2018 - Scott Morgan - Stability of Periodically Modulated Rotating Disk Boundary Layers

Stability of Periodically Modulated
Rotating Disk Boundary Layers
Scott Morgan
Cardiff University
Submitted for the degree of Doctor of Philosophy
August 2018
Declaration
This work has not been submitted in substance for any other degree or award at
this or any other university or place of learning, nor is being submitted concurrently in
candidature for any degree or other award.
Signed . . . . . . . . . . . . . . . . . . . . . . . . (candidate) Date . . . . . . . . . . . .
STATEMENT 1
This thesis is being submitted in partial fulfilment of the requirements for the degree
of PhD.
Signed . . . . . . . . . . . . . . . . . . . . . . . . (candidate) Date . . . . . . . . . . . . . . . . . .
STATEMENT 2
This thesis is the result of my own independent work/investigation, except where
otherwise stated. Other sources are acknowledged by explicit references. The views
expressed are my own.
STATEMENT 3
I hereby give consent for my thesis, if accepted, to be available for photocopying and
for inter-library loan, and for the title and summary to be made available to outside
organisations.

Abstract
The linear stability properties of the boundary layer generated above a disk of infinite
extent which rotates around its azimuth are explored for a novel configuration. The
rotation rate is taken to be temporally periodic, motivated by findings from Thomas et.
al. (Proc. Royal Soc. A, 2011) that the addition of an oscillatory component to an
otherwise steady flow has stabilising effects.
The vorticity-based methods that were first adopted by Davies & Carpenter (J. Com-
put. Phys., 2001) are utilised in a novel way for the solution of steady and temporally pe-
riodic eigenvalue dispersion relations. Validation of this method is provided by archetypal
flow configurations such as the steady Blasius boundary layer and the temporally periodic
Stokes layer, where Floquet theory is incorporated.
Floquet stability theory is applied to the periodically modulated rotating disk for fixed
wavenumber and fixed frequency disturbances, where it is shown that the addition of a
modulated rotation rate has a stabilising effect on the boundary layer across a range of
modulation frequencies. Confirmation is provided by frozen profile analyses and direct
numerical simulations of the subsequent flow development. An energy analysis of the
perturbation quantities is conducted to provide insights into the physical mechanisms for
the stabilisation.
The flow response to impulsive excitations in the periodically modulated rotating disk
boundary layer is explored. Direct numerical simulations of radially homogeneous and in-
homogeneous configurations are conducted and global stability behaviour is investigated.
Acknowledgements
The acknowledgements of any thesis are amongst my favourite parts to read, providing
the author with a rare opportunity to express deserving gratitude and invariably lending
some humanity to what is otherwise an impersonal scientific document. Throughout the
research which has led to the creation of this thesis, there have been a great many family,
friends and colleagues who have contributed either directly or indirectly and I express my
warmest thanks to everyone who has had even a small part to play. There are, however,
several people who I will take this opportunity to personally acknowledge, without whom
this thesis would simply not exist.
Firstly, I would like to thank the Cardiff School of Mathematics and its staff for
accommodating me throughout the project and the EPSRC for funding this research.
My deepest gratitude goes to my supervisor, my friend and mentor Dr. Christopher
Davies, whose guidance and support throughout this project has been outstanding. I
cannot express enough how much appreciation I have for your advice and direction through
the last four years. I have thoroughly enjoyed our conversations.
Secondly, to the friends and colleagues who I have had the pleasure of working along-
side throughout my time at Cardiff School of Mathematics. To Alex, for all of your
support and for being my right-hand-man. To my academic sister Martina, for your in-
valuable advice and for the endless entertainment both at work and outside. To James,
for some great conversations and distractions through squash. To Danny, for your help
and support during events, conferences and academic travelling.
I would like to express my heartfelt thanks and appreciation to Claire George, whose
commitment and determination in all walks of life is inspiring. I would not be where I
am today without your support.
For her phenomenal company in the final stages of this research, I will be eternally
grateful to Chelsea. You have an unnatural ability to distract me and make me laugh,
even when the pressures of writing a thesis are greatest. Your passion is inspirational and
thank you for everything.
Finally, for all of her support throughout my life, I thank my Mum. From revising
with me through undergraduate studies to being a rock in my life through this research, I
cannot express the gratitude I have towards you. Also to my wonderful sister Charlotte,
who is simply the best person in the world.
Contents
1 Linear Stability Analysis of Two-Dimensional Boundary Layers using a

Velocity-Vorticity Formulation 5
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.1 About this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 The Orr-Sommerfeld Equation . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.1 Eigenvalue Solutions of the Orr-Sommerfeld Equation . . . . . . . . 14
1.2.2 Inviscid Theory and Rayleigh’s Equation . . . . . . . . . . . . . . . 15
1.3 Numerical Solutions of the Orr-Sommerfeld Equation using Chebyshev Se-
ries Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.1 The Fast Fourier Transform for Chebyshev Series . . . . . . . . . . 17
1.3.2 Numerical Solution of the Orr-Sommerfeld Equation . . . . . . . . . 21
1.4 Linear Stability of Two-Dimensional Parallel Flows using a Velocity-Vorticity
Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4.1 The Velocity-Vorticity Formulation of the Navier-Stokes Equations 24
1.4.2 Two-Dimensional Steady Base Flow . . . . . . . . . . . . . . . . . . 27
1.4.3 Application of Solution Method to the Blasius Boundary Layer . . . 30
1.4.4 Linear Stability of the Temporally Periodic Stokes Layer . . . . . . 38
1.5 The Neutral Stability Curve . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.5.1 Arclength Continuation . . . . . . . . . . . . . . . . . . . . . . . . 49
2 Local Linear Stability Analysis of the Steady Rotating Disk Boundary

Layer 52
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.1.1 Numerical Solution of the Base Flow Equations . . . . . . . . . . . 54
2.1.2 High Frequency Limit for Small Amplitude Modulation . . . . . . . 63
1
2.1.3 Average Deviation from the Basic State . . . . . . . . . . . . . . . 70
2.1.4 Review of the Local Stability Properties of Steady Rotating Disk
Boundary Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.1.5 Steady Local Eigenvalue Analysis Using a Primitive Variable For-
mulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.2 Local Stability of the Steady Rotating Disk Boundary Layer using a Velocity-
Vorticity Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.2.1 Numerical Methods for Linear Stability - Eigenvalue Analysis . . . 79
2.2.2 Results and Validation for the Steady Case . . . . . . . . . . . . . . 84
2.2.3 Absolute Instability in the Steady Rotating Disk Boundary Layer . 91
2.2.4 Numerical Methods for Linear Stability - Monochromatic Direct
Numerical Simulations . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.2.5 Numerical Methods for Linear Stability - Direct Numerical Simula-
tions with Radial Dependence . . . . . . . . . . . . . . . . . . . . . 100
3 Local Stability of the Periodically Modulated Rotating Disk Boundary

Layer using a Velocity-Vorticity Formulation 104
3.1 Eigenvalue Analysis via Floquet Theory . . . . . . . . . . . . . . . . . . . 106
3.1.1 Discussion of Numerical Computation . . . . . . . . . . . . . . . . . 109
3.1.2 Temporal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.1.3 Spatial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.1.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.2 Quasi-Steady Stability Theory and the Frozen-flow Approximation . . . . . 122
3.2.1 Temporal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.2.2 Spatial Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.3 Energy Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.3.1 Energy Balance for the Steady Rotating Disk Boundary Layer . . . 129
3.3.2 Energy Balance for the Rotating Disk Boundary Layer with a Pe-
riodically Modulated Rotation Rate . . . . . . . . . . . . . . . . . . 132
3.4 Monochromatic Direct Numerical Simulations . . . . . . . . . . . . . . . . 135
3.4.1 Results and Comparison with Floquet Analysis . . . . . . . . . . . 136
3.5 Direct Numerical Simulations of Stationary Forcing with a Radial Depen-
dence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
2
3.5.1 Results and Comparison against Floquet Analysis . . . . . . . . . . 138
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4 Global Behaviour and Absolute Instabilities in Periodically Modulated

Rotating Disk Boundary Layers 144
4.1 Radially Homogeneous Steady Case . . . . . . . . . . . . . . . . . . . . . . 145
4.1.1 Convective Instability . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.1.2 Absolute Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.2 Radially Homogeneous Periodically Modulated Case . . . . . . . . . . . . . 151
4.2.1 Convective Instability . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.2.2 Absolute Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.3 Global Stability of the Periodically Modulated Rotating Disk Boundary
Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.3.1 Steady Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
4.3.2 Periodically Modulated Case . . . . . . . . . . . . . . . . . . . . . . 164
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5 Future Directions 170

5.1 The Torsionally Oscillating Disk Boundary Layer . . . . . . . . . . . . . . 170
5.1.1 Local Linear Stability of the Torsionally Oscillating Disk Boundary
Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.1.2 Hydrodynamic Voltammetry at Rocking Disk Electrodes . . . . . . 178
5.2 The Rotating Disk Boundary Layer with Surface Roughness . . . . . . . . 179
5.2.1 The Partial Slip Model (MW) . . . . . . . . . . . . . . . . . . . . . 180
5.2.2 The Surface Geometry Model (YHP) . . . . . . . . . . . . . . . . . 183
5.2.3 Future Directions for Surface Roughness . . . . . . . . . . . . . . . 186
6 Conclusions 187
Appendices 200
A Convergence of Eigenvalue Routines 201

A.1 Blasius Boundary Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
A.2 Stokes Boundary Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
A.3 Steady Rotating Disk Boundary Layer . . . . . . . . . . . . . . . . . . . . 204
3
A.4 Periodically Modulated Rotating Disk Boundary Layer . . . . . . . . . . . 208
B Algorithms and Numerical Schemes 210

B.1 Numerical Solution of the Base Flow Equations . . . . . . . . . . . . . . . 210
B.2 Monochromatic Numerical Simulations . . . . . . . . . . . . . . . . . . . . 212
B.3 Numerical Simulations with Radial Dependence . . . . . . . . . . . . . . . 214
B.3.1 Inflow and Outflow Conditions . . . . . . . . . . . . . . . . . . . . . 214
B.3.2 Finite Difference Discretisation Schemes . . . . . . . . . . . . . . . 215
B.3.3 Poisson Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
4
Chapter 1
Linear Stability Analysis of

Two-Dimensional Boundary Layers
using a Velocity-Vorticity
Formulation
1.1 Introduction
The following exposition broadly describes the study of instability mechanisms in flows of
so-called Newtonian fluid. Stability theory is a large and important area of fluid dynamics,
with the essential problems being recognised, among others, by Reynolds, Rayleigh, Kelvin
and Helmholtz towards the end of the nineteenth century.
We will make reference to stability throughout and it will prove useful to have a con-
cept of what it means for a flow to be stable or unstable from the outset. In what follows,
we will take stability as an umbrella term for describing the subsequent behaviour of a
fluid flow after having been subjected to a small perturbation or variation. James Clerk
Maxwell, as explained in Drazin [32], described stability concisely in the late nineteenth
century as
When...an infinitely small variation of the present state will alter only by an infinitely
small quantity the state at some future time, the condition of the system, whether at rest
or in motion, is said to be stable; but when an infinitely small variation in the present
5
state may bring about a finite difference in the state of the system in a finite time, the
condition of the system is said to be unstable.
This explains that following some sort of disturbance, the flow is unstable if it never
returns to its original state but stable if the disturbance decays over time. In almost
all applications, stable flows are preferable and instabilities are usually associated with
greater energy consumption, less efficiency and greater cost. Unstable flows can often
evolve into a chaotic state of motion called turbulence which is extremely irregular and
often undesirable. Turbulence is very poorly understood and in 2006, Richard Feynman
[34] described it as the most important unsolved problem of classical physics.
For the last century or so, many mathematicians and physicists have attempted to
understand the complexities of stability theory and turbulence. Osborne Reynolds’ [62]
famous series of experiments in 1883 studying the stability of fluid flow in a pipe were
among the first of their kind, and still present one of the clearest ways of introducing
stability theory.
Reynolds considered flow in three pipes at varying velocities, studying the behaviour
of dye streaks in the liquid, and figure 1.1 shows the configuration where two distinct
cases develop over time.
The first image shows the case of laminar flow, where the streak extends in a straight
line through the tube. In the other case, the dye has mixed with the fluid and filled
the tube with colour, indicating a turbulent situation. The third drawing shows the
second case in more detail, illustrating the colour as a mass of curls and eddies during
the transition to turbulence.
Reynolds proceeded to demonstrate that the different cases of smooth laminar flow
and turbulence are dependent on the quantities V , the maximum velocity of the fluid in
the pipe, a, the radius of the pipe and ν, the kinematic viscosity of the fluid. He noted
Va
that the laminar flow started to break down when the ratio ν
exceeded a certain value.
This dimensionless quantity is now known as the Reynolds number and can be defined
differently depending on the flow configuration. Throughout this report we will denote
the Reynolds number by R. It should be noted however, that many texts and references
cited herein use the notation Re instead, with no interesting distinction between the two.
This idea of a somewhat critical Reynolds number above which the flow is unstable is
6
Figure 1.1: Reynolds’ experiment of flow in a pipe, taken from Drazin [32].
in fact a general idea and can be applied to many flows. Typically, a fluid flow is stable
if the Reynolds number R is small enough, although there might exist a critical value
Rc such that the flow is unstable with respect to infinitesimal perturbations for R > Rc .
Finding this critical value Rc is often a computationally difficult task and must be carried
out numerically.
On the topic of computation, both experimental and analytic solution techniques have
drawbacks, with experiments being almost always subject to disruption by noise in the
system and analytical approaches being hindered by the complexity of the Navier-Stokes
equations. These limitations have been somewhat alleviated in more recent times with
the rapid development of technology and efficient numerical methods have been developed
which can, in some cases, model even the fine complexities of turbulence with sufficient
accuracy in a practical time frame.
7
There are several types of stability analysis that exist and these can be loosely cate-
gorised into linear or non-linear analysis and local or global analysis and all must be used
in conjunction with each other if a flow configuration is to be fully understood.
Linear theory is usually employed as a first stage and will predict whether sufficiently
small disturbances grow or remain bounded, although further investigation is often neces-
sary to fully understand the outcome. There are famous examples of linearly stable flows
that actually become turbulent such as Poiseuille pipe flow, largely due to the fact that
as disturbances grow they become large enough to invalidate the linear theory. Weakly
non-linear theories predict the next stages of disturbance evolution if growth rates are
weak but not infinitesimal and in recent years, with an increase in computational power,
it has been possible in some cases to conduct fully non-linear simulations.
Local analyses may be carried out by separating the disturbances into Fourier-type
travelling wave modes, a superposition of which gives an idea of the stability properties.
This is the background to the Orr-Sommerfeld equation and its relatives introduced in
section 1.2 and discussed through chapters 2-3. Global analyses are typically conducted
using time-dependent simulations of the Navier-Stokes equations, and offer deeper insights
into the real-world flows. We will discuss a global approach to stability theory in chapter
4.
The classical, historical approach has been to study the local, linear development of
infinitesimal disturbances to a known steady base flow field, and attempt to glean some
understanding of the flow configuration as a whole. Under this theory, regions of a fluid
flow domain may be categorised into one of three categories:
• Stable
• Convectively unstable
• Absolutely unstable
and an illustration of convective and absolute instabilities can be seen in figure 1.2. These
schematic diagrams are taken with reference to x as the streamwise direction and the time
t. If at all streamwise locations x, the disturbance decays with time, then the flow is said
to be stable. In such flows all disturbances always decay with time. If a disturbance
grows as it propagates upstream or downstream but is such that at any given x, the
flow at that point eventually returns to the undisturbed basic state, the flow is said to
8
(a) Absolute instability (b) Convective instability
Figure 1.2: Schematic diagram of convective and absolute instabilities for illustration
purposes.
be convectively unstable. On the other hand, if at that fixed streamwise location, the
magnitude of the disturbance grows with time, it is said to be absolutely unstable. In such
flows, the disturbance grows at all spatial locations and at all times.
1.1.1 About this Thesis
Throughout this presentation we detail the instability mechanisms for several flow con-
figurations in a variety of situations. We discuss numerical techniques, some of which
are novel, before setting the scene for a thorough examination of a particular three-
dimensional temporally periodic boundary layer. This boundary layer is formed in a
configuration which has not previously been studied, over a disk which rotates with a
modulated rotation rate.
Prior to our examination of the local and global stability properties of such a con-
figuration, chapter 1 provides a thorough introduction to stability theory and introduces
the novel numerical methods in application to archetypal two-dimensional flow configura-
tions. We present discussions of standard equations such as the Orr-Sommerfeld equation
in section 1.2 and provide an overview of the numerical solution methods used throughout
this thesis based on Chebyshev series and their corresponding integral operators.
9
Chapter 2 provides a comprehensive review of the local linear stability properties of
the boundary layer formed over a disk which rotates with a constant rotation rate and
applies the two-dimensional methods of chapter 1 to this three-dimensional scenario. This
configuration was first studied by von Kármán [80] in 1921 due to its admission of an exact
similarity solution to the Navier-Stokes equations and later by Gregory et al. [38] due to
its similarities with fluid flow over swept-wings. Further interest in the configuration
was generated by Lingwood [43] after her discovery of a local absolute instability in
the flow and further still by Davies and Carpenter [27] upon their realisations that the
local absolute instability did not generate a linearly unstable global mode, leading to the
conclusion that convective instabilities persist at all Reynolds numbers, at least for the
range of azimuthal mode numbers considered by Davies and Carpenter [27]. However,
more recently, Thomas and Davies [75] have explored a higher range of azimuthal mode
numbers than what was previously studied and have found evidence for global instability
at these levels. Further work has been done on the non-linear behaviour of disturbances
via direct numerical simulations in studies such as Appelquist et al. [4, 5, 6]. Research
into the global stability properties of the steady rotating disk boundary layer is still highly
active and Lingwood and Alfredsson [46] provide a thorough review of the field prior to
2015.
In chapter 3, we present a discussion of a novel flow configuration and analyse the
stability properties of a disk of infinite extent which rotates with a periodically modu-
lated angular velocity about the azimuth. The motivation for the addition of a modulated
rotation rate originates with the Thomas et al. [76] study of an oscillatory Stokes layer
added to a Poiseuille flow. The authors show that inducing this small amount of oscil-
lation can have stabilising effects, complementing the work of Kelly and Cheers [42] and
Von Kerczek [81] who studied modulation of plane Couette flow and also found stabilisa-
tion. Thomas et al. [76] found that the steady flow was stabilised for certain frequencies
of the oscillation but destabilised for others. Wise and Ricco [82] also show that turbulent
drag reduction may be achieved using oscillation of flat plates and rotating disks.
Chapter 4 extends the discussion of the local analysis of the periodically modulated
rotating disk boundary layer to a global setting, and introduces techniques for direct nu-
merical simulations of a response to an impulsive forcing. We briefly review the global
behaviour in the steady case, before presenting archetypal results for the modulated con-
10
figuration. Some discussion is given to the aforementioned discovery by Davies and Car-
penter [27] that the steady flow is linearly globally stable and some exploratory results
are presented for the modulated case. However, a detailed examination of the subtleties
involved in this configuration is reserved for future work.
Aside from stabilising the steady rotating disk boundary layer, the periodically mod-
ulated rotating disk is an interesting problem in its own right, and has applications in
electrochemical engineering. As discussed in Bard and Faulkner [8], the steady rotating
disk electrode provides a convenient configuration for conducting hydrodynamic voltam-
metry, where convection of an analyte controlled by the rotation is designed to increase
the rate of controlled mass transport of ions at an electrode surface. This technique is
commonly used for electroplating, and Schwartz et al. [66] discuss the effects of a pe-
riodically modulated rotation rate on the mass transfer. These concepts will be briefly
discussed in the context of future work in chapter 5.
Also discussed in chapter 5 are the parallels and similarities between our proposed
stabilising technique and periodically distributed surface roughness. Several recent studies
such as Garrett et al. [36] and Cooper et al. [21] have investigated radially iso- and
anisotropic roughness in the context of linear stability and found significant stabilising
effects. The significance of our method in the context of theirs is discussed, and some
advantages of our technique are presented.
The following sections are intended to be introductory, and provide an overview to
the interested reader of the fundamental equations that underpin stability theory. Some
of the derivations are standard and discussed in many introductory textbooks so we will
omit the finer details for brevity. However, an overview will be presented at this stage
thereby setting the scene for the forthcoming chapters.
1.2 The Orr-Sommerfeld Equation

Consider flows of an incompressible viscous fluid in a given domain Π. Let ρ be the
density of the fluid, and ν the kinematic viscosity. Let also u* and p∗ be the dimensional
velocity and pressure respectively of the fluid at a given point x* and time t∗ . Then the
Navier-Stokes equations which govern the fluid motion in the absence of body forces are
given by
11
∂u*
∗
+ (u* · ∇∗ )u* = −∇∗ p∗ + ν∇∗ 2 u* (1.1a)
∂t
∇∗ · u* = 0 (1.1b)
Non-dimensionalisation is carried out by some characteristic velocity scale V and length

scale L, to give the dimensionless quantities
x* u* V t∗ p∗
x= , u= , t= , p= (1.2)
L V L ρV 2
Substitution of (1.2) into the equations (1.1) gives the familiar non-dimensional Navier-
Stokes equations
∂u 1
+ (u · ∇)u = −∇p + ∇2 u (1.3a)
∂t R
∇·u=0 (1.3b)
VL
where R is the Reynolds number defined by R = ν
. Equation (1.3a) is known as the
momentum equation, while equation (1.3b) is known as the continuity equation and is
equivalent to conservation of mass in the system. It is sometimes useful to interpret the
Reynolds number as a measure of the inertial effects relative to the viscous effects of the
fluid. If R is large, the viscous term is typically much smaller than the inertial term,
indicating that it may be safe to treat the flow as essentially inviscid. Some care must
be taken when doing this and conducting stability analyses however, as is highlighted in
section 1.2.2.
To provide an illustration of common linear stability concepts, we consider a steady
two-dimensional base flow of the form
U = (U (z), 0, 0), P = constant for z1 ≤ z ≤ z2
where z1 , z2 ∈ R ∪ {±∞} bound the region of the flow. The analysis proceeds by
introducing a small perturbation to the flow so that
u(x, t) = U(z) + u0 (x, t), p(x, t) = P + p0 (x, t)
where x = (x, y, z), the 0 denotes the perturbation quantities and << 1. After substitut-
ing these representations for u and p into the equations (1.3) and linearising by retaining
12
O() terms, we get
∂u0 1
+ (U · ∇)u0 + (u0 · ∇)U = −∇p0 + ∇2 u0 (1.4a)
∂t R
∇ · u0 = 0 (1.4b)
which are often referred to as the perturbation equations. Indeed, the perturbation equa-
tions referred to throughout this document should be understood as the momentum and
continuity equations for the perturbation quantities.
The following method relies on the base flow being parallel, which means that the
velocity does not vary in the direction of the flow. This property is satisfied exactly for
simple flows such as Poiseuille channel flow, but the method carries through for other flows
provided we apply the so-called parallel flow approximation. This amounts to treating the
streamlines as being approximately parallel and proceeding with the method anyway.
Used with caution, non-parallel contributions can often be ignored in a system with few
adverse effects, although this is not true for all flow configurations as will be discussed
later in the context of the rotating disk boundary layer. Assuming for now that we are
at liberty to apply it, the parallel flow approximation allows us to assume that the base
flow is independent of x and y, thereby making the solution separable and enabling us to
consider a normal mode solution in the form of a travelling wave given by
u0 = û(z)ei(αx+βy−ωt) , p0 = p̂ei(αx+βy−ωt) (1.5)
The parameters α, β and ω can be interpreted as the wavenumber in the x-direction, the
wavenumber in the y-direction and the disturbance frequency respectively. The wavenum-
ber is the spatial frequency of the travelling wave solution, and can be interpreted as the
number of waves that exist over a specified distance. The full derivation of the Orr-
Sommerfeld equation is standard and described thoroughly in many introductory texts
to hydrodynamic stability so is omitted here, but the interested reader is referred to a
very well presented derivation in [25]. Another standard result, Squire’s theorem, which
enables us to consider only two-dimensional disturbances when investigating the stabil-
ity properties of a flow, is also stated, proven and described in reference [25]. Thus, we
merely state here only the Orr-Sommerfeld equation for two-dimensional perturbations
to the steady base flow U (z) as
(D2 − α2 )2 − iR (αU − ω)(D2 − α2 ) − αU 00

φ=0 (1.6)
13
d2
where D2 = dz 2
and the perturbation stream function φ is of the form
φ̂(x, z, t) = φ(z)ei(αx−ωt) (1.7)
The boundary conditions that are imposed on φ are
φ(z1 ) = 0, φ(z2 ) = 0, φ0 (z1 ) = 0, φ0 (z2 ) = 0 (1.8)
where z1 ≤ z2 . Similar conditions hold for a semi-infinite or unbounded flow where either
one or both of z1 and z2 are infinite.
1.2.1 Eigenvalue Solutions of the Orr-Sommerfeld Equation
The Orr-Sommerfeld equation (1.6) is a fourth order differential equation, which upon
solving would admit four linearly independent solutions φ1 , φ2 , φ3 , φ4 which could be
written as
φ = A1 φ1 + A2 φ2 + A3 φ3 + A4 φ4
subject to the boundary conditions (1.8). This can be recast in the usual way as a matrix
problem     
φ (z ) φ2 (z1 ) φ3 (z1 ) φ4 (z1 ) A 0
 1 1  1   
    
 φ1 (z2 ) φ2 (z2 ) φ3 (z2 ) φ4 (z2 )   A2   0 
  = 
 0 0 0 0    
 φ1 (z1 ) φ2 (z1 ) φ3 (z1 ) φ4 (z1 ) A   0 
 3   
 

φ01 (z2 ) φ02 (z2 ) φ03 (z2 ) φ04 (z2 ) A4 0
which has non-zero solutions if the determinant of the matrix on the left hand side is
zero. Hence, the solution for the eigenvalues of the Orr-Sommerfeld equation (1.6) can be
expressed in the form of the eigenvalue relation
D(ω, α, Re) = 0 (1.9)
This generalised eigenvalue problem (1.9) is called the Orr-Sommerfeld problem and is
attributed independently to Orr [55] and Sommerfeld [68]. This equation is central to
stability theory since the analysis of the eigenvalues of this equation for a fixed Reynolds
number provides insight into the linear stability properties of flow configurations. There
are two main types of stability analysis usually performed, temporal and spatial, which
impose conditions on α and ω and are outlined below.
14
• If α is fixed real and ω the complex eigenvalue then the disturbances are periodic
in x and grow or decay in time depending on the sign of the imaginary part of ω.
In this analysis, the eigenvalue ω appears linearly in the stability equation. This is
called temporal stability analysis.
• If ω is prescribed real and α the complex eigenvalue then the disturbances are
periodic in time and grow or decay exponentially with x. This is termed spatial
stability analysis and the corresponding disturbance growth may be observed in
reality.
Inspection of the normal mode approximation (1.7) reveals that we can interpret the real
and imaginary parts of α as the spatial frequencies and growth rates respectively, and
similarly for ω. The original comparisons between experimental results and the growth
rates predicted by linear theory were based on converting temporal growth rates into
spatial growth rates using a downstream convection speed. However, as argued by Gaster
[37], it is more physical to calculate the spatial growth rates directly and this is now com-
putationally feasible. The difficulty with this approach is that the eigenvalue α appears
non-linearly with a fourth power, forming a polynomial eigenvalue problem. A method
for dealing with this issue is given by Bridges and Morris [14] which involves an indefi-
nite integration of the Orr-Sommerfeld equation and the use of Chebyshev polynomials
to solve the eigenvalue problem. Other methods of solving spatial eigenvalue problems
exist, such as those based on a differential formulation, as seen in Trefethen [79] or on
a system of first order equations, see Malik [50], but the integral formulation of Bridges
and Morris [14] has some advantages over these other methods which will be discussed
in section 1.3. This integral formulation provides the solution method for the majority
of analyses conducted through this work and will be discussed at length throughout this
thesis.
1.2.2 Inviscid Theory and Rayleigh’s Equation
In the interest of completeness, we briefly consider the limit of inviscid flows and Rayleigh’s
equation, which is the counterpart of the Orr-Sommerfeld equation in this limit. Fixing
z and α and letting R → ∞, the Orr-Sommerfeld equation (1.6) becomes
(U − c)(D2 − α2 )φ − U 00 φ = 0 (1.10)
15
which is known as Rayleigh’s equation. Since this equation is only of second order, we
only require the two boundary conditions
φ(z1 ) = 0, φ(z2 ) = 0
This equation is important in the study of flows for which the use of an inviscid fluid gives
a good approximation to the stability characteristics of a viscous fluid at large Reynolds
numbers. Care must be taken however and in fact, viscosity, although it dissipates energy,
may destabilise a flow which is stable for an inviscid fluid. Introducing viscosity introduces
a viscous layer near the wall that brings this inviscid slip velocity down to zero at the wall
in what is known as boundary layer behaviour. Prandtl [58] suspected that the inclusion of
a viscous wall layer could be destabilizing and Tollmien [77] found asymptotic solutions to
the Orr-Sommerfeld equation for profiles with no inflection point predicting instability but
their counter-intuitive results were not widely accepted until this behaviour was verified in
wind tunnel experiments on boundary layers in the 1940s. It had been previously assumed
that the additional friction associated with viscosity would be stabilising, but this turned
out not to be the case. Therefore, while inviscid theory can be used to investigate stability,
it must be used with caution and indeed if it predicts stability then this may not give
the correct behaviour once viscosity is incorporated. However in general, if the inviscid
theory predicts instability then it will usually be correct with small viscous corrections.
1.3 Numerical Solutions of the Orr-Sommerfeld Equa-

tion using Chebyshev Series Expansions
This section turns attention back to the Orr-Sommerfeld equation (1.6) and discusses in
detail the procedure taken to solve the resulting eigenvalue problem. A standard tech-
nique, detailed in Bridges and Morris [14], will be briefly discussed to provide context
before we introduce a method based on a velocity-vorticity formulation of the Navier-
Stokes equations which to the extent of our literature search has not been described
previously. This approach will be described in detail for the standard Blasius flow config-
uration which will serve as an introduction to the added three-dimensional complexities
later studied in the context of the linear stability of the rotating disk boundary layer. The
solution procedure relies on expanding the disturbance functions in terms of Chebyshev
16
polynomials and a form of the Chebyshev-tau method. In the interest of concision we will
assume throughout that the base flow and disturbance functions live in an appropriate
functional space and can thus be expanded in terms of Chebyshev polynomials.
The Chebyshev polynomials Tn are defined in the standard way as
Tn (x) = cos(n arccos(x)), x ∈ [−1, 1] (1.11)
and using basic trigonometric identities, we can derive the recursion relations
T0 (x) = 1
T1 (x) = x
Tk+1 (x) = 2xTk (x) − Tk−1 (x) for k ∈ {1, 2, ...}
These polynomials have many properties which have made them a favourable choice for
spectral expansions in computational mathematics for many years. They are orthogonal
functions in the domain [−1, 1] with respect to their defining inner product and hence, in
the appropriate space, any function can be expressed via the expansion
∞
X
f (x) = ck Tk (x)
k=0
while the coefficients ck are determined from some appropriate inner product relation.
While not unique in their ability to do so, they also alleviate some of the computational
phenomena associated with either interpolation by trigonometric polynomials or interpo-
lation on equally spaced points.
The most important advantage of the Chebyshev series representations in terms of the
work conducted for this report is their close links with Fourier series and the availability
of a Fast Fourier Transform (FFT) for computationally transforming between physical
space and Chebyshev space. The FFT has been discussed extensively by several authors
including Boyd [13] and Trefethen [79] and so only the main points will be discussed
here. In particular, since throughout this work we will work mainly with even and odd
representations separately, we lean towards discussing these cases in what follows.
1.3.1 The Fast Fourier Transform for Chebyshev Series
We begin by introducing Fourier spectral methods in general, before moving towards

describing the methods used in this report regarding Chebyshev expansions. It is well
17
known that the Fourier transform of a function f (θ) is the function F (k) defined by
Z ∞
F (k) = e−ikθ f (θ)dθ, k ∈ R (1.12)
−∞
and we can interpret the transformed variable F as a decomposition of the physical

variable f into sinusoids of different frequencies. Similarly, we can recover the physical
variable f from F by way of the inverse Fourier transform
Z ∞
1
f (x) = eikθ F (k)dk, θ ∈ R (1.13)
2π −∞
As one might imagine, these transforms have discrete analogues which are appropriate
for computational problems on discretised domains. If we consider a bounded, periodic
interval [0, 2π] and divide it into an equally spaced grid with N points, labelled θj and a
2π
grid spacing h = N
, then the discrete Fourier transform of a function f (θ) evaluated at
the points θj is given by
N
X N N
ck = h e−ikθj f (θj ), k=− + 1, . . . , (1.14)
j=1
2 2
while the inverse discrete Fourier transform is given by

