheme
41
0.11
ν=0.1t=200
Figure 2-8. The behavior of the error in function of the wavelength: comparison between
DRP and space-time a-ε schemes
Short Wave: b/∆x = 3
Figure 2-9 summarizes various aspects of error accumulation in time, along with
selected solution profiles for both schemes. It is apparent that the space time a-ε scheme
introduces both dissipation and dispersion errors, while the DRP scheme shows mainly
dispersion errors with less dissipation.
In addition, the two schemes exhibit different levels and growth rates. For the
space-time a-ε scheme, at the beginning, the rate of accumulation is slower. Then, around
t ≅ (103), the dissipation becomes more substantial. Around t ≅(5⋅103), both dissipation
58
and dispersion characterize the solution, but the overall error growth rate decreases
slightly in comparison to the previous state.
Initially, the DRP scheme produces lower error level and grows almost linearly
with time. After t ≅ (103), the error growth rate slows down; the dispersion becomes
substantial, but the solution is not smeared. Around t ≅ (2⋅104), dispersion becomes
highly visible while dissipation is also present at a modest level.
Intermediate Wave: b/∆x = 6
As shown in Figure 2-10, for this case, the DRP scheme has lower error than the
space-time a-ε scheme. The space-time a-ε scheme exhibits similar growth trends in both
cases, intermediate and short wave. Again, both dispersion and dissipation errors are
noticeable.
The DRP scheme behaves differently between short and intermediate waves. The
error growth rate at the beginning is low, and then becomes higher. Throughout the entire
computation, only minor dispersion error appears.
Long Wave: b/∆x = 20
Figure 2-11 presents the solutions for long wave computations. Both schemes
perform well over the interval of time. But the space-time a-ε scheme exhibits noticeably
slower rate of error accumulation. Neither dispersion nor dissipation errors are
significant.
59
x
u(x,
t)
19980 20000 20020-0.2
0
0.2
0.4
0.6
0.8
1numerical solution - t=0.20E+05analytical solution t=0.20E+05
D
x
u(x,
t)
880 890 900 910 9200
0.2
0.4
0.6
0.8
1numerical solution - t=0.90E+03analytical solution t=0.90E+03B
x
u(x,
t)
5260 5280 5300 5320
0
0.2
0.4
0.6
0.8
1
numerical solution - t=0.53E+04analytical solution t=0.53E+04
C
x
u(x,
t)
80 90 100 110 1200
0.2
0.4
0.6
0.8
1
numerical solution - t=0.10E+03analytical solution t=0.10E+03A
t103 104
1.0E-02
1.5E-02
2.0E-02
2.5E-023.0E-02
b/ x=3∆ν=ε=0.99
Space - Time scheme
AB
C
D
1
1
errora -ε
a)
x
u(x,
t)
1475 1500 1525
0
0.2
0.4
0.6
0.8
1
numerical solution t=0.150E+04analytical solution t=0.150E+04B
x
u(x,
t)
80 100 120
0
0.2
0.4
0.6
0.8
1
numerical solution t=0.100E+03analytical solution t=0.100E+03A
x
u(x,
t)
19950 20000-0.2
0
0.2
0.4
0.6
0.8
1numerical solution t=0.200E+05analytical solution t=0.200E+05
C
t102 103 104
2.0E-03
1.0E-02
1.8E-022.6E-02
1
0.8
1
0.65
A
B
CDRP schemeb/ x=3.∆ν=0.2η=0.85
error
b) Figure 2-9. Accumulation of the error in time for short wave - b/∆x = 3; a) space-time
a-ε scheme; b)DRP scheme
60
x
u(x,
t)
20000-0.2
0
0.2
0.4
0.6
0.8
1numerical solution - t=0.20E+05analytical solution t=0.20E+05C
x
u(x,
t)
80 100 120
0
0.2
0.4
0.6
0.8
1numerical solution - t=0.10E+03analytical solution t=0.10E+03A
x
u(x,
t)
3575 3600 36250
0.2
0.4
0.6
0.8
1
numerical solution - t=0.36E+04analytical solution t=0.36E+04B
t103 1045.0E-03
1.0E-02
1.5E-02
2.0E-02
2.5E-02b/ x=6∆ν=ε=0.99
Space - Time scheme
A
B
C
1
1
errora - ε
a)
x
u(x,
t)
80 90 100 110 120
0
0.2
0.4
0.6
0.8
1
numerical solution t=0.100E+03analytical solution t=0.100E+03
A
x
u(x,
t)
19980 20000 20020-0.2
0
0.2
0.4
0.6
0.8
1
numerical solution t=0.200E+05analytical solution t=0.200E+05
B
t102 103 104
1.0E-03
3.0E-03
5.0E-03
b/ x=6.∆
DRP scheme
A
B
ν=0.1 1
1η=0.6
error
b)
Figure 2-10. Accumulation of the error in time for intermediate wave - b/∆x = 6; a) space-time a-ε scheme; b) DRP scheme
61
x
u(x,
t)
50 100 1500
0.2
0.4
0.6
0.8
1numerical solution - t=0.10E+0analytical solution t=0.10E+03A
x
u(x,
t)
19950 20000 200500
0.2
0.4
0.6
0.8
1numerical solution - t=0.20E+05analytical solution t=0.20E+05
B
t103 104
5.0E-03
5.2E-03
5.4E-03
5.6E-03
5.8E-03b/ x=20∆ν=ε=0.99
Space - Time scheme
A
Berror
1
0.12
a - ε
a)
x
u(x,
t)
50 75 100 125 150
0
0.2
0.4
0.6
0.8
1
numerical solution t=0.100E+03analytical solution t=0.100E+03
A
x
u(x,
t)
19950 19980 20010 20040-0.2
0
0.2
0.4
0.6
0.8
1
numerical solution t=0.200E+05analytical solution t=0.200E+05
B
t102 103 104
5.0E-04
1.5E-03
2.5E-03
3.5E-034.5E-03
b/ x=20.∆
DRP scheme
A
B
ν=0.1 10.9
error
η=0.6
b)
Figure 2-11. Accumulation of the error in time for long wave - b/∆x = 20, a) space-time a-ε scheme; b) DRP scheme
62
Summary and Conclusions
It should be noted that he DRP scheme is a multi-step method, which requires more
boundary conditions and initial data, while the space-time a-ε scheme is a one-step
method. Combined with the fact that the DRP scheme performs better with smaller ν, it
can be more expensive to compute than for space-time a-ε scheme. Tam and Webb
(1993) show by means of the Laplace transform that the DRP scheme can be constructed
to use the same number of initial data as the original PDE and the single time step
method. For the boundary treatment, the DRP scheme requires additional points outside
the computational domain. A combination of ghost points and backward difference
operators, based on similar optimization procedures, is employed.
The present study is restricted to the investigation of a simple 1-D linear wave
equation. Obviously, more issues will arise when multi-dimensional geometry,
nonlinearity of the physics, and coupling of the dependent variables need to be
considered. Nonuniform grid also creates varying CFL number even with a constant
convection speed. Nevertheless, it seems useful to consider, in a well-defined framework,
the merits of the two schemes in a well-defined context. In this sense, the present study
can be viewed to establish an upper bound of the performance characteristics of the DRP
and space-time CE/SE methods.
63
CHAPTER 3 FINITE VOLUME TREATMENT OF DISPERSION-RELATION-PRESERVING AND OPTIMIZED PREFACTORED COMPACT SCHEMES FOR WAVE PROPAGAION
In computational aero-acoustics (CAA) accurate prediction of generation of sound
is demanding due the requirement of preserving the shape and frequency of wave
propagation and generation. Furthermore, the numerical schemes need to handle multiple
scales, including long and short waves, and nonlinear governing laws arising from
sources such as turbulence, shocks, interaction between fluid flows and elastic structures,
and complex geometries. It is well recognized (Hardin and Hussaini, 1992; Tam and
Webb, 1993; Tam et al., 1993) that in order to conduct satisfactory CAA, numerical
schemes should induce minimal dispersion and dissipation errors. In general, higher-
order schemes would be more suitable for CAA than the lower-order schemes since,
overall, the former are less dissipative. That is why higher-order spatial discretization
schemes have gained considerable interest in computational acoustics (Hixon, 1997; Kim
et al., 1997; Lin an Chin, 1995). Table 3-1 summarizes several approaches proposed in
the literature.
For longer wavelengths, the formal order of accuracy is sufficient to indicate the
performance of a scheme. However, for shorter waves relative to the grid size, it is
known that the leading truncation error terms are not good indicators (Shyy, 1985; Shyy,
1997). To handle broad band waves, the idea of optimizing the scheme coefficients via
minimizing the truncation error associated with a particular range of wave numbers has
been used over the years by many researchers, e.g., Refs.(Hu et al., 1996; Stanescu and
64
Habashi, 1998; Nance et al., 1997; Wang and Sankar, 1999; Cheong and Lee, 2001;
Wang and Chen, 2001; Ashcroft and Zhang, 2003) . A successful approach is the
Dispersion-Relation-Preserving (DRP) finite difference scheme proposed by Tam and
coworkers (Tam and Webb, 1993; Tam et al., 1993). The basic idea in the DRP scheme is
to optimize coefficients to satisfactorily resolve short waves with respect to the
computational grid, namely, waves with wavelengths of 6-8∆x (defined as 6-8 points per
wave or PPW) or shorter. It maximizes the accuracy by matching the wave number and
frequency characteristics between the analytical and the numerical operators in the range
of resolvable scales. Recently, Ashcroft and Zhang (2003) have reported a strategy for
developing optimized prefactored compact (OPC) schemes, requiring smaller stencil
support than DRP. The prefactorization strategy splits the central implicit schemes into
forward and backward biased operators. Using Fourier analysis, they have shown that it
is possible to select the coefficients of the biased operators such that their dispersion
characteristics match those of the original central compact scheme. Hixon and Turkel
(1998) proved that the “prefactored scheme is equivalent to the initial compact scheme if:
i) the real components of forward and backward operators are equal to those at the
corresponding wavenumber of the original compact scheme; ii) the imaginary
components of the forward and backward operators are equal in magnitude and opposite
in sign”.
Both DRP and OPC schemes are originally designed based on the finite difference
approach. In order to satisfy the governing laws of the fluid physics, it can be
advantageous to adopt the finite volume approach (Udaykumar et al., 1999; Yang et al.,
1999; Udaykumar et al., 1999), which ensures that fluxes estimated from different sides
65
of the same surface are identical, i.e., no spurious source/sink is generated due to
numerical treatment. Such a requirement is particularly important when nonlinearity is
involved, as is typically the case in shock and turbulence aspects of the aeroacoustic
computations. Furthermore, a finite volume formulation can offer an easier framework to
handle the irregular geometry and moving boundaries. In this work, we extend the
concept embodied in the original, finite difference-based DRP scheme (which we call
DRP–fd) to a finite volume formulation (which we call DRP-fv). Similarly, for the OPC-
scheme, we extend the basic concepts of the original, finite difference-based OPC (OPC-
fd) scheme, to a finite volume formulation, called OPC-fv. Our overall goal is to develop
the finite volume version of DRP and OPC schemes into a cut-cell type of Cartesian-grid
computational technique that we have developed earlier for moving and complex
boundary computations (Yang et al., 1999; Udaykumar et al., 1999; Ye et al., 1999) to
treat aero-acoustic problems needed for engineering practices.
We present the finite volume formulation of both DRP and OPC schemes, and
assess both fd and fv versions of the DRP and OPC schemes, using well defined test
problems to facilitate systematic evaluations. Both linear and nonlinear wave equations
with different wavelengths and viscous effects are utilized for direct comparisons. In the
following, we first summarize the essence of the individual schemes, including
derivations, then present assessment of the test cases.
Numerical Schemes
We use the following one-dimensional wave equation to facilitate the development
and presentation of the concept and numerical procedures:
0u uct x
∂ ∂+ =
∂ ∂ (3.1)
66
The equation contains time and space derivative. In our work the space derivative
term is treated with DRP or OPC scheme and the time derivative by a Low –Dissipation
and Low-Dispersion Runge-Kutta (LDDRK) scheme (Hu et al., 1996).
The first scheme used for space derivative is DPR scheme. We present the finite
volume version of DRP, and the boundary treatment of the DRP.
The OPC scheme is the second method considered for the space derivative. The
finite difference procedure of the OPC scheme is offered. This presentation is follow by
the extension of this approach to a finite volume and the specific boundary treatment of
OPC .
LDDRK (Hu et al., 1996) scheme is used to approximate time derivative. The
principal characteristics will be presented
DRP Scheme
Finite volume-based DRP scheme (DRP-fv)
To incorporate the DRP-fd concept into a finite volume framework, let us consider
a one-dimensional linear wave equation:
0ct xφ φ∂ ∂
+ =∂ ∂
(3.2)
To derive the discredized equation, we employ the grid point cluster shown in
Figure 3-1. We focus on the grid point i, which has the grid points i -1, and i+1 as its
neighbors. The dashed lines define the control volume, and letters e and w denote east
and west faces, respectively, of the control volume. For the one-dimensional problem
under consideration, we assume a unit thickness in the y and z directions; thus, we obtain
( ) ( )( ) 0w
e we
dx c A Atφ φ φ∂
+ − =∂∫
(3.3)
67
where (Aφ)e and (Aφ)w are the flux across the east and west face, respectively.
Figure 3-1.Grid points cluster for one-dimensional problem
Hence, the discretized wave equation (3.2) can be written as
( ) ( )( ) 0e w
x c A Atφ φ φ∂
∆ + − =∂
(3.4)
where φ is the averaged value of φ over a control volume.
Taking into account Eq. (3.4) we describe the general form of the approximation of
xφ∂
∂ in 1-D using the control volume concept:
` ( ) ( )( )1e w
A Ax xφ φ φ∂
→ −∂ ∆
(3.5)
The general form of the DRP scheme is:
3
3
1 k
k l kkl
ax xφ φ
=
+=−
∂⎛ ⎞ ≅⎜ ⎟∂ ∆⎝ ⎠∑ (3.6)
where ∆x is the space grid, and coefficients aj are constant.
