MASTER THESIS

TIME-DOMAIN IMPEDANCE BOUNDARY

CONDITIONS IN COMPUTATIONAL
FLUID DYNAMICS FOR USE IN
THERMOACOUSTIC MODELING

Bart van der Poel

FACULTY OF ENGINEERING TECHNOLOGY

LABORATORY OF THERMAL ENGINEERING

EXAMINATION COMMITTEE
Prof.dr.ir. T.H. van der Meer (chair)
Ir. J.P. Oosterhuis (member)
Dipl.-Ing. S. Bühler (member)
Dr. A.R. Thornton (external member)

AUGUSTUS 2013
 Time-Domain Impedance Boundary Conditions in
Computational Fluid Dynamics for use in Thermoacoustic
Modeling

Bart van der Poel - University of Twente

August, 2013
 Abstract

Thermoacoustic devices convert heat to work or vice versa using acoustic waves, and consist
of several parts that partially reflect acoustic waves. These reflections can be represented by
an acoustic impedance, which is a complex quantity that takes into account a phase difference
between the pressure and velocity oscillation. Usually one is only interested in a single part
of the thermoacoustic device, but to correctly numerically predict these wave phenomena the
whole device should be modeled. As solving acoustics using Computational Fluid Dynamics
(CFD) requires comparatively high computational resources, using the complete computational
domain of the whole device is unattractive. The reduction of the computational domain might
be possible with a so-called Time-Domain Impedance Boundary Condition (TD-IBC), which
relates the inward traveling acoustic wave to the outward traveling acoustic wave. Standard
commercial CFD software does not feature such a boundary condition as of yet. This master
thesis discusses the feasibility, theory and implementation of both a non-reflective boundary
condition (NRBC) and a TD-IBC. The NRBC and TD-IBC are implemented in ANSYS CFX
CFD software using FORTRAN subroutines. The results showed that for a complex reflection
coefficient of R = 0.5 + 0.5i, and an excitation frequency of f = 100Hz, the measured absolute
reflection coefficient |R| was off 1.05%, and the phase differed φ = 0.0334rad. From the work
done it can be concluded that it is indeed possible to implement a TD-IBC for application in
thermoacoustics.
Preface

The start of this master thesis was characterized by uncertainty, as the exact topic was not
known yet. What was known, was that it would be a numerical assignment in the field of
thermoacoustics. After some preliminary research, and input from Joris and Simon, the topic
of the assignment became clear. Joris and Simon guided me throughout the whole thesis, and
despite often being busy, have always found time to discuss matters. Later in the project, a
weekly progress meeting with everybody related to the thermoacoustics project started. This
meeting provided me with invaluable feedback.

As Joris and Simon were the most prominent contributors to the succes of this thesis, I
would like to thank them first. I also would like to thank Theo for helping me choose a master
thesis assignment, assigning me to the project, and guiding me throughout the project.

A special note for my colleagues in room N248. Although the ubiquitous presence of my
fellow graduation students in room N248 have impaired my work efficiency from time to time,
their unique characters have supplied me with plentiful laughter that prevented me from falling
into utter despair at numerous stages of my research. During my stay we became accustomed
to various avocations, such as “klaverjassen” at 12 o’clock, and “blik-time” at 3 o’clock in the
afternoon. While the rudimentary knowledge of some “klaverjassen” players jeopardized the
efficacy of the teams during play, their assiduous efforts and lavish endeavors were often eulo-
gized. Thus I would like to thank Daan, Johan, Frans, Sjoerd, Gerald, Maarten, and Sam for
creating a collegial and comfortable atmosphere.

B.J. van der Poel (s0217972)

University of Twente
Faculty of Engineering Technology
Laboratory of Thermal Engineering
Sustainable Energy Technology

The exam committee is comprised of the following members:

Prof.dr.ir. T.H. van der Meer (Chair)

Ir. J.P. Oosterhuis (Member)
Dipl.-Ing. S. Bühler (Member)
Dr. A.R. Thornton (External Member)

2
Contents

1 Introduction 5
1.1 Thermoacoustics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Thermoacoustics in computational fluid dynamics . . . . . . . . . . . . . . . . . . 12
1.3 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Theory 14
2.1 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Navier-Stokes Characteristic Boundary Conditions . . . . . . . . . . . . . . . . . 18
2.3.1 Characteristic analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Determination of outgoing wave amplitude variation . . . . . . . . . . . . 22
2.4 Linear relaxation method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Plane-Wave Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6 Time-Domain Impedance Boundary Conditions . . . . . . . . . . . . . . . . . . . 24
2.6.1 Determination of filter coefficients . . . . . . . . . . . . . . . . . . . . . . 26
2.6.2 Determination of normalized angular frequency . . . . . . . . . . . . . . . 27
2.6.3 Stability of filter coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.7 Reflection measurement theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.7.1 Multi-microphone method . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3 Method 30
3.1 Domain model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.1 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.2 Space and time discretization . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Fluid models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.1 Initial conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.2 Heat energy equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Implementation of a TD-IBC in ANSYS CFX . . . . . . . . . . . . . . . . . . . . 36
3.3.1 Fortran programming in ANSYS CFX . . . . . . . . . . . . . . . . . . . . 36
3.3.2 Parallelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 Solver data structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.4 Memory precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.5 Memory Management System . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.6 Initialization routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.7 Non-reflective part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.8 Plane-wave masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.9 Time-domain impedance boundary condition part . . . . . . . . . . . . . 53
3.4 Reflection measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.1 Measurement in test-cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.1 Solver settings and convergence criteria . . . . . . . . . . . . . . . . . . . 59

3
 3.5.2 Simulation time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.3 Relaxation factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.4 Non-reflective boundary test cases . . . . . . . . . . . . . . . . . . . . . . 60
3.5.5 Time-domain impedance boundary condition test cases . . . . . . . . . . 62

4 Results 66
4.1 Non-reflective boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.1.1 Heat transfer equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.2 Time integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1.3 Relaxation factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1.4 Mean velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1.5 Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.1.6 Linear refinement near boundary . . . . . . . . . . . . . . . . . . . . . . . 77
4.1.7 Sharp gradients in pressure and velocity . . . . . . . . . . . . . . . . . . . 79
4.2 Time-domain impedance boundary condition . . . . . . . . . . . . . . . . . . . . 81
4.2.1 Acoustically hard wall and pressure release surface . . . . . . . . . . . . . 82
4.2.2 Non-reflecting outlet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2.3 Complex reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5 Conclusions and Recommendations 84

References 88

Nomenclature 92

Appendix 93

4
Chapter 1

Introduction

The twentieth century has seen unprecedented growth in the fields of science, technology, and
medicine, but also huge advances in civil and political rights. These fundamental and far reach-
ing changes throughout society, more often seen as beneficial than not, increased life span,
well-being and prosperity for people from all walks of life. But the vastly increased quality of
life over the last centuries does come at a price. The dependence on finite reserves of fossil fuels
and the increased pollution are two of the most prominent caveats of these changes. To mitigate
these negative side effects, one can look at possible (partial) solutions from three angles. In
order of priority: 1. To increase efficiency on the demand side (i.e. use less). 2. To use more
sustainable energy sources. And 3. to make more efficient use of fossil fuels. This strategy has
been formulated by E. Lysen [12] and is called Trias Energetica.

Figure 1.1: Trias Energetica

This master thesis is based on the premise that thermoacoustic devices can function as a po-
tential solution to some of the aforementioned problems society faces today. A thermoacoustic
engine is a type of thermoacoustic device that uses sound waves as a means to transport heat

5
(heat pump) or convert heat into mechanical work (prime mover). The mechanical work done
in a prime mover could be used to drive electrical generators and produce electricity. These
type of engines could possibly fulfill roles in both point 2 and 3 of the Trias Energetica strategy.
Point 2 by using a sustainable heat sources, and point 3, for example, by using part of the gas
to produce work or electricity while the waste heat is used to heat a building.

While the thermoacoustic engine and principle has been around for a while, the recently
increased awareness on the topic of sustainability has sparked new interest in the technology
which had previously been practically dormant for years. The technology features some promis-
ing characteristics in the light of sustainability. Firstly, thermoacoustic devices can operate
with a small temperature difference between heat source and sink. This opens possibilities for
application with sustainable heat sources, as these are generally of relatively low temperature
compared to fossil fuel combustion. Secondly, the devices usually have minimal or no moving
parts, depending on application. Which can be beneficial to production and maintanance costs,
as well as reliability (especially in rough conditions).

The technology has not enjoyed the amount of research that the piston and turbine engines
had. In this respect it lags behind considerably, and much experimental, analytical, as well
as numerical research needs to be done to bring the technology on par. In this master thesis
focus lays on the numerical modeling aspect. To create a physically realistic numerical model
of a thermoacoustic engine based on a fluid dynamics code, often the complete domain must be
calculated, even if only the performance and flow field of a particular section needs to be deter-
mined. Special boundary conditions called “Time-Domain Impedance Boundary Conditions”
(TD-IBC) can potentially solve these problems. These boundary conditions are, however, not
provided by commercial CFD codes. A literature research is done on these special boundary
conditions, and subsequently are implemented using Fortran subroutines in a commercial CFD
package called Ansys CFX, and their performance is tested.

6
1.1 Thermoacoustics
The first observations and descriptions about the thermoacoustic effect are already over two
centuries old. In 1777, B. Higgins showed how to produce acoustic oscillations excited purely
by a single heat source. His device was simple, just an open glass tube with a hydrogen flame
at a suitable position (quarter length) inside, and was called Higgins‘ singing flame. A similar
experiment was devised by P.L. Rijke, but now the flame was replaced by a heated gauze (see
Fig. 1.2). Rijke also observed that the convective air current was imperative for the acoustic
oscillations to occur [7].

Figure 1.2: a) Higgins’, b) Rijke, and c) Sondhauss tubes [7]

Another device is the Sondhauss tube, which has a closed end with a bulb attached to it
(see Fig. 1.2). The bulb side was heated and did not need convective air current to operate.
Nowadays, Higgins‘ singing flame and the Rijke tube can be characterized as a quarter wave
resonator, while the Sondhauss tube is a half wave resonator. The thermoacoustic effect does
not only occur under heating, but also under cooling. Taconis observed, when cooling a gas-
filled tube from room temperature to cryogenic temperature, spontaneous oscillations of the gas
under extreme cooling. The first, and correct, qualitative explanation about the thermoacoustic
effect has been done by Lord Rayleigh in The Theory of Sound [20], saying:

“If heat be given to the air at the moment of greatest condensation (compression)
or taken from it at the moment of greatest rarefaction (expansion), the vibration is
encouraged.”

Thermoacoustics, as the name implies, is a combination of the fields of thermodynamics

and acoustics. Acoustic sound waves are a combination of pressure and particle velocity waves.
Temperature variations are usually very small for sound mostly experienced by people, and
therefore often neglected in acoustics. In thermoacoustics however, it is essential in describing
the thermoacoustic effect.
Thermoacoustic effects are often explained by using the notion of fluid parcels. Fluid parcels
can be seen as the smallest volume possible in the medium for which the continuum assumption
is still valid. The continuum assumption is the assumption that the behavior of the materials
can be modeled as a continuous mass rather than as discrete particles (e.g. molecules), and is
the basis of many equations in fluid dynamics (e.g. Navier-Stokes equations).

7
 In free space, gas parcels exposed to an acoustic disturbance behave in a such a way that they
predominantly expand and compress adiabatically. The temperature variations as a function
of pressure in an adiabatic process can be described by equation 1.1. Where γ is the adiabatic
index (1.4 for air), T the temperature, and P the pressure. This means that when the pressure
increases, so does the temperature, and vice versa. (under the assumption that γ > 1, which it
always the case in an adiabatic process).

T γ = Constant · P γ−1 (1.1)

When gas parcels are allowed to interact thermally with a solid boundary, two types of
thermoacoustic effects may occur. Energy (heat) can be supplied to the acoustic wave, which
increases acoustic power (called prime mover ), or energy (heat) can be extracted and pumped,
which decreases acoustic power (called heat pump). These parts can be combined with an
electroacoustic transducer to supply (e.g. loudspeaker) or extract (e.g. microphone) acoustic
energy from and to a mechanical/electrical source. Note that it does not necessarily need to be
a loudspeaker or microphone to supply or extract acoustic energy, for example, a bi-directional
impulse turbine could also be used to extract acoustic energy. Different useful combinations
of parts are summed up in Table 1.1. In Fig. 1.3 an example of a heat driven refrigerator is
shown. For example, to use industrial waste heat to cool another medium.

Table 1.1: Different thermoacoustic applications (non-exhaustive list)

# Acoustic source Acoustic sink Description
1 Prime mover Heat pump Heat driven refrigerator
2 Prime mover Microphone Heat conversion to work/electricity
3 Prime mover Environment Heat induced loudspeaker
4 Loudspeaker Heat pump Work/electricity driven refrigerator

Figure 1.3: Thermoacoustic device that functions as a refrigerator [7]

The acoustic pressure and particle velocity do not necessarily be in phase. When the acous-
tic wave is a purely traveling wave, the phase angle between the pressure and velocity is zero
(see Fig. 1.4b), while a purely standing wave has a phase angle of 90◦ (see Fig. 1.4a). A stand-
ing wave is caused by two traveling waves propagating in opposite directions, and has nodes
at certain positions. Nodes are positions in space where the acoustic pressure or velocity is
constantly zero, and are called pressure nodes and velocity nodes respectively. Positions where
the acoustic pressure or particle velocity periodically reaches the maximum amplitude are called
anti-nodes. Often, acoustic waves are not purely the one or the other, but a combination of both.

8
 At a pressure anti-node the pressure variations, and thus the temperature variations, are
largest, while at a pressure node there is no temperature variation. At a velocity anti-node the
displacement is large, while at a velocity node there is no variation. For a standing wave device
this means that at a velocity node the temperature variations are largest, while at a velocity
anti-node there is no temperature variation. Between the nodes both the temperature and
the velocity oscillates (and thus displaces the parcel), which results in a temperature gradient
(visualized in Fig. 1.4c), which is known as the critical temperature gradient.
For a traveling wave the situation is somewhat different because the nodes coincide and move
with the speed of sound. At the moment of greatest compression (hottest), the fluid parcels
moves to the right, it then stops and starts moving to the left while expanding (cooling). This
gives circle like temperature-position cycles (see Fig. 1.4d).
In practice, a device will be designed either as a standing or a traveling wave device, but
both effects will occur. This will result in oval shaped temperature-positions paths (see Fig.
1.4e).

Figure 1.4: Standing and traveling wave temperature position cycles [7]

The heat pump and prime mover parts of a thermoacoustic device are in basic structure
very similar, and often comprised of two heat exchangers with a porous structure between them.
One heat exchanger is cold, the other hot. If the part functions either as a heat pump or prime
mover depends on the actual temperature gradient achieved between the hot and cold heat
exchangers, and the critical temperature gradient. If the temperature gradient at the position
of the porous structure is smaller than the critical temperature gradient, the part will function
as a heat pump (see Fig. 1.5a). If the temperature gradient is larger, than the part will function
as a prime mover (see Fig. 1.5b). A heat pump will cause heat to be “pumped” from the cold
to the hot heat exchanger, while a prime mover will cause the heat to flow from the hot to the
cold heat exchanger.

The porous structure in between the heat exchangers serves two main purposes. For both
traveling and standing wave devices the porous structure provides much more surface area for
heat transfer to take place, and thus increasing the amount of power that can be transferred.
The other porpuse has everything to do with maximizing the encircled area within the P-v
diagram, which will be explained later in this section.

9
 Figure 1.5: Refrigerator and prime mover temperature gradients [7]

The physical principle of a heat pump and a prime mover can be explained from a thermo-
dynamical perspective by taking into account the first and second law of thermodynamics. The
first law states that energy is always conserved, which is mathematically described by the sum
of the incoming and outgoing energy flows being equal equal to the change in internal energy
(see eq. 1.2). Where ∆U is the change in internal energy, Q the heat added to the system, and
W the work done by the system. If steady operation is considered, the change in internal energy
∆U is zero. The second law states that the entropy of an isolated system never decreases. In
practive this implies that for a hot and a cold heat exchanger, without any work done, the heat
will always flow from the hot to the cold heat exchanger.

∆U = Q − W (1.2)

An overview of the resulting energy flows for a heat pump and prime mover are shown in
Figure 1.6. From which it can be seen that (acoustic) work is needed to extract heat from a
cold reservoir and move it to the hot reservoir (heat pump function). The other way around,
(acoustic) work can be generated when heat can flow from the hot reservoir to the cold reservoir
(prime mover function).

A more detailed explanation of the thermoacoustic effect, and the conversion between acous-
tic energy and heat in both a standing-wave and traveling wave device, can be given by looking
at the thermodynamic cycle the process cycles through. A thermodynamic cycle is a collection
of thermodynamic processes of both heat and work, which always returns to its initial state. A
pressure-volume diagram (see Fig. 1.7) is a useful way to optimize the efficiency and operation
of the cycle. Because the process is a cycle, the process always encircles an area within this
diagram by definition. This surface area (within A-B-C-D in Figure 1.7) is equal to the power
absorbed or produced. This power, in respect to the heat supplied or extracted, is a figure of
merit for efficiency.

For a traveling wave device, the porous structure in between the heat exchangers needs to
have good thermal contact (heat is exchanged quickly) for the area (and thus power) of the
thermodynamic cycle in the P-v diagram to be maximized. The thermodynamic cycle of the
traveling-wave device now resembles a so-called Stirling cycle (see Fig. 1.7a), and the porous
structure is now called a regenerator. A standing wave device needs a time-lag for it to maxi-
mize the surface area in the P-v diagram. The time-lag can be accomplished by an imperfect

10
 Figure 1.6: Heat pump and prime mover energy flows [7]

Figure 1.7: Standing and traveling wave thermodynamic cycles in P-v diagram [7]

(slow heat transfer) thermal contact. The porous structure in a standing wave device is called
a stack, and it resembles a Brayton thermodynamic cyle (see Fig. 1.7b).

The difference between the Stirling and Brayton cycles is the manner in which the path
between states is traversed. A state can be seen as a point within the thermodynamic cycle
for which the state variables are known (for example: pressure and temperature), and between
different states the type of thermodynamic process varies. A gas turbine or a jet engine are
examples of Brayton cycles. A stirling engine is an example of a Stirling cycle. Without go-
ing into too much detail about these cycles, the Brayton cycle in the standing-wave device is
theoretically considered the least efficient of the two, because of the imperfect heat transfer
accompanying the time-lag that increases entropy. An increase in entropy results in a new state
that is “less able” to do work compared to a reference state.

The stack and regenerator also disturb the flow in other ways. The increased surface area
causes extra viscous losses. Also, because of the increased viscous effects, part of the acoustic
waves are reflected. These reflections are usually unwanted in a traveling wave device, as the

11
reflections will cause standing waves to form.

To give a more intuitive explanation, the thermoacoustic effect can be thought of as fluid
parcels that move energy over a small distance each, but together over a large distance. A bit
like firefighters did long ago, standing in a long line passing the buckets of water over from
one person to another. From this analogy comes the term ”Bucket-brigade effect”, which is
graphically represented in Figure 1.8.

Figure 1.8: Bucket-brigade effect [7]

A fluid parcel interacts thermally with the boundary and heat is transferred to the fluid
parcel. The fluid parcel then moves a small distance to either a hotter or colder region of the
boundary, before it transfers the heat back to the boundary. If the region is colder, the system
functions as a prime mover. If the heat is released at the hotter region, the system functions as
a heat pump.

1.2 Thermoacoustics in computational fluid dynamics

In numerical analysis of thermoacoustic devices often frequency domain solvers are used. For
example, thermoacoustic software DeltaEC uses the second order Helmholtz equation by assum-
ing sinusoidal time dependence of the wave equation, and an energy equation added optionally
[25]. Many other assumptions were also made by using these equations as a basis, among which
laminar and one dimensional flow are two. For a lot of thermoacoustic research, frequency
domain solvers are sufficient, but for certain applications a time-domain solver is needed, or
preferred. Such applications include analysis of complex flow patterns, and streaming effects.

The Helmholtz equation has frequency and position as its independent variables, and can be
evaluated using impedance, or the complex reflection coefficient, at bounding surfaces or points
of the domain. Complete thermoacoustic parts can be represented by an impedance or reflection
coefficient. This simplifies the thermoacoustic problems considerably, and decreases domain size.

Impedance is a quantity that relates the amount of sound pressure that is generated to acous-
tic particle velocity. It can be a characteristic of the medium, called characteristic impedance,
or from an acoustic component. The reflection coefficient relates the pressure of the incoming
acoustic wave to the reflected acoustic wave at a bounding surface of an acoustic component.
Impedance and reflection are complex quantities defined in the frequency domain.

12
 For time-domain fluid dynamics equations such as the Navier-Stokes equations, and its de-
rived Euler equations and wave equation, these quantities cannot be used directly because they
are only defined in the frequency domain. Conventional boundary conditions, such as a con-
stant pressure outlet, usually behave acoustically fully reflective. One can circumvent these
problems in CFD by modeling the whole domain. This is computationally very expensive, es-
pecially considering one is usually only interested in a certain part of the thermoacoustic device.

A potential solution to decrease domain size, while maintaining correct acoustical proper-
ties, is to use so-called Time-Domain Impedance Boundary Conditions (TD-IBC). These type
of boundary conditions mimic the impedance, or the complex reflection coefficient, in the time-
domain. Several papers are written about this topic in literature, but none are actually available
in commercially available time-domain CFD packages such as ANSYS CFX and ANSYS FLU-
ENT. The goal of this master thesis is to implement and test a time-domain impedance bound-
ary condition in ANSYS CFX, and provide a theoretical reference on the matter for future work.

1.3 Thesis outline

This report is sectioned in a very conventional structure, in chronological order: theory, method,
results, and conclusions. The actual process during the research did not strictly follow this order,
especially between method and results. Within the method chapter often references towards
the “future” results chapter will be made, while the method chapter itself is (i.e. the boundary
conditions), in a way, also a result of this thesis.

Chapter 2 is the Theory chapter, which gives the underlying theory of constructing a
TD-IBC. Also related theory such as the Navier-Stokes Characteristic Boundary Conditions
(NSCBC), the Linear Relaxation Method (LRM), Plane-Wave Masking (PWM), and the multi-
microphone method are discussed.

Chapter 3 covers the methodology of simulations and construction of the boundary condi-
tions in ANSYS CFX. An overview of the test cases are given, and their rationale is explained.
Also an overview and the main design choices of the programmed Fortran subroutines in AN-
SYS CFX is discussed.

The last two chapters, chapters 4 and 5, covers the results, conclusions, and recommen-
dations. The results chapter features the influence of the relaxation factor on the reflection
coefficient, the influence of mean flow, an analysis of pressure drift, and an analysis of the error
in the complex plane of the TD-IBC as the most important topics. To prevent clutter and a
plethora of numbers, results are mainly presented by graphs. The actual data used to construct
the graphs can be found in the appendix. In the last chapter a short overview of the conclusions
and recommendations are given.

References to the appendix are not given specifically in the report, but just referenced to
the appendix as a whole. The specific page numbers and headings can be found in the table of
contents of the appendix, at the beginning of the appendix.

13
Chapter 2

Theory

Different methods for constructing a Time-Domain Impedance Boundary Condition (TD-IBC)

exist. A method created by Fung uses a direct inversion of the reflection coefficient R from
the frequency domain to the time-domain, instead of the (equivalent) impedance Z (see section
2.2), but is not always stable and relatively difficult to implement [4][5]. For this master thesis
a different method is used, a method created by Polfike et al and based on a so-called filter
method [9][6]. The reason behind this choice is the inherent modularity of the approach. This
modularity allows for a step-wise implementation and debugging, and can more easily be tested
against theory and other implementations at various points in the process. The TD-IBC consists
of two main parts, a non-reflective part, and a TD-IBC specific part. The non-reflective part is
based on a method for constructing boundary conditions called the Navier-Stokes Characteristic
Boundary Condition (NSCBC) formalism (see section 2.3) [22][15]. The non-reflective part can
function independently as a non-reflective boundary condition when the TD-IBC specific part
is not added to the formulation. The TD-IBC specific part consists of the filter method (see
section 2.6).

Optionally other terms can be added to the boundary specification. To solve stability prob-
lems discussed in section 2.4, a so-called Linear-Relaxation Method (LRM) can be used [22].
The downside of the LRM method is additional reflections. These additional reflections can sub-
sequently be mitigated by adding a Plane-Wave Masking (PWM) term [16]. Strictly speaking
the LRM method is not required, for example for certain short simulations which remain stable
within its simulation time, but in practice the LRM term must be added to prevent stability
problems.

The construction of the TD-IBC is presented graphically in Figure 2.1. All contributions
can be added as a sum of terms, similar to the superposition principle in linear systems.

Figure 2.1: Construction of a TD-IBC boundary condition

14
 ANSYS CFX already provides a non-reflective boundary condition as a beta feature. Con-
veniently, this built-in non-reflective boundary condition is also based on the NSCBC formalism
using the Locally One-Dimensional Inviscid (LODI) relations, and the LRM method for stabil-
ity [26]. See section 2.3.1 and 2.3.2 for details about the LODI relations. The ANSYS CFX
non-reflective boundary condition (NRBC) will be used as a reference in the results chapter.

2.1 Boundary Conditions

Differential equations in fluid dynamics need additional constraints at the boundaries to be able
to find a solution, thermoacoustic applications in CFD will be no exception to this. These type
of differential equations are called boundary value problems, and the accompanying constraints
are called boundary conditions. Mathematically, a problem that has a solution that is unique,
exists, and does not behave chaotically is called a well-posed problem. To state that a problem
is well-posed in CFD problems based on the Navier-Stokes equations is in general not possible,
because it has never been proven that a solution always exists. One can therefore, in this case,
only say that a problem is more well-posed, or less well-posed, compared to a reference problem
based on the Navier-Stokes equations.
Different kind of boundary conditions exist of which certain types are considered “stronger”
than others, i.e. resulting in a problem being more well-posed. In practice this statement means
that a solver (software that calculates the solution) is usually more able to find a solution to the
problem if the boundary condition is considered stronger. A few important boundary conditions
are summed up in Table 2.1. The boundary conditions are explained using a model differential
equation where y is the dependent variable of the independent variable x, Ω is the domain, ∂Ω
is the boundary of the domain, α(x) and β(x) are given values on the boundary, and a and b
are constants which may, or may not, be a function of x.

Table 2.1: A few important types of boundary conditions

Boundary Condition Formula
Dirichlet y(x ∈ ∂Ω) = α
Neumann y 0 (x ∈ ∂Ω) = α
Cauchy y(x ∈ ∂Ω) = α and y 0 (x ∈ ∂Ω) = β
Robin a · y(x ∈ ∂Ω) + b · y 0 (x ∈ ∂Ω) = α

The Dirichlet boundary condition specifies the values of the dependent variable (y) at the
boundary, and the Neumann boundary condition specifies the derivative of the dependent vari-
able (y 0 ) at the boundary. A Dirichlet boundary condition is considered stronger than a Neu-
mann boundary condition (in differentiation information is lost, the integration constant). The
others are combinations of aforementioned boundary conditions. The Cauchy boundary con-
dition specifies both y and y 0 , while the Robin boundary condition specifies y via a linear
combination of both y and y 0 .

After derivation of the theory about non-reflective and TD-IBC boundaries later on in the
this chapter, it will become clear that the resulting boundary conditions are of a “weaker” type
than a constant Dirichlet boundary condition often used in CFD (e.g. constant pressure outlet).
This can result in stability and convergence problems.

15
2.2 Governing equations
The most important set of equations in fluid dynamics are the Navier-Stokes equations. These
equations can describe many flow problems, and lie at the basis of many derived and simpler
equations. Important equations in acoustics, such as the wave equation, Helmholtz equation,
and the Euler equations are also derived from the Navier-stokes equations. The notation and
form of the Navier-Stokes equations differ. In this thesis a compact differential form (eq. 2.2)
using four column vectors (eq. 2.1) is used to enhance readability, but other forms exist. The
differential form is also used in the characteristic analysis later on in section 2.3.

~0 
       
ρ  ρ(~u − ~u∂V )    0 
inv visc
U = ρ~u ~
F = ρ~u(~u − ~u∂V ) + pI¯ ~
F = τ̄¯ J= ρf~ (2.1)
¯  ~
ρE ρE(~u − ~u∂V ) + p~u τ̄ ~u − ~q ρf · ~u + Q̇
     

∂U ~ ~ inv ~ visc
+ ∇ · (F − F ) = J (2.2)
∂t
In equation 2.1 and 2.2, U is a column vector with the conserved quantities, F ~ inv are the
~ visc are the viscous flux terms, and J are the source terms. ρ is the density,
inviscid flux terms, F
~u is the velocity vector, E the total specific energy (eq.2.3), p the pressure, τ̄¯ the stress tensor,
~q the heat conduction, f~ the gravity (or any other arbitrary volumetric force field), and Q̇ as
a volumetric heat source. The three-dimensional Navier-Stokes equations contain 7 unknowns
and 5 equations, leaving the system underdetermined. To solve the system of equations, two
additional equations need to be supplied, which are usually state equations such as the ideal
gas law (eq. 2.4).

1
E = e + |u|2 (2.3)
2

P = ρRT (2.4)
In acoustics, the viscous terms and heat conduction terms are generally small, and can be
neglected. From the compact form of the Navier-Stokes equations (eq. 2.2), the Euler equations
~ visc vector.
(eq. 2.5) can easily be constructed by removing the viscous flux terms F

∂U ~ ~ inv
+∇·F =J (2.5)
∂t
From the Euler equations (eq. 2.5) a partial differential equation important to acoustics can
be derived, which is called the wave equation (2.6). An important form of the wave equation
is the scalar velocity potential form in equation 2.7, because only one variable instead of three
needs to be solved for the three spatial dimensions. Apart from its application in acoustics, the
wave equation can also be used to describe other wave phenomena such light waves and water
waves.

∂2u
= c2 ∇2 u (2.6)
∂t2

~ 2 φ − 1 φtt = 0
∇ ~ = ~u
∇φ (2.7)
c20
The wave equation is a hyperbolic second-order linear Partial Differential Equation (PDE).
Solutions of hyperbolic equations exhibit wave-like behavior. Any disturbance made to the

16
initial conditions are not felt immediately throughout the domain, but instead are propagated
at a finite propagation speed. For the wave equation this propagation speed is equal to the
characteristic speed c0 . The wave equation is also a linear PDE, which means solutions can
be added to form other solutions (called superposition principle). The general solution to the
wave equation is given by equation 2.8. The functions f and g represent right and left traveling
waves respectively, and the solution is thus a sum of both (which can be a solution according
to the superposition principle).

u(x, t) = f (x − ct) + g(x + ct) (2.8)

Another important equation in acoustics is the Helmholtz equation. The Helmholtz equation
can be seen as a time-independent version of the wave equation. Using separation of variables,
and assuming time-harmonic behavior, the solution can be written in the form shown in equation
2.9. Substitution of this form of the solution in the wave equation (eq. 2.7) results in the
Helmholtz equation (eq. 2.10) (see appendix for detailed derivation). The wave equation can
also be expressed in the pressure amplitude δ p̂ instead of the velocity potential φ (eq. 2.11).
The solution of the Helmholtz equation for the pressure, is of the form given by equation 2.12,
which is, unlike the solution of the wave equation, not an equation of space x and time t, but of
space x and (angular) frequency ω. An important advantage of this equation is that frequency
dependent quantities such as impedance and reflection can be used directly.