N/2
1 X
f (θj ) = eikθj ck , j = 1, . . . , N (1.15)
2π
k=−N/2+1
We have assumed here that N is even, as we will throughout, although similar results
hold for odd N . The periodicity refers to the fact that any data on the grid points comes
from evaluating a periodic function, while periods of lengths other than 2π are easily
accommodated by some scaling factor. Of course, this periodic restriction is a serious
one, and not all problems can be treated as such. If we were to periodically extend
a smooth but inherently non-periodic function then the contamination caused by the
discontinuities would be global if using equally spaced points as we have above. This is
known as the Gibbs phenomenon and destroys the spectral accuracy of the scheme.
This problem can be alleviated by using unevenly spaced points in the periodic domain
an example of which we illustrate by considering the domain [−1, 1]. Various different
sets of points can be used, but they mainly share the common property that
N
density ≈ √ (1.16)
π 1 − x2
This has the effect of clustering the points near the boundaries, with the average grid
spacing being O(N −2 ) for x ≈ ±1 and O(N −1 ) away from the boundaries. Our example
18
of such points are the well-known Chebyshev-Lobatto or Chebyshev-extreme collocation
points in [−1, 1] given by

jπ
xj = cos , j = 0, . . . , N (1.17)
N
These points are naturally associated with the Chebyshev polynomials Tn , since they are
the extreme values of Tn in [−1, 1]. If, as we will do often in what follows, we choose to
work only with odd and even Chebyshev representations separately then we can define
our collocation points on the half-grid (0, 1] as

jπ
xj = cos , j = 0, . . . , N − 1 (1.18)
2N
We can also derive direct relations between Fourier series and Chebyshev series by ap-
plying some transformations to the variable θ in equation (1.14). As we have seen, the
Fourier transform is applicable on the domain R and so taking some θ ∈ R, we can recover
the domain x ∈ [−1, 1] by writing x = cos(θ). Thus, combining this with equation (1.11),
the Chebyshev polynomials on [−1, 1] can be written as
Tn (x) = cos(nθ) (1.19)
and it can be shown using trigonometric identities that Tn defines a polynomial of degree
n and that Tn is odd when n is odd and even when n is even. Since Tn is exactly degree
n for each n, any N degree polynomial can be written uniquely as a linear combination
of Chebyshev polynomials. Thus, if we have a degree N polynomial p such that
N
X
p(x) = pn Tn (x), x ∈ [−1, 1] (1.20)
n=0
then there is the equivalent representation for the 2π-periodic even trigonometric polyno-
mial P given by
N
X
P (θ) = pn cos(nθ), θ∈R (1.21)
n=0
Hence, given an arbitrary function f defined on [−1, 1], we can draw the comparison
between the Chebyshev and Fourier series by saying that P (θ) interpolates f at the
equispaced points θj while p(x) interpolates f at the Chebyshev points xj .
Similar ideas can be employed when considering odd and even Chebyshev representa-
π
tions separately. Any function f (x) which is symmetric about 0 in x or about 2
in θ can
be written as
∞
f0 X
f (x) = + fn T2n (x), x ∈ (0, 1] (1.22)
2 n=1
19
since all even degree cosines have the same property. The halving of the first term in a
Chebyshev expansion is standard practice, and helps with consistency for several identi-
π
ties. Similarly, a function g(x) which is anti-symmetric about 0 in x or about 2
in θ takes
the form
∞
X
g(x) = an T2n−1 (x), x ∈ (0, 1] (1.23)
n=1
since all odd degree cosines have the same property. When talking about these repre-
sentations throughout this report, we will describe them as even and odd representations
respectively. The reader should be aware of a possible confusion. The labelling even and
odd refers to the degree of the expansion polynomials, and thus to the representation in
terms of x ∈ (0, 1].
The transformations between physical space and Chebyshev space are necessary for
many situations, including several outlined in this report. A solution method which travels
back and forth between physical space and Chebyshev space is generally called a pseudo-
spectral method, where the advantages of each space are exploited at each stage. For
example, integration and differentiation are far easier and more accurate when performed
in Chebyshev space, while products between functions are easiest performed in physical
space. Since the Chebyshev series are so closely related to Fourier series, there is the
availability of a highly efficient technique called the Fast Fourier Transform (FFT) for
converting from physical space to Chebyshev space and vice-versa. This transform can be
computed in O(N log N ) steps, as opposed to the O(N 2 ) required for a standard Fourier
transform. The algorithm for the FFT is standard and largely irrelevant for the purpose
of this report, since most numerical algorithms libraries such as NAG [53] and MATLAB
have built in routines that carry out the transforms. The precise details of the algorithm
are therefore omitted here for brevity, while the interested reader is directed to Boyd [13]
or Trefethen [79] where a thorough introduction and explanation is given.
Given the parities which we are interested in utilising, we can adapt the FFT routine
slightly and define the discrete cosine transform as
n−1
1 X jkπ 1
f k = x0 + xj cos + (−1)k xn , k = 0, . . . , n − 1 (1.24)
2 j=1
n 2
which is equivalent up to an overall scale factor of 2 to a discrete Fourier transform of

length 2(N −1) with even symmetry. This is its own inverse, when multiplied by 2/(N −1).
This routine can be used directly to calculate the even Chebyshev coefficients of a function
20
which can be approximated by a sum of even Chebyshev polynomials and likewise in the
opposite direction. In practise, it is accomplished using the proprietary NAG [53] routine
C06RBF or the open-source routine dcost from dfftpack [54].
The transformations for the odd representations are not so obvious, but can be
achieved by way of the discrete quarter-wave sine transform which can be defined as
n−1
X j(2k − 1)π 1
gk = xj sin + (−1)(k−1) xn , k = 0, . . . , n (1.25)
j=1
2n 2
with inverse
n
X j(2k − 1)π
xk = 2 gj sin k = 0, . . . , n (1.26)
j=1
2n
Given a representation
∞
X
g(xj ) = gn T2n−1 (xj ), x ∈ (0, 1] (1.27)
n=1
we can write
∞
X (2n − 1)jπ
G(θj ) = gn cos (1.28)
n=1
2N
where the collocation points xj are defined by equation (1.18). Using the trigonometric
identity
π π
cos(φ) = sin φ + = − sin φ − (1.29)
2 2
and setting hn = (−1)n−1 gn we get
∞
X (2n − 1)jπ
G(θN −j ) = hn sin (1.30)
n=1
2N
which is in the correct form for a discrete quarter-wave sine transform. In practise, this
transform is accomplished using the proprietary NAG routine C06RCF [53] or the open-
source routine dsinqt from dfftpack [54].
1.3.2 Numerical Solution of the Orr-Sommerfeld Equation
There appear to be two standard methods for solving the temporal and spatial eigenvalue
problems associated with the Orr-Sommerfeld equation. The first, based on derivatives of
Chebyshev polynomials is discussed in many texts including Trefethen [79] and Schmid
and Henningson [65] and utilises the identity
1 0 1 0
2Tk (x) = Tk+1 (x) − Tk−1 (x) (1.31)
k+1 k−1
21
dk
where T00 (x) = 0 and T10 (x) = 1 to derive the so-called differentiation matrices Dk ≈ dxk
which operate on the physical values of the function being differentiated. This approach is
discussed in great detail by Trefethen [79] and has been formalised in a MATLAB library
called chebfun [78].
While this approach is effective and has long been used for this type of problem, we
will adopt an alternative technique, described in Bridges and Morris [14], in which the
Orr-Sommerfeld equation is integrated indefinitely and solved using matrix representa-
tions of the integral operators. The integrated form of the Chebyshev polynomials takes
a somewhat more convenient form than the differentiated form and so has been chosen
as the solution method for all cases in this report. The main advantage of the integral
operators is that they can be expressed as n-diagonal banded matrices, for relatively
small n, as opposed to the fully populated ones present in the differential formulation. In
time-dependent simulations, and in particular the ones carried out for the rotating disk
in chapter 2, this allows for a modified Thomas algorithm to be implemented at each
time step, rather than a full matrix inversion. The matrix representations of the oper-
ators in the differential formulation are fully populated and would thus greatly increase
the computational cost. The same advantage is not present in the eigenvalue problem
however, since products of the base flow profile with the disturbance variable lead to fully
populated matrices, but the integral formulation is retained since the operator definitions
are identical in both cases and thus only need to be derived once.
As an illustration of the banded representation, we can derive the following identity
for the integration of Chebyshev polynomials from equation (1.31)

Tn+1 (x) Tn−1 (x)
 2(n+1) − 2(n−1) n ≥ 2






Z 


Tn (x)dx = 1 (1.32)
[T (x) + T2 (x)] n = 1
4 0









 T (x)
1 n=0
Considering the interpolating polynomial p to an arbitrary appropriate function f defined

on [−1, 1] with Chebyshev representation
N
X
p(x) = pn Tn (x)
n=0
22
we can write
Z N
X Z
f (x)dx ≈ pn Tn (x)dx
n=0
N
1 1X Tn+1 (x) Tn−1 (x)
= aT0 (x) + p0 T1 (x) + [T0 (x) + T2 (x)]p1 + − pn
4 2 n=2 n+1 n−1
N
1 1X1
= aT0 (x) + +p0 T1 (x) + [T0 (x) + T2 (x)]p1 + (pn−1 − pn+1 )Tn (x)
4 2 n=2 n
where a is an arbitrary constant of integration. This results in a tridiagonal banded

R
matrix representation for the integral operator I1 := .
1.4 Linear Stability of Two-Dimensional Parallel Flows

using a Velocity-Vorticity Formulation
The method of Bridges and Morris [14] proceeds by indefinitely integrating Orr-Sommerfeld
equation (1.6) four times with respect to the single flow coordinate variable and rearrang-
ing to give a polynomial eigenvalue equation in α given by
ZZZZ ZZZZ ZZ ZZZZ
4 3 2
α φ+α iR U φ + α −2 φ − iωR φ +
b1 y 3 b2 y 2
ZZZ ZZ ZZ
0
α 2iR U φ − iR U φ + φ + iωR φ + + + b3 y + b4 = 0
6 2
(1.33)
where bi , i = 1, . . . , 4 denote constants of integration. We note that, throughout this

thesis, repeated integrals are taken with respect to the same variable as the goal of the
integration procedure is to eliminate higher order derivatives in the equations in prepa-
ration for the numerical solution method. This notation is standard in the literature,
see Bridges and Morris [14] and Cooper and Carpenter [19]. Whilst no further mention
of this exact solution method will be given in this report, many of the same ideas carry
through to our new formulation which we present in detail. The results of Bridges and
Morris [14] were used as a validation tool for the new method and are described in detail
for the Poiseuille channel flow by the authors of [14]. The interested reader is referred to
the references for further details.
23
1.4.1 The Velocity-Vorticity Formulation of the Navier-Stokes
Equations
In 2001, Davies and Carpenter [26] developed a velocity-vorticity formulation of the un-
steady, three-dimensional Navier-Stokes equations which is particularly suitable for sim-
ulating the evolution of disturbances in three-dimensional boundary layers.
Velocity-vorticity methods provide a convenient format for dealing with unsteady flow
fields and have several advantages over traditional methods, a key one being that there
are only three governing equations for three dependent variables, the so-called primary
variables.
In this formulation, the dependent variables are perturbations to a known, undisturbed
flow field where the three primary variables comprise two vorticity components and the
wall normal velocity component. These are governed by two vorticity transport equations
and a Poisson equation. The secondary variables consist of the remaining velocity and
vorticity components and the pressure and can be determined explicitly from the primary
variables by means of an integral. It is shown by Davies and Carpenter [26] that, subject to
some fairly general conditions at infinity, the formulation is fully equivalent to the Navier-
Stokes equations. It is also worth noting that the only fundamental boundary conditions
are those that must be imposed on the velocity field, and in Davies and Carpenter’s [26]
formulation, no boundary conditions at the wall are needed for the vorticity. Instead,
integral conditions can be derived from the definition of vorticity, which will be discussed
later.
To introduce this method, we consider the velocity-vorticity formulation for the Blasius
boundary layer. This flow is two-dimensional and has been included here to serve as
a stepping-stone to the analysis of the three-dimensional rotating disk boundary layer.
Although there is no necessity to use the velocity-vorticity formulation for the spatial
stability analysis of this flow, there is a significant advantage to doing so. This is due to
the fact that the perturbation equations are reduced to two second order equations, as
opposed to one fourth order one, making the integral operator definitions much neater
and more amenable to calculations. It should also be noted that the following derivation
is base flow independent, provided the base flow in question is either steady or temporally
periodic, two dimensional and parallel. Validation of this new approach was also carried
out against the well-documented Poiseuille channel flow but these results are not presented
24
here.
Consider a base flow UB = (UB , VB , WB ) with vorticity Ξ = ∇ × U = (Ξx , Ξy , Ξz ).
Let u denote the usual velocity perturbation components and ξ the vorticity perturba-
tion. Then the velocity-vorticity formulation equations for the three primary variables
{ξx , ξy , w} take the form
∂ξx ∂Nz ∂Ny 1
+ − = ∇2 ξx (1.34a)
∂t ∂y ∂z R
∂ξy ∂Nx ∂Nz 1
+ − = ∇2 ξy (1.34b)
∂t ∂z ∂x R
∂ξx ∂ξy
∇2 w = − (1.34c)
∂y ∂x
where N = (Nx , Ny , Nz ) is defined as
N = Ξ × u + ξ × UB + ξ × u (1.35)
Linearisation can be performed if necessary by dropping out the ξ × u term in N.

The convective quantity N can only be evaluated if the secondary variables {ux , uy , ξz }
are known. Thus, we make the explicit definitions
Z ∞
∂w
ux = − ξy + dz (1.36)
z ∂x
Z ∞
∂w
uy = ξx − dz (1.37)
z ∂y
Z ∞
∂ξx ∂ξy
ξz = + dz (1.38)
z ∂x ∂y
which are obtained by integrating the appropriate definitions of vorticity via the usual
relation ξ = ∇ × u.
Boundary Conditions
We will present here an overview of the imposed boundary conditions in the case of an
unbounded flow contained by a stationary flat plate located at z = 0. The boundary con-
ditions imposed on the system are derived from the no-penetration and no-slip conditions
at the wall and the integral constraint imposed on the vorticity. Hence, at the stationary
boundary wall z = 0, we immediately have the boundary conditions
Z ∞
∂w
ux (0) = ξy + dz = 0 (1.39a)
0 ∂x
Z ∞
∂w
uy (0) = ξx − dz = 0 (1.39b)
0 ∂y
w(0) = 0 (1.39c)
25
As discussed in Davies and Carpenter [26], there are no issues in implementing the condi-
tion w → 0 as z → ∞ as the vanishing of w provides a boundary condition for the solution
of the Poisson equation (1.42b). It remains to constrain the other two primary variables
ξx and ξy in the far field, which is achieved in practise by enforcing vanishing conditions
at infinity. The validity of this approach is explained in detail in Davies and Carpenter
[26] and essentially reduces to the requirement that the z-derivatives of the three primary
variables vanish for z → ∞. In practise, it is possible to satisfy this condition by making
use of the coordinate transformation
l
η= (1.40)
z+l
which maps the semi-infinite physical domain z ∈ [0, ∞) to the computational domain
η ∈ (0, 1] for some stretching factor l. This mapping transforms the z-derivative operators
to
∂f η 2 ∂f
=− (1.41)
∂z l ∂η
and it is clear that the requirement that the z-derivatives of the primary variables vanish
for z → ∞ becomes a condition that the η-derivative remains bounded as η approaches
zero. Further discussion about this condition is given by Davies and Carpenter [26] but
for now we will assume it will be satisfied for our perturbation variables and proceed to
derive the corresponding formulation for two dimensional flows.
Reduction of the Full System to Two Dimensions
In two dimensions, consider a parallel base flow UB = UB (z, t). If, as for the three-
dimensional case, we let u and w denote the usual streamwise and wall-normal velocity
perturbation components and ξ the vorticity perturbation then the velocity-vorticity for-
mulation equations take the form
∂ ξˆ ∂ ξˆ ∂(ûω) ∂(ŵξ) 1 1 000

+U + + +wU 00 = ∇2 ξˆ + U (1.42a)
∂t ∂x ∂z ∂y R |R{z }
| {z }
non−linear non−parallel
∂ ξˆ
∇2 ŵ = − (1.42b)
∂x
where the single secondary variable û is given by

Z ∞
ˆ ∂ ŵ
û = − ξ+ dz
z ∂x
26
and we have written U 0 (z) := ∂U
∂z
. Linearisation can be obtained by dropping out the
terms with underbraces in the first of the above equations, equivalent to the linearisation
achieved by dropping the ξ × u in equation (1.35). The non-parallel term also indicated
by underbraces is removed under the parallel flow approximation. Having done this, we
arrive at the two-dimensional velocity-vorticity perturbation equations
∂ ξˆ ∂ ξˆ 1
+U + wU 00 = ∇2 ξˆ (1.43a)
∂t ∂x R
∂ ξˆ
∇2 ŵ = − (1.43b)
∂x
At this stage we divide our discussion into two parts to consider steady and temporally
periodic two-dimensional base flows separately, beginning with the steady case.
1.4.2 Two-Dimensional Steady Base Flow
Considering the base flow in question to be steady, we may assume a separable form for
the perturbation and, as usual, take a normal mode approximation of the form
ŵ = wei(αx−ωt) ξˆ = ξei(αx−ωt)
where α and ω have the same interpretations as in section 1.2.1. Substitution into the
perturbation equations (1.43) gives the system of two second order ordinary differential
equations
1 2
−iωξ + iαU ξ + wU 00 = (D − α2 )ξ (1.44)
R
(D2 − α2 )w = −iαξ (1.45)
d2
where D2 = dz 2
. Following a similar procedure to that described by Bridges and Morris
[14], we integrate the equations twice indefinitely and rearrange to get the polynomial
eigenvalue equations in α
ZZ ZZ ZZ Z Z
2 1 1 00
α ξdz +α i U ξdz + −iω ξdz − ξ + U wdz + a1 x + a0 = 0
R R
(1.46a)
ZZ Z Z
α i ξdz + α2 − wdz + w + b1 x + b0 = 0
(1.46b)
where a0 , a1 , b0 and b1 are arbitrary constants of integration.
27
Eigenvalue Solutions
Given a base flow UB , we are able to expand both the perturbation variables and the
base flow profile in terms of Chebyshev polynomials and derive matrix representations of
the operators ZZ ZZ
f dz and U f dz (1.47)
where f is an arbitrary perturbation variable. The boundary conditions are included

in the system by replacing the matrix rows that would otherwise determine the four
constants a0 , a1 , b0 and b1 . This can be interpreted as a form of the tau-method and a
similar method is described in Bridges and Morris [14]. The boundary conditions along
with the operator form of the velocity-vorticity perturbation equations thus give a system
of (2N × 2N ) equations of the form
    
A0 α2 + A1 α + A2 B0 α2 + B1 α + B2 ξ 0
   =  
C0 α2 + C1 α + C2 D0 α2 + D1 α + D2 w 0
where A, B, C and D should be clear from (1.46). Bridges and Morris [14] present two
methods for solving this polynomial eigenvalue equation, namely the companion and the
lambda matrix methods. Both of these are now largely obsolete and can be accomplished
by a simple call of MATLAB’s polyeig function.
Eigenvalue solution methods are often prone to spurious eigenvalues, those that are
not true eigenvalues of the perturbation equations. These spurious eigenvalues may be
attributed to the solution method of the problem and may be either stable or unsta-
ble. Stable spurious eigenvalues are of little importance because the stable eigenvalues
are rarely of interest, however spurious unstable eigenvalues are highly undesirable since
they could wrongly predict an instability. There are two types of spurious eigenvalues
which must be distinguished in order to be sure of their accuracy. The first are physi-
cally spurious eigenvalues, which are numerically computed eigenvalues in error because
of misapplication of boundary conditions or some other misrepresentation of the physics.
Boyd [13, p. 139] presents an example of how these physically spurious eigenvalues can
show up in a system and so the reader is referred to the references for details. The other
type of spurious eigenvalue are those which are numerically spurious. These are poor
approximations to exact eigenvalues because the mode may be oscillating too rapidly to
be resolved using a given discretisation, and can always be computed accurately using a
28
sufficiently large degree Chebyshev approximation. The physically spurious eigenvalues
can often be eliminated by using basic row and column operations on the matrix repre-
sentations of the integral operators. If the four boundary conditions are confined to the
top four rows of the matrix operators then we can apply row and column operations in
the following way
 
a0 ··· aN −3 aN −2 aN −1 aN
 
 b0 ··· bN −3 bN −2 bN −1 bN 
 
 
··· cN  = 0
 
 c0 cN −3 cN −2 cN −1
 
···
 
 d0 dN −3 dN −2 dN −1 dN 
 
··· ··· ··· ···
| {z }
 
0 ··· ãN −3 ãN −2 ãN −1 ãN
 
 0 ··· 0 b̃N −2 b̃N −1 b̃N 
 
 
→ 0 ···
 
0 0 c̃N −1 c̃N 
 

 0 ··· 0 0 0 ˜
dN 

 
··· ··· ··· ··· ··· ···
Neglecting these rows from consideration will give a non-singular matrix of order 2(N −
1) × 2(N − 1) in the polynomial eigenvalue problem and is shown in Boyd [13] and Bridges
and Morris [14] to eliminate physically spurious eigenvalues. The other variety, numer-
ically spurious eigenvalues, have very large magnitudes compared with genuine modes
and will usually scale like O(N 4 ) where N is the degree of the Chebyshev polynomial
approximation. This makes them easily distinguishable in practise, and any extremely
large magnitude eigenvalues are often irrelevant for stability calculations.
An alternative method of eliminating spurious eigenvalues is to use two different orders
of polynomial expansion and compare the results. The genuine eigenvalues are usually
picked up by both expansions provided the difference in discretisation order is large enough
to achieve genuine convergence. Both types of spurious eigenvalue may be noticed and
discarded in this case as they fail to converge as the orders of expansion are increased.
Local Eigenvalue Iteration
Given an appropriate initial guess αk , a cubically convergent method described in Bridges

and Morris [14] can be used to evaluate the eigenvalue α. We can iteratively improve the
29
guess by the formula
2f (αk )
αk+1 = αk − (1.48)
f (αk )2 − f1 (αk )
where
f (αk ) = Tr[D−1 (αk )D0 (αk )]
f1 (αk ) = Tr{D−1 (αk )D00 (αk ) − [D−1 (αk )D0 (αk )]2 }
and D0 (α) and D00 (α) are the derivatives of D with respect to α.
1.4.3 Application of Solution Method to the Blasius Boundary

Layer
Figure 1.3: Schematic of Blasius boundary layer flow configuration.
In the interest of validation and to provide an exemplar application of the new method,
we present now the solution procedure for the Blasius boundary layer configuration. This
configuration has several features in common with the rotating disk which will be the
main focus of this report. Since the formulation for the disk is particularly involved, this
case will illustrate the methods while keeping the algebra relatively accessible. Assume
we have a base flow UB = (Ux , 0, Uz ) which describes the steady flow over a semi-infinite
plate, forming the standard two dimensional Blasius boundary layer at the plate, as seen
30
in figure 1.3. Assume the free-stream velocity of the fluid is U∞ . Then from a standard
order of magnitude analysis, we can derive the boundary layer equations in dimensional
form which are given by
∂Ux ∂Uz
+ =0
∂x ∂z
∂Ux ∂Ux ∂ 2 Ux
Ux + Uz =ν
∂x ∂z ∂z 2
Defining a streamfunction ψ such that
∂ψ ∂ψ
Ux = , Uz = −
∂z ∂x
and substituting into the boundary layer equations gives
∂ψ ∂ 2 ψ ∂ψ ∂ 2 ψ ∂ 3ψ
− = ν
∂z ∂x∂z ∂x ∂z 2 ∂z 3
Following the standard derivation, introducing a similarity variable ζ such that

r
U∞ p
ζ=z , ψ = νU∞ xF (ζ)
νx
we arrive at the third order ordinary differential equation
1
F 000 + F F 00 = 0 (1.49)
2
subject to the boundary conditions arising from no slip and no penetration F (0) = F 0 (0) =
0 and F → 0 as ζ → ∞, derived using the relationship between ψ and F .
Due to the relationship between ζ, z and x, we can see that the base flow, expressed
in terms of the single similarity variable ζ, is UB = F 0 (ζ) where we have set U∞ = 1 and
scaled the other variables accordingly.
Numerical Solution of the Base Flow Equations
For the stability analysis which follows in the subsequent sections based on the velocity-
vorticity formulation described previously, we are required to calculate the coefficients of
the base flow when expanded as a Chebyshev series. Thus, we require the solution of the
base flow equation (1.49) evaluated at the Chebyshev collocation points in (0, 1] given by
equation (1.18). Since the Blasius configuration is naturally defined on the semi-infinite
domain [0, ∞), we must first map our equations to the computational domain (0, 1] by
means of the coordinate transformation (1.40).
31
This choice of domain is used primarily because of the symmetries enforced on the
disturbance variables and the vanishing condition of the z-derivatives as discussed in
section 1.4.1 and in Davies and Carpenter [26]. The standard Chebyshev domain [−1, 1]
could also be used, with the different mapping
ζ −l
η̂ =
ζ +l
as discussed in references such as Cooper and Carpenter [19] but in the interest of consis-
tency and the advantage of less explicitly imposed boundary conditions by splitting the
odd and even cases separately we use only the mapping (1.40) for the remainder of this
report.
The choice of l is facilitated by a wish to balance the degree of the Chebyshev approx-
imation with computational time and achieve a high level of accuracy in the shortest time
possible. It was found by Davies and Carpenter [26] that in the case of the time-dependent
simulations of rotating disk, the variation of the primary perturbation variables across
the boundary layer could be fully resolved using a Chebyshev expansion involving N = 48
polynomials. Although not explicitly mentioned in the reference, they also found that the
most appropriate choice of l for computational efficiency was l ≈ 4.
The base flow equations (1.49) can be solved fairly simply using MATLAB’s bvp4c
boundary value solver, and evaluated at the collocation points. It is worth noting that
out-of-the-box, bvp4c has issues with directly solving the equation on a large physical
domain. In fact, it is numerically unstable over [0, a] for a > 14. This was circumvented
by introducing an iterative scheme which used the solution over [0, K] as an initial guess
for the solution over [0, K + 1] and proceeded for K ≥ 1. The resulting base flow profile
is shown in figure 1.4.
Boundary Layer Thickness
An additional scaling factor must be taken into account when calculating derivatives of
the Blasius boundary layer base flow profile. Consider z ∗ to be the non-dimensional wall-
distance in the Blasius configuration and the non-dimensional displacement thickness δ
defined in the classical way as
Z ∞
UB (η)
δ= 1− dζ
0 U∞
32
1
0.8
0.6
0.4
0.2
0
0 2 4 6 8 10
Figure 1.4: Blasius base flow profile UB for ζ ∈ [0, 10].
Defining the dimensional displacement thickness as δ ∗ =

p νx
U∞
, we get
z z δz
z∗ = =⇒ η = ∗ which gives η = ∗ = Cz for constant C
δ δ δ
Thus, since UB = F 0 and U∞ = 1, we have

Z ∞
δ
C= ∗ = (1 − F 0 )dζ
δ 0
and by the chain rule

dUB d2 UB
= CF 00 , = C 2 F 000 (1.50)
dz dz
Chebyshev Coefficients of the Base Flow
As will be the case for the rotating disk study of chapter 2, the base flow UB and perturba-
tion variables ξ and w are expanded in terms of either odd or even Chebyshev polynomials.
It is acceptable to split the parities in this way since we are only working with half of
the usual Chebyshev domain under the mapping (1.40). If an odd representation were
to be used then it is assumed implicitly that the function decays at η = 0 since all odd
Chebyshev polynomials have this property. This is equivalent to the condition that the
flow profiles in physical space decay as z → ∞. Since the base flow velocity UB does
not decay at infinity, it is sensible to use an even Chebyshev expansion for this variable
and its second derivative. The Chebyshev coefficients of the base flow variable and its
33
derivatives can be calculated from the collocation values using the FFT method described
in section 1.3.1.
Perturbation Equations
After applying the mapping (1.40), the derivative operators become
−η 2 d d2 η3 d2

d d
= , = 2 2 +η 2
dz l dη dz 2 l dη dη
and so the perturbation equations (1.44) after taking the usual normal mode approxima-
tion become
η3 d2

00 d
−iωξ + iαU ξ + wU = 2 2 + η 2 ξ − α2 ξ (1.51)
Rl dη dη
3 2

η d d
2
2 + η 2 w − α2 w = −iαξ (1.52)
l dη dη
while the integral definition of the secondary variable u becomes

Z η
ξ w
u=l + iα 2 dη̃
0 η̃ 2 η̃
Boundary Conditions
As described in section 1.4, the boundary conditions (1.39) imposed on the system are
derived from the no-slip condition at the wall and the integral constraint imposed on the
vorticity. Under the mapping given by equation (1.40), these become
w(1) = w(0) = ξ(0) = 0

Z 1
ξ 1 ∂w
+ dη = 0
0 η 2 η 2 ∂x
Since we are now working with the odd and even Chebyshev representations separately,
the vanishing conditions as z → ∞, or equivalently η → 0, are assumed implicitly if we
expand the perturbation variables in terms of odd Chebyshev polynomials. Thus, if we
apply the normal mode approximation, the two remaining boundary conditions are
w(1) = 0
Z 1
ξ iαw
+ 2 dη = 0
0 η2 η
The second boundary condition can be implemented by means of a tridiagonal matrix

operation, details of which are given in the appendices of Davies and Carpenter [26] and
34
in Thomas [71]. However, we may simplify things slightly by dividing the perturbation
equations by η 2 , provided we take care when doing so. We could thus, in principle, work
entirely with new variables defined by fˆ(η) = f (η)
η2
for an arbitrary f , so that the integral
condition becomes Z 1
ξˆ + iαŵ dη = 0
0
which is simple to implement numerically as it only involves the first integral of a Cheby-
shev representation. The apparent singularity at η → 0 caused no problems in practise,
and validation tests were carried out, both for the eigenvalue problem and the time-
dependent simulations of later sections to ensure that this division did not alter any
results. The expectation that this division is non-consequential may be justified by anal-
ysis of the perturbation quantities. Typically, for an arbitrary perturbation quantity f ,
we have f ∼ e−az as z → ∞. Thus, under the mapping (1.40), we have
f 1 −al
∼ e η2 as η → 0 (1.53)
η2 η2
which decays to 0 as η → 0 because of the dominance of the exponential. Thus, we
f
should expect η2
to be well behaved as η approaches zero, precisely what was observed in
practise. Agreement was found between the divided and non-divided versions in all cases
tested.
Thus, this division by η 2 in the perturbation equations and application of the mapped
derivative operators gives
∂2

ˆ ˆ 00 η ∂ ˆ + α2 ξˆ = 0
−iω ξ + iαU ξ + U ŵ − 2 2 + η 2 (η 2 ξ) (1.54)
Rl ∂η ∂η
∂2

η ∂
2 + η 2 (η 2 ŵ) − α2 ŵ + iαξˆ = 0 (1.55)
l2 ∂η ∂η
Following the same method as described in section 1.4 and integrating indefinitely with
respect to the mapped variable η gives

1 1
α 2
I2 + α(iI2 UB ) + − 2 J − iωI2 ξˆ + [I2 UB00 ]ŵ = 0 (1.56)
R l R

1
[α(iI2 )]ξˆ + α (−I2 ) + 2 J ŵ = 0
2
(1.57)
l
where ZZ ZZ
1 2 2
I2 f := f dη and Jf = D (η f (η))dη (1.58)
η2
∂ ∂ 2
with D2 := 6η 2 + 6η 3 ∂η + η 4 ∂η 2 . After integration by parts, we get
Z
Jf = η f − 2 η 3 f dη
4
(1.59)
35
Operator Representations in Matrix Form
As mentioned previously, we will expand the perturbation variables in terms of odd Cheby-
R
shev polynomials, and using the matrix representation of , we can derive operators which
RR
approximate I2 , J and UB p for an arbitrary perturbation variable f . Thus, if
X
f= fn T2n−1
n
then by the representation (1.32), we can integrate f twice and equate coefficients of T2n−1
to get

1X fn−1 fn fn+1
If = aT0 + bT1 + − + T2n−1
8 n (2n − 1)(n − 1) n(n − 1) n(2n − 1)
where a and b are arbitrary constant of integration. Consider now the operator J given by
equation (1.59). Expanding the multiplicative parts η 4 and η 3 in terms of the Chebyshev
polynomial basis, we get
1
η 4 = (T4 + 4T2 + 3T0 )
8
1
η 3 = (T3 + 3T1 )
4
which gives
" Z #
1 1 X 1 X
Jf = (T4 + 4T2 + 3T0 ) fn T2n−1 − 2 (T3 + T1 ) fn T2n−1
l2 8 n
4 n
Finally, making use of the identity
2Tm (η)Tn (η) = Tm+n (η) + Tm−n (η) (m ≥ n ∈ Z) (1.60)
we get
X 1
Jf = cT0 + dT1 + fk−2 + 4fk−1 + 6fk + 4fk+1 + fk+2
n
16

2
− fk−2 + 2fk−1 − 2fk+1 − fk+2 T2k−1
2k − 1
which is the pentadiagonal matrix form of the operator J.