The DRP scheme has a general form similar to the central difference
approximation. Hence, one can adopt a central difference scheme to express φe in the
neighborhood:
( ) 1 2 2 1 3 4 1 5 2 6 3i i i i i ieφ β φ β φ β φ β φ β φ β φ− − + + += + + + + + (3.7)
( ) 1 3 2 2 3 1 4 5 1 6 2i i i i i iwφ β φ β φ β φ β φ β φ β φ− − − + += + + + + + (3.8)
i e i+1(W) i+2 i+3wi-1(E) i-2
∆x
(δx)w (δx)e
68
Taking into consideration Eqs. (3.5), (3.6), (3.7) and, (3.8) we obtain the values of
the βi, i = 1,…,6 by imposing that the value of φ at the same locations has the same
values as that of the DRP-fd.
3
3
k
k l k e wk
a φ φ φ=
+=−
= −∑ (3.9)
Hence, the values of coefficients β’s are
1 6 3
2 5 2 3
3 4 1 2 3
aa aa a a
β ββ ββ β
= =⎧⎪ = = +⎨⎪ = = + +⎩
(3.10)
To illustrate the above-described concept, we consider the following equation:
1 2 0u u uc ct x y
∂ ∂ ∂+ + =
∂ ∂ ∂ (3.11)
If we integrate Eq. (3.11) on the surface we have (see Figure 3-2):
,1
i jabcd
u Ft S
∂=
∂ (3.12)
where the resulting DRP-fv scheme is
, 1 , , , ,
2 , , , ,
( )
( )
s e n wi j i j s i j e i j n i j w
s e n wi j s i j e i j n i j w
F c u y u y u y u y
c u x u x u x u x
⎡= − ∆ + ∆ + ∆ + ∆ +⎣⎤∆ + ∆ + ∆ + ∆ ⎦
(3.13)
, 1 2, 2 1, 3 , 4 1, 5 2, 6 3,ei j i j i j i j i j i j i ju u u u u u uβ β β β β β− − + + += + + + + + (3.14)
, 1 3 2 2, 3 1, 4 , 5 1, 6 2,wi j i i j i j i j i j i ju u u u u u uβ β β β β β− − − + += + + + + + (3.15)
, 1 , 2 2 , 1 3 , 4 , 1 5 , 2 6 , 3ni j i j i j i j i j i j i ju u u u u u uβ β β β β β− − + + += + + + + + (3.16)
, 1 , 3 2 , 2 3 , 1 4 , 5 , 1 6 , 2si j i j i j i j i j i j i ju u u u u u uβ β β β β β− − − + += + + + + + (3.17)
69
( ) ( )
( )
1 2 1
2
abcd s s e e n n w wV S
s s e e n n w w
u udv c udy c udx S c u y u y u y u yt t
c u x u x u x u x
∂ ∂+ − = + ∆ + ∆ + ∆ + ∆ +
∂ ∂
+ − ∆ − ∆ − ∆ − ∆ +
∫ ∫ (3.18)
d c
e
s
w
nP
ba S
W E
Nd c
e
s
w
nP
ba S
W E
N
Figure 3-2. Grid notation for two-dimensional problem, where (i) P denotes the center of
a cell, (ii) E, W, N, and S denote, respectively, the nodes corresponding to the east, west, north and south neighbors, (iii) e, w, n and s denote, respectively, the center of the east, west, north and south face of the cell, and (iv) a, b, c, and d denote, respectively, the corners of the cell
Boundary treatment of the DRP scheme
The current version of the DRP scheme requires seven grid points in space.
Consequently, it is necessary to impose some supplementary condition for boundary
treatments. In this regard, Tam and Webb (1993) devise ghost points. The minimum
number of ghost points is equal to the number of boundary conditions. For example, for
an inviscid flow the condition of no flux through the wall requires a minimum of one
ghost value per boundary point on the wall. It is desirable to use a minimum number of
ghost points to maintain simplicity in coding and structuring data.
In this work we use only backward difference for grid points near the
computational boundary and a ghost point is used only for wall boundary condition.
70
OPC Scheme
Finite-difference-based optimized prefactored compact (OPC-fd) scheme
To derive the factorized compact scheme Ashcroft and Zhang (2003) define
forward and backward operators FiD and B
iD , such that
( )12
B Fi i
i
u D Dx
∂⎛ ⎞ = +⎜ ⎟∂⎝ ⎠ (3.19)
The generic stencil for the forward and backward derivative operators are then
defined as:
1 2 1 1 21F F F F F F F F F
i i i i i i iD D a u b u c u d u e ux
η β+ + + − −⎡ ⎤⋅ + ⋅ = ⋅ + + ⋅ + ⋅ +⎣ ⎦∆ (3.20)
1 2 1 1 21B B B B B B B B B
i i i i i i iD D a u b u c u d u e ux
β γ − + + − −⎡ ⎤⋅ + ⋅ = ⋅ + + ⋅ + ⋅ +⎣ ⎦∆ (3.21)
The coefficients of the scheme are chosen such that: i) the wavenumber of the
scheme is close to the important wavenumber of the exact solution; ii) the imaginary
components of the forward and backward stencil are equal in magnitude and opposite in
sign, and the real components are equal and identical to original compact scheme; iii) the
scheme preserves a certain order of accuracy. The authors (Ashcroft and Zhang, 2003)
define the integrated error (weighted deviation) as:
( ) ( )0
rE x x W x d x
πα α α α= ∆ − ∆ ∆ ∆∫ (3.22)
where W(α∆x) is a weighting function, and r is a factor to determine the optimization
range (0< r < 1). The integrated error, defined in Eq.(3.22), is different from the one of
Tam and Web (1993) in that it contains the weighting function. The coefficients are
obtained by imposing that, within a given asymptotic order, the error is minimal. In space
71
discretization, one sacrifices formal order of accuracy in favor of wide-band
performance, especially for the short wave components.
Finite volume-based OPC scheme (OPC-fv)
Taking into account Eq.(3.3) that describes the approximation of the first derivative
in the finite volume formulation, equations that describe the OPC scheme, Eq.(3.20) and
Eq.(3.21), and the idea that the general form of approximation of the function for points
at the center of the cell face, namely, e and w assumes similar forms :
0.5( )
0.5( )
Fe Bee
Fw Bww
u u u
u u u
⎧ = +⎪⎨
= +⎪⎩ (3.23)
where uFe, uBe, uFw and uBw are determined from:
1 1Fe Fei i i iu u bu duη β+ ++ = − (3.24)
1 1Fw Fwi i i iu u bu duη β+ −+ = − (3.25)
1 1Be Bei i i iu u bu duβ η − ++ = − (3.26)
1 1Bw Bwi i i iu u bu duβ η − −+ = − (3.27)
where the coefficients are the same as those in the OPC-fd scheme: η = ηF= γB,
β = βF= βB, b = bF= -dB, d= dF. = -bB. These relationships among forward and backward
operators are obtained by Ashcroft and Zhang (2003).
The boundary treatment of the OPC scheme
Boundary Formulation of the OPC scheme employs a biased explicit stencil.
Ashcroft and Zhang (2003) design OPC-fd scheme with the follow boundary stencil:
4
11 3
1 1,N
B Bj j N j j
j j N
D s u D e ux x= = −
= =∆ ∆∑ ∑ (3.28)
and
72
4
1 1 11 3
1 1,N
F FN j j N N j j
j j N
D e u D s ux x+ − + −
= = −
= − = −∆ ∆∑ ∑ (3.29)
where the coefficients sj and ej are determined by matching the Taylor series of the
forward and backward compact interior stencils to third-order accuracy.
The boundary treatment in case of OPC-fv approach is similar to that of OPC–fd,
but the boundary stencil is computed on the face:
Di = (uA)ie –(uA)i
w (3.30)
3 3
11 1
3 3
1 1 11 1
w wi i N i N i
i i
e ei i N i N i
i i
u a u u ru
u a u u ru
−= =
+ − += =
⎧ ⎧= =⎪ ⎪⎪ ⎪
⎨ ⎨⎪ ⎪= =⎪ ⎪⎩ ⎩
∑ ∑
∑ ∑
(3.31)
where the value of the coefficients are:
11 1
2 1 2 2 1
3 1 2 3 3 21
FBN
B FN N
B FN N N
a ea s
a s s a e e
a s s s a e e e−
−−
⎧⎧ == −⎪⎪
= − − = +⎨ ⎨⎪ ⎪= − − − = + −⎩ ⎩
(3.32)
1 1 1
2 2 1 21
3 1 2 33 21
B FN
B FN N
FBN N N
r e r s
r e e r s s
r s s sr e e e−
−−
⎧ ⎧= = −⎪ ⎪
= + = − −⎨ ⎨⎪ ⎪ = − − −= + − ⎩⎩
(3.33)
Time Discretization – The Low Dispersion and Dissipation Runge-Kutta (LDDRK) Method
Hu et al.(1996) consider time integration using the Runge-Kutta algorithm of the
differential equation
( )u F ut
∂=
∂ (3.34)
where the operator F is a function of u. An explicit p-stage algorithm advances the
solution of Equation (3.34) from the nth to the (n + 1)th iteration as
73
( )
( )( )
( )( )( )
( )
0
0(1)
1( )
( )
1
...
1,...,...
n
ii
i n ii
pn
u u
K tF u
K tF u
u u b K i p
u u
−
+
=
= ∆
= ∆
= + =
=
(3.35)
where
• bp = 1,
• u(p), where p indicates the stage in algorithm advances
• un+1, where n indicates the number of iterations for time dependent computation
The vale of the un+1 can be written on short like
1
1 1
j
j nppn n j
l jj l p j
uu u b tt
γ
+
= = − +
=
∂= + ∆
∂∑ ∏14243
(3.36)
The resulting algorithm is obtained by optimizing the dispersion and dissipation
properties. Assuming F(u) is linear and applying temporal Fourier transform to (3.36),
the amplification factor is given by
( )1
*
1
1n p j
jnj
ur i tu
γ ω+
=
⎛ ⎞= = + − ∆⎜ ⎟
⎝ ⎠∑%
% (3.37)
The exact amplification factor is
*i t i
er e eω σ− ∆ −= = (3.38)
The numerical amplification factor r in (3.37) is viewed as an approximation of the
exact factor. The order of the optimized Runge-Kutta scheme is indicated by the leading
74
coefficient in (3.37) that matches the Taylor series expansion of e-iσ. For instance, the
third order algorithm is obtained by setting γj = 1/j! for j = 1, 2 and 3.
To compare the numerical and exact solutions we take into consideration the ratio:
i
e
r r er
δ−= (3.39)
where |r| represents the dissipation rate (obviously, the correct value should be 1), and δ
represents the phase error (or dispersive error) where the correct value should be zero.
Hu et al (1996) obtained coefficients of the low dispersion and dissipation Runge-
Kutta (LDDRK) scheme by imposing that: i) the scheme has certain order of accuracy,
ii) the error of the amplification factor of the scheme is minimized, which means that
both dispersion and dissipation errors are minimized. In other words the following
integral is minimized:
( )2
01
1 minp
j ij
j
c i e dσσ σΓ −
=
+ − − =∑∫ (3.40)
and iii) the amplification factor of the scheme is less than 0ne within the given stability
limit
t
|r|-
1
0 1 2 3
-0.04
-0.02
0
0.02
0.04
LR
LLDDRK = 1.64RLDDDR = 2.52
∆
LDDRK
ω∗
ω ∆
δ
0 1 2 3
-0.04
-0.02
0
0.02
0.04
* t
LDDRK
LLDDRK = 1.85
L
a) b) Figure 3-3. Four-six–stage optimized Runge-Kutta of order four scheme: a) dissipation
error; b) phase error
75
In this work we use a two-step alternating scheme: in odd steps we use four stages
and in the even steps we use six stages. The scheme is a fourth-order accurate scheme in
time for a linear problem and second-order accurate for a nonlinear problem. The
advantage of the alternating schemes is that, when two steps are combined, the dispersion
and the dissipation errors can be reduced and higher order of accuracy can be maintained.
The specific procedure is given bellow.
• four-stage
( ) ( )( ) ( )
( ) ( )
( ) ( )
( )
1
2 1
3 2
4 3
41
141312
n
n
n
n
n n
K tF u
K tF u K
K tF u K
K tF u K
u u K+
= ∆
⎛ ⎞= ∆ +⎜ ⎟⎝ ⎠⎛ ⎞= ∆ +⎜ ⎟⎝ ⎠⎛ ⎞= ∆ +⎜ ⎟⎝ ⎠
= +
(3.41)
• six-stage
( )( )( )( )( )
( )
( )
( )
( )
(1)
1(2)
2(3)
3(4)
4(5)
5(6)
61
0.17667
0.38904
141312
n
n
n
n
n
n
n n
K tF u
K tF u K
K tF u K
K tF u K
K tF u K
K tF u K
u u K+
= ∆
= ∆ +
= ∆ +
⎛ ⎞= ∆ +⎜ ⎟⎝ ⎠⎛ ⎞= ∆ +⎜ ⎟⎝ ⎠⎛ ⎞= ∆ +⎜ ⎟⎝ ⎠
= +
(3.42)
76
In the follow we will give an implementation example of LDDRK scheme when we
use OPC and DRP scheme for space discretization. Base on Eq.(3.1), the value of F in
point l is defined as following:
• DRP – fd:
3
3l i l i
i
cF a ux +
=−
= −∆ ∑ (3.43)
• DRP-fv:
Fl = -c( ul e – ul w )/∆x (3.44) where
1 2 2 1 3 4 1 5 2 6 3el l l l l l lu u u u u u uβ β β β β β− − + + += + + + + + (3.45)
1 3 2 2 3 1 4 5 1 6 2wl l l l l l lu u u u u u uβ β β β β β− − − + += + + + + + (3.46)
In the linear case, the fv and fd schemes are equivalent.
• OPC-fd
( )2B F
l l lcF D D= − + (3.47)
where DlB and Dl
F are obtained from the following system of equations:
( ) ( )1 1 11 , 1,...,F F
i i i i i iD D b u u d u u i Nx
η β+ + −+ = − + − =⎡ ⎤⎣ ⎦∆ (3.48)
( ) ( )1 1 11 , 1,...,B B
i i i i i iD D d u u b u u i Nx
β η − + −+ = − + − =⎡ ⎤⎣ ⎦∆ (3.49)
where N represent the number of grid points in space.