φ(~x, t) = φ(~x, ω)eiωt (2.9)

~ 2 φ(~x, ω) + k 2 φ(~x, ω) = 0 ω
∇ where k= (2.10)
c

∂ 2 δ p̂
+ k 2 δ p̂ = 0 (2.11)
∂x2

δ p̂(x, f ) = A(f ) · e−ikx + B(f ) · eikx (2.12)

An often used property to describe the acoustic behavior of media or acoustic components
is the acoustic impedance Z (eq. 2.13). The acoustic impedance relates the acoustic pressure
amplitude δ p̂ to the acoustic velocity amplitude δ û, and is used for acoustic components. A
special type of impedance is the so-called characteristic acoustic impedance Z0 , and is used to
describe the acoustic behavior of a medium (eq. 2.13). A high impedance means that a lot of
acoustic pressure is needed to move the fluid parcels, while a low impedance means the fluid
parcels are relatively easily moved.

δ p̂
Z= Z0 = ρ0 c0 Z∈C (2.13)
δ û
The acoustic impedance is equivalent to the reflection coefficient R, given by equation 2.14.
Where f and g are so-called characteristic wave amplitudes given by equation 2.8, and explained
in section 2.3. The reflection coefficient R is the ratio of the acoustic pressure that is reflected
(δ p̂− ) from an acoustic component over the outward traveling acoustic pressure (δ p̂+ ), and is
a measure for reflectivity. The part that is not reflected is transmitted (δ p̂tr ). The reflection
coefficient R can become complex, signifying a phase difference between the incoming and
outgoing waves.

δ p̂−
R= R∈C (2.14)
δ p̂+

17
 Z − Z0
R= (2.15)
Z + Z0
In acoustics three special cases can be distinguished, a hard wall, a pressure release surface,
and a matched impedance. These cases are explained using Figure 2.2, which represents two
domains connected by an interacting surface (interface). When it will take a lot more acoustic
pressure to move the fluid parcels in domain 2 compared to domain 1 (Z2 >> Z1 ), the interface
is called a hard wall and the reflection coefficient R will be 1. When on the other hand the fluid
parcels move far more easily in domain 2 compared to 1 (Z2 << Z1 ), the interface is called a
pressure release surface and the reflection coefficient R will be -1. The last special case is the
matched impedance, in which the impedance on both sides are equal (Z2 = Z1 ), and is thus
non-reflective (R = 0).

Figure 2.2: Reflection caused by an impedance difference

A non-reflective boundary is equivalent to a matched impedance surface. A hard wall can

be simulated in CFD by using a constant velocity outlet (e.g. wall boundary condition), and a
pressure release surface by using a constant pressure outlet boundary condition.

2.3 Navier-Stokes Characteristic Boundary Conditions

As is mentioned in the introduction (sec. 1.2), the conventional boundary conditions in CFD
software fully reflect acoustic waves back into the domain. These conventional boundary con-
ditions are however quite “strong” in the sense of well-posedness (see sec. 2.1 for more infor-
mation). Ideally a non-reflective boundary condition, as the name implies, should let acoustic
waves propagate through the boundary undisturbed, but at the same time be as well-posed
enough for the solver to be able to find a solution. In practice however, a non-reflective bound-
ary condition is always a concession between well-posedness and reflectivity, which will become
clear in section 2.4.

The NSCBC formalism used in the non-reflective part of the TD-IBC is a method to con-
struct boundary conditions based on the characteristic wave relations of the (modified) hyper-
bolic Euler equations. The characteristic wave relations can be derived by a method called
method of characteristics, and applies to hyperbolic PDEs such as the Euler equations. The
method of characteristics is a technique to decouple a set of coupled partial differential equations
into a set of decoupled ordinary differential equations (ODEs), in order to solve the problem
more easily. The hyperbolic one-dimensional wave equation (eq. 2.6) is considered one of the
simplest hyperbolic equations, and is used in the next paragraph as an example to explain a
few important properties of hyperbolicity and the method of characteristics. The solution of
hyperbolic PDEs are considered “wave-like”, which means that disturbances made to the initial
conditions propagate at a finite speed (called characteristic speed ) trough space and time, and
hence, are not felt immediately in every point of the domain. In case of the one-dimensional wave
equation, the characteristic speed is equal to the sound speed c0 and couples both independent
variables x (space) and t (time), of which so-called characteristic curves (or just characteris-
tics) can be drawn ,originating from an initial condition in a diagram called the characteristic

18
diagram (see Fig. 2.3b). By using the method of characteristics a set of characteristic waves
(also called Riemann invariants) can be constructed, which are variables that are constant
along the accompanying characteristic curves. An expression for these characteristic waves are
comprised of the dependent variables, and their coefficients from the originating equation. For
the one-dimensional wave equation the characteristic waves f and g are given by equation 2.16.
f and g are equivalent to the same variables in equation 2.8 for the general solution to the wave
equation.
   
1 δp 1 δp
f= + δu g= − δu (2.16)
2 ρ0 c0 2 ρ0 c0
Where δp is the acoustic contribution to the pressure (or simply acoustic pressure), and δu
the acoustic contribution to the velocity (or simple acoustic velocity).
Characteristics and characteristic waves can also be shown graphically, as is done in Figure
2.3. In Figure 2.3a one can see the acoustic wave traveling to the right, and in Figure 2.3b one
can see the accompanying characteristic in the characteristic diagram.

Figure 2.3: Characteristics for the wave-equation

The Euler equations are more complex than the above wave equation, and characteristic
analysis in section 2.3.1 will result in five different characteristics speeds and characteristic
wave expressions. The advantage of the method of characteristics is that different contributions
to the solution can be separated, of which each can be integrated in time via relatively easy
expressions to obtain the next value in time. As one will observe in the characteristic analysis
in section 2.3.1, two of the five characteristic wave expressions are essentially the same as the
characteristic wave expressions given in equation 2.16 for the wave equation, except for the used
NSCBC form being a function of the gradients of the primitive quantities pressure and veloc-
ity. The constructed non-reflective boundary condition, the NSCBC formalism exploits these
properties to calculate and set the pressure at the boundary for the next timestep accordingly.

More extensive analysis and a more detailed derivation of the NSCBC formalism is added
to the appendix.

19
2.3.1 Characteristic analysis
To be able to derive the characteristic wave relations, a modified version of the Euler equations
must be used. One might think that the use of the Euler equations to construct the NSCBC
method severely limits the types of flow that can be modeled, but this is not necessarily the
case as the NSCBC method is only used locally (near the boundary) to calculate and specify
boundary values, and does not bother with the calculation of the interior of the domain. The
approximation at the boundaries is accurate for flows that are pre-dominantly hyperbolic in
nature, which is the case for flows with a Reynolds number (see eq. 2.17) significantly higher
than one. In the asymptotic limit of Re → ∞, the momentum equation of the Navier-Stokes
equations leads to the hyperbolic Euler equations. On the other hand, if the Reynolds number
goes to zero, the flow is dominated by viscosity µ and will result in the elliptic Stokes equation
[24]. Often thermoacoustic devices use low viscocity fluids such as air or helium as a medium,
to minimize viscous losses, and so the approximation is justified. It is therefore possible to use
the full Navier-Stokes equations as the governing equations for the interior of the domain, and
use the NSCBC method based on the modified Euler equations to calculate the values at the
boundary.

ρU∞ L
Re = (2.17)
µ
Where U∞ is the free stream velocity, and L is the characteristic length of the domain.
In the characteristic analysis leading to the NSCBC formalism, Thompson modifies the one-
dimensional Euler equation in such a way that it fits the quasi-linear advection equation (see
eq. 2.18) [22]. This quasi-linear advection equation is a model equation often used to determine
if a system of PDEs is hyperbolic, and subsequently decouple the set of PDEs and put in
characteristic form.

∂Up ∂Up
+ A1 (Up ) =0 (2.18)
∂t ∂x1
Where Up denotes the primitive quantity vector instead of the conserved quantity vector
U, because it simplifies derivation [22]. The primitive quantity vector is given in the following
expression: Up = {ρ, p, ~u}−1 . What exactly constitutes “modified” in the modified Euler
equations are three assumptions and modifications, listed below.

1. A quasi-linear form realised by using a primitive quantity vector Up .

2. Neglecting transverse terms, because else the system cannot be put in characteristic form
for mathematical reasons (see appendix). This basically results in a system with one
dominant direction of propagation: the x-axis.

3. Neglecting source terms because they do not significantly contribute to the analysis, and
only complicate matters.

A more detailed explanation about the assumptions, modifications (i.e. quasi-linear form)
and the full derivation of the characteristic form of the above modified quasi-linear Euler equa-
tion (eq. 2.18) is added to the appendix.

Eventually Thompson arrives at five expressions for the temporal derivative of the primitive
field quantities (see eq. 2.19). These temporal derivatives are a function of the so-called ampli-
tude variations Li , which are time variations in the amplitude of the characteristic waves (see
eq. 2.20). The amplitude variations are in their turn functions of the gradients of the primitive
quantities at the boundary. The complete set of equations given by equations 2.19 and 2.20 are

20
called Locally One-Dimensional Inviscid (LODI) relations.

∂ρ
∂t = − c12 [L2 + 12 (L5 + L1 )]
∂p
∂t = − 12 (L5 + L1 )
∂u1 1
∂t = − 2ρc (L5 − L1 ) (2.19)
∂u2
∂t = −L3
∂u3
∂t = −L4

∂p
L1 = λ1 ( ∂x1
− ρc ∂u
∂x1 )
1

∂ρ ∂p
L2 = λ2 (c2 ∂x1 − ∂x1 )
L3 = λ3 ∂u
∂x1
2
(2.20)
∂u3
L4 = λ4 ∂x1
∂p
L5 = λ5 ( ∂x1 + ρc ∂u
∂x1 )
1

The L1 and L5 amplitude variations are the relevant variations for the construction of a
non-reflective boundary condition. It is assumed that the non-reflective boundary is flat and
has a surface normal pointing outwards (of the domain) along the positive x-axis. If this is
not the case, the equations must be transformed in terms of a local coordinate system with
one coordinate perpendicular to the boundary. This does not add anything to the analysis,
but does require extra work, and is a potential source of error. Based on the aforementioned
assumptions, L1 is the amplitude variation of the incoming characteristic wave, and L5 of the
outgoing characteristic wave. In Figure 2.4 the incoming and outgoing amplitude variations of
the characteristic waves are visualized.

Figure 2.4: Amplitude variation waves L1 and L5

The temporal derivatives of the different primitive variables can be used in conjunction
with a numerical time integration scheme to calculate the values at the next timestep. These
variables can be set in a CFD code to provide a Dirichlet boundary condition. In the case of
a non-reflective boundary condition, the most ideal primitive quantity to set is the pressure,
because of two reasons. The first being the fact that only the “acoustic” characteristics L1 and
L5 are used the specification of the temporal derivative of the pressure. The second reason is
that specifying the pressure instead of the velocity at an outlet boundary condition is consid-
ered more stable in ANSYS CFX [23]. However, this does not restrict non-reflective boundary
conditions to be of the “pressure outlet” type, velocity specified and other type of boundaries
could also be created based on the NSCBC method.

As L1 is the amplitude variation of the incoming, and thus reflected, wave, it should ideally
be zero in a non-reflective outlet boundary condition. By explicitly assigning a value of zero
to L1 , and calculating L5 via one-sided differences of the gradients of the pressure and velocity
near the boundary, an expression for the time-derivative of the pressure for a non-reflective
boundary condition is constructed.

21
2.3.2 Determination of outgoing wave amplitude variation
The method based on the LODI relations to determine the outgoing amplitude variation L5
is not the only method of determining L5 . Another method to determine L5 uses a slightly
different, but equivalent, expression for the characteristic wave f (eq. 2.21, based on eq. 2.16).
The characteristic wave f , which now has units of velocity (m/s), cannot be determined at the
boundary because both values for pressure and velocity are needed. As the whole boundary
condition is constructed such that pressure is specified, the pressure itself cannot be used in
calculating the pressure, because it would create an infinite recursive loop. The pressure and
velocity is retrieved on a sample plane near the boundary, and offset in time (using the sound
speed) towards the boundary. The method is not mentioned by Thompson [22] and Poinsot
[15], but mentioned by Polifke [16] in conjunction with the Plane-Wave masking method.

 
1 δp
f= + δu (2.21)
2 ρ0 c0
The quantities used in the expression for characteristic wave f are not pressure and velocity,
but the acoustic pressure and velocity. These quantities cannot be directly retrieved from the
flow solution, and must be derived from expressions. The expressions for determining the acous-
tic contributions of the pressure and velocity are the same as used in the Plane-Wave masking
method, and treated in section 2.5.

2.4 Linear relaxation method

The NSCBC method to construct a Non-Reflecting Boundary Condition (NRBC) will result
in a drift of the mean pressure with time, and ultimately convergence problems [15]. Funda-
mentally, boundary conditions carry information from the boundary to the interior in order
to find a solution to the problem. By eliminating the amplitude variation L1 of the incoming
characteristic wave, no information about the pressure is able propagate from the boundary to
the interior [17]. In other words, the boundary is “invisible” to the interior of domain, and the
problem effectively becomes ill-posed.

The most used, and relatively simple, solution to the problem is the Linear Relaxation
Method (LRM) [15][17]. This method is also used by ANSYS CFX in its NRBC to prevent
pressure drift [26]. In the LRM method, the expression for L1 is adjusted to include a term that
restores a deviation from a target pressure, represented by p∞ (eq. 2.22). The target pressure
p∞ , sometimes called the far-field pressure, is the pressure to which one wants the pressure
to be restored. The relaxation factor K is a factor that represents the amount by which the
deviation is restored.

L1 = K · (p − p∞ ) (2.22)
Of course, this term will by definition cause reflections. The non-reflective boundary con-
dition using the LRM method for specifying L1 will always be a compromise between drift
(convergence) and reflection. The better the target pressure is enforced using relaxation factor
K, the higher the reflection coefficient R (eq. 2.14), and vice versa. But also frequency plays
a part, the higher the frequency, the less time the LRM term has to correct the deviation in
pressure, resulting in a lower reflection coefficient, while for a low frequency the LRM term
can easily correct for the pressure deviation within the period of the wave, and results in a
higher reflection coefficient. Selle devised an analytical expression to determine the theoretical
complex reflection coefficient R as a function of frequency (see Fig. 2.5) [17]. The magnitude

22
||Rtheo || and phase ψ of the complex reflection coefficient R are given by equation 2.23.

1 2ω
||Rtheo || = q and ψ = −π − arctan( ) (2.23)
1 + ( 2ω 2 K
K)

Figure 2.5: Theoretical reflection coefficient vs. frequency (K = 501/s)

ANSYS CFX uses an expression proposed by Selle to determine an optimal relaxation factor
K (eq. 2.24) [17]. The parameters that can be used to determine the optimal value for K are
the free stream mach number M, the sound speed c0 , a characteristic length L, and a coupling
parameter σ. The scaling method from Selle is based on domain dimensions and eigenfrequen-
cies of the domain. Two optimal coupling parameters σ are mentioned by Selle, a value of 0.58
based on numerical simulations, and a value of 0.27 based on theory. See appendix for more
information.

(1 − M2 )
Kopt = σ (2.24)
L
The ANSYS CFX NRBC is tested and compared to the theoretical values by Oosterhuis
[14]. Results showed a good agreement with theoretical values for higher relaxation factors, but
showed a significantly higher reflection coefficient for lower relaxation factor. Proposed hypoth-
esis for the phenomenon where a loss in accuracy due to numerical discretization schemes used,
and a built-in mechanism in ANSYS CFX that is deliberately adjusting the boundary values
to prevent pressure drift and convergence problems.

2.5 Plane-Wave Masking

As is discussed in section, 2.4 the LRM method is always a compromise between pressure drift
and reflection. Polifke introduces a method called Plane-Wave Masking (PWM) to minimize the
reflection caused by the LRM method, especially at lower frequencies (see Fig. 2.5). The PWM
method determines the amplitude of the characteristic wave f (eq. 2.21) at a sample plane near
the boundary, and off-sets it in time until it reaches the boundary. The acoustic contribution of
the characteristic wave f traveling towards the boundary is then subtracted from the LRM term
(eq. 2.25), effectively “masking” the pressure deviation caused by acoustics. What theoretically
remains are pressure deviations caused by mean pressure drift and turbulence.

23
 L1 = K · (p − p∞ − ρ0 c0 f ) (2.25)
The acoustic pressure and velocity cannot be directly retrieved from the flow solution. To
determine the acoustic pressure and velocity, Polifke et al proposes to separate the pressure
into three different contributions (see eq. 2.26). Where p0 is the mean pressure, δp is the
acoustic pressure, and p0 is the turbulent component of the pressure. The velocity u is separated
analogously in equation 2.27.

p(~x, t) = p0 (~x) + δp(x, t) + p0 (~x, t) (2.26)

u(~x, t) = u0 (~x) + δu(x, t) + u0 (~x, t) (2.27)
Polifke assumes that the characteristic wave f is a plane-wave that travels perpendicular to
the sample plane. Mathematically this is equal to saying that the integrated area average over
the sample plane, signified by the angled brackets, is equal to f (see eq. 2.28). The acoustic
wave can be assumed a plane-wave if the acoustic boundary layer is thin, which it is in the
numerical problem defined in section 3.

f ≈ hf i (2.28)
Then Polifke reasons that if the turbulent length scale is sufficiently small, the area average
of the turbulent contribution to both pressure and velocity practically vanishes (see eq. 2.29).
No turbulence models will be used in the setup of the numerical problem, and thus this is indeed
the case in this report (see Chapter 3). With the turbulent contribution assumed to be zero,
equations 2.26 and 2.27 can be rewritten for the area averaged acoustic pressure and velocity
(see eq. 2.30).

hp0 i ≈ 0 hu0 i ≈ 0 (2.29)

p̃ ≡ hp − p̄i ũ ≡ hu − ūi (2.30)

The area averaged acoustic pressure and velocity are now used in the expression for charac-
teristic wave f (eq. 2.21), which results in an area averaged expression for f (eq. 2.31).
 
1 p̃
f= + ũ (2.31)
2 ρ0 c0
The mean pressure can be taken as the target pressure at the boundary p∞ , or as a time
averaged pressure. The pressure is retrieved from the solution field.

2.6 Time-Domain Impedance Boundary Conditions

The TD-IBC specific part is based on the LRM model and PWM for specifying L1 (eq. 2.25),
but extended by an external perturbation g ∗ which is calculated using so-called filter methods
[9][6]. Using filter-methods (explained later on) the new expression for L1 is given by equation
2.32. The external perturbation g ∗ within the brackets, preceded by the relaxation factor K, is
for “masking” the contribution of g ∗ to the LRM term, i.e. reduce unwanted reflections caused
by g ∗ . The temporal derivative of g ∗ is the actual TD-IBC contribution.

∂g ∗
L1 = K(p − ρc(f + g ∗ ) − p∞ ) − ρc (2.32)
∂t

24
 From an arbitrary reflection coefficient, or equivalently, impedance (see sec. 2.2), an external
perturbation g ∗ with the desired properties mimicking the aforementioned reflection coefficient
can possibly be obtained. Possibly, because for certain reflection coefficients that are a function
of multiple frequencies, the process of determining g ∗ can become very difficult, of which the
details will be explained in section 2.6.1 and 2.6.3.
To construct g ∗ , first the definition of the reflection coefficient 2.14 in the frequency domain
is rewritten with the backward traveling wave ĝ expressed as the product of the reflection
coefficient R and the forward traveling characteristic wave amplitude fˆ in equation 2.33.

gb(ω)
R(ω) = → gb(ω) = R(ω)fb(ω) (2.33)
fb(ω)
Then the product (eq. 2.33) is transformed to the time domain using the property that a
product of two functions in the frequency domain is equivalent to a convolution in the time-
domain, resulting in equation 2.34. Where τ is a time coordinate, comparable to t. The integral
is evaluated from 0 to ∞ instead of −∞ to ∞, because values for τ lower than 0 correspond to
future values, which are not available in the present. This is equivalent to stating the system
must be causal in time.
Z ∞
g(t) = (R ∗ f )(t) = R(τ )f (t − τ )dτ (2.34)
0
The evaluation of a convolution integral (see eq. 2.34) is computationally very expensive
and difficult. To be able to solve this convolution integral, equation 2.33 is converted into the
Laplace domain by exchanging the arguments via the relation s = iω, and the Z-transform is
used to solve the equation at discrete timesteps. The Z-transform is also called the “discrete
Laplace transform”, and allows for a relatively efficient evaluation of equation 2.33. Using the
Z-transform, equation 2.33 can be rewritting in the z-domain as equation 2.35. It is then as-
sumed that R(z) can be approximated by the transfer function syntax in equation 2.36.

G(z) = R(z)F (z) (2.35)

a0 + a1 z −1 + a2 z −2 + . . . + an z −n
R(z) = (2.36)
1 − b1 z −1 + b2 z −2 + . . . + bm z −m
The expression for R(z) in equation 2.36 can be substituted in equation 2.35, and rewritten
into equation 2.37. By using the time shift property of the z-transformation (see eq. 2.38) one
arrives at a time-domain expression for the external perturbation g ∗ in equation 2.39.

(1 − b1 z −1 + b2 z −2 + . . . + bm z −m )G(z) = (a0 + a1 z −1 + a2 z −2 + . . . + an z −n )F (z) (2.37)

Z{x(i − k)} ↔ z −k X(z) (2.38)

g(i) = a0 f (i) + a1 f (i − 1) + . . . + ai f (i − n) − b1 g(i − 1) − . . . − bm g(i − m) (2.39)

Equation 2.39 is rewritten in short summation notation in equation 2.40, and expressed in
time t and time increment ∆t.
n
X m
X
g(t) = ai f (t − iδt) − bj g(t − jδt) (2.40)
i=0 j=1

25
 Which is the sum of the present and past incoming and outgoing characteristic wave ampli-
tudes multiplied by a constant. The constants are called filter coefficients, and are determined
based on the desired complex reflection coefficient as a function of frequency in a process called
complex-curve fitting, which is discussed in section 2.6.1.

2.6.1 Determination of filter coefficients

The filter coefficients can be determined from the complex reflection in different ways, two of
which are discussed and used. If only a certain reflection coefficient at one given frequency is
needed, one can relatively easily calculate the corresponding filter coefficients from the time
delay tdelay corresponding to the phase delay at the aforementioned frequency (see eq. 2.41) by
noting that the different filter coefficients correspond to a different time delay of the incoming
and outgoing waves (see eq. 2.40). The filter coefficient in question can be calculated via
equation 2.42. Equation 2.42 can also be used to estimate the minimum number of filter
coefficients.

ψ
tdelay (f, φ) = ( ) · f −1 where ψ = phase(R) [rad] (2.41)
2π

tdelay = n · ∆t → an = abs(R) (2.42)

It must be noted however, that a constant time delay causes the phase of the reflection
coefficient to shift linearly with frequency, as Figure 2.6 shows qualitatively. But as the number
of filter coefficients increases, so do the memory requirements and susceptibility to quantization
error.

Figure 2.6: Phase shift at different frequencies

From equations 2.41 and 2.42 the filter coefficients needed to mimic an acoustically hard
wall (R = 1) and pressure release surface (R = −1) can also be obtained (see sec. 2.2). For an
acoustically hard wall the reflection coefficient has no phase, and thus time delay, and results
in setting filter coefficient a0 to 1. For a pressure release surface the phase ψ is −π, resulting
in a time delay of half the period of the incoming wave. For a non-reflecting outlet the filter
coefficients are all zero.

It becomes more difficult if the reflection coefficient is specified as a function of frequency at

multiple frequencies. It is assumed that the reflection coefficient R can be approximated as a
rational function (ratio of two polynomials) in equation 2.36, these polynomials are then fitted
to the given points by means of minimizing the error squared. An example of such a fit, done
by Polifke et al, is shown in Fig. 2.7 for a short duct [9]. This process of complex-curve fitting
is discussed in detail by E.C. Levi in [11].

In the construction of filters based on the transfer function (eq. 2.36) two types of filters
can be distinguished:

26
 Figure 2.7: Coefficient fit for a short duct [9]

ˆ Infinite Response Filter (IIR): History of both inputs and outputs are considered (ai , bi 6=
0).

ˆ Finite Response Filter (FIR): Only the history of the inputs are considered (ai 6= 0, bj = 0).

FIR filters are naturally stable and easier to fit (discussed in section 2.6.3). The IIR filters
can be unstable, but need (far) less historical information and are more accurate. The merits
of both filters will be investigated for constructing a TD-IBC in chapter 3.

2.6.2 Determination of normalized angular frequency

The argument z of the Z-transform is an important quantity in determining the filter coefficients.
The argument z reflects the frequency at which a reflection coefficient R should be enforced in
the filter method. The bilateral Z-transform is given by equation 2.43. Where n is an integer,
and z a complex number defined by equation 2.44. Because the system must be causal in time,
the sum is taken from n = 0 to n = ∞ in equation 2.43. This is called a unilateral Z-transform.
∞
X
X(z) = Z{x[n]} = x[n]z −n (2.43)
n=−∞

z = r · ejωn (2.44)
The complex argument ωn can be seen as a normalized frequency, which runs from 0 to π for
the causal discrete time systems in question. The frequency of the sampled wave f is normalized
by the sampling frequency fs (fs = ∆t−1 ) (see eq. 2.45). The Nyquist frequency states that
the frequencies above 0.5 times the sampling frequency aliases with lower frequencies, and can
therefore not be uniquely distinguished. From this fact the upper bound of π (in radians) can
be understood, as it corresponds to the maximum frequency that can be uniquely identified on
the grid.

f
ωn = · 2π [rad/sample] (2.45)
fs

2.6.3 Stability of filter coefficients

Unfortunately not all combinations of filter coefficients are unconditionally stable. The ques-
tion of stability is basically a question if all the poles of the transfer function (with which R
is approximated) reside within the unit circle in the complex plane (i.e. |z| ≤ 1). For simple
coefficient sets this is often the case, for more complex coefficient sets the problem becomes

27
more difficult. However, it might not lead to stability problems if the frequencies corresponding
to the poles are hardly or not present in the domain. Also, as FIR filters have no poles by defi-
nition, they are more stable. Results by Polifke showed that stable results can be obtained via
this method with the numerical results within an error margin of around 3% for laminar flow [9].

2.7 Reflection measurement theory

Different methods exist to quantify reflection or impedance of a sample or interface in numerical
simulations (but also in experimental set-ups). Among the most used are the Two-Microphone
Method (TMM), the Multi-Microphone Method (MMM), and the Standing-Wave Ratio Method
(SWR Method). The MMM is the method of choice, because the other two methods have cer-
tain disadvantages that make them less suitable for this application. The TMM is simple, but
very sensitive to sensor positioning, especially if the microphone separation is near one half
wavelength [18][8]. The SWR method is on the other hand very accurate and reliable, but
very time-consuming, because a probe with a pressure sensor needs to moved through the tube
[18][8]. As ANSYS CFX does not support moving monitor points (to measure pressure), and
approximating the Standing-Wave Ratio as a function of position might be a cumbersome task,
the SWR-Method is at a disadvantage.

The multi-microphone method is discussed briefly in section 2.7.1. All three methods are
discussed in more detail in the appendix.

2.7.1 Multi-microphone method

The MMM method calculates the reflection coefficient R (eq. 2.14) by solving a linear sys-
tem of equations for forward traveling pressure amplitude δ p̂+ and backward traveling pressure
amplitude δ p̂− , which constitute the definition of R. First the time-series of three or more
microphones, a certain distance d from the boundary, must be recorded (see Fig. 2.8 for place-
ment). On these pressure time-series a Fourier transform must be performed to extract the
pressure amplitude δ p̂ as a function of frequency at the different locations.

Figure 2.8: Multi Microphone Method in impedance tube

The linear system of equations, that will be constructed later on, is based on the solution
of the Helmholtz equation (eq. 2.12). Matters are simplified by introducing a new spatial
coordinate d originating from the boundary with the relation L − d = x (see Fig. 2.8), and
recognizing that δ p̂+ = A · e−ikL and δ p̂− = B(f ) · eikL (evaluated at the boundary). This
results in the solution of the Helmholtz equation given by equation 2.47.

δ p̂(x, f ) = A(f ) · e−ikL eikd + B(f ) · eikL e−ikd (2.46)

28
 δ p̂(di , f ) = p̂+ (f )eikdi + p̂− (f )e−ikdi (2.47)
Using the known pressure amplitude spectra of the microphones an overdetermined linear
system of equations of the form A~x = ~b can be constructed (eq. 2.48). The known pressure
amplitudes δ p̂1 ,δ p̂2 ,..., δ p̂n are in the right-hand side (RHS) vector ~b. The coefficient matrix A
consists entirely of complex powers of the euler number e. The only unknowns are δ p̂+ and δ p̂−
in the ~x vector.

e−ikd1   
 ikd1   
e p̂1 
 eikd2 e−ikd2  p̂+

 
p̂2


 ...
 = (2.48)
...  p̂− 
 ... 

eikdn e−ikdn p̂n
 

The overdetermined system is solved for ~x using matrix inversion. But because A is not a
square matrix, an inverse A−1 of the matrix in the general sense cannot be created. The problem
is solved by using a so-called Moor-Penrose pseudo-inverse A+ of A, which is determined using
a least-squares method (eq. 2.49).

~x = A+~b (2.49)
Sensor positioning has an influence on the sensitivity to measurement errors, and is quan-
tified in the Singularity Factor (SF). The lower the SF value, the lower the estimation error.
Seung-Ho et al showed that the SF is minimal when an equidistant spacing between micro-
phones is used [18]. In this case of equidistant sensor spacing, the optimal ratio of the wave
number corresponding to freqeuncy to be measured, over the critical wave number kcr , is 0.5.
Using this formula, and the known design frequency, the optimal microphone spacing can be
calculated using the formula given in equation 2.50.

kcr · ∆xmic = π · (1 − M2 ) (2.50)

A more detailed description about the multi-microphone method, Moor-Penrose inverse,
attenuation effects, and sensor positioning sensitivity are added to the appendix.

29
Chapter 3

Method

The TD-IBC must be tested against a well-defined numerical problem. This numerical prob-
lem consists of a discretized domain, boundary conditions, initial conditions, fluid models, and
other properties such as spatial and temporal integration schemes. Time discretization, space
discretization, and boundary conditions are discussed in section 3.1. Fluid models and initial
conditions are discussed in section 3.2. The actual implementation of the TD-IBC in ANSYS
CFX using FORTRAN subroutines is discussed in section 3.3. The manner in which the reflec-
tion coefficient R is determined is discussed in section 3.4. Lastly an overview and reasoning
behind the test cases is discussed in section 3.5, of which the results are discussed in the next
chapter, chapter 4.

3.1 Domain model

To eliminate problems not inherent to the implemented TD-IBC, the numerical problem should
be defined as simple as possible, but is capable of capturing essential characteristics. To capture
the plane waves correctly, and exclude any turbulent, shear, and boundary layer contributions,
the domain is modeled one-dimensional. Taking into account the fact that the reflection coef-
ficient must be determined in post-processing, and an optimal configuration would be to use
four microphones per wavelength in an equidistant array (see sec. 2.7.1), the domain should be
at least one wavelength long. The domain size thus depends on the expected frequencies in the
domain, which is predominantly the excitation frequency at the inlet of fin = 100Hz, resulting
in an expected wavelength of λ100 = 3.43m (see sec. 3.2). A domain size L = 4m seems a
good compromise between accuracy of measurement and domain size (and thus computational
effort). The domain size with respect to wavelength is visualized in Figure 3.1.

Figure 3.1: Wavelength of excited wave at fin = 100Hz w.r.t. domain size L = 4m

The applied boundary conditions are discussed in section 3.1.1, and the space and time
discretization is discussed in section 3.1.2.