Thus, it remains to evaluate the product operators of the type
ZZ
UB f dη (1.61)
36
The method outlined below is discussed in Bridges and Morris [14] and can be applied to
any base flow profile desired. The method is presented only for the case of no enforced
parities, although similar formulae can be derived for the odd and even cases separately.
Expanding some arbitrary base flow profile UB (η) and the disturbance variable f (η) in
terms of full Chebyshev representations gives
N
U0 X
UB (η) = + Un Tn (η)
2 n=1
N
f0 X
p(η) = + fn Tn (η)
2 n=1
where the first term is halved for convenience as discussed previously. Thus, again using
the identity (1.60), we can deduce that
N
c0 X
UB (η)p(η) = + cn Tn (η)
2 n=1
where
N
fn U0 1 X
cn = + (fm+n + f|m−n| )Um , n = 1, . . . , N
2 2 m=1
This gives a matrix representation for the Chebyshev coefficients of the product UB p
which may then be integrated by multiplication with the matrix integral operators defined
previously.
Numerical Results
Having derived matrix operators for the constituent parts of the perturbation equations,
we form a matrix dispersion relation
D(R, α, ω) = 0 (1.62)
which can be solved in either a spatial or temporal setting by prescribing either ω or α

respectively, as described in section 1.2.1.
This dispersion relation is solved in MATLAB and the numerical eigenvalue spectrum
for R = 500 and α = 0.2 is plotted in figure 1.5. The order of Chebyshev discretisation was
set to be N = 64 to ensure convergence and appendix A.1 shows the convergence prop-
erties for varying Chebyshev discretisation orders. The diagram shows the numerically
spurious eigenvalues which are identifiable by their sensitivity to the order of discretisa-
tion and the genuine critical eigenvalue at ωr ≈ 0.0730. This figure can be compared
favourably with a similar one in Schmid and Henningson [65].
37
0.2
-0.2
-0.4
-0.6
-0.8
-1
0 0.05 0.1 0.15 0.2
Figure 1.5: Discrete approximation to the continuous spectrum of the Blasius boundary
layer dispersion relation for R = 500 and α = 0.2. The critical eigenvalue is shown as a
filled in circle, while the × and ◦ show the numerical spectra for N = 64 and N = 96
respectively.
1.4.4 Linear Stability of the Temporally Periodic Stokes Layer
Since the main part of the rotating disk study in chapter 3 of this thesis deals with
undisturbed basic flow profiles with periodic time-dependencies, we present in this sec-
tion an overview of the machinery required to deal with this modification. We will use
the quintessential purely oscillatory flow that is generated when a flat plate oscillates in
the plane beneath a semi-infinite expanse of fluid, namely the Stokes layer, to provide this
illustration. Time-periodic oscillatory flows occur in many types of physical and physio-
logical processes and in particular, Womersley [83] shows that high frequency oscillatory
blood flow in an artery can be described by invoking a Stokes layer adjacent to the wall,
along with a region of inviscid flow at the centre of the artery. The stability of the model
plane Couette and plane Poiseuille problems for time-periodic flows has been thoroughly
reviewed by Davis [29] in one of the earliest pieces of work on the subject, and several au-
thors including Blennerhassett and Bassom [10], Blennerhassett and Bassom [11], Thomas
et al. [76] and Ramage [59] have since extended the understanding of the stability of such
flows. In order to introduce the methods employed in this report to analyse the stability
38
of the periodically modulated rotating disk boundary layer, we present first the derivation
of the formulation applied to the semi-infinite flat Stokes layer, as previously studied by
the aforementioned authors. While the flow configuration presented here is identical to
previous work, the formulation which will be outlined in the following section is distinctly
different from that which appears in previous literature and has not, to the best of my
knowledge, been discussed previously. Since the basic flow under consideration is now
time-dependent, we cannot use the eigenvalue methods employed thus far in this report,
and so we begin with an overview of two different approaches for investigating the stabil-
ity of time-periodic flows; the instantaneous or frozen-flow approximation and a method
related to Floquet theory. In particular, given this time-dependence, we cannot readily
utilise the standard normal mode form, and thus at this stage, a flow variable φ can only
be decomposed as
φ(x, z, t) = φ̂(z, t)eiαx (1.63)
where the temporal dependence of the solution is still held in φ̂.
Stokes Layer Base Flow
Consider the Stokes layer that is generated by the motion of an infinitely long and flat
rigid wall located at z = 0, which oscillates in its own plane with a velocity U0 cos(ϕt)
beneath an unbounded body of viscous fluid that would otherwise remain stationary. The
q
boundary layer thickness is described by δ = 2νϕ
and we can define a Reynolds number
associated with the flow as
U0 δ U0
R= =√
2ν 2νϕ
q
2ν
If all lengths are scaled on the boundary layer thickness δ = ϕ
and we introduce a
non-dimensional time τ = ϕt then in the absence of disturbances, the basic Stokes flow
is two-dimensional and given by
UB (z, τ ) = e−z cos(τ − z), VB (z, τ ) = 0
We can study the development of disturbances to the basic flow in the velocity-vorticity
formulation described in section 1.4.3 by introducing, as usual, perturbations of the form
(U, V ) = (UB , 0) + (u, v), Ξ = ΞB + ξ (1.64)
39
where ΞB = UB0 is the undisturbed vorticity field associated with the base flow UB . Due
to the chosen non-dimensionalisation, this gives the slightly different linearised velocity-
vorticity formulation for this flow configuration as
2
∂2

1 ∂ξ ∂ξ 00 1 ∂
+ UB + UB w = + ξ (1.65a)
R ∂τ ∂x 2R ∂x2 ∂z 2
2
∂2

∂ ∂ξ
2
+ 2 w=− (1.65b)
∂x ∂z ∂x
with the secondary variable u being defined by the integral relation

Z ∞
∂w
u=− ξ+ dz
z ∂x
Instantaneous Frozen-Flow Approximation
Instantaneous instability theory, as explained in Luo and Wu [48], can be characterised by

freezing the base flow at a specific time instant and treating it as if it were steady. This
allows for the time to be treated as a parameter, and has the advantage of allowing an Orr-
Sommerfeld equation to be considered at each time instant. Using the Orr-Sommerfeld
solutions at each time interval leads to a prediction of an instantaneous growth rate at each
point in the cycle of the base flow. We can formalise this by considering a decomposition
of (1.63) into the form
φ(x, z, t) = φ̂(z; t)ei(αx−ωt) (1.66)
∂ φ̂
where φ̂ is a function which is slowly varying in time, allowing us to neglect the ∂t
term
that arises in the perturbation equations after substitution.
While this is an attractive method, it must be used with caution as it can only be
strictly justified in the limit R → ∞ and may give misleading results when applied to the
necessarily finite Reynolds numbers that we consider. The modal interaction which will be
discussed in the forthcoming section is also not accounted for by this approach, and while
the frozen-flow method is useful for tracking a single mode across a period, it cannot fully
resolve the disturbance structure. Therefore a more appropriate and rigorous method
for studying the stability of periodic time-dependent base flows is to be taken within
the framework of Floquet theory. The study by Luo and Wu [48] attempted to explain
the relationship between the results of an instantaneous approach and the more rigorous
Floquet theory and found that for a given wavenumber, numerical simulation results from
the instantaneous approach agreed well with those from Floquet theory over significant
40
parts of the cycle. However, there are certain intricacies in the stability properties that
are not accounted for by these approaches and we leave our discussion of them for now in
favour of formulating the Floquet analysis in detail. We will return to the instantaneous
approach with our discussion of the modulated rotating disk, although the approaches of
Luo and Wu [48] were used in preliminary work as a numerical validation of the code used
in the rotating disk scenario.
Floquet Theory
In order to illustrate the concepts of Floquet theory in a general light, we first make some
standard definitions relating to ordinary differential equations and provide a theoretical
background to the methods involved. Since we are interested in general in the theory
from the standpoint of hydrodynamic stability, we omit any laborious analytical detail in
the derivations of theorems. However it should be noted that Floquet theory is a very
well developed area of mathematics and extended details are available for the interested
reader in Jordan and Smith [41].
From a very general standpoint, consider only the system of ordinary differential equa-
tions
u0 (t) = A(t)u(t), (t ∈ R)
for some matrix A(t). Then we can define the fundamental matrix of the system as follows.
Definition. (Fundamental Matrix) Take some t0 ∈ [a, b] ⊆ R and let A : [a, b] → Rn,n be
continuous. Then the matrix-valued function Φ : [a, b] → Rn,n such that
Φ0 = AΦ, Φ0 (t0 ) = I
is called the t0 -canonical fundamental matrix of the ordinary differential equation u0 = Au.
Here, I represents the identity matrix.
Assume now that we impose an additional constraint on A(t), and require that A is
T -periodic. Then we have A(t + T ) = A(t) and it can be shown as in Jordan and Smith
[41] that
Φ(t + T ) = Φ(t)Φ(T )
The usefulness of this result can be seen in Floquet’s theorem, which is stated below.
41
Theorem. (Floquet’s Theorem) The system u̇ = A(t)u, where A is an n × n matrix-
valued function with minimal period T , has at least one non-trivial solution u = P(t) such
that
P(t + T ) = λP(t), t ∈ R
where λ is a constant.
The constants λ are eigenvalues of the matrix M = Φ−1 (t0 )Φ(t0 + T ) and are called
the Floquet multipliers for the periodic linear system. The corresponding eigenvectors are
called the Floquet solutions.
A final result can be shown which will provide a convenient format for our discussion
of stability.
Theorem. If M = Φ−1 (t0 )Φ(t0 + T ) has N distinct eigenvalues µi , then the system
u0 = A(t)u admits N linearly independent solutions of the form
ui (t) = pi (t)eµi t
where pi has minimal period T and we have set λi = eµi t . The constants µi are called the
Floquet exponents of the system.
This framework provides the most convenient format for the study of disturbances
in time-periodic flows and since both the Stokes layer and the modulated rotating disk
system are inherently periodic, we will utilise these ideas to attempt to provide an insight
into the stability characteristics. As previously alluded to, several studies of the Floquet
stability properties of the semi-infinite Stokes layer, such as Hall [39], Blennerhassett and
Bassom [11, 10], Thomas and Davies [73] and Ramage [59] exist and the results presented
in the following section are not intended to be novel. However, I am unaware of any
Floquet eigenvalue solver which exists for the velocity-vorticity formulation of Davies and
Carpenter [27] and in the interest of algebraic brevity and as numerical validation, we
present the formulation of the Floquet problem in detail, and outline the methods used
by previous authors for the Stokes layer study.
Derivation of the Floquet Stability Equations
To consider the development of small disturbances to the unsteady base flow UB =

e−z cos(τ − z), we introduce a perturbation and write
u(x, τ ) = UB (z, τ ) + u0 (x, τ ), ξ(x, τ ) = Ξ(z, τ ) + ξ 0 (x, τ )
42
Following Hall [39], while making use of the theorems discussed previously, we can take
a slightly different normal mode solution than that of the steady case to incorporate the
temporal dependence. This normal mode solution is taken to be of the form
f (x, z, τ ) = fˆ(z, τ )eµτ eiαx + c.c. (1.67)
where fˆ(z, τ ) is the now time-periodic primary variable with the same period as that of
the wall oscillation. All exponential growth of fˆ(z, τ ) is incorporated into the Floquet
exponent eµτ and the term c.c. denotes the complex conjugate, which is added to ensure
that the disturbance is real. Clearly the quantities of interest here will be R(µ) and I(α)
as these specify the temporal and spatial growth rates of the disturbance and whether or
not we have temporal and spatial stability.
Substituting this normal mode approximation into (1.65a), multiplying through by R
and rearranging gives
α2 1 ∂ 2

µ + iαRUB + − ξ + RUB00 w = 0 (1.68a)
2 2 ∂z 2
∂2

2
iαξ + −α + 2 w = 0 (1.68b)
∂z
The boundary conditions on the system, arising from no-slip and no-penetration are given
by the representation of the secondary variable u and become
Z ∞
w(0) = 0, (ξ + iαw) dz = 0 (1.69)
0
Since these equations are now time dependent, the local ideas of stability analysis for the
steady case are not strictly applicable and we must approach the problem from a slightly
different angle. The following method was derived by Hall [39], and first successfully
implemented by Blennerhassett and Bassom [11] for the Stokes layer and involves the
decomposition of the perturbation variables into harmonics such that
∞
X
fˆ(z, τ ) = fn (z)einτ (1.70)
n=−∞
A representation for the time dependent part of the base flow is also required in a similar
form, and this is given, by definition of cosine, by
UB = e−z cos(τ − z) = u1 eiτ + u2 e−iτ (1.71)
43
where u1 = 12 e−(1+i)z and u2 = ū1 . To illustrate the method of Hall [39], we substitute
(1.70) into (1.68) to get
∞
α2 1 ∂ 2
X
iτ −iτ −iτ
µ + ik + iαR(u1 e + u2 e ) + − 2
iτ
ξk + R(2iu1 e + 2iu2 e )wk eikτ = 0
k=−∞
2 2 ∂z
∞
∂2
X
2
iαξk + −α + 2 wk eikτ = 0
k=−∞
∂z
and comparing coefficients of eikτ for each k gives
α2 1 ∂ 2

µ + ik + − ξk + iαR (u1 ξk−1 + u2 ξk+1 ) + R (2iu1 wk−1 + 2iu2 wk+1 ) = 0
2 2 ∂z 2
∂2

2
iαξk + −α + 2 wk = 0
∂z
This gives a system of eigenvalue equations which can be solved in a similar way to the
steady case. As has been done in previous sections, we introduce the mapping (1.40) to
map the physical domain [0, ∞) to the computational domain (0, 1], divide through by η 2
and integrate twice with respect to η to get
Lk1 ξˆk + Mξˆk−1 − M̄ξˆk+1 + P ŵk−1 − P̄ ŵk+1 = −µI2 ξˆk (1.72a)
iαI2 ξˆk − α2 I2 + J2 ŵk = 0

(1.72b)
where
α2 1
Lk1 = I2 − J2 + ikI2 , M = iαRI2 u1 , P = 2iRI2 u1 (1.73)
2 2
For consistency with the odd Chebyshev expansions of the primary variables, the base flow
profile and its second derivative are also represented in terms of odd polynomials, thereby
implicitly, and physically appropriately, assuming decay as z → ∞. Using the integral
representation (1.32), we can create an infinite system of generalised matrix eigenvalue
problems of the form
Lk (α, R)fˆk = −µI2 fˆk (1.74)
where the boundary conditions are included by replacing the rows which would otherwise
determine the arbitrary constants of integration. Clearly, to enable a computational
implementation of this method, we must truncate this system above and below by some M ,
giving a (2M N +M )×(2M N +M ) where N is the degree of Chebyshev discretisation. This
system is assembled and solved using MATLAB’s sparse matrix routine eigs. As explained
in Blennerhassett and Bassom [11], this formulation only determines µi modulo unity.
44
Thus, for each eigenvalue µ, we can expect µ ± ik to also be an eigenvalue. Additionally,
left and right propagating waves with the same growth rate are possible, implying the
conjugate µ̄ should also be an eigenvalue. These symmetries allow us to restrict the search
for the most unstable mode to the interval µi ∈ [0, 12 ].
Thus, using the integral representation for Chebyshev polynomials as discussed in
detail in section 1.3.2, the full matrix form of this equation becomes
    
−M
L −M̄ 0 ··· ··· ··· ··· 0 ξ R
 1   −M 
. .
 
 0 M L−M −M̄ 0 · · · · · · 0   ..   .. 
    
1
  ..   .. 
    
.. .. ..
 0 0 . . . 0 ··· 0  .  .

    
M
··· ··· · · · · · · 0 M L1   ξM  R
     
 0
   = µ 
−P̄ · · · · · · · · · · · · 0  w−M 
    
 0 0 0
  ..   .. 
    
P −P̄ 0 ··· ··· 0  . 

 0 0 .
.  .. 
    
.. .. ..
. . . 0 · · · 0   .. 
   
 0 0 .
    
0 ··· ··· ··· ··· 0 P 0 wM 0
Results and Validation
Eigenvalue solutions for certain parameters examined in previous literature are sum-
marised in table 1.1. In each case, µ is calculated from (1.74) for prescribed α. While
in principle α can be calculated for prescribed µ, in practise the number of harmonics
required led to prohibitive computational requirements at this time.
Evidently for the cases tested, the disparity between the results obtained via the
method described above and those from previous studies is negligible. It should be noted
that Blennerhassett and Bassom [10] and Ramage [59] were conducting their studies in a
confined channel with a large width, so the small disparity between the truly semi-infinite
case and theirs is expected. Table 1.2 shows the variation in µ depending on how many
harmonics are taken in the truncation of the infinite series. As can be seen from the
table, this method unfortunately does not improve on the bound set by Blennerhassett
and Bassom [11] of M ≈ 0.8αR. Figures 1.6 and 1.7 show the temporal eigenvalue spectra
for stable and unstable parameter choices, and also illustrate the symmetry properties in
µ as discussed in the previous section.
45
R α Reference µ
Blennerhassett and Bassom [11] 0.08174 + 0.35096i
800 0.3
Ramage [59] 0.08238 + 0.34582i
Current Work 0.081743+0.35096i
847.5 0.38
Ramage [59] 0.67620 + 0.14880i
Current Work 0.675937 + 0.148056i
Table 1.1: Comparison between eigenvalues computed from the dispersion relation (1.74)
for the semi-infinite Stokes layer in this study against identical parameter sets in the
literature. Excellent agreement is found in all cases tested.
Variation of µ with Number of Harmonics M

R α 0.8αR M µ
48 -0.045047-2.0829e-06i
64 0.30537+0.49254i
128 0.63373+0.1559i
800 0.3 192
156 0.19995+0.47735i
192 0.081743+0.35096i
256 0.081743+0.35096i
64 0.19942+0.46809i
128 1.2531-0.23944i
156 1.2797-0.30737i
847.5 0.38 257.64
192 1.0066-0.1865i
256 0.675937 + 0.148056i
300 0.675937 + 0.148056i
Table 1.2: Variation in the temporal eigenvalue µ calculated from the dispersion relation
(1.74) for varying harmonics. Results agree well with the assertion by Blennerhassett and
Bassom [11] that M & 0.8αR for convergence.
46
2
-1
-2
-1 -0.5 0 0.5 1
Figure 1.6: Eigenvalue spectra for the dispersion relation relating to the semi-infinite
Stokes layer for a stable configuration with R = 700 and α = 0.3.
-1
-2
-1 -0.5 0 0.5 1
Figure 1.7: Eigenvalue spectra for the dispersion relation relating to the semi-infinite
Stokes layer for an unstable configuration with R = 800 and α = 0.3. The unstable
eigenmodes are visible as those with µr > 0.
47
1.5 The Neutral Stability Curve
Having read several introductory texts, lecture courses and articles relating to hydrody-
namic stability theory, it is my belief that the fundamental concepts of the neutral stability
curve are not particularly well documented in standard literature and I have been unable
to find many formulaic approaches to calculating them. Although they appear in some
form in most papers and articles, the method by which they are obtained is rarely, if ever,
presented. The goal of the following section is to introduce an otherwise unfamiliar reader
to this concept, and describe in detail the methods used in this report for calculating neu-
tral stability curves. We will use the steady, two-dimensional case of section 1.4.3 for our
exposition, although similar concepts apply to each scenario considered in this document.
The neutral curve is meant as a tool for quickly identifying linearly stable and unstable
regions in the parameter space of a flow. They are usually displayed with the Reynolds
number along the horizontal axis and the streamwise wavenumber α on the vertical. They
can similarly be displayed showing the temporal frequency ω on the vertical although this
is less common. The neutral curve for the Blasius flow configuration is shown in figure
1.8.
0.4
0.3
0.2
0.1
0
0 1000 2000 3000 4000 5000
Figure 1.8: Neutral stability curve for the Blasius boundary layer.
This illustrates that inside the curve, the flow is linearly, convectively unstable with
respect to infinitesimal perturbations. Similarly, outside the region, the flow is stable with
48
respect to infinitesimal perturbations. If the normal mode approximation were taken to
be
φ(x, z, t) = φ̂(z)ei(αx−ωt)
where α = αr + iαi , then the region inside the curve would correspond to αi < 0, while
the region outside the curve would correspond to αi > 0. The curve itself is precisely the
line of αi = 0. Similar concepts apply for the temporal curves, with the signs reversed
due to the negative sign appearing in front of ω in the normal mode solution. Either way,
regardless of whether a temporal or spatial analysis is conducted, the curve corresponds
to αi = ωi = 0.
It remains to formulate an algorithm for recursively calculating points on the curve
given an initial starting guess. A simple algorithm is described very briefly in Bridges
and Morris [14] but for complex curves with many corners and turns, a more effective
technique is required. For our purposes, this technique is a form of pseudo-arclength
continuation similar to that described in Dickson et al. [31].
1.5.1 Arclength Continuation
Consider a parameter dependent non-linear equation of the form
F (u, λ) = 0 (1.75)
as λ varies. Assuming that we know the solution u0 for a particular value of λ0 , then
simple parameter continuation [31] looks to solve the corresponding problem
F (u0 , λ0 + δλ) = 0 (1.76)
using the known solution as an initial starting guess. We can interpret this for our case
as knowing
αi (ω0 , R) = 0 (1.77)
and looking to solve the corresponding problem
αi (ω0 , R0 + δR) = 0 (1.78)
having solved the global eigenvalue problem for some initial parameter pair (ω0 , R0 ). The
perturbed problem (1.78) can be solved using Newton iteration from the base problem
(1.77) and taking δR sufficiently small can fully resolve a neutral curve until a singularity
49
is encountered. These singularities occur precisely when the Jacobian Fu is singular, or
∂αi
in our case when ∂ω
= 0. In practical terms, this occurs at a turning point of the neutral
curve and unfortunately, this simple method cannot deal with the more complex neutral
curves with many turns and corners without running into difficulty. This can be dealt
with by a slightly more complex method described in Dickson et al. [31], termed arclength
continuation. The major advantage of arclength continuation over simple parameter con-
tinuation is that the choice of δR is not a pre-defined input into the algorithm and so does
not rely on any a priori idea about the curvature of the line. As explained in Dickson
et al. [31], one way to remedy the failure of parameter continuation at singularities is
to introduce an approximate arclength parameter s, so that both ω and R depend on
s. This idea is known as pseudo-arclength continuation. We will present the method in
our specific case for the neutral stability curve, although Dickson et al. [31] gives a more
general approach involving arbitrary non-linear operators.
Assuming that both ω and R depend smoothly on s, we can differentiate αi (ω, R) = 0
with respect to s to give

dαi (ω(s), R(s)) ∂αi ∂αi
= ω̇ + Ṙ = 0 (1.79)
ds ∂ω ∂R
where the dot notation denotes differentiation with respect to s. Since s is the arclength,
we must also have
||(ω̇, Ṙ)||2 = |ω̇|2 + |Ṙ|2 = 1 (1.80)
which can serve as an extra equation to supplement the original equation αi (ω, R) = 0 and
fully specify the unknowns αi and s. We can approximate condition (1.80) by introducing
 
ω̇0
N (ω, R, s) =   (ω − ω0 , R − R0 ) − ds = 0 (1.81)
Ṙ0
which says that new point, (ω, R), on the path lies on a hyperplane orthogonal to the
tangent vector through the current point (ω0 , R0 ), and the intersection of that hyperplane
with the tangent vector is a distance ds from (ω0 , R0 ). We can thus extend the system
(1.77) to  
 
α (ω, R) 0
 i =  (1.82)
N (ω, R, s) 0
which can be solved for the new point (ω, R) while specifying only the arclength ds. Note
that this is not equivalent to specifying an increment in either ω or R and does not rely
50
on any pre-determined notion of the direction of the curve. Ramage [59] gives a thorough
overview of this process.
In order to compute the N term given by equation (1.81), we require values of the
˙ These can be derived using a combination of the normalisation
derivatives ω̇ and Re.
requirement (1.80) and the original dispersion relation
D(α; ω, R)φ = 0 (1.83)
By equation (1.79), we see that

∂α
− ∂R Ṙ
ω̇ = ∂α
(1.84)
∂ω
p
and since |ω̇|2 + |Ṙ|2 = 1, we can thus specify R as R = ± 1 − |ω̇|2 . The sign of R
determines which direction along the curve we are travelling using the method, either
in the direction of increasing or decreasing R. Having derived these formulae, we may
∂α ∂α
compute the remaining requirements ∂ω
and ∂R
by making use of the dispersion relation
∂α
(1.83). We present only the case for obtaining ∂ω
but the other is essentially identical.
Differentiating (1.83) with respect to ω gives
∂D ∂φ
φ+D =0 (1.85)
∂ω ∂ω
A similar dispersion relation must hold for a left eigenvalue ϕ† such that
ϕ† D(α; ω, R) = 0 (1.86)
where † denotes the complex conjugate transpose. Multiplying through by ϕ† in equation

(1.85) eliminates the second term to give
∂D
ϕ† φ=0 (1.87)
∂ω
Finally, since ϕ† φ = I, we have

∂D
∂α ∂α ∂D ∂ω
= = ∂D
(1.88)
∂ω ∂D ∂ω ∂α
and
∂α ∂α ϕ† ∂D φ
= ϕ† φ = † ∂D
∂ω
(1.89)
∂ω ∂ω ϕ ∂α φ
∂D ∂D
The quantities ∂ω
and ∂α
are available directly from the perturbation equations, and the
eigenvectors φ and ϕ† are determined quickly in practise by recursive iteration from a
non-zero guess.
51
Chapter 2
Local Linear Stability Analysis of

the Steady Rotating Disk Boundary
Layer
2.1 Introduction
Figure 2.1: Rotating disk flow configuration.
52
This following chapter will detail the application of the numerical methods described
in chapter 1 to a canonical three-dimensional steady boundary layer, and discuss the
modifications to the techniques required therein. We begin by considering a disk of
infinite extent rotating at a constant rate beneath an otherwise stationary fluid such as
air or water. The motion of the disk creates a thin boundary layer on the surface where
the flow is directly affected by the rotation. Viscous stresses act to drag the fluid near
the disk in circular paths but the absence of a pressure gradient to hold fluid elements
causes them to spiral outwards, to be replaced by an axial flow downwards towards the
disk. The resulting flow has radial, azimuthal and vertical components and admits an
exact similarity solution to the Navier-Stokes equations, first derived by von Kármán [80]
in 1921.
For this reason, the flow as described over a rotating disk is often referred to as von
Kármán flow, and provides a canonical example of a three-dimensional boundary layer.
Boundary layers over swept wings also have flow components in two directions - along
the span of the wing as well as in the direction of flight, and the crossflow inflexion point
instability mechanism is common to both the rotating disk boundary layer and the flow
over a swept wing. Thus the investigation of strategies for controlling the behaviour
of disturbances that develop in the rotating disk flow may prove to be helpful for the
identification and assessment of technologies that have the potential to maintain laminar
flow over swept wings.
Additionally, the von Kármán problem has certain useful practical advantages over
the swept wing configuration; it admits an exact similarity solution of the Navier-Stokes
equations and is more amenable to experiments. However, it is important to note that
there are differences between the two flow configurations. The boundary layer over a swept
wing is not affected by Coriolis forces whereas the rotating disk layer is and the former
does not include the azimuthal periodicity of the rotating disk configuration. We will
review the implications of these differences later when discussing the stability properties
of the rotating disk configuration.
As stated by Lingwood and Alfredsson [46], the rotating disk flow and its related
flows are relevant not only to swept wing flows but to a wider range of complex three-
dimensional configurations such as atmospheric and oceanic flows, rotating-cavity flows
and computer storage devices. Additionally, in electrochemistry, the rotating disk elec-
53
trode is a common instrument utilised for performing a technique known as hydrodynamic
voltammetry. In part, it is the simplicity of the base flow equations that has made the
von Kármán flow an attractive candidate for studies of some of these more general three-
dimensional boundary layers.
2.1.1 Numerical Solution of the Base Flow Equations
In 1921, Theodore von Kármán formulated the exact similarity solution to the Navier-
Stokes equations for the steady laminar flow over a rotating disk of infinite radius. We will
provide an overview of this solution method while adding the modifications required to
deal with a non-constant rotation rate. Following von Kármán’s method, let U ∗ , V ∗ and
W ∗ denote the dimensional radial, azimuthal and axial similarity velocities respectively
and r∗ and z ∗ denote the dimensional radial position and wall-normal height. Let also ν ∗
denote the kinematic viscosity and Ω∗ the angular velocity of the disk.
Thus, using the regular notation, the Navier-Stokes equations in cylindrical polar
coordinates, with the absence of a θ-dependence due to inherent symmetry of the problem,
reduce to
∂U ∗ ∗ V ∗2
∗
+ U · ∇U ∗
− ∗
− 2Ω∗ V ∗ − Ω∗ 2 r∗
∂t r
1 ∂P ∗ U∗ 2 ∂V ∗

2 ∗
=− + ν ∇ U − ∗2 − ∗2
ρ ∂r∗ r r ∂θ
∗ ∗ ∗
∂V U V
∗
+ U∗ · ∇V ∗ − + 2Ω∗ U ∗
∂t r∗
1 ∂P ∗ 2 ∂U ∗ V∗

2 ∗
=− ∗ + ν ∇ V + ∗2 − ∗2
r ρ ∂θ r ∂θ r
∂W ∗ 1 ∂P ∗
∗
+ U∗ · ∇W ∗ = − ∗
+ ν∇2 W ∗
∂t ρ ∂z
∂U ∗ U ∗ 1 ∂V ∗ ∂W ∗
+ ∗ + ∗ +
∂r∗ r r ∂θ ∂z ∗
subject to the boundary conditions
U ∗ (0) = W ∗ (0) = 0, V ∗ (0) = 0; U ∗ → 0, V ∗ → −Ω∗ r as z ∗ → ∞
as a result of no-slip and no-penetration at the disk surface. In this formulation we have
chosen to work in a frame of reference that is rotating with the disk, thereby requiring us
to add Coriolis and streamwise curvature effects into the governing equations and alter
the boundary conditions on the azimuthal base flow component. Most calculations and
54
simulations throughout this report will be conducted in this rotating frame, although one
can switch between frames if necessary by removing appropriate terms and redefining the
base flow boundary conditions accordingly.
In von Kármán’s original derivation, the rotation rate of the disk, Ω∗ , was considered
constant, however for what follows we will assume that the disk rotates in the usual flow
configuration at an arbitrary rotation rate Ω∗ (t∗ ). If we let U = (U ∗ , V ∗ , W ∗ ) denote
the dimensional base flow then the boundary conditions associated with the system in a
non-rotating frame of reference are given by
U ∗ (0, t∗ ) = W ∗ (0, t∗ ) = 0, V ∗ (0, t∗ ) = r∗ Ω∗ (t∗ )
U∗ → 0 V ∗ → 0 as z ∗ → ∞
Initially, von Kármán [80] scales out the radial dependence of the base flow and writes
U ∗ (r∗ , z ∗ , t∗ ) V ∗ (r∗ , z ∗ , t∗ ) W ∗ (r∗ , z ∗ , t∗ )

f ∗ (z ∗ , t∗ ) = , g ∗ ∗ ∗
(z , t ) = , h∗ ∗ ∗
(z , t ) = (2.1)
r∗ r∗ δ∗
where the lower case (f ∗ , g ∗ , h∗ ) now have the dimension of frequency, namely s−1 . Drop-
ping the ∗ for notational simplicity gives the system of equations
∂f ∂f ∂ 2f
= g2 − f 2 − h +ν 2 (2.2a)
∂t ∂z ∂z
∂g ∂g ∂ 2g
= −2f g − h + ν 2 (2.2b)
∂t ∂z ∂z
∂h
= −2f (2.2c)
∂z
with boundary conditions
f (0, t) = h(0, t) = 0, g(0, t) = Ω(t)
f →0 g → 0 as z → ∞
The modulation we propose is of a periodic nature, similar to that of the Stokes layer
discussed in section 1.4.4. In order to draw parallels between this work and that of Thomas
et al. [76], the rotation rate is thus chosen to have the form
Ω∗ (t∗ ) = Ω∗0 + φ∗ cos(φ∗ t∗ ) (2.3)
where Ω∗0 may be thought of as analogous to the constant rotation rate in the steady case,
while and φ∗ denote the angular displacement and angular velocity of the modulation
respectively. It should be noted that the steady system is recovered for = 0. Depending
55
on the frequency of oscillation, the flow can be seen to evolve on two length scales; the
boundary layer thicknesses based on the constant rotation rate Ω∗0 and on the modulation
frequency φ∗ . We will refer to these two length scales as the Kármán layer and the Stokes
layer respectively, and they take the familiar forms
r ∗ r ∗
∗ ν ∗ ν
δk = ∗
, δs = (2.4)
Ω0 φ∗
There are also three temporal scales to consider, two associated with the local and global
rotation rates of the disk and one pertaining to the modulation frequency. We can thus
elect to non-dimensionalise the time-scale in either of the following ways
rL∗ ∗ ∗
τkl = ∗ Ω0 t , τkg = Ω∗0 t∗ , τs = φ∗ t∗ (2.5)
δk
The differences incurred by the choice of scaling will be discussed later in chapter 4 in
conjunction with the global stability but since we are only at this stage interested in
dealing with the local stability analysis, we neglect the global time-scale τkg . Thus, we are
left with a choice of non-dimensionalisation based on either one of the scales
(δk∗ , τkl ), (δs∗ , τs ) (2.6)
Either scaling could equally be chosen without adverse effects, provided we are consistent
in our solution method. However, since we are dealing primarily with a small-amplitude
modulation of the otherwise steady system and with a necessity of validation in mind, we
choose to non-dimensionalise on the Kármán scales (δk∗ , τkl ) for the following procedure.
A local velocity scaling is obtained by using the circumferential disk velocity rL∗ Ω∗0
at some dimensional reference radius rL∗ , chosen over its global counterpart based on the
boundary layer thickness, δk∗ Ω∗0 . The local scaling gives a Reynolds number R, associated
with the steady rotation of the disk as
rL∗ Ω∗0 δk∗ rL∗

R= = = rL (2.7)
ν∗ δk∗
where we identify rL as a local, non-dimensional radial position. We can also define a

Reynolds number, Rs associated with the Stokes layer, in a similar manner to section
1.4.4 given by s
rL∗ φ∗0 δs∗ φ∗ √
Rs = ∗
= rL ∗
= ϕR (2.8)
ν Ω0
56
φ∗
where we have defined ϕ := Ω∗0
. We interpret ϕ as a non-dimensional frequency term, and
identify it with the number of periods of modulation during one disk rotation. Scaling
the velocities on the local scale rL∗ Ω∗0 gives
F (z, τ ) = Ω∗0 f (z, τ ), G(z, τ ) = Ω∗0 g(z, τ ), H(z, τ ) = Ω∗0 h(z, τ ) (2.9)
and results in the system of equations
∂ 2F