• OPC-fv Fl = - c( ul
e – ul w )/∆x (3.50)
where ul
e = 0.5(ul Be + ul
Fe), and ulw = 0.5(ul
Bw + ulFw) (3.51)
The value of ulBe, ul
Fe, ulBw, ul
Fw are obtained by solving the follow system of equations 1 1 1,...,Fe Fe
i i i iu u bu du i Nη β+ ++ = − = (3.52)
1 1 1,...,Fw Fwi i i iu u bu du i Nη β+ −+ = − = (3.53)
1 1 1,...,Be Bei i i iu u bu du i Nβ η − ++ = − = (3.54)
1 1 1,...,Bw Bwi i i iu u bu du i Nβ η − −+ = − = (3.55)
77
Analytical Assessment of DRP and OPC Schemes
Operation Count
We will compare the cost between the alternative approaches only for the
approximation of the first derivative, because we employ the same time stepping scheme
for both scheme.
The efficient form of general formula for the discretization in space of the DRP–fd
scheme is
( ) ( ) ( )3 3 3 2 2 2 1 1 11
i i i i i i iF a x x a x x a x xx + − + − + −= − + − + −⎡ ⎤⎣ ⎦∆
(3.56)
This scheme requires a total of three multiplications and five additions to evaluate
the first derivatives in a certain point. In case of DRP-fv the most efficient form of the
computations scheme is
( ) ( ) ( )1 3 2 2 2 1 3 11
e i i i i i iu u u u u u ux
β β β+ − + − += + + + + +⎡ ⎤⎣ ⎦∆ (3.57)
( ) ( ) ( )1 3 2 2 2 1 3 11
w i i i i i iu u u u u u ux
β β β− + − + −= + + + + +⎡ ⎤⎣ ⎦∆ (3.58)
DRP–fv requires a greater number of operations than DRP-fd: six multiplications
and eleven additions to compute the first derivatives at a given point.
To see the computational cost of the OPC-fd scheme we adopt the most efficient
form that is
( ) ( )1 1 1
1 12 2 2
F Fi i i i i iD b u u d u u D
xη
β β+ − +⎡ ⎤= − + − −⎣ ⎦∆ (3.59)
( ) ( )1 1 1
1 12 2 2
B Bi i i i i iD b u u d u u D
xη
β β− + −⎡ ⎤= − + − −⎣ ⎦∆
(3.60)
78
where the relation between the coefficients of the forward and backward stencils have
been substituted to highlight the equivalent terms in the two stencils. The operation count
is then four multiplications and five additions per point (Ashcroft and Zhang, 2003).
OPC-fv can be written in the form:
( )1 12 2
F Fe Fwi i iD u u
x= −
∆ (3.61)
( )1 12 2
B Be Bwi i iD u u
x= −
∆ (3.62)
where
1 1
1 1
1
1
Fe Fei i i i
F
Fw Fwi i i i
u bu du u
u bu du u
ηβ
ηβ
+ +
− +
⎧ ⎡ ⎤= − −⎪ ⎣ ⎦⎪⎨⎪ ⎡ ⎤= − −⎣ ⎦⎪⎩
(3.63)
1 1
1 1
1
1
Be Bei i i i
Bw Bei i i i
u bu du u
u bu du u
ηβ
ηβ
+ −
− −
⎧ ⎡ ⎤= − −⎣ ⎦⎪⎪⎨⎪ ⎡ ⎤= − −⎣ ⎦⎪⎩
(3.64)
In this case the operation count is eleven additions and six multiplications per point.
So we can see also, in Table 3-2 the finite volume approach is computationally
more expensive.
Dispersion Characteristics
The characteristics of the OPC and DRP schemes, in the finite difference form over
the interval 0 to π, are shown in
Figure 3-4. One can see that the difference between the effective wave number of
the scheme and the real wave is maintained to be within 2% if α∆x < 1.30 for the DRP
scheme, and α∆x < 1.84 for the OPC scheme. The dispersive characteristics of these
79
schemes can be more clearly seen in Figure 3-5, which shows the phase speed error,
1d xabsd x
αα
∆⎛ ⎞−⎜ ⎟∆⎝ ⎠, as a function of wave number on a log-arithmetic scale. We see that the
DRP scheme has a somewhat larger error than the OPC scheme until around 3π/4. The
error is maintained to be within 2% for α∆x less than 0.85 for the DRP scheme, and less
than 1.53 for the OPC scheme. Overall, the OPC scheme yields slightly less dispersion
error than the DRP scheme.
The dispersive characteristics of LDDRK are obtained by studying the value of |r|
and δ, i.e., dissipation rate and dispersion error (see Eq.(3.39)), respectively. In Figure 3-
3 we can see conditions of stability: |r| < 1 for ω*∆t ≤ 2.52 .To obtain an accurate solution
the dispersive characteristics (|r| and δ) should be close to the exact solution (|r| close to
one and δ close to zero). Hu et al. (1996) considered time accurate criterion ||r| -1| ≤ 0.001
(i.e ω*∆t ≤ 1.64), and δ ≤ 0.001 (i.e., ω*∆t ≤ 1.85). These two conditions are satisfied if
ω*∆t ≤ 1.64.
Stability
The Fourier-Laplace transformation of the wave equation (Eq.(3.1)) is
% % %12
initiali u ci u uω απ
∗− = − + (3.65)
where α, ω* characterize the PDE . For the long wave we can approximate wavenumber
of the scheme with wavenumber of the PDE
α ≅ α (3.66)
which leads to
% % %*initialu c u kuω α= + (3.67)
80
Hence
* c kkω α= + (3.68)
0 0.5 1 1.5 2 2.5 3 3.50
0.5
1
1.5
2
2.5
3
3.5
IdealDRPOPC
Figure 3-4. Dispersive characteristics of the schemes:. xα∆ versus α∆x
0 0.5 1 1.5 2 2.5 3 3.5
10-4
10-3
10-2
10-1
100
101
α∆x
DRP OPC
Figure 3-5. Phase speed error on a logarithmic scale
The condition of the numerical stability is that amplification factor for time
discretization is less than 1, and hence ω*∆t ≤ 2.52 (see Figure 3-3a). It is also noted from
Figure 3-4 that
1.8 for DRP scheme2.1 for OPC scheme
xx
αα
∆ ≤∆ ≤
(3.69)
Phas
e sp
eed
erro
r
81
hold true. By introducing Eq. (3.69) into (3.68) and upon multiplying by ∆t it is found
that
[ ]
[ ]
1.8 M 1 for DRP scheme
2.1 M 1 for OPC scheme
ct kk tx
ct kk tx
ω
ω
∗
∗
∆ ≤ ⋅ + ⋅∆∆
∆ ≤ ⋅ + ⋅∆∆
(3.70)
where M is mach number. From Figure 3-3a it is clear that the condition of stability is
satisfied if |ω*∆t| is less than 2.52. Therefore to ensure numerical stability it is sufficient
by Eq. (3.70) to restrict ∆t to be less than ∆tmax, where ∆tmax is given by
[ ]
[ ]
max
max
2.52 for DRP scheme1.8 M 1
2.52 for OPC scheme2.1 M 1
xtkk c
xtkk c
∆∆ =
⋅ +
∆∆ =
⋅ +
(3.71)
Therefore, for ∆t < ∆tmax the schemes are numerically stable. Consequently, the
schemes yield the following criteria for numerical stability:
1.4 for DRP scheme1.2 for OPC scheme
CFLCFL
<<
(3.72)
Although it is clear that CFL≤1.4 is the stability condition for DRP scheme, this
limit does not assure accuracy of the solution. In the previous analysis we have
established that the solution is time accurate for 4-6 LDRRK if ||r|-1| ≤ 0.001 and | δ | ≤
0.001. But this limit is not fixed, but depends on the scheme that is used for space
discretization. For example, in the case of the DRP scheme, the solution is considered
time accurate as long as ||r|-1| ≤ 0.02 and | δ | ≤ 0.02, or ω*∆t ≤2.0. Hence, in this case the
condition of being both accurate and stable is CFL ≤ 1.1
82
The OPC scheme is less sensitive to the dispersive characteristics of the LDDRK
scheme; hence CFL<1.2 is a condition of the stability and accuracy for the OPC scheme.
This limit is in concordance with the stability analysis of Ashcroft and Zhang (2003).
Computational Assessment of the DRP and OPC Schemes
To investigate the behavior of the schemes, we will use four test problems. First,
we consider a one-dimensional wave equation with constant speed. The purpose of this
test is to check the accuracy, stability, dissipation and dispersion of the scheme. The
second test problem is one-dimensional nonlinear wave equation with no viscous
dissipation. The purpose of this test case is to i) check the influence of singularities on the
performance of the scheme, and ii) analyze dispersion properties when waves are
coupled. In the third test problem, we consider the one-dimensional viscous Burgers
equation, which contains unsteady, nonlinear convection and viscous terms. In this case
we pay attention to the influence of the viscosity on the solution accuracy. The last test
problem is a 2D acoustic scattering problem from the second CAA Workshop (Tam and
Hardin, 1997). This problem tests the curved wall boundary and the capacity of the
scheme to reproduce different wavelengths.
Test problem 1: One-Dimensional Linear Wave Equation
To assess the behavior of the DRP and OPC schemes the following simple test
problem is studied first.
0.u uct x
∂ ∂+ =
∂ ∂ (3.73)
2
0exp ln 2 x xur
⎡ ⎤−⎛ ⎞= −⎢ ⎥⎜ ⎟⎝ ⎠⎢ ⎥⎣ ⎦
; at t = 0 (3.74)
83
which is a Gaussian profile. This is one of the test problems offered in the second CAA
Workshop (Tam and Hardin 1997)]
The exact solution is
2
0exp ln 2 x x ctUr
⎡ ⎤− −⎛ ⎞= −⎢ ⎥⎜ ⎟⎝ ⎠⎢ ⎥⎣ ⎦
(3.75)
In this study we evaluate the performance of the schemes in short, intermediate, and long
waves relative to the grid spacing, which is assured by the value r/∆x.
For time discretization, we previously presented the detailed formulas for the 4-6
LDDRK, see equations (3.41), and (3.42).
Tam (Tam and Webb,1993; Tam et al., 1993) show that xα∆ is related to α∆x, and
in function of α∆x they divided the wave spectrum into two categories; i) the long waves
(waves for which xα∆ , in this case α∆x is less than α∆xc, ii) the short waves (waves for
which α is not close to α ). This difference between long and short waves is totally
dependent upon the grid space. Hence, by inspecting the number of grid points on the
wavelength, we can decide that we have a certain category of wave.
In the following, we will present the results based on three categories
• long wave (r/∆x = 20)
• intermediate wave (r/∆x = 6)
• short wave (r/∆x = 3)
The categories are defined according to the ratio r/∆x, where r is a parameter that
characterizes the wavelength of this problem. This test problem is linear; hence we do not
expect differences between finite difference and finite volume approach.
84
In regard to the time step selection, the CFL number (ν) limit is similar for all
schemes. We can see from Figure 3-6 that the critical CFL number of both schemes is
close to 1.1. From the study of the error in time for linear equations with constant
convection speed it is clear that the DRP-fv and OPC-fv schemes have essentially the
same behavior as the corresponding finite difference approach; hence, we only present
comparison for DRP–fv and OPC-fv schemes.
The error decreases when the grid size in space decreases until a critical value is
reached. For all schemes the errors have slopes consistent with the formal order of
accuracy in space. This conclusion is confirmed in Figure 3-7, where the CFL number is
maintained at 0.5. For the long time scale solution, the accumulation of error for both
DRP and OPC schemes is very close (as seen in Figure 3-7b). Here, we consider:
i) different grid space, so both schemes have almost the same initial error, and ii) the
same CFL number (0.5). This behavior is expected because both schemes present the
same discretization in time.
t
ER
RO
R
0.1 0.20.30.410-4
10-3
10-2
∆
DRP - fdDRP - fvOPC - fdOPC - fv
t
ER
RO
R
0.05 0.1
10-5
10-4
10-3
∆
DRP - fdDRP - fvOPC - fdOPC - fv
a) r/∆x = 3 (short wave); b) r/∆x = 10 – long wave
Figure 3-6. Errors with respect to the time step size under a fixed space ∆x, at t = 50 - linear wave equation
85
x
ER
RO
R
10-2 10-1 100
10-7
10-6
10-5
10-4
10-3
10-2
∆
DRP-fvDRP-fdOPC-fvOPC-fd
4
1
t
ER
RO
R
1000 2000
0.001
0.002
0.0030.004
1
1
DRPOPC
a) b)
Figure 3-7. Errors under a fixed CFL = 0.5, at t = 50 - linear wave equation: a) error with respect to the space size; b) accumulation of the error in time
Test problem 2: One-Dimensional Nonlinear Wave Equation
The finite volume and finite difference schemes are equivalent for a linear
equation. The difference between them appears for the nonlinear convective equation. To
observe the merits and similarities of DRP, and OPC schemes, we restrict ourselves to the
1-D case. In this test, a nonlinear wave equation with a different speed is solved
0u uut x
∂ ∂+ =
∂ ∂ (3.76)
This equation is solved in the conservative form
( )2
0.5 0uu
t x
∂∂+ =
∂ ∂ (3.77)
To better understand the effect of high gradients and discontinuities, we chose the
following initial conditions
0 0
( ,0)1 0
xu x
x≤⎧
= ⎨ >⎩ (3.78)
The solution for this problem can be written
86
0 0
( , ) 0
1
xxu x t x tt
x t
≤⎧⎪⎪= < <⎨⎪
≥⎪⎩
(3.79)
In this case, for both DRP and OPC schemes, the finite difference version behaves
differently from the finite volume version. In the Eqs.(3.41) and (3.42) the function F
takes the form:
• DRP-fd
( )23
30.5i k i k
kF a u +
=−
= − ∑ (3.80)
• DRP-fv
( ) ( )( )2 20.5 e w
i i iF u u= − − (3.81)
where ue and uw are as defined before
• OPC-fd ( )0.25 B F
i i iF D D= − +
(3.82) where Di
B and DiF is backward and forward derivative of u2 in place of u
• OPC-fv
( ) ( )( )220.25
i
e wi iF u u= − − (3.83)
where ue and uw are defined by (3.52) - (3.55)
The similarities and differences for all three categories (short waves [∆x/U = 1.0],
intermediate waves [∆x/U = 0.25], and long waves [∆x/U = 0.06]) are first presented. It
should be noted again that the short, intermediate and long waves are defined based on
the numerical resolution. Here, U is defined as the jump (umax - umin); in our case U = 1,
hence in the following we discuss only the effect of the grid space step (∆x).