30
3.1.1 Boundary conditions
An overview of the applied boundary conditions is given in Figure 3.2. At the inlet a sinu-
soidally variying velocity of fin = 100Hz is specified. The choice for the excitation frequency is
somewhat arbitrary, as spatial and temporal discretization can easily be adjusted to account for
different frequencies. However, an excitation freqeuncy of 100Hz is also often used in literature
by Oosterhuis and Bühler, which aids in the comparison of results [14][3]. The other bound-
aries at the sides of the domain are symmetry boundaries, used to emulate a one-dimensional
problem. The solver in ANSYS CFX cannot be set to use a one-dimensional set of equations,
regardless of domain shape and settings, ANSYS CFX always uses a three-dimensional set of
equations to solve the problem at hand. The boundary on the right side of the domain is the
pressure specified TD-IBC boundary condition.

Figure 3.2: Applied boundary conditions in domain model

The sinusoidally varying velocity is specified according to equation 3.1, using Command
Expression Language (CEL) in ANSYS CFX.1

u = u0 + ua · sin(2πf t) (3.1)

3.1.2 Space and time discretization

The influence of discretization on the solution of thermoacoustic problems in CFD has been
tested by Bühler [3], and is used as a guideline in discretizing the domain. Results showed
that 100 elements per wavelength and 100 timesteps per period leads to an error, compared
to the analytical solution, smaller than 0.5%. These values are taken as minimum values in
determining the element size ∆x and timestep ∆t.

To determine the exact values of ∆x and ∆t used in simulations, the Courant-Friedrichs-

Lewy number, or simply the CFL number, is used. To give an intuitive explanation about what
the CFL number actually represents, it can be seen as the ratio between the speed in which
information innate to the physical problem travels, over the speed at which information can
travel over the numerical grid (see eq. 3.2). If physical information travels faster than the grid
can handle, stability and convergence problems can arise. This limit is given by the maximum
CFL number CF Lmax , and depends on the numerical schemes used to solve the governing
equations.
For example, if a compact stencil2 is used to solve the Navier-Stokes equations (or derived
fluid dynamics equations), and an explicit first order time marching scheme is used, informa-
tion is only taken from neighboring points to calculate the next value at the point in question,
meaning that information can only travel one element size ∆x over one timestep ∆t, resulting
in a maximum CFL number CF Lmax = 1. Implicit methods are usually unconditionally stable,
which means that the schemes used (and thus the CFL number) are not a factor in stability.
1
See section 3.3.1 for more information on CEL.
2
A compact stencil is a stencil that includes only directly neighboring nodes

31
In accuracy however, it is a factor. The solver used by ANSYS CFX is implicit, however, the
implemented boundary conditions later on are explicit (see sec. 3.3). To rule out any conver-
gence problems due to the CFL number being to high, a maximum CFL number CF Lmax = 1
is chosen.

uphys · ∆t
CF L = < CF Lmax (3.2)
∆x
Next, the maximum physical speed at which information travels in the governing problem
needs to be determined. Mean flow velocity u0 and the acoustic velocity amplitude δ û will be
relatively small compared to the sound speed c0 = 346.13m/s in the test cases defined in section
3.5 (see sec. 3.2). The sound speed c0 is therefore the limiting physical velocity uphys . Taking
an acoustic CFL number CF Lacoustic ≤ 1, and keeping the minimum criteria for the element
size and timestep discussed at the beginning of this section in mind, a conforming element size
∆x = 4 · 10−3 m and a timestep ∆t = 1 · 10−5 s is chosen. The minimum and maximum design
criteria leave room open to adjust the element size and timestep, therefore rounded numbers are
chosen to minimize user error in calculations and derivations, and maximize readability. The
resulting discretized domain is visualized in Figure 3.3.

Figure 3.3: Discretization of the domain

For the spatial discretization of the domain, perfectly cubical elements are chosen.3 All
elements of size ∆x = 4 · 10−3 m in each direction fit exactly 1000 times within the domain
of L = 4m, resulting in perfect mesh quality for different ICEM mesh quality checks, such as
determinant, volume, angle, and warpage [1]. ICEM software is the software used to create the
mesh. A timestep of ∆t = 1 · 10−5 s results in 1000 timesteps per wavelength for a frequency
of fin = 100Hz. The element size ∆x = 4 · 10−3 m results in 857.5 elements per wavelength
λ100 = 3.43m. Both well above the set minimum of 100 elements per wavelength set out at the
beginning of this section.

3
A cube is a form of a hexahedron element: a regular hexahedron with all faces square.

32
 The PWM method as well as the TD-IBC need a characteristic wave amplitude f to be de-
termined at a sample plane near the boundary (see sec. 2.5 and 2.6). The sample plane should
be far enough from the boundary to prevent feedback caused by the used schemes (stencil), but
close enough so that the f determined at the sample plane is very close to the actual f when it
hits the boundary (especially amplitude attenuation due to viscous losses). Viscous losses will
be neglected in the heat transfer equations because its influence can hardly be measured (see
sec. 4.1.1). The sample plane is placed at a distance xpwm = 0.4m, or 100 elements, from the
boundary, which should be out of reach for any stencil. See Figure 3.4 for placement of the
boundary within the domain.

Figure 3.4: Sample plane position within the domain

Preliminary tests showed that increasing time increment ∆t, until less than 100 timesteps
per period are attained, indeed caused significantly worse results, and even stability issues (i.e.
unusable results), as is predicted by Bühler [3]. The hypothesis is that increasing the mesh
linearly near the boundary will result in a better reflection coefficient R, as the gradients for
pressure and velocity near the boundary might be determined more accurately by ANSYS CFX,
i.e. the discretization error near the boundary decreases. However, a smaller ∆x also increases
the CFL number near the boundary, and causes the hexahedron elements to be non-cubical,
which might decrease accuracy and cause stability problems as mentioned earlier in this section.
See test-cases in section 3.5.4 for the test-cases including linear refinement.

The mesh is linearly refined from a distance of xref ine = 0.4m towards the boundary (see
Figure 3.5). Instead of 100 elements within this section, now 300 elements are used, totaling
1200 elements for the complete domain. The last element against the boundary has a edge ratio
of 1/5, meaning that the x-length of the element is ∆xmin = 0.0008m.

Figure 3.5: Mesh refined near boundary

33
3.2 Fluid models
The fluid used as a medium in all test cases is air, and its properties vary according to the ideal
gas law. No turbulence model is used in the simulations, as it is not a fundamental part in
describing acoustic waves, making the flow problem purely laminar.

3.2.1 Initial conditions

The domain is initialized using air at standard pressure and temperature (STP). Unfortunately,
STP definitions vary, and to prevent ambiguity, the values of the properties used in initialization
are mentioned in Table 3.1.

Table 3.1: Fluid properties Air @ STP

p0 = 101325.0P a ρ0 = 1.1839kg/m3 c0 = 346.13m/s T0 = 298.15K
γ = 1.4 Z0 = 409.78Rayl Cp = 1007J/kg · K R = 287.058J/kg · K

The initial velocity is equal to the inlet velocity at initialization (eq. 3.1) to prevent dis-
continuities at initialization. The inlet velocity at initialization (t = 0s) is always equal to the
mean velocity u0 due to a sine function being 0 when the argument is zero. The mean velocity
is set per test case, which are defined in section 3.5.

3.2.2 Heat energy equations

There is a distinct difference in isothermal and adiabatic sound speeds. In the isothermal case,
heat exchange between the wave and the environment need to take place at moments of rarefac-
tion and compression to keep the temperature constant, which is only accurate when the period
of the wave is of the same order or longer than the time it takes to cool the medium. Acoustic
waves on earth generally have periods much shorter than it takes to cool the medium, which
causes the fluid parcels to compress and expand adiabatically by approximation. In practice this
means that the speed of sound is higher in the adiabatic case compared to the isothermal case.
Mathematically this results in an adiabatic index γ = 1.4 in the adiabatic case, and γ = 1.0 in
the isothermal case. The sound speed can be calculated using equation 3.3 (see appendix for
derivation).

r
γp0
c0 = (3.3)
ρ0
In ANSYS CFX calculations can be done both with a heat transfer equation and without
(isothermal). Using the heat transfer equation, the acoustic waves will behave adiabatically.

Depending on the heat model that is used, there are consequences for the boundary specifica-
tion of the temperature. For isothermal calculations the temperature does not vary throughout
the domain, and need not at the boundary (i.e. constant temperature boundaries). It is a
different story when a heat transfer equation is used. The acoustic waves vary in temperature
cyclically through each period. This temperature variation should also apply to the boundary
condition to prevent reflections. The temperature can be varied on the boundary by using the
adiabatic equation for temperature variation given by equation 3.4, derived from the ideal gas
law (eq. 2.4).

34
   γ−1
P2 γ
T2 = T1 (3.4)
P1
Preliminary tests showed that varying temperature using this adiabatic temperature vari-
ation equation (eq. 3.4) resulted in large pressure drift and large reflection coefficients. An
explanation could be that even though the acoustic wave behaved adiabatically, in reality al-
ways some losses and heat transfer to neighboring fluid parcel occurs. Another explanation
could be that something goes wrong in the coupling between boundary values and internally
calculated temperatures. Fortunately, the option exists in ANSYS CFX to specify an opening
pressure at initialization, and let the solver do the calculations using conservation laws. The
results improved significantly, but does still have some slight problems with pressure drift com-
pared to isothermal calculations (see sec. 4).

Two types of heat transfer equations can be used in ANSYS CFX to take into account the
varying temperature, ”Thermal Energy” and ”Total Energy”. The thermal energy equations
only convects enthalpy h with the flow, while the total energy equations also adds kinetic energy
0.5 · u2 (needed for example to take into account stagnation temperatures). The total energy
equation also automatically includes the pressure transient term in the set of equations, while
for the thermal energy equation it is an option. The total energy heat equation is generally
used for high speed flow problems (M > 0.3) [23].

Both heat transfer equations have the options of including viscous work terms in ANSYS
CFX. This option adds the viscous work term to the energy equation, necessary when large
viscous shear forces are to be expected. Viscosity is expected to have a very small influence
because the domain is small compared to the wavelength and sound speed, and the flow field
is laminar (no turbulent viscosity losses). Viscosity causes a decay in amplitude of the acoustic
wave over distance traveled. It can also influence the coupling between acoustic pressure δp
and acoustic velocity δu, which is assumed in phase in many equations (e.g. in the linearized
wave equation, eq. 2.7). The difference is tested in a pragmatic manner by running the same
simulation with the viscous work terms both included and excluded, and comparing the solution
of the flow field via the reflection coefficient R (see sec. 3.5).

The thermal energy equation also has explicitly the option to include a pressure transient
term for describing the influence of the time-varying behavior of the pressure on the tempera-
ture. In the total energy equation it is automatically included. This pressure transient term is
essential for acoustic waves which behave adiabatically (and has a varying temperature). The
influence of this term is also tested (see sec. 3.5).

The theoretical amplitude of the resulting pressure wave p̂ throughout the domain is different
for the isothermal and the thermal energy case because the sound speed differs. The sound speed
c0 can be calculated using equation 3.3, and results in 346.21m/s and 295.86m/s for thermal
energy and isothermal calculations respectively. The pressure will, as a linear approximation,
vary in space as a sine according to equation 3.5.
x
δp = p̂ · sin(2π ) (3.5)
λ
To verify the calculations in ANSYS CFX, it is also important to check if the expected
amplitudes of the pressure and velocity gradients match the theoretical values. As the gradients
are of crucial importance in the LODI-relations (eq. 2.19). By differentiating eq. 3.5 in x using
the chain rule, the resulting expressing for the pressure and velocity gradient can be derived
(eq. 3.6 and eq. 3.7).

35
 x 2π
δpx = p̂x · sin(2π ) where p̂x = p̂ (3.6)
λ λ

x 2π
δux = ûx · sin(2π ) where ûx = û (3.7)
λ λ
The properties that pertains to the difference in isothermal and heat transfer equation sim-
ulations are summed up in Table 3.2. The wavelength λ is calculated using the well-known
formula that relates wavelength, frequency, and sound speed in linear media: c = f · λ.

Table 3.2: Wave amplitude expressions at the boundaries

Isothermal Heat transfer Unit
fin 100 [Hz]
γ 1.0 1.4 [−]
c0 295.86 346.13 [m/s]
λ 2.9586 3.4613 [m]
ua 0.5 [m/s]
p̂ 175.1 204.9 [P a]
p̂x 371.86 371.86 [ a/m]
P
ûx 1.062 0.91 [m/s/m]

3.3 Implementation of a TD-IBC in ANSYS CFX

In the first sections some general topics concerning programming in FORTRAN for ANSYS
CFX applied to the non-reflective part (including LRM), the optional PWM part, and the TD-
IBC part, is discussed. In section 3.3.7, details specific to the non-reflective part are treated.
In section 3.3.8, the same is done for the PWM part, and in section 3.3.9 the TD-IBC imple-
mentation is discussed.

3.3.1 Fortran programming in ANSYS CFX

There are two methods to create custom behavior in ANSYS CFX. The first being a script
language called Common Expression Language (CEL). CEL can be used to construct relatively
simple mathematical statements which can directly be evaluated, and used as variables, but
also certain field quantities and other variables can be retrieved. For example, a vibrating inlet
velocity can be specified as a function of time t.

For more complex applications involving solutions to be stored in memory, gradients or

time derivatives to be retrieved/calculated from certain locales, output to screen, and complex
arithmatic, CEL is not suitable. For these more demanding applications FORTRAN subroutines
are used. Subroutines work comparable to a normal function, with arguments that can be
supplied (both user defined, and fixed solver arguments), and return values that are returned.4
The custom built FORTRAN subroutines are programmed and compiled before the solver runs,
and the solver calls them during execution. Compiled subroutines are platform specific, which
means that a subroutine compiled for a Linux environment will only run in a Linux environment
[23]. Two types of FORTRAN subroutines can be distinguished in ANSYS CFX :
4
See ANSYS CFX documentation for more information on FORTRAN subroutines [23].

36
 1. User (Defined) CEL function: Function that can be called using standard CEL rules.
Solver determines when and how often the subroutine is called.

2. Junction Box Routines. Called at certain points during solver execution.

Both type of subroutines are used in the developed boundary conditions. A total of three
subroutines are used to construct both boundary conditions (NRBC and TD-IBC), two of which
are junction box routines, and one a user CEL routine. The user CEL subroutine is used to
return the pressure, and is automatically called by the solver during its solution process. The
two junction box subroutines are called only at certain stages within the solution process, of
which both positions are given in Figure 3.6. The User Input junction box is only called once,
at the start of the simulation, while the End of Timestep junction box is called at the end of
every timestep when the solution is converged.
The PLinit subroutine at the junction box location User Input creates data areas and
directories, retrieves user parameters, and checks if some important prerequisites for the code
to work are met. The user CEL function Pset sets the pressure at the current locale. To
calculate the returned pressure (to the solver), the subroutine retrieves gradient values for
pressure and velocity, historical information from memory, and does time integration various
quantities. The junction box routine PTloop at the junction box location End of Timestep
calculates the pressure to be set the next timestep and reorders memory, as well as retrieving
area averaged quantities to calculate characteristic wave f .

Figure 3.6: Junction box and user CEL locations of subroutines used.

Note that the graph in Figure 3.6 is not a depiction of all of the available junction box
routines in ANSYS CFX, which far exceeds the shown, but merely shows junction box locations
relevant to the programmed code. See ANSYS CFX docomumentation for all junction box
locations [23].

37
3.3.2 Parallelization
Often multiple processors are used at once to solve a numerical problem that places stringent
demands on computational power within a reasonable time-frame. If the developed boundary
conditions are to be used in such a setting in the future, it is important that they are compat-
ible with running on multiple processors. This is however not as straightforward as it might
seem, especially memory allocation might create problems when constructing the subroutines
in FORTRAN.
ANSYS CFX runs on the Single-Program-Multiple-Data (SPMD) model for distributing
memory. Before the solver starts with the solution routine, the computational domain is split
into many regions. Each region, called locale, is assigned to a different (partition) process
and is handled by a certain processing core. One process is called master, and the rest of the
processes are called slaves. This distinction between master and slave is important in memory
management, because the user-data memory is not shared among processes (see sec. 3.3.4).
Regardless of the type of partitioning that is chosen in ANSYS CFX, both Message Passing
Interface (CHameleon) (MPI(CH)) and Parallel Virtual Machine (PVM) do not result in a
shared memory model for user-data.5

Figure 3.7: Difference “every partition” and “master only” memory handling

Based on the junction box location, or user CEL routine, at which the subroutine is called,
the solver handles memory management differently. Distinction is made between “Every Par-
tition” and “Master only” memory management, which is visualized in Figure 3.7. “Every
partition” means that subroutine called from a partition has its own memory assigned space
throughout the simulation. From this subroutine variables set on other partitions cannot be
read or changed, and vice versa. “Master only” means that when the subroutine is called, the
user-data memory area of the master process overwrites the user-data memory of slave processes.
Any change in the user-data memory will only hold until such a “master only” subroutine is
called. As can be seen from Figure 3.6, the PLinit subroutine at junction box location User
Input is of type “Master only”. The PLinit subroutine is thus only called once from the master
process in a parallel environment, and written data in memory is automatically copied to the
slave processes at the end of the routine. the PTloop subroutine at junction box location End
of Timestep is of type “Every partition”, and is called on every partition once every timestep.
5
The difference between parallel models are discussed in theANSYS CFX solver modeling guide [23]

38
The user CEL function is always of “Every partition” type. However, this type of memory man-
agement does not need to be a problem, but should be taken into account when programming.
The subroutine PTloop called at junction box location End of Timestep determines for each
partition if any data is stored in its user-data area, and does the calculation for this partition.

3.3.3 Solver data structure

ANSYS CFX structures data and calls User CEL subroutines in a way that is not fully doc-
umented by the ANSYS CFX documentation [23]. ANSYS CFX also distinguishes internal
solver names from user names. User names often have descriptive names such as “Outlet” for
a boundary condition, internal solver names do not. User names can be specified in ANSYS
CFX PRE, whereas internal solver names are fully controlled by the solver. For example, the
user name for a boundary condition could be “Outlet”, then the corresponding solver name
is “BCPn”, which is an abbreviation of Boundary Condition Patch, and where n is an integer
referring to a boundary condition number. An overview of the different types is given in Table
3.3, where ’-’ under heading User signifies that is only used by the solver and cannot be set in
any way in ANSYS CFX PRE.

Table 3.3: Ansys CFX model structure, n:identifying (natural) number

User-name Name Solver-name Format
Domain “BODY” Zone ZNn
Phase “Fluid 1” Phase FLn
Boundary Condition “Outlet” Boundary Condition Patch BCPn
- - Interior Element Group IELGn
- - Boundary Element Group BELGn
- - Integration Point IPn
- - Boundary Integration Point BIPn
- - Element Centre CENTREn
- - Boundary Centre BCENTREn

A boundary condition patch is subdivided multiple in boundary element groups, which com-
prises of element faces that coincide with the boundary. These element faces each have one face
centre (BCENTRE), and, depending on the type of element, multiple boundary integration
points (BIP). In the domain model (see sec. 3.1) cubical elements are used with quadrilateral
faces. The quadrilateral faces that coincide with the boundary have one BCENTRE and four
BIPs, of which the placement is shown in Figure 3.8.

39
Figure 3.8: Integration points and Element center, left: quadrilateral element, right: triangular
element

The Pset routine is only called from these BIP and BCENTRE entities from within a bound-
ary element group (BELGn), other entities are not used directly in the developed subroutines.
The velocity and pressure gradient are needed to construct the amplitude variation L5 based on
the LODI relations (see sec. 2.3.2). The built-in ANSYS CFX boundary condition uses BIPs
for retrieval of the pressure gradient, and BCENTREs for retrieval of the velocity gradient, for
reasons that are not known publicly [19].

3.3.4 Memory precision

Initial tests with the boundary conditions showed a large amount of pressure drift. An example
of the mentioned pressure drift is given by Figure 3.9. Before the excited wave reaches the
boundary after 0.012s, already 40P a pressure drift is observed. After two periods this has
accumulated to over 100P a pressure drift. Loss in precision due to multiplication of single
precision numbers several orders of magnitude apart seemed to be the cause. The solution is
to use double precision numbers. Both the solver needs to be set to double precision, as well as
the FORTRAN subroutines needs to be coded and compiled as such.

Figure 3.9: Example of pressure drift (K = 101/s, ∆t = 1 · 10−5 )

40
 Depending on the value of the decimal number to be stored, single precision numbers give 6
to 9 significant decimal digits precision, while double precision numbers give 15 to 17 significant
decimal digits precision. When multiplying or adding numbers orders of magnitude apart, the
error in the resulting number can be substantial. This is often the case when multiplying a
time derivative with a very small time increment to retrieve the value at the next time step. Or
when adding a small pressure increment onto the current pressure. These type of mathematical
operations occur in the FORTRAN subroutines written. This results in an error that adds up
every timestep, and can cause the pressure to drift significantly, or even become unstable.

More details about single and double precision numbers are added to the appendix.

3.3.5 Memory Management System

To make things easier, ANSYS created a memory structure that can be accessed using a “di-
rectory” type of hierarchical structure called Memory Management System (MMS). This keeps
data ordered and easy to manipulate. However, on a low level all data of a certain type (e.g.
integer, real, etc) is placed on a stack and managed by the memory management of the host
system. The global stacks on which the pressure information is stored can be read and written
to using two ways, the first by calling a so-called peek or poke routine with the MMS path as
argument, the second by directly accessing the stacks. A peek routine retrieves values from the
stack, while a poke routine stores values on the stack. However, for double precision floating
point numbers the peek and poke routines are not defined, and not in any of the includes of
ANSYS CFX. The stacks need to be directly addressed to access and store values.

To alter data on the stacks, first a pointer to the data area in question needs to be retrieved.
A pointer is an integer data type memory address to the first value of the data structure it points
to. By using a MMS function called LOCDAT, and the MMS path as argument, a pointer can
be retrieved. Multi dimensional arrays are stored as a one-dimensional vector in memory, with
all memory positions reserved in one consecutive data area. For a 2D matrix A of size V × W ,
each cell A(I, J) can be accessed on stack DZ using pointer pPOINT (see eq. 3.8). This is
done in a row for row fashion. For example, for a 4 × 3 matrix A cell A(3, 2) can be accessed
on the stack via eq. 3.9. Pointers are, by convention, written starting with a small p.

A(I, J) = DZ(pP OIN T + W ∗ (I − 1) + (J − 1)) (3.8)

A(3, 2) = DZ(pP OIN T + 3 ∗ (3 − 1) + (2 − 1)) (3.9)

The complete directory structure used to store variables and arrays is shown in Figure 3.10.
The USER data area is used to store user parameters and values returned by various utility
routines defined by ANSYS CFX. User parameters are variables that are read in from a text-
file at initialization, and stored in memory. For the developed subroutines the user parameters
are used to import values such as the relaxation factor K, and the target pressure at infinity
p∞ . Utility routines are built-in routines from ANSYS CFX, for example to get area integrated
quantities of field variables. The USER DATA data area is used to store all variables that are
specific to the programmed code.

41
Figure 3.10: Memory structure used in subroutines

42
3.3.6 Initialization routine
The PLinit routine, of which the flowchart is shown in Figure 3.11, is equal for both the non-
relflective (with or without PWM) and the TD-IBC implementation, except for a few added
user parameters in the case of the TD-IBC. In the first section of the code the user parameters
are retrieved and stored in the MMS. Also other important variables and directories are created.
The second section retrieves global mesh information, such as the number and names of zones in
the problem definition. These are needed for certain utility routines in the other two developed
subroutines PTloop and Pset. If there is more than one zone, the exit-sequence is called. More
than one zone means that multiple domains are defined, which is not supported. The exit-
sequence stops the code, and signals to the solver that it needs to stop the simulation. In the
zonal mesh info section the number and names of BCPs are retrieved, and checked for the “LRM”
prefix needed to mark which boundary is designated to function as a non-reflective or TD-IBC
boundary. If no boundary is found, the exit-sequence is called. The LRM BCP mesh info
section checks which physical type (“OUTLET”, “WALL”, etc) the designated LRM-boundary
has. Only “INLET”, “OUTLET”, and “OPENING” are supported, or else the exit-sequence is
called. The next section is the Zonal Phase Info section that checks if only one phase is present,
as multi-phase flow is not supported in the code. If multi-phase flow is present, the exit-sequence
is called. The last section stores information retrieved regarding the defined problem, such as
domain names, BCPs, and Phase names.

43
44
Figure 3.11: Simplified PLinit subroutine flowchart

45
3.3.7 Non-reflective part
The non-reflective part uses the LODI method to construct amplitude variation L5 , which is
distinct from the TD-IBC part that uses the characteristic wave f retrieved from a sample
plane near the boundary (see sec. 2.3.2). The PTloop and Pset subroutines are more distinct
for both the non-reflective and TD-IBC boundary conditions. The flowchart representing the
PTloop routine is given by Figure 3.12. The orange part is the TD-IBC specific part, the rest is
equal for both boundary conditions. The flowchart in Figure 3.13 is the Pset subroutine for the
non-reflective boundary condition. The LRM and PWM methods are both implemented in the
non-reflective boundary condition, but can be used or neglected based on the user parameters
given. By specifying a relaxation factor K = 01/s, the LRM method is effectively not used. The
inclusion of the PWM term is controlled by the PWM flag in the user parameters, which can
be set to 0 (off) or 1 (on).

3.3.7.0.1 PTloop subroutine description The PTloop subroutine described by the flowchart
in Figure 3.12 is called at the end of each timestep on every partition. First, transient data
such as the timestep and time increment ∆t are retrieved and stored, then it is checked if the
subroutine is in its initial phase (meaning first timestep in the simulation). If the subroutine
is in initial phase, an array that holds data concerning the characteristic wave f sampled on
the sample plan is constructed. How this array is constructed is discussed in section 3.3.8.2.
Then the ISINIT boolean flag is set to false, so the arrays in question are not constructed again
within this partition. As data is stored in memory according to the SPMD model (each partition
isolated), it is checked (using LCALS ) if the current partition is tied to the BCP in question
(and thus holds any relevant data). If it is not tied to the LRM BCP, the routine exists and the
solver continues. Next a memory optimization routine is performed, of which the details are
explained in section 3.3.7.2. In the “Calculate new FAMP” section, the characteristic wave f is
calculated using area average pressure and velocity at a sample plane near the boundary. (see
sec. 3.3.8.1 for more details). The last section is TD-IBC specific, and is explained in section
3.3.9.

3.3.7.0.2 Pset subroutine description The non-reflective version of subroutine Pset is

more distinct from the TD-IBC version than for the PTloop subroutine, because of the different
manner in which amplitude variation L5 is determined. The flowchart of the non-reflective
version of the Pset subroutine is shown in Figure 3.13. First locale and entity information is
gathered (see sec. 3.3.3), and certain other important quantities such as the relaxation factor
K and the time increment ∆t are retrieved. If the solver is in its initial phase, arrays contain-
ing historical information about the gradients and pressure are constructed, and given initial
values. If the subroutine is not in initial phase, then the gradients are retrieved from utility
routine user getvar, the amplitude variations L1 and L5 calculated using the LODI relations
(see sec. 2.3.2 and 2.4), and the new pressure for the current timestep is determined. Lastly
this new pressure is returned to the solver to be set at the boundary.

46
.

47
Figure 3.12: Simplified PTloop subroutine flowchart

48
Figure 3.13: Simplified Pset subroutine flowchart (NRBC version)

49
3.3.7.1 Time integration

As only temporal derivatives of the pressure are given by the LODI relations (see eq. 3.10),
the pressure derivative must first be integrated using a numerical time integration scheme to
determine a pressure value to be set at the next timestep.

∂p 1
=L where L = − (L5 + L1 ) (3.10)
∂t 2
Z t+∆t
n+1 n
p =p + L dt (3.11)
t

Different time integration schemes can be used to approximate the integral in (3.11), the
simplest being explicit forward (eq. 3.12) and implicit backward Euler (eq. 3.13).

pn+1 = pn + ∆t · L(n) (3.12)

pn+1 = pn + ∆t · L(n + 1) (3.13)

The Euler method is only first order in time, and the accuracy and stability might bene-
fit from higher order schemes. Well-known integration schemes such as Runge-Kutta will not
work, because one can use a Euler-Forward predictor to predict p(n + 12 ), but is not able to
evaluate L at future timesteps, because L does not depend on t and p explicitly. However, for
more accuracy (lower truncation error), one can use higher order backward (finite) differences
to approximate the time derivative, and a predictor of L at n + 21 .

4 1 2 1
pn+1 = pn − pn−1 + ∆t · L∗ (n + ) + O(∆t2 ) (3.14)
3 3 3 2
18 n 9 2 6 1
pn+1 = p − pn−1 + pn−2 + ∆t · L∗ (n + ) + O(∆t3 ) (3.15)
11 11 11 11 2
The time integration schemes given by equation 3.14 and 3.15 are second order and third
order respectively. These higher order integration schemes also require more historical informa-
tion, and thus higher memory requirements. The L∗ (n + 21 ) term is a predictor for the LODI
function at n + 21 , created by using either a (explicit) Lagrange polynomial extrapolation (see
eq. 3.16), or by using an implicit predictor by taking the average left and right of n + 12 (see
eq. 3.17).

1 3 1
L∗ (n + ) = L(n) − L(n − 1) (3.16)
2 2 2

1 1
L∗ (n + ) = (L(n) + L(n + 1)) (3.17)
2 2
Using the implicit predictor for L∗ (n + 12 ) (see eq. 3.17) results in a Crank-Nicolson scheme.
The caveat is how to evaluate L at n + 1, which is not known yet at the beginning of the
timestep. By using a predictor for L(n + 1) at the beginning of the timestep and first coefficient
loop iteration (for example, constant extrapolation: L(n) = L(n + 1)), the loop can be started
and at the second loop the value for L(n + 1) can be calculated on the boundary. This is a
semi-implicit scheme which uses the latest information available.

50
3.3.7.2 Memory optimization
As can be seen in section 3.3.7.1, more historical information is needed at the boundary for
higher order time integration. For a third order approximation of the pressure derivative, pres-
sure data is needed at n, n − 1, and n − 2, and for function L it is L(n) and L(n − 1). This will
most probably causes no big increase in memory requirements, because the number of nodes at
the boundary increases by an order of 2, while the number of nodes in the interior increases by
an order of 3. And when the domain is very small, memory requirements are usually not an
issue. Not only memory requirements increases, but also memory operations if the data is just
shifted over each timestep loop. The number of memory operations can be reduced by using a
modulo operation (see Fig. 3.14).

Figure 3.14: Optimized memory storage

At each timestep a new pressure needs to be calculated, but the pressure at n − 1 in memory
is not n − 1 anymore (as was the previous timestep), so the value must be moved to n − 2,
and n − 2 to n − 3, and so on. For a third order scheme three memory operations must be
performed per node to store the new value. Each arrow in the Figure (3.14) represents an
operation. By using a new way of managing the memory locations, this can be reduced to just
one. By taking the modulo of 3 (three locations) for timestep n, the correct memory location (0,
1, or 2) for the storage of the new pressure can be determined at the beginning of each timestep.

3.3.8 Plane-wave masking

The plane-wave masking method is used to mask the characteristic wave f from the LRM term
(eq. 2.25), in this way deviations from the target pressure p∞ resulting from acoustic contribu-
tions are not corrected, and thus do not cause reflective behavior (see sec. 2.5). To construct
characteristic wave f , the area averaged pressure and velocity must be retrieved at a sample
plane near the non-reflective boundary. How these quantities are retrieved is explained in section
3.3.8.1. Next, the calculated incoming characteristic wave f (see eq. 2.31) must be stored for at
least the time delay until the wave hits the boundary, which is elaborated upon in section 3.3.8.2.