∂F 1 2 2 ∂F
= G −F −H + (2.10a)
∂τ R ∂z ∂z 2
∂G ∂ 2 G

∂G 1
= −2F G − H + (2.10b)
∂τ R ∂z ∂z 2
∂H
= −2F (2.10c)
∂z

ϕ
F (0, τ ) = H(0, τ ) = 0, G(0, τ ) = 1 + Uw cos τ (2.11a)
R
F → 0, G → 0 as z → ∞ (2.11b)
√
Rs ϕ
where Uw := R
. Thus, we see that the flow properties are fully specified by the varia-
tion of the three parameters (R, Uw , ϕ), and that the steady case is recovered for Uw = 0.
Keeping our laminar flow control motivation in mind, it is sensible to constrain the vari-
ation in the wall velocity to be relatively small with respect to the rotation rate, and so
we impose the condition Uw < 0.2 for the remainder of this work. In turn, for sufficiently
high ϕ and R in the region of interest, this keeps the Reynolds number associated with
the Stokes layer, Rs, to be small compared with the Reynolds number associated with in-
stability onset in the semi-infinite Stokes layer of Blennerhassett and Bassom [11], namely
Rs ≈ 350. We may, therefore, proceed with cautionary confidence that the introduction
of the modulation will not introduce any additional modes of instability arising from the
Stokes part of the flow, and will endeavour to ensure that this is indeed the case in the
following solution procedures. Figure 2.2 shows the variation in Rs with ϕ for various
wall modulation velocities Uw . For the majority of results discussed in chapter 3, we will
use Uw < 0.2 and ϕ > 1 so as to appropriately constrain the value of Rs .
57
100
80
60
40
20
0
0 20 40 60 80 100
Figure 2.2: Variation in Rs with ϕ for wall modulation velocities Uw = 0.1 (-), Uw = 0.2
(- - -) and Uw = 0.25 (· · · ), for R fixed at R = 500.
Numerical Solution of the Base Flow Equations
In the steady case with Uw = 0, the system (2.10) reduces to the von Kármán system of
ODEs
G2 − F 2 − F 0 H + F 00 = 0 (2.12a)
−2F G − G0 H + G00 = 0 (2.12b)
H 0 = −2F (2.12c)
Similarly to the Blasius configuration, this system can be solved with a relatively straight-
forward application of MATLAB’s bvp4c solver, and evaluated at the collocation points.
It is worth noting that out-of-the-box, bvp4c has issues with directly solving the equation
on a large physical domain. In fact, it is numerically unstable over [0, a] for a > 12. This
was avoided by introducing an iterative scheme which used the solution over [0, K] as an
initial guess for the solution over [0, K + 1] and proceeded for K ≥ 1. The base flow
profiles for the non-rotating frame are shown in figure 2.3.
To facilitate our aim of solving the full time-dependent equations (2.10), we use similar
matrix operators to those already derived in section 1.4.3, and enforce similar parities on
the base flow variables. Choosing to solve the equations in the non-rotating frame would
ensure both F and G decay as z → ∞ and allow these flow variables to be represented
58
1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8
Figure 2.3: Steady rotating disk base flow velocity profiles F (-), G (- - -) and −H (· · · )
in terms of odd Chebyshev polynomials, while leaving H to be expressed as even since

by definition it has opposite parity. Curtailing our discussion here and proceeding would
give rise to a small problem, since we would require the value of H as z → ∞ which
is determined by the equations and not pre-specified. Fortunately this issue may be
alleviated by the usual mapping (1.40), which gives, in the non-rotating frame,

∂F 1 2 2 1 2 ∂F 2
= G −F + η H +D F (2.13a)
∂τ R l ∂η

∂G 1 1 2 ∂G 2
= −2F G + η H +D G (2.13b)
∂τ R l ∂η
∂H 2l
= 2F (2.13c)
∂τ η

ϕ
F (1, τ ) = H(1, τ ) = 0, G(1, τ ) = 1 + Uw cos τ (2.14a)
R
F → 0, G → 0 as η → 0 (2.14b)

2 η2 ∂ ∂2
where D = l
2η ∂η + η 2 ∂η 2 .
Since the flow variable H only appears in the equations when multiplied by η 2 , we
may expand the modified variable H̃ := η 2 H in terms of even Chebyshev polynomials,
thereby satisfying the required far-field decay. The continuity equation may be adapted
59
to take this into account by noticing that
∂ H̃ 2
= H̃ + 2lF (2.15)
∂η η
which can be integrated analytically by means of an integrating factor to give the explicit
definition Z η
2 F
H̃ = 2lη (2.16)
0 η2
In practise we are able to calculate H̃ given the Chebyshev coefficients of the expansion
F
for F̂ := η2
. The multiplications and divisions by η 2 are performed in collocation space
after a fast Fourier transform while the integration may be evaluated using the integral
representations of the Chebyshev polynomials (1.32). Thus, we can write
" N ! N
#
X ˆ1
f X 1
H̃(η) = 2lη 2 ak fˆk T0 (ζ) + T2 (η) + (fˆk − fˆk+1 )T2k (η) (2.17)
k=1
4 k=2
4k
where
(−1)k (1 − 2k)
a1 = 0, ak = (k > 1) (2.18)
4k(k − 1)
and thus, H̃ = SF̂ where S is the matrix form of (2.17).
The temporal integration is performed via a backward three-level scheme which takes
the form l
∂f 1
3f l − 4f l−1 + f l−2

= (2.19)
∂t 2∆t
where we have used the notation
fl = f
t=l∆t
The majority of terms on the right hand side of (2.10) are treated explicitly via a predictor-
corrector method in the temporal integration, although for numerical stability reasons the
second z-derivative term is treated implicitly. The predictor step for the explicit terms is
given by
(N l )p = 2N l−1 − N l−2
while the corrector step is defined by
(N l )c = (N l )p
where (N l )p denotes the velocity fields determined from the predictor stage.
60
We may simplify notation slightly by defining the operators RF and RG to be the right
hand sides of the transport equations (2.13a) and (2.13b) without the second derivative
terms, giving
∂F 1
− D2 F = RF
∂τ R
∂G 1
− D 2 G = RG
∂τ R
while H is now defined in terms of F by (2.16). Discretising the time derivative using
(2.19) and integrating the transport equations twice with respect to the mapped variable
η gives
ZZ ZZ ZZ ZZ ZZ
3 2 l+1 2 1
− D F dη = F l dη − l−1
F dη + RlF dη (2.20a)
2∆τ ∆τ 2∆τ
ZZ ZZ ZZ ZZ ZZ
3 2 l+1 2 1
− D G dη = Gl dη − l−1
G dη + RlG dη (2.20b)
2∆τ ∆τ 2∆τ
The Chebyshev polynomial representation allows the integrals to be expressed as matrix

operators and the time-marching is achieved by means of matrix inversions and multipli-
cations at each time-step.
The nonlinear product terms are treated with a pseudo-spectral scheme which uses
a Fast Fourier Transform to convert between Chebyshev coefficients and collocation val-
ues. This method was explained in section 1.3.1 but essentially can be described by the
following diagram:
Set up equations
FFT: Chebyshev → Collocation
Temporal loop Calculate (F l , G l ) Multiply collocation values
FFT: Collocation → Chebyshev
Reset equations
61
A similar idea is also employed to calculate the η-derivatives of F and G and is accom-
plished using the NAG library routines from the C06 series or the open-source versions
from dfftpack [54]. A slight gain in computational cost may be achieved by inverting the
constant matrix on the left hand side once during initialisation of the method, and simply
multiplying at each time-step.
Illustrative Results
Figure 2.4 shows a typical variation of the modulated base flow over one period of mod-
ulation. The profiles are shown for the non-rotating frame with boundary conditions on
the azimuthal velocity field given by
ϕ
G(z = 0, τ ) = 1 + Uw cos τ , where R = 500, Uw = 0.2 and ϕ = 10
R
Clearly there is much less variation in the radial and vertical flow components than in the
azimuthal component. In fact it is shown in section 2.1.2 that the variation in F and H
are of a lower order of oscillation magnitude than the variation in G.
1.2
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8
Figure 2.4: Typical variation of the modulated velocity profiles F (- - - ), G (—), −H

(· · · ) with R = 500, Uw = 0.2 and ϕ = 10. Profiles have been sampled at 10 evenly spaced
time locations over a period of modulation to demonstrate the variation in the three base
flow components.
62
Figure 2.5 shows the temporal evolution of the base flow for vertical positions z = 0,
z = 0.1 and z = 0.25 over one period of modulation with R = 500, Uw = 0.2 and ϕ = 10.
1.2
1.1
0.9
0.8
0.7
0 0.2 0.4 0.6 0.8 1
Figure 2.5: Temporal evolution of G for vertical positions z = 0 (-), z = 0.1 (- - -) and
z = 0.25 (· · · ) over one period of modulation, with R = 500, Uw = 0.2 and ϕ = 10.
2.1.2 High Frequency Limit for Small Amplitude Modulation
To begin the exploration of alterations introduced by the modulation, it is informative to

consider the base flow velocity fields in the high frequency limit. Throughout the following
discussion in this section we will assume that the oscillations are of high frequency, ϕ 1,
√
Rs ϕ
and small amplitude so that ˆ := R
1, allowing for non-linear interactions in ˆ to
be neglected.
Consider the modulated equations in the non-rotating frame given by (2.10), with
boundary conditions
ϕ
F (0, τ ) = H(0, τ ) = 0, G(0, τ ) = 1 + ˆ cos τ
R
F → 0, G → 0 as z → ∞
and small amplitude oscillations, ˆ 1 with fixed R. We expand the flow variables in
63
terms of this small parameter ˆ to give
F = F0 + ˆf1 + ˆ2 f2 + O(ˆ3 )
G = G0 + ˆg1 + ˆ2 g2 + O(ˆ3 )
H = H0 + ˆh1 + ˆ2 h2 + O(ˆ3 )
and derive the first and second order expansion equations to be

∂f1 ∂ 2 f1

∂f1 1 ∂F0
= 2 (G0 g1 − F0 f1 ) − h1 − H0 + (2.21a)
∂τ R ∂z ∂z ∂z 2
∂g1 ∂ 2 g1

∂g1 1 ∂G0
= −2 (F0 g1 + G0 f1 ) − h1 − H0 + (2.21b)
∂τ R ∂z ∂z ∂z 2
∂h1
= −2f1 (2.21c)
∂z
ϕ
f1 (0, t) = h1 (0, t) = 0, g1 (0, t) = cos τ
R
f1 → 0 g1 → 0 as z → ∞
and
∂F0 ∂ 2 f2

∂f2 1 2 2 ∂f2 ∂f1
= 2g2 G0 + g1 − 2f2 F0 − f1 + H0 + h1 + h2 + (2.22a)
∂τ R ∂z ∂z ∂z ∂z 2

∂g2 1 ∂g2 ∂g1 ∂G0
= f2 G0 + f1 g1 + F0 g2 + H0 + h1 + h2 (2.22b)
∂τ R ∂z ∂z ∂z
∂h2
= −2f2 (2.22c)
∂z
with homogeneous boundary conditions on f2 , g2 and h2 .
Since we are looking specifically at the high frequency limit, the primary lengthscale
on which the flow will evolve will be the Stokes layer thickness δs . In the current non-
q
dimensionalisation, this takes the form δs = ϕ2 and we can scale
z ϕ
z̃ = τ̃ = τ
δs R
to give

1 ∂f1 1 1 0 0 1 00
= 2 (G0 g1 − F0 f1 ) − (F0 h1 + H0 f1 ) + 2 f1 (2.23a)
Rδ 2 ∂ τ̃ R δ δ

1 ∂g1 1 1 0 0 1 00
= −2 (F0 g1 + G0 f1 ) − (G0 h1 + H0 g1 ) + 2 g1 (2.23b)
Rδ 2 ∂ τ̃ R δ δ
h01 = −2δf1 (2.23c)
64
0 ∂
where we have set = ∂ z̃
and dropped the subscript s from the definition of δs for
notational simplicity. We have homogeneous boundary conditions on f1 and h1 together
with
g1 (0, τ̃ ) = cos(τ̃ ), g1 (z̃ → ∞, τ̃ ) = 0 (2.24)
At this stage, it is possible to develop an expansion for f1 , g1 and h1 in terms of the small
parameter δ, so we set
f1 = f10 + δf11 + δ 2 f12 + O(δ 3 )
g1 = g10 + δg11 + δ 2 g12 + O(δ 3 )
h1 = h01 + δh11 + δ 2 h21 + O(δ 3 )
which, after substitution into (2.23), gives
∂f1k ∂f1k−1 ∂ 2 f1k

= 2 G0 g1k−2 − F0 f1k−2 − F00 hk−2

1 + H 0 + (2.25a)
∂ τ̃ ∂ z̃ ∂ z̃ 2
k k−1
∂g1 ∂g1 ∂ 2 g1k
= −2 F0 g1k−2 + G0 f1k−2 − G00 hk−2

1 + H 0 + (2.25b)
∂ τ̃ ∂ z̃ ∂ z̃ 2
∂hk1
= −2f1k−1 (2.25c)
∂ z̃
for some integer k. Note that the superscript refers to the order in terms of the bound-
ary layer thickness δ while the subscript refers to the order in terms of the modulation
amplitude ˆ. If k = 0, this gives the dominant terms to first order in δ as
∂f10 ∂ 2 f10
= (2.26a)
∂ τ̃ ∂ z̃ 2
0
∂g1 ∂ 2 g10
= (2.26b)
∂ τ̃ ∂ z̃ 2
0
∂h1
=0 (2.26c)
∂ z̃
which, subject to the boundary conditions (2.24), gives
f10 = 0 (2.27a)
g10 = exp(−z̃) cos(τ̃ − z̃) (2.27b)
= e−(1+i)z̃ eiτ̃ + e−(1−i)z̃ e−iτ̃ (2.27c)
h01 = 0 (2.27d)
65
If k = 1, the second order in δ equations reduce to
∂f11 ∂ 2 f11 ∂f10

= + H 0 (2.28a)
∂ τ̃ ∂ z̃ 2 | {z∂ z̃}
=0
∂g11 ∂ 2 g11 ∂g10
= + H0 (2.28b)
∂ τ̃ ∂ z̃ 2 ∂ z̃
∂h11
= −2f10 (2.28c)
∂ z̃ | {z }
=0
with homogeneous boundary conditions. This demonstrates that f1 and h1 are of at least
an order of magnitude in δ smaller than g1 in this limit. Equations (2.28) also provide
the non-linear correction term g11 .
This process suggests that the dominant behaviour for small amplitude, high frequency
modulations amounts to the addition of a Stokes layer to the steady rotating disk bound-
ary layer in a similar fashion to Thomas et al. [76] for the channel flow scenario. Therefore,
the configuration we are presenting throughout this report may be viewed as a combina-
tion of the archetypal two-dimensional oscillatory boundary layer, the Stokes layer, and
one of the canonical three-dimensional boundary layer configurations, the steady rotating
disk.
The results of the analysis may be confirmed numerically by solving the linearised
system of equations (2.21) in a similar fashion to the solution of the basic state equations
in section 2.1.1. Figure 2.6 shows a snapshot at a frozen time instant of the comparison
between g10 and the time-dependent part of the azimuthal component of the base flow,
Gosc , calculated by the relation.
Gosc = Gtotal − Gsteady (2.29)
Figures 2.7 and 2.8 show a similar comparison across a single period of modulation at two
values of z, namely z = 0.1 and z = 0.25. Figure 2.9 shows the maximum error
max |Gosc − g10 | (2.30)

τ̃ ,z̃
across the boundary layer and across a modulation period between the analytic Stokes
layer contribution g10 and the time-dependent part of the azimuthal component of the base
flow Gosc for ˆ = 0.1 and ˆ = 0.2 across a range of values for ϕ. As would be expected, the
maximum error is highest for low frequency modulations, although still relatively small
66
when compared against the modulation amplitude ˆ. Thus, it should be appropriate to
utilise the approximation that
G(z̃, τ̃ ) ∼ G0 (z̃) + exp(−z̃) cos(τ̃ − z̃) (2.31)
for any stability analyses conducted using the modulated rotating disk boundary layer,
provided the oscillation amplitude and frequency are constrained appropriately. Through-
out the stability analyses conducted in chapter 3, this approximation is not used, and the
base flow is always computed numerically in full. However, for future studies that re-
quire a more restrictive computational load, the use of this approximation would aid in
a reduction in computational cost and simplify the analysis. As will be explained in the
forthcoming chapter, this approximation is in no way necessary for our analysis, and is
intended only as an illustration of the dominant behaviour in the system.
2.5
1.5
0.5
0
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
Figure 2.6: Comparison at frozen time instant between the Stokes layer contribution g10
(—) and the time-dependent part of the azimuthal component of the base flow Gosc (- -
-) for ˆ = 0.2 and ϕ = 10.
67
0.2
0.1
-0.1
-0.2
0 0.2 0.4 0.6 0.8 1
Figure 2.7: Comparison between the analytic Stokes layer contribution g10 (—) and the
time-dependent part of the azimuthal component of the base flow Gosc (- - -) for ˆ = 0.2
and ϕ = 10 at z = 0.1.
0.2
0.1
-0.1
-0.2
0 0.2 0.4 0.6 0.8 1
Figure 2.8: Comparison between the analytic Stokes layer contribution g10 (—) and the
time-dependent part of the azimuthal component of the base flow Gosc (- - -) for ˆ = 0.2
and ϕ = 10 at z = 0.25.
68
0.06
0.05
0.04
0.03
0.02
0.01
0
1 5 10 15 20 25 30 35 40 45 50
Figure 2.9: Maximum error max |Gosc − g10 | across the boundary layer and across a mod-
τ̃ ,z̃
ulation period between the analytic Stokes layer contribution g10 and the time-dependent
part of the azimuthal component of the base flow Gosc for ˆ = 0.1 (—) and ˆ = 0.2 (- - -
) across a range of values for ϕ.
69
2.1.3 Average Deviation from the Basic State
An important feature of the modulated rotation rate when compared with other stabilising
techniques such as the surface roughness imposed by Cooper et al. [21], Garrett et al. [36]
is that there is no inherent average deviation from the base flow across a period. This
may be checked by computing the average of a flow field over a period of modulation and
comparing against the steady state solution. We compute the average using the expression
1 T
Z
F̄ (z) = F (z, t)dt
T 0
and figure 2.10 shows the maximum deviation between the averaged azimuthal flow com-
ponent and the steady solution across a range of ϕ and Uw values. We can thus see
that for small amplitude modulation, the deviation from the steady state is negligible, in
contrast to other stabilising methods studied previously. We will revisit this averaging
procedure of Garrett et al. [36] in our discussion of future directions in chapter 5.
10 -3
1.2
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100
Figure 2.10: Maximum deviation of averaged modulated base flow profiles from the steady
state solution for varying ϕ and Uw = 0.1 (-), Uw = 0.2 (- - -) and Uw = 0.25 (· · · ).
70
2.1.4 Review of the Local Stability Properties of Steady Rotat-
ing Disk Boundary Layers
The local linear stability of the steady rotating disk boundary layer was first studied by
Gregory et al. [38] in 1955 with a view to approximating the flow over a swept wing,
before Malik [50] and Mack [49] provided solutions to the eigenvalue dispersion relations
and calculated neutral stability curves in the early 1980s. As discussed in section 2.1,
both flow configurations are three-dimensional and have distinct flow components in two
directions along the surface. The disk has flow components in the radial and azimuthal
directions, while the swept wing configuration has flow components along the span of
the wing and in the direction of flight. It was shown by Gregory et al. [38] that both
configurations are also prone to crossflow instabilities, which give rise to crossflow vortices.
For this reason, the rotating disk was classically studied as a canonical example for a
three-dimensional flow and as an approximation to the flow over a swept wing. The facts
that the rotating disk equations admit an exact similarity solution to the Navier-Stokes
equations, and that the disk is easier to work with experimentally has inspired a great
deal of stability studies over the past few decades.
There are three types of instability present in the rotating disk boundary layer. A
crossflow instability or type I mode was originally discovered by Gregory et al. [38]. This
instability is inviscid in nature, and similar mechanisms can be found on swept wings and
other rotating objects. This is the most unstable of the three types of instability, in the
sense that it occurs at the lowest Reynolds number for stationary disturbances. It was
shown first by Malik [50], and then confirmed by Lingwood [44] and others such as Cooper
and Carpenter [19], Thomas [71], Dhanak et al. [30], Garrett et al. [36] that the onset of
this instability occurs around R = 286 for disturbances which are stationary with respect
to the disk surface.
In addition to the crossflow instability, the boundary layer is convectively unstable
to a second mode, type II, which is viscous in nature and caused by the streamwise
curvature and Coriolis effects. In the neutral curves present in Malik [50], Lingwood
[47] and this document, the type I and type II instabilities can be clearly seen as the
upper and lower sections respectively. Mack [49] mentions a third mode, now known
as the type III mode, which propels energy towards the disk centre, but is spatially
damped. Travelling instabilities in the rotating disk configuration were discussed and
71
traced in detail by Balakumar and Malik [7] and their importance to the full spatio-
temporal stability structure of the flow was realised with Lingwood’s [43] discovery of an
absolute instability present in the configuration. As explained in section 2.2.3, it is the
coalescence of the type I and III modes which causes the absolute instability to arise.
As well as the convectively unstable type I and type II instabilities, where distur-
bances grow spatially as they convect radially outward, Lingwood [43] showed, using a
local spatio-temporal linear stability analysis, that above a critical Reynolds number of ap-
proximately 507 the flow becomes absolutely unstable in the radial direction. Lingwood’s
[43] absolute instability was found using Briggs’ method [15] and occurs for travelling,
rather than stationary disturbances.
Given the similarities between the rotating disk and swept-wing boundary layers, it
seems natural to consider whether such a mechanism exists for the latter. However, as
discussed in Lingwood and Alfredsson [46], one major difference between the rotating
disk and swept-wing flows lies with the inherent azimuthal periodicity. The possibility
that disturbances can circulate in the azimuthal direction has significant implications for
the long term behaviour. In order for there to exist a physically significant absolute
instability in the swept-wing boundary layer, a wavepacket would need to grow in time
simultaneously at fixed chordwise and spanwise locations. Lingwood and Alfredsson [46]
mention the asymptotic studies of Lingwood [45], Ryzhov and Terent’ev [64] and Taylor
and Peake [70] which demonstrate the possibility of chordwise, but not spanwise, absolute
instability, and without spanwise periodicity analogous to the azimuthal periodicity of the
rotating disk flow, chordwise absolute instability does not prevent disturbances convecting
out of the domain of interest.
One of the original motivations for Lingwood’s [43] spatio-temporal analysis was the
relatively consistent experimental evidence of transition to turbulence around a Reynolds
number of R = 513 ± 3%. The critical Reynolds number for the onset of absolute instabil-
ity was found by Lingwood [43] to be R = 507.3 and she thus suggested that the absolute
instability may be a cause for transition to turbulence. However, there are limitations as
to what conclusions relating to the connection between absolute instability and transition
can be drawn from this analysis, due to the fact that this analysis is purely linear and
local in radius while experimental studies include non-linear effects and are spatially de-
veloping. In particular, Davies and Carpenter [27] showed, using linear DNS of impulsive
72
disturbances, that when the radial inhomogeneity of the flow is incorporated, disturbances
do not grow in time at fixed radial locations above the critical Reynolds number for local
absolute instability, at least for the azimuthal mode numbers in the range that Lingwood
[43] focused upon. The conclusion drawn, therefore, was that the rotating disk boundary
layer flow has no unstable linear global mode, and is linearly globally stable, even where
the flow is locally absolutely unstable. This describes the first numerical global study of
the flow and was contrary to what many others believed at the time. We will review the
global stability results and present an explanation of the effects of incorporating radial
inhomogeneity in chapter 4, but for now will concentrate on the local approximation in
the context of establishing an eigenvalue dispersion relation analogous to equation (1.9).
2.1.5 Steady Local Eigenvalue Analysis Using a Primitive Vari-

able Formulation
A brief account of the perturbation analysis undertaken by other authors such as Malik
[50], Lingwood [47, 43] and Appelquist [3] will be given here, with the view of introducing
the normal mode approximation taken, and the non-dimensionalisations that will be em-
ployed throughout the remainder of this chapter. Following the methods of these previous
authors, we non-dimensionalise using the local radial position rL , so that the timescale is
non-dimensionalised by δ ∗ /(Ω∗ rL∗ ) and the Reynolds number becomes
R = rL (2.32)
as described in section 2.1.1. This gives the mean flow in the non-dimensional form as

r r 1
U(r, z) = F (z), G(z), H(z) (2.33)
R R R
Under this scaling, the so-called parallel-flow approximation amounts to the replacement
of the radial coordinate r by the constant rL in the equations, which homogenises the
mean flow along the radial direction. By selecting the mean flow at one given radial
position and then artificially replicating it at all other radial locations, this parallel-flow
approximation allows the disturbance to be separable along the radial direction and take
a normal-mode form. As mentioned in Lingwood and Alfredsson [46], this is a useful
and common approximation to make, and is known to give accurate results for the local
analysis provided the disturbance wavelength is short compared to the scale the basic flow
73
varies on. However, on using this approach we must neglect terms which are the same
order as others that are retained. In particular, some R−1 terms are kept because the
no-slip boundary condition can create instability, whereas other R−1 terms are neglected
because they would lead to a partial instead of an ordinary differential equation. While
this approach has been shown by Balakumar and Malik [7], among others, to have little
effect on the eigenvalue analysis, it is important to note that the inhomogeneity of the
disk arising from including the non-parallel terms has a profound effect on the global
stability of the flow. This will be revisited during the discussion of the global stability
in chapter 4 but for the time being we will assume that the local effects of applying this
approximation are negligible. The rotational symmetry of the mean flow, together with
the inherent cylindrical geometry of the problem, ensures that the disturbance variation
along the azimuthal direction can also be taken to have a normal-mode form. Similarly
to usual stability analyses, we can decompose the total flow into mean and perturbation
quantities
r
F (r, θ, z, t) = F (z) + u(r, θ, z, t)
R
r
G(r, θ, z, t) = G(z) + v(r, θ, z, t)
R
1
H(r, θ, z, t) = H(z) + w(r, θ, z, t)
R
1
P (r, θ, z, t) = 2 P (z) + p(r, θ, z, t)
R
and take a normal mode form of the disturbance to get
u = û(z)ei(αr+βRθ−ωt) , p = p̂(z)ei(αr+βRθ−ωt) (2.34)
where û = (û, v̂, ŵ) and p̂ are the spectral representations of the perturbation fields. We
interpret, as usual, ω as the frequency of the travelling wave disturbance and α, β as the
radial and azimuthal wavenumbers respectively. The azimuthal mode number n = βR
can be identified as the integer number of complete cycles of the disturbance around the
azimuth and can be compared experimentally with the number of spiral vortices around
the disk surface. Although by definition n is integer-valued, for practical purposes we will
do as others have done and treat β, and therefore n, as continuously varying. Substituting
the normal mode solution (2.34) into the full perturbation equations and applying the
74
parallel-flow approximation by replacing r by R gives the perturbation equations

1 iβ
iα + û + v̂ + ŵ0 = 0
R R
iβG α2 β2

F H 1 2
iαF + + + 3+ û + û0 − û00 − (1 + G)v̂ + F 0 ŵ + iαp̂ = iûω
R R R R R R R
2 2

2 iβG α β F H 1 iβ
(1 + G)v̂ + iαF + + + 3+ v̂ + v̂ 0 − v̂ 00 + G0 ŵ + p̂ = iv̂ω
R R R R R R R R
2 2 0

iβG α β F H 0 1 00 0
iαF + + + 3+ ŵ + ŵ − ŵ p̂ = iŵω
R R R R R R
The stability properties of this flow have been studied using these equations in various
works by Malik [50], Lingwood [47, 43] and Appelquist [3] and others and there is a wealth
of literature describing their findings. Their solution method amounts to expressing these
equations as an eigenvalue relation of the form
D(ω, α, β, R) = 0 (2.35)
and solving for α once ω, β and R are specified. We will not attempt to solve the
eigenvalue equations in this form, but will instead make use of the velocity-vorticity ideas
from chapter 1.
2.2 Local Stability of the Steady Rotating Disk Bound-

ary Layer using a Velocity-Vorticity Formulation
The velocity-vorticity formulation of Davies and Carpenter [26] is discussed in section 1.4.1
for the general three-dimensional Cartesian case, and solved for a special two-dimensional
case; the Blasius boundary layer. Since, in this section, we are interested in the rotating
disk boundary layer, we will formulate the governing equations for this flow configuration
and discuss the advantages over the methods of analysis based on the primitive variable
perturbation equations. In previous studies, such as Davies and Carpenter [26, 27], Davies
et al. [28] and Thomas and Davies [73], the velocity-vorticity formulation has been utilised
to conduct direct numerical simulations (DNS) of the steady rotating disk boundary layer.
The study described in the following section, however, marks the first linear normal mode
analysis using said formulation. Direct numerical simulations are widely used in the fluid
dynamics community to achieve numerical solutions of the Navier-Stokes equations which
fully resolve the spatial and temporal evolution of velocity variables. These simulations
75
are often very computationally costly, and thus have become far more popular in recent
years with the rapid advancement of computational power. Simulations that only in the
last decade would have taken days on a departmental server can now be accomplished in a
reasonable time frame on a personal workstation. This has led to many simulation based
studies being carried out to understand the stability of the rotating disk boundary layer.
Direct numerical simulations allow for a more complete understanding of the stability
properties of a flow than eigenvalue-based approaches. One can view DNS as a stepping
stone, bridging the gap between theory and experiment and indeed has allowed far more
of an insight into spatially developing flows than was previously possible. A recent paper
by Lingwood and Alfredsson [46] gives a review of the current state of the field and
overviews are also given in section 2.1.4 and chapter 4. Some notable simulation studies
are described in Davies and Carpenter [27], Davies et al. [28], Appelquist et al. [5, 6] and
Appelquist et al. [4].
When considering the Blasius case, we used a Cartesian coordinate system to present
the velocity-vorticity formulation. As stated in Davies and Carpenter [26], we can readily
change to a cylindrical polar system and a non-inertial frame of reference which rotates at
a constant angular velocity about the z-axis in conjunction with the disk. The Reynolds
number R is given by (2.32), and under the scalings (2.6), the non-dimensional rotation
1
rate is equal to Ω = R
.
In the same notation as section 2.1.5, let F , G and H denote the non-dimensional
radial, azimuthal and axial similarity velocities respectively, so that the undisturbed base
flow is given by (2.33). In the usual manner, perturbations to the velocity and vorticity
fields are introduced, denoted by
u = (ur , uθ , w), ξ = (ξr , ξθ , ξz )
If we define the primary variables as being (ξr , ξθ , w) then the following equations for the
primary variables in the rotating frame of reference are fully equivalent to the Navier-
Stokes equations

∂ξr 1 ∂Nz ∂Nθ 2 ∂w 1 2 1 2 ∂ξθ
+ − − ξθ + = ∇ − 2 ξr − (2.36a)
∂t r ∂θ ∂z R ∂r R r r2 ∂θ

∂ξθ ∂Nr ∂Nz 2 1 ∂w 1 2 1 2 ∂ξr
+ − + ξr − = ∇ − 2 ξθ + (2.36b)
∂t ∂z ∂r R r ∂θ R r r2 ∂θ

2 1 ∂ξr ∂(rξθ )
∇w= − (2.36c)
r ∂θ ∂r
76
where
N = (Nr , Nθ , Nz ) = (∇ × UB ) × u + ξ × UB + ξ × u
and
∂2 1 ∂ 1 ∂2 ∂2
∇2 = 2 + + 2 2+ 2
B B
∂r r ∂r
B
r ∂θ ∂z
∂UzB ∂UrB