87
∆
ER
RO
R
10-2 10-1 100
10-3
10-2
DRP - fvDRP - fdOPC - fvOPC - fd
x
21
Figure 3-8. Errors with respect to the space step size under a fixed CFL = 0.5, at t = 5;
nonlinear wave equation
The evolution of the error as a function of grid spacing (∆x) is similar for both DRP
and OPC schemes; the difference between the finite volume and finite difference versions
are far greater, as shown in Figure 3-8. In the case of finite volume, error decreases with
decreasing grid space (see Figure 3-10 and Figure 3-12). For finite difference, a totally
different behavior is seen. The error not only does not decrease when grid spacing
decreases, but in fact increases, as seen in Figure 3-9 and Figure 3-11).
X
U
-4 -2 0 2 4 6 8 10
-0.2
0
0.2
0.4
0.6
0.8
1
DRP - fdExact
X = 1.∆
X
U
-4 -2 0 2 4 6 8 10
-0.5
-0.25
0
0.25
0.5
0.75
1
DRP - fdExact
X = 0.25∆
X
U
-4 -2 0 2 4 6 8 10-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
DRP - fdExact
X = 0.06∆
a)∆x = 1 b) ∆x = 0.25 c) ∆x = 0.06 Figure 3-9. DRP–fd solution - nonlinear wave equation; t = 5; CFL = 0.5
88
X
U
-4 -2 0 2 4 6 8 100
0.25
0.5
0.75
1
DRP - fvExact
X = 1.∆
X
U
-4 -2 0 2 4 6 8 100
0.25
0.5
0.75
1
DRP - fvExact
X = 0.25∆
X
U
-4 -2 0 2 4 6 8 100
0.25
0.5
0.75
1
DRP - fvExact
X = 0.06∆
a) x = 1 b) -∆x = 0.25 c) ∆x = 0.06 Figure 3-10. DRP–fv solution - nonlinear wave equation; t = 3; CFL = 0.5
For short waves, all solutions show substantial errors, but the finite difference
schemes perform noticeably worse. In the case of intermediate or long waves, the finite
volume schemes exhibit satisfactory or better performance than the finite difference
schemes.
X
U
-4 -2 0 2 4 6 8 10-0.25
0
0.25
0.5
0.75
1
OPC - fdExact
X = 1.∆
X
U
-4 -2 0 2 4 6 8 10-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
OPC - fdExact
X = 0.25∆
X
U
-4 -2 0 2 4 6 8 10-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
OPC - fdExact
X = 0.06∆
a ) ∆x = 1 b) ∆x = 0.25 c) ∆x = 0.06 Figure 3-11. OPC-fd solution - nonlinear wave equation; t = 5; CFL = 0.5
X
U
-4 -2 0 2 4 6 8 100
0.25
0.5
0.75
1
OPC - fvExact
X = 1.∆
X
U
-4 -2 0 2 4 6 8 100
0.25
0.5
0.75
1
OPC - fvExact
X = 0.25∆
X
U
-4 -2 0 2 4 6 8 100
0.25
0.5
0.75
1
OPC - fvExact
X = 0.06∆
a) ∆x = 1 c) ∆x = 0.25 d) ∆x = 0.06 Figure 3-12. OPC-fv solution - nonlinear wave equation; t = 5; CFL = 0.5
89
Test problem 3: One-Dimensional Nonlinear Burgers Equation
In this test the solution for the one-dimensional nonlinear Burgers equation is
evaluated.
2
2
u u uut x x
µ∂ ∂ ∂+ =
∂ ∂ ∂ (3.84)
The numerical solution will approach equation (3.84) in conservative form;
2 2
20.5u u ut x x
µ∂ ∂ ∂+ =
∂ ∂ ∂ (3.85)
The initial condition is:
( ) 0,0 1 tanh2
x xu xµ
⎛ ⎞−= − ⎜ ⎟
⎝ ⎠ (3.86)
In this case the exact solution is:
( ) 0, 1 tanh2
x x tu x tµ
⎛ ⎞− −= − ⎜ ⎟
⎝ ⎠ (3.87)
The scheme described earlier for inviscid Burgers’ equation can also be applied to
the current equation. This is accomplished by simply adding a second-order central-
difference expression for the viscous term uxx. In other words Fi is replaced by Hi
1 1( 2 ) /i i i i iH F u u u xµ − −= + − + ∆ (3.88)
Because of the viscosity that characterizes the scheme in this case, it is expected
that the solution of both approaches would be stable and similar. Hence this term will
have a large influence over the value of the error.
In our discussion, we will distinguish the following three categories of results:
• short wave (∆x/µ = 10)
• intermediate wave (∆x/µ = 3 )
90
• long wave (∆x/µ = 1)
In this case the numerical performance is affected by two parameters: the CFL
number and the Peclet number (Pe = U∆x /µ).
First we compare the solution of all four schemes as function of the Peclet number
(Pe) under constant CFL number (0.2). The value of CFL number is fixed at 0.2, because
the critical value for all schemes is much lower in the present case than for the linear
case. The behavior of the error is similar among DRP-fv, DRP-fd, OPC-fd and OPC-fv:
the error increases with increasing Peclet number, until a certain value beyond which the
schemes can no longer perform satisfactorily (see Figure 3-13).
Pe
ER
RO
R
100 10110-4
10-3
10-2
10-1 DRP - fvDRP - fdOPC - fvOPC - fd
**
**
***
Figure 3-13. Error as a function of Pe - nonlinear Burgers equation; ∆x = 0.25;
CFL = 0.2; t = 20
For the four schemes (DRP-fv, DRP-fd, and OPC-fv, OPC-fd), the solution and
error are very similar for all categories of wave, as shown in Figure 3-13, Figure 3-14 and
Figure 3-15. For long waves the solution is reproduced with high accuracy with all four
schemes, but the finite volume approach presents a slightly higher accuracy than the
finite difference schemes. The error for the intermediate wave is nearly the same with all
four approaches.
91
X
U
16 18 20 220
0.5
1
1.5
2
2.5
DRP - fvDRP - fdExact
X
U
16 18 20 220
0.5
1
1.5
2DRP - fvDRP - fdExact
X
U
16 18 20 220
0.5
1
1.5
2DRP - fvDRP - fdExact
a) Pe = 10 b) Pe = 3 c) Pe = 1 Figure 3-14. Numerical solution obtained by DRP schemes - nonlinear Burgers quation;
∆x = 0.25; CFL = 0.2; t = 20
X
U
16 18 20 220
0.5
1
1.5
2
2.5
DRP - fvDRP - fdExact
X
U
16 18 20 220
0.5
1
1.5
2OPC - fvOPC - fdExact
X
U
16 18 20 220
0.5
1
1.5
2
OPC - fvOPC - fdExact
a) Pe = 10 b) Pe = 3 c) Pe = 1 Figure 3-15. Numerical solution obtained by OPC schemes - nonlinear Burgers equation;
∆x = 0.25; CFL = 0.2; t = 20
Test problem 4: Two-Dimensional Acoustic Scattering Problem
To check the accuracy of the finite volume schemes in multi-dimensional
situations, we consider a test problem from the Second CAA Workshop (Tam and
Webb,1993; Tam et al., 1993): the two-dimensional acoustic scattering problem. The
physical problem is to find the sound field generated by a propeller scattered off by the
fuselage of an aircraft. The pressure loading on the fuselage is an input to the interior
noise problem. The fuselage is idealized by a circular cylinder and the noise source
(propeller) as a line source so that the computational problem is two-dimensional. The
cylinder has a radius of R = 0.5 and is located at the center of the domain.
The linearized Euler equations in polar coordinates are:
92
0 0 0
1 10 0 00
r
r r
u pu p
t r r ru u up
θ
θ
θ
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤∂ ∂ ∂⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥+ + + =⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥∂ ∂ ∂
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎣ ⎦
(3.89)
At time t = 0, the initial conditions are:
ur = uθ = 0 (3.90)
( )2 24( , ,0) exp ln 2
0.04x y
p x y⎡ ⎤⎛ ⎞− +⎢ ⎥= − ⎜ ⎟
⎜ ⎟⎢ ⎥⎝ ⎠⎣ ⎦ (3.91)
The test problem asks for the unsteady pressure time history at three points A(r = 5,
θ = 900), B(r =5, θ = 1350) and C(r = 5, θ = 1800), over the interval t = 5 → 10.
The numerical computations were performed over the domain: R∈ [0.5, 10.5] and θ
∈ [0, 2π]. For this problem three kinds of the boundary conditions are needed:
• Wall condition on the wall of the cylinder at R = 0.5
• Periodic condition along both azimuthally boundaries at θ = 0 and θ = 2π
Outfield boundary condition, along of the far field boundary, is the acoustic
radiation of Bayliss and Turkel (1982)].
The wall condition is based on the wall condition of Tam and Dong (1994). This
requires that:
0rdv dpdt dr
= − = (3.92)
This condition is satisfied by imposing the pressure derivatives on the wall to be
zero, and vr = 0 on the wall.
93
X
Y
-10 -5 0 5 10-10
-5
0
5
10
X
Y
-10 -5 0 5 10-10
-5
0
5
10
a) DRP-fv b)OPC-fv Figure 3-16. Instantaneous pressure contours at time t = 7 – two-dimensional acoustic
scattering problem
x
Y
-10 0 10-10
-5
0
5
10
AB
C
time
P
6 8 10
-0.02
0
0.02
0.04
0.06 OPC - fvDRP - fvExact
a) position of the testing points b) A: R = 5 ,θ = 900
time
p
6 8 10
-0.02
0
0.02
0.04
0.06OPC - fvDRP - fvExact
time
P
6 8 10-0.02
-0.01
0
0.01
0.02
0.03
0.04 OPC - fvDRP - fvExact
c) θ = 1350 d) θ = 900 Figure 3-17. The pressure history at point A, B and C – two-dimensional acoustic
scattering problem: finite volume approach
94
For this calculation, a uniformly spaced grid of 101 radial points and 153
azimuthally points was used, with a time step of CFL = 0.5. Figure 3-16 shows an
instantaneous pressure at t = 7. In this figure, the acoustic pulse is reflected by the
cylinder and reaches the outer boundary. We can see that two transients are shown: the
first and larger transient travels directly from the source; the second and smaller transient
is reflected from the cylinder. Both schemes reproduce both transients with acceptable
accuracy. Figure 3-17 compares the solution given by the fourth order schemes: DRP-fv
and OPC-fv. Between the two schemes the OPC-fv scheme performs better.
Summary and Conclusions
The DRP and OPC schemes, originally proposed in the finite difference form, have
been assessed. To better handle nonlinearity and geometric complexities, the finite
volume versions of both schemes have also been developed. Linear and nonlinear wave
equations, with and without viscous dissipation, have been adopted as the test problems.
For the linear wave equation with constant convection speed, the numerical
stability bound posed by the CFL number is comparable between the DRP and OPC
schemes. Both OPC and DRP produce solutions of a comparable order of accuracy, but
the magnitude of the error of the OPC scheme is lower.
For the nonlinear wave equation, the finite volume schemes can produce noticeably
better solutions and can handle the discontinuity or large gradients more satisfactorily.
However, as expected, all schemes have difficulties when there is insufficient mesh
resolution, as reflected in some of the short wave cases.
In conclusion, the finite volume version of both DRP and OPC schemes improve
the capabilities of the original version of the finite difference formulas in regard to
nonlinearity and high gradients. They can enhance performance of the original DRP and
95
OPC schemes for many wave propagation problems encountered in engineering
applications.
96
Table 3-1. A summary of proposed CAA algorithms Scheme The philosophy of the scheme Applications DRP (Tam and Webb, 1993; Tam et al., 1993)
In this scheme a central difference is employed to approximate first derivative. The coefficients are optimized to minimize a particular type of error
Wave propagation
LDDRK (Hu et al., 1996 ; Stanescu et al., 1998)
Traditionally, the coefficients of the Runge-Kutta scheme are optimized to minimize the dissipation and propagation waves.
Wave propagation problem
LDFV (Nance et al, 1997; Wang et al., 1999)
Scheme minimizes the numerical dispersion errors that arise in modeling convection phenomena, while keeping dissipation errors small: special high-order polynomials that interpolate the properties at the cell centers to the left and right sides of cell faces. A low pass filter has been implemented
Shock noise prediction
GODPR (Cheong and Lee, 2001)
It is derived, based on optimization that gives finite difference equations locally the same dispersion relation as the original partial differential equations on the grid points in the nonuniform Cartesian or curvilinear mesh
- Acoustic radiation from an oscillating circular cylinder in a wall - Scattering of acoustic pulse from a cylinder
OWENO (Wang and Chen, 2001)
The idea is to optimize WENO in wave number schemes, following the practice of DRP scheme to achieve high resolution for short wave. But in the same time it retains the advantage of WENO scheme in that discontinuity are captured without extra numerical damping.
Simulation of the shock/broadband acoustic wave
CE/SE (Chang, 1995; Chang et al., 1999)
The method is developed by imposing that: (i) space and time to be unified and treated as a single entity; (ii) local and global flux conservation in space and time to be enforced;
Flow involving shock; acoustic wave
FDo, RKo (Bogey and Bailly, 2004)
Optimized schemes are obtained by similar approach as DRP (space discretization), respectively LDDRK (time discretization). The difference consists that: i) error is minimized taking into account logarithm of the wavenumber; ii) the error is minimized on an interval that starts from ln(π/16). The stability and accuracy increase for these schemes
a)convective wave equation b)subsonic flows past rectangular open cavities; c) circular jet
97
Table 3-2. The computational cost for DRP and OPC schemes Scheme Number of
operation DRP-fd 8 DRP-fv 17 OPC-fd 9 OPC-fv 17
98
CHAPTER 4 A FINITE VOLUME-BASED HIGH ORDER CARTESIAN CUT-CELL METHOD
FOR COMPUTATIONAL AEROACOUSTICS
A finite volume-based high-order scheme with optimized dispersion and dissipation
characteristics in cooperation with the Cartesian cut-cell technique is presented for
aeroacoustic computations involving geometric complexities and nonlinearities. The field
equation is solved based on an optimized prefactored compact finite volume (OPC-fv)
scheme. The cut-cell approach handles the boundary shape by sub-dividing the
computational cells in accordance with the local geometric characteristics and facilitates
the use of numerical procedures with a desirable level of accuracy. The resulting
technique is assessed by several test problems that demonstrate satisfactory performance.