3.3.8.1 Retrieval of pressure and velocity from sample plane

To evaluate the characteristic wave f , the area averaged pressure hpi and velocity hui on a
sample plane near the boundary are needed. The utility routine user get gvar can retrieve area
averaged field quantities on any user defined location. User locations can include 2D regions

51
(e.g. boundaries) and 3D regions (e.g. domains). However, one cannot just create an arbitrary
2D region within the mesh. The mesh needs a face to work on, and this cannot be an internal
element face. The approach is to create two blocks in a mesh editor (e.g. ICEM), and create
a named part from the interface between them (see Fig. 3.15). This interface (shown in red)
is used as the sample plane. In ANSYS CFX a 1:1 general connection interface needs to be
constructed, which automatically creates two boundaries. These are now two locations (on
exactly the same location) from which the area averaged field quantities can be retrieved. Also
the mean density and sound speed are needed, but these remain constant during the simulation
and are evaluated at initialization.

Figure 3.15: Plane-wave masking sample plane construction using blocks in ICEM

3.3.8.2 Storage of integrated quantities

The evaluated characteristic wave f at the sample plane must be stored in memory until the
wave reaches the boundary, where it is used in the specification of L1 (see eq. 2.25). The sample
plane is positioned at a distance xpwm from the boundary (see Fig. 3.15). The corresponding
time delay tpwmdelay is calculated using eq. 3.18.

xpwm
tpwmdelay = (3.18)
c0
The characteristic wave f can only be retrieved at discrete time intervals, and is stored
in a vector of length nf , where nf is calculated by dividing the time delay tpwmdelay by the
time increment ∆t, and adding 1 (see eq. 3.19). If one would choose a fixed position in the
vector for the newest and oldest value (the oldest being the one that should be retrieved to
calculate L1 ), all data must shift one position every timestep, resulting in nf read and write
operations. By using a modulo operation the index of the newest and oldest data shifts one
cell (see eq. 3.20). This results in just one read and write operation per timestep (see Fig. 3.16).

tpwmdelay

nf = lower +1 (3.19)
∆t

Q1 = M OD(tstep , nf ) + 1 and Q2 = M OD(tstep + 1, nf ) + 1 (3.20)

Figure 3.16: Plane-Wave Masking memory storage

52
3.3.9 Time-domain impedance boundary condition part
The TD-IBC part is constructed by adding the temporal derivative of an external perturbation
g ∗ to the amplitude variation L5 , which is determined through filter-methods and historical in-
formation about f and g ∗ (see sec. 2.6). How the filter coefficients are determined is discussed
in section 3.3.9.1. Next, an alternate formulation for L1 and L5 is discussed in section 3.3.9.2,
and in section 3.3.9.3) the storage of quantities and the calculation of the first time derivative
of g ∗ and f is discussed. In the last section (sec. 3.3.9.4) the TD-IBC specific user CEL routine
Pset is explained.

3.3.9.1 Calculation of filter coefficients

MATLAB includes a function that calculates the filter coefficients (see sec. 2.6.1) based on
the method by E.C. Levy [11]. This function is called invfreqz, and returns the coefficients of
the numerator and denominator of the rational expression for R in equation 2.36. The func-
tion takes the complex reflection R at different normalized frequencies ωn , and the order of
the numerator and denominator as arguments. A script is written that takes the frequency
range f , reflection coefficient R, and the timestep ∆t as input, and calculates the order n and
normalized frequency ωn as input for the invfreqz function (see sec. 3.4). The script is based on
the assumption that the complex reflection coefficient stays constant with the given frequency
range. Based on this, the order can be calculated by means of the largest time delay needed
(see eq. 2.41 and 2.42), which belong to the lowest frequency.

To assess if the filter coefficients will be stable, a pole-zero map is used (see Fig. 3.17 for
an example pole-zero map). Any pole (marked: ×) outside the unit circle in the complex plane
might cause unstable behavior. Preliminary simulations indeed show that filter coefficient sets
with poles outside the unit-circle indeed become unstable. Iterative runs, and slight adjusting
of the reflection coefficient and number of coefficients, might result in better positioning of the
poles, and is discussed in section 3.5.5.

3.3.9.2 Altered LODI formulation

The locally reacting nature within the definition of L5 , as given by equation 2.20 gives rise
to problems in combination with the modified expression (with external perturbation) for L1
∗
as given by equation 2.32. Because the ∂g ∂t term in the modified L1 expression changes the
pressure and velocity gradients such that the incoming f wave cannot be uniquely and correctly
distinguished anymore. The problem can be overcome by giving up some properties of the LODI
relations. The characteristic wave f is no longer determined implicitly by boundary gradients,
but taken from the sample plane and offset in time towards the boundary. By doing this it is
assumed that between the sample plane and the boundary, the characteristic wave f is a pure
1D plane wave, does not exhibit dispersion behavior, has a constant sound speed, and does not
decay in amplitude. The new expression for L5 is given by equation 3.21.

∂f
L5 = − (3.21)
∂t

3.3.9.3 Storage and time differentials

The forward travelling wave f is sampled at a sample plane near the boundary. As plane-wave
masking is also used for the TD-IBC implementation using filter methods, the same time de-
layed value of f can be used. The value of f when it reaches the boundary is retrieved from

53
 Figure 3.17: Pole-Zero map for R = 0.5 + 0.5i, o:zero, x:pole

the plane wave masking array, and stored in the TD-IBC array for f (see Fig. 3.18). Also g ∗
needs to be stored, but is calculated directly at the boundary using equation 2.40. The size of
both arrays for f and g ∗ are equal to the number of filter coefficients n used.

Figure 3.18: TD-IBC Masking memory storage

As can be seen from the expressions for L1 and L5 (see eq. 2.32 and 3.21), the time derivatives
of both f and g are needed. These are approximated using one-sided sixth order finite difference
expressions (see eq. 3.22). Which demands historical information of six timesteps back.

 
∂f 1 49 15 20 15 6 1
= ft − 6ft−∆t + ft−2∆t − ft−3∆t + ft−4∆t − ft−5∆t + ft−6∆t (3.22)
∂t ∆t 20 2 3 4 5 6

54
3.3.9.4 User CEL routine Pset for TD-IBC
The Pset subroutine for the TD-IBC differs from the NRBC version mainly in the manner in
which the amplitude variation L5 is determined (see sec. 3.3.9.2). The flowchart of the TD-IBC
version of the Pset subroutine is given in Figure 3.19. In the purple section of the flowchart
the forward and backward traveling characteristics f and g are retrieved (which were calculated
after each timestep in the PTloop subroutine using area averaged integrated quantities). From
these quantities the temporal derivatives are determined, which are used in the definition of L1
and L5 . The new pressure of the current timestep can be calculated using equation 3.10 and
3.11, which is subsequently returned to the solver.

55
Figure 3.19: Simplified Pset subroutine flowchart (TD-IBC version)

56
3.4 Reflection measurement

The multi-microphone method discussed in section 2.7.1 is programmed in Matlab using object
oriented programming instead of classing programming. This is done to make the code more
general, easier to understand, use, and extend by others. A class called MultiMicMethod is
written that stores all the data, and harbors all functionality. The class is structured in such a
way that other users can easily extend the functionality by adding methods to the class, or use
class inheritance and keep the original class intact.

Figure 3.20: UML diagram of MultiMicMethod Class (incomplete, see appendix for full UML)

The most important properties and methods are shown in the Unified Modeling Language
(UML) diagram in Figure 3.20. This is only a short list of important properties and methods,
the full UML diagram is added to the appendix. The work-flow in short, first an MultiM-
icMethod object needs to be created, and then the pressure time-series can be imported from
ANSYS CFX using the loadPArray function. The frequency spectrum for the pressure time
series for each microphone must be created using the calcFreqSpectrum function before one can
calculate the reflection coefficient R as a function of frequency with the calcReflection function.
With the function reflectionTargetFreq one can obtain the reflection coefficient nearest to the
given target frequency (as argument). Be advised that one has determined the positioning of
the pressure sensors such as to minimize the error in the Multi-Microphone Method for a given
frequency (see sec. 2.7.1). Further away from this “design” frequency the error in determining
the reflection coefficient R will increase. The plotReflection function plots the reflection coeffi-
cient as a function of frequency.

Losses due to turbulence and viscosity can also be taken into account using the MultiM-
icMethod class, but is not used in determining the reflection coefficient R in all the defined test
cases in section 3.5. The inclusion of viscosity in the heat transfer equations (see sec. 3.2.2)
resulted in a negligible difference in the non-viscous reflection coefficient and solution (see sec.
4.1.1), therefore only the non-reflective version is used. More detail about taking into account
losses in the multi-microphone method is added to the appendix.

57
3.4.1 Measurement in test-cases
In the test-cases defined in section 3.4.1 the reflection coefficient is measured using four equidis-
tant microphones per wavelength, as this is seen optimal according to MMM theory (see sec.
2.7.1). The wavelength varies depending on whether the simulation is performed isothermal,
or the heat equation is included (see sec. 3.2.2). The inclusion of a heat transfer equation
is imperative in thermoacoustics as the change in temperature is what drives thermoacoustic
devices. Therefore this wavelength of λ = 3.472m is chosen as a basis to calculate the micro-
phone spacing ∆xmic , which results in ∆xmic = 0.0825m (see Table 3.4 for absolute position
of microphones, coordinate x and d as shown in Figure 2.8). The pressure time-series of the
microphones are exported from ANSYS CFX to comma-separated values (CSV) format. These
timeseries are then used as input for the Matlab class MultiMicMethod which calculated the
reflection coefficient R based on the MMM method.

Table 3.4: Microphones used for determining reflection

Microphone P1 P2 P3 P4
x position 1.196 2.064 2.932 3.8
d position 2.804 1.936 1.068 0.2

Figure 3.21: Microphone placement within domain

58
3.5 Test cases
Several test cases are formulated to test the performance and stability in different situations.
First general solver settings and convergence criteria are discussed in section 3.5.1.Then the rea-
soning behind the simulation time is discussed in section 3.5.2, and the relaxation factor K used
in the LRM term is discussed in section 3.5.3. Distinction is made between the non-reflective
and TD-IBC test cases. Not only the phase and absolute value of the reflection coefficient can
explicitly be set in the TD-IBC boundary, but also the way in which the non-reflective character
is being established is different (see sec. 3.3.9.2). The non-reflective boundary test-cases are
discussed in section 3.5.4, and the TD-IBC test-cases are discussed in section 3.5.5.

3.5.1 Solver settings and convergence criteria

For all simulations the same solver settings and convergence criteria are used. Because the
acoustic CFL number used is on the save side (less than one) and a very basic domain and
models are used, a high resolution advection scheme is used instead of an upwind scheme. This
minimizes numerical dissipation and maximize accuracy. The transient scheme used is sec-
ond order backward Euler with a timestep initiliazation being extrapolated from the previous
timestep. The residual is set to max 1 · 10−6 .

3.5.2 Simulation time

On the basis of the length of the domain and the characteristic sound speed, the wavefront is
4
calculated to arrive at the boundary after 343,2 = 0.011655s. To see a standing wave appear-
ing in the tube the calculation time should be at least twice this time. To be able to exclude
transient effects from influencing the reflection coefficient calculated using the multi-microphone
method, the simulation time should be such that at least a few periods in a periodic steady state
is attained. Preliminary simulations showed that some simulations need at least ten periods to
attain a periodic steady state (due to pressure drift). On the basis of this observation, a total
simulation time of t = 0.2s, and a sampled range from t = 0.015s to 0.2 is chosen, which results
in 15.87 periods that are sampled.

3.5.3 Relaxation factor

The relaxation factor K influences the reflection coefficient R. Theoretically this relation is
given by equation 2.23 in section 2.4. The reflection coefficient R of the boundary conditions
are tested against K, and compared to the theoretical value of R as a function of K. It is ex-
pected that the measured reflection coefficient is near the theoretical reflection coefficient, and
behaves in the same way. The relaxation factor K is varied logarithmically between K = 1s−1
and 105 s−1 . In ANSYS CFX the coupling parameter σ is used instead of K, but can be con-
verted using equation 2.24. The optimal values 0.27 and 0.58, as discussed in section 2.4, have
the corresponding relaxation factors of K = 23.4361/s and K = 50.3441/s respectively. These
values for K are rounded to K = 231/s and K = 501/s, and used across all other test-cases as
defined in section 3.5.4 and 3.5.5. All values for relaxation factor K, and corresponding coupling
parameters σ, are summed up in Table 3.5.

59
 Table 3.5: K values and corresponding σ values for a mean flow u0 = 0
K 100 101 23 50 102 2 · 102 5 · 102 103 104 105
σ 0.0115 0.1152 0.2650 0.5760 1.152 2.304 5.760 11.52 115.2 1152

3.5.4 Non-reflective boundary test cases

The test-cases defined in this section are used to compare three NRBCs: the built-in ANSYS
CFX NRBC, the custom NRBC based on the NSCBC method including the LRM term, and
the custom NRBC including both LRM and PWM terms. Three important parameters are
varied among the test-cases, which are summed up in Table 3.6. Calculation of all possible
combinations would result in 120 test cases per boundary, and a grand total of 360 for all three
boundaries. To decrease the number of test-cases, only certain combinations are tested, which
are elaborated upon in this section. The influence of mean flow u0 on the non-reflective bound-
aries is tested for optimal relaxation factors K = 231/s and K = 501/s only. Main figures of
merit during simulations are the reflection coefficient R and the pressure drift Pdrif t .

Table 3.6: Possible combinations test cases

Heat model Isothermal, Thermal energy, Total energy
Relaxation factor K 100 , 101 , 23, 50, 102 , 2 · 102 , 5 · 102 , 103 , 104 , 105 [1/s]
u0 0, 0.5, 2.5, 5 [1/s]

Before the standard test cases are performed, first the influence of different time integration
schemes and the type of heat transfer equation are analyzed. The time integration schemes (see
sec. 3.3.7.1) are tested using the total energy heat equation without the viscous work term, and
a relaxation factor of K = 501/s, and are summed up in Table 3.7.

Table 3.7: Time integration test cases

Integration order L∗ predictor
Influence of time integration 1st L∗ = Ln
2 nd L∗ = Ln
3rd L∗ = Ln
∗
Influence of L predictor 1 st L∗ = Ln+1 (implicit)
1st L∗ = 12 (Ln + Ln+1 )

The influence of the heat transfer equation, and their options, are tested. The test cases
are summed up in Table 3.8. The heat transfer test cases are only performed for the built-in
ANSYS CFX NRBC and the custom NRBC. A relaxation factor of K = 501/s is used.

Table 3.8: Heat transfer test cases, K = 501/s

Heat transfer equations Pressure transient term Viscous terms
Thermal energy Off Off
Off On
On Off
On On
Total energy N/A Off
N/A On

The best performing time integration scheme is determined to be second order with an

60
explicit Lodi function: L∗ = Ln (see sec. 4.1.2 for results and analysis). The best perform-
ing heat transfer equation is the total energy equation without viscous terms (see sec. 4.1.1 for
results and analysis). These options are used in all other simulations, unless specified otherwise.

3.5.4.1 Smooth sine construction

The wavefront of a sine impinging on the boundary shows a discontinuity in its gradient, which
can be seen in Figure 3.22. This causes an overshoot and wiggly behavior in the numerical
solution of the gradient (see sec. 4.1.7). A hypothesis is that this initiates drift. To test this,
test-cases are constructed with a sine that is smoothed at the front.

Figure 3.22: Normal sine and gradient

The sine is smoothed by multiplying a polynomial f (x) for the first quarter of the wave.
To prevent introducing yet another discontinuity, the polynomial should match certain criteria.
First, the product of the functions should equal zero at the onset of the wave, and be equal to
the sine amplitude at a quarter wave-length (at maximum amplitude). This corresponds with
f (0) = 0 and f (1) = 1 (normalized 1 for 0.5π). Also, to prevent a discontinuity, the gradient of
the product should be zero at both the wave-front and at maximum amplitude, which results
in f 0 (0) = 0 and f 0 (1) = 0. Given these four constraints, the polynomial should be of third
order to have a fully determined problem. After solving the system of equations, the smoothing
function f and its derivative are found, and given by equation 3.23.

f (x) = 3x2 − 2x3 f 0 (x) = 6x − 6x2 (3.23)

The resulting sine and corresponding gradient are given by Figure 3.23. One can observe
a small “kink” in the gradient, the gradient is therefore not completely smooth and causes a
discontinuity in the second spatial derivative. This might cause some inaccuracies, but should
be far less severe than with an normal sine. The second spatial derivatives are not used in the
construction of the boundary conditions, therefore no problems are expected.

The smooth sine is added to definition file in ANSYS CFX PRE using the CEL statement
in equation 3.24.

u = u0 + [(3t2norm − 2t3norm ) · step(1 − tnorm ) + step(tnorm − 1)] · ua · sin(2πf t) (3.24)

The step function is zero when its argument is smaller than zero, and one when it is larger
than zero. tnorm is the normalized time, from beginning to end of smoothing: tnorm = t/tend .

61
 Figure 3.23: Smooth sine and gradient

The influence of the smooth sine is tested for all three NRBCs (ANSYS CFX NRBC, custom
NRBC, and custom NRBC+PWM) using the ten relaxation factors mentioned at the start of
the section.

3.5.4.2 Linear refinement near boundary

The influence of the linearly refined mesh, as discussed in section 3.1.2, is tested for all three
NRBCs (ANSYS CFX NRBC, custom NRBC, and custom NRBC+PWM) using the ten re-
laxation factors mentioned at the start of the section. This results in thirty simulations to be run.

3.5.4.3 Influence of partitioning

Last, the influence of multiple partitions on the performance and stability of all three NRBCs
is tested. Both serial, 2 partitions, and 4 partitions are tested. Using a relaxation factor of
K = 50s−1 and the total energy equation.

3.5.5 Time-domain impedance boundary condition test cases

A lot of different test cases can be thought of for a TD-IBC because it can basically acousti-
cally mimic any type of domain behind the boundary. The test-cases defined in Table 3.9 are
designed to cover different relevant types of reflective behavior.

Table 3.9: TD-IBC test cases

Nr. Test case R Frequencies [Hz]
1 Acoustically hard wall 1 All
2 Pressure release surface -1 100
3 Non-reflecting outlet 0 All
4 Complex R at fixed frequency 0.5 + 0.5i 100
5 Complex R over range frequencies 0.5 + 0.5i 100-400
6 Levine-Schwinger non-ideal open end R(f ) 100-400

All test-cases are tested against theory. Only the non-reflective test-case can be tested
against the NRBC boundary conditions. The pressure release surface TD-IBC can be tested
against a constant pressure outlet in ANSYS CFX, and an acoustically hard wall TD-IBC can
be tested against a wall in ANSYS CFX. The reflection coefficient R of the test-cases is also
determined using the Multi-Microphone method (see sec. 2.7.1), because it also captures phase

62
information.

3.5.5.1 Acoustically hard wall and pressure release surface

The acoustically hard wall and pressure release surface have relatively simple coefficient sets,
only one non-zero coefficient for both a and b. The acoustically hard wall causes pressure
doubling at the wall, and a reflection coefficient that is real and positive (R = 1). This results
in two non-zero coefficients; a0 = 1 and b0 = 1. For the pressure release surface the reflection
coefficient is R = −1. The phase of R is π, which results in a time delay of half a period.
At a frequency of f = 100Hz this is 0.005 seconds. Using equation 2.42 and ∆t = 10−5 , n is
determined to be 500. The non-zero filter coefficients are a500 = 1 and b0 = 1.

3.5.5.2 Non-reflecting outlet

The non-reflecting outlet in question is different from the ANSYS CFX and custom NRBC
based on the NSCBC formalism, as L5 is specified using the characteristic wave f determined
at the sample plane (see eq. 3.21) instead of the gradients at the boundary. The corresponding
filter coefficients are relatively simple, b0 must be equal to 1.

3.5.5.3 Complex reflection at fixed and varying frequency

A complex reflection of R = 0.5 + 0.5i is tested with filter coefficients derived from one fixed
frequency, and fitted along different frequencies using the method by E.C. Levy [11]. √ The com-
plex reflection coefficient of R = 0.5 + 0.5i represents an absolute reflection of 0.5 · 2, and a
phase of 45◦ . This is done to mimic a partially reflective outlet with a phase change somewhere
between an acoustically hard wall and a pressure release surface. Using equation√2.41 and 2.42,
n is determined to be 125, which results in two filter coefficients: a125 = 0.5 · 2 and b0 = 1,
the rest being zero.

The second case which is fitted among different frequencies has a more complex coefficient
set. The reflection coefficient is fitted from 100 Hz to 400 Hz. At least 125 filter coefficients
are needed. However, calculating a stable IIR filter (also called minimum-phase filters) for this
many filter-coefficients proved to be a difficult task. Even a damped Gauss newton iteration
method (see [13]) does not result in all poles to reside within the unit circle. Varying parame-
ters, such as the order (number of filter coefficients) and frequency, does not result in a stable
filter. It seems that the invfreqz function in Matlab cannot find stable solutions for a high order,
higher than 10, transfer function. The complex reflection coefficient R is nonetheless fitted very
accurately. See Figure 3.24 for fitted R and pole-zero plot using 303 coefficients.

Filter coefficient sets with only a few poles slightly outside the unit-circle, for frequencies
far from the excitation frequency of f = 100Hz, have been obtained. These filter coefficient
sets also proved to be unstable. The hypothesis is, that although the dominant frequency is
f = 100Hz, practically all frequencies that can be represented on the grid are excited, however
small. An extremely small, but non-zero, amplitude corresponding to a freqeuncy of a (un-
stable) pole slightly outside unit-circle causes the amplitude to grow each timestep, until the
amplitudes becomes extreme, and the solver crashes. When looking at the Fourier transform of
the pressure time-series (see Fig. 3.25), it becomes clear that indeed the amplitude of practically
all frequencies are excited, albeit usually extremely small compared to the dominant excitation
frequency of f = 100. These frequencies can be excited by numerous sources, such as solver
and boundary condition characteristics, the finite precision of numbers stored in memory, and

63
 Figure 3.24: Unstable IIR configuration for R = 0.5 + 0.5i, x:pole, o:zero

the non-smooth sine as its wavefront arrives at the boundary.

Figure 3.25: Spectra for pressure sensor P1 , the custom NRBC, K = 50s−1 , Total energy,
u0 = 0m/s.

An obvious alternative would be to use a FIR filter which has no poles by definition. After
simulation this also seems to result in instabilities. This time not because of poles outside the
unit circle, but probably because of the large coefficients which can cause quantization error
[6]. Calculated coefficients are order O(105 ), while for IIR filters the coefficients are generally
of order O(100 ).

By reducing the order of the denominator to 11, and reducing order of numerator to 302, a
stable system in the sense of no poles outside the unit circle can be retrieved (see Fig. 3.26).
However, even this system proves to be unstable during simulations. This might also have to
do with quantification error, as coefficients are generally of order O(101 ). Another cause might
be that the low denominator order might not include enough historical information about g∗.
Another problem with this coefficient set, although not a stability problem, is that the complex
reflection coefficient R is not fitted very accurately.

64
 Figure 3.26: “Stable” FIR configuration for R = 0.5 + 0.5i, x:pole, o:zero

3.5.5.4 Levine-Schwinger non-ideal open end

Levine constructed a set of equations to analytically calculate the complex reflection coefficient
of an unflanged open pipe [10]. These equations are quite complicated, and therefore a rational
equation from Norris is used to approximate the Levine-Schwinger equations [2]. The equations
for the absolute reflection |R| and phase angle φ are given by equations 3.25 and 3.26 respectively.

1 + 0.2 · (ka) − 0.084 · (ka)2

|R| = (3.25)
1 + 0.2 · (ka) + 0.416 · (ka)2

0.6133 + 0.027 · (ka)2

φ = π − 2πa · (3.26)
1 + 0.19 · (ka)2
But also here, as in the complex reflection case, no stable filter coefficient sets are found
after varying various parameters.

65
Chapter 4

Results

The results are separated into two sections, one for the results of the non-reflective boundary
conditions, and the other for the results of the Time-Domain Impedance Boundary Condition
(TD-IBC). In the first section the custom Non-Reflective Boundary Condition (NRBC) with
and without Plane-Wave Masking (PWM) is compared to the ANSYS CFX NRBC and theory.
In the second section the TD-IBC is compared to theory, and for certain complex reflection
coefficients they are compared to other boundary conditions. Those boundary conditions in-
clude a wall and a constant pressure outlet, which are fully reflecting. Also a non-reflective
TD-IBC is tested, as its determination of L5 differs from the other NRBCs (see sec. 2.3.2), and
is compared to the custom NRBC, the PWM NRBC, and the ANSYS CFX NRBC.

The phase angle φ of the measured complex reflection coefficient is not analyzed (but it is
registered in the appendix) for the non-reflective boundary conditions. The goal of non-reflective
boundary conditions, as the name implies, is to be non-reflective, this means that the phasor of
the complex reflection coefficient R is generally small, and is therefore prone to measurement
error of the phase angle. It is also less relevant, as it is not the goal of a non-reflective boundary
condition to influence the phase angle. In contrary to the non-reflective boundary conditions,
the phase-angle φ is fundamental for TD-IBCs, and is therefore taken into account in this section.

The main idea of this chapter is to characterize the different boundary conditions, and to
asses their performance. Several unexpected trends, odd behavior, and lacking performance in
certain areas were observed in various simulations. Not all of the observations can be explained
with absolute certainty, as research would be time-consuming and uncertain. For those obser-
vations only hypothesis are postulated based on theory on other observations.

The results in the following sections are mainly presented by graphs, as showing all the
figures of merit for all simulations would not aid in the readability, given the number of simu-
lations that were run. The numbers behind all the graphs in the results chapter are added to
the appendix.

4.1 Non-reflective boundary conditions

First the influence of different simulation parameters such as the heat transfer equation and
time integration scheme is tested in section 4.1.1 and 4.1.2, of which the best option is used
in subsequent test cases. The influence of the relaxation factor K on reflection coefficient R is
tested in section 4.1.3. Other test cases include the influence on the solution of a mean flow u0
(sec. 4.1.4), partitioning (sec. 4.1.5), analysis of the influence of discontinuous gradients (sec.
4.1.7), and linear refinement near the boundary (sec. 4.1.6).

66
4.1.1 Heat transfer equation
The influence of the heat transfer equation options in ANSYS CFX is tested for both the
custom NRBC as well as the ANSYS CFX NRBC. The results are presented in Figure 4.1 for
the reflection coefficient R, and in Figure 4.2 for the CPU time. The thermal energy equation
is abbreviated as “Thermal E”, and the total energy equation as “Total E”. The acronyms PT
stands for Pressure Transient term on, and VT stands for Viscous work Terms on. Note that the
total energy equation already includes pressure transient behavior on temperature. Explanation
about these terms and equations can be found in section 3.2.2.
What immediately stands out from the reflection coefficient graph (Fig. 4.1) are the high
reflection values when the pressure transient term is not included in the set of equations, as was
expected in section 3.2.2. Although the ANSYS CFX NRBC performs a lot better in this case,
the reflection coefficient R is still very high with R = 22.8% compared to the cases in which
the pressure transient term is included. The CPU time also approximately doubles compared
to the other cases (see Fig. 4.2). While the reflection coefficient R is be lower for the ANSYS
CFX NRBC, the CPU time is higher. The average amount of coeffient loop iterations are
approximately 11.5 for both the ANSYS CFX NRBC as well as the custom NRBC in this case.
The cause remains unknown, but this does not matter as simulations are concerned, as these
parameters (i.e. the exclusion of the pressure transient term) perform abysmally for both the
ANSYS CFX NRBC and the custom NRBC, and thus will not be used. As expected (see sec.
3.2.2), the pressure transient term is essential for correct function of both NRBCs.

Figure 4.1: Influence of heat transfer equations on the reflection coefficient for the ANSYS CFX
NRBC and custom NRBC, K = 50s−1

The reflection coefficient is between R = 3.18% and R = 3.9% for the other four cases tested.
A slight, but almost negligible, difference in favor of the custom NRBC is observed compared
to the ANSYS CFX NRBC. The differences in CPU time between the custom NRBC and the
ANSYS CFX NRBC are negligible. The inclusion of the viscous work terms do not create any
difference, which is to be expected because hardly any attenuation will occur for such short
domains with the used simulation parameters.
The difference in the reflection coefficient between thermal energy equation and total energy
equation is very small. The total energy equation is a more complete equation, but actually
takes approximately an hour shorter to run than the thermal energy simulations. For both the
ANSYS CFX NRBC and the custom NRBC the average number of coefficient loops is approx-
imately three for the total energy simulations, while it is approximately five for the thermal

67
energy simalations. The total energy simulations thus converges faster. The main difference
between the thermal energy and total energy equation is the inclusion of the kinetic energy term
and pressure transient term in the total energy equation (see sec. 3.2.2). It might be that the
pressure transient term is added as an extra equation in the set of equations when the thermal
energy equation is used, while in the total energy equation it is included in the equations, and
therefore takes more time to converge. Based on the results, -a shorter simulation time and
comparable reflection-, the choice is made to use the total energy equation excluding the viscous
work term over the thermal energy equation including the pressure transient term.

Figure 4.2: Influence of heat transfer equations on CPU time [hours : min], K = 50s−1

4.1.2 Time integration

The influence of the time-integration schemes for the custom NRBC are tested, as it only applies
to the custom NRBC and one does not have such level of control over the ANSYS CFX NRBC.
The five cases tested are shown in Figure 4.3. It is observed that a third order approximation
to the time derivative does not result in a lower reflection coefficient over the second order ap-
proximation. The most noteworthy observation however, is that against expectations both the
implicit as well as the half-way predictor showed a higher reflection coefficient R compared to
the explicit first order expression. This might have something to do with the manner in which
the solver constructs and calculates its conservation equations, but is outside the scope of this
research.

68
Figure 4.3: Influence of the time integration scheme on reflection for the custom NRBC, K =
50s−1

4.1.3 Relaxation factor

The reflection of the ANSYS CFX NRBC, and the custom NRBC, both including and excluding
the PWM term, are tested using various relaxation factors K, and compared to the analytical
solution. Both the isothermal and total energy equation are used because of distinct properties
discussed in section 3.2.2. A lower relaxation factor K should result in a lower reflection coeffi-
cient R, but is also more prone to pressure drifting (see sec. 2.4).

4.1.3.1 Reflection coefficient

The results for the ANSYS CFX NRBC are shown in Figure 4.4. The black line is the analytical
solution. For the total energy simulations (red dots) the results are practically equal to the ana-
lytical prediction, and has a minimum reflection coefficient of R = 0.12% for a relaxation factor
K = 1s−1 . More peculiar is the large reflection coefficient in the order of 8% for isothermal sim-
ulations (blue dots) at lower relaxation factors, while one would expect values in the same order
as the total energy equation. The custom NRBC does not show this behavior (see Fig. 4.5)
for isothermal calculations. The problem might possibly reside in the wrong assumption of the
sound speed for isothermal calculations by ANSYS CFX. The returned sound speed, by either
monitor points within the domain as well as via a Fortran utility routine called user getvar, is
equal to the sound speed in which adiabatic expansion can occur. This is physically incorrect,
as the sound speed should be equal to the isothermal sound speed, which is 1.4−0.5 lower (see
sec. 3.2.2). It seems that an adiabatic index of γ = 1.4 is taken to calculate the sound speed
from the primitive field quantities, while this should be γ = 1.0. For both isothermal and total
energy equations no pressure drift is observed, even at a relaxation factor K = 1s−1 .