B 1 ∂Uz ∂Uθ ∂Ur 1 ∂ B
∇×U = − er + − eθ + rUθ − ez
r ∂θ ∂z ∂z ∂r r ∂r ∂θ
are the usual Laplacian and curl operators in cylindrical polar coordinates. Similarly to
the Cartesian case, the secondary variables can be defined in terms of integrals of the
primary variables as follows
Z ∞
∂w
ur = − ξθ + dz (2.37a)
z ∂r
Z ∞
1 ∂w
uθ = ξr − dz (2.37b)
z r ∂θ
1 ∞ ∂(rξr ) ∂ξθ
Z
ξz = + dz (2.37c)
r z ∂r ∂θ
For the purposes of the present analysis, we linearise the governing perturbation equations
by dropping the ξ × u term in the definition of N, leaving
N = (Nr , Nθ , Nz ) = (∇ × UB ) × u + ξ × UB (2.38)
and assume that we are dealing with a rigid disk. The linearised no-slip conditions on the
perturbation variables thus become
ur = uθ = w = 0 at z = 0
and substitution of these boundary conditions into the definitions of the secondary vari-
ables gives the following integral constraints on the primary variables
Z ∞
∂w
ξθ + dz = 0
0 ∂r
Z ∞
1 ∂w
ξr − dz = 0
0 r ∂θ
The far-field vanishing conditions on the velocity and vorticity fields as z → ∞ discussed
for the Cartesian case in section 1.4 are satisfied for this configuration.
At this stage, we may proceed to gain insights into the stability of the flow configura-
tion by solving the system (2.36) in one of three ways, each of which will be described in
the following sections.
77
1. Eigenvalue Analysis: Firstly, as has been done in chapter 1, we may take a normal
mode approximation of the form
ˆ
u = û(z)ei(αr+βRθ−ωt) , ξ = ξ(z)e i(αr+βRθ−ωt)
(2.39)
and solve the resulting eigenvalue dispersion relation
D(ω, α, β, R) = 0 (2.40)
for α or ω having specified the other parameters.
2. Monochromatic DNS: Secondly, we could retain the time dependence in the

stability equations and only decompose the variables as
ˆ t)ei(αr+βRθ)
u = û(z, t)ei(αr+βRθ) , ξ = ξ(z, (2.41)
leaving a set of time-dependent ODEs to be solved in a similar manner to the base

flow equations in section 2.1.1. This allows us to effectively calculate ω given a
prescribed α and β.
3. DNS with Radial Dependence: Finally, we could genuinely include the radial
dependence and decompose as
ˆ t)eiβRθ
u = û(z, t)eiβRθ , ξ = ξ(z, (2.42)
leaving us to discretise the system in the radial direction. The solution method
for this scenario is described in detail in Davies and Carpenter [26] and will be
discussed in this report in section 2.2.5. Removing the normal mode solution in
the radial direction removes the necessity for the parallel flow approximation and
allows an exploration of the effects of the true radial inhomogeneity of the flow. As
briefly alluded to in section 2.1.5 and discussed in greater detail later in chapter
4, the inclusion of this inhomogeneity has profound effects on the global stability
properties of the flow and the reader should be aware of the potential pitfalls of the
parallel flow approximation throughout.
Each of these solution methods will be visited in turn throughout the following sections
and we begin our discussion with the eigenvalue analysis.
78
2.2.1 Numerical Methods for Linear Stability - Eigenvalue Anal-
ysis
The following sections illustrate the attempt to extend the work of chapter 1 to the
rotating disk configuration. We look to develop a temporal and spatial eigenvalue solver
for the velocity-vorticity formulation of the perturbation equations discussed in section
2.2. While studies exist in which the eigenvalue stability problem has been solved using
alternative methods to those presented in this work, no method for solving the velocity-
vorticity formulation has yet been found in the available literature.
There appear to be three distinct approaches taken to solve the eigenvalue problem in
the literature. Two of these, one based on a sixth order system of ODEs and one based
on a collocation method, are outlined in Malik [50], Lingwood [43] and Appelquist [3]
among other places and so the reader is referred to the references for details. The other
is similar to methods already employed in this report which involves integral operators
on the primitive variable formulation and is detailed in Cooper and Carpenter [19]. Due
to the facts that we are working in a mapped domain and that the primitive formulation
contains fourth order operators, this method leaves the operator definitions and the result-
ing equations inelegant and difficult to work with. Thus we turn to the velocity-vorticity
formulation as a starting point for an eigenvalue solver since the operators in question
take a much simpler form.
It is worth remarking that in deriving the velocity-vorticity formulation perturbation
equations, we apply the parallel-flow approximation at a different stage of the formulation
to other authors such as Malik [50], Lingwood [43] and Cooper and Carpenter [19], and
thus retain some terms that are not present in the standard analysis due to derivatives
taken with respect to the radial coordinate r. A study by Dhanak et al. [30] uses a
different vorticity formulation to study effects of uniform suction at the surface, and the
discrepancy between his and Malik’s [50] formulations is discussed in Lingwood [44]. It
has also been shown by Balakumar and Malik [7] that the additional terms arising from
the radial inhomogeneity have little effect on the local stability characteristics. Thus, if
we convince ourselves that we are at liberty to, we make the parallel flow approximation
as usual, replace r by R and take a normal mode approximation of the form
ˆ
u = û(z)ei(αr+βRθ−ωt) , ξ = ξ(z)ei(αr+βRθ−ωt)
(2.43)
79
1

After neglecting the non-parallel O R2
terms, as per the previous literature and for
consistency with validation, we get
∂2

∂Nθ 2 1
−iωξr + iβNz − − (ξθ + iαw) = −α2 − β 2 + ξr (2.44a)
∂z R R ∂z 2
∂2

∂Nr ∂Nz 2 1
−iωξθ + − + (ξr − iβw) = −α2 − β 2 + ξθ (2.44b)
∂z ∂r R R ∂z 2
2

∇ˆ2 + ∂ w = iβξr − iαξθ −
ξθ
(2.44c)
∂z 2 R
with conditions
w(0) = 0 (2.45)
Z ∞
(ξθ + iαw) dz = 0 (2.46)
0
Z ∞
(ξr − iβw) dz = 0 (2.47)
0
Treating the Convective Terms
The main difference between this formulation and the two-dimensional version discussed
for the Blasius configuration in section 1.4 is the inclusion of all six flow variables in the
convective terms. In order to create a well-posed eigenvalue problem we must eliminate
the secondary variables {ur , uθ , ξz } from consideration by expressing them in terms of
the primary variables {ξr , ξθ , w}. This is done in a similar fashion to the method for the
implementation of the boundary conditions in section 1.4 and utilises the mapping
l
η= (2.48)
z+l
where l is the scaling factor discussed in section 1.4. We can write an arbitrary secondary
variable s in terms of an arbitrary primary variable f as
Z ∞
s(z) = f (z̃)dz̃
z
Transforming this integral to the computational domain under the mapping (2.48), we
get Z η
f (η̃)
s(η) = l dη̃
0 η̃ 2
and if we divide through our perturbation equations by η 2 as we did for the Blasius case
in section 1.4, then we can define fˆ(η) = f (η)
η2
and thus get
Z η
s(η) = l fˆ(η̃)dη̃ (2.49)
0
80
The definitions of the secondary variables in terms of the primary variables are given by
equations (2.37) and on introducing the mapping (2.48) become
Z η
ur = −l ξˆθ + iαŵ dη̃ (2.50a)
Z η0
uθ = l ξˆr − iβ ŵ dη̃ (2.50b)
Z0 η
1ˆ ˆ ˆ ˜
ξz = l ξr + iαξr + iβ ξθ deta (2.50c)
0 R
To implement the solution method, we split the convective terms (2.38) into components
and write
1
Nr = (rF 0 w − 2Guθ + Hξθ − rGξz )
R
1
Nθ = (rG0 w + 2Gur − Hξr + rF ξz )
R
r
Nz = (Gξr − G0 uθ − F 0 ur − F ξθ )
R
By inspection of the equations and with the aid of hindsight, we can see that we will also
require the r-derivative of Nz , given by

∂Nz r ∂ξr 0 ∂uθ 0 ∂ur ∂ξθ 1
= G −G −F −F + (Gξr − G0 uθ − F 0 ur − F ξθ )
∂r R ∂r ∂r ∂r ∂r R
Since there are no other r-derivatives of the convective terms present in the stability
equations, we can at this stage apply the parallel-flow approximation and replace r by R
in our definitions of N, leaving
2 1
Nr = F 0 w − Guθ + Hξθ − Gξz (2.51a)
R R
2 1
Nθ = G0 w + Gur − Hξr + F ξz (2.51b)
R R
Nz = Gξr − G0 uθ − F 0 ur − F ξθ (2.51c)
∂Nz 1
= iα(Gξr − G0 uθ − F 0 ur − F ξθ ) + (Gξr − G0 uθ − F 0 ur − F ξθ ) (2.51d)
∂r R
81
Substitution of (2.50) into equations (2.51) gives
2lG η ˆ
Z Z η
0
1 1ˆ ˆ ˆ
Nr = F w − ξr − iβ ŵ dη̃ + Hξθ − lG ξr + iαξr + iβ ξθ dη̃
R 0 R 0 R
Z η Z η
0 2lG ˆ
1 1ˆ ˆ ˆ
Nθ = G w − l ξθ + iαŵ dη̃ − Hξr + lF ξr + iαξr + iβ ξθ dη̃
R 0 R 0 R
Z η Z η
Nz = Gξr − lG 0 ˆ
ξr − iβ ŵ dη̃ + lF 0
ξˆθ + iαŵ dη̃ − F ξθ
0 0
Z η Z η
∂Nz 0 ˆ

0 ˆ

= iα Gξr − lG ξr − iβ ŵ dη̃ + lF ξθ + iαŵ dη̃ − F ξθ
∂r 0 0
Z η Z η
1 0 ˆ

0 ˆ

+ Gξr − lG ξr − iβ ŵ dη̃ + lF ξθ + iαŵ dη̃ − F ξθ
R 0 0
The integral operator (2.49) maybe transformed to a matrix operator if we expand the
primary variables fˆ in terms of Chebyshev polynomials. As for the Blasius case, we use an
odd representation for the primary variables and an even representation for the secondary
variables. This parity choice implicitly assumes that the primary variables and all of their
even order derivatives vanish as η → 0, or equivalently, z → ∞ in the physical domain.
As discussed in section 1.4, it is reasonable to assume that this property holds for both
velocity and vorticity variables and so we proceed to expand as described. It should be
noted that a reverse parity can similarly be chosen, with an even representation for the
primary variables and an odd representation for the secondary ones. However, the first
case is subtly simpler as there is no halving of the first term, making the integral operators
slightly more straightforward and easier to deal with. Truncating the expansion to finite
order, we have, for an arbitrary primary variable fˆ,
N
X
fˆ(η) = fˆk T2k−1 (η)
k=1
so that
XN Z η
s=l fˆk T2k−1 (η̃)dη̃
k=1 0
N
l l X T2k (η) T2k−2 (η) T2k (0) T2k−2 (0)
= (T0 (η) + T2 (η) − T0 (0) − T2 (0))fˆ1 + fˆk − − +
4 4 k=2 k k−1 k k−1
( N
) N
lfˆ1

l X ˆ T2k (0) T2k−2 (0) X l ˆ
= − fk − T0 (η) + T2 (η) + (fk − fˆk+1 )T2k (η)
4 k=2 k k−1 4 k=2
4k
N
! N
X lfˆ1 X l
= ak fˆk T0 (ζ) + T2 (η) + (fˆk − fˆk+1 )T2k (η)
k=1
4 k=2
4k
where
(−1)k l(1 − 2k)

l T2k (0) T2k−2 (0)
a1 = 0, ak = − − = (k > 1)
4 k k−1 4k(k − 1)
82
In practise, for the computational implementation of the routine, since integration in-
creases the order of the polynomial approximation, the highest order term is set equal
to zero. The requirement that there is exponential decay of the disturbance variables
is sufficient to ensure that the coefficients of the Chebyshev expansion converge to zero
faster than any inverse power of their order, referenced to Canuto et. al. [12] in Davies
and Carpenter [26]. Thus, taking a large enough N should make the effect of the trunca-
tion negligible. In computational tests, it was found that the eigenvalues could be fully
resolved when using a discretisation of order N = 64. As for the Blasius case, tables of
convergence are given in appendix A.3.
Solving the Eigenvalue Problem
To illustrate the process of dividing through by η 2 and subsequently solving the dispersion
relation, we present the solution procedure only for the vorticity transport equation of the
radial component. Dividing through by η 2 and applying the mapping to equation (2.44b)
gives
2
1 1 η ∂Nθ 2 ˆ 1
−iω ξˆr + iβ −α2 − β 2 ξˆr

Nz − − − ξθ + iαŵ −
η2 η2 l ∂η R R
2

1 d 2 d 2ˆ

− 2 2η + η η ξr = 0
l dη dη 2
| {z
}
∂ ∂2
= 6η 2 +6η 3 ∂η +η 4 ξ̂r
∂η 2
which can be condensed to
1 ∂Nθ 2 ˆ 1 1
−iω ξˆr + iβ N̂z + α2 + β 2 ξˆr − 2 D2 ξˆr = 0

− ξθ + iαŵ + (2.52)
l ∂η R R l
where
∂ ∂2 1
D2 = 6η 2 + 6η 3 + η4 2 and N̂z = Nz
∂η ∂η η2
∂ N̂z
The N̂z and ∂r
terms are calculated in practise by incorporating the division by η 2
directly into the collocation values for the base flow profiles. Thus,
Z η Z η
N̂z = Gξˆr − lĜ0
ξˆr − iβ ŵ dη̃ + lF̂ 0
ξˆθ + iαŵ dη̃ − F ξˆθ
0 0
∂ N̂z
where F̂ and Ĝ represent the divided mean flow quantities. The procedure for the ∂r
term is similar, and likewise, a similar procedure is employed for the multiplications by η 2
that occur in Nr and Nθ due to the single integral operator. Following the same method
83
as for the two-dimensional configuration in section 1.4, we integrate the perturbation
equations (2.44) twice with respect to η to give

2 1 1 1
I2 −iω ξˆr + iβ N̂z − ξˆθ + iαŵ + α + β ξˆr +
2 2
I1 Nθ − 2 Jξˆr = 0

R R l Rl
" #
∂ N̂z 2 1 1 1
I2 −iω ξˆθ − ξˆr − iβ ŵ + α2 + β 2 ξˆθ − I1 Nr − 2 Jξˆθ = 0

+
∂r R R l Rl
!
iα ξˆθ 1
I2 − α2 − β 2 − iβ ξˆr + iαξˆθ + + 2 Jŵ = 0
R R l
where, similarly to equations (1.58) and (1.59), the operators I1 f , I2 f and Jf are defined
as
Z
I1 f = f
ZZ
I2 f = f
Z
4
Jf = η f − 2 η3f
and the boundary conditions (2.45) transform under the division by η 2 to
ŵ(1) = 0
Z 1
ξˆθ + iαŵ dη = 0
Z0 1
ξˆr − iβ ŵ dη = 0
0
We can write this system as a polynomial eigenvalue problem in α as

  
α2 F0 + αF1 + F2 ξˆ
   r
 α S0 + αS1 + S2  ξˆθ  = 0
 2
(2.53)
 
  
2
α T0 + αT1 + T2 ŵ
where {F, S, T}k , (k = 1, . . . , 3) are (N × N ) matrices corresponding to the first, second

and third equations respectively and N is the order of the Chebyshev discretisation. The
solution method is identical to that discussed in section 1.4.3 and a discussion of spurious
eigenvalues and local eigenvalue iteration is given there.
2.2.2 Results and Validation for the Steady Case
In the interest of validation of the results of the velocity-vorticity formulation, we present

some established results and compare them against the new method. For the validation
84
procedure, we will concentrate mainly on the spatial setting, where wavenumbers α are
calculated from a prescribed real ω. In the alternative scenario, the so called temporal
analysis, the dispersion relation (2.53) is simplified to a generalised linear eigenvalue
problem and becomes   
ω F̂ + F̂2 ξ
 1   r
 ω Ŝ1 + Ŝ2  ξθ  = 0 (2.54)
  
  
ω T̂1 + T̂2 w
We will use this solution method in section 2.2.4 when talking about monochromatic DNS,
although presently, the following section will describe the careful numerical validation of
the new formulation in the spatial setting.
Direct Eigenvalue Comparisons
Figure 2.11 shows the global spectrum of radial eigenvalues for stationary disturbances
with R = 286 and β = 0.077. These parameters were chosen to exactly match a similar
diagram in Thomas [71], but are important in their own right as they are the close to the
critical parameters for the onset of convective instability. The numerical eigenvalues can
be identified as those which differ for the two discretisation orders N = 96 and N = 128,
while the genuine eigenvalues are shown on the figure as filled-in circles. This diagram
shows excellent agreement with Thomas [71].
Table 2.1 gives an overview of previously published eigenvalue results relating to the
critical parameters for the onset of both type I and type II instabilities and comparison of
these with the current work. Some small discrepancies are clearly visible between several of
the results, which could be due to several factors relating to the difference in formulations
or numerical methods by different authors, or to the fact that some authors round up
or down to use only physical integer values of azimuthal mode number n, while some
treat β, and therefore n, as continuously varying. Garrett et al. [36] provide a relatively
up-to-date discussion of these inconsistencies in their appendix. As mentioned in section
2.2.1, in deriving the velocity-vorticity formulation perturbation equations, we apply the
parallel-flow approximation at a different stage to other authors, and as a result should
expect some small deviation in our results than those in the literature. The non-parallel
correction to the local, eigenvalue analysis is shown by Malik and Balakumar [51] not to
be significant, and in practise the deviations between our formulation and that of others
85
0.2
0.1
-0.1
-0.2
-1 -0.5 0 0.5 1
Figure 2.11: Radial wavenumber spectrum for R = 286, β = 0.077 and ω = 0. The most
unstable eigenvalue is shown as a filled in circle, while the ◦ and × correspond to the
Chebyshev discretisation orders N = 64 and N = 96 respectively.
1

have turned out to be small. One would expect the correction to be O R
, and this is
certainly apparent in our results, as larger discrepancies are seen when R is smaller in
magnitude. In each case presented below, the temporal frequency ω was prescribed while
neutral values of α, β and R were calculated using the arclength continuation method
described in section 1.5.1. For the non-stationary (ω 6= 0) disturbance validation, the
parameter ω was selected to be identical to that in Thomas [71], to give appropriate
comparisons.
86
Eigenvalue Validation against Literature
Mode Reference ω Rc βc αc
Malik [50] 0 285.36 0.07759 0.38402
Cooper & Carpenter [19] 0 285.36 0.0776 0.38451
Thomas [71] 0 290 0.077 0.3779
Type I
Garrett et. al. [36] 0 286.05 0.0775 0.38419
Appelquist [3] 0 286.05 0.0776 0.38338
Current Work 0 285.55 0.0772 0.3818
Malik [50] 0 440.88 0.04672 0.13228
Cooper & Carpenter [19] 0 440.87 0.0466 0.13159
Thomas [71] 0 451 0.04641 0.1336
Type II
Garrett et. al. [36] 0 450.95 0.04634 0.13067
Appelquist [3] 0 452.97 0.0468 0.13227
Current Work 0 439.95 0.0468 0.13186
Thomas [71] -0.01023 284 0.09379 0.36064
Type I
Current Work -0.01023 282.49 0.0949 0.37277
Lingwood [43] 0.1225 64.4 -0.106 0.276
Type II Thomas [71] 0.1225 65 -0.106 0.2661
Current Work 0.1225 71.958 -0.1047 0.2940
Table 2.1: The temporal frequencies ω for this data are prescribed while neutral values of
α, β and R were calculated using the arclength continuation method described in section
1.5.1. For the non-stationary (ω 6= 0) disturbance validation, the parameter ω was selected
to be identical to that in Thomas [71], to give appropriate comparisons.
Neutral Curve
The algorithm for calculating the neutral curve for the rotating disk configuration is iden-
tical in most respects to that for the Blasius configuration described in section 1.5. The
only notable difference is the inclusion of the azimuthal mode number β, but the algorithm
can be applied analogously with ω replaced everywhere by β, provided we impose some
additional constraint on ω. In practise this additional constraint is to specify ω as the
desired temporal frequency of the disturbance, with the stationary case corresponding to
87
ω = 0. The radial and azimuthal wavenumber neutral curves for stationary disturbances
are shown in figures 2.12 and 2.13.
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
250 300 350 400 450 500 550
Figure 2.12: Radial growth rate contours for stationary disturbances, with the neutral
curve αi = 0 (—) and αi = −0.25 (- - -), αi = −0.5 (· · · ), αi = −0.75 (◦−).
80
60
40
20
0
250 300 350 400 450 500 550
Figure 2.13: Azimuthal wavenumber contours for stationary disturbances, with the neu-
tral curve αi = 0 (—) and αi = −0.25 (- - -), αi = −0.5 (· · · ), αi = −0.75 (◦−).
88
To aid with validation of the new formulation, data from Appelquist [3] was made available
to the author and a small discrepancy between the two neutral curves was noticed in the
particularly sensitive area around the onset parameters for the type II instability, as seen
in figure 2.14.
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
250 300 350 400 450 500 550
Figure 2.14: Comparison between radial neutral curves for stationary disturbances from
Appelquist [3] (- - -) and current work (—).
Initially, it was believed that this discrepancy was due to the difference in formulations
and the inclusion of the extra non-parallel terms but in fact it can be accounted for by
examining the difference in the spatial growth rates, αi , from the data sets of Appelquist
[3] and the current work. Figure 2.15 shows these frequencies. Clearly, the frequencies
from Appelquist [3] are much larger in magnitude than those calculated here and by
definition, on the neutral curve we should have αi = 0. The tolerance for convergence of
αi is set during the numerical arclength continuation procedure to be 10−8 but increasing
this tolerance to around 10−3 shows that we can recover Appelquist’s [3] curve almost
exactly, as in figure 2.16.
89
10 -4
-5
450 452 454 456 458 460
Figure 2.15: Spatial growth rates near the onset of the type II instability. Comparison
between Appelquist [3] (◦) and current work (×).
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
250 300 350 400 450 500 550
Figure 2.16: Comparison between radial neutral curves for stationary disturbances from
Appelquist [3] (- - -) and current work (—) when tolerance is increased to investigate
sensitive area around type II onset.
90
2.2.3 Absolute Instability in the Steady Rotating Disk Bound-
ary Layer
This section will very briefly illustrate the apparent existence of a local absolute instabil-
ity in the rotating disk boundary layer. In order to distinguish between convectively and
absolutely unstable disturbances, we must study the time evolution of an impulsive exci-
tation of the disk surface. An impulse centred at re and t = 0 for a particular azimuthal
wavenumber β can take the form of a boundary condition on the wall-normal component
w such that
w(0; r, θ, t) = δ(r − re )δ(t)eiβRθ (2.55)
where δ denotes the Dirac delta function. For this particular azimuthal wavenumber β,
the response to the impulsive forcing (2.55) is given in terms of a Green’s function such
that
∂ ∂
D(i , −i ; β, Re)G(r, t) = δ(r − re )δ(t) (2.56)
∂t ∂r
where D is the dispersion relation (2.35). In a frame of reference fixed by the perturbation,
we can quantify this response into the following conditions for absolute and convective
instability:
• G(t → ∞, r) = 0 =⇒ convective instability
• G(t → ∞, r) = ∞ =⇒ absolute instability
Without entering into the extensive analytic detail which can be found in Lingwood [47]
& Lingwood [43], we will briefly explain the method for finding absolute instability and
detail the critical parameters for its onset.
Absolute instabilities can be identified numerically by following the paths of two solu-
tion branches in the α-plane. A so-called pinch point occurs when waves propagating in
opposite directions coalesce, producing a branch-point singularity in the ω-plane. Certain
criteria must be satisfied in order for branch-point singularities in the dispersion rela-
tion to correspond to an absolute instability, and these conditions are encapsulated in
the Briggs [15] criterion. If the two solution branches originate in distinct halves of the
α-plane when ωi is large and positive and if coalescence occurs before ωi = 0 then an
absolute instability is present. For example, if this pinch point occurs at some temporal
91
wavenumber ω0 then at the pinch point we must have
∂ω0
=0 (2.57)
∂α
and by Briggs’ criterion, the flow is absolutely unstable if I(ω0 ) > 0.

Figure 2.17 shows these pinch points for some values of ωi chosen for comparison
with Lingwood [47]. The lower branch can be identified as what is known as the type
III mode, discovered by Mack [49] and mentioned in section 2.1.4. Before Lingwood’s
[43] discovery of this absolute instability, little attention had been paid to this mode,
since it has relatively large positive values of αi and so is heavily damped. However,
the coalescence between this and the type I mode is crucial for the onset of absolute
instability. Lingwood [47] states that above a critical Reynolds number of R = 507.3,1
and azimuthal wavenumber β ≈ 0.126 (or n = 68), these pinch points have I(ω0 ) > 0
as required by the absolute instability criterion. A finite range of β values were found
by Lingwood [43] to define the absolutely unstable region, but in figure 2.17, the value
of β = 0.126 has been chosen simply to demonstrate the existence of absolute instability.
These plots can be compared with similar ones in Cooper and Carpenter [19] and also in
Lingwood [47]. While this analysis appears to predict an absolute instability in the flow
configuration, it is important to note the global subtleties in the radial homogenisation
procedure, which are first noted by Davies and Carpenter [27]. A brief overview of the
global stability properties of the flow is given in chapter 4 and detailed discussions of the
global behaviour are given in many references such as Davies and Carpenter [27], Davies
et al. [28], Thomas and Davies [72], Healey [40], Thomas and Davies [74], Appelquist et al.
[4, 6]. The true behaviour of this absolute instability is as yet unresolved and is still the
focus of a great deal of ongoing work by the aforementioned authors, as can be seen in
Thomas and Davies [75].
We will refer back to this absolute instability at several points throughout the forth-
coming report and as mentioned, it will be particularly important when discussing global
stability in chapter 4. However, for the time being the reader should simply be aware of
its existence and the implications for the local analysis.
1
Stated as R = 510.625 in Lingwood [47] but subsequently corrected.
92
0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
(a) ωi = 0.01 (b) ωi = 0.004
0.2
0.1
-0.1
-0.2
-0.3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
(c) ωi = 0.000289
Figure 2.17: Coalescence of type I and type III spatial branches for azimuthal mode
number β = 0.126 and Reynolds number R = 530. A pinch point can be seen to occur
for ωi ≈ 0.000289 > 0. Note that the two solution branches originate from the upper and
lower parts of the complex α plane, as is required for a pinch point to be associated with
absolute instability by Briggs’ criterion.
93
2.2.4 Numerical Methods for Linear Stability - Monochromatic
Direct Numerical Simulations
As discussed in the introduction to section 2.2, there are three approaches which we will
take to analyse the stability of the boundary layer. This section describes the second
of these three, namely the alteration induced by retaining the time dependence in the
stability equations and only decomposing the perturbation variables as
ˆ t)ei(αr+βRθ)
u = û(z, t)ei(αr+βRθ) , ξ = ξ(z, (2.58)
Substituting into (2.36) and applying the parallel flow approximation replacing r by R
leaves the set of time-dependent ODEs

∂ξr ∂Nθ 2 1 ˆ −
2 1 2iβ
+ iβNz − − (ξθ + iαw) = ∇ ξr − ξθ (2.59a)
∂t ∂z R R r2 R

∂ξθ ∂Nr ∂Nz 2 1 ˆ 2 1 2iβ
+ − + (ξr − iβw) = ∇ − 2 ξθ + ξr (2.59b)
∂t ∂z ∂r R R r R

ˆ 2 1 ∂ξr ∂(rξθ )
∇w= − (2.59c)
r ∂θ ∂r
where
2
ˆ 2 = −α2 + iα − β 2 + ∂
∇ (2.60)
R ∂z 2
The convective terms N, and the boundary and integral conditions may still be described
by (2.51) and (2.45) respectively while the secondary variables (2.37) become
Z ∞
ur = − (ξθ + iαw) dz (2.61a)
Z ∞z
uθ = (ξr − iβw) dz (2.61b)
Zz ∞
ξr
ξz = + iαξr + iβξθ dz (2.61c)
z R
The flow is disturbed for a given azimuthal mode number n = βR by a temporally
localised impulsive forcing of the form
η(θ, t) = b(t)eiβRθ (2.62)
where b defines the time-dependent amplitude, given by

2 2
b(t) = (1 − e−σt )e−σt (2.63)
The first term of b acts as a continuous Heaviside function to scale the forcing up from
zero and σ is a parameter which dictates the timescale over which this takes place. Figure
94
0.25
0.2
0.15
0.1
0.05
0
0 2 4 6 8 10 12
Figure 2.18: Typical temporal development of impulsive excitation.
2.18 shows the temporal evolution of this impulse at the forcing location for some σ. The
solution method employed is similar to that described in section 2.1.1 for the solution
of the time-dependent base flow fields and involves a spectral scheme in the wall-normal
direction with a predictor-corrector scheme for the temporal integration. The spectral
expansions in the wall-normal direction are given in terms of Chebyshev polynomials as
described in full in section 2.2 and are identical to the ones used for the linear normal
mode stability analysis described there. We state them here for completeness however,
as there is a subtle difference involving the azimuthal mode number worth mentioning.
As explained in the previous section, the primary perturbation variables {ξr , ξθ , w} are
expanded in terms of odd Chebyshev polynomials and mapping the physical coordinate
z ∈ [0, ∞) to the computational coordinate η ∈ (0, 1] by the usual mapping (1.40) leads
to an expansion of the form
N
!
X
f (r, θ, η, t) = fk (r, t)T2k−1 (η) einθ (2.64)
k=1
which has been restricted to an individual azimuthal wavenumber n = βR.

There are many similarities between the time-dependent simulation and the eigenvalue
problem presented in section 2.2, and much of what has already been discussed will be
relevant here. The important difference is the inclusion of temporal integration, which is
95
achieved by the backward three-level scheme of section 2.1.1, namely
l
∂f 1
3f l − 4f l−1 + f l−2

= (2.65)
∂t 2∆t
where we have used the notation
fl = f
t=l∆t
Similarly to section 2.1.1, the convective terms contained in N, as well as the Coriolis
terms are treated explicitly via a predictor-corrector method in the temporal integration,
while the remaining viscous terms including the second z-derivative term are treated
implicitly. The predictor step for the convective terms is given by
(Nl )p = 2Nl−1 − Nl−2
= 2(ΞB × ul−1 + ξ l−1 × UB ) − (ΞB × ul−2 + ξ l−2 × UB )
while the corrector step is defined by
(Nl )c = (Nl )p = ΞB × (ul )p + (ξ l )p × UB
where (ul )p and (ξ l )p are the disturbance velocity and vorticity fields determined from the
predictor stage. The products involving the base flow quantities UB and ΞB are treated
with a pseudo-spectral scheme which uses a Fast Fourier Transform to convert between
Chebyshev coefficients and collocation values for the multiplications.
Validation Against Eigenvalue Solution Procedure
Figure 2.19 shows the temporal development of ξθ (z = 0), excited by an impulsive forcing
of the form (2.62) for R = 500, n = 32 and α = 0.3. These parameters were chosen as
an example, so that the system would be unstable by the eigenvalue analysis of section
2.2.1.
In order to get viable data against which we may compare our eigenvalue solutions,
we must decompose a perturbation variable f as
f = fˆ(z)ei(αr+βRθ−ωt) (2.66)
so that the dominant temporal growth rate and frequency ω can be computed from the
simulation data by means of the formula
i ∂A
ω= (2.67)
A ∂t
96
10 5
1.5
0.5
-0.5
-1
-1.5
0 0.2 0.4 0.6 0.8 1
Figure 2.19: Temporal response (—) to impulsive disturbance of ξθ (z = 0) for R = 500,

n = 32 and α = 0.3 including wavepacket envelopes (- - -). T denotes the disk rotation
period.
where A is the amplitude of some flow variable. The disturbance variable which was
selected for the results presented below was the azimuthal vorticity component evaluated
at the disk surface but other specifications for the disturbance amplitude could likewise
be used. Provided that ω is found not to vary too rapidly in time, it can be interpreted
as the complex frequency displayed by the disturbance at a particular instant of time.
Figure 2.20 shows the convergence of the temporal growth rate calculated by (2.67) to
the genuine eigenvalue calculated by the dispersion relation (2.54).
Figure 2.21 shows comparisons between the complex frequencies ω calculated from the
eigenvalue dispersion relation (2.54) and the simulation data for varying R. The simula-
tion was conducted until a steady growth rate was achieved and all transient behaviour
had passed out of the domain of interest. Throughout the following discussions, we will
refer to this time after which the growth rates are constant as T∞ , which is identified by

i ∂A
ω(T∞ ) ≈ lim (2.68)
t→∞ A ∂t
Typically, T∞ is chosen to be between two and four periods of disk rotation.
Before proceeding with our discussion of the analysis, it is important at this stage to
note a distinct difference between this impulse response and that presented in chapter
97
1
0.8
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5
Figure 2.20: Convergence of growth rates from monochromatic DNS calculated by (2.67)
(—) to genuine eigenvalue calculated by the dispersion relation (2.54) (- - -) as t → T∞
for R = 500, n = 32 and α = 0.3. T denotes the disk rotation period.
4 and the works on the global stability of this configuration associated to Davies and
Carpenter [27], Lingwood [43] and [6]. Here, we are, in effect, setting up an initial con-
dition f (z, t0 )eiαr einθ and following the development in the space of normal modes with
a prescribed radial dependence. Therefore, this impulse is localised in time but not in
space with regard to the radius; an important distinction to that of chapter 4.
98
10 -3
6
-2
-4
250 300 350 400 450 500
Figure 2.21: Comparison between eigenvalues calculated by the dispersion relation (·)
(2.54) and growth rates from monochromatic DNS (—) calculated by (2.67) for n = 32
and α = 0.3. The simulations were conducted until the growth rates have settled to the
dominant value where t ≥ T∞ .
99
2.2.5 Numerical Methods for Linear Stability - Direct Numeri-
cal Simulations with Radial Dependence
Finally, we discuss the third method mentioned in section 2.2, which retains the radial
dependence in the DNS. This numerical method is described in detail in Davies and
Carpenter [26] although in reality the schemes have since evolved to attempt to alleviate
numerical instabilities.
The main alteration here over section 2.2.4 is the lack of a normal-mode approximation
in the radial direction, leading to the need to calculate r-derivatives of the perturbation
variables. The finite difference discretisations for the first and second radial derivatives
are standard and described at length in Fasel et al. [33]. All that it is pertinent to
mention here is that the schemes used here have been chosen to be non-compact and
centred. The more precise details of the schemes are relegated to appendix B.3.2 and
the reader is referred there or to Fasel et al. [33] for more details. The inner and outer
radial boundaries are both placed well away from the area of interest and a discussion of
inflow and outflow conditions is given in B.3.1. Wavelike or null outflow conditions may
be used, as described in Davies et al. [28] although in practise, we usually ensure that the
simulations are terminated before any significant disturbance reaches the location of the
radially outward computational boundary.
Validation Against Eigenvalue Solution Procedure
Following Davies and Carpenter [26], in order to have simulation data comparable to the
local eigenvalue results, we excite a periodic disturbance on the disk surface by introducing
a forcing of the form
η(r, θ, t) = a(r − re )b(t)einθ (2.69)
to give a localised Gaussian pulse centred at position re , where λ specifies the radial extent
of the forcing.
The time-dependent amplitude (2.63) may be adapted so as to give a time-periodic
excitation, for instance by setting
b(τ ) = h(τ )e−iω0 t (2.70)
where ω0 is a prescribed temporal frequency and h(τ ) = (1−e−στ ) is the continuous Heav-
iside function used to scale the forcing up from zero. Disturbances which are stationary
100
with respect to the surface may be considered by setting ω0 equal to zero.
Locally defined spatial wavenumbers α may be computed from the simulation data by
means of the formula
i ∂A
α=− (2.71)
A ∂r
where A is the amplitude of some flow variable. Similarly to section 2.2.4, the disturbance
variable which was selected for the results presented below was the azimuthal vorticity
component evaluated at the disk surface but other specifications for the disturbance am-
plitude could likewise be used. Provided that α is found not to vary too rapidly in
space, it can be interpreted as the complex wavenumber displayed by the disturbance at
a particular radial position and instant of time.
Figure 2.22 shows the radial evolution of a disturbance generated by stationary forcing
for R = 500 and n = 32. These parameter choices ensure that the configuration is
unstable according to the eigenvalue analysis. For this figure, the position of maximum
forcing is located at re = 500, although this is somewhat arbitrary since we have applied
the parallel flow approximation, replacing r by R in the equations and hence treating all
radial locations as equivalent. As already alluded to, this homogenisation of the radial
direction is, strictly speaking, unphysical and further discussion of this is given in chapter
4. Similarly to section 2.2.4, the simulations were conducted until a steady growth rate
was achieved and all transient behaviour had passed out of the domain of interest so that
T∞ can be identified with
i ∂A
α(T∞ ) ≈ lim − (2.72)
t→∞ A ∂r
Typically, T∞ is chosen to be between two and four periods of disk rotation. Davies and
Carpenter [26] state that the variation of the primary variables may be fully resolved
using a Chebyshev expansion involving N = 48 polynomials and a radial resolution of
about ∆r ≈ 1. Each of the following simulation plots was produced using N = 48, a
radial resolution of ∆r = 1.25 and a time discretisation of ∆t = 0.625. In all of the
simulations, the computational domain extends well beyond the limits suggested by the
figures in order to ensure no computational edge effects creep into affect the results.
Figure 2.23 shows radial growth rates for azimuthal mode number n = 32 and max-
imum wall displacement located at r = 350 for a stationary disturbance, with the true
radial inhomogeneity incorporated. Away from radial locations near the greatest wall
deformation, the results from the DNS and the eigenvalue dispersion relation (2.53) agree
101
10 4
1.5
0.5
-0.5
-1
-1.5
500 550 600 650
Figure 2.22: Radial evolution of ξθ (z = 0, τ > T∞ ) for R = 500 and n = 32 (—) along
with wavetrain envelope (- - -).
well with each other, providing confidence in the newly developed solution procedure for
the eigenvalue analysis. The oscillations in the curve near the maximum forcing at r = 350
are due to near-field effects and the fact that in these locations the disturbances do not
have a well-defined radial wavenumber that can be computed using equation (2.71) for
t > T∞ .
102
0.02
-0.02
-0.04
-0.06
-0.08
350 400 450 500
Figure 2.23: Plots of αr and αi for a stationary disturbance with azimuthal wavenumber
n = 32, showing comparisons between DNS (·) and local results (—).
103
Chapter 3
Local Stability of the Periodically