Introduction
In computational aero-acoustics (CAA), one needs to resolve dispersion and
dissipation characteristics in order to preserve form and amplitude of the wave (Hardin
and Hussaini (1992), Tam and coworkers (Tam and Webb, 1993; Tam et al., 1993)).
Furthermore, special care needs to be exercised in treating the boundaries to prevent the
creation of spurious waves while still accounting for wave reflection and/or scattering. In
order to meet these expectations, there are specific issues and challenges associated with
the employment of a structured multiblock grid, an overlapping grid, or an unstructured
grid ((Delfs, 2001), (Shyy et al., 2001), (Henshaw, 2004), (Sherer, 2004), (Dumbser,
2003), (Basel and Grünewald, 2003)), other alternatives should also be pursued. In this
work, we present an approach utilizing the Cartesian cut-cell approach based on the
99
finite-volume - based scheme aimed at optimizing the dispersion and dissipation
treatment. The Cartesian, cut-cell approach has been developed extensively for moving
and complex boundary computations ( see Table 4-1). It uses a background Cartesian grid
for the majority of the flow domain while creating sub-cells in the boundary regions to
meet the geometric requirements. Special algorithms for complex geometries are
constructed using flow properties and appropriate interpolation and extrapolation
procedures.
The low dispersion and dissipation scheme is based on an extension of the finite
difference-based, optimized prefactored compact (OPC) scheme originally developed by
Ashcroft and Zhang (2003) that we call OPC-fd. In the previous chapter we present the
extension of this method using the finite volume concept that we call OPC-fv, creating a
scheme to better satisfy the nonlinearity and conservation laws.
The present technique (which consists of the optimized dispersion and dissipation
characteristics and the cut-cell technique) offers desirable capabilities in both interior and
boundary regions. Several test problems will be presented to highlight the performance
characteristics of the proposed approach.
Cut-Cell Procedure
The cut-cell method rearranges the computational cells in the vicinity of the
interface via sub-division to conform to the specified boundary shape. Depending on the
intersection between the grid line and the interface, the subdivided, or cut, cells can
remain independent or can be merged into a neighboring cell in a given direction, e.g., a
direction approximately normal to the solid face. Accordingly, the interfacial cells are
reorganized along with their neighboring cells to form new cells with triangular,
trapezoidal, or pentagonal shapes.
100
The flux across the cell boundaries can be approximated by
1
m
k kkS
f n f n=
⋅ ≈ ∑∫r (4.1)
where S
f n⋅∫r is contour integral on the closed path S
The flux on the face is computed based on the multi-dimensional interpolation
method (Ye et al., 1999; Ye, 2001). Since the OPC scheme considered here is fourth
order, it is desirable to preserve the same order of accuracy around the boundary.
Figure 4-1. Illustration of the interfacial cells and cut-and-absorption procedures
The Cartesian cut cell mesh approach follows the subsequent steps:
• Locate the intersection of the boundary using a Cartesian mesh.
• Construct the background Cartesian mesh: The cells are flagged as solid cells, flow cells, or boundary cells. The boundary cells are those that either intersect the boundary or have a face in common with the boundary.
• Determine the geometric characteristics of the boundary cells such as cell volume, the direction normal to the boundary, and other information.
• Merge cells as necessary. A minimum acceptable cell area Smin is specified, and when the area of a cut cell is smaller than this value, it is merged into a neighboring cell. Determine the new characteristics of the merged cells.
interface interface
domain 1
domain 2 domain 2
domain 1
101
A
B
C
E
D
2928272625
19 222120 23
31 3332
37
35
34
13 161514 17
7 1098 11
1 432 5 6
12
18
2430
36
A
B
C
E
D
2928272625
19 222120 23
31 3332
37
35
34
13 161514 17
7 1098 11
1 432 5 6
12
18
2430
36
B D
A E
Cfsw
fefw
G
(xi,yj)
B D
A E
Cfsw
fefw
G
(xi,yj)
a) b) Figure 4-2. Modified cut – cell approach for CAA: a) Cartesian cut cell approach;
b) Detail around of the cut cell
For the trapezoid ACDE from Figure 4-2, the finite volume approach can be used
to approximate the wave equation:
0ABCDE
u E F dvt x y
⎛ ⎞∂ ∂ ∂+ + =⎜ ⎟∂ ∂ ∂⎝ ⎠
∫ (4.2)
where dv is a volume element. Appling Stokes’s theorem to Eq.(4.2), we have
∂uc/∂t SABCDE + ( )ABCDE
Edy Fdx−∫ (4.3)
The value of the integral can be approximated by
( ) ( )ABCDE AC CD DE EA
Edy Fdx Edy Edy Fdx Edy Fdx− = + − + −∫ ∫ ∫ ∫ ∫ (4.4)
where ( )ABCDE
Edy Fdx−∫ is contour integral on closed path ABCDE.
The function F and E can be represented generically by the function f. The fluxes
on the faces AC and DE can be approximate by
102
AC AB BC
fdy fdy fdy= +∫ ∫ ∫ (4.5)
An approximation of the value of the flux is given by:
( ) ( )w A B sw B CAC
fdy f y y f y y≈ − + −∫ (4.6)
The value of the flux at point w is given by the specific formula for the boundary
cell. The value of fsw is approximated using a polynomial that is fourth order in the x and
y direction:
4 4
0 0
i jsw ij
i jf b x y
= =
= ∑∑ (4.7)
where the coefficients bij are unknown. This interpolation has fourth-order accuracy in
the evaluation of the flux on the cut cells. In this case, the value of the coefficients is
obtained using the values of f in 25 grid points. An example is given in Figure 4-2 where
the value of the function is approximated using the following 25 points.
To solve bij , we use the following system of equations by expressing the function f
at 25 locations:
4 4 3 4 4 31 1 1 1 1 1 1 11 14 4 3 4 4 3
2 22 2 2 2 2 2 2 2
4 4 3 4 4 325 2525 25 25 25 25 25 25 25
1
1
1
x y x y x y x yf bf bx y x y x y x y
f bx y x y x y x y
⎧ ⎫⎧ ⎫ ⎧ ⎫⎪ ⎪⎪ ⎪ ⎪ ⎪
⎪ ⎪ ⎪ ⎪⎪ ⎪=⎨ ⎬ ⎨ ⎬⎨ ⎬⎪ ⎪ ⎪ ⎪⎪ ⎪⎪ ⎪ ⎪ ⎪⎪ ⎪⎩ ⎭ ⎩ ⎭⎩ ⎭
L
L
M ML L L L L L
L
(4.8)
The coefficients bij from Eq.(4.7) become the coefficients b1, b2, …, b25 in Eq. (4.8)
. These coefficients can now be expressed in terms of values of f at the twenty five points
by inverting Eq.(4.8), i.e.,
25
1, 1,...25n nj j
jb a f n
=
= =∑ (4.9)
103
where anj are the elements of the inverse of the matrix in Eq.(4.8).
After bn is obtained, the value of f at the center of BC is expressed in the form
4 4 3 4 4 31 2 3 23 24 25sw sw sw sw sw sw sw sw swf b x y b x y b x y b x b y b= + + + + + +L (4.10)
and using Eq.(4.9), the value of fsw can be rewritten as
25
1sw j j
jf fα
=
= ∑ (4.11)
Note that bi, i = 1, 2, …, 25 are coefficients that depend only on the mesh, the
location, and the orientation of the boundary. Therefore, with a fixed geometry, these
coefficients can be computed once at the beginning of the solution procedure.
Now we turn to the calculation of the flux on the immersed face CD of the cell
(i, j). To compute the flux on a solid boundary, we use the reflection boundary condition.
In Figure 4-2b, (xi, yj) is the mass center of the boundary cell. We introduce a point across
the boundary that is symmetrically opposite of the center of mass of face CD, and note it
with G. The variable values in the ghost point G are
( )2G ij
G ij ij
p p
u u u n n
=⎧⎪⎨
= − ⋅⎪⎩r r r r r (4.12)
where nr is the normal vector of the solid boundary, the index ij indicates the value of
variable at point (xi, yj), and the index G indicates values at the ghost point G. The value
of the flux on face CD is approximate by
2
2
G ijCD
G ijCD
p pp
u uu
+⎧=⎪⎪
⎨ +⎪ =⎪⎩
r rr
(4.13)
104
Following simple algebra, we can observe that Eq.(4.13) assures that the velocity is
zero in the normal direction, and the variation of pressure in the normal direction around
the boundary is zero.
Test Cases
Three test cases have been selected to assess the performance of the present
approach: i) radiation from a baffled piston in an open domain; ii) reflection of an
acoustic pulse on an oblique wall; iii) a wave generated by a baffled piston and reflected
on an oblique wall.
The numerical methods used in the test problem are: OPC-fv for discretization in
space, and LDDRK for discretization in time.
Radiation from a Baffled Piston
A piston in a large baffle is a good starting approximation for investigating the
radiation of sound from a boxed loudspeaker. The physical problem is to find the sound
field generated by a piston that is baffled. The problem is solved in a two-dimensional
planar domain. This problem is chosen to evaluate the field equation solver in the
absence of the rigid wall.
General description
We will use a system of coordinate centered in origin of the piston as shown in
Figure 4-3. Dimensionless variables with respect to the following scales are to be used
• Length scale = diameter of piston, 2a
• Velocity scale = speed of sound, c
• Time scale = 2a/c
• Density scale = undisturbed density, ρ0
• Pressure scale = ρ0c2
105
• Frequency = c/2a
The linearized Euler equations are used:
0
0.
p utu pt
∂+ ∇ ⋅ =
∂∂
+ ∇ =∂
r
r (4.14)
The initial conditions are:
( )( )( )
, 0.
, 0.
, 0.
u x y
v x y
p x y
=
=
=
(4.15)
y
2a
y
2a
x
y
Outflow BCRigid bafflePiston
x
y
x
y
Outflow BCRigid bafflePiston
Outflow BCRigid bafflePiston
a) b) Figure 4-3. Radiation from a baffled piston (test problem (ii)): a) general description;
b) boundary condition
For this problem, two boundary conditions are used. The boundary condition on the
wall with piston is
( ) ( ) ( )0 cos ,0,0,
0p
V t x pistonu x t
otherwiseω ∈⎧⎪= ⎨
⎪⎩ (4.16)
where ω is the frequency of the piston, and V0 is the amplitude of the displacement. The
solution is obtained using linear equation; hence the value of V0 influences only the
106
amplitude of the solution and not its behavior. From this reason we take the value of V0
equal to one.
The outflow boundary condition is based on the acoustic radiation condition of
Tam and Web (1994):
0.
cos sin 0.2
u ptp p p pt x y r
θ θ
∂+ ∇ =
∂∂ ∂ ∂
+ + + =∂ ∂ ∂
r
(4.17)
where θ is the angular coordinate of the boundary point, and r is distance from the origin
(center of the piston) to the boundary point.
An analytical solution exists for this problem (Williams, 1999; Morse and Ingard,
1968):
( )
2 2
02 2
',0,
. . '2
p
s
u x tc
p x y t d dx
γ ϕ
ρ ϕπ γ ϕ
∞
−∞
⎛ ⎞+−⎜ ⎟
⎜ ⎟⎝ ⎠=
+∫ ∫
&
(4.18)
where the x′ values are points on the source, γ2 = (x-x’)2 + y2 is distance from a point of
the piston, and pu& is time derivative of the displacement of the piston.
The behavior of the piston is presented in terms of the Helmholtz number, ka,
where k is wavenumber, and a is the radius of the piston
akac
ω= (4.19)
In the following we will present two cases: low frequency (ka = 2) and high
frequency (ka = 7.5).
107
Directive factor D
The directional characteristic of a source is described by the amplitude directivity
factor D, defined as the pressure at any angle to the pressure on the angle of maximum
pressure.
( ) ( )( )
,, / 2
p rD
p rθ
θπ
= (4.20)
where the pressure is computed at any arbitrary time t. It is clear that the radiation is
strongest on the y axis, that is why we took maximum pressure at θ = π/2.
We note that r and θ are spherical coordinate. Because wavenumber kr
is in the
same direction as rr then the same spherical angles describe both of them. Thus in
spherical coordinate we have:
coscossinsin
x
y
k kx rk ky r
θθθθ
=⎧=⎧ ⎪⎨ ⎨ == ⎪⎩ ⎩
(21)
It is clear that the strongest radiation is on the y axis. In case that 1ka , the
directive function has nulls, and between the nulls are secondary radiation maxima, of
monotonically decreasing prominence. The number of nulls and secondary maxima is
determined by the size of ka, where a = radius of the piston, and ka = 2π a/λ . In other
words the number of the lobs increases with the value of ka. In our calculation we
compare the analytical and numerical values of beam pattern for ka = 2 and ka = 7.5. As
shown in Figure 4-2 and Figure 4-3, the numerical solution recovers with high accuracy
the analytical beam pattern, which proves the capacity of the approach to recover the
solution.
108
Figure 4-3 and Figure 4-4 represent, respectively, the beam pattern for ka = 7.5
and ka = 12.5. In these figures we can see that the number of lobs increases in the same
time with Helmholtz number (ka).
Figure 4-4. Radiation from a baffled piston: Beam patterns |D(θ)|; ka = 2
Figure 4-5. Radiation from a baffled piston: Beam patterns |D(θ)|; ka = 7.5
Figure 4-6. Radiation from a baffled piston: Beam patterns |D(θ)|; ka = 12.5
109
Pressure on the face of the piston
In the following the radiance impedance of the piston is calculated. The radiation
impedance of the piston is given as:
avp
av
pZu
= (4.22)
where the average values of the pressure and velocity on the face of the piston are given
by:
/ 2
/ 2/ 2
/ 2
( , )
( , )
a
ava
a
ava
p p x y dx
u u x y dx
−
−
⎧=⎪
⎪⎨⎪ =⎪⎩
∫
∫ (4.23)
The function R1, the resistance is defined by:
( )( )1 0Re /pR Z cρ= (4.24)
To contrast the behavior of resistance is in the domain of “low frequency” (ka « 1)
and “high frequency” (ka » 1), we can see in Figure 4-7 that the resistance is close to zero
at low frequency, and close to one at high frequency. Overall, the numerical solution
closely follows the analytical solution.