The CPU time is approximately 3:25 hour for the total energy simulations, and 8:30 hour for
the isothermal simulations. The custom NRBC, both including and excluding the PWM term,
shows the same order of CPU time, and thus does not seem boundary condition dependent. The
total energy simulations need on average three coefficient loops per timestep to meet the con-
vergence criteria, while the isothermal simulations need twelve coefficient loops per timestep.
The isothermal simulations are harder to solve, which might be caused by the heat transfer
needed to maintain a constant temperature (see sec. 3.2.2).

In Figure 4.5 the results for the custom NRBC are shown. The reflection coefficient is con-
sequently slightly below the theoretical value for relaxation factors between K = 23s−1 and
K = 1000s−1 , but follows the same general trend nonetheless. Both the isothermal and total

69
 Figure 4.4: Absolute reflection coefficient vs. relaxation factor, for the ANSYS CFX NRBC

energy simulations bottom out at R = 1.31%, with a minimum of R = 1.156% for the total
energy simulation using a relaxation factor K = 10s−1 . The ANSYS CFX NRBC does not
show this bottoming out for the total energy simulations. It seems that something else is limit-
ing the accuracy of the boundary condition, which could be caused by the discretization error
in the returned pressure and velocity gradients near the boundary, that are determined using
one-sided difference approximations. Another cause could be that the used (fluid) properties are
not accurate. The bottom at around R = 1.3% is low nonetheless, so for practical application
this is probably acceptable.

Figure 4.5: Absolute reflection coefficient vs. relaxation factor, for the custom NRBC

The results for the custom NRBC including the PWM term are shown in Figure 4.6. The
reflection coefficient, for both the isothermal and the total energy simulations, remain fairly
constant at around R ≈ 1.2% throughout most of the relaxation factors, with a minimum of
R = 0.07% for a relaxation factor of K = 1000s−1 . There is no relation with the analytically
determined values for the reflection coefficient R as a function of the relaxation factor K,
which can be expected because in the PWM method the pressure difference due to acoustics is
“masked” from the LRM term, while the LRM term without PWM is the basis of the analytically
determined equation for R as function of K (eq. 2.23).
The minimum around K = 104 s−1 looks impressive due to the logarithmic scale, but the
reflection coefficient is actually roughly only 1.1% lower.

70
Figure 4.6: Absolute reflection coefficient vs. relaxation factor, for the custom NRBC+PWM

4.1.3.2 Pressure drift

The custom NRBC, with and without PWM, showed pressure drift for certain relaxation factors,
the ANSYS CFX NRBC on the other hand, did not. Two figures of merit are introduced to
characterize this behavior, a maximum pressure drift Pdrif t−max and a stabilized pressure drift
Pdrif t−sta , as shown in Figure 4.7. The pressure drift for the total energy equation with a
relaxation factor of K = 50s−1 is shown in Figure 4.7. The pressure time-series of microphone
P2 is plotted to show the pressure drift (see sec. 3.4.1 for the placement of the microphone).

Figure 4.7: Pressure drift figures of merit. Example pressure is P2 (see sec. 3.4) for the custom
NRBC, total energy equation, and K = 50s−1

For comparison of different simulations a pressure drift versus relaxation factor K graph
is constructed, and shown in Figure 4.8. Certain values are missing, which means that they
cannot be determined, for example when the pressure did not stabilize within the simulation
time.
Comparing the isothermal and total energy simulations, it becomes apparent that the
isothermal simulations show much less pressure drift than the total energy simulations. For
the total energy simulations the trend between the pressure drift and relaxation factor is clear,
the pressure stabilizes with increasing relaxation factor K. This trend is expected, as it is the
reason the LRM term is introduced (see sec. 2.4). For the isothermal equations the trend is
less clear and remains fairly constant throughout the relaxation factors, eventually becoming
negligible (lower than 1P a) for relaxation factors higher than K = 500s−1 . The stabilized pres-
sure drift is also close, or equal, to the maximum pressure drift for the isothermal simulations,
meaning that no or hardly any overshoot occurs. It is a different story for the total energy
simulations, which does show significant overshoot, especially at lower relaxation factors. The

71
 Figure 4.8: Pressure drift vs. relaxtion factor K for the custom NRBC and NRBC+PWM

total energy equations seems to stabilize within the simulation time of 0.02s only for relaxation
factors K = 50s−1 or higher.

Using the total energy heat equation, the custom NRBC including the PWM term performs
better and stabilizes sooner than the custom NRBC. But still considerable maximum pressure
drift is observed, see Figure 4.8. When the isothermal simulations are run in combination with
the PWM NRBC however, practically no pressure drift is observed. The small pressure drift
that is observed is in the order of 2P a to 4P a, but all stabilize to 0P a within the simulation time.

The mean pressure of certain simulations seems to asymptotically converge to a value other
than the expected value of 0P a relative pressure, while a relative pressure of 0P a is expected
as the LRM term always “forces” the solution to go to this value. An additional simulation is
run to test if the observed non-zero asymptote continuous after t = 0.2s. This simulation is
five times longer than the standard simulation time, tsim = 1.0s instead of tsim = 0.2s. Apart
from the changed simulation time, the same properties are used as for the simulation shown
in Figure 4.7. The solution of the P2 microphone at x = 2.064m is plotted in Figure 4.9. It
can be seen that the observed mean pressure asymptote did continue after t = 0.2s. This odd
behavior seems to be caused by another restoring “force” (other than the LRM term) with
its equilibrium position elsewhere. It could be that certain properties such as the sound speed
or density are not accurate enough, or that the explicit time integration scheme causes this error.

Figure 4.9: Pressure drift for a simulation of 1 second long. Pressure sensor P2 at x = 2.064m
(see sec. 3.4). Custom NRBC, total energy equation, and K = 50s−1

Although the exact reason behind the observed pressure drift is unclear, there seems to
be a correlation between pressure drift and the varying temperature in combination with the

72
LRM term. The pressure drift decreases when isothermal simulations are run, which suggests
the temperature variation is related to pressure drift. When subsequently the PWM term is
included in the custom NRBC to mask the incoming characteristic pressure perturbation from
the LRM term, the pressure drift basically vanishes. It could be that the restoring force of
the LRM method reacting to the acoustic pressure perturbation in combination with varying
temperature at the boundary is causing the pressure and/or velocity gradients to be incorrectly
returned, and create a “tendency” (or force) for either a positive or negative relative mean
pressure to be attained, which constitutes pressure drift. See section 4.1.3.3 for a more in-depth
analysis of the varying temperature and temperature trends on pressure drift.

Another explanation could be that pressure drift of this kind is expected because the prob-
lem is not “well-posed” enough, and that the ANSYS CFX NRBC has a built-in mechanism
that prevents such ill-posedness.

Based on the results of the reflection coefficient and pressure drift, the conclusion could be
drawn that isothermal simulations using the custom NRBC inlcuding PWM performs the best,
and should therefore be used. But this conclusion is only valid within this idealized numerical
problem. Isothermal simulations would not produce useful results when simulating a stack in
a thermoacoustic device, as it is the temperature gradient that induces sound waves to form,
and thus a heat transfer equation must be used. The custom NRBC including PWM might
be influenced negatively in various situations, such as when very viscous fluids are used, with
significant turbulence, or when the sampled wave is not traveling perpendicular from the sample
plane towards the boundary. In these situations the sampled wave will not be the same as the
actual wave at the boundary. But even when this does happen, it is not expected to cause
major problems as it is only used to mask the acoustic contribution to the pressure from the
LRM term. Only when the phase difference is off by more than a quarter wavelength, problems
can be expected, as it will give positive feedback to the LRM restoring term.

4.1.3.3 Temperature trends

One of the possible causes of the previously mentioned problems of pressure drift is the varying
temperature. Because ANSYS CFX itself determines the temperature at the boundary using
its conservation equations (via the option opening temperature in ANSYS CFX ), one does not
have direct control over the temperature that influences the solution. The temperature evolu-
tion in time for all three NRBCs, using the total energy equations and a relaxation factor of
K = 50s−1 , is plotted in Figure 4.10. The temperature does not vary in isothermal simulations
by definition, therefore it is useful to look at isothermal simulations.

All three NRBCs seem to converge to a mean temperature that is ∆T = 0.175K higher than
the expected T = 298.15K. The ANSYS CFX NRBC converges the fastest to this higher mean
temperature, the custom NRBC comes in second, while the custom NRBC including PWM
lags the most. Maybe some very slight attenuation of the acoustic waves takes place, of which
the energy changes its form to heat by virtue of energy conservation. However, no definitive
explanation can be given about the cause of this phenomenon, but it might influence, or even
instigates, pressure drift.

The increased temperature also influences the density and sound speed in the domain, albeit
slightly. The density decreases with temperature
√ (ideal gas law: ρ = p/RT ), but the sound speed
increases (sound speed, ideal gas: c = γRT ). In the definition of the amplitude variation L5
(eq. 2.20), which is used in the custom NRBC and custom NRBC+PWM to determine the
pressure at the next timestep, both the mean density ρ0 and the small signal sound speed c0

73
Figure 4.10: Temperature vs. time for the ANSYS CFX NRBC, custom NRBC, and custom
NRBC+PWM. Total energy, K = 50s−1 .

are used as factors in the equation. These quantities are determined at initialization of the
simulation, using the pressure and temperature at initialization. The hypothesis is that that
due to a slightly changed mean density ρ0 and small signal sound speed c0 , an error is intro-
duced at each timestep. This error builds-up and causes (more) pressure drift. The fact that
isothermal simulations show much less pressure drift underlines this hypothesis. An algorithm
that updates the mean density and small signal sound speed could potentially decrease this
error from building up, and thus preventing pressure drift.

4.1.4 Mean velocity

The influence of a mean flow on the reflection and pressure drift is tested for the ANSYS CFX
NRBC and the custom NRBC, the last both including and excluding the PWM term. Both the
isothermal and total energy equation are used, for a relaxation factor K = 23s−1 and K = 50s−1
(these relaxation factors are considered optimal, see sec. 3.5.3).

4.1.4.1 Reflection coefficient

The reflection coefficient R of the isothermal simulations is shown in Figure 4.11, from which
it can be seen that for the custom NRBC and PWM boundaries, the reflection coefficient R
increases significantly with mean velocity. The ANSYS CFX NRBC performs abysmally for
all mean flows tested, which can be expected from the isothermal results discussed in section
4.1.3, although a slightly decreasing trend can be observed. A possible explanation could be
that increasing the mean flow actually makes up for the wrongly calculated sound speed by
shear coincidence. With additional mean flow the sound wave travels faster than the isothermal
sound speed c0 = 295.86m/s towards the boundary, which is closer to the adiabatic sound speed
c0 = 346.13m/s that seems to be incorrectly used by ANSYS CFX in the LODI-relations for
isothermal calculations (see sec. 4.1.3). When linearly extrapolating the reflection coefficient
to a mean velocity of 50m/s, which is approximately the difference between the isothermal and
adiabatic sound speed, the reflection coefficient is R ≈ 0.8%, which is in the same order as
the ANSYS CFX NRBC using the total energy equation and without mean flow. To test this
hypothesis two additional test cases were run, one with a mean flow of u = 25m/s, and one with
a mean flow of u = 50m/s. The decreasing reflection coefficient trend indeed continued, with a
reflection coefficient of R = 5% for the u = 25m/s simulation, and R = 0.49% for the u = 50m/s

74
simulation. This observation supports the hypothesis that the sound speed is indeed incorrectly
calculated by ANSYS CFX for isothermal simulations, and the mean flow makes up for it.

Figure 4.11: Influence of mean flow on the reflection coefficient for isothermal simulations

In Figure 4.12 the reflection coefficient R as a function of mean flow u is shown for the total
energy simulations. The custom NRBC and PWM boundary conditions show a comparable
trend and values with respect to the isothermal case. The reflection coefficient R of the ANSYS
CFX NRBC increases only slightly with mean flow, outperforming the custom NRBC and
PWM boundaries when mean flow is introduced.

Figure 4.12: Influence of mean flow on the reflection coefficient for total energy simulations

The custom NRBC increases at approximately the same rate as the custom NRBC including
PWM, but the reflection coefficient of the custom NRBC exhibits an unexpected minimum at
u = 0.5m/s, resulting in the custom NRBC including PWM to have the highest reflection
coefficient for simulations with u = 2.5m/s and u = 5.0m/s mean flow. Possibly an error is
introduced in the custom NRBC method that inadvertently is corrected by a small positive
mean flow, of which a slightly inaccurate sound speed seems a plausible cause.
The reason behind the significantly sharper increase of the reflection coefficient of the custom
NRBC and PWM boundary conditions compared to the ANSYS CFX NRBC is unclear. The
sound speed is adjusted for the mean drift in the governing equations of the NRBCs via c =
c0 + u0 . It might be the case that higher mean flow velocities influence the retrieved pressure
and velocity gradients in an unfavorable manner. The “dip” being only present in the custom
NRBC, and not the PWM boundary condition, suggests that the sharp increase is not caused
by the LRM term, as the pressure difference due to acoustics is masked. On the other hand,
the minimum at u = 0.5m/s might be caused by the LRM method, as it is not present in the
PWM boundary condition.

75
 The custom NRBC including PWM showed a very small difference between K = 23s−1 and
K = 50s−1 , which is expected because the LRM term is effectively masked. See the results of
section 4.1.3 for more details.

4.1.4.2 Pressure drift

Not only the reflection coefficient increases with increasing mean flow, also the pressure drift is
influenced considerably. For comparison of simulations between simulations with and without
mean flow, pressure drift graphs similar to Figure 4.8 are created. The pressure drift without
mean flow is subtracted from the pressure drift at mean flow using the following formula:
∆|P | = |Pdrif t−f low | − |Pdrif t−u=0 |. A positive value means that it performs worse compared to
the pressure drift without mean flow.

Figure 4.13: Pressure drift vs. mean flow velocity, isothermal simulations, K = 23s−1 and
K = 50s−1 , ∆|P | = |Pdrif t−f low | − |Pdrif t−u=0 |

The pressure drift difference for the isothermal simulations is shown in Figure 4.13. Practi-
cally all values are positive, which means that the pressure drift is larger. The trend is decreasing
with mean flow velocity, in contrary to what is observed with the reflection coefficient. The
cause might be sought in the propagation of pressure information from the boundary inward.
As the reflection coefficient increases with mean flow velocity, so does propagation of pressure
information from the boundary to the interior, and thus causing a decrease in pressure drift
(i.e. the problem is more “well-posed”).

Figure 4.14: Pressure drift vs. mean flow velocity, total energy simulations, K = 23 and K = 50,
|∆P | = |Pdrif t−f low | − |Pdrif t−u=0 |

In Figure 4.14 the pressure drift difference for the total energy simulations are shown. The
difference is smaller, and even becomes negative for higher mean flow velocities. The pressure
drift for isothermal simulations without mean flow are relatively low compared to the total

76
energy simulations (see sec. 4.1.3). The higher pressure drift difference for the isothermal
simualations therefore work out to be quite similar to the total energy simulations. The con-
clusion that can be drawn based on these observations is that the isothermal simulations are
more sensitive to mean velocity compared to a simulation without mean velocity, but performs
very similar in absolute pressure drift.

4.1.5 Partitioning
Partitioning is often used in computational fluid dynamics to speed up simulations by enabling
the numerical problem to be divided into different sections (partitions), of which each can run
on a single processor core. This enables multiple processor cores to be assigned to solving the
numerical problem, instead of just a single processing core. The influence of partitioning is
tested for the ANSYS CFX NRBC, and the custom NRBC, both including and excluding the
PWM term. The simulations are run serial, on two partitions, and on four partitions. Both
the reflection coefficient and the simulation time is measured. For the test case, with the given
discretization and fluid properties set out in section 3, no difference in the reflection coefficient is
observed between simulations. The simulation time did show a difference between simulations,
of which the speed-up is shown in Figure ??. Using two partitions speeds up the simulation
about 1.6 times, while using four partitions is actually slower than using two partitions, with a
speed-up of 1.38. Assembling and disassembling partitions after each timestep will cost extra
simulation time, for such a small domain of 1000 elements this will probably take more effort
than the gain in processing power [23]. Although the benefit of partitioning on the speed-up of
simulations is slim, the custom NRBC can handle multiple partitions without any performance
hit or stability problems.

4.1.6 Linear refinement near boundary

Preliminary simulations showed an influence of the spatial discretization on the accuracy of the
returned gradients for pressure and velocity by ANSYS CFX. As incorrectly returned gradients
by ANSYS CFX could be the cause of pressure drift, the influence of increasing the mesh
resolution near the NRBC on the reflection coefficient and pressure drift is studied.

4.1.6.1 Difference in reflection coefficient

Using the linearly refined mesh near the NRBC discussed in section 3.1.2, and the total energy
equation as the heat transfer equation, the difference in reflection coefficient is determined. As
the difference in reflection coefficient R is generally small, it will not be clearly visible in the
logarithmic plots such as 4.4. Therefore the difference is plotted in Figure 4.15. The difference
is defined as |∆R| = |Rlinref | − |R|. A negative value means the linearly refined mesh performs
better than the homogeneous mesh, and vice versa.

From the plot (Fig. 4.15) it can be seen that the ANSYS CFX NRBC profits slightly from
the refined mesh. The custom NRBC benefits from the refinement between K = 231/s and
K = 5001/s, while the custom NRBC including the PWM term is negatively influenced by the
refined mesh. It could be that the gradual refinement influences sound speed in the linearly
defined section of the domain because mesh size is not equal for each element, which causes the
arrival of the characteristic wave f from the sample plane to the boundary to be incorrectly
approximated, and therefore influencing the custom NRBC including PWM negatively.

77
Figure 4.15: Difference in reflection coefficient between the homogeneous mesh and the linearly
refined mesh

Another potential source of increased reflection is the increased acoustic CFL number. For
the normal mesh, the CFL number is CF Lnorm = 0.865325, while the CFL number is five times
higher near the boundary for the linearly refined mesh: CF Lref = 4.326625. This is because the
mesh size ∆x is five times smaller near the boundary, while the timestep ∆t has not changed.
The increased CFL number might cause decreased accuracy and stability near the boundary,
especially in an explicit time marching scheme, such as is used in the custom NRBC. Decreasing
the timestep ∆t might solve this problem, at the cost of simulation time. See section 3.1.2 for
more information about the CFL number.

A simulation run with a relaxation factor of K = 50s−1 , total energy, and using the custom
NRBC, but with a timestep of ∆t = 2 · 10−6 , showed a slight decrease in the reflection coeffi-
cient from R = 2.07% to R = 1.8%. The pressure drift was hardly influenced by the increased
temporal resolution, in the order of a few Pascal.

4.1.6.2 Pressure drift

In Figure 4.16 the pressure drift difference is shown as a function of the relaxation factor K.
The pressure drift difference ∆P is defined in section 4.1.4. While the reflection coefficient in
the previous section showed mixed results, the pressure drift has significantly improved over
the homogeneous mesh in section 4.1.3, for most relaxation factors. The only exception is the
custom NRBC including PWM for a relaxation factor higher than K = 1000s−1 .

Figure 4.16: Difference in pressure drift between the homogeneous mesh and the linearly refined
mesh

The decreased pressure drift is probably due to a decreased discretization error of the one-

78
sided finite difference approximation to the first spatial derivative of the pressure and velocity
at the boundary. In conclusion, the increased mesh resolution near the non-reflective boundary
is only significantly helpful for the custom NRBC. The custom NRBC including PWM and the
ANSYS CFX NRBC are hardly influenced, or even negatively influenced.

4.1.7 Sharp gradients in pressure and velocity

The wavefront of the sine has an infinitely sharp jump in the theoretical solution of the gradient,
from 0P a/m to 371.86P a/m (see sec. 3.2.2). Such infinitely sharp jumps are often not accurately
represented by numerical methods, as such jumps cause numerical wiggles in the solution and
are usually mitigated by flux limiters in CFD [21]. The LODI-relations (see eq. 2.20) use the
pressure and velocity gradients to construct the amplitude variation L5 . Thus potential prob-
lems can be expected when the wave arrives at the boundary. The pressure gradients are plotted
at the sample plane location (∆d = 0.4m from the boundary) and at the designated NRBC
in Figure 4.17. The “jump” slightly after 0.01s is “smoothed”, overshoots by approximately
19.6%, and exhibits numerical wiggles that damp out after approximately 0.002s. The small
“bumps” near zero can be explained by the governing sine being at its maximum amplitude,
and therefore more prone to error and noise in the gradients.

Figure 4.17: Pressure gradient for the normal sine velocity excitation, at the sample plane and
custom NRBC boundary

The hypothesis is that these jumps might initiate pressure drift (see sec. 4.1.3), as it seems
that pressure drift starts when the traveling wave hits the boundary. This is tested using a
smooth transition at the wave-front, as defined in section 3.5.4. The resulting pressure gradi-
ents at the sample plane location and custom NRBC boundary is shown in Figure 4.18, from
which it can be seen that the transition is much smoother and continuous. A wiggle at around
t = 0.014s is observed, which might be due to the theoretical gradient not being smooth (but
continuous) at that point (see sec. 3.5.4). The second derivative is discontinuous, which ap-
pearently has an influence on the solution, but the wiggly behavior is far less pronounced than
for the normal sine.

79
Figure 4.18: Pressure gradient for the smooth sine velocity excitation, at the sample plane and
custom NRBC boundary

4.1.7.1 Difference in reflection coefficient

Using the normal sine results, and the total energy equation as the heat transfer equation, the
difference in reflection coefficient is determined. A similar definition for the reflection coeffi-
cient difference is used as in section 4.1.6, in this case: ∆|R| = |Rsmooth | − |Rnormal |. Where a
negative difference means that the smooth sine performs better than the normal sine.

Figure 4.19: Difference in reflection coefficient between the smooth sine and the normal sine,
∆|R| = |Rsmooth | − |Rnormal |

The plot in Figure 4.19 shows the reflection coefficient difference ∆|R| as a function of the
relaxation factor K, for all tested NRBC boundaries. The smooth sine shows lower reflections
for higher relaxation factors for both the ANSYS CFX NRBC and the Custom NRBC. The
PWM NRBC shows a different story, and hardly profits from a smooth sine. For the high
relaxation factors the reflection is already close to 1, which reduces the practical usability of a
smooth sine.

4.1.7.2 Pressure drift

In Figure 4.20 the pressure drift difference is shown as a function of the relaxation factor K.
The pressure drift difference ∆P is defined in section 4.1.4. For the PWM NRBC the pressure
drift hardly differs from the normal sine case, while the custom NRBC is definitely influenced.
For relaxation factors of K = 23 and K = 50 the maximum pressure drift difference is neg-
ative, meaning that the pressure overshoots less. For very high relaxation factors, K > 500,
the custom NRBC performs worse for the smooth sine. This contradicts the hypothesis that
pressure drift is caused by “discontinuity” introduced by the normal sine. The reason behind
this observation is not clear, but in practice one would not use such high relaxation factors

80
with the custom NRBC because it is almost fully reflecting, and therefore not encounter such
pressure drift.

Figure 4.20: Difference in pressure drift of the smooth sine compared to the normal sine, K = 50,
Total energy, ∆|P | = |Pdrif t−f low | − |Pdrif t−u=0 |

In conclusion, it does not seem very helpful to use a smooth sine to prevent drifting or
decrease the reflection coefficient.

4.2 Time-domain impedance boundary condition

It is difficult to compare the TD-IBC directly to the other NRBC’s discussed and tested in this
paper, because a TD-IBC can influence the phase angle φ of the complex reflection coefficient
R. However, three cases are directly comparable. The non-reflective TD-IBC to the ANSYS
CFX NRBC and the custom NRBC both including and excluding PWM, the acoustically hard
wall TD-IBC to a hard wall in ANSYS CFX, and the pressure release surface with a constant
pressure outlet in ANSYS CFX. Furthermore a complex reflection coefficient of R = 0.5 + 0.5i
is tested against theory.
The TD-IBC test cases are all tested using a relaxation factor K = 50s−1 . The relaxation
factor of K = 50s−1 is seen as an optimal value for the custom NRBC (see sec. 2.4), and
therefore also used throughout all previous test cases, which makes comparison more easily
done. Other values for the relaxation factor could also have been chosen, as the TD-IBC uses
PWM that causes the reflection coefficient R to be basically “decoupled” from the relaxation
factor K (see sec. 4.1.3). The influence of the relaxation factor K is thus expected to be small.
Two figures of merit are introduced to quantify the performance of the test cases, which
are the normalized absolute difference with theory ∆|R̄| (eq. 4.1), and the phase difference ∆φ
between the measured and theoretical reflection coefficient (eq. 4.2).

|Rmeasured | − |Rtheory |
∆|R̄| = (4.1)
|Rtheory |

∆φ = φmeasured − φtheory (4.2)

For all TD-IBC simulations no pressure drift is observed, irrespective of the relaxation fac-
tor K used, which can probably be attributed to the manner in which the amplitude variation
L5 is determined (see sec. 3.3.9.2). This method does not rely on the pressure and velocity
gradients on the boundary, but uses an expression for characteristic wave f determined at a
sample plane near the boundary. This supports the hypothesis that the pressure drift in the
custom NRBC and PWM NRBC originates in the determination of the one-sided pressure and

81
velocity gradients near the boundary.

4.2.1 Acoustically hard wall and pressure release surface

The theoretical reflection coefficients of both a hard wall and a pressure release surface are
compared to the measured reflection coefficient of aforementioned boundaries in Table 4.1.
Theoretically both boundary conditions have an absolute reflection of |R| = 1, but the phase
of the acoustically hard wall is φ = 0 rad while the pressure release surface has a phase of
φ = ±π rad.

Table 4.1: Reflection coefficient results for acoustically hard and pressure release surfaces
Boundary condition |R| φ [rad] ∆|R̄| ∆φ [rad]
Total Energy Calculations
Acoustically hard wall TD-IBC 1.0069 -0.0818 0.69% −0.0818
Wall ANSYS CFX 0.9998 -0.1036 −0.02% −0.1036
Pressure release surface TD-IBC 0.9675 3.1057 −3.25% −0.0359
Constant pressure outlet ANSYS CFX 0.9775 -3.1020 −2.25% 0.0396
Isothermal Calculations
Acoustically hard wall TD-IBC 1.0114 0.0617 1.14% 0.0617
Wall ANSYS CFX 1.0151 0.0433 1.51% 0.0433
Pressure release surface TD-IBC 0.8779 3.0893 −12.21% −0.0523
Constant pressure outlet ANSYS CFX 0.9661 3.0585 −3.39% −0.0831

The measured complex reflection coefficient for the TD-IBC boundaries are close to the
theoretical values, for both absolute reflection coefficient R as well as the phase φ. The “hard”
(Dirichlet) constant pressure and constant velocity (hard wall) boundaries in ANSYS CFX
show errors in the same order of magnitude. The error for the TD-IBC is small and comparable
to constant pressure and velocity boundaries. Pressure drifting is not an issue with constant
pressure and constant velocity boundaries.

4.2.2 Non-reflecting outlet

The reflection of the non-reflective TD-IBC is compared to the ANSYS CFX NRBC, custom
NRBC, and custom NRBC including PWM for both a value of K = 50s−1 , and for the best
reflection values in Table 4.2.

Table 4.2: Absolute reflection coefficient |R| for different non-reflecting outlets
ANSYS CFX NRBC Custom NRBC Custom NRBC+PWM TD-IBC
Total E 3.90%/0.12% 3.18%/1.16% 1.21%/0.08% 1.00%
Isothermal 9.07%/8.16% 3.76%/1.31% 1.26%/0.61% 0.45%
K = 50s−1 / Kbest

The nonreflective TD-IBC shows a lower reflection coefficient compared to the other NRBCs
for a reflection coefficient of K = 50s−1 . Taking the best values of the different models, the
TD-IBC is generally within the same order of error.

82
4.2.3 Complex reflection
The complex reflection of R = 0.5 + 0.5i is fitted to one frequency (f = 100Hz), as the
designed complex curve fitted filter coefficient sets of high order
√proved unstable (see sec. 3.5.5).
Theoretically the absolute reflection coefficient is |R| = 0.5 2, and the phase angle is φ =
0.25π rad.

Table 4.3: Reflection results for complex R = 0.5 + 0.5i

Heat transfer f [Hz] |R| φ [rad] ∆|R̄| ∆φ [rad]
Total Energy 100 0.7146 0.7520 0.75% −0.0333
Isothermal 100 0.6994 0.8010 0.77% 0.0156

The results are given in Table 4.3. The measured complex reflection comes very close to
the designed complex reflection coefficient. The isothermal simulation showed a slightly smaller
error than the total energy equations. In conclusion, it is indeed possible to implement an
accurate and stable TD-IBC for a single frequency.

83
Chapter 5

Conclusions and Recommendations

The Time-Domain Impedance Boundary Condition (TD-IBC) based on theory set out by Po-
lifke et al [9][6] has been implemented successfully for a numerical problem where a single
frequency is dominant. No stability problems were observed, and pressure drift was absent for
all TD-IBC simulations. Different boundaries have been simulated, among which are the acous-
tically hard wall, the pressure release surface, a non-reflecting outlet, and a complex reflection
of R = 0.5 + 0.5i. The last mentioned boundary showed a small difference in the absolute
reflection coefficient with theory of |∆R̄| = 0.75%, and a phase angle difference with theory
of ∆φ = −0.0333 rad. In the other simulations the absolute reflection coefficient differed with
theory ranging from |∆R̄| = 0.69% for an acoustically hard wall using the total energy equation
to |∆R̄| = 12.21% for a pressure release surface in an isothermal simulation.

If the impedance or complex reflection coefficient is defined at multiple frequencies, find-

ing the correct filter coefficients to simulate the complex reflection coefficient can become a
cumbersome and daunting task. No stable filter configurations have been found for broadband
reflection coefficients. Thermoacoustic phenomena in gas-turbines are examples were a broad-
band reflection coefficient could be used. If the TD-IBC is to be used in broadband applications
in the future, more research on finding stable filter coefficients must be done. Fortunately,
thermoacoustic devices usually have one dominant operating frequency, and other frequencies
are often negligible.

The non-reflective part of the TD-IBC, called the custom Non-Reflective Boundary Condi-
tion (NRBC), can work independently from the TD-IBC. In addition, a Plane-Wave Masking
(PWM) term is added to the custom NRBC to improve low-frequency behavior, as is proposed
by Polifke [16]. These two boundary conditions were tested against theory, and a built-in NRBC
by ANSYS CFX. For all three NRBCs a low reflection coefficient is attainable in problems with-
out mean flow, but all have distinct peculiarities. The ANSYS CFX NRBC showed abysmal
performance regarding the reflection coefficient in isothermal simulations, which could be traced
back (using mean flow simulations) to an incorrectly assumed isothermal sound speed by AN-
SYS CFX. The custom NRBC seems to “bottom out” at lower relaxation factors to R ≈ 1.3%,
while the ANSYS CFX NRBC shows a minimum of R = 0.12%, and the PWM NRBC even
has a mimimum of R = 0.07%.

Another observation is that the custom NRBC and the custom NRBC+PWM both showed
considerable pressure drift, especially at lower relaxation factors, while the ANSYS CFX NRBC
never showed any significant pressure drift. The pressure drift seems to decrease either when the
PWM term is used, or when the simulation is run isothermal. When both parameters are set, the
pressure drift is reduced to practically zero. The fact that the isothermal simulations show much
less pressure drift suggests that the (varying) temperature is a contributor to the pressure drift.