Modulated Rotating Disk Boundary
Layer using a Velocity-Vorticity
Formulation
This forthcoming exposition details the extension of the methods of chapter 2 to the flow
configuration that has been the main focus of the work conducted throughout this thesis.
We will, as an introduction to the chapter and in the interest of clarity, briefly review the
notation and key concepts introduced previously.
Consider a disk of infinite extent rotating with a periodically modulated angular ve-
locity beneath an otherwise stationary fluid. The resulting flow has radial, azimuthal
and vertical components and admits the exact similarity solution to the Navier-Stokes
equations, as explained in section 2.1.1. The motion of the disk surface is described by
the rotation rate
Ω∗ (t∗ ) = Ω∗0 + φ∗ cos(φ∗ t∗ ) (3.1)
where Ω∗0 may be thought of as analogous to the constant rotation rate in the steady case,
while and φ∗ denote the angular displacement and angular velocity of the modulation
respectively. We use spatial and temporal scalings
rL∗ ∗ ∗
r ∗
∗ ν
δk = , τ = Ωt (3.2)
Ω∗0 δk∗ 0
and a local velocity scaling is obtained by using the circumferential disk velocity rL∗ Ω∗0
104
at some dimensional reference radius rL∗ . This local scaling gives a Reynolds number R,
associated with the steady rotation of the disk as
rL∗ Ω∗0 δk∗ rL∗

R= = = rL (3.3)
ν∗ δk∗
where we identify rL as a local, non-dimensional radial position. As in section 2.1.1, we

also define a Reynolds number, Rs associated with the Stokes layer, given by
s
rL∗ φ∗0 δs∗ φ∗ √
Rs = = rL ∗
= ϕR (3.4)
ν∗ Ω0
q
φ∗
where we have defined ϕ = Ω∗0
. The resulting three-dimensional base flow may be
expressed as the solution to the system
∂ 2F

∂F 1 2 2 ∂F
= F −G +H − (3.5a)
∂τ R ∂z ∂z 2
∂G ∂ 2 G

∂G 1
= 2F G + H − (3.5b)
∂τ R ∂z ∂z 2
∂H
= −2F (3.5c)
∂z

ϕ
F (0, τ ) = H(0, τ ) = 0, G(0, τ ) = 1 + Uw cos τ (3.6a)
R
F → 0, G → 0 as z → ∞ (3.6b)
√
Rs ϕ
where Uw := R
. With a laminar flow control motivation in mind, we constrain the
variation in the wall velocity to be relatively small with respect to the rotation rate,
and so we impose the condition Uw < 0.2 for the remainder of this work. In turn, for
sufficiently high ϕ and R in the region of interest, this condition keeps the Reynolds
number associated with the Stokes layer, Rs, to be small compared with the Reynolds
number associated with instability onset in the semi-infinite Stokes layer of Blennerhassett
and Bassom [11].
The motivation for the addition of a periodically modulated rotation rate is discussed
in section 1.1.1 and originates with Thomas et al. [76], who show that inducing a small
amount of oscillation into an otherwise steady flow can have stabilising effects. The in-
troduction of this modulation creates a time-modulated unidirectional flow, and Thomas
et al. [76] detail their stabilising effects using a Stokes layer added to plane Poiseuille
and Hagen-Poiseuille flows. They complement the work of Kelly and Cheers [42] and
105
Von Kerczek [81] who studied modulation of plane Couette flow and also found stabilisa-
tion. Thomas et al. [76] found that the steady flow was stabilised for certain frequencies
of the oscillation but destabilised for others. Wise and Ricco [82] also show that turbulent
drag reduction may be achieved using oscillation of flat plates and rotating disks. The
work presented in this report will be of a more mathematically fundamental nature than
that described in Wise and Ricco [82], and will attempt to explain the underlying concepts
of the stabilisation achieved by this modulation.
The investigation into the local stability properties of the modulated system can be
executed in several ways, and we will describe the different possible methods and attempt
to unify the results to give a detailed understanding of the effects of this modulation on
stability. The following procedure encompasses three distinct investigatory approaches.
Linearised direct numerical simulations using the vorticity-based methods of previous
sections are complemented by a local in time linear stability analysis, which is made
possible by imposing an artificial frozen base-flow approximation. This localised analysis
is deployed together with a more exact global treatment based upon Floquet theory,
which avoids the need for any simplification of the temporal dependency of the base-flow.
Before moving on to the time-dependent simulations, we begin the discussion with Floquet
theory, and with laminar flow control techniques and suppression of crossflow vortices on
swept wings foremost in mind, we will focus on disturbances which are stationary with
respect to the disk surface.
3.1 Eigenvalue Analysis via Floquet Theory

Similarly to the Stokes layer scenario discussed in section 1.4.4, an understanding of the
disturbance structures may be gleaned by a consideration of Floquet theory. We will use
the same temporal and spatial scalings as in section 2.1.1, which, with the boundary layer
q ∗
thickness δk∗ = Ων ∗ , gives rise to the spatial and temporal scalings
0
z∗ rL∗ ∗ ∗
z= τ= Ωt
δk∗ δk∗ 0
To consider the development of small disturbances to the unsteady base flow UB , we

perturb as for the Stokes layer in section 1.4.4 and write
u(x, τ ) = UB (z, τ ) + u0 (x, τ ), ξ(x, τ ) = Λ(z, τ ) + ξ 0 (x, τ )
106
Following the method developed by Hall [39], while making use of the theorems and ideas
discussed in section 1.4.4, we take the Floquet mode solution
f (r, θ, z, τ ) = fˆ(z, τ )eµτ ei(αr+βRθ) + c.c. (3.7)
where fˆ(z, τ ) is a time-periodic function with the same period as that of the modulation
and all exponential growth of fˆ(z, τ ) incorporated into the Floquet exponent eµτ . In this
case, since the period of modulation, given by the boundary condition on the azimuthal
velocity component, is T = ϕ
2πR
, we have that fˆ(z, τ ) is periodic with period T . The c.c.
denotes the complex conjugate, which is added to ensure that the disturbance is real. As
for the Stokes case, the quantities of interest here will be µ, as its real part specifies the
temporal growth of the disturbance, and α, whose imaginary part gives the spatial growth
rate.
Substituting into the perturbation equations (2.36), linearising as usual as neglecting
the non-parallel O R12 terms as for the steady case gives

∂2

∂ξr ∂Nθ 2 1 2 2
+ µξr + iβNz − − (ξθ + iαw) = −α − β + 2 ξr (3.8a)
∂τ ∂z R R ∂z
∂2

∂ξθ ∂Nr ∂Nz 2 1 2 2
+ µξθ + − + (ξr − iβw) = −α − β + 2 ξθ (3.8b)
∂τ ∂z ∂r R R ∂z
2

ˆ +
2 ∂ ξθ
∇ w = iβξ r − iαξ θ − (3.8c)
∂z 2 R
which is similar to equations (2.44) aside for the replacement of the terms multiplied by
∂
−iω with the Floquet temporal derivative structure ∂τ
+ µ.
As discussed for the Stokes layer in section 1.4.4, the method of Hall [39] involves the
decomposition of the perturbation variables into harmonics such that
∞
X
fˆ(z, τ ) = fm (z)eimτ̃ (3.9)
m=−∞
ϕ ϕ
where τ̃ = R
τ and the factor R
is introduced to ensure the harmonics have the same
period as fˆ(z, τ ). A representation for the time dependent part of the base flow is also
required in a similar form, and this is given by
UB = US (z) + UM (z, τ̃ ) (3.10a)

∞
X
S
= U (z) + uk (z)eikτ̃ (3.10b)
k=−∞
107
where US (z) is the steady state base flow solution. In practise, we can calculate the
Fourier coefficients uk (z) by applying a fast Fourier transform on the known values
UM (z, τ̃ ) := UB (z, τ̃ ) − US (z) (3.11)

∞
X
= uk (z)eikτ̃ (3.12)
k=−∞
Validation of this method of calculating the base flow contribution to the system is pro-
vided by use of the analytic Stokes solution of section 2.1.2. Provided the Stokes layer
solution is an appropriate approximation for the parameters in question, we can simply
set u1 (z) = ū−1 (z) = e−(1+i)z , with uk (z) = 0 (k 6= 1) and proceed to solve.
The system (3.8) is solved by a similar method to that of the Stokes layer. In the case
for the steady rotating disk discussed in section 2.2.1, we had a problem of the form
−iωξr + Lr {ξr , ξθ , w} = 0
−iωξθ + Lθ {ξr , ξθ , w} = 0
Lw {ξr , ξθ , w} = 0
where L{r,θ,w} represents all of the terms which are not included in the temporal derivative.
The alteration that the inclusion of the Floquet normal mode structure introduces can
thus be written similarly as
∞ ∞ ∞
!
X ϕ X ∂ X θ
z
im ξr,m + µξr,m + iβ − Nm+k Nm+k + Lr {ξr,m , ξθ,m , wm } eimτ̃ = 0
m=−∞
R k=−∞
∂z k=−∞
∞ ∞ ∞
!
X ϕ ∂ X z ∂ X r
im ξθ,m + µξθ,m + N + N + Lθ {ξr,m , ξθ,m , wm } eimτ̃ = 0
m=−∞
R ∂r k=−∞ m+k ∂z k=−∞ m+k
∞
X
Lw {ξr,m , ξθ,m , wm }eimτ̃ = 0
m=−∞
where we have defined Nk to be the unsteady contributions to the convective terms.

Comparing coefficients of eimτ̃ gives an infinite system of ordinary differential equations
which can be truncated and solved for the eigenvalues using MATLAB’s eigs package.
Essentially, this problem reduces to a system of dispersion relations of the form
∞
X
Dm {µ, α R, n}eimτ̃ = 0 (3.13)
m=−∞
and as for the steady case, there are two distinct types of analysis which can be conducted,
namely temporal and spatial. The temporal approach consists of specifying α ∈ R and
108
calculating µ from (3.13) while the spatial approach specifies µ ∈ R and calculates α from
(3.13). At the time of writing, no literature which deals with the spatial implementation
of Floquet theory could be found and from a physical viewpoint, its interpretation is in
need of some clarification. For the majority of results presented below, unless otherwise
stated, we have fixed the Reynolds number associated with the disk rotation at R = 500
and the azimuthal mode number at n = 32, so as to be sure that instability would be
present in the steady, unmodulated scenario. As has been alluded to previously, the base
flow may be fully specified by the variation of three parameters, namely (R, Uw , ϕ) defined
by
√ s
Rs ϕ φ∗
R = rL , Uw = , ϕ= (3.14)
R Ω∗0
where rL and Rs are defined by equations (3.3)-(3.4) and φ∗ , Ω∗0 denote the dimensional
modulation frequency and unmodulated disk rotation rate respectively.
As for the steady case, the spatial analysis of the dispersion relation (3.13) introduces
two more parameters (n, µ). The restriction to the analysis of disturbances which are
stationary with respect to the constant rotation rate is achieved by setting µ = 0, which
we will do for all spatial analyses unless specifically otherwise stated.
There are several complications involved in the numerical computation of these quan-
tities, so prior to our presentation and discussion of the results, it is pertinent to provide
an overview of these at this stage.
3.1.1 Discussion of Numerical Computation
Temporal vs. Spatial - Computational Times
In general terms, the most physically relevant quantities extractable from (3.13) will be
the spatial growth rates, and we will focus on these for much of the discussion in this
section. However, since the solution procedure is computationally more feasible in the
temporal setting, we consider this first. While presumably not immediately obvious to
the reader, the increase in computational feasibility is as a result of the difference in
order of the polynomial eigenvalue problem (3.13). The temporal analysis, where µ is
calculated, is a linear generalised eigenvalue problem, whereas a spatial analysis would
see the eigenvalue α appear to second order, thereby increasing the complexity of the
problem considerably. This issue was clearly also present in sections 2.2.1, although we
109
paid it little attention since it did not have a great effect on the computational time.
The added complexity in this problem due to the coupled system of differential equations
required to solve the dispersion relation changes this considerably, and we first present a
brief discussion of the eigenvalue routines and associated computational times.
The three relevant eigenvalue routines available in MATLAB are eig, polyeig and eigs.
The routines eig and eigs are capable of solving linear generalised eigenvalue problems of
the form
Aφ = λBφ (3.15)
while polyeig is optimised to solve polynomial eigenvalue problems of the form
λn An + λn−1 An−1 + · · · + λA1 + A0 = 0 (3.16)
While polyeig performs reasonably well for relatively small matrix operators, it pro-
hibitively struggles for the large matrix operators we are attempting to deal with in this
problem. The linear routine eigs has an additional flag which allows a user to instruct
the routine to find the N closest eigenvalues in absolute value to a given initial guess,
although the current version of polyeig unfortunately has no such flag and requires a full
calculation of every numerical eigenvalue in the solution. Therefore, for many of the re-
sults presented for the spatial analysis in later sections, we use the local iteration method
(1.48) to calculate the eigenvalues, while conducting regular checks using the global solver
to ensure convergence to the correct eigenvalue at all times. Table 3.1 illustrates the com-
putational difficulties encountered using when using polyeig for a typical set of rotation
and modulation parameters, and compares the spatial and temporal cases. In each case
for the linear solver, the MATLAB routine eigs was provided with the steady eigenvalue
as an initial guess, and instructed to return the 5 closest eigenvalues in absolute value.
In each case, the order of the Chebyshev discretisation is taken to be NC = 64 and five
terms in the Fourier expansion of the base flow (3.10) were taken, which we label NB = 2.
This selection allowed for max |uNB (z)| < 1e−10 in all cases. The simulations were per-
z∈[0,∞)
formed for a typical set of rotation and modulation parameters, namely R = 500, n = 32,
Uw = 0.2 and ϕ = 10. In the temporal analysis, we have set α = 0.3 and in the spatial
analysis, we have set µ = 0.
110
Computational Time (s)
Number of Harmonics, NF
Linear (eigs) Quadratic (polyeig)
2 6.72 247.82
4 40.08 1358.60
6 80.14 3494.68
10 175.29 N/A
Table 3.1: Comparison between computational times for the solution of the temporal
(linear) and spatial (quadratic) eigenvalue problems arising from the dispersion relation
(3.13). Simulations were performed for a typical set of rotation and modulation param-
eters, namely R = 500, n = 32, Uw = 0.2 and ϕ = 10. In the temporal analysis, we
have set α = 0.3 and in the spatial analysis we have set µ = 0. This was performed with
Matlab version 2017a on a workstation with an Intel Core i5-3470T CPU at 2.90GHz.
Number of Harmonics in Floquet Solution
Section 1.4.4 presented a discussion of the stability properties of the quintessential purely
oscillatory flow that is generated when a flat plate oscillates in the plane beneath a
semi-infinite expanse of fluid, namely the Stokes layer, and provided an overview of the
solution procedure. Table 1.2 shows the number of harmonics required in the truncation of
the infinite system of differential equations that determine the dispersion relation (3.13).
Throughout this work, we will refer to this computational truncation parameter as NF ,
and identify it in the following sense
∞
X NF
X
imτ̃
Dm {µ, α R, n}e =0 → Dm {µ, α R, n}eimτ̃ = 0 (3.17)
m=−∞ m=−NF
In order to achieve eigenvalue convergence for the Stokes layer in section 1.4.4 we required
NF & 0.8αR, which for the Reynolds numbers close to critical is typically in the region
of NF ≈ 200. It should not be expected that the same number would be required here,
since the modulation constitutes a small modification to a steady rotation rate and we do
not have the same separation of temporal scales that we do in the Stokes case.
In fact, as detailed in table 3.2, in almost all cases, NF ≈ 4 is sufficient to achieve
convergence. In practise, when carrying out the calculations, two routines were performed
and compared with differing NF to ensure convergence is always achieved on each run.
111
Variation of µ with Number of Harmonics NF
R α Uw ϕ NF µ
1 0.005289 + 0.009825i
2 0.005289 + 0.019824i
500 0.3 0.2 5 3 0.005289 + 0.029824i
4 0.004229 - 0.011160i
5 0.004229 - 0.011160i
1 0.004779 + 0.019120i
2 0.004206 - 0.020812i
500 0.3 0.2 10 3 0.004206 - 0.000811i
4 0.004206 - 0.000811i
5 0.004206 - 0.000811i
1 0.004779 + 0.0291144i
2 0.004497 - 0.000840i
500 0.3 0.2 15 3 0.004497 - 0.000840i
4 0.004497 - 0.000840i
5 0.004497 - 0.000840i
(3.13) for varying numbers of harmonics NF . For all cases, the azimuthal mode number
n was set to n = 32.
Number of Terms in Base Flow Expansion
The final parameter to be addressed in this section is the number of terms taken in the
Fourier expansion of the base flow. Since the Stokes layer discussed in section 1.4.4 can
be represented exactly as a finite combination of Fourier modes, as displayed in equation
(1.71), this parameter does not exist in that case. However, since we cannot readily
express our base flow exactly with a finite combination such as this, we must introduce a
further truncation parameter NB which we identify with
∞
X NB
X
M ikτ̃ M
U (z, t) = uk (z)e → U (z, t) ≈ uk (z)eikτ̃ (3.18)
k=−∞ k=−NB
112
As explained in section 2.1.2 when discussing the high frequency, small amplitude limiting
behaviour, the dominant behaviour comes in at first order, the expectation is that NB
should be fairly small to achieve convergence. This is indeed the case, as detailed in table
3.3, and in all cases NB ≈ 2 was sufficient for well-resolved results. In practise, when
carrying out the calculations, two routines were performed and compared with differing
NB to ensure convergence is always achieved on each run.
Variation of µ with Number of Harmonics NB

R α Uw ϕ NB µ
1 0.004229 - 0.011160i
500 0.3 0.2 5 2 0.004229 - 0.011160i
3 0.004229 - 0.011160i
1 0.004207 - 0.000812i
500 0.3 0.2 10 2 0.004206 - 0.000812i
3 0.004206 - 0.000812i
1 0.004779 + 0.0291144i
500 0.3 0.2 15 2 0.004497 - 0.000840
3 0.004497 - 0.000840i
(3.13) for varying numbers of terms, NB , in the Fourier expansion of the base flow. For
all cases, the azimuthal mode number n was set to n = 32.
3.1.2 Temporal Analysis
We begin the presentation of the modifications introduced by the periodically modulated

rotation rate by examining the temporal setting, where we specify the radial wavenumber
α as real and use the dispersion relation (3.13) to calculate the eigenvalues µ. In general,
µ will be complex, with the real part µr corresponding to the disturbance growth rate
and the imaginary part µi its frequency.
We will consider the configuration to be stabilised in the temporal setting if
µM S
r < µr (3.19)
113
where µSr and µM
r denote the steady and modulated growth rates respectively. Given
the discussions presented in section 3.1.1, the numbers of harmonics and terms in the
Fourier expansions of the base flow were chosen to ensure convergence of the eigenvalues
in each case. This was achieved in practise by comparing the relative errors between two
distinct orders of approximation and ensuring that the output was negligible. For the vast
majority of cases presented, a choice of NF = 4 and NB = 2, in addition to a Chebyshev
discretisation order of N = 64 was sufficient to ensure convergence.
Figure 3.1 shows the temporal eigenvalue spectrum for R = 500, n = 32, α = 0.3,
Uw = 0.2 and ϕ = 10. The corresponding steady eigenvalue µS = −iωS is indicated on
the diagram, which clearly shows a shift towards 0, or temporal stabilisation, in the real
part of µ in the modulated case.
The parameter choices of Uw = 0.2 and ϕ = 10 were chosen partly with the aid of
hindsight to provide an illustrative example of the stabilisation, but also to lie within a
physically significant range for the applications to laminar flow control. As alluded to in
the opening of this chapter, we can identify Uw as a modulation amplitude which, under
the scalings, becomes a percentage deviation from the steady state. Uw = 0 would corre-
spond to the steady case, while Uw = 0.1 would constitute a swing of 10% back-and-forth
from the steady rotation rate in one modulation period. The quantity ϕ may be inter-
preted as a non-dimensional frequency, and is identified with the number of oscillations
about the steady state in one disk rotation period. As will be the case with all results
presented in this report, the amplitude parameter will be constrained to be small, so as
to provide a valid comparison against the steady case and an application to laminar flow
control. It is important to note, however, that this small Uw assumption is not an inher-
ent feature of our solution procedure, and Uw > 0.2 is easily achievable with the same
method.
This demonstrates our first glimpse thus far of the stabilising behaviour achieved by
the addition of the modulated rotation rate. As we progress through the following sections
we will provide further examination of this behaviour, and attempt to explain some of the
underlying physical processes behind it. While figure 3.1 shows only an isolated parameter
case, figure 3.2 shows the difference in temporal growth rates µSr −µM
r for R = 500, n = 32,
α = 0.3 with varying ϕ and Uw ∈ {0.1, 0.2}.

Intriguingly, we see a peak stabilisation around ϕ ≈ 8, while very low frequencies
114
and very high frequencies contribute little effect to the system. Potential explanations
for this phenomena will be given in later sections, although it should be noted that the
explanations will be conjecture, and that the true reasoning behind this behaviour is
unknown at this time. These figures demonstrate that the stabilisation achieved by the
addition of the modulated rotation rate is more robust than being restricted to an isolated
parameter set.
0.01
-0.01
Eigenvalues
-0.02 Steady
Most Unstavble
-0.03
-0.02 -0.015 -0.01 -0.005 0 0.005 0.01
Figure 3.1: Temporal eigenvalue spectrum for α = 0.3 with Uw = 0.2 and ϕ = 10.
The steady eigenvalue is shown as (×), while the filled in circle corresponds to the most
unstable eigenvalue in the modulated case
115
10 -4
8
0
0 20 40 60 80 100
Figure 3.2: Variation in temporal growth rates for the most unstable mode, with R = 500,
n = 32, α = 0.3, Uw = 0.1 (—), Uw = 0.2 (· · · ) and varying ϕ ∈ (0, 100]. ∆µr :=
µSr − µM
r denotes the difference in the growth rate between steady and modulated cases
with positive ∆µr indicating a weaker growth rate.
116
3.1.3 Spatial Analysis
As explained in the previous section, a global spatial analysis is computationally expensive

and thus the following results will, in the most part, correspond to calculations carried out
using a local iteration procedure first discussed as equation (1.48). This amounts to using
an initial guess for the eigenvalue and using the cubically convergent method described
in Bridges and Morris [14] to iterate towards α. We can iteratively improve this guess by
the formula
2f (αk )
αk+1 = αk − (3.20)
f (αk )2 − f1 (αk )
where
f (αk ) = Tr[D−1 (αk )D0 (αk )]
f1 (αk ) = Tr{D−1 (αk )D00 (αk ) − [D−1 (αk )D0 (αk )]2 }
and D(α) refers to the system of differential equations (3.13) which determine the disper-
sion relation while D0 (α) and D00 (α) are the derivatives of D with respect to α.
The exception to this locally computed method is figure 3.3, which shows a typical
global α-eigenvalue spectrum for stationary disturbances with Uw ∈ {0.1, 0.2} and ϕ ∈
{10, 25}. The corresponding steady, Uw = 0, eigenvalue αS is indicated, with the diagrams
clearly showing a shift towards 0, or stabilisation, of the imaginary part of α in the
modulated cases. Unlike for the steady configuration, the physical interpretation of the
reduction in this radial growth rate is in need of clarification as this study presents the
first implementation of Floquet theory in the spatial setting to the best of the author’s
knowledge. It is anticipated that the analyses and interpretations of the radial growth
rate in the steady case will persist in the modulated scenario but a rigorous mathematical
justification of this is not presented here.
Figure 3.4 shows the difference in spatial growth rates αiS − αiM for stationary distur-
bances with Uw ∈ {0.1, 0.2} and ϕ ∈ (0, 100]. Similarly to the temporal case shown in
figure 3.2, we see a peak stabilisation around ϕ ≈ 8, while very low very high frequencies
contribute little effect to the system. This alludes to an optimal range of frequencies ϕ
for maximum stabilisation in the system.
Heretofore, we have only considered isolated data points in R and n, chosen as exem-
plar parameters to demonstrate the stabilising effects of modulation on the radial growth
rates αi . Figures 3.5 and 3.6 show the radial and azimuthal neutral curves respectively
117
0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
(a) Uw = 0.1, ϕ = 10 (b) Uw = 0.1, ϕ = 25

0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
(c) Uw = 0.2, ϕ = 10 (d) Uw = 0.2, ϕ = 25
Figure 3.3: Radial wavenumber spectrum for stationary disturbances with R = 500,
n = 32 and varying ϕ and Uw . The steady eigenvalue is shown as (×), while the filled in
circle corresponds to the most unstable eigenvalue in the modulated case.
for stationary disturbances with Uw = 0.2 and ϕ = 20. This demonstrates that the sta-
bilising behaviour is not an isolated occurrence for a particular parameter set and shows
the stabilisation across a large range of R and n. Figure 3.7 shows the variation in the
critical Reynolds number Rc for Type I disturbances with Uw = 0.2 across a range of ϕ.
Interestingly, and contrary to other methods of stabilisation studied by other authors
such as radially isotropic and anisotropic surface roughness (see Cooper et al. [21], Garrett
et al. [36]) or surface compliance (see Cooper and Carpenter [19, 20]), the type II mode
seems also to be stabilised by the addition of modulation, at least for the stationary
disturbances considered here. Further investigation of the effect of the modulation on
non-stationary disturbances is reserved for future work, as is a comprehensive study of
the behaviour of the type II mode in the various configurations.
118
10 -3
20
15
10
-5
0 20 40 60 80 100
Figure 3.4: Variation in radial growth rates for most unstable mode calculated from (3.13)
for stationary disturbances, with Uw = 0.1 (—), Uw = 0.2 (· · · ) and varying ϕ ∈ (0, 100].
∆αi denotes the difference in the growth rate between steady and modulated cases with
positive ∆αi indicating a weaker growth rate.
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
250 300 350 400 450 500
Figure 3.5: Radial neutral curves for stationary disturbances, comparing the steady case
(—) with the modulated scenario when Uw = 0.2 and ϕ = 20 (· · · ).
119
80
60
40
20
0
250 300 350 400 450 500
Figure 3.6: Azimuthal neutral curves for stationary disturbances, comparing the steady
case (—) with the modulated scenario when Uw = 0.2 and ϕ = 20 (· · · ).
305
300
295
290
285
0 20 40 60 80 100
Figure 3.7: Variation in the critical Reynolds number Rc for stationary type I disturbances
with varying ϕ and Uw = 0.1 (—), Uw = 0.2 (· · · ). The critical Reynolds number for
onset of instability in the steady case (Rc ≈ 285.55) is shown as (- - -).
120
3.1.4 Discussion
Sections 3.1.2 and 3.1.3 have detailed our first encounters with the stabilisation of the
rotating disk boundary layer by the addition of a modulated rotation rate, and we have
presented and discussed several results which appear to demonstrate a significant sta-
bilisation in both temporal and spatial analyses. We have demonstrated that there is
an intriguing range of somewhat optimal stabilisation in terms of the non-dimensional
modulation frequency parameter ϕ and that the stabilisation persists across a range of
Reynolds numbers, azimuthal mode numbers and modulation parameters.
Similar behaviour is demonstrated by other authors such as Cooper et al. [21], Garrett
et al. [36] when discussing radially isotropic and anisotropic surface roughness or by
Cooper and Carpenter [19, 20] when studying surface compliance, although a distinct
difference and advantage of the method presented here is that there is no modelling of
the surface of the disk involved at any point. We do not discard first order effects as
the modelling of Cooper et al. [21] and Garrett et al. [36] does, and over a period of
modulation, the base flow has a zero average deviation from the steady state, further
details and explanations of which are given in section 2.1.3.
The following sections discuss two further methods for analysing the linear stability
of stationary disturbances in the modulated rotating disk boundary layer, which will
act as validation for the Floquet theory analysis and provide further interesting results
in their own right. We proceed first with an instantaneous frozen profile analysis before
progressing to an examination of fixed radial wavenumber monochromatic direct numerical
simulations. This chapter concludes with an energy analysis intended to illuminate the
fine details of the stabilising behaviour and results relating to direct numerical simulations
of stationary forcing with the full radial dependence of the mean flow included.
121
3.2 Quasi-Steady Stability Theory and the Frozen-
flow Approximation
As discussed in section 1.4.4 when considering the semi-infinite Stokes layer, an instan-
taneous stability investigation may be conducted by freezing the base flow at a specific
time instant and treating it as if it were steady. This allows for the time to be treated as
a parameter, and suggests a normal-mode type solution of the form
f (r, θ, z, τ ) = fˆ(z; τ )ei(αr+βRθ− ω)

R
τ (3.21)
where fˆ is a function which is slowly varying in time. This assumption allows us to neglect
∂ fˆ
the temporal derivative ∂τ
≈ 0 and derive a set of perturbation equations similar to those
in section 2.2.1 given by
∂2

∂Nθ (τ ) 2 1
−iωξr + iβNz (τ ) − − (ξθ + iαw) = −α2 − β 2 + ξr (3.22a)
∂z R R ∂z 2
∂2

∂Nr (τ ) ∂Nz (τ ) 2 1
−iωξθ + − + (ξr − iβw) = −α2 − β 2 + ξθ (3.22b)
∂z ∂r R R ∂z 2
∂2