Figure 4-7. Radiation from a baffled piston: Real of piston radiation impedance
110
Low frequency (ka = 2)
The computation is done on a uniformly spaced grid ∆x = ∆y = 0.05 and a time step
∆t based on CFL = 0.2, where CFL = tcx
∆∆
.
The challenge of this problem is to solve the discontinuity at time t and at the
points y c t= , when maxc t y≤ , where c is the speed of sound. In this test case, the
variables are non-dimensionalized. Hence, c = 1, and the point of discontinuity will be at
y = t, for t < ymax (see Figure 4-8).
Figure 4-9 compares the solution given by the OPC-fv scheme with the analytical
solution demonstrating that there is good agreement between them.
Figure 4-8. Radiation from a baffled piston: Numerical solution on Oy axis (ka = 2),
t = 7.3 < ymax
Figure 4-10 shows contour plots of pressure at t = 20 and t = 40. The piston
radiation starts as a collimated plane-wave beam that moves from the face to r ≅ radius of
piston, beyond which the beam propagates spherically. This behavior is in accordance
with Blackstock’s study (2000).
111
Figure 4-9. Radiation from a baffled piston: Comparison between analytical and
computed solutions on axis (ka = 2)
a) b) Figure 4-10. Radiation from a baffled piston: Contour plot of pressure (ka = 2) –
numerical solution at: a) t = 20.0; b) t = 40.0
High frequency (ka = 7.5)
The computation is done on a uniformly spaced grid ∆x = ∆y = 0.02 and a time step
∆t based on CFL = 0.2, where CFL = tcx
∆∆
. In this case we increase the resolution to
better capture the wave characteristics. Figure 4-11 compares the solution given by the
OPC-fv scheme with the analytical solution demonstrating that there is good agreement
between them. Figure 4-12 shows contour plots of pressure at t = 20, highlighting the
entire domain as well as the region close to the piston.
112
Figure 4-11. Radiation from a baffled piston: Comparison between analytical and
computed solutions on axis (ka = 7.5)
a) b) Figure 4-12. Radiation from a baffled piston: Contour plot of pressure (ka = 7.5) –
numerical solution at t = 20: a) entire domain; b) reduced domain around piston
Reflection of a Pulse on an Oblique Wall
A simple example that enables us to check the performance of the cut-cell approach
is the reflected sound on an oblique wall. Even though this problem can be solved by
placing the wall parallel to the grid line, we purposely arrange the wall so it is at an
oblique angle to the grid line. This offers a direct evaluation of the cut-cell technique.
113
In this example, the sound hits the wall and reflects. The problem is characterized
by the linearized Euler equations, Eq. (4.14).
The initial condition is
2 2
0 0( , ) exp ln(2) x x y yp x yb b
⎧ ⎫⎡ ⎤− −⎪ ⎪⎛ ⎞ ⎛ ⎞= − +⎢ ⎥⎨ ⎬⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠⎢ ⎥⎪ ⎪⎣ ⎦⎩ ⎭
(4.25)
where b = 1./6. The values x0 and y0 are chosen so that the distance from the point (x0, y0)
to the oblique solid wall is equal to 1.5.
Figure 4-13. Reflection of the pulse on an oblique wall (test problem (ii)): general
description
The outflow boundary condition is the same as that used in the previous test
problem. The solid wall boundary conditions for pressure and velocity are given by
reflecting boundary condition.
The computation is done on a Cartesian grid that is characterized by ∆x = ∆y = 0.05
and CFL = 0.5. Figure 4-12 illustrates the solid boundary intersecting a Cartesian mesh.
The boundary cell can remain independent, like in ABCK, or it can merge with a
neighboring cell. For example, CID merges with KCIJ. The computation of the parameter
in the cut cell CDIJK can be written for pressure from wave Eq.(4.14) in the form:
114
( )1 0.CD CD JK JK DI DI IJ IJ CK CK CD CDCDIJK
p u dy u dy v dx v dx v dx v dxt S
∂+ + − − − − =
∂ (4.26)
where the values of u and v represent the values of the function in the middle of the
segment. The value of the parameter is approximated using i) the OPC scheme on faces
IJ, JK and KC, ii) fourth order polynomial on faces DI , and iii) reflecting boundary
condition on face DF (boundary).
Figure 4-14. Reflection of the pulse on an oblique wall (test problem (ii)): cell around
boundary
To evaluate the performance of the cut-cell approach, Figure 4-15 shows the
pressure history from time = 0 - 4 and for three wall angles: 900, 810, and 630. As
apparent from the figure, there is a good agreement between these solutions. Figure 4-16
shows the pressure contours at three time instants by placing the solid wall at 630
inclinations.
115
Figure 4-15. Reflection of the pulse on an oblique wall: history of pressure for different
angle of wall
a) b) c)
Figure 4-16. Reflection of the pulse on an oblique wall: α = 630 a) t = 0.8; b) t = 1.6; c) t = 3.2
Wave Generated by a Baffled Piston and Reflected on an Oblique Wall
This problem is based on the combined characteristics of the previous two cases.
The wave generated by a piston is reflected by an oblique wall. The linearized Euler
equations (Eq. (4.14)) is used in this test case, and the initial condition is the same as that
of the baffled piston, Eq. (4.16).
The rectangular domain over which we do the computation is:
( ) [ ] [ ], 6, 6 0,15x y ∈ − × . The bottom of the domain is a piston mounted on a plane rigid
116
baffle; hence, the boundary conditions are given by Eq.(4.16). The velocity and pressure
on the solid and open boundaries are identical to those used in the previous case.
Figure 4-17. Wave generated by a baffled piston and reflects on an oblique wall: General
description of the domain
The piston presents the following characteristics: V0 = 1 and ω = 4 (ka = 2). The
solution is obtained using the uniform grid ∆x = ∆y = 0.05, and with a time step of
CFL = 0.5. Figure 4-18 highlights the pressure contours at different time instants.
a) b)
Figure 4-18. Wave generated by a baffled piston and reflects on an oblique wall: α = 630; a) t = 9; b) t = 14;
The challenges of this problem are to recover:
117
• The wave generate by the piston
• The reflection wave
• The combination of the two waves: wave generated by the piston and reflected on the oblique wall.
Figure 4-18 shows the outcome of the computation, exhibiting a series of local
maxims and nulls.
Conclusion
A method based on OPC-fv scheme aimed at optimizing the dispersion and
dissipation properties, and the cut-cell technique aimed at handling geometric variations
is presented. The approach is motivated by the need for handling acoustic problems with
nonlinearities (using finite volume technique) and a complex geometry (using the cut-cell
technique). Selected test cases have been used to demonstrate the performance of the
method. The present approach can offer accurate and versatile treatment to some
important and challenging aspects of acoustic problems.
118
Table 4-1. Published cut-cell approach for different problems Authors Test Cases Objective De Zeeuw et al, 1999 ;
Transonic NACA Airfoil “Subsonic” Three-Element Airfoil Supersonic Double Ellipse Supersonic Channel Flow
A method for adaptive refinement of the steady equation
Udaykumar et al, 1997
Deformation of viscous droplets in axisymetric Extensional Stakesian flow Deformation of Droplets in extensional flows with Inertia Effects Deformation of droplets in constricted tubes
The ELAFINT algorithm is developed and applied to compute flows with solid-fluid and fluid-fluid interface
Yang et al, 1999; Causon et al, 1999; Ingram et al, 2003
Open-Ended shock tube Fifteen-degree wedge flow at Mach 2 Muzzle brake flowfields Muzzle Sabot/Projectile flowfields Shallow water flow
The flow is an upwind scheme of the Godunov-type based on MUSCL reconstruction and a suitable approximate Riemann solver
Udaykumar et al, 1999
Inviscid flow around circular cylinder Track solid-liquid boundaries on a fixed underlying grid. The interface is treated as a discontinuity conditions and explicitly tracked
Ye et al, 1999 Two-dimensional stokes flow past a circular cylinder Flow past a circular cylinder immersed in a freestream Flow past a circular cylinder in a channel Application to complex geometries: (i) flow past a random array of cylinders (ii) flow past a cascade of airfoils
A Cartesian grid method has been developed for simulating two-dimensional unsteady, viscous, incompressible flows with complex immersed boundaries
Lahur et al, 2000 Moving piston Moving cylinder
Treat moving body problem
D.Calhoun et al, 2000
Advection and diffusion of a plane wave in a channel Advection and diffusion in an annulus Advection through a field of irregular objects
A fully conservative, high-resolution, finite volume algorithm for advection-diffusion equations in irregular geometries
Verstappen et al., 2000
Flow over a circular cylinder To introduce a novel cut-cell Cartesian grid method that preserve the spectral properties of convection and diffusion
Ye et al, 2001 Bubble dynamics with large liquid-to vapor density ratio and phase change
Treating sharp discontinuity interface for bubble dynamics and phase change
119
CHAPTER 5 SUMMARY AND FUTURE WORK
Predicting noise radiation associated with unsteady flows is the central theme in
aeroacoustics. Since the generation of the sound by the flow typically involves complex
physical mechanisms, approximate models are often employed. Colonius and Lele (2004)
have reviewed the modeling as well as the numerical aspects based on recent progress.
Although it is not possible to separate the flow from sound, oftentimes, we can separate
the flow prediction from noise prediction, based on pressure, frequency and length scale
separations.
In these regards, CAA presents a series of challenges. In particular, it is important
to extract small quantities relative to other phenomena in the course of computations. For
example, the pressure fluctuations in a turbulent flow are on order of 10-2 of the mean
pressure (Blake, 1986). However, loud noise can be produced by pressure fluctuation at
the level of 10-5 of the mean pressure (Hardin, 1996). In order to accurately capture the
acoustic field, numerical techniques need to yield low artificial dissipation. In addition,
the numerical dispersion related to frequency and phase constitutes another concern. For
this reason, many tests of “goodness” of CAA solutions have been developed. In light of
these issues, we have investigated schemes aimed at offering refined performance in
numerical dissipation and dispersion, including the dispersion-relation-preservation
(DRP) scheme, proposed by Tam and Webb (1993), the space-time a-ε method,
developed by Chang (1995), and the optimized prefactored compact (OPC) scheme,
proposed by Ascroft and Zhang (2003). The space-time a-ε method directly controls the
120
level of dispersion and dissipation via a free parameter, ε, the DRP scheme minimizes the
error by matching the characteristics of the wave, and the OPC scheme aims at
employing a compact stencil to achieve optimized outcome. Insight into the dispersive
and dissipative aspects in each scheme is gained from analyzing the truncation error.
Another challenge in CAA consists in the reproduction of the nonlinear
propagation, including nonlinear steepening and decay, viscous effects in intense sound
beams (e.g., acoustic near field of high-speed jets), sonic boom propagation, hydrofoil
and trailing-edge noise at low speed (Goldstein, 1976; Wang and Moin, 2000; Manoha
and Herraero, 2002; Takeda et al, 2003; Howe , 1978, 1999, 2000, 2001; Casper and
Farassat, 2004; Roger and Moreau, 2005) and high speed jet noise ( Davies et al, 1963;
Hong and Mingde, 1999; Shur et al, 2003; Fisher et al, 1998; Morris et al, 1997; Bogey et
al, 2003). In this regard, we have developed finite volume treatments of the DRP and
OPC schemes. These schemes, originally based on the finite difference, attempt to
optimize the coefficients for better resolution of short waves with respect to the
computational grid while maintaining pre-determined formal orders of accuracy. In the
present study, finite volume formulations of both schemes are presented to better handle
the nonlinearity and complex geometry encountered in many engineering applications.
Linear and nonlinear wave equations, with and without viscous dissipation, have been
adopted as the test problems. Highlighting the principal characteristics of the schemes
and utilizing linear and nonlinear wave equations with different wavelengths as the test
cases, the performance of these approaches is documented. For nonlinear wave equations,
the finite volume version of both DRP and OPC schemes offer substantially better
solutions in regions of high gradient or discontinuity.
121
Boundary treatment is another critical issue. Instead of employing boundary
conforming grid techniques, we propose a Cartesian grid, cut-cell approach, aimed at
aero-acoustics computations involving geometric complexities and nonlinearities. The
cut-cell approach handles the boundary shape by sub-dividing the computational cells in
accordance with the local geometric characteristics and facilitates the use of numerical
procedures with desirable accuracy. The resulting technique is assessed by several test
problems that demonstrate satisfactory performance.
In the following we summarize the outcome of the present research, as presented in
the previous chapters
Assessment of DRP and Space-Time CE/SE Scheme
First we investigate two numerical methods for treating convective transport: the
dispersion-relation-preserving (DRP) by Tam and Webb (1993), and the unified space-
time a-ε method, developed by Chang (1995). Based on the investigation of the DRP
CE/SE scheme we offer the following summary:
• The two methods exhibit different performance with regard to the CFL number, ν. Both truncation error analysis and numerical testing indicate that better solutions can be obtained for the DRP scheme method if ν is close to 0.2 for short wave, and close to 0.1 for intermediate and long wave. For the space-time a-ε method it is preferable if ν is less than but close to 1.
• Both schemes are dispersive for short waves. The space-time a-ε method directly controls the level of dispersion and dissipation, via the free parameter, ε. The DRP scheme, on the other hand, minimizes the error by adjusting the scheme to match the characteristics of the wave.
• Formally, based on the truncation error analysis, the DRP scheme is fourth order in space and third order in time, while the space-time a-ε scheme is second order.
• Evidence based on the truncation error analysis and the test problems indicate that in order to reduce numerical dispersion and maintain satisfactory resolution, for short wave ( e.g., b/∆x = 3), ν and ε are preferable to be close to each other. For long and intermediate waves there is virtually no need to introduce much numerical
122
dissipation and hence ε can be reduced close to zero. On the other hand, mismatched ε and ν can substantially worsen the performance of the scheme. Hence to achieve the best performance of the CE/SE scheme, one chooses the largest possible ν and then a matching ε.