84
The mean temperature also increases by ∆T = 0.175K in total energy simulations, which alter
the mean density ρ0 and small signal sound speed c0 , that are determined at initialization. These
quantities, that remain constant during the simulation, are used as factors in the definition of the
amplitude variation L5 . L5 is used in the custom NRBC, both including and excluding PWM,
to determine the pressure at the next time-step via time-integration of the LODI relations.
The slight error introduced in the mean density and small signal sound speed might cause a
small error in the temporal derivative of the pressure in the LODI relations, which, after time-
integration, builds up every time-step and induces pressure drift. Three other explanations of
pressure drift could be that the ANSYS CFX NRBC has a built-in “mechanism” to prevent
pressure drift, uses an algorithm to update the mean density and small signal sound speed,
or uses a higher order finite difference expression at the boundaries to construct the pressure
and velocity gradients. The gradients that are used by the custom NRBC are returned by the
ANSYS CFX solver, and determined using a one-side finite difference expression.
Based on the aforementioned explanations on pressure drift, it is recommended to run sim-
ulations with a built-in algorithm to update the mean density and small signal sound speed
during simulation. This might potentially reduce pressure drift.

When simulations are run using a non-zero mean flow velocity, the custom NRBC and PWM
NRBC show a sharp increase in the reflection coefficient, while the ANSYS CFX NRBC showed
much less sensitivity to mean flow velocity. A possible explanation could be that the returned
pressure and/or velocity gradients are negatively influenced by the mean flow. Mean flow is,
however, less important for thermoacoustic applications.

Two optimizations were investigated to potentially positively influence the reflection coeffi-
cient and pressure drift. The first being a linear increase in mesh resolution towards the NRBC
in question, the other being a smoothed sine in the excitation velocity with a continuous first
spatial derivative. The linear refined mesh showed mixed results regarding the reflection coef-
ficient, which might be related to the increased CFL number due to the smaller mesh size ∆x.
The pressure drift on the other hand decreased, especially at lower relaxation factors. For the
smooth sine simulations the reflection coefficient slightly decreased for higher relaxation factors
for both the custom NRBC and the ANSYS CFX NRBC, while the PWM NRBC was hardly
influenced. The pressure drift on the other hand was, unexpectedly, increased by the smooth
sine. This contradicts the postulated hypothesis that the pressure drift is instigated by the
discontinuity in the pressure and velocity gradient at the wave front.

A non-reflective version of the TD-IBC is also tested, which differs from the other NRBCs
in the expression for the outgoing amplitude variation L5 (see sec. 2.3.2). This non-reflective
TD-IBC showed a reflection coefficient of R = 1% for the total energy equation, and unlike the
custom NRBC and PWM NRBC, does not show any pressure drift. For all NRBCs, including
the non-reflective TD-IBC, a reflection coefficient in the order of R = 1% to R = 3% is attain-
able using the correct parameters.

All boundary conditions have been tested using a one-dimensional problem with purely plane
waves. What has not been done is extensive analysis of two-dimensional and three-dimensional
domains, only the correct working of the NRBCs and TD-IBC have been verified for these
meshes. The boundary conditions have not been tested against other, more complex, numerical
problems. It is therefore recommended that the boundary conditions are tested in other, more
demanding, numerical problems to test their performance and robustness.

Other CFD software could use better approximations to the pressure and velocity gradients
at the boundary, and differ in other fundamental aspects relevant to the application of the

85
developed boundary conditions. Implementing the TD-IBC in other CFD packages, and assess-
ing its stability and performance, is therefore recommended. Also, ANSYS FLUENT uses the
C++ programming language, which is generally considered more user friendly compared to the
antiquated FORTRAN77 language ANSYS CFX uses.

Although the TD-IBC is only extensively tested in a very idealized numerical problem, the
results seem very promising for application in thermoacoustic CFD simulations in which a sin-
gle frequency is dominant. In this regard, subsequent testing in more demanding and complex
simulations to characterize the behavior of the TD-IBC in such settings is recommended.

86
Bibliography

[1] ANSYS ICEM User Manual. pages 724–746, 2009.

[2] I.C. Sheng A.N. Norris. Acoustic Radiation from a Circular Pipe with an Infinite Flange.
Journal of Sound and Vibration, 135:85–93, 1989.

[3] S Bühler and T.H. van der Meer. Numerical thermoacoustic experiment for verification of
commercial CFD code. 2009.

[4] K. Fung and J. Hongbin. Impedance and Its Time-Domain Extensions. AAIA, 38(1), 2000.

[5] K.Y. Fung and J. Hongbin. Time-domain Impedance Boundary Conditions for Computa-
tional Acoustics and Aeroacoustics. International Journal of Computational Fluid Dynam-
ics, 18(6):503–511, August 2004.

[6] A. Huber, P. Romann, and W. Polifke. Filter-Based Time-Domain Impedance Boundary

Conditions for CFD Applications. Proceedings of ASME Turbo Expo, 2008.

[7] P. in ’t Panhuis. Mathematical Aspects of Thermoacoustics. 2009.

[8] M.G. Jones and T.L. Parrot. Evaluation of a Multi-Point Method for Determining Acoustic
Impedance. Mechanical Systems and Signal Processing, 3:15–35, 1989.

[9] R. Kaess, A. Huber, and W. Polifke. A Time-Domain Impedance Boundary Condition for
Compressible Turbulent Flow. American Institute of Aeronautics and Astronautics, pages
1–15, 2008.

[10] H. Levine and J. Schwinger. On the Radiation of Sound from an Unflanged Circular Pipe.
Physical Review Letters, 73(4):24, 1948.

[11] E.C. Levy. Complex-Curve Fitting. pages 37–43, 1959.

[12] E.H. Lysen. The Trias Energica : Solar Energy Strategies for Developing Countries. Eu-
rosun Conference, pages 16–19, 1996.

[13] Matlab. Matlab Manual.

[14] J. Oosterhuis. Performance of non-reflective boundary condition in ANSYS CFX. Technical

report, 2012.

[15] T.J. Poinsot. Boundary conditions for direct simulations of compressible viscous flows.
Journal of Computational Physics, 99(2):352, April 1992.

[16] W. Polifke, C. Wall, and P. Moin. Partially reflecting and non-reflecting boundary con-
ditions for simulation of compressible viscous flow. Journal of Computational Physics,
213(1):437–449, March 2006.

[17] L. Selle. The actual impedance of non-reflecting boundary conditions : implications for the
computation of resonators. pages 1–21, 2003.

87
[18] J. Seung-ho and I. Jeong-guon. On the multiple microphone method for measuring in-duct
acoustic properties in the presence of mean flow. 103(3):1520–1526, 1998.

[19] P. Stopford. E-Mail communication with ANSYS. pages 1–3, 2013.

[20] J.W. Strutt Baron Rayleigh. The Theory Of Sound. MacMillan and Co, 1877.

[21] C.K.W. Tam. Computational Aeroacoustics : Issues and Methods. 33(10), 1995.

[22] W. Thompson. Time- Dependent Boundary Conditions for Hyperbolic Systems , II. Journal
of Computational Physics, 439461:439–461, 1990.

[23] Ansys Tutorials. ANSYS CFX-Solver Modeling Guide. 15317(November):724–746, 2011.

[24] C.H. Venner. On the significance of the Laplace equation and the advection diffusion
equation as model problems for Computational Fluid Dynamics. pages 1–5, 2012.

[25] J. Ward, B. Clark and G. Swift. Design Environment for Low-amplitude Thermoacoustic
Energy Conversion DeltaEC Users Guide. 2012.

[26] P. Zwart. Non-reflective Boundary Conditions in CFX-5. 2005.

88
Nomenclature

CEL Command Expression Language

CFD Computational Fluid Dynamics

CFL number Courant-Friedrichs-Lewy number

FIR Finite response filter

IIR Infinite response filter

LODI Locally One-Dimensional Inviscid

LRM Linear Relaxation Method

MMM Two-Microphone Method

MMS Memory Management System

MPI-CH Message Passing Interface CHameleon

NRBC Non-Reflecting Boundary Condition

NSCBC Navier-Stokes Characteristic Boundary Conditions

ODE Ordinary Differential Equation

PDE Partial Differential Equation

PVM Parallel Virtual Machine

PWM Plane-Wave Masking

PWM Plane-Wave Masking

SF Singularity Factor

SPMD Single-Program Multiple-Data

STP Standard Temperature and Pressure

SWR Standing Wave Ratio

TD-IBC Time-Domain Impedance Boundary Condition

TMM Two-Microphone Method

E Total specific energy [J/kg]

K Relaxation factor [1/s]

89
L Characteristic length of domain [m]

P Pressure [P a]

Q Heat (energy) [J]

Re Reynolds number [−]

T Temperature [K]

U Internal energy [J]

W Work (energy) [J]

Z Impedance [Rayl]

Z0 Characteristic impedance [Rayl]

∆t Time increment [s]

Ω Domain space

I¯ Identity tensor

τ̄¯ Stress tensor [N/m2]

δ p̂ Acoustic pressure amplitude [P a]

δ p̂+ Amplitude of forward traveling acoustic pressure wave [P a]

δ p̂− Amplitude of backward traveling acoustic pressure wave [P a]

δ p̂tr Amplitude of transmitted acoustic pressure wave [P a]

δ û Acoustic velocity amplitude [m/s]

δp Acoustic pressure [P a]

δu Acoustic velocity [m/s]

Q̇ Volumetric heat source [J/m3 s]

γ Temperature [K]

fˆ Amplitude of characteristic wave f [m/s]

ĝ Amplitude of characteristic wave g [m/s]

p̂x Amplitude of pressure gradient [P a/m]

ûx Amplitude of velocity gradient [m/s/m]

λi Characteristic speed [m/s]

h...i Instantaneous area average of variable

A+ Moor-Penrose pseudo-inverse of A

J Column vector of source terms

U Column vector of conserved quantities

90
Up Column vector of primitive quantities

Li Amplitude variations of characteristic waves [varies]

M Mean flow Mach number [−]

ω Angular frequency [rad/s]

ωn Normalized angular frequency (w.r.t. grid) [rad/rad]

∂Ω Domain boundary

φ Velocity potential [m2/s]

ψ Phase of reflection coefficient R [rad]

ρ Density [kg/m3 ]

ρ0 Mean density [kg/m3 ]

σ Coupling parameters in LRM method [m/s]

τ Time coordinate in convolution [s]

~ inv
F Inviscid flux vector
~ visc
F Viscid flux vector

f~ Volumetric force field [N/kg]

~q Heat conduction [J/m2 s]

~u Velocity vector [m/s]

~u∂v Velocity vector of boundary [m/s]

ai Filter coefficients of f [−]

bi Filter coefficients of g [−]

c Sound speed [m/s]

c0 Characteristic sound speed [m/s]

d Space coordinate used in MMM [m]

e Internal energy [J/kg]

f Forward traveling characteristic wave amplitude [m/s]

fs Sampling frequency [1/s]

fin Excitation frequency at inlet [1/s]

g Backward traveling characteristic wave amplitude [m/s]

g∗ External perturbation [m/s]

k Angular wave number [rad/m]

p Pressure [P a]

91
p0 Turbulent contribution to pressure [P a]

p0 Mean pressure [P a]

p∞ Pressure at infinity [P a]

t Time coordinate [s]

u Velocity in x-direction [m/s]

u0 Turbulent contribution to velocity [m/s]

u0 Mean velocity [m/s]

v Specific volume [m3/kg]

x Space coordinate [m]

z Complex argument in z-transform [−]

92
Appendix

93
Contents

1 Acoustic theory 2
1.1 Derivation of the small signal sound speed . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Derivation of the wave equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Fourier and Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Derivation of the Helmholtz equation . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Reflection and impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Reflection measurement theory 7

2.1 Two Microphone Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Multi-Microphone Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Standing Wave Ratio Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Non-reflective boundary condition theory 11

3.1 Navier-Stokes Characteristic Boundary Conditions . . . . . . . . . . . . . . . . . 11
3.1.1 Characteristic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.2 LODI-relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.3 Boundary conditions/treatments . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Linear Relaxation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.1 Performance of the LRM method . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Performance of the non-reflective BC in Ansys CFX . . . . . . . . . . . . 21

4 Numerical modeling 23
4.1 General workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 Sample plane creation and use . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 Problems encountered in programming for ANSYS CFX . . . . . . . . . . . . . . 25
4.4 Single and double precision numbers . . . . . . . . . . . . . . . . . . . . . . . . . 28

5 Simulation data 29

6 Source code 37
6.1 MultiMicMethod MATLAB class . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.2 CCL file with user parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.3 Filter coefficient sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

References 40

Nomenclature 42

1
Chapter 1

Acoustic theory

In this chapter an overview is given on important equations and derivations relevant to acoustics
in general.

1.1 Derivation of the small signal sound speed

The sound speed is the speed at which an acoustic wave propagates through a medium. The
sound speed for an isentropic process is defined by equation 1.1. To evaluate the sound speed,
the pressure p needs to be expressed in terms of the density ρ via an equation of state. The
isentropic equation of state used, is the adiabatic gas law (eq. 1.2) and is subsequently expanded
by using a Taylor series around the static density ρ0 (eq. 1.4). The acoustic density δρ, pressure
δp, and velocity δu are defined in equation 1.3.

2 ∂p 
c ≡ (1.1)
∂ρ s=const
p ρ
( ) = ( )γ (1.2)
p0 ρ0
δρ = ρ − ρ0 δp = p − p0 δu = u − u0 (1.3)
δρ B δρ C δρ
p(ρ0 + δρ) = p0 + A( ) + ( )2 + ( )3 + O(δρ4 ) (1.4)
ρ0 2! ρ0 3! ρ0
where A = p0 γ, B = p0 γ(γ − 1), C = p0 γ(γ − 1)(γ − 2) (1.5)
However, the adiabatic gas law is only valid for perfect gases. In the case of liquids, the
coefficients (A,B,C, etc) are determined using experiments or other analysis. By differentiating
equation 1.4 an expression for the sound speed is derived (see eq. 1.6). By noting that fluctua-
tions of the density due to acoustics are generally very small, the terms after the constant term
A
ρ0 are practically zero in these cases and can be neglected. One arrives at an expression from
which the adiabatic bulk modulus can be extracted (eq. 1.7).

A B ρ − ρ0 C ρ − ρ0 2
c2 = + ( )+ ( ) + O(δρ3 ) (1.6)
ρ0 ρ0 ρ0 2ρ0 ρ0
A
c20 = → A = ρ0 c20 (1.7)
ρ0
To construct an expression for c0 in gases, equation 1.4 is rewritten in form 1.8 with the
adiabatic bulk modulus A substituted, and in form 1.9 which is the ideal gas form. When
compared, one can identify c0 (see eq. 1.10).
 
2 B δρ C δρ 2 3
δp = c0 δρ 1 + ( )+ ( ) + O(δρ ) (1.8)
2!A ρ0 3!A ρ0

2
  
γp0 (γ − 1) δρ (γ − 1)(γ − 2) δρ 2
δp = δρ 1 + ( )+ ( ) + O(δρ3 ) (1.9)
ρ0 2! ρ0 3! ρ0
r
γp0
c0 = (1.10)
ρ0
Important to mention is the difference between isothermal and adiabatic sound speeds. For
isothermal calculations the adiabatic index is γ = 1 instead of γ = 1.4, resulting in a lower
sound speed as calculated by equation 1.10. In the isothermal case, heat exchange between
the wave and the environment need to take place at moments of rarefaction and compression
to keep the temperature constant, which is only accurate when the period of the wave is of
the same order or longer than the time it takes to cool the medium. Generally this is not the
case for acoustic waves on earth, and these compress and expand by approximation adiabatically.

1.2 Derivation of the wave equations

To arrive at the linearized wave equation (eq. ??) from the Euler equations (eq. 1.11), the
following assumptions are made:

1. Homogeneous fluid, the fluid composition is uniform throughout the domain.

2. No mass, momentum, and energy source terms.

3. Assuming one dominant direction of propagation (i.e. one-dimensional).

4. In the absence of sound, the fluid is at static pressure, static density, and with a mean
flow of zero.

5. Acoustic disturbances are small compared to static pressure, density, and sound speed.
|δρ|  ρ0 |δpitemize|  ρ0 c20 |δu|  c0

Using the first three assumptions equations 1.12 and 1.13 are constructed based on the mass
and momentum Euler equations.

∂U ~ ~ inv
+∇·F =J (1.11)
∂t

ρt + (ρu)x = 0 (1.12)

(ρu)t + (ρu2 )x + Px = 0 (1.13)

Using equation 1.8, and neglecting higher order terms, a relation for acoustic pressure δp
can be derived (see eq. 1.14). Expanding the 1D continuity (eq. 1.12) and momentum (eq.
1.13) equations using the product rule, substituting the expression for the density in terms of
a static density and the acoustic density (eq. 1.3), and substituting acoustic velocity δu (note:
u0 = 0), one arrives at equations 1.15 and 1.16.

δp = c20 δρ (1.14)
δρt + δρx δu + ρ0 δux + δρδux = 0 (1.15)
ρ0 δut + ρ0 δuδux + δρδut + δρδuδux + δpx = 0 (1.16)
For equation 1.15 the second and fourth term are products of small signal quantities, and
therefor extremely small and can be neglected (i.e. linearized). The same argumentation is valid

3
for the second, third, and fourth terms in equation 1.16, and can be neglected. This results in
a linearized set of equations (see 1.17).

δρt + ρ0 δux = 0 (1.17a)

ρ0 δut + δpx = 0 (1.17b)

The usual convention in acoustics is to use pressure and velocity as the primary field vari-
ables. Taking the time-derivative of equation 1.14 and combining it with equation 1.17a to
eliminate the density, one arrives at equation 1.18.

δpt + ρ0 c20 ux = 0 (1.18)

Using equation 1.17b and 1.18, taking the time- and spatial derivatives, one can choose to
eliminate either pressure or velocity and arrive at form 1.19 or 1.20 of the linear wave equation
respectively. The three dimensional form is derived analogously, and is often written in scalar
velocity potential form (see eq. 1.21). The velocity is defined in equation 1.22.

c20 uxx − utt = 0 (1.19)

c20 pxx − ptt = 0 (1.20)

~ 2 φ − 1 φtt = 0
∇ (1.21)
c20
∇φ = ~u (1.22)

1.3 Fourier and Laplace Transform

Fourier and Laplace transforms are often used to represent signals and functions in other do-
mains, especially periodic functions. In this report the Fourier transform is used to transform
a signal into its Fourier componenents, i.e. frequency content, in the frequency domain. The
Laplace transform is related, but somewhat different to the Fourier transform, and is often
used to solve equations that are hard to solve in the time domain but relatively easy in the
s-domain. The s-domain can be seen as a complex frequency domain. The Fourier and inverse
Fourier transforms are given in equations 1.23 and 1.24. The Laplace and its inverse are given
in equations 1.25 and 1.26.

Z ∞
ū(x, ω) = F[u(x, t)] = u(x, t) · e−iωt dt (1.23)
−∞
Z ∞
−1 1
u(x, t) = F [ū(x, t)] = ū(x, ω) · eiωt dω (1.24)
2π −∞
Z ∞
F (s) = L[f (t)] = f (t) · e−st dt (1.25)
0
Z γ+iT
1
f (t) = L−1 [F (s)] = lim F (s)ds (1.26)
2πi T →∞ γ−iT

Also discrete forms of the Fourier and Laplace transform exist, which are called the Discrete-
Time Fourier Transform (DTFT) and the Z-transform respectively.

4
Convolution integral in Laplace domain An important property is that convolution in
the time-domain is multiplication in the Laplace domain.

Time-domain Laplace domain

Rt
(f ∗ g)(t) = 0 f (τ )g(t − τ )dτ F (s) · G(s)

1.4 Derivation of the Helmholtz equation

Many acoustic quantities like impedance and reflection are defined as a function of frequency.
However, the wave equation and other time-domain methods cannot use these quantities directly.
To be able to use and characterize these quantities a frequency-domain method must be used.
The Helmholtz equation is an equation that works in the frequency domain instead of the
time-domain. In steady-state analysis of acoustics, the time-series behave in a predictable
periodic manner in time, and can be eliminated by assuming time-harmonic behavior. The
time-harmonic behavior assumption used (eq. 1.27) is substituted in the wave equation (eq.
1.21) to obtain, after further derivation, the Helmholtz equation (eq. 1.31). This method of
decoupling ~x and t is effectively the method of separation of variables.
For more complex signals a sum of Fourier coefficients can be used via a Fourier transform.

φ(~x, t) = φ(~x, ω)eiωt (1.27)

By calculating the gradients and time-derivatives (eq. 1.28), substituting them into the
velocity potential form of the wave-equation (eq. 1.21), an equation (eq. 1.29) can be derived
from which the time-dependent part can be eliminated by bringing the term outside the brackets
(see eq. 1.30). Finally arriving at the Helmholtz equation (eq. 1.31).

~ x, t) = [∇φ(~
∇φ(~ ~ x, ω)]eiωt → ~ 2 φ(~x, t) = [∇
∇ ~ 2 φ(~x, ω)]eiωt (1.28a)
φt (~x, t) = φ(~x, ω)ieiωt → φtt (~x, t) = −φ(~x, ω)ω 2 eiωt (1.28b)

~ 2 φ(~x, ω)]eiωt + [ 1
[∇ φ(~x, ω)ω 2 ]eiωt = 0 (1.29)
c2
~ 2 φ(~x, ω) + ω2
[∇ φ(~x, ω)]eiωt = 0 ∀t (1.30)
c2
~ 2 φ(~x, ω) + k 2 φ(~x, ω) = 0 ω
∇ where k= (1.31)
c
The Helmholtz equation is a function of position and frequency rather than of position and
time like the wave-equation. In the equations, k is the wave number, which can be seen as a
spatial frequency. The Helmholtz equation is a PDE of the elliptic type.

5
1.5 Reflection and impedance
The definition for reflection and impedance are given in equations 1.32 and 1.33 respectively.
Where p̂ is the acoustic pressure phasor, and û is the acoustic particle velocity phasor. These
variables can become complex, signifying a phase difference. This means also the reflection and
impedance can become a complex quantity. The + and − subscript signify the complex pressure
of the forward and backward traveling waves respectively. These quantities can be derived from
the solution of the Helmholtz equations, and is discussed in section 2.1.

p̂− (f )
R(f ) = (1.32)
p̂+ (f )
p̂(f )
Z(f ) = (1.33)
û(f )
When the same assumptions as for the construction of the wave equations apply to the
problem, the characteristic impedance (eq. 1.34) can be used. Which is an inherent property
of the medium.

Z0 = ρ0 · c0 (1.34)
Using the definition of the impedance (eq. 1.33) and the characteristic impedance (eq. 1.34)
the acoustic pressure δp can be easily expressed as a function of the acoustic velocity δu and
the characteristic impedance Z0 (eq. 1.35).

δp = ρ0 c0 δu = Z0 δu (1.35)
The impedance is a complex quantity, and often expressed in a resistive (R) and reactive
(X ) part (see eq. 1.36).

Z = R + iX R = R(~x, ω) and X = X (~x, ω) (1.36)

6
Chapter 2

Reflection measurement theory

Different methods exist to quantify reflection or impedance of a sample or interface. Among the
most used are the Two-Microphone Method (TMM), the Multi-Microphone Method (MMM),
and the Standing-Wave Ratio Method (SWR Method). The MMM is the method of choice
because the other two methods have certain disadvantages that make them less suitable for this
application. The TMM is simple, but very sensitive to sensor positioning, especially if the mi-
crophone separation is near one half wavelength [1][2]. The SWR method is on the other hand
very accurate and reliable, but very time-consuming because a probe with a pressure sensor
needs to moved through the tube [1][2]. As ANSYS CFX does not support moving monitor
points (to measure pressure), and approximating the Standing-Wave Ratio as a function of
position might be a cumbersome task, the SWR-Method is at a disadvantage. Also, the MMM
method is the only method that takes into account mean flow explicitly.

2.1 Two Microphone Method

The Two-Microphone Method (TMM) uses, as the name implies, two microphones positioned
a certain distance from each other within the domain (see fig. 2.1).

Figure 2.1: Two Microphone Method in impedance tube

The method is based on the Helmholtz equation (2.1) and its solution (2.2). The Helmholtz
equation works in the spatial frequency domain.
ω
∇2 p̂ + k 2 p̂ = 0 where k= (2.1)
c0
p̂(x, f ) = A(f ) · e−ikx + B(f ) · eikx (2.2)
By evaluating the solution of the Helmholtz equation (eq. 2.2) for the position of sensor
1 and 2 (eq. 2.3 and 2.4), a fully defined system of equations (eq. 2.5) can be created with
which the forward and backward travelling pressure phasors p̂+ and p̂− at the boundary can
be calculated using either matrix inversion or Gaussian elimination. With this information,
the calculation of the complex reflection coefficient is straightforward by using eq. 1.32. The

7
introduction of a new coordinate d simplifies matters a little bit because p̂+ and p̂− can be
evaluated at d=0 by using the fact that L − d = x.

p̂(x1 , f ) = A(f )e−ikx1 + B(f )eikx1 = A(f )e−ikL eikd1 + B(f )eikL e−ikd1 (2.3)
p̂(x2 , f ) = A(f )e−ikx2 + B(f )eikx2 = A(f )e−ikL eikd2 + B(f )eikL e−ikd2 (2.4)
e 1 e−ikd1
 ikd    
p̂+ p̂1
ikd −ikd = (2.5)
e 2 e 2 p̂− p̂2
Where p̂+ = A · e−ikL and p̂− = B(f ) · eikL . By normalizing equation 2.3 and 2.4 by p̂(x1 , f ),
a version of the formula for the reflection coefficient which is not dependent on the absolute
position of the pressure sensors, but just the distance s between the sensors can be derived(eq.
2.6). H21 = p̂p̂21 .

H21 − e−iks
R(f ) = (2.6)
eiks − H21

2.2 Multi-Microphone Method

In the Multi-Microphone Method (MMM) three or more microphones are used, rendering the
system of equations over-determined. The system is solved using a Least-Squares method.

Figure 2.2: Multi Microphone Method in impedance tube

Instead of just the wave numbers, plane wave propagation constants Γ (see eq. 2.7) are
used which correct for the effects of mean flow and attenuation effects due to viscosity and
turbulence. When these effects are not present, Γ just evaluates to the wave number k.

k − iδ
Γ± = ∓ (2.7)
1±M
Where k is the wave number, M is the Mach number, and δ is an attenuation constant. The
new system of equations A~x = ~b (eq. 2.8) resembles equations 2.5, but is now over determined.
 −ikx1
eikx1   
  
e p̂1 
 e−ikx2 eikx2  p̂+

 
p̂2


 ...
 = (2.8)
...  p̂− 
 ... 

e−ikxn eikxn p̂n
 

The Least-Squares solution can be obtained using a Moor-Penrose inverse A+ (eq. 2.9).
Where AH is the Hermitian matrix of A.

~x = A+~b = (AH A)−1 AH~b (2.9)

The sensitivity to measurement errors is expressed in the Singularity Factor SF. The lower
the SF value, the lower the estimation error. Seung-Ho et al [1] showed that the SF is minimal at
a kratio (see eq. 2.10) of 0.5. Where kcr is a critical wave number corresponding to the maximum

8
wave number that can be uniquely determined by the set-up. Also the spacing between different
sensors seems to have an influence on the SF value, where a uniform equidistant spacing between
the sensors resulted in the lowest SF. There are no limitations on the spacing regime in the test
cases, therefore a uniform equidistant spacing is used. Because of the use of uniform equidistant
spacing, a simpler formula to determine the optimal distance between sensors can be used (eq.
2.11). The wave number k of the excited wave is known at forehand, from this kcr can be
calculated using eq. 2.10. Subsequently ∆xmic can be determined using 2.11.

k
kratio = (2.10)
kcr

kcr · ∆xmic = π · (1 − M2 ) (2.11)

When no mean flow is present, four sensors per wavelength are optimal.

Attenuation effects Two effects contribute to the attenuation constant δ (eq. 2.12): the
viscothermal effects (δν ) and the turbulent effects (δt ). The viscothermal effects are described
by equation 2.13, and the turbulent effects are described by equation 2.14.

δ = δν + δt (2.12)
r 
ω µ κ
r
δν = √ + (γ − 1) (2.13)
2 2a0 c0 ρ0 ω ρ0 ωcp
Ψ0 Re
 
ΨM
δt = 1+ (2.14)
2a0 2Ψ
Where µ is the shear viscosity constant, κ is the heat conduction coefficient, and Re is
the Reynolds number. The coefficient of friction for turbulent flow Ψ is calculated from the
Prandtl universal resistance law (eq. 2.15). See literature from Sueng-Ho [1] for more detailed
information on the subject.

1 √
√ = 2 · log10 (Re Ψ) − 0.8 (2.15)
Ψ

2.3 Standing Wave Ratio Method

In practice the standing wave field is measured by moving a microphone along the tubes length
(see fig. 2.3), which retrieves the pressure maximum at each point. From this the Standing
Wave Ratio (SWR) (eq. 2.16) can be determined, which is in words the absolute value of the
pressure maximum divided by the absolute value of the pressure minimum within the domain.
The amplitude of the reflection coefficient can then be calculated using equation eq. 2.17.

Figure 2.3: SWR Method in impedance tube

9
 | p̂max |
SW R ≡ (2.16)
| p̂min |
SW R − 1
|R| = (2.17)
SW R + 1
However, with this method one loses phase information of the standing wave field. By
measuring the distance d from the boundary to the first extremum (be it minimum or maximum),
the complex reflection coefficient including phase information can be calculated (eq. 2.18)

SW R − 1 i2(kd)max SW R − 1 i2(kd)min
R= e R=− e (2.18)
SW R + 1 SW R + 1

10
Chapter 3

Non-reflective boundary condition

theory

This chapter discusses the assumptions made, and the derivation of, the Navier-Stokes Charac-
teristic Boundary Conditions (NSCBC). In the last section a formalism to construct boundary
conditions based on the NSCBC method is discussed.

3.1 Navier-Stokes Characteristic Boundary Conditions

It is possible to decouple the quasi-linear first order Euler equations and extract expressions
for characteristic waves (Riemann invariants). Using this Riemann invariant transformation to
construct a set of relations called the Locally One-Dimensional Inviscid (LODI) relations, the
dependent flow variables on the boundary can be determined in terms of characteristic waves.

In short, the methodology’s assumptions are:

ˆ Ideal gas assumption as equations of state

ˆ The Viscous and heat terms are neglected (Euler equations)

ˆ Locally one-dimensional (thus one direction of wave propagation), and the transverse
terms are neglected.
The NSCBC is derived from the Euler equations, however, this does not limit the calculation
of the interior domain to the Euler equations. The interior of the domain can be solved using
the full Navier-Stokes equations by adding boundary conditions that account for the viscous
dissipation and thermal diffusion. To obtain the missing dependent variables on the boundary
one can extrapolate from the inner domain, or use a conservation equation on the boundary.
The NSCBC method uses the latter [6].