ˆ 2 ξθ
∇ + 2 w = iβξr − iαξθ − (3.22c)
∂z R
In this system, the base flow UB and thus N, is replaced with the frozen base flow at the
given time instant. This system is complemented by the definition of the secondary vari-
ables (2.37) and the boundary conditions (2.45) of section 2.2.1 to fully specify eigenvalue
dispersion relations for each τ of the form
D(ω, α, β, Re; τ ) = 0 (3.23)
Section 1.4.4 and Ramage [59] give detailed descriptions of the subtleties and difficul-
ties in using this approximation to specify the disturbance evolution properties of the
semi-infinite Stokes layer, indicating that any predictions or conclusions drawn from this
analysis must be treated with caution. However, some insight may be gained by a careful
treatment of this method and in section 3.3 we will utilise this approximation to discuss
the energy balance of the perturbed flow in the modulated rotation rate scenario.
The τ -parameter dependent dispersion relation (3.23) is solved in an identical manner
to the steady equations (2.44) for each frozen base flow field UB (z; τ ), where each time-
instant is treated as if it were steady and τ is taken as a parameter. Similarly to the
122
steady case, we may solve this dispersion relation (3.23) for either α or ω while specifying
the other in a spatial or temporal setting. In keeping with the methodology of section 3.1,
for the results below, we have fixed the Reynolds number associated with the disk rotation
at R = 500 and the azimuthal mode number at n = 32, so as to be sure that instability
would be present in the steady, unmodulated scenario. Throughout the following sections,
the temporal growth rates and frequencies ω are non-dimensionalised so that the period
of disk rotation is 2π. Certain authors and results in the literature may use a time-scale
such that the locally defined growth rates are a factor R larger, leading the disk rotation
period to be 2πR. Since the perturbation quantities are periodic with the period of the
τϕ
modulation, which we will denote Tϕ = 2πR
here.
3.2.1 Temporal Analysis
As for the Floquet analysis, we begin our discussion in the temporal setting, and figure
3.8 shows the variation of µ = −iω in the Argand plane over a period of modulation for
α = 0.3, Uw = 0.2 and ϕ = 10. These parameters were chosen as exemplary, with the
particular intention of lying near to the peak of the optimal stabilisation in terms of the
non-dimensional frequency ϕ shown in figure 3.2. The corresponding steady and Floquet
eigenvalues are also shown on figure 3.8. A reduction in the real part is consistent with
stabilising behaviour, as R(µ) denotes the temporal growth rate here. Figure 3.9 shows
the time histories of the real and imaginary parts of µ over one modulation period together
with the corresponding steady and Floquet counterparts which are necessarily constant
across the period. An important distinction between these results and those of Luo and
Wu [48] for the semi-infinite Stokes layer should be noted, namely that the eigenvalues
of this configuration are indeed periodic, and do not exhibit the head and tail behaviour
described in Luo and Wu [48].
123
10 -3
4
-2
-4
-6
1 2 3 4 5 6 7
-3
10
Figure 3.8: Temporal evolution of µ = −iω across a period of modulation in the Argand
plane for disturbances with R = 500, n = 32, α = 0.3, Uw = 0.2 and ϕ = 10. The
temporally periodic quantity µ is shown as (×) along with the Floquet (+) and steady
(o) eigenvalues.
10 -3 10 -3
7 4
6
2
5
0
4
-2
3
-4
2
1 -6
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) µr (b) µi
Figure 3.9: Evolution of temporal eigenvalue µ across a period of modulation for distur-
bances with R = 500, n = 32, α = 0.3, Uw = 0.2 and ϕ = 10. The temporally periodic
quantities R(µ) and I(µ) are shown as (· · · ) along with the Floquet (- - -) and steady
(—) eigenvalues.
124
3.2.2 Spatial Analysis
Similarly to the results presented when discussing the Floquet approach in section 3.1,
we will consider stationary (ω = 0) disturbances with typical parameter values R = 500
and n = 32. As throughout the rest of this and previous sections, these parameters are
chosen as exemplary, with the particular intention of lying near to the peak of the optimal
stabilisation in terms of the non-dimensional frequency ϕ shown in figure 3.2.
A reduction in the imaginary part of α is consistent with stabilising behaviour, as
I(α) denotes the spatial growth rate here. Figure 3.11 shows the time histories of the real
and imaginary parts of µ over one modulation period together with the corresponding
steady and Floquet counterparts which are necessarily constant across the period. Since
the quasi-steady eigenvalues calculated from (3.23) are periodic and do not exhibit the
head and tail behaviour described in Luo and Wu [48], it is expected that the Floquet
eigenvalue should be identifiable with an average of the instantaneous eigenvalue across a
period of modulation. We can thus expect
Z Tϕ
1
αF ∼ α(τ )dτ (3.24)
Tϕ 0
tϕ
where µF is the Floquet eigenvalue calculated using (3.13) and Tϕ = 2πR
is the modulation
period. This quantity is also shown on figures 3.10 and 3.11 for comparison purposes where
it can be seen that the two quantities match well.
While providing validation for the Floquet analysis, this method provides an alter-
native approach for calculating the spatial growth rates in a substantially more compu-
tationally feasible manner. There is no requirement to work with the particularly large
matrices present in the Floquet theory and the computation is only limited by the number
of temporal locations required to sample the data. In effect, this method is of comparable
computational feasibility to the steady case and is advantageous over the Floquet analysis
for this reason. However, great care must be taken on using this approach in isolation, as
it is not rigorously guaranteed that equation (3.24) holds in all cases.
125
0.02
-0.02
-0.04
-0.06
-0.08
-0.1
0.2 0.25 0.3 0.35
Figure 3.10: Temporal evolution of α across a period of modulation in the Argand plane
for stationary disturbances with R = 500, n = 32, Uw = 0.2 and ϕ = 10. The temporally
periodic quantity µ is shown as (×) along with the Floquet (+), steady (o) and period-
averaged () eigenvalues.
0.35 0.02
0.3 -0.02
-0.04
0.25 -0.06
-0.08
0.2 -0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) αr (b) αi
Figure 3.11: Evolution of spatial eigenvalue α across a period of modulation for stationary
disturbances with R = 500, n = 32, Uw = 0.2 and ϕ = 10. The temporally periodic
quantities R(α) and I(α) are shown as (· · · ) along with the Floquet (- - -), steady (—)
and period-averaged (◦ —) eigenvalues.
126
3.3 Energy Analysis
This section will explore an energy balance approach which will enable the assessment of
the relative influences of the various energy transfer mechanisms affecting the stability
properties of the disturbances. Similar approaches have been used by several authors in
previous studies, with Cooper and Carpenter [19] intending to measure how the intro-
duction of wall compliance modified the energy balance and Cooper et al. [21], Garrett
et al. [36] investigating radially iso- and anisotropic surface roughness. The aim of the
forthcoming section will be to conduct a similar analysis, and following these previous
studies, we begin by deriving an integral energy equation to analyse the fundamental
physical mechanisms behind the apparent stabilising effects of periodic modulation on
the stability of the boundary layer.
The energy equation is obtained by multiplying the velocity transport equations by
ur , uθ and w respectively, and subsequently summing to give an equation for the kinetic
energy, K := 12 (u2r + u2θ + w2 ) given by
U u2θ

∂K ∂K V ∂K ∂K ∂U
+U + +W = −ur wU 0 − uθ wV 0 − w2 W 0 − u2r −
∂t ∂r r ∂θ ∂z ∂r r

∂ 1 ∂ ∂ ur p
− (ur p) + (uθ p) + (wp) +
∂r r ∂r ∂z r

∂ ∂uj 1
+ (uj σij ) − σij +O
∂xi ∂xi R2
where j ∈ {r, θ, z}, the 0 signifies ∂

∂z
, repeated indices represent summation and σij is an
anti-symmetric tensor representing the viscous stress terms, namely

1 ∂ui ∂uj
σij = − (3.25)
R ∂xj ∂xi
We have neglected the O R12 for consistency with the eigenvalue analysis, and can further

simplify the equations by averaging over a single azimuthal mode number n, thus removing
∂ ∂
the ∂θ
terms, while clearly by definition, σii = 0. Collecting some ∂r
terms on the left
127
hand side and integrating across the boundary layer gives
Z ∞
∂K ∂K ∂ ∂
+U + (ur p) − (uθ σ12 − wσ13 ) dz
0 ∂t ∂r} |∂r {z } |∂r
| {z {z }
A B C
Z ∞
= −ur wU 0 − uθ wV 0 − w2 W 0 dz (I)
0
Z ∞
∂uθ ∂w ∂ur ∂uθ
− σ12 + σ13 + σ31 + σ32 (II)
0 ∂r ∂r ∂z ∂z
Z ∞
ur p
− dz + (wp)w (III)
0 r
− (ur σ31 + uθ σ32 )w (IV)
Z ∞
∂K ∂U 2 U 2
− W + u + uθ dz (V)
0 ∂z ∂r r r
where the subscript w denotes a variable evaluated at the wall, z = 0 and ab = ab∗ + a∗ b
a quantity averaged across a perturbation period. The terms in the energy equation are
calculated using the eigenvalues and eigenfunctions from the dispersion relation (3.13)
given R, n and ω, while the pressure is obtained from the primary and secondary variables
via the relation
Z ∞
∂w iα 1 1 1
p= + Fw + iαξθ + ξθ dz + Hw (3.27)
z ∂t R R R R
as in Davies and Carpenter [26].
As described in Cooper et al. [21], each term has a physical interpretation in terms of
its contributions to the energy of the system. On the left hand side, the term (A) gives the
average kinetic energy convected by the radial mean flow, while (B) gives the work done
by the perturbation pressure. (C) denotes the work done by viscous stresses across some
internal boundary in the fluid and for the majority of results presented below is negligible
in comparison to the others. On the right hand side we have the Reynolds stress energy
production term (I), the viscous dissipation energy removal term (II), pressure work terms
(III) contributions from work done on the wall by viscous stresses (IV) and terms arising
from streamline curvature effects and the three-dimensionality of the mean flow (V). The
energy balance may be carried out for any eigenmode, and positive terms may be identified
with energy production while negative ones remove energy from the system. A mode is
amplified, giving αi < 0 in a spatial analysis, when energy production outweighs the
energy dissipation in the system. Calculation of each term in the energy equation allows
for the identification of where the effects of the modulation are the greatest.
128
The following section will provide an analysis of the steady case, and produce results
analogous to those in Cooper et al. [21] for validation purposes. Following this analysis,
we describe the process applied to the modulated scenario and detail the modifications
to the energy balance found therein.
3.3.1 Energy Balance for the Steady Rotating Disk Boundary

Layer
∂
In the steady case, we are able to further eliminate the ∂t
terms by averaging across a
single perturbation period and as in Cooper et al. [21], we normalise the energy equation
by the integrated mechanical energy flux which contributes the left hand side. Applying
the same nomenclature for each term as Cooper et al. [21], we write
−2αi = (P1 + P2 + P3 ) + |{z}

D + (P W1 + P W2 ) + (S1 + S2 + S3 ) + (G1 + G2 + G3 )
| {z } | {z } | {z } | {z }
(I) (II) (III) (IV) (V)
(3.28)
The energy balance may be carried out for any eigenvalue α, and figures 3.12 and 3.13
are given for validation against similar figures in Cooper et al. [21], showing variation
in the Type I and II energy production and viscous dissipation terms, P2 and −D, with
azimuthal mode number n for stationary disturbances with R = 400. In all cases tested,
the energy functions as calculated by our novel numerical methods described throughout
this thesis agree well with those in the literature of Cooper et al. [21] and Garrett et al.
[36].
Figure 3.14 shows that the energy balance in the system is essentially between the two
terms detailed in figures 3.12 and 3.13, with the rest of the terms in (3.28) contributing
negligibly to the overall balance. The terms that are not shown, namely S1 , S2 , S3 and
P W2 , are identically zero in this situation due to the boundary conditions. The following
section will analyse whether this balance remains the case in the modulated scenario, and
give an overview of the modifications to the energy terms when the modulation is added.
129
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 10 20 30 40 50 0 10 20 30 40 50
(a) P 2 (b) −D
Figure 3.12: Variation in the Type I energy production and viscous dissipation terms P2
and −D with azimuthal mode number n for stationary disturbances with R = 400. A
clear peak in energy production can be seen around n ≈ 28, consistent with the location
of maximum disturbance amplification described in Cooper et al. [21].
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(a) Energy production term P 2 (b) Viscous dissipation term −D
Figure 3.13: Variation in the Type II energy production and viscous dissipation terms P2
and −D with azimuthal mode number n for stationary disturbances with R = 400. A
clear peak in energy production can be seen around n ≈ 19, consistent with the location
of maximum disturbance amplification described in Cooper et al. [21].
130
0.25
0.2
0.15
0.1
0.05
-0.05
-0.1
Figure 3.14: Balance of terms in the energy analysis for stationary disturbances with
R = 500 and n = 32 in the steady rotating disk boundary layer. This figure shows
that the balance is essentially between the energy production term P 2 and the viscous
dissipation term −D, consistent with that found in Cooper et al. [21]. The other terms
contribute negligibly to the system.
131
3.3.2 Energy Balance for the Rotating Disk Boundary Layer
with a Periodically Modulated Rotation Rate
The modulated rotation rate comes into the system by way of a temporally periodic mod-
ification of the disk rotation rate, as described by equation (2.3) and in detail throughout
this thesis. Therefore, with the full time-dependent structure of the modulated rotation
∂
rate incorporated, we are no longer able to eliminate the ∂t
terms as we did for the steady
case, and thus may only write
∂K
− 2αi (t) = − + (P1 + P2 + P3 )(t) + D(t) + (P W1 + P W2 )(t)
∂t | {z } |{z} | {z }
(I) (II) (III)
+ (S1 + S2 + S3 )(t) + (G1 + G2 + G3 )(t) (3.29)

| {z } | {z }
(IV) (V)
where each variable, including the mechanical energy flux by which the equations are
normalised, is time-dependent and in particular periodic with the same period as the
modulation. We will deal with this in a similar fashion to that of section 3.2, where we
effectively freeze the flow at each time-step and march through the period as if t were
a parameter. This allows us to get an idea of how the energy changes across a period
of modulation, and provides insight into how the modulation affects the balance in the
system. Since the energy terms in (3.29) are periodic with the period of the modulation,
tϕ
which we will denote Tϕ = 2πR
here, we may average the variation across Tϕ to give an
appropriate quantity to be compared against the steady case.
Figure 3.15 shows the evolution of the temporal growth rate αi for R = 500, n = 32,
Uw = 0.2 and ϕ = 10 across two periods of modulation Tϕ , along with the quantity
−1 Tϕ
R
Tϕ 0
αi averaged across the modulation period. The stabilisation, as should be ex-
pected for this parameter set by the analyses of section 3.1, is clear by the reduction in
magnitude of −αi by the modulation. Figure 3.16 shows the evolution of the energy pro-
duction and viscous dissipation terms, P2 and −D, across a period of the flow modulation
RT R Tϕ
for the same parameter set, along with the quantities T1ϕ 0 ϕ P and −1Tϕ 0
D averaged
across the modulation period. Clearly, the averaged energy production term is smaller
than the steady equivalent and likewise the averaged dissipation term is larger than its
steady counterpart. This is similar to results given by Cooper and Carpenter [19], Cooper
et al. [21] and Garrett et al. [36] when studying wall compliance and surface roughness
and alludes to potentials for further analysis of the exact mechanisms for this via a care-
132
ful study of the contributions to each term. Figure 3.17 shows the full energy balance
for this parameter set, comparing the modulated scenario against the steady case, and
demonstrates that the main contributions to the system are still the energy production
and viscous dissipation terms when modulation is added. The other terms contribute
negligibly to the system, as in the steady case.
While this discussion does not give a robust explanation of the stabilising behaviour,
it may be utilised to pinpoint the important terms in the equations which contribute to
it. A more thorough examination of the energy production and viscous dissipation terms
could be used to shed light on the exact mechanism for the stabilisation, but such an
analysis is reserved for future work.
0.1
0.08
0.06
0.04
0.02
-0.02
0 0.5 1 1.5 2
Figure 3.15: Variation in the spatial growth rate −αi across a period for stationary
disturbances with with R = 500, n = 32, Uw = 0.2 and ϕ = 10 (—), compared with the
R Tϕ
steady case (· · · ) and quantity −1
Tϕ 0
αi (- - -) averaged across the modulation period Tϕ .
133
0.25 0.085
0.08
0.2
0.075
0.07
0.15
0.065
0.1 0.06
0.055
0.05 0.05
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(a) Energy production term P 2 (b) Viscous dissipation term −D
Figure 3.16: Variation in the energy production and viscous dissipation terms P2 and −D
across a period for stationary disturbances with with R = 500, n = 32, Uw = 0.2 and
RT R Tϕ
ϕ = 10 (—), compared with the steady case (- - -) and quantities T1ϕ 0 ϕ P and −1
Tϕ 0
D
(· · · ) averaged across the modulation period Tϕ .
0.25
0.2
0.15
0.1
0.05
-0.05
-0.1
Figure 3.17: Balance of terms in the energy analysis for stationary disturbances with
R = 500 and n = 32. The steady case is compared against the modulated scenario with
Uw = 0.2 and ϕ = 10. This figure shows the reduction in the energy production term
P 2 and expansion in the viscous dissipation term D produced by the modulation and
that the balance is still essentially between these terms, as for the steady case. The other
terms contribute negligibly to the system.
134
3.4 Monochromatic Direct Numerical Simulations
The following section describes a further method of analysis which we will use for vali-
dation against the Floquet theory and to provide an alternative approach to calculating
temporal growth rates. A similar idea was discussed in section 2.2.4 when considering the
steady rotating disk boundary layer. This will be the analysis of monochromatic distur-
bances, and the solution method is essentially identical to that described in section 2.2.4.
The primary difference is the fact that the base flow is now time-dependent, and therefore
updated at each time-step in the calculation. As in section 2.2.4, we apply a temporally
impulsive, but not radially localised, disturbance of the form
η(θ, t) = b(t)eiβRθ (3.30)
where b defines the time-dependent amplitude, given by
2 2
b(t) = (1 − e−σt )e−σt (3.31)
The first term of b acts as a continuous Heaviside function to scale the forcing up from
zero and σ is a parameter which dictates the timescale over which this takes place. We
decompose the perturbation quantities as
ˆ t)ei(αr+βRθ)
u = û(z, t)ei(αr+βRθ) , ξ = ξ(z, (3.32)
which gives the system

∂ξr ∂Nθ 2 1 ˆ 2 1 2iβ
+ iβNz − − (ξθ + iαw) = ∇ − 2 ξr − ξθ (3.33a)
∂t ∂z R R r R

∂ξθ ∂Nr ∂Nz 2 1 ˆ −2 1 2iβ
+ − + (ξr − iβw) = ∇ ξθ + ξr (3.33b)
∂t ∂z ∂r R R r2 R

ˆ w=
2 1 ∂ξ r ∂(rξ θ )
∇ − (3.33c)
r ∂θ ∂r
with
2
ˆ 2 = −α2 + iα − β 2 + ∂
∇ (3.34)
R ∂z 2
where the convective quantities N, given by (2.51), are time-dependent in line with the
time-dependent base flow UB . We solve the system using the identical predictor-corrector
scheme described in section 2.2.4.
135
3.4.1 Results and Comparison with Floquet Analysis
In order to obtain results which are comparable to the Floquet Analysis of section 3.1, we
must invoke a similar normal mode hypothesis and assume we can write a perturbation
variable f as
f (r, θ, z, τ ) = fˆ(z, τ )eµτ ei(αr+βRθ) (3.35)
and all exponential growth of fˆ(z, τ ) is incorporated into the Floquet exponent eµτ . Thus,
following similar ideas in Thomas et al. [76], we can calculate the dominant eigenvalue µ
directly from the simulation data by

1 p̂(z = 0, τ0 + T )
µ ≈ log (3.36)
T p̂(z = 0, τ0 )
for some starting value τ0 . With this operation, we are effectively calculating numerically
estimated asymptotic values, for large τ , and implicitly assuming that there is a most
unstable mode that dominates the behaviour. The growth rates calculated via this method
may then be identified with the temporal Floquet eigenvalues discussed in section 3.1.2,
thereby adding confidence to the reliability of the results presented in that section.
As for previous sections, the following results are shown for R = 500 and n = 32 so
that an instability would exist in the corresponding steady scenario. Figure 3.18 shows the
temporal development of ξθ (z = 0), excited by an impulsive forcing of the form (2.62) for
typical parameters α = 0.3, Uw = 0.2 and ϕ = 10, compared against the steady, Uw = 0,
case with the same parameters. The reduction in the growth rate for this parameter set
is such that we have used the natural logarithm of the absolute value of the disturbance
wavetrain to illustrate the stabilisation. A clear dampening effect can be seen in the
modulated case, as would be expected by the corresponding reduction in growth rate as
calculated by the Floquet analysis.
Figure 3.19 shows the temporal growth rates for a range of ϕ, together with a com-
parison of the results against those of the Floquet analysis. This diagram clearly shows
the temporal stabilisation achieved by adding the modulation, and its increasing effect
for φ ≈ 10. The agreement between the two distinct methods of growth rate calculation
is apparent, with the very slight discrepancy attributed to the inevitable errors that are
involved in estimating asymptotic values using results obtained for a finite simulation
length.
136
20
-20
-40
-60
-80
-100
0 0.5 1 1.5 2
Figure 3.18: Log plot of the temporal response to impulsive disturbance of ξθ (z = 0)

for α = 0.3, Uw = 0.2 and ϕ = 10 (- - -), together with the corresponding solutions for
the steady case (—). A clear reduction in the temporal growth rate is visible for the
modulated case.
10 -3
5
4.8
4.6
4.4
4.2
4
0 10 20 30 40
Figure 3.19: Temporal growth rates µr for α = 0.3, Uw = 0.2 and varying ϕ. Comparison
shown between monochromatic simulations (- - -) and Floquet analysis (—)
137
3.5 Direct Numerical Simulations of Stationary Forc-
ing with a Radial Dependence
Similarly to the analysis of the equivalent forcing method for the steady case presented
in section 2.2.5, we may incorporate a radial dependence into our simulations and include
the full r-derivatives of the perturbation variables. This analysis will provide independent
validation of the spatial Floquet analysis conducted in section 3.1 and lend confidence to
the methods presented there. Concentrating our exposition on stationary forcing as in
section 3.1, we look to excite disturbances in the flow by prescribing radially localised
motion around some location re of the form
ζ(r, θ, τ ) = a(r − re )b(τ )einθ (3.37)
for some azimuthal mode number n. We can pin down the radial variation of the forcing
by employing the parallel-flow approximation and choose
2
a(r) = e−λr (3.38)
to give a localised Gaussian pulse, where λ specifies the radial extent of the forcing. The
time-dependent amplitude may be chosen so as to give a time-periodic excitation, for
instance by setting
b(τ ) = h(τ )e−iω0 t (3.39)
here ω0 is a prescribed temporal frequency and h(τ ) = (1 − e−στ ) acts like a continuous
Heaviside function to scale the forcing up from zero. In the steady case, disturbances
which are stationary with respect to the surface can be considered by setting ω0 equal to
zero. There is a concern that since the modulation is time-dependent, this naı̈ve approach
could potentially introduce modes at different frequencies, thereby giving spurious results.
However, after thorough validation, the subsequent evolution transpired to be independent
of the form of stationary forcing employed, and rigorous checks were carried out to ensure
the frequency of forcing introduced was consistent with that of a stationary disturbance
analogous to the steady case.
3.5.1 Results and Comparison against Floquet Analysis
The general idea behind the solution method described in section 2.2.5 is preserved here,
with the modification introduced by the modulation incorporated by simply altering the
138
mean flow at each time-step. This is achieved in practise by solving equations (2.10) in
tandem with the time-marching procedure of the DNS and using the resulting mean flow
when calculating the convective terms (2.38). Computational efficiency can be enhanced
by utilising the periodicity of the mean flow and only computing the mean flow profiles
once over a modulation period, then replicating these values throughout the simulation.
For the majority of the results presented in this section, the primary variables may be
fully resolved using a Chebyshev expansion involving N = 64 polynomials and a radial
resolution of about ∆r ≈ 1. Each of the following simulation plots was produced using
N = 64 and a radial resolution of ∆r = 1.25. In all of the simulations, the computational
domain extends well beyond the limits suggested by the figures in order to ensure no
computational edge effects creep into affect the results. The temporal discretisation is
such that there are no less than 100 time-steps in one period of modulation. For small
ϕ . 15, it was necessary to use up to 1000 time-steps per modulation period to fully resolve
the disturbance quantities under consideration. No formal mathematical analysis of the
exact number required was undertaken during this work, and convergence was determined
empirically from comparisons between differing values of temporal discretisation.
The following figures show the radial evolution of a disturbance generated by stationary
forcing for R = 500 and n = 32. These parameter choices ensure that the steady disk
configuration is unstable, and similarly to previous sections, provide us with a base to
compare modulation effects against. The position of maximum forcing is located at r =
R = 500, again somewhat arbitrary since the parallel flow approximation treats all radial
locations as equivalent. Similarly to section 2.2.4, the simulations were conducted until a
steady growth rate was achieved and all transient behaviour had passed out of the domain
of interest so that T∞ can be identified with

i ∂A
α(T∞ ) ≈ lim − (3.40)
t→∞ A ∂r
Typically, T∞ is chosen to be between two and four periods of disk rotation. Further
discussions of the inflow and outflow criteria are given in Thomas [71] and briefly in
appendix B.3.1.
Figure 3.20 shows this evolution for Uw = 0.2 and ϕ = 10 on a log plot showing
the differences in the growth rates in each case. These parameters are chosen to give
an indication of the type of behaviour observed in the simulations and relatively small
amplitude oscillations. An initial inspection would suggest that the modulation heavily
139
12
10
0
550 600 650 700
Figure 3.20: Radial evolution of a disturbance generated by stationary forcing for R = 500,
n = 32, U w = 0.2 and ϕ = 10. The simulations were conducted until the growth rates
has settled to the dominant value where t ≥ T∞ .
dampens the radial growth rates of the disturbance, indicating a stabilising effect. This
effect may be quantified by calculating, retrospectively, the radial growth rates using the
normal mode approximation. As dictated by linear stability theory and the normal mode
hypothesis, after a sufficient amount of time and sufficiently far away from the point of
maximal forcing, we would expect the radial growth rates to be the same at each radial
position and be prescribed by a single wavenumber α which may be calculated directly
from the simulation data. If, since we are considering parallel flow in the radial direction,
we decompose our perturbation variables as
A = Â(z, t)ei(αr+nθ)
then we can identify the complex wavenumber displayed by the disturbance at a particular
radial position and instant of time with
1 ∂A
α≈ (3.41)
A ∂r
In this case, we take A to be ξθ (z = 0, t > T∞ ) for some T∞ such that the disturbance at
1 ∂A
fixed r has reached a steady state, provided A ∂r
is found not to vary too rapidly in space.
For the purposes of this investigation, we choose to study the azimuthal component of
140
vorticity at the wall, although other selections could equally be made without adverse
effects on the results. For the same parameters as in figure 3.20, figure 3.21 gives the
temporal evolution of the wavenumbers αs and αm corresponding to the unmodulated and
modulated scenarios respectively, and shows a steady growth rate being achieved after
around 25 periods of modulation, or 2.5 disk rotations under this non-dimensionalisation.
The imaginary parts of the wavenumbers, I(αs ) and I(αm ) are the quantities of interest
here as these give the expressions for the radial growth rates. Clearly we have
I(αs ) > I(αm ) (3.42)
which confirms our conclusions of stabilisation from figure 3.20.

0.34 0
0.33
0.32 -0.02
0.31
-0.04
0.3
0.29
-0.06
0.28
0.27 -0.08
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
(a) αr (b) αi
1 ∂A
Figure 3.21: Temporal evolution of real and imaginary parts of A ∂r
for R = 500, n = 32,
Uw = 0.2 and ϕ = 10, with comparison between steady (—) and modulated cases (- - -).
Stabilisation in terms of growth rate (αi ) reduction is clear.
This apparent stabilisation seems not to be an isolated case in terms of the parameters
shown, and figure 3.22 shows a plot of various modulation frequencies ϕ against the growth
rates I(αm ) for R = 500 and n = 32, with Uw ∈ {0.1, 0.2}. Validation of these results is
provided by calculation of α from the Floquet dispersion relation (3.13) while specifying
µ = 0. The results of these calculations are also shown in figure 3.22 which agree well
with the simulation data. The agreement between the two distinct methods of growth
rate calculation is apparent, with any discernible discrepancy attributed to the inevitable
errors that are involved in estimating asymptotic values using results obtained for a finite
simulation length. For values of ϕ . 6, the computational time required for the growth
rates to settle to a constant, namely T∞ , increased greatly. For very small values of ϕ it
was not possible to fully resolve the growth rates using the simulations to the required
141
accuracy in a reasonable time-frame. It is, as yet, unclear to the author whether there is a
physical process preventing the settling of these quantities for such values of ϕ or whether
the phenomenon is numerical. Such an analysis is beyond the scope of this project and is
thus reserved for future work.
0.02
0.015
0.01
0.005
0
5 10 15 20 25 30 35 40 45 50
Figure 3.22: Variation of I(αm ) with ϕ, for R = 500 and n = 32. Data from the Floquet
dispersion relation for Uw = 0.1 (—) and Uw = 0.2 (· · · ) is compared against growth rates
calculated from the simulations (◦). Values of ϕ ≤ 3 are intentionally omitted due to the
restrictive computational time required for t to pass T∞ in these cases.
3.6 Discussion
Throughout the preceding sections, we have detailed four distinct approaches to calcu-
lating the linear stability properties of a rotating disk boundary layer with a modulated
rotation rate. The approaches have involved an application of Floquet theory in the
temporal and spatial settings along with quasi-steady frozen profile analyses, monochro-
matic direct numerical simulations to calculate temporal growth rates and simulations
with a radial dependence capable of calculating spatial growth rates. Consistency in the
results between each method has been demonstrated in all cases, lending strength to the
confidence in our approaches. Each method constitutes an essentially alternative way of
presenting the same result, and have been illustrated to give validation to the others. Each
142
approach also presents both advantages and disadvantages in terms of computational cost
and mathematical rigour and the choice of solution method in any future study should
depend on the parameter set in question.
The main result of this chapter is to have demonstrated that the addition of a mod-
ulated rotation rate stabilises the boundary layer formed above a rotating disk to distur-
bances which have a fixed temporal frequency and spatial wavenumber. In particular, we
demonstrated that the spatial growth rates may be reduced when modulation is incorpo-
rated for disturbances which are stationary with respect to the motion of the disk surface.
This stabilisation has been shown for isolated parameter cases in the sense of the Reynolds
number and azimuthal mode number and across a range of parameters in the sense of
critical values for instability onset. Several results were presented which demonstrated
fairly strong stabilising behaviour for even small amplitude modulations. There is also an
indication of a somewhat optimal region of the non-dimensional frequency term ϕ where
the strongest stabilisation is present.
The underlying physical mechanisms for such behaviour are as yet unclear, although an
analysis of the energy balance of the disturbance was described in an attempt to illustrate
the finer details of the stabilising behaviour shown by the addition of the modulation.
Certain terms in the energy balance equations were identified as being dominant, and
comparisons were made against similar studies by Cooper et al. [21] and Garrett et al. [36]
during their investigations of surface roughness and by Cooper and Carpenter [19] during
their investigations of wall compliance. Similar behaviour to these alternative stabilising
mechanisms was identified for the periodically modulated boundary layer, indicating the
potential for parallels between each configuration and the requirement for a substantial
future study to ascertain the exact nature of the stabilisation.
143
Chapter 4
Global Behaviour and Absolute

Instabilities in Periodically
Modulated Rotating Disk Boundary
Layers
Chapters 1, 2 and 3 have thus far only considered the analysis of locally defined dispersion
relations and results from radially simplified direct numerical simulations. These give only
limited insight into the true behaviour exhibited by the rotating disk boundary layer and
in reality, rotating flows can exhibit rich, complex disturbance structures including abso-
lute instability and turbulence. Heretofore, we have, in all cases, described disturbance
properties from the standpoint of either a temporal analysis or a spatial one. Genuine
behaviour, however, develops both spatially and temporally and some sense may be made
of this by considering the spatio-temporal analysis that was briefly discussed in section
2.2.3. This may be studied by an analysis of the flow response to a radially localised
impulsive forcing, as will be performed in this chapter.
During our discussions in section 2.2.3, we remarked on an analysis conducted by
Lingwood [43] which appeared to show the existence of a local absolute instability in the
rotating disk boundary layer. Lingwood’s [43] analysis utilises the so-called parallel flow
approximation which amounts to neglecting the radial dependence of the basic state, with
which it is useful to refamiliarise ourselves with at this stage. Section 2.1.5 first introduced
144
the concept of this homogenising technique, where we defined the base velocity field as

r r 1
U(r, z) = F (z), G(z), H(z) (4.1)
R R R
with F, G and H being the solutions to the Kármán ODEs given by equation 2.12. The
parallel flow approximation amounts in this case, as discussed in more detail in section
2.1.5, to the replacement of r by R in the equations, thereby allowing the radial evolution
of the perturbation to be separable into normal mode form. Using this approximation,
Lingwood [43] demonstrates the existence of the absolute instability using Brigg’s cri-
terion. While we will not reproduce the analytical methods used in this study, we will
present numerical simulation results for the steady rotating disk configuration which are
consistent with Lingwood’s [43] and were first discussed in Davies and Carpenter [26].
These results, outlined in the first part of this chapter, utilise the same parallel flow ap-
proximation as in Lingwood [43] and thereby allow for the homogenisation of the radial
direction. As background to the study we first present the steady case before discussing
the modifications to the behaviour introduced by the periodic modulation.
The removal of the parallel flow approximation and thereby the radial homogenisation
was first studied by Davies and Carpenter [27] who discovered that the locally defined
absolute instability described in Lingwood [43] does not trigger a linearly unstable global
mode and convective behaviour dominates at all radii and azimuthal mode numbers, at
least within the ranges discussed in Lingwood [43]. A further explanation of the subtleties
involved in this approximation and the current state of research in this area is provided
later in section 4.3 where simulation results are given to illustrate the behaviour. We
discuss the steady inhomogeneous configuration as an introduction to the presentation of
the results pertaining to the modulated rotation rate.
4.1 Radially Homogeneous Steady Case