• The DRP scheme exhibits mainly dispersive errors, while the space-time a-ε scheme exhibits both dispersive and dissipative errors.
• It is advisable to use space-time a-ε scheme for long wave computation because its error grows slower.
• For short and intermediate waves, the DRP scheme yields errors with lower level and slower accumulation rate.
• DRP does not offer an accurate solution for a short wave, α∆x greater than αc∆x (in our case greater than 1.1), because the wavelength is not adequately to condition of the method. Hence the DRP scheme offers a good guidance in regard to temporal and spatial sizes.
It should also be noted that he DRP scheme is a multi-step method, which requires
more boundary conditions and initial data, while the space-time a-ε scheme is a one-step
method. Combined with the fact that the DRP scheme performs better with smaller ν
(ν = CFL), it can be more expensive to compute than for space-time a-ε scheme.
Finite-Volume Treatment of Dispersion-Relation-Preserving and Optimized Prefactored Compact Schemes
The DRP and OPC schemes, originally proposed in the finite difference form, have
been further developed. To better handle nonlinearity and geometric complexities, the
finite volume version of both schemes has also been constructed. Linear and nonlinear
wave equations, with and without viscous dissipation, have been adopted as the test
problems.
For the linear wave equation, the numerical stability bound posed by the CFL
number is comparable between the DRP and OPC schemes. Both OPC and DRP schemes
123
produce solutions of comparable accuracy, but the magnitude of the error of the OPC
scheme is somewhat lower.
For the nonlinear wave equation, the finite volume schemes can produce noticeably
better solutions and can handle the discontinuity or large gradients more satisfactorily.
However, as expected, all schemes have difficulties when there is insufficient mesh
resolution, as reflected in some of the short wave cases.
In conclusion, the finite volume version of both DRP and OPC schemes improve
the capabilities of the original version of the finite difference formulas in regard to
nonlinearity and high gradients. They can enhance performance of the original DRP and
OPC schemes for many wave propagation problems encountered in engineering
applications.
Cartesian Grid, Cut-Cell Approach for Complex Boundary Treatment
A method based on high order, finite-volume schemes aimed at optimizing the
dispersion and dissipation properties, and the Cartesian grid, cut-cell technique aimed at
handling geometric complexities is presented. The approach is motivated by the need for
handling acoustic problems for practical problems involving realistic geometries.
The finite volume-based Optimized Prefactored Compact scheme and the Cartesian
cut-cell approach are combined to offer 4th order accuracy and geometric flexibility. The
computational overhead of the cut-cell approach is modest because the following
information needs to be computed only once, unless, of course, if the geometry is time
dependent:
• Data communication between cells affected by the boundary treatment
• Calculation of area and other geometric information
• Interpolation procedures required for the flux computation in the boundary region.
124
Based on the evaluation of the test cases investigated, we conclude that the present
approach can be effective in treating aeroacoustics problems with irregular geometry.
Future Work
Even though the radiant pressure fluctuation is less than 10-4 of the ambient
pressure, there are situations when the nonlinearity behavior can not be neglected.
Examples include intense sound beam resulting from the near field high-speed after
burner flows, sonic boom, and scattering of high intensity noise caused by the blade and
vortex interaction. For example, the interaction of unsteady disturbances with leading and
trailing edges of fan and compressor blades has motivated numerous studies (Goldstein,
1976; Wang and Moin, 2000; Manoha and Herraero, 2002; Takeda et al., 2003; Howe ,
1978, 1999, 2000, 2001; Casper and Farassat, 2004; Roger and Moreau, 2005). In this
problem, the edge is usually a source of high frequency sound associated with smaller-
scale boundary layer turbulence.
A general description of high frequency, nonlinear acoustics involves
nonequilibrium effects, which arises when, e.g., the acoustic time scale becomes
comparable to that of the vibrational relaxation process of polyatomic gases (Hamilton
and Blackstock, 1997).
Another important example of nonlinear behavior is landing gear noise. This
problem is recognized as one of the major components of airframe noise for commercial
aircraft. The computational investigations should take into consideration the following
aspects: the flow conditions with varying Mach number from 0.18 to 0.24 (Guo, 2005),
the noise radiation involves complex flows around a complex geometry. To be
computationally efficient, adaptive grid refinement techniques are attractive. Recent
efforts reported by Singh et al. (2005) are directly applicable.
125
LIST OF REFERENCES
Ashcroft, G. and Zhang, X., 2003, “Optimized Prefactored Compact Scheme,” J. Comput. Phys., Vol.190, pp.459-477
Basel, B., and Grünewald, M., 2003,“High Order Unstructured Finite Difference Method in Aeroacoustics,” Presented at CAA: From Acoustic Sources Modeling to Far-Field Radiated Noise Prediction, EUROMECH Colloquim no. 449, Chamonix, France
Bayliss, A., Turkel, E. and Manthey, J., 1982, “Far Field Boundary Conditions for Compressible Flows,” J. Comput. Phys., Vol.48, pp.182-199
Blacke, W.K., 1986, Mechanics of Flow-Induced Sound and Vvibration, Vols.I and II, Academic Press, New York
Blackstock, D.T., 2000, Fundamentals of Physical Acoustic, A Wiley-Interscience publication John Wiley & Sons inc, USA
Bogey, C., Bailly, C. and Juve, D., 2003, “Noise Investigation of a High Subsonic, Moderate Reynolds Number Jet Using a Cmpressible Large Eddy Simulation,” Theor Comput Fluid Dyn, Vol.16, pp.273-297
Bogey, C. and Bailly, C.,2004, “A Family of Low Dispersive and Low Dissipative Explicit Schemes for Flow and Noise Computations,” J. Comput. Phys. Vol.194, pp.194-214
Calhoun, D. and Le Veque, R.J., 2000, “A Cartesian Grid Finite-Volume Method for the Advection-Diffusion Equation in Irregular Geometries,” J. Comput. Phys., Vol.157, pp.143-180
Casper, J. and Farassat, F, 2004, “Broadband Trailing Edge Noise Predictions in the Time Domain,” J. Sound Vib, Vol.271, pp.159-176
Chang, S.C., 1995, “The method of Space- Time Conservation Element and Solution Element – A New Approach for Solving the Navier –Stokes and Euler Equations,” J. Comp. Phys., Vol.119, pp.295-324
Chang, S.C., Wang, X.Y. and Chow, C.Y., 1999, “The Space-Time Conservation Element and Solution Element Methods: a New High-Resolution and Genuinely Multidimensional Paradigm for Solving Conservation Laws,” J. Comput. Phys., Vol.156, pp.89-136
126
Cheong, C. and Lee, S., 2001, “Grid-Optimized Dispersion-Relation-Preserving Schemes on General Geometries for Computational Aeroacoustics,” J. Comput. Phys., Vol.174, pp.248-276
Chu, B.T. and Kovasznay, L.S.G., 1958, “Non-Linear Interaction in Viscous Heat-Conducting Compressible gas,” J. Fluid Mech., Vol.3 (5), pp.494-514
Colonius, T. and Lele, S.K., 2004, “Computational Aeroacoustics: Progress on Nonlinear Problems of Sound Generation,” Progress in Aerospace Sciences, Vol.40, pp.345-416
Davies, P., Fisher, M.J. and Baratt, M.J., 1963, “The Characteristics of Turbulence in the Mixing Region of a Round Jet,” J. Comput. Phys., Vol.15, pp.337-367
Delfs, J.W., 2001, “An Overlapping Grid Technique for High Resolution CAA Schemes for Complex Geometries,” AIAA Paper 2001-2199
Dowling, A.P. and Ffows Williams, J.E., 1983, Sound and Source of Sound, Ellis Horwood Limited, Chicester, England, UK
Dumbser, M., 2003, “ADER Discontinuous Galerkin Schemes for Aeroacoustics,” CAA: From Acoustic Sources Modeling to Far-Field Radiated Noise Prediction, EUROMECH Colloquim no. 449
Efraimsson, G. and Kreiss, G., 1996, “A Note on the Effect of Artificial Viscosity on Solution of Conservation Laws,” Applied Numerical Mathematics, Vol.21, pp.155-173
Efraimsson, G., 1997, ”A 2D Analysis of the Influence of Artificial Viscosity Terms on Solutions of the Euler Equation,” J. Comput. Phys., Vol.138, pp.103-120
Ffowcs-Williams, J.E., 1969, “Sound Generation by Turbulence and Surfaces Arbitrary Motion,” Phil. Trans. Roy. Soc. A264
Ffowcs-Williams, J.E., 1992, Modern Methods in Analytical Acoustics, D.G.Crighton et al. Springer Verlag
Ffowcs-Williams, J.E., and Hawkings, D.L., 1969, “Theory Relating to the Noise of Rotating Machinery,” J. Sound and Vib. Vol.10(1)
Fisher, M.J., Preston, G.A. and Bryce, W.D., 1998, “A Modelling of the Noise from Simple Coaxial Jets, Part I: With Unheated Primary Flow,” J. Sound Vibr., Vol.209(3), pp.385-403
Gloerfelt, X., Bailly, C., Juvé, D., 2003, “Direct Computation of the Noise Radiated by A Subsonic Cavity Flow and Application of Integral Methods,” J. Sound and Vibration, Vol.266, pp.119-146
127
Goldstein, M.E., 1976, Aeroacoustics, McGraw-Hill International Book Company, New York
Goodrich, J.W., 1995, “An Approach to the Development of Numerical Algorithm for First Order Linear Hyperbolic System in Multiple Space Dimensions: The Constant Coefficient Case,” NASA TM 106928
Guo, Y., 2005, ”A Statistical Model For Landing Fear Noise Prediction,” J.Sound and Vibration, Vol.282, pp.61- 87
Gutin, L., 1948, “On the Sound Field of a Rotating Propoller,” NACA TM 1195
Hamilton, M.F. and Blackstock, D.T., 1997, Nonlinear Acoustic, Academic Press Limited, London
Hardin, J.C., 1996, “Introduction to Computational Aero-Acoustic,” Applied Aero-Acoustic Lecture, Lecture Series 1996-2004, von Karman Institute for Fluid Dynamics, Chaussé de Waterloo, 72, Rhode Saint Genèse, Belgium
Hardin, J. and Hussaini, M.Y., 1992, Computational Aeroacoustics, Springer-Verlag, New York,Berlin
Hardin, J.C and Pope, D.S., 1995, “Sound Generation by Flow over a Two-Dimensional Cavity,” AIAA Journal, Vol.33, No.3, pp.407-412
Harten, A., 1989, “ENO Schemes with Subcell Resolution,” J. Comput. Phys., Vol.83, pp.148-184
Harten, A., 1983, “High Resolution Schemes for Hyperbolic Conservation law,” J. Comput. Phys., Vol.49, pp.357-393
Hataue, I., 2003, “On Analogy and Dissimilarity of Dependence of Stability on Several Parameters in Flow Simulation,” J. Computational and Applied Mathematics, Vol.159, pp.45-53
Henshaw, B., 2004, “A high-Order Accurate Solver for Maxwell’s Equations on Overlaping Grids,” 7th Symposium on Overset Composite Grids and Solution Technology, Huntington Beach, Ca
Hirschberg, A., 1994, “Applied Aeroacoustics,” Presented at Applied Aero-Acoustic Lecture, Lecture Series 1994-2004, von Karman Institute for Fluid Dynamics, Chaussé de Waterloo, 72, Rhode Saint Genèse, Belgium
Hirschberg, A. and Rienstra, S.W., 1994, “Elements of Aeroacoustics,” Applied Aero-Acoustic Lecture, Lecture Series 1994-2004, von Karman Institute for Fluid Dynamics, Chaussé de Waterloo, 72, Rhode Saint Genèse, Belgium
128
Hirsh, C., 1990, Numerical Computation of Internal and External Flows, Vol.II: Computational Methods for Inviscid and Viscous Flows, Wiley Series in Numerical Methods in Engineering, Massachusettes
Hixon, R., 1997, “Evaluation of High-Accuracy MacCormack – Type Scheme Using Benchmark Problems,” NASA Contractor Report 202324, ICOMP-97-03
Hixon, R. and Turkel, E., 1998, “High-Accuracy Compact MacCormack-Type Schemes for Computational Aeroacoustics,” NASA CR 1998 – 208672
Hong, Y. and Mingde, S.U., 1999, “Application and Comparison of Two SGS Models in Large Eddy Simulation of Free Turbulent Jet Flow,” Communication in Nonlinear Science & Numerical Simulation, Vol.4(1)
Howe, M.S., 1978, “A Review of the Theory of Trailing Edge Noise,” J. Sound Vib., Vol.62, pp.437-465
Howe, M.S., 1999, “Trailing Edge Noise at Low Mach Numbers,” J. Sound Vib., Vol.225(2), pp.211-238
Howe, M.S., 2000, “Trailing Edge Noise at Low Mach Numbers, Part 2: Attached and Separated Edge Flows,” J. Sound Vib., Vol.234(5), pp.761-775
Howe, M.S., 2001, “On the Hydroacoustics of Trailing Edge Noise with a Detached Flap,” J. Sound Vib., Vol.239(4), pp.801-817
Hu, F.Q., Hussaini, M.Y. and Manthey, J.L., 1996” Low Dissipation and Dispersion Runge-Kutta for Computational Acoustics,” J.Comput. Phys., Vol.124, pp.177-191
Hu, Z.W., Morfey, C.L. and Sandham, N.D., 2002, “Aeroacoustics of Wall-Bounded Turbulent Flows,” AIAA J., Vol.40(3), pp.465 -473
Ingram, D.M., Causon, D.M., and Mingham, C.G., 2003, ” Development in Cartesian Cut Cell Methods,” Mathematics and Comput. In Simulation, Vol.61, pp.561-572
Kim,. J.W., and Lee, D.J., 1996, “Optimized Compact Finite Difference Schemes with Maximum Resolution,” AIAA J., Vol.34, No. 5, pp.887-893
Kim, C., Roe, P.L. and Thomas, J.P., 1997, “Accurate Schemes for Advection and Aeroacoustics,” Technical Paper 97-2091, AIAA Press, Washington DC
Lele, S.K., 1997, “Computational Aeroacoustics: A Review,” AIAA paper 97-0018
Leonard., B.P., 1988, “Simple High-Accuracy Resolution Program for Convective Modeling of Discontinuities,” Int. J. Numer. Methods in Fluids, Vol.8, pp.1291-1318
129
Lighthill, M.J., 1952, “On Sound Generated Aerodynamically. I. General Theory,” Proc. Roy. Soc. (London), 211A, 1107, pp.564-587
Lighthill, M.J., 1954, “On Sound Generated Aerodynamically. II. Turbulence as a Source of Sound,” Proc. Roy. Soc. (London), 22A, 1148, pp.1-32
Lin, S.Y. and Chin, Y.S., 1995, “Comparison of Higher Resolution Euler Schemes for Aeroacoustics Computation,” AIAA J., Vol.33, pp.237
Loh C.Y., Hultgren, L.S., Chang, S.C. and Jorgenson, P.C.E., 2000, “Noise Computation of a Shock-Containing Supersonic Axisymmetric Jet by the CE/SE Method,” AIAA Paper 2000-0475
Loh, C.Y., Hultgren, L.S. and Chang, S.C., 1998, “Computing Waves in Compressible Flow Using the Space-Time Conservation Element and Solution Element Method,” AIAA Paper 98-0369
Loh, C.Y., Hultgren, L.S., Chang, S.C. and Jorgenson, P.C.E., 1999, “Vortex Dynamics Simulation in Aeroacoustics by the Space- Time Conservation Element and Solution Element Method,” AIAA Paper 99-0359
Loh, C.Y., Hultgren, L.S., Wang, X.Y., Chang, S.C. and Jorgenson, P.C.E., 2000, “Aeroacoustics Computation for Nearly Fully Expanded Supersonic Jets Using the CE/SE Method,” NASA/TM 210225
Liou, M.S. and Steffen, C.J.,1993, “ A New Flux Splitting Scheme,” J.Comput. Phys., Vol.107, pp.23-39
Margot, G. and Olsson, P., 1996, “Designing an Efficient Solution Strategy for Fluid Flows – 1. A Stable High Order Finite Difference Scheme and Sharp Shock Resolution for the Euler Equations,” J. Comput. Phys., Vol.129, pp.245-262
Manoha, E. and Herraero, C., 2002, “Numerical Prediction of Airfoil Aerodynamic Noise,” AIAA Paper 2002-2573
Morris, P.J., Long, L.N., Bangalore, A. and Wang, Q., 1997, “A Parallel Three-Dimensional Computational Aeroacoustics Method Using Nonlinear Disturbance Equations,” J.Comput. Phys., Vol.133, pp.56-74
Morse, P.M., and Incard, K.U., 1968, Theoretical Acoustic, McGraw-Hill Book Company, New York
Nance, D.V., Viswanathan, K. and Sankar, L.N., 1997, “Low-Dispersion Finite Volume Scheme for Aeroacoustic Application,” AIAA Journal, Vol.35, No.2, pp.255-262
Oran, E.S. and Boris, J., 2002, Numerical Simulation of Reactive Flow, Cambridge University Press, DC/New York, second edition
130
Osher, S. and Chakravarthy, S., 1983, “Upwind Schemes and Boundary Conditions with Applications to Euler Equations in General Geometries,” J. Comput. Phys., Vol.50, pp.447-481
Popescu, M. and Shyy, W., 2002 “Assessment of Dispersion-Relation-Preserving and Space-Time CE/SE Schemes for Wave Equations,” Numerical Heat Transfer, Vol.42, No 2, pp.93-118
Popescu, M., Shyy, W., and Garbey, M., 2004, “A Study of Dispersion-Relation-Preserving and Optimized Prefactored Compact Schemes for Wave Equation,” AIAA paper 2004-0519, also accepted for publication in J. Comput. Phys.