3.1.1 Characteristic Analysis

The characteristic analysis used in the construction of the hyperbolic boundary condition for-
malism by Thompson [6], uses the following general hyperbolic equation as a basis:

∂U ∂F1 ∂F2 ∂F3

+ + + +D=0 (3.1)
∂t ∂x1 ∂x2 ∂x3
Where U are the conserved quantities, F the fluxes, and D the inhomogeneus source terms.
From this equation it can be observed that it has the same form as the Euler equations in

11
conservative PDE form (1.11). The hyperbolic model equation (3.1) is in conservative form,
however, boundary conditions further on are greatly simplified by using the primitive form
(primitive quantity vector denoted by a p superscript: Up ). For example: it is easier to prescribe
the velocity u at the boundary than the momentum ρu. Thompson transforms the conserved
form to a primitive form before characteristic analysis:

∂U ∂Up ∂Ui
=P where Pij ≡ (3.2)
∂t ∂t ∂Ujp
Which can be obtained by using the chain rule, after all the conserved quantities U can be
written as a function of primitive quantities Up : U = f (Up ). For example, momentum ρu is a
function of ρ and u. For the flux terms one may do the same:
∂Fk ∂Up
= Qk where k = 1, 2, 3 (3.3)
∂xk ∂xk
The matrix Qk has elements:
∂Fik
Qkij ≡ (3.4)
∂Ujp
Substitution into eq. (3.1) and left multiply by P −1 results in the primitive form of the
model equation:
∂Up ∂Up ∂Up ∂Up
+ A1 + A2 + A3 + Dp = 0 (3.5)
∂t ∂x1 ∂x2 ∂x3
Where Ak = P−1 Qk and Dp = P−1 D. The system of equations is subsequently lumped
into two parts (3.6). The Cp part is composed of the transverse and source terms which do not
substantially add to the characteristic analysis, but are just carried along passively.
∂Up ∂Up ∂Up ∂Up
+ A1 +C=0 C = A2 + A3 +D (3.6)
∂t ∂x1 ∂x2 ∂x3
The system of partial differential equations (PDE) (3.5) is quasi-linear and first order, such
a system is hyperbolic iff the eigenvalues of the matrices Ak are real. However, these multiple
A matrices in multi-dimensional problems are not simultaneously diagonalizable, and there-
for cannot be decoupled and consequently do not have an unique direction of propagation.
Multi-dimensional problems cannot be put into characteristic form, however, one-dimensional
problems with only one A matrix can be diagonalized and thus put into characteristic form.
For multi-dimensional problems locally one-dimensional wave propagation is assumed and the
transverse terms C drops out:
∂Up ∂Up
+ A1 =0 (3.7)
∂t ∂x1
The coefficient matrix A1 in the system of equations (3.7) can be used for eigenvalue ex-
traction. The left l and right r eigenvectors have the following properties:
lTi A1 = λi lTi A 1 ri = λ i ri where i = 1, . . . , m (3.8)
The eigenvalues can be determined via the well-known formula det(A1 − λI) = 0. By
using a similarity transformation the matrix A1 can be transformed to a diagonal matrix with
eigenvalues:
S−1 A1 S = Λ (3.9)
The matrix S and S−1 can be constructed by adding the right eigenvectors as columns or the
left eigenvectors as rows respectively. By substituting eq. (3.9) into eq. (3.6), and subsequently
left multiply with S−1 one arrives at:
∂U ∂U
S−1 + L + S−1 C = 0 where L = ΛS−1 (3.10)
∂t ∂x1

12
In component form:
∂U ∂U
lTi + Li + lTi C = 0 where Li ≡ λi lTi (3.11)
∂t ∂x1
The original hyperbolic equations are now put into a characteristic form, as a function of
the amplitudes of the characteristic waves Li .

3.1.2 LODI-relations
By using the characteristic analysis procedure in section 3.1.1 on the Euler equations (1.11) the
so-called Locally One-Dimensional Inviscid (LODI) relations can be derived. First the column
vectors constructed:
       
 ρ   ρ  ρuk   0 
       P3 
e p (e + p)u −ρ k=1 uk gk 
       
k 

 
 
   
   
U ≡ ρu1 Up ≡ u1 F k ≡ ρuk u1 + δk1 p D≡ −ρg1 (3.12)
ρu u ρu u + δ p −ρg
       
 2  2  k 2 k2 

 
 
 
 
 
 
 2


    
ρu3 u3 ρuk u3 + δk3 p −ρg3
    

Where δ is the Kronecker delta function. Next the Jacobian matrix P (3.2) and Q (3.4) are
constructed:
1 0 0 0 0
 
 21 3i=1 u2i 1
P
(γ−1) ρu1 ρu2 ρu3 


P=  u1 0 ρ 0 0 
 (3.13)
 u2 0 0 ρ 0 
u3 0 0 0 ρ
Using the two equations of state (eq. 3.14), the matrix Qk can be determined from (3.4):

e ≡ 21 ρ 3k=1 u2k + 
P
(3.14)
p ≡ (γ − 1)

uk 0 δk1 ρ δk2 ρ δk3 ρ

 
1 P3 2 1
i=1 ui uk ( γ−1 + 1)uk fk1 fk2 fk3
 
2
k

Q =
 uk u1 δk1 (1 + δk1 )ρuk δk2 ρu1 δk3 ρu1   (3.15)
 uk u2 δk2 δk1 ρu2 (1 + δk2 )ρuk δk3 ρu2 
uk u3 δk3 δk1 ρu3 δk2 ρu3 (1 + δk3 )ρuk

Where fki is defined as:

3
1 X
fki = ρ((2 + δki )ui uk + δki (1 − δij )u2j ) (3.16)
2
j=1

In which Ak = P−1 Qk can be used to find A1 . Matrix Q1 becomes:

u1 0 ρ 0 0
 
P3
1 2
i=1 ui u1
1
( γ−1 + 1)u1 1 2 2 2
2 ρ(3u1 + u2 + u3 ) ρu2 u1 ρu3 u1 

1
2
Q = 2
u1 1 2ρu1 0 0  (3.17)
 
 u1 u2 0 ρu2 ρu1 0 
u1 u3 0 ρu3 0 ρu1

13
 The inverse of the Jacobian matrix P is found to be:
 
1 0 0 0 0
 1 3 u2 (γ − 1) (γ − 1) −(γ − 1)u1 −(γ − 1)u2 −(γ − 1)u3 
P
−1
 2 i=1 i 
P =  −u1 ρ−1 0 ρ−1 0 0 
 (3.18)
 −u2 ρ −1 0 0 ρ−1 0 
−u3 ρ −1 0 0 0 ρ−1

Which results in A1 :
 
u1 0 ρ 0 0
 0 u1 γp 0 0 
 0 ρ−1 u1 0 0 
 
  (3.19)
0 0 0 u1 0 
0 0 0 0 u1
From which the eigenvalues can be determined using cofactor expansion (where Z is defined
for convenience):
 
u1 − λ 0 ρ 0 0
 0 u1 − λ γp 0 0 
1 −1
 
det(Z) = det(A − λI) = 0 =   0 ρ u1 − λ 0 0  (3.20)
 0 0 0 u1 − λ 0 
0 0 0 0 u1 − λ

det(Z) = (u1 − λ) · det(Z2:5,2:5 ) = 0

(u1 − λ) · [(u1 − λ) · det(Z3:5,3:5 ) − ρ−1 · det(Z2,3:5 : Z4:5,3:5 )] = 0
(u1 − λ) · [(u1 − λ)4 − ρ−1 γp(u1 − λ)2 ] = 0
(u1 − λ) · [(u1 − λ)4 − c2 (u1 − λ)2 ] = 0 (3.21)
3 2 2
(u1 − λ) · [(u1 − λ) − c ] = 0
(u1 − λ)3 → λ2,3,4 = u1
(u1 − λ)2 − c2 = [(u1 − λ) + c][(u1 − λ) − c] → λ1,5 = (u1 ∓ c)
The eigenvalues λ are sorted in increasing order, and represent the characteristic velocity at
which a particular wave mode propagates. The speed of sound is c2 = γp ρ (ideal gas, adiabatic).
The (left) eigenvectors are determined using (3.8), putting them into the form lTi (A1 − λI) = ~0,
and solve for l corresponding to each eigenvalue.
The left eigenvectors are:
lT1 = (0, 1, −ρc, 0, 0)
lT2 = (c2 , −1, 0, 0, 0)
lT3 = (0, 0, 0, 1, 0) (3.22)
lT4 = (0, 0, 0, 0, 1)
lT5 = (0, 1, ρc, 0, 0)
From eq. (3.11) the characteristic waves can be determined:
∂p
∂U
L1 = λ1 (0, 1, −ρc, 0, 0) ∂x 1
= λ1 ( ∂x 1
− ρc ∂u
∂x1 )
1

∂U ∂ρ ∂p
L2 = λ2 (c2 , −1, 0, 0, 0) ∂x 1
= λ2 (c2 ∂x 1
− ∂x 1
)
∂U ∂u2
L3 = λ3 (0, 0, 0, 1, 0) ∂x 1
= λ3 ∂x1 (3.23)
∂U
L4 = λ4 (0, 0, 0, 0, 1) ∂x 1
= λ4 ∂u
∂x1
3

∂p
∂U
L5 = λ5 (0, 1, ρc, 0, 0) ∂x 1
= λ5 ( ∂x 1
+ ρc ∂u
∂x1 )
1

14
 Rewriting these equation in terms of the spatial derivative of the primitive flow quantities
along propagation x1 :
∂ρ 1 L2 1 L5 L1
∂x1 = c2 [ u1 + 2 ( u1 +c + u1 −c )]
∂p 1 L5 L1
∂x1 = 2 ( u1 +c + u1 −c )
∂u1 1 L L1
∂x1 = 2ρc ( u1 +c − u1 −c ) (3.24)
5

∂u2 L3
∂x1 = u1
∂u3 L4
∂x1 = u1

These two sets of equations (3.23)(3.24) form the LODI-relations. Albeit, different forms of
(3.24) may be constructed, depending on the application. Using eq. (3.7) the time derivatives
of the primitive flow quantities can easily be obtained using the spatial derivatives (3.24) and
the matrix A1 (3.19), which results in the set of LODI relations in temporal derivatives (3.26).

∂Up ∂Up
= −A1 (3.25)
∂t ∂x1
∂ρ
∂t = − c12 [L2 + 12 (L5 + L1 )]
∂p
∂t = − 12 (L5 + L1 )
∂u1 1
∂t = − 2ρc (L5 − L1 ) (3.26)
∂u2
∂t = −L3
∂u3
∂t = −L4
The characteristic velocities λi and the corresponding characteristic waves Li are for num-
bers 1 and 5 the sound waves moving in the negative and positive directions respectively, number
2 the entropy advection, and numbers 3 and 4 are the transverse waves for u3 and u4 respectively.

3.1.3 Boundary conditions/treatments

In the formalism for the treatment of boundary conditions by Thompson [6], a clear distinction
is made between a “boundary condition” and a “boundary treatment”.

ˆ Boundary Condition

1. Single mathematical expression

2. External to the calculation of U within interior of domain Ω.
~
∂U
3. Contributes to, but does not itself define ∂t

ˆ Boundary Treatment
∂U
1. Complete algorithm for defining the values of ∂t along the boundary.
2. Incorporates information of U from both interior Ω and the boundary conditions at
∂Ω.

Where U are the primitive flow quantities and Ω is the interior of the domain. Directly
defining U at the boundary is prohibited in the formalism, the boundary condition only con-
tributes to the determination of ∂U
∂t at the boundary. To overcome this, one could set certain
primitive flow quantities at the boundary as an initial condition, and force the time derivative
of the primitive flow quantity to be zero at the boundary at all times.
To construct boundary conditions the characteristic velocities λ are used to determine
whether a characteristic wave travels from the interior Ω outwards or into the volume. When
the wave travels outwards trough the boundary, Li on the boundary can just be calculated using
one-sided derivative approximations and the definitions for Li (3.23). When the wave travels

15
inwards, the corresponding Li must be determined using the boundary condition specified on
the boundary. From this general procedure all kind of boundary conditions can be derived.
Some frequently used boundary conditions are discussed by Poinsot and Thompson [4][6], but
only the subsonic nonreflecting outflow boundary condition is treated here.

A Subsonic Nonreflecting outflow A nonreflecting boundary condition is seen within this

formalism as a wave traveling back into the domain Ω that does not exhibit any time-varying
behavior, i.e. is constant in time. As wave motion is by definition time-varying, this is equivalent
to stating that there is no reflected wave. Using equation (3.11) and neglecting transverse terms
to derive an expression for the ’reflected’ wave L1 :
∂p ∂u1
( − ρc ) + L1 + ρcg1 = 0 (3.27)
∂t ∂t
The expression within the parenthesis is the time derivative of the well-known Riemann in-
variant in acoustics p = ρ0 c0 u. Rewriting the invariant to p − ρ0 c0 u = 0, one can easily see that
the term within the parenthesis should equal zero. Eq. (3.27) reduces to L1 = −ρcg1 , where
g1 are the specific volumetric forces. Often the only volumetric force terms used are those due
to gravitation, which are usually more or less homogeneous over the domain (especially with a
nearly incompressible fluid). Usually one does not have to account for the volumetric forces, as
they are often not present in the direction of propagation x1 . The creation of a nonreflective
boundary is equal to stating: L1 = 0.

Subsonic
x1 = a1 x1 = b1
With volumetric forces L5 = ρcg1 L1 = −ρcg1
Without volumetric forces L5 = 0 L1 = 0

Table 3.1: Wave amplitude expressions at the boundaries

16
3.2 Linear Relaxation Methods
The L1 amplitude can be set to zero to effectively eliminate reflection, but results often in
an ill-posed problem. Then a drift of the mean flow field quantities occurs because far–field
information about pressure cannot propagate into the domain. To overcome the problem of
ill-posedness one could set the far-field pressure p∞ , but this would create an acoustically hard
surface and thus reflection. One way or another, information from the far–field pressure must
be able to travel into the domain, while maintaining a more or less non-reflecting boundary.
The LRM model is the simplest boundary treatment for L1 that accommodates both:

L1 = K(p − p∞ ) (3.28)
The relaxation factor K can be defined in multiple ways. An expressions for K is:

σ(1 − M2 )c
K= (3.29)
L
For small values of K the problem might become ill-posed again, while for large values of
K the boundary looses its non-reflective character. This behavior also seems frequency depen-
dent as at lower frequencies the boundary treatment causes reflectivity, as will be looked at in
section 3.2.1. Which will also deal with optimal values for coupling parameter σ, and other
performance issues. Section 3.2.2 will look at the performance of the nonreflective boundary in
ANSYS CFX. The last section ?? will look at an extension to the LRM model which in theory
is fully non-reflective while maintaining a well-posed problem.

3.2.1 Performance of the LRM method

For the LRM method, Poinsot showed using numerical calculations that one-dimensional waves
exit the domain (thus normal incidence) with very small reflection in the order of R = 10−6 [4].
A vortex in a parallel flow field exits the domain without much reflection, albeit more than the
one-dimensional wave, in the order of R = 10−4 .

Two aspects of the LRM model are looked at by Selle in more detail [5]:

ˆ The effect of the LRM model on acoustic behavior.

ˆ Optimal choices for the proportionality constant K.

The value of K determines the degree to which the far–field pressure p∞ is enforced. A small
value allows for a large pressure difference and reflects only a small fraction, but is not very
able to control the mean pressure. A large value of K enforces the difference between p and
p∞ to be small, but reflects a larger fraction. Thus, the value of K has a direct relation to the
reflection coefficient R.
pref l
R= R∈C (3.30)
pout
The incoming wave L5 is assumed to be harmonic, and is expressed as:

L5 = 2ρcûiωe−iωt (3.31)

And the reflected wave is equal to (3.28), the LRM model. Substituting the two above
expressions in the second and third LODI relations (3.26) leads to the following system of
equations:

17
 ∂u 1
+ (2ρcûiωe−iωt − K(p − p∞ )) = 0 (3.32)
∂t 2ρc
∂p 1
+ (2ρcûiωe−iωt + K(p − p∞ )) = 0 (3.33)
∂t 2
The second equation depends only on P and can be solved analytically:
Kt ρcûiω −iωt
p(t) = p∞ + A0 e− 2 − K
e (3.34)
2 − iω

Looking at the limit behavior when the transient term vanishes for t → ∞ and K > 0,
which results in:
ρcûiω −iωt
p(t) = p∞ − K e (3.35)
2 − iω
Subsituting the relations for L1 and L5 into the reflection coefficient:

pref l K(p − p∞ )
R= = (3.36)
pout 2ρcûiωe−iωt

The expression for p (3.35) is substituted in (3.36):

ρcûiω −iωt
K(P∞ − K
−iω
e − P∞ ) −K −1
2
R= = = (3.37)
2ρcûiωe−iωt K − 2iω 1 − 2iω
K

Because R is a complex number it can be expressed in a magnitude kRk and a phase angle
φ.
1
kRk = q (3.38)
1 + ( 2ω
K)
2

2ω
φ = −π − arctan( ) (3.39)
K
To illustrate how the reflection coefficient R evolves through the complex plane as a function
of relaxation factor K a plot is made 3.1.

18
 Figure 3.1: Reflection coefficient R in the complex plane for varying K

Where A is defined as:

2ω
A= (3.40)
K
A increases proportional with ω, and inversely proportional with K.
Limit behavior of the reflection for different values of relaxation factor K are shown in table
3.2.1.

Table 3.2: Limit behavior of reflection coefficient

A kRk φ
A=0 1 −π
A=1 √1 −π − π4
2
A→∞ 0 − 3π
2

For high frequencies (and low K) the reflection factor R tends to zero, and the boundary
condition can be called non-reflecting (see fig. 3.2). For low frequencies (and high K), however,
the reflection coefficient R tends to 1, and the phase to −π, basically acting as a pressure release
surface. Selle defines a cut-off frequency at A = 1, below which most of the acoustic energy is
reflected back into the domain [5].

ωc K
fc = = (3.41)
2π 4π

19
 In practice different cut-off frequencies may be defined depending on problem characteristics.

Figure 3.2: Modulus and phase of reflection coefficient as a function of frequency

Scaling strategy for K A scaling strategy for the relaxation factor K is proposed [5]. Good
values of K should allow all duct modes to leave the domain. Using the cut-off frequency, which
defines the reflected from the non-reflected part, lower than the lowest eigenvalue. The lowest
eigen-mode is the quarter-wave mode, which translates to the frequency domain via:
c
f0 = (1 − M2 ) (3.42)
4L

1
Using eq. (3.41) and eq. (3.42), the highest value of K which still significantly damps the 4
wave mode can be determined:
c
K max = π(1 − M2 ) (3.43)
L
When comparing eq. 3.43 to eq. 3.29 it can be seen that σ is bounded above by π. The
lower bound is not fixed by acoustics, but by other parameters such as the Reynolds number
and geometry. A value for σ lower than 0.1 often results in increased convergence times or no
convergence at all due to too much pressure drift. This results in an adequate range of values
for σ as shown in figure 3.3.

Optimal values for σ from literature range from 0.27 based on theory to 0.58 based on nu-
merical simulations [5]. One only needs to verify if the frequency of interest (e.g. excitation
frequency) is higher than the quarter-wave length frequency (see eq. 3.44). This is because eq.
(3.29) is based on this frequency and the cut-off frequency. Important values for σ are summed
up in table 3.2.1.

c0
fquarter = (3.44)
4·L
Using numerical analysis two values for sigma were tested. One value is way to high,
with σ = 10π. At this value the first eigenmodes are unable to leave the domain, and remain
asymptotically as a standing wave. The second value is equal to the maximum value σ = π. The

20
 Figure 3.3: Optimal values for sigma

σ min ≈ 0.1 σ max = π σ opt1 = 0.27 σ opt2 = 0.58

Table 3.3: Important sigma values

standing waves within the tubes die out rapidly, although the eigenfrequencies are significantly
shifted towards a lower frequency.

3.2.2 Performance of the non-reflective BC in Ansys CFX

A study to compare the perfomance predicted by theory with the (beta) implementation by
ANSYS CFX is investigated in [3]. Deviations are stumbled upon, and possible explanations
are sought. The non-reflecting boundary condition in ANSYS CFX is also based on the LODI-
relations and the LRM model for the reflected wave. Using a long tube with acoustically hard
cylinder walls, a prescribed harmonically varying velocity inlet, and a nonreflecting outlet, the
performance of the non-reflecting boundary condition is evaluated.
The numerically determined values for the reflection coefficient agrees with theory for the
higher K values. For the lower K values, however, the values do not match. Two possible reasons
are proposed. The first is loss in accuracy due to the numerical scheme and discretization. The
second is that a build-in mechanism in Ansys is preventing drift of the mean pressure which
results with low values of K, and thus preventing convergence problems.

21
Figure 3.4: Numerically determined R (dots) vs. analytical model (line)

22
Chapter 4

Numerical modeling

In this chapter workflow in using the developed boundary conditions and problems encountered
during programming in FORTRAN for ANSYS CFX are discussed.

4.1 General workflow

General workflow for setting up a simulation to be run with the developed boundary conditions:

1. Create a mesh with a flat boundary perpendicular to the dominant direction of propaga-
tion.

2. Optional (for PWM and TD-IBC): Create a sample plane within the mesh in a mesh
creation package (see next section).

3. Call the boundary to be used as a NRBC or TD-IBC with the pre-fix LRM*.

4. Create a CCL file with the necessary options (see example CCL files in appendix), and
add to ANSYS CFX PRE file.

5. Compile the *.f files using the built-in Command Editor, and write: !system(”cfx5mkext
double NAMA.f”) or die;. Then process (assuming files are in running directory).

6. Add the junction box and user cell routines in ANSYS CFX PRE.

7. Call the pset routine from user cell in the pressure specification field of the boundary with
the correct arguments and pressure as return dimension.

8. Specify the junction box routines in the Solver Control dialogue box.

9. Run the simulation.

23
4.2 Sample plane creation and use
The creation of a sample plane from which the area averaged pressure and velocity are obtained
by the solver is not as straightforward as one might think. Monitor points cannot be used
to retrieve the aforementioned quantities as these are apparently retrieved directly from the
solution that is written to the disk, which can be seen as a sort-of real-time post-processing.
The utility routine user get gvar can retrieve and calculate area averaged quantities like pres-
sure and velocity, but need a user named location within the mesh. These locations include
boundary surfaces. But it is not possible to define a surface within the domain using ANSYS
CFX PRE only, as these are only recognized as unnamed internal element surfaces. To create
a named surface at the location where the sample plane needs to be, mesh modeling software
needs to be used. Follow the step-by-step process listed below to implement the sample plane in
question. The process is written for mesh software ICEM, and assumes the geometry is already
fully designed and opened. Also, the naming is important and should be copied exactly as they
are hard-coded in the Fortran subroutines.

1. Create a surface within the geometry, a certain distance from and parallel to the boundary
in question.

2. Create a part from the surface and call it SAMPLEP.

3. Create two blocks which together enclose the geometry, and of which the separating face
is aligned and associated to the SAMPLEP surface.

4. Create the mesh and save.

5. Import the mesh in ANSYS CFX PRE, and check if the SAMPLEP 1 and
SAMPLEP 2 principle 2D regions are shown under the mesh item.

6. Create a 1:1 conservative general connection interface between SAMPLEP 1 and

SAMPLEP 2, and call it sampleplane.

7. Safe, close, and re-open the ANSYS CFX-PRE file (or else the automatically created
boundaries do not show up).

8. Now two boundaries should appear, call them sampleplane Side 1 and
sampleplane Side 2.

The sampleplane Side 1 boundary is used by the Fortran subroutines to retrieve the area
averaged pressure and velocity, and is hard-coded.

24
4.3 Problems encountered in programming for ANSYS CFX
Several problems and bugs were stumbled upon during modeling and programming for ANSYS
CFX. An overview of them is given in table 4.3, and is discussed in detail in the subsections.

Table 4.1: Problems encountered in programming for ANSYS CFX

No. Type Applies to Problem Status
1 CFX ucf template.F template Gives errors S
2 CFX Message to master process Stops solver inside slave partition N
3 CFX Single precision numbers Accumulation of error W
4 CFX Pressure gradient Bug in user getvar routine W
5 CFX Velocity gradient Bug in user getvar routine W
6 CFX Utility routines CACTION argument can cause crash S
7 CFX User parameters user peek* fails for big lists W
8 CFX Sound speed Wrong sound speed in isothermal simulations W
9 Doc Array shape user getvar returns wrong array shape S
10 Fort iFort compiler iFort compiler for Windows cause problems W
11 Fort Other compilers Give problems, flags are difficult W
12 Linx X-Connection X-Connection in Windows is slow/buggy W
13 Linx Closing SSH connection Closing connection causes solver termination S

In the Type column CFX means ANSYS CFX specific, Doc means ANSYS documentation,
Fort means Fortran specific, and Linx means Linux specific. In the Status column S means
Solved, W means no solution is found but a workaround method can be used to overcome the
problem, and N means No solution.

ucf template.F gives errors

The template file ucf template.F for user defined CEL functions causes a crash when no utility
function is called within the subroutine that uses the argument CRESLT, and the CRESLT
argument is set to GOOD at the end of the subroutine. The reason is of this appearant bug is
not known, but is easily solved by erasing (or make a comment of) the line of code in question.
However, remember to un-comment the line when utility routines are used in the code, so the
solver knows the subroutine did not encounter any errors. This also applies to a subroutine
called TStat Control.F supplied in an ANSYS CFX tutorial. The tutorial can be found under
help in ANSYS CFX-PRE by the name Chapter 19: Air Conditioning Simulation.

Message to master process

This bug only applies to partitioned solver runs. Printing a message to the solver output window
with the command MESAGE(’WRITE’, <text >) in a DO-LOOP when the subroutine is called
by a slave process (often the case in user defined CEL functions) will cause the solver to stop
outputting all messages to the solver output window. The source of this bug is unknown.

Single precision numbers accumulate error

If the pressure or velocity at a boundary is calculated via a user defined CEL function, and
the mathematical expressions used to calculate these quantities include numbers that vary
significantly in order of magnitude, the output may contain an error of such magnitude that
the solution becomes inaccurate and eventually unstable (error adds up). This is caused by the

25
fact that single precision numbers loose significant digits in mathematical operations involving
numbers which vary greatly in order of magnitude.
The solution (at least for the code used in this master thesis) is to use double precision
(8 byte) numbers instead of single precision (4 byte) numbers. Double precision numbers in
Fortran have a significand of 52 bit while a single precision number has a significand of 23 bit.

Error in retrieving pressure gradient

A utility routine called user getvar can retrieve the gradient of different quantities at a certain
location. According to the documentation the variable description argument <locale >.Pres-
sure.Gradient can be used to retrieve the pressure gradient at the given locale, but results in
a crash of the solver when used. According to the error description the function MAKDIR
called within the user getvar utility function cannot create a directory because it has an illegal
character in it. The illegal character is a space. No space is present in the name of the variable
description argument. This cannot be solved as it is a bug in the ANSYS CFX software.
However, when one uses the variable description argument Pressure.Gradient instead of
the aforementioned, the correct gradients are retrieved without problems. Another method of
retrieving gradients is by using additional variables, these can be created in ANSYS CFX PRE.
First create an additional variable, call it Px for example. Go to Body domain, and then to tab
Fluid Models. Go to additional variable and set option to Algebraic Equation, then fill in the
Add. Var. Value box: p.Gradient x. Now the pressure gradient variable Px can be accessed
via the user getvar routine via the variable description argument Px. The pressure gradient
value is calculated using the latest solution field.

Error in retrieving velocity gradient

Basically the same problem as for retrieving the pressure gradient using the user getvar utility
routine. But now the variable description argument Velocity.Gradient causes the same type of
crash, while the variable description argument <locale >.Velocity.Gradient functions correctly.
Also the velocity gradient, like the pressure gradient, can be retrieved using the additional
variable method.

Utility routines crash

Certain utlitity routines take a CACTION argument. One retrieves data by setting this argu-
ment to RETURN. However, if the memory area occupied by the data is not released by the end
of the subroutine the solver crashes because the memory area cannot be accessed. To overcome
this issue the utility routine call for retrieval should always be paired with a RELEASE value
as argument afterwards. This is not a bug, but important to mention because this is easily
overseen.

User parameter retrieval fails for big lists

Large lists of values (comma separated) as user parameters in the loaded CCL files can cause
errors when they are retrieved using the user peek* function when the list is too long. This
is probably because the solver loads the whole list as a character string of finite length before
parsing the data to an array. This bug can be overcome by using separate variables for each
item, and call them ’A1’, ’A2’, ... , ’AN’, and load them into an array (via a DO loop) at
initialization.

26
Wrong sound speed in isothermal simulations
When using isothermal simulations the sound speed returned by both monitor points√ as well
as returned by the utility function user getvar is wrong. Surprisingly this is exactly 1.4 too
large. The hypothesis is that the sound speed does not follow directly from √ the solution field,
but is calculated afterwards via equations for the sound speed such as γRT . The adiabatic
index γ used is then still 1.4, while it should be 1.0 for isothermal calculations. It can be worked
around by calculating the sound speed in the subroutines or CEL via the other field quantities
(pressure, temperature, etc).

Wrong array shape for user getvar

According to the documentation the shape of the returned array of scalar gradient should be
[N COM P T, N LOC] ([row, column]), where N COM P T represents the number of components
(cartesian coordinates in 3D, so 3 cells), and N LOC is the number of locations for which the gra-
dient is calculated. However, the format is actually [N LOC, N COM P T ]. The documentation
is wrong [7].

iFort compiler for Windows causes problems

The Intel iFort Fortran compiler would not work in Windows. The compiler also costs $ 1900 for
Windows, while it is free for Linux. Therefore it is recommended to use Linux as the platform
of choice. The drawback is that the Fortran code is compiled platform specific, and thus the
compiled code for Linux will not work on Windows PC’s.

Other compilers cause problems

Other compilers such as f77 and g77 would not work for standard settings in the cfx5mkext.ccl
file. The number of flags that can be set runs in the hundreds, and comes with little docu-
mentation and description. The standard flags are written for the iFort compiler. It might be
possible to use the other compilers with the correct flags, but these could not be found and
could be a cumbersome task to set by oneself.

X-Connection emulation in Windows is slow and buggy

To create a connection with the X Window system on a Linux pc in Windows, special software
on the Windows PC is needed. Two applications, Cygwin and Exceed, are tested. Both show
problems with graphical OpenGL viewports (which are used in ANSYS CFX ). This is not
solved. A workaround is to do the graphical part of pre-processing and post-processing on a
Windows PC, and transfer files between computers.

Closing SSH connection causes solver termination

When an ANSYS CFX solver process runs which outputs data to the terminal, and the terminal
window is closed, the connection is also closed and sends a SIGHUP signal to the process
automatically. The process can then decide to kill itself or continue. In an apparently random
way the process sometimes decides to kill itself, but usually continue. The way to work around
this problem is to start ANSYS CFX with nohup as a prefix from the terminal window (e.g.
nohup cfx5 ). This prevents HUP (hangup) signal to arrive at the process.

27
4.4 Single and double precision numbers
Single precision numbers use 32 bit of data, where 23 bits are used for the fraction part, 8 bits
for the exponent, and 1 bit for the sign (plus or minus). 23 bits for the fraction results in 24 bits
of total precision due to a implicit leading bit with a value of 1 (which is assumed zero when
the exponent part is zero) (see fig. 4.1). Depending on the value of the decimal number to be
stored, single precision numbers give 6 to 9 significant decimal digits precision. The decimal
value can be calculated from the stored bits using eq. 4.1. Where e is the decimal value of the
exponent. Not all decimal fractions can be exactly represented in binary, for example 0.1, but
0.125 can (2−3 ).

Figure 4.1: Single precision in memory [9]

−i
value = (−1)sign (1 + Σ23
i=1 b23−i 2 ) · 2
e−127
(4.1)
Double precision numbers use 11 bits instead of 8 for the exponent, and 52 bits instead of
23 bits for the significand. Off course also here is an implicit bit used, so 24 bits effectively.
Double precision gives 15 to 17 significant decimal digits precision, which is a lot more precise
than 6 to 9 significant digits with single precision.

Figure 4.2: Double precision in memory [8]

28
Chapter 5

Simulation data

In this section all results from simulations (test cases) are shown. Table 5 shows all simulation
parameters for both isothermal and total energy simulations.