For completeness, and in an attempt to provide a base from which to elaborate, we detail
several important results regarding the stability of the steady rotating disk boundary
layer, and will use these to illustrate the local absolute instability which is present in this
configuration. Similarly to section 2.2.5, we excite the impulsive disturbance by means of
prescribed wall motion which takes the form
η(r, θ, t) = a(r − re )b(t)einθ (4.2)
145
where a describes the radial variation centred at the location re and the function b defines
the time-dependent amplitude. Since the simulations of this section are radially homoge-
neous, the forcing location is somewhat arbitrary since it is the Reynolds number which
dictates the behaviour. The notation re is retained however, to allow for direct compar-
isons between the homogeneous and inhomogeneous simulations conducted in section 4.3.
The radial distributions of the impulsive disturbance (4.2) take the form
2
a(r) = e−λr
and the impulsive excitation for b takes the form
2 2
b(t) = (1 − e−σt )e−σt
where σ is a parameter chosen such that √1 gives the timescale of the forcing up from
σ
zero amplitude at t = 0.
Davies and Carpenter [26] state that the variation of the primary variables may be
fully resolved using a Chebyshev expansion involving N = 48 polynomials and a radial
resolution of about ∆r ≈ 1. Each of the following simulation plots was produced using
N = 48, a radial resolution of ∆r = 1.25 and a time discretisation of ∆t = 0.625. In all of
the simulations, the computational domain extends well beyond the limits suggested by
the figures in order to ensure no computational edge effects creep into affect the results.
4.1.1 Convective Instability
Figure 4.1 shows the temporal development of the azimuthal component of the vorticity
ξθ (z = 0, t) for a disturbance with azimuthal mode number n = 32 triggered by an
impulse with Reynolds number R = 350. It is clear that the disturbance decays rapidly
at r = re and subsequently, for r > re , there is an initial period of inactivity before the
disturbance takes hold, followed by a decay as would be expected for this convectively
unstable configuration.
The radially localised impulsive forcing is located at r = R = 500. The radial position
of the forcing is arbitrary in the homogenised flow configuration as it it the Reynolds
number which dictates the flow behaviour, and the figure shows the temporal development
at four radial locations, namely r = R, r = R + 25, r = R + 50 and r = R + 100. The
temporal axis is scaled by the disk rotation period 2πR.
146
Figure 4.2 shows the spatio-temporal contour plots of |ξθ (z = 0, t)| for a convectively
unstable configuration with R = 350 and n = 32. It is possible to identify both the leading
and trailing edges of a disturbance wavepacket, from which it is clearly determined that
both ends of the wavepacket propagate outwards. This is characteristic of convective
instability, as would be expected for this parameter choice.
0.4 0.04
0.2 0.02
0 0
-0.2 -0.02
-0.4 -0.04
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
(a) r = R = 350 (b) r = 375

0.1 0.3
0.2
0.05
0.1
0 0
-0.1
-0.05
-0.2
-0.1 -0.3
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
(c) r = 400 (d) r = 450
Figure 4.1: Temporal evolution of ξθ (z = 0, t) (—) for a disturbance with n = 32 and

an impulse excited with Reynolds number R = 350 in the steady configuration. The
wavepacket envelopes |ξθ (z = 0, t)| are also shown (- - -). The temporal axis is scaled by
the disk rotation period 2πR.
147
0
-2
1.5
-4
1
-6
0.5
-8
-10
300 400 500 600 700 800
Figure 4.2: Spatio-temporal development of |ξθ (z = 0, t)| for an impulsively excited dis-
turbance with Reynolds number R = 350 and azimuthal mode number n = 32. The
temporal axis is scaled by the disk rotation period 2πR. Contours are drawn using a
logarithmic scale, with levels separated by factors of two.
148
4.1.2 Absolute Instability
The critical values for the onset of absolute instability as stated in Lingwood [44] are
given by (R, n) ≈ (507, 68). Figure 4.3 shows the temporal evolution of a wavepacket in
response to radially localised impulsive forcing located at r = R = 525 with prescribed
azimuthal wavenumber of n = 68. The figure shows the temporal development at four
radial locations, namely r = R, r = R + 25, r = R + 50 and r = R + 100 and the
temporal axis is scaled by the disk rotation period 2πR. Since R > 507 and n = 68, this
is within the absolutely unstable region according to Lingwood [43], as illustrated by the
exponential growth in time of the wavepacket seen at all radial locations at all times.
Figure 4.4 show the spatio-temporal development of the disturbance and from the
contours we can see that over the time interval considered, the trailing edge of the distur-
bance propagates along the inward radial direction, confirming the existence of the local
absolute instability.
0.5 1
0.5
0 0
-0.5
-0.5 -1
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
(a) r = R = 525 (b) r = 550

15
2000
10
1000
5
0 0
-5
-1000
-10
-15 -2000
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
(c) r = 575 (d) r = 625
Figure 4.3: Temporal evolution of ξθ (z = 0, t) (—) for a disturbance with n = 68 and

an impulse excited with Reynolds number R = 525 in the steady configuration. The
wavepacket envelopes |ξθ (z = 0, t)| are also shown (- - -). The temporal axis is scaled by
the disk rotation period 2πR.
149
1.2
10
1
5
0.8
0.6 0
0.4
-5
0.2
-10
500 600 700 800 900
Figure 4.4: Spatio-temporal development of |ξθ (z = 0, t)| for an impulsively excited dis-
turbance with Reynolds number R = 350 and azimuthal mode number n = 32. The
temporal axis is scaled by the disk rotation period 2πR. Contours are drawn using a
logarithmic scale, with levels separated by factors of two.
150
4.2 Radially Homogeneous Periodically Modulated
Case
The goal of this following section will be to provide a demonstration of the response of
the periodically modulated rotating disk boundary layer to radially localised impulsive
forcing, and contrast this behaviour to the steady case. We will, for the time being,
employ the same parallel flow approximation discussed in section 4.1 and homogenise the
radial direction by replacing r by R in the governing equations for the base flow and
perturbations. The impulsive forcing in all cases is taken to be equivalent in amplitude
to that of the steady case, the form of which is given by equation (4.2). We begin our
presentation with a typical configuration which is convectively unstable in the steady
case, before exploring a parameter set which would exhibit absolutely unstable behaviour
without the added modulation.
Similarly to Davies and Carpenter [26], the variation of the primary variables in the
modulated case may be fully resolved using a Chebyshev expansion involving N = 48
polynomials and a radial resolution of about ∆r ≈ 1. Each of the following simulation
plots was produced using N = 48, a radial resolution of ∆r = 1.25 and a time discreti-
sation of ∆t = 0.625. In all of the simulations, the computational domain extends well
beyond the limits suggested by the figures in order to ensure no computational edge effects
creep into affect the results.
4.2.1 Convective Instability
Figure 4.5 shows the temporal development of the azimuthal component of the vorticity
|ξθ (z = 0, t)| for a disturbance with azimuthal mode number n = 32 triggered by an
impulse with Reynolds number R = 500 at varying radial locations for Uw = 0.2 and
various values of ϕ.
The impulsive forcing is taken to be equivalent in amplitude across all parameter
choices, with the position of maximum forcing located at r = R = 500. The radial
position of the forcing is arbitrary in the homogenised flow configuration as it it the
Reynolds number which dictates the flow behaviour, and the figure shows the temporal
development at four radial locations, namely r = R, r = R + 25, r = R + 50 and
r = R + 100. The temporal axis is scaled by the disk rotation period 2πR.
151
Clearly, sufficiently far from the centre of excitation, the addition of a modulated
rotation rate has a dampening effect on the disturbance evolution, leading to significant
stabilisation in terms of suppression of the wavepacket maximum for a certain parameter
range. While this isolated parameter set has been chosen for illustrative purposes, similar
behaviour was observed in the preliminary work to this study across a wide range of
Reynolds and azimuthal mode numbers. Intriguingly, we see a stronger reduction in the
wavepacket maximum for ϕ = 8 than for ϕ = 12, consistent with the idea of a somewhat
optimal stabilisation parameter range first remarked upon in section 3.1.3, figure 3.4.
0.5 0.15
0.1
0.05
0 0
-0.05
-0.1
-0.5 -0.15
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(a) r = R = 500 (b) r = 525

0.5 10
0 0
-5
-0.5 -10
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(c) r = 550 (d) r = 600
Figure 4.5: Temporal evolution of wavepackets |ξθ (z = 0, t)|. The disturbance is excited
with an azimuthal mode number n = 32 and Reynolds number R = 500 for Uw = 0.2
and ϕ = 4 (- - -), ϕ = 8 (◦−), ϕ = 12 (—). The steady case is shown as (· · · ) and the
4.2.2 Absolute Instability
To demonstrate the modifications to the behaviour of the disturbance evolution subject

to an impulsive forcing in an absolutely unstable configuration, we consider R = 525
and n = 68. Figure 4.6 shows the temporal development of the azimuthal component
152
of the vorticity |ξθ (z = 0, t)| for a disturbance excited with these parameters along with
Uw = 0.2 and various ϕ.
Similarly to the convective case, the impulsive forcing is taken to be equivalent in
amplitude across all parameter choices, with the position of maximum forcing located at
r = R = 525. The figure shows the temporal development at four radial locations, namely
r = R, r = R + 25, r = R + 50 and r = R + 100 and the temporal axis is scaled by the
disk rotation period 2πR.
Again, clear stabilisation may be seen for the modulated simulations, in the sense
that the wavepacket maximum is smaller across the range of r. Of particular interest is
the apparent suppression of the characteristic behaviour indicating absolute instability.
Figure 4.7 shows the spatio-temporal development of the disturbance in a similar manner
to figure 4.4, which no longer exhibits contours indicative of absolute instability, implying
that this is also stabilised by the modulation. Furthermore, as may be seen by the
range of values shown on the colourbar accompanying figures 4.7 and 4.4, the wavepacket
maximum is a factor 24 smaller in the modulated case, indicating a particularly strong
stabilisation for this parameter set.
Given the apparent strength of this stabilisation in this case, it is pertinent to present
similar indicative figures (4.8-4.9) for very low values of modulation parameter Uw = 0.01
and Uw = 0.05. It appears, from this parameter set at least, that the stabilisation of the
global behaviour is particularly strong and while a thorough parametric investigation of
has not been conducted through the process of this work, such a study would make for a
particularly interesting avenue for future investigation.
153
3
0.5
0 0
-1
-2
-0.5 -3
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(a) r = R = 525 (b) r = 550

40 10 4
1
20 0.5
0 0
-20 -0.5
-40 -1
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(c) r = 575 (d) r = 625
with an azimuthal mode number n = 68 and Reynolds number R = 525 for Uw = 0.2 and
ϕ = 4 (- - -), ϕ = 8 (◦−), ϕ = 12 (—). The steady case is shown as (—) and the temporal
axis is scaled by the disk rotation period 2πR.
154
8
1.2 6
1 4
2
0.8
0
0.6 -2
-4
0.4
-6
0.2 -8
-10
500 600 700 800 900
Figure 4.7: Spatio-temporal development of |ξθ (z = 0, t)| for an impulsively excited distur-
bance with Reynolds number R = 525 and azimuthal mode number n = 68, for Uw = 0.2
and ϕ = 10. The temporal axis is scaled by the disk rotation period 2πR. Contours are
drawn using a logarithmic scale, with levels separated by factors of two.
155
3
0.5
0 0
-1
-2
-0.5 -3
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(a) r = R = 525 (b) r = 550

10 4
40 1
20 0.5
0 0
-20 -0.5
-40 -1
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(c) r = 575 (d) r = 625
and ϕ = 4 (- - -), ϕ = 8 (◦−), ϕ = 12 (—). The steady case is shown as (—) and the
156
3
0.5
0 0
-1
-2
-0.5 -3
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(a) r = R = 525 (b) r = 550

40 10 4
1
20 0.5
0 0
-20 -0.5
-40 -1
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(c) r = 575 (d) r = 625
157
4.3 Global Stability of the Periodically Modulated
Rotating Disk Boundary Layer
In the local stability analyses of chapters 2-3, we applied the so-called parallel flow ap-
proximation and simplified the mean flow by taking it to be homogeneous along the radial
direction. Additionally, we have remarked several times throughout this thesis that this
process is somehow unphysical and should be treated with caution. If we allow the mean
flow to vary with the radius, then the separability of the solution into normal mode form
is destroyed for the radial direction and the r-derivatives must be retained to specify the
disturbance structure. For the radially inhomogeneous flow, Davies and Carpenter [27]
discovered that the absolute instability predicted by the local analysis does not in fact give
rise to any global linear instability. Instead, convective behaviour dominates the flow at
all Reynolds numbers, even those taken well within the region of absolute instability. This
was the first numerical global study of the flow and was contrary to what many other
authors believed at the time. Experimental studies since undertaken by Othman and
Corke [56] have confirmed this conclusion, and provided the disturbance amplitudes can
be forced to remain small enough to avoid non-linearity, convective transient behaviour
wins out at all Reynolds numbers. An exposition of the global behaviour is given in
Davies and Carpenter [27], later explained by Davies et al. [28] and Thomas and Davies
[72], and has been referred to in many studies since, for example during the investiga-
tion of secondary instabilities by Pier [57] and the influence of an axial magnetic field by
Thomas and Davies [74]. Healey [40] also investigated the effect of a finite disk radius
using the Ginzburg-Landau equation. The most recent studies have been undertaken by
Appelquist et al. [4, 5, 6] and have investigated non-linear behaviour and the influence of
a finite radius disk on the global stability by means of full direct numerical simulations.
A fairly recent review by Lingwood and Alfredsson [46] gives a detailed overview of the
up-to-date state of the field which it would be superfluous to repeat.
To illustrate the concept of global stability, it is useful to consider an example. Figure
4.10 shows the response to an impulsive disturbance centred at r = 520 with azimuthal
mode number n = 68 in both the radially homogeneous and inhomogeneous cases.
These parameters were chosen so as to be inside the absolutely unstable region de-
scribed by Lingwood [43]. It can be seen that while the initial growth is similar in both
158
15 20
10
10
5
0 0
-5
-10
-10
-15 -20
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(a) Homogeneous (b) Inhomogeneous
Figure 4.10: Comparison between temporal evolution of ξθ (z = 0, t) (—) for a distur-

bance with azimuthal mode number n = 68 and an impulse excited by radially localised
impulsive forcing centred at re = 520 in both the homogeneous and inhomogeneous con-
figurations. The wavepacket envelopes |ξθ (z = 0, t)| are also shown (- - -). In the inhomo-
geneous configuration, the temporal axis is scaled by the disk rotation period T = 2πri
where ri is the interior of the computational boundary, and by 2πR in the homogeneous
case.
the homogeneous and inhomogeneous cases, after around t/T ≈ 1.3, the magnitudes of
this flow variable in the inhomogeneous case start to decrease. This behaviour is not what
had been anticipated on the basis of the absolute instability predicted from the local the-
ory, where it had been expected that for these parameters, the magnitude should increase
without bound at all fixed radial locations. It was thus inferred by Thomas and Davies
[72] that once the radial inhomogeneity of the base flow is incorporated, the flow is linearly
globally stable. This phenomenon was also shown to be reproducible using a linearised
Ginzburg-Landau equation with spatially varying coefficients, as detailed in Thomas [71],
Davies et al. [28] and Thomas and Davies [72]. The expectation that the non-parallel cor-
rections should not have had any effect on the absolute instability properties arose from
Lingwood’s [43] inviscid analysis which demonstrated that the absolute instability per-
sisted in an appropriate inviscid limit. This suggested that the absolute instability should
be robust even when it is embedded in the non-parallel flow, since non-parallel corrections
ought to eventually become insignificant at large enough radii. However, as explained in
Davies et al. [28], there is a subtlety to the localised viscous stability analysis, as well as
to its inviscid counterpart, in the time scalings used for the non-dimensionalisation of the
159
time variation and the corresponding disturbance frequencies. In the local analysis, where
the absolute instability is present, the time scale is constructed from the ratio between
the constant boundary-layer thickness of the mean flow and the circumferential speed of
the disk surface, which increases linearly with the radius. But there is also a global time
scale, independent of the radius, that is defined by simply taking the inverse of the con-
stant angular rotation rate of the disk. Thus if ωl and ωg represent the local and global
non-dimensional temporal frequencies corresponding to the same physical frequency, then
we have the relation
ωg = ωl R
which corresponds to ωl = const. in the inviscid analysis. We can thus see that the
fixed local frequency characterising the absolute instability actually corresponds, in the
inhomogeneous flow, to a collection of radially varying globally non-dimensionalised fre-
quencies. It was this subtlety that, as explained by Davies et al. [28], was shown to cause
the inhomogeneity to be globally stabilising, contrary to intuition.
In the homogeneous simulations of section 4.1, the disk rotation period T was set to
be 2πR where R specifies the Reynolds number of the forcing location which is locally
equivalent to the radius. Similarly, the azimuthal wavenumber β was set to be β = 68/R.
In the forthcoming results for the inhomogeneous simulations, since the disturbance is
evolving in a genuine radially inhomogeneous flow, the disk rotation period T is set to
be 2πri where ri is the radius of the inner computational boundary and the azimuthal
wavenumber β is set to be β = 68/ri . This allows for direct comparisons between the ho-
mogeneous and inhomogeneous configurations, as has been done in Davies and Carpenter
[27], Davies et al. [28], Thomas and Davies [73] and Thomas [71].
4.3.1 Steady Case
Similarly to section 4.1, we excite the impulsive disturbance by means of prescribed wall
motion which takes the form of equation (4.2). The important difference with this radially
inhomogeneous configuration is that the forcing location is no longer arbitrary, and the
radial extent of the computational domain is not approximated by the replacement of r
by R in this case.
Figure 4.11 shows the temporal development of the azimuthal component of the vor-
ticity ξθ (z = 0, t) for a disturbance with azimuthal mode number n = 68 triggered by
160
an impulse at re = 510. According to the local analysis of Lingwood [43], this param-
eter choice should be within the absolutely unstable range and the temporal evolutions
should, therefore, grow at all spatial locations at all times. This is not the case, as first
demonstrated by Davies and Carpenter [27], and the disturbances show a clear decay in
finite time after an initial period of growth. This is indicative of a convectively unstable
disturbance, consistent with the results presented in Davies and Carpenter [27].
As illustration of the behaviour exhibited by this steady configuration, figure 4.2 shows
the spatio-temporal contour plots of |ξθ (z = 0, t)| for the same parameter set of n = 68 and
re = 510. The temporal axis is scaled by the disk rotation period T , set to be 2πri where ri
is the radius of the inner computational boundary. Similarly, the azimuthal wavenumber
β is fixed as β = 68/ri . It is possible to identify the turn towards convectively unstable
behaviour around t/T ≈ 1.3 of the disturbance wavepacket, from which it is clearly
determined that both ends of the wavepacket propagate outwards. This is characteristic of
convective instability, not absolute instability as would have been predicted by Lingwood’s
[43] local analysis.
161
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(a) r = rf = 510 (b) r = 550

4
10
20 2
10 1
0 0
-10 -1
-20 -2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(c) r = 575 (d) r = 600
Figure 4.11: Temporal evolution of ξθ (z = 0, t) (—) for a disturbance with azimuthal mode
number n = 68 and an impulse excited by radially localised impulsive forcing centred at
re = 510 in the steady configuration. The wavepacket envelopes |ξθ (z = 0, t)| are also
shown (- - -). The temporal axis is scaled by the disk rotation period T = 2πri where ri
is the interior of the computational boundary.
162
50
1.5
40
30
1
20
10
0.5
0 -10
500 600 700 800 900
Figure 4.12: Spatio-temporal development of |ξθ (z = 0, t)| for azimuthal mode number
n = 68 and radially localised impulsive forcing centred at re = 510. The temporal axis is
scaled by the disk rotation period T = 2πri where ri is the interior of the computational
boundary. Contours are drawn using a logarithmic scale, with levels separated by factors
t
of two. The trailing edge can be seen to turn around T
≈ 1.3, indicating convective
behaviour as discussed in Davies and Carpenter [27].
163
4.3.2 Periodically Modulated Case
Figure 4.5 shows the temporal development in the periodically modulated boundary layer
of the azimuthal component of the vorticity |ξθ (z = 0, t)| for a disturbance with azimuthal
mode number n = 68 triggered by an impulsive forcing centred at re = 510 at varying
radial locations for Uw = 0.2 and various values of ϕ.
The impulsive forcing is taken to be equivalent in amplitude across all parameter
choices, with the position of maximum forcing located at r = re = 510. The radial
position of the forcing is no longer arbitrary in the inhomogeneous flow configuration.
The figure shows the temporal development at four radial locations, namely r = re ,
r = re + 25, r = re + 50 and r = re + 100. The temporal axis is scaled by the disk rotation
period 2πri where ri is the interior of the computational boundary.
Clearly, sufficiently far from the centre of excitation, the addition of a modulated
rotation rate has a dampening effect on the disturbance evolution, leading to significant
stabilisation in terms of suppression of the wavepacket maximum for a certain parameter
range. While this isolated parameter set has been chosen for illustrative purposes, similar
behaviour was observed in the preliminary work to this study across a wide range of
Reynolds and azimuthal mode numbers. Intriguingly, we see a stronger reduction in the
wavepacket maximum for ϕ = 8 than for ϕ = 12, consistent with the idea of a somewhat
optimal stabilisation parameter range first remarked upon in figure 3.4, section 3.1.3.
Additionally, the radially outward movement of the trailing edge is clearly much
stronger for the modulated case than for the unmodulated scenario, indicating a strong
stabilisation to impulsive forcing for the parameter choices considered in this study.
Similarly to the radially homogeneous case, given the apparent strength of this stabili-
sation, it is pertinent to present similar indicative figures (4.15-4.16) for very low values of
modulation parameter Uw = 0.01 and Uw = 0.05. It appears, from this parameter set at
least, that the stabilisation of the global behaviour is particularly strong and while a thor-
ough parametric investigation of has not been conducted through the process of this work,
such a study would make for a particularly interesting avenue for future investigation.
164
5
1.5
0.5
0 0
-0.5
-1
-1.5 -5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(a) r = rf = 510 (b) r = 550

200 5000
100
0 0
-100
-200 -5000
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(c) r = 575 (d) r = 600
with an azimuthal mode number n = 68 and impulse centred at re = 510 for Uw = 0.2
temporal axis is scaled by the disk rotation period T = 2πri where ri is the interior of the
computational boundary.
165
50
1.5
40
30
1
20
10
0.5
0 -10
500 600 700 800 900
Figure 4.14: Spatio-temporal development of |ξθ (z = 0, t)| for azimuthal mode number
n = 68 and radially localised impulsive forcing centred at re = 510 with Uw = 0.2 and
ϕ = 10. The temporal axis is scaled by the disk rotation period T = 2πri where ri is the
interior of the computational boundary. Contours are drawn using a logarithmic scale,
with levels separated by factors of two. The radially outward propagation of trailing edge
is much clearer than in the steady case, indicating stabilisation.
166
6
1
4
0.5
2
0 0
-2
-0.5
-4
-1 -6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(a) r = rf = 510 (b) r = 550

200 6000
4000
100
2000
0 0
-2000
-100
-4000
-200 -6000
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(c) r = 575 (d) r = 600
167
5
1
0.5
0 0
-0.5
-1 -5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(a) r = rf = 510 (b) r = 550

200 5000
100
0 0
-100
-200 -5000
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
(c) r = 575 (d) r = 600
168
4.4 Discussion
This chapter has detailed the study of radially localised impulsive forcing in both the
steady and periodically modulated rotating disk boundary layers. Split into two sections,
we examined the behaviour of disturbances in both radially homogenised and inhomoge-
neous configurations and presented results showing the differences in behaviour between
the steady and modulated cases. Similarly to that shown in chapter 3, stabilising be-
haviour was demonstrated by the simulations across a range of modulation parameters and
Reynolds numbers. For a certain absolutely unstable configuration in the homogenised
flow, it was also shown that the modulation suppresses the absolute instability and con-
verts the behaviour to convective for even very small amplitude modulations. Similarly,
in the inhomogeneous configuration, we demonstrated that small amplitude modulations
have a significant dampening effect on the wavepacket evolution across a range of modu-
lation frequencies.
Since this spatio-temporal analysis is, in general, more indicative of physical, experi-
mental behaviour than the fixed-frequency forcing of chapter 3, it may provide confidence
that the stabilisation shown by the addition of the modulated rotation rate is more robust
than being isolated to stationary or fixed wavenumber disturbances. Impulsive excitation
has been used throughout this chapter, thereby creating disturbances that take the form of
wavepackets, initially containing a wide range of frequencies, and the form of the pertur-
bations in our simulations is not specified in any way. Just like in a physical experiment,
the perturbations are initially excited by a radially localised displacement of the disk
surface with the subsequent evolution of the disturbance governed purely by the pertur-
bation equations. Therefore, we may have confidence that any results presented here
should be ratifiable by experimental data. However, it should be noted that these global
studies were not the main focus of the work carried out during this thesis, and thorough
parametric testing and experimental confirmation is reserved for future studies.
169
Chapter 5
Future Directions
The following sections provide an insight into the author’s intentions for future directions,
and will introduce an interesting inter-disciplinary application to the work conducted
throughout this thesis. We begin the discussion with purely torsional oscillations and a
calculation of the base state in this configuration, before progressing through a further
exploration of the Floquet stability theory to some fundamental techniques in electro-
chemistry. We conclude the discussion of future directions with alternative stabilising
methods such as designed surface roughness and some parallels to our work contained
therein.
5.1 The Torsionally Oscillating Disk Boundary Layer

The previous chapters have dealt primarily with a disk of infinite extent rotating with
an unsteady rotation rate and we have discussed the stabilising effects of a periodically
modulated rotation rate on the local stability of the boundary layer. We now turn our
attention to a slightly different configuration, and consider the case where the disk per-
forms torsional oscillations around the azimuth, which is an important configuration in
several electrochemical engineering processes such as chemical deposition, hydrodynamic
voltammetry and electroplating, see Cummings et al. [22, 23], Ahn et al. [1, 2]. To the
best of my knowledge, there has been very little mathematical study of this flow, and
only a few analytical studies exist which calculate the basic state in various configura-
tions. Rosenblat [63] was the first to study this problem, and Benney [9] subsequently
improved on the results. Since then, there have been several studies on variants of the
170
configuration, including Secomb and Rosenblat [67], Reddy et al. [60] and Srivastava [69]
who study various non-Newtonian effects and dual-disk configurations. It seems, however,
that there has been no formal exploration of the hydrodynamic stability of such a flow,
and we thus proceed to present an introductory examination.
We begin by considering a disk of infinite extent, torsionally oscillating in the plane
z = 0 about the axis r = 0. Let the oscillations have angular displacement and frequency
φ∗ so that the disk moves with angular velocity φ∗ cos(φ∗ t∗ ). Thus, the base flow UB =
(U ∗ , V ∗ , W ∗ ) obeys the usual Navier-Stokes equations in a cylindrical polar coordinate
system with boundary conditions
U ∗ (0) = W ∗ (0) = 0, V ∗ (0) = r∗ φ∗ cos(φ∗ t∗ )
U ∗ → 0, V ∗ → 0 as z ∗ → ∞
When discussing the periodically modulated rotation rate in chapter 2, we noted two
length scales on which the flow evolves, namely
r ∗ r ∗
∗ ν ∗ ν
δk = ∗
, δs = (5.1)
Ω0 φ∗
where Ω∗0 was the underlying steady rotation rate and φ∗ the frequency of modulation.
In that case, we selected the Kármán scale δk∗ for the non-dimensionalisation. In this
scenario, we do not have a mean rotation rate and so must choose the Stokes length scale
q ∗
δs∗ = φν ∗ which leads to the natural selection for the temporal scale of
τs = φ∗ t∗ (5.2)
Following Rosenblat [63], we can take a similar solution to von Kármán [80] for the steady
case, and scale the velocities as
U ∗ (r, z, t) V ∗ (r, z, t) W ∗ (r, z, t)

F (z, t) = , G(z, t) = , H(z, t) = (5.3)
φ∗ r∗ φ∗ r∗ φ∗ δ ∗
giving the system of equations for the base flow as
∂ 2F

∂F 2 2 ∂F
= G −F −H + (5.4a)
∂τ ∂z ∂z 2
∂ 2G

∂G ∂G
= −2F G − H + (5.4b)
∂τ ∂z ∂z 2
∂H
= −2F (5.4c)
∂z
171
F (0, τ ) = H(0, τ ) = 0, G(0, τ ) = cos(τ )
F → 0, G → 0 as z → ∞
This system is solved in an identical way to that of the periodically modulated disk in
section 2.1.1, so the reader is referred to the explanation presented there for further details.
Figure 5.1 shows the typical evolution of the velocity flow components over one period of
oscillation for = 0.25. In addition, figures 5.2, 5.3 and 5.4 show the time history of the
base flow profiles at selected locations z = 0 (-), z = 0.1 (- - -) and z = 0.25 (· · · ) over
one period of oscillation.
0.5
-0.5
-1
0 1 2 3 4 5 6 7 8
Figure 5.1: Torsional oscillation velocity profiles F (- - - ), G (—), −H (· · · ) with = 0.25.
172
0.025
0.02
0.015
0.01
0.005
-0.005
0 0.2 0.4 0.6 0.8 1
Figure 5.2: Time history of radial velocity at selected locations z = 0 (-), z = 0.1 (- - -)
and z = 0.25 (· · · ) over one period of oscillation with = 0.25.
1.5
0.5
-0.5
-1
0 0.2 0.4 0.6 0.8 1
Figure 5.3: Time history of azimuthal velocity at selected locations z = 0 (-), z = 0.1 (-
- -) and z = 0.25 (· · · ) over one period of oscillation with = 0.25.
173
10 -3
0
-2
-4
-6
-8
0 0.2 0.4 0.6 0.8 1
Figure 5.4: Time history of vertical velocity at selected locations z = 0 (-), z = 0.1 (- - -)
and z = 0.25 (· · · ) over one period of oscillation with = 0.25.
174
5.1.1 Local Linear Stability of the Torsionally Oscillating Disk
Boundary Layer
We proceed here to develop the methodology required to analyse the stability of the
oscillatory flow configuration discussed in the previous section, and begin by deriving an
appropriate Reynolds number based on the Stokes length scale δs∗ and the local velocity
scale rL∗ φ∗ where rL is a local radial position. This gives
r∗ φ∗ δs∗
R= = rL (5.5)
ν∗
∗
rL
for rL = δs∗
the non-dimensional radial position. Thus we can fully specify the motion by
supplying the two non-dimensional parameters (R, ). The base flow can be represented
similarly to the steady case as
r r
UB = F, G, H (5.6)
R R R
This gives a similar velocity-vorticity formulation to that for the rotating disk in chapter
2,

1 ∂ξr 1 ∂Nz ∂Nθ 1 2 1 2 ∂ξθ
+ − = ∇ − 2 ξr − (5.7a)
R ∂t r ∂θ ∂z R r r2 ∂θ

1 ∂ξθ ∂Nr ∂Nz 1 2 1 2 ∂ξr
+ − = ∇ − 2 ξθ + (5.7b)
R ∂t ∂z ∂r R r r2 ∂θ

2 1 ∂ξr ∂(rξθ )
∇w= − (5.7c)
r ∂θ ∂r
1
where N is given by (2.38) and the R
factor in the temporal derivative term arises from
the temporal scale of choice. The secondary variables (ur , uθ , ξz ) can be represented in
terms of the primary variables in an identical way to chapter 2, given by (2.37). The
notable lack of the Coriolis terms in (5.7) is a consequence of working in a non-rotating
laboratory frame, in contrast to the methods presented in chapter 2. Since we do not
have a mean rotation rate in this scenario, the concept of a rotating frame of reference is
illogical and thus, for the remainder of this section, we will work only in the non-rotating
frame.
As in chapter 3, we have several methods at our disposal for analysing the stability
of these time-dependent flow configurations. We could use a combination of eigenvalue
analyses from Floquet theory and frozen flow approximations, in conjunction with more
global treatments from both monochromatic simulations and simulations with the full ra-
dial dependence incorporated. Since this exposition is intended to be merely an indication
175
to future directions, we present only the formulation of the Floquet analysis, and leave
any corresponding results to future work. The methods of Floquet analysis are virtually
identical to those already presented in chapter 2 and so in the interest of concision, we
will only reiterate the main points here and the interested reader is directed there for a
more thorough explanation.
Floquet Theory
As in chapter 3, the Floquet stability analysis may be undertaken by introducing small

disturbances to the unsteady base flow UB and writing
u(x, τ ) = UB (z, τ ) + u0 (x, τ ), ξ(x, τ ) = Λ(z, τ ) + ξ 0 (x, τ )
Again following chapter 3 and in particular the method developed by Hall [39], we assume
a Floquet-mode approximation of the form
f (r, θ, z, τ ) = fˆ(z, τ )eµτ ei(αr+βRθ) + c.c. (5.8)
and all exponential growth of fˆ(z, τ ) is incorporated into the Floquet exponent eµτ . The
c.c. denotes the complex conjugate, which is added to ensure that the disturbance is real.
The quantities of interest here will be µ, as its real part specifies the temporal growth of
the disturbance, and α, whose imaginary part gives the spatial growth rate. The βR term
is introduced into the normal-mode solution so as to maintain the structure of the steady
rotating disk. Whether or not the value n = βR retains its experimental importance is as
yet unclear, and any radial homogenisation, parallel flow approximation or replacement of
R by R should be treated with extreme caution. At the time of writing, no exploration of
the effects of the radial homogenisation has been conducted, and such a study is reserved
for future work.
Substitution of (5.8) into (5.7), along with linearisation and neglect of the non-parallel
O R12 terms gives

2 !
2

1 ∂ξr iβR ∂Nθ 1 βR ∂
+ µξr + Nz − = −α2 − + 2 ξr (5.9a)
R ∂τ r ∂z R r ∂z
2 !
2

1 ∂ξθ ∂Nr ∂Nz 1 βR ∂
+ µξθ + − = −α2 − + 2 ξθ (5.9b)
R ∂τ ∂z ∂r R r ∂z
2

ˆ +
2 ∂ iβR ξθ
∇ w = ξr − iαξ θ − (5.9c)
∂z 2 r r
176
Similarly to chapter 2, using the method of Hall [39], we decompose the perturbation
variables and base flow velocity fields into harmonics so that
∞
X
fˆ(z, τ ) = fn (z)einτ (5.10)
n=−∞
and
∞
X
B
U = uk (z)eikτ