Powell, A., 1964, “Theory of Vortex Sound,” JASA , Vol.33(1), pp.177-195
Roe, P.L., 1981, “Numerical Algorithms for the Linear Wave Equation,” Technical Report 81047, Royal Aircraft Establishment, Bedford, UK
Roger, M., 1995, “Applied Aeroacoustics Methods for the Acoustic Design of Shrouded Tail Rotor,” AIAA Aeroacoustics Conference, Münschen
Roger, M. and Moreau, S., (accepted October 2004), “Back-Scattering Correction and Further Extensions of Amiet’s Trailing-Edge Noise Model. Part 1: Theory” J.Sound Vibr.
Sherer, S., 2004, “High-Order Overset- Grid Acitivities in AFRL/VAAC,” Air Forces Research Laboratory, Presented at 7th Symposium on Overset Composite Grids and Solution Technology, Huntington Beach, California
Shu, C. and Oscher, S., 1988, “Efficient Implementation of Essentially Non-Oscillatory Shock-Capturing Schemes II,” J. Comput. Phys., Vol.83, pp.32-78
Shur, M.L., Spalart, P.R., Strelets, M.Kh. and Travin, A.K., 2003, “Towards of Noise from Jet Engine,” Int.J.Heat and Fluid Flow, Vol.24, pp.551-561
Shyy, W., 1985 “A Study of Finite Difference Approximations to Steady-State, Convection-Dominated Flow Problems,” J. Comput. Phys., Vol.57, pp.415-438
Shyy, W., 1994, revised printing 1997, Computational Modeling For Fluid Flow and Interfacial Transport, Elsevier, Amsterdam, The Netherlands
Shyy, W., Francois, M. and Udaykumar, H.S., 2001, “Cartesian and Curvilinear Grid Method for Multi-Domain, Moving Boundary Domain” Thirteenth International Conference on Domain Decomposition Methods, CIMNE, Barcelona, M. Debit et al. editors
Shyy, W., Francois, M., Udaykumar, H.S., N’Dri, N, and Tran-Son-Tay, R., 2001, “Moving boundaries in micro-scale biofluid dynamics,” Applied Mechanics Reviews, Vol.54, pp.405-453
131
Shyy, W., Thakur, S.S., Ouyang, H., Liu, J. and Blosch, E., 1977, Computational Techniques for Complex Transport Phenomena, Cambridge University Press, New York
Singh, R.K., N’Dri, N.N., Uzgoren, E., Shyy, W., and Garbey, M., 2005 “Three-Dimensional Adaptive, Cartesian Grid Method for Multiphase Flow Computations,” AIAA Paper 2005-1389, 43rd Aerospace Sciences Meeting & Exhibit
Stanescu, D. and Habashi, W.G., 1998 “2N-Storage Low Dissipation and Dispersion Runge-Kutta for Computational Acoustics,” J. Comput. Phys., Vol.143, 674-681
Sun, M. and Takayama, K., 1999, “Conservative Smoothing on an Adaptive Quadrilateral Grid,” J. Comput. Phys., Vol.150, pp.143-180
Tam, C. K. W. and Dong, Z., 1994, “Wall Boundary Condition for High-Order Finite Difference Schemes in Computational Aeroacoustics ,” Theor. Comput. Fluid Dynam. Vol.6, pp.303
Tam, C. K. W. and Hardin, J. C., 1997 (Eds), Second Computational Aeroacoustics Workshop on Benchmark Problems, NASA CP-3352
Tam, C.K.W. and Burton, D.E., 1984, “Sound Generation by Instability Waves of Supersonic Flows. Part2. Axisymetric Jets,” J. Fluid. Mech., Vol.138 pp.273-295
Tam, C.K.W. and Morris, P.J., 1980, “The Radiation of Sound by Instability Waves of a Compressible Plane Turbulent Shear Layer,” J. Fluid. Mech., Vol.98(2), pp.349-381
Tam, C.K.W., and Kurbatskii, K.A., 2000, “A Wavenumber Based Extrapolation and Interpolation Method for Use in Conjunction with High-Order Finite Difference Schemes,” J. Comput. Phys., Vol.157, pp.588-617
Tam, C.K.W., and Webb, J.C,., 1994, “Radiation Boundary Condition and Anisotropy Correction for Finite Difference Solutions of the Helmholtz Equation,” J. Comput. Phys., Vol.113, pp.122-133
Tam, C.K.W., and Webb, J.C., 1993, “Dispersion-Relation-Preserving Finite Difference Schemes for Computational Acoustics,” J. Comput. Phys., Vol.107, pp.262-281
Tam, C.K.W., Webb, J.C., and Dong, Z., 1993, “A study of the Short Wave Components in Computational Acoustics,” J. Comput. Acoust., Vol.1, pp.1-30
Takeda, K., Zhang, X. and Nelson, P.A., 2003, “Computational Aeroacoustic Simulation of Leading-Edge Slat Flow,” J.Sound and Vibration, Vol.270, pp.559-572
Tang, L. and Baeder, J., 1999, “Uniformly Accurate Finite Difference Schemes for p-Refinement,” SIAM J. Sci. Comput, Vol.20, No.3, pp.1115-1131
132
Tannehill, C., Anderson, D.A. and Pletcher, R.H., 1997, Computational Fluid Mechanics and Heat Transfer, Taylor & Francis Group, Philadelphia, PA
Tennekes, H. and Lumley, L., 1972 ,A First Course in Turbulence, The MIT Press, Cambridge, Massachusetts, and London, England
Thakur, S., Shyy, W. and Liou, M.-S., 1996a, "Convection Treatment and Pressure Splitting for Sequential Solution Procedures. Part II: Pressure-Based Algorithm,” Numerical Heat Transfer; Part B, Vol.29 pp.29-42
Thakur, S.S., Shyy, W., and Liou, M.S., 1996b, “Convection Treatment and Pressure Splitting for Sequential Solution Procedures. Part I: Theory and One- Dimensional Test Cases,” Numerical Heat Transfer, Part B, Vol.29, pp.1-27
Toro, E.F. and Billet, S.J., 1996 “Centred TVD Schemes for Hyperbolic Conservation Laws”. Technical Report MMU-9603, Department of Mathematics and Physics, Manchester Metropolitan University, UK
Udaykumar, H.S., Kan, H.C., Shyy, W. and Tran-Son-Tay, R., 1997 “Multiphase Dynamics in Arbitrary Geometries on Fixed Cartesian Grids” , J. Comp. Phys., Vol.137, pp.366-405
Udaykumar, H.S., Mittal, R. and Shyy, W., 1999 “Computational of Solid-Liquid Phase Fronts in the Sharp Interface Limit on Fixed Grids,” J. Comput. Phys., Vol.153, pp.535-574
van Leer, B., 1979, “Towards the Ultimate Conservation Difference Scheme V. A Second Order Sequel to Godunov’s Method,” J. Comput. Phys., Vol.32, pp.101-136
Verstappen, R.W.C.P and Veldman, A.E.P., 2000 “Numerical Computation of a Viscous Flow around a Circular Cylinder on a Cartesian Grid,” European Congress Methods in Applied Sciences and Engineering, Barcelona, 11-14 September
Vichnevetsky, R., 1987, “Wave propagation and Reflection in Irregular Grids for Hyprebolic Equations,” Applied Numerical Mathematics, Vol.3, pp.133-166
Wang, G. and Sankar, L.N., 1999, “Prediction of Rotorcraft Noise with a Low-Dispersion Finite Volume Scheme,” AIAA-99-0480
Wang, X.Y., Chang, S.C. and Jorgenson, P.C.E., 2000, “Accuracy Study of the Space- Time CE/SE Method for Computational Aeroacoustics Problems Involving Shock Waves,” AIAA Paper No.2000-0474, 38th Aerospace Sciences Meeting & Exhibit, Reno, NV, January 10-13
Wang, Z.J. and Chen, R.F., 2001, “Optimized Weighted Essentially Nonoscillatory Schemes for Linear Waves with Discontinuity,” J. Comput. Phys., Vol.174, pp.381-404
133
Wang, M. and Moin, P., 2000, “Computationa of Trailing-Edge Flow and Noise Using Large-Eddy Simulation,” AIAA J., Vol.38, pp.2201-2209
Williams, E.G., 1999, Fourier Acoustics Sound Radiation and Nearfield Acoustical Holography, Academic Press, Dan Diego
Yang, G. and Ingram, D.M., 1999 “Cartesian Cut-Cell Method for Axisymmetric Separating Body Flows,” AIAA Journal, Vol.37, No.8, pp.905-911
Ye, T., 2001, Direct Numerical Simulation of a Translating Vapor Bubble With Phase Change, Ph.D. Dissertation, University of Florida, Department of Mechanical Engineering, Gainesville FL
Ye, T., Mittal, R., Udaykumar, H.S. and Shyy, W., 1999 “An Accurate Cartesian Grid Method for Viscous Incompressible Flows with Complex Immersed Boundaries,” J. Comput. Phys., Vol.156, pp.209-240
Ye, T., Shyy, W. and Chung,J.N., 2001, “A Fixed-Grid, Sharp-Interface Method for Bubble Dynamics and Phase Change,” J. Comput. Phys. Vol.174, pp.781-815
Yu, S.T., Chang, S.C., 1997, “Treatments of Stiff Source Terms in Conservation Laws by the Methods of Space- Time Conservation Element and Solution Element,” AIAA Paper No.97-0435, 35th Aerospace Sciences Meeting & Exhibit, Reno, NV, January
134
BIOGRAPHICAL SKETCH
Mihaela Popescu was born in Romania, in a small village, Berca, known for rare
natural phenomena- mud volcanoes. Mihaela received two bachelor degrees from the
University of Bucharest, Romania. Her first BS was in mathematics, specializing in
mathematics applied in mechanics. Her second BS was in biology, specializing in
ecology. While pursuing her degree in biology, Mihaela worked with modeling of
Chironomid community based on ecosystem characteristics. In adition, for over a year,
Mihaela worked for the Nuclear Energy Reactor Institute in Bucharest, Romania.
Mihaela’s other qualification also entail working as a teacher for two years.
In 1999, she enrolled in the University of Florida to pursue a Doctor of Philosophy
degree in Aerospace Engineering under the guidance of Dr. Wei Shyy.
During her Ph.D. studies she was the recipient of the Alumni Fellowship at the
University of Florida.
Mud volcano cones in the Berca Anticline Depression of the Curvature
Subcarpathians. [Photograph Credit: Dr. Dan Balteanu, Romanian Academy]