Table 5.1: Parameters used in test cases

Parameter Isothermal Total energy
Simulation time 0.2s
Characteristic length 4m
Temperature 25◦ C (fixed) 25◦ C (initial)
Frequency 100Hz
Characteristic sound speed 346.13m/s 292.53m/s
Sampled frequency range 0.015 − 0.2s

The sampled frequency range parameter signifies the part of the time-series pressure data
that is used in the multi-microphone method to determine the reflection. The first part is dis-
carded because the wave has not arrived at the boundary yet, and mainly contributes to the
first coefficient of the Fourier transform associated with a frequency of zero, which is constant
(mean pressure over range).

Table 5.2 and 5.3 show how the different simulations are grouped and in which table they
can be found for the nonreflective and TD-IBC simulations respectively.

Table 5.2: Non-reflective test cases

# Group No. Cases Table
1 ANSYS CFX NRBC (Total energy) 16 5.4
2 ANSYS CFX NRBC (Isothermal) 16 5.5
3 Custom NRBC (Total energy) 16 5.6
4 Custom NRBC (Isothermal) 16 5.7
5 Custom NRBC+PWM (Total energy) 16 5.8
6 Custom NRBC+PWM (Isothermal) 16 5.9
7 Time integration schemes, custom NRBC (Total energy) 5 5.10
8 Heat transfer equations, ANSYS CFX NRBC 6 5.11
9 Heat transfer equations, custom NRBC 6 5.12
10 Influence of geometry 9 5.13
11 Influence of partitioning 9 5.14

29
 Table 5.3: TD-IBC test cases
# Group No. Cases Table
1 TD-IBC (Total energy) 4 5.22
2 TD-IBC (Isothermal) 4 ??
3 ANSYS CFX NRBC, hard wall/pressure release (Total energy) 4 5.24

Table 5.4: (1) ANSYS CFX NRBC (Total energy) test cases
Parameters Reflection
# K σ u0 R I |R| φ Cpu
1.1 1 0.0115 0 -0,0011 -0,0005 0,0012 -2,7486 3:13
1.2 10 0.1152 0 -0,0014 -0,0075 0,0077 -1,7536 3:16
1.3 23 0.265 0 -0,0020 -0,0177 0,0178 -1,6829 3:15
1.4 50 0.576 0 -0,0039 -0,0388 0,0390 -1,6713 3:16
1.5 100 1.152 0 -0,0098 -0,0772 0,0778 -1,6964 3:13
1.6 200 2.304 0 -0,0299 -0,1508 0,1537 -1,7668 3:12
1.7 500 5.76 0 -0,1412 -0,3278 0,3569 -1,9775 3:12
1.8 1000 11.52 0 -0,3748 -0,4600 0,5933 -2,2545 3:14
1.9 10000 115.2 0 -0,9493 -0,1228 0,9572 -3,0130 3:14
1.10 100000 1152 0 -0,9719 -0,0480 0,9731 -3,0923 3:15
1.11 23 0.265 0.5 -0,0028 -0,0185 0,0187 -1,7205 3:09
1.12 23 0.265 2.5 -0,0062 -0,0214 0,0223 -1,8535 3:36
1.13 23 0.265 5 -0,0102 -0,0252 0,0272 -1,9540 3:48
1.14 50 0.576 0.5 -0,0047 -0,0396 0,0399 -1,6890 3:12
1.15 50 0.576 2.5 -0,0081 -0,0424 0,0432 -1,7590 3:34
1.16 50 0.576 5 -0,0120 -0,0463 0,0478 -1,8245 3:45

30
 Table 5.5: (2) ANSYS CFX NRBC (Isothermal) test cases
Parameters Reflection
# K σ u0 R I |R| φ Cpu
2.1 1 0.0115 0 0,0813 -0,0068 0,0816 -0,0833 11:34
2.2 10 0.1152 0 0,0811 -0,0137 0,0822 -0,1679 11:35
2.3 23 0.265 0 0,0806 -0,0238 0,0840 -0,2869 11:36
2.4 50 0.576 0 0,0790 -0,0446 0,0907 -0,5142 11:32
2.5 100 1.152 0 0,0737 -0,0828 0,1109 -0,8434 11:29
2.6 200 2.304 0 0,0549 -0,1571 0,1665 -1,2347 8:24
2.7 500 5.76 0 -0,0556 -0,3436 0,3481 -1,7313 8:18
2.8 1000 11.52 0 -0,3093 -0,4971 0,5854 -2,1274 8:23
2.9 10000 115.2 0 -0,9898 -0,0762 0,9927 -3,0648 8:30
2.10 100000 1152 0 -0,9656 0,0658 0,9678 3,0735 8:27
2.11 23 0.265 0.5 0,0798 -0,0236 0,0832 -0,2875 9:50
2.12 23 0.265 2.5 0,0766 -0,2240 0,0798 -0,2839 9:58
2.13 23 0.265 5 0,0731 -0,0211 0,0761 -0,2812 9:54
2.14 50 0.576 0.5 0,0781 -0,0444 0,0899 -0,5168 9:55
2.15 50 0.576 2.5 0,0750 -0,0431 0,0865 -0,5215 9:47
2.16 50 0.576 5 0,0715 -0,0418 0,0828 -0,5291 9:53

Table 5.6: (3) Custom NRBC (Total energy) test cases

Parameters Reflection P. drift
# K u0 R I |R| φ Cpu Max. Sta.
3.1 1 0 -0,01150 0,00640 0,01310 2,63420 3:37 CBD CBD
3.2 10 0 -0,01140 -0,00160 0,01156 -3,00068 3:36 193.5 193.5
3.3 23 0 -0,01180 -0,01130 0,01630 -2,37627 3:36 120 CBD
3.4 50 0 -0,01270 -0,02910 0,03177 -1,98289 3:35 85 20
3.5 100 0 -0,01590 -0,06120 0,06320 -1,82530 3:11 70 32.5
3.6 200 0 -0,02850 -0,12360 0,12680 -1,79740 3:08 58.5 35
3.7 500 0 -0,10530 -0,28350 0,30240 -1,92630 3:12 36 36
3.8 1000 0 -0,29400 -0,43580 0,52570 -2,16430 3:12
3.9 10000 0 -0,94820 -0,13870 0,95830 -2,99640 3:15
3.10 100000 0 -0,97140 -0,04980 0,97270 -3,09040 3:15
3.11 23 0.5 -0,00990 0,01190 0,01550 2,26500 3:06 151 151
3.12 23 2.5 0,04920 0,07420 0,08900 0,98540 3:35 81 81
3.13 23 5 0,11420 0,06390 0,13090 0,51030 3:41 37 37
3.14 50 0.5 -0,01060 -0,00610 0,01220 -2,61420 3:07 96 85
3.15 50 2.5 0,05120 0,05410 0,07450 0,81230 3:35 56 40
3.16 50 5 0,11610 0,04110 0,12310 0,34000 3:39 45 18

31
 Table 5.7: (4) Custom NRBC (Isothermal) test cases
Parameters Reflection P. drift
# K u0 R I |R| φ Cpu Max. Sta.
4.1 1 0 -0,0126 -0,0033 0,0131 -2,8859 7:18 6 6
4.2 10 0 -0,0127 -0,0092 0,0157 -2,5163 11:44 29.5 29.5
4.3 23 0 -0,0131 -0,0176 0,0219 -2,2124 11:49 33.5 33.5
4.4 50 0 -0,0140 -0,0349 0,0376 -1,9520 11:49 26.5 26.5
4.5 100 0 -0,0178 -0,0669 0,0692 -1,8307 11:42 29 29
4.6 200 0 -0,0315 -0,1291 0,1329 -1,8099 11:33 24 20
4.7 500 0 -0,1162 -0,2909 0,3133 -1,9507 11:36 12.5 12.5
4.8 1000 0 -0,3257 -0,4407 0,5480 -2,2072 11:38
4.9 10000 0 -0,9895 -0,0767 0,9925 -3,0642 11:29
4.10 100000 0 -0,9657 0,0658 0,9680 3,0735 11:34
4.11 23 0.5 -0,0111 0,0062 0,0127 2,6294 10:06 167 167
4.12 23 2.5 0,0473 0,0719 0,0861 0,9888 9:52 100 44
4.13 23 5 0,1126 0,0686 0,1318 0,5472 9:56 48 48
4.14 50 0.5 -0,0111 -0,0110 0,0156 -2,3584 10:07 88 66
4.15 50 2.5 0,0499 0,0529 0,0728 0,8145 9:59 53 41
4.16 50 5 0,1151 0,0465 0,1242 0,3836 10:00 26 18

Table 5.8: (5) Custom NRBC+PWM (Total energy) test cases

Parameters Reflection P. drift
# K u0 R I |R| φ Cpu Max. Sta.
5.1 1 0 -0,0115000 0,0070000 0,0135000 2,5928000 3:31 CBD CBD
5.2 10 0 -0,0114000 0,0051000 0,0125000 2,7221000 3:35 206 206
5.3 23 0 -0,0115000 0,0040000 0,0122000 2,8051000 3:37 100 CBD
5.4 50 0 -0,0115000 0,0036000 0,0121000 2,8353000 3:38 53 CBD
5.5 100 0 -0,0114000 0,0039000 0,0121000 2,8110000 3:38 29 0
5.6 200 0 -0,0109000 0,0046000 0,0119000 2,7448000 3:37 14.5 0
5.7 500 0 -0,0091000 0,0060000 0,0109000 2,5591000 3:37 6.5 0
5.8 1000 0 -0,0059000 0,0066000 0,0089000 2,3003000 3:32
5.9 10000 0 -0,0002675 -0,0007301 0,0007759 -1,9220000 3:32
5.10 100000 0 -0,0008000 -0,0023000 0,0024111 -1,8982062 3:36
5.11 23 0.5 -0,0095000 0,0275000 0,0291000 1,9024000 3:11 203 CBD
5.12 23 2.5 0,0486000 0,0905000 0,1027000 1,0783000 3:36 135 CBD
5.13 23 5 0,1135000 0,0811000 0,1395000 0,6205000 3:45 64 CBD
5.14 50 0.5 -0,0089000 0,0269000 0,0284000 1,8902000 3:13 103 CBD
5.15 50 2.5 0,0502000 0,0895000 0,1026000 1,0593000 3:38 68 68
5.16 50 5 0,1153000 0,0788000 0,1396000 0,5992000 3:43 32 32

32
 Table 5.9: (6) Custom NRBC+PWM (Isothermal) test cases
Parameters Reflection P. drift
# K u0 R I |R| φ Cpu Max. Sta.
6.1 1 0 -0,0126 -0,0026 0,0129 -2,9362 11:38 4 0
6.2 10 0 -0,0126 -0,0026 0,0128 -2,9411 11:47 4 0
6.3 23 0 -0,0125 -0,0025 0,0128 -2,9482 11:53 3 0
6.4 50 0 -0,0124 -0,0023 0,0126 -2,9627 11:54 3 0
6.5 100 0 -0,0123 -0,0019 0,0124 -2,9888 11:50 3 0
6.6 200 0 -0,0117 -0,0012 0,0118 -3,0370 11:49 3 0
6.7 500 0 -0,0099 0,0003 0,0099 3,1161 11:47 3 0
6.8 1000 0 -0,0066 0,0009 0,0067 3,0009 7:18
6.9 10000 0 -0,0006 -0,0061 0,0061 -1,6744 7:20
6.10 100000 0 -0,0011 -0,0077 0,0078 -1,7091 7:21
6.11 23 0.5 -0,0112 0,0210 0,0238 2,0583 10:14 177 CBD
6.12 23 2.5 0,0462 0,0870 0,0985 1,0822 10:07 165 CBD
6.13 23 5 0,1113 0,0848 0,1399 0,6515 9:57 80 80
6.14 50 0.5 -0,0098 0,0215 0,0236 2,0007 10:09 95 95
6.15 50 2.5 0,0481 0,0869 0,0993 1,0650 10:04 85 85
6.16 50 5 0,1132 0,0831 0,1404 0,6333 9:56 40 40

Table 5.10: (7) Time integration schemes (Total energy)

Params. Reflection
# K u0 R I |R| φ Description
7.1 50 0 -0,0128 -0,0316 0,0341 -1,9559 First order, explicit L
7.2 50 0 -0,0127 -0,0291 0,0318 -1,9829 Second order, explicit L
7.3 50 0 -0,0127 -0,0291 0,0318 -1,9827 Third order, explicit L
7.4 50 0 -0,0131 -0,0367 0,0390 -1,9132 First order, implicit L
7.5 50 0 -0,0130 -0,0342 0,0366 -1,9333 First order, implicit predictor L

Table 5.11: (8) Influence of heat transfer options, ANSYS CFX NRBC
Params. Reflection Heat eqs. options
# K σ u0 R I |R| φ Heat equation PT VT Cpu
8.1 50 0.576 0 0,1765 0,1443 0,2280 0,6854 Thermal Off Off 9:27
8.2 50 0.576 0 0,1765 0,1443 0,2280 0,6854 Thermal Off On 9:29
8.3 50 0.576 0 -0,0038 -0,0376 0,0378 -1,6725 Thermal On Off 4:27
8.4 50 0.576 0 -0,0038 -0,0376 0,0378 -1,6725 Thermal On On 4:25
8.5 50 0.576 0 -0,0039 -0,0388 0,0390 -1,6713 Total N/A Off 3:26
8.6 50 0.576 0 -0,0039 -0,0388 0,0390 -1,6713 Total N/A On 3:25

Table 5.12: (9) Influence of heat transfer options, Custom NRBC

Params. Reflection Heat eqs. options
# K σ u0 R I |R| φ Heat equation PT VT Cpu
9.1 50 0.576 0 0,2491 0,2498 0,3528 0,7868 Thermal Off Off 8:06
9.2 50 0.576 0 0,2491 0,2498 0,3528 0,7868 Thermal Off On 8:07
9.3 50 0.576 0 -0,0122 -0,0294 0,0319 -1,9646 Thermal On Off 4:30
9.4 50 0.576 0 -0,0122 -0,0294 0,0319 -1,9646 Thermal On On 4:33
9.5 50 0.576 0 -0,0127 -0,0291 0,0318 -1,9829 Total N/A Off 3:21
9.6 50 0.576 0 -0,0127 -0,0291 0,0318 -1,9829 Total N/A On 3:25

33
 Table 5.13: (10) Influence of geometry (Total energy)
Params. Reflection
# K u0 R I |R| φ Boundary Description
10.1 50 0 -0,0037 -0,0382 0,0383 -1,6671 ANSYS NRBC Mesh refined near BC
10.2 50 0 ? ? ? ? ANSYS NRBC 2D Mesh
10.3 50 0 ? ? ? ? ANSYS NRBC 3D Mesh
10.4 50 0 -0,0123 -0,0166 0,0207 -2,2064 Custom NRBC Mesh refined near BC
10.5 50 0 ? ? ? ? Custom NRBC 2D Mesh
10.6 50 0 ? ? ? ? Custom NRBC 3D Mesh
10.7 50 0 -0,0143 0,0166 0,0219 2,2815 Custom NRBC+PWM Mesh refined near BC
10.8 50 0 ? ? ? ? Custom NRBC+PWM 2D Mesh
10.9 50 0 ? ? ? ? Custom NRBC+PWM 3D Mesh

Table 5.14: (11) Influence of partitioning (Total energy)

Params. Reflection
# K u0 R I |R| φ Cpu Boundary Description
11.1 50 0 -0,0039 -0,0388 0,0390 -1,6713 3:16 ANSYS NRBC Serial
11.2 50 0 -0,0039 -0,0388 0,0390 -1,6713 2:02 ANSYS NRBC 2 Partitions
11.3 50 0 -0,0039 -0,0388 0,0390 -1,6713 2:23 ANSYS NRBC 4 Partitions
11.4 50 0 -0,0127 -0,0291 0,0318 -1,9829 3:12 Custom NRBC Serial
11.5 50 0 -0,0127 -0,0291 0,0318 -1,9829 2:00 Custom NRBC 2 Partitions
11.6 50 0 -0,0127 -0,0291 0,0318 -1,9829 2:19 Custom NRBC 4 Partitions
11.7 50 0 -0,0115 0,0036 0,0121 2,8353 3:16 Custom NRBC+PWM Serial
11.8 50 0 -0,0115 0,0036 0,0121 2,8353 2:03 Custom NRBC+PWM 2 Partitions
11.9 50 0 -0,0115 0,0036 0,0121 2,8353 2:21 Custom NRBC+PWM 4 Partitions

Table 5.15: (12) Linear refinement for ANSYS CFX NRBC

Parameters Reflection
# K σ u0 R I |R| φ
12.1 1 0.0115 0 -0,00090273 0,00017813 0,00092014 2,9468
12.1 10 0.1152 0 -0,00120000 -0,00690000 0,00700000 -1,7420
12.1 23 0.265 0 -0,00180000 -0,01710000 0,01720000 -1,6750
12.1 50 0.576 0 -0,00370000 -0,03820000 0,03830000 -1,6671
12.1 100 1.152 0 -0,00950000 -0,07660000 0,07720000 -1,6943
12.1 200 2.304 0 -0,02960000 -0,15020000 0,15310000 -1,7656
12.1 500 5.76 0 -0,14080000 -0,32740000 0,35640000 -1,9770
12.1 1000 11.52 0 -0,37450000 -0,45980000 0,59310000 -2,2543
12.1 10000 115.2 0 -0,94940000 -0,12220000 0,95720000 -3,0136
12.1 100000 1152 0 -0,97190000 -0,04740000 0,97310000 -3,0929

34
 Table 5.16: (13) Linear refinement for custom NRBC
Parameters Reflection P. drift
# K u0 R I |R| φ Max. Sta.
13.1 1 0 -0,0180 0,0133 0,0224 2,5033 220 CBD
13.2 10 0 -0,0152 0,0091 0,0177 2,6041 98 98
13.3 23 0 -0,0136 0,0010 0,0136 3,0655 51 39
13.4 50 0 -0,0123 -0,0166 0,0207 -2,2064 29 12
13.5 100 0 -0,0119 -0,0491 0,0505 -1,8085 34 res
13.6 200 0 -0,0175 -0,1122 0,1135 -1,7258 23 res
13.7 500 0 -0,0768 -0,2800 0,2903 -1,8387 37 res
13.8 1000 0 -0,2673 -0,4585 0,5307 -2,0986 51 res
13.9 10000 0 -0,9498 -0,1390 0,9599 -2,9963
13.10 100000 0 -0,9714 -0,0492 0,9727 -3,0910

Table 5.17: (14) Linear refinement for custom NRBC+PWM

Parameters Reflection P. drift
# K u0 R I |R| φ Max. Sta.
14.1 1 0 -0,0182 0,1390 0,0229 2,4902 230 CBD
14.2 10 0 -0,0167 0,0150 0,0228 2,4100 155 155
14.3 23 0 -0,0154 0,0159 0,0221 2,3380 99 99
14.4 50 0 -0,0143 0,0166 0,0219 2,2815 59 53
14.5 100 0 -0,0132 0,0171 0,0216 2,2284 29 29
14.6 200 0 -0,0115 0,0177 0,0212 2,1475 19 16
14.7 500 0 -0,0070 0,0183 0,0196 1,9367 5 5
14.8 1000 0 -0,0009 0,0161 0,0161 1,6253 1 1
14.9 10000 0 0,0016 -0,0020 0,0026 -0,9086
14.10 100000 0 -0,0007 -0,0039 0,0039 -1,7500

Table 5.18: (15) Smooth sine for ANSYS CFX NRBC

Parameters Reflection
# K σ u0 R I |R| φ
15.1 1 0.0115 0 -0,00110000 -0,00050000 0,00120000 -2,7479
15.1 10 0.1152 0 -0,00140000 -0,00750000 0,00760000 -1,7493
15.1 23 0.265 0 -0,00190000 -0,01770000 0,01780000 -1,6786
15.1 50 0.576 0 -0,00370000 -0,03870000 0,03890000 -1,6670
15.1 100 1.152 0 -0,00940000 -0,07710000 0,07770000 -1,6923
15.1 200 2.304 0 -0,02930000 -0,15060000 0,15340000 -1,7630
15.1 500 5.76 0 -0,14010000 -0,32780000 0,35650000 -1,9746
15.1 1000 11.52 0 -0,37340000 -0,46010000 0,59260000 -2,2525
15.1 10000 115.2 0 -0,94590000 -0,12130000 0,95360000 -3,0140
15.1 100000 1152 0 -0,96650000 -0,04970000 0,96780000 -3,0902

35
 Table 5.19: (16) Smooth sine for custom NRBC
Parameters Reflection P. drift
# K u0 R I |R| φ Max. Sta.
16.1 1 0 -0,0115 0,0063 0,013112 2,6437 414 CBD
16.2 10 0 -0,0114 -0,0017 0,011566 -2,9953 208 208
16.3 23 0 -0,0117 -0,0113 0,016288 -2,3742 112 59
16.4 50 0 -0,0126 -0,0291 0,031701 -1,9793 75 22
16.5 100 0 -0,0157 -0,0611 0,063090 -1,8217 70 34
16.6 200 0 -0,0280 -0,1235 0,126582 -1,7936 60 38
16.7 500 0 -0,1042 -0,2835 0,302091 -1,9231 51 38
16.8 1000 0 -0,2927 -0,4359 0,525088 -2,1622 70 43
16.9 10000 0 -0,9456 -0,1369 0,955453 -2,9979
16.10 100000 0 -0,9660 -0,0514 0,967383 -3,0884

Table 5.20: (17) Smooth sine for custom NRBC+PWM

Parameters Reflection P. drift
# K u0 R I |R| φ Max. Sta.
17.1 1 0 -0,0115 0,0069 0,013429 2,6031 403 CBD
17.2 10 0 -0,0114 0,0050 0,012460 2,7299 204 CBD
17.3 23 0 -0,0115 0,0040 0,012192 2,8108 103 80
17.4 50 0 -0,0116 0,0036 0,012115 2,8401 55 13
17.5 100 0 -0,0114 0,0039 0,012076 2,8157 35 0
17.6 200 0 -0,0110 0,0045 0,011873 2,7496 19 0
17.7 500 0 -0,0092 0,0060 0,010921 2,5641 5 0
17.8 1000 0 -0,0060 0,0066 0,008886 2,3055 2 0
17.9 10000 0 -0,0003 -0,0007 0,000769 -1,9256
17.10 100000 0 -0,0008 -0,0023 0,002405 -1,8957

Table 5.21: (18) Special simulations

Params. Reflection P. drift
# K u0 R I |R| φ Max. Sta. Boundary
Mean flow convergence of isothermal simulations
18.1 50 25 0,04080000 -0,02900000 0,05000000 -0,6189 ANSYS NRBC
18.2 50 50 -0,00150000 -0,00470000 0,00490000 -1,8782 ANSYS NRBC
Very long simulation time to check convergence of drift. 1 second sim. time. Total E.
18.3 50 0 -0,00980000 -0,02930000 0,03090000 -1,8946 85 20 custom NRBC
Linearly refined mesh, but with ∆t = 2 ∗ 10−6 , Total E.
18.1 50 0 -0,00830 -0,01600 0,01800 -2,05010 27 15 custom NRBC

Table 5.22: (1) TD-IBC test cases (Total energy)

Params. Reflection
# K u0 R I |R| φ Description
1.1 50 0 1,0035 -0,0823 1,0069 -0,0818 Acoustically hard wall
1.2 50 0 -0,9669 0,0347 0,9675 3,1057 Pressure release surface
1.3 50 0 -0,0001 0,0100 0,0100 1,5841 Non-reflecting outlet
1.4 50 0 0,5219 0,4882 0,7146 0,7520 Complex R = 0.5 + 0.5i

36
 Table 5.23: (2) TD-IBC test cases (Isothermal)
Params. Reflection
# K u0 R I |R| φ Description
2.1 50 0 1,0095 0,0623 1,0114 0,0617 Acoustically hard wall
2.2 50 0 -0,8767 0,0459 0,8779 3,0893 Pressure release surface
2.3 50 0 -0,0005 0,0045 0,0045 1,6906 Non-reflecting outlet
2.4 50 0 0,4868 0,5022 0,6994 0,8010 Complex R = 0.5 + 0.5i

Table 5.24: (3) Ansys hard wall and constant pressure

Reflection
# u0 R I |R| φ Description
3.1 0 -0,9767 -0,0387 0,9775 -3,1020 Constant pressure (Total energy)
3.2 0 -0,9628 0,0802 0,9661 3,0585 Constant pressure (Isothermal)
3.3 0 0,9944 -0,1034 0,9998 -0,1036 Hard wall (Total energy)
3.4 0 1,0142 0,0439 1,0151 0,0433 Hard wall (Isothermal)

37
Chapter 6

Source code

The source code of the FORTRAN subroutines, the MATLAB scripts and classes, the CCL
files, and the used filter coefficient arrays for the construction of the TD-IBCs, are added to
the supplied CD-ROM. The UML diagram and two example CCL files are added to this chapter.

Supplied on CD-ROM:

ˆ PLinit FORTRAN subroutine source

ˆ PTloop FORTRAN subroutine source, NRBC version

ˆ PTloop FORTRAN subroutine source, TD-IBC version

ˆ Pset FORTRAN subroutine source, NRBC version

ˆ Pset FORTRAN subroutine source, TD-IBC version

ˆ MultiMicMethod MATLAB class, including example run script.

ˆ CCL files, including filter coefficient sets for TD-IBC.

ˆ ANSYS CFX PRE files

ˆ All P1 ... P4 pressure time series in *.csv file.

38
6.1 MultiMicMethod MATLAB class
The complete UML diagram. The source code is added to the supplied CD-ROM.

Figure 6.1: Complete UML diagram for MultiMicMethod class

39
6.2 CCL file with user parameters
The CCL file (with example settings) for both the NRBC as well as the TD-IBC.

CCL file example for NRBC

USER:
# ====== Model P r o p e r t i e s ======
# R e l a t i v e f a r − f i e l d p r e s s u r e [ Pa ]
pinf = 0.0
# Mean f l o w s p e e d [m/ s ]
u = 0.5
# Relaxation f ac t o r K [1/ s ]
K = 50.0
# D i s t a n c e o f sample p l a n e from Non−r e f l e c t i v e BC [m]
DXSAMPLE = 0 . 4
# Use Plane−Wave Masking ? ( 0 : No , 1 : Yes )
PWM = 1
# ====== F l u i d P r o p e r t i e s ======
# A d i a b a t i c i n d e x gamma
gamma = 1 . 4
# S p e c i f i c gas constant
RSPEC = 2 8 7 . 0 5 8
END
CCL file example for TD-IBC

USER:
# ====== Model P r o p e r t i e s ======
# R e l a t i v e f a r − f i e l d p r e s s u r e [ Pa ]
pinf = 0.0
# Mean f l o w s p e e d [m/ s ]
u = 0.0
# Relaxation f ac t o r K [1/ s ]
K = 50.0
# D i s t a n c e o f sample p l a n e from Non−r e f l e c t i v e BC [m]
DXSAMPLE = 0 . 4
# Use Plane−Wave Masking ? ( 0 : No , 1 : Yes )
PWM = 1
# ====== F l u i d P r o p e r t i e s ======
# A d i a b a t i c i n d e x gamma
gamma = 1 . 4
# S p e c i f i c gas constant
RSPEC = 2 8 7 . 0 5 8
# ====== TD−IBC f i l t e r c o n s t a n t s ======
# Number o f f i l t e r c o n s t a n t s
NFC = 10
# A f i l t e r constants
AFC = 0 , 0 , 0 , 0 , 0 , 1 , 0 , 0 , 0 , 0
# B f i l t e r constants
BFC = 1 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0
END

40
6.3 Filter coefficient sets
If not mentioned, the coefficient is equal to zero. A minimum of ten coefficients need to be
specified, and all coefficients need to be specified until the last non-zero coefficient (equal for
both A and B). If only coefficient # 1 is non-zero, the others need to be specified with zero.
If coefficient # 350 is non-zero, every coefficient to # 350 need to be specified, even if they are
zero. Coefficient B1 always needs to be equal to 1.

Acoustically hard wall:

A0=1.0000
B0=1.0000
Pressure release surface:
A499 =1.0000
B0=1.0000
Non-reflecting outlet:
B0=1.0000
Complex R = 0.5 + 0.5i at fixed frequency of f = 100Hz.
A124 =0.7071
B0=1.0000

41
Bibliography

[1] Seung-ho Jang and Jeong-guon Ih. On the multiple microphone method for measuring
in-duct acoustic properties in the presence of mean flow. 103(3):1520–1526, 1998.

[2] Michael G Jones and Tony L Parrot. Evaluation of a Multi-Point Method for Determining
Acoustic Impedance. Mechanical Systems and Signal Processing, 3:15–35, 1989.

[3] Joris Oosterhuis. Performance of non-reflective boundary condition in ANSYS CFX. Tech-
nical report, 2012.

[4] T.J. Poinsot. Boundary conditions for direct simulations of compressible viscous flows.
Journal of Computational Physics, 99(2):352, April 1992.

[5] Laurent Selle. The actual impedance of non-reflecting boundary conditions : implications
for the computation of resonators. pages 1–21, 2003.

[6] W Thompson. Time- Dependent Boundary Conditions for Hyperbolic Systems , II. Journal
of Computational Physics, 439461:439–461, 1990.

[7] Ansys Tutorials. ANSYS CFX-Solver Modeling Guide. 15317(November):724–746, 2011.

[8] Wikipedia. Double-precision floating-point format. 10(2):6–9, 2013.

[9] Wikipedia. Single-precision floating-point format. 24:4–9, 2013.

42
Nomenclature

BC Boundary Condition

DTFT Discrete-Time Fourier Transform

EOS Equation(s) Of State

LODI Locally One-Dimensional Inviscid

LRM Linear Relaxation Method

MMM Two-Microphone Method

NSCBC Navier-Stokes Characteristic Boundary Conditions

PDE Partial Differential Equation

SWR Standing Wave Ratio

SWR Standing Wave Ratio

TMM Two-Microphone Method

K Relaxation factor [1/s]

L Characteristic length (e.g. domain length) [m]

R Reflection coefficient [−]

Re Reynolds number

Γ Plane-wave propagation constants

Ψ Coefficient of friction for turbulent flow

δρ Acoustic density [kg/m3 ]

δp Acoustic pressure [P a]

δu Acoustic velocity [m/s]

δ Attenuation constant in Multi-Microphone Method

δt Attenuation due to turbulent effects

δν Attenuation due to viscothermal effects

γ Adiabatic index

p̂ Acoustic pressure phasor [P a]

43
p̂+ Forward traveling pressure phasor [P a]

p̂− Backward traveling pressure phasor [P a]

û Acoustic particle velocity phasor [m/s]

κ Heat conduction coefficient

λ Eigenvalues of A / Characteristic speed LODI-relations [m/s]

C Set of complex numbers

A+ Moor-Penrose psuedo-inverse of A

AH Hermitian matrix of A

F[] Fourier transform

Li Amplitude variation of characteristic waves [varies]

L[] Laplace transform

M Mean flow Mach number

R Resistive part of impedance [Rayl]

X Reactive part of impedance [Rayl]

Z Impedance [Rayl]

Z0 Characteristic impedance [Rayl]

µ Shear viscosity constant

ω Angular frequency

φ Velocity potential [m2/s]

ρ0 Mean density [kg/m3 ]

σ Coupling parameter [1/s]

τ Time coordinate [s]

a0 Ratio of duct area to perimeter

c0 Small signal sound speed [m/s]

d Position coordinate (like x) [m]

k Wave number [rad/m]

p0 Mean pressure [P a]

u0 Mean velocity [m/s]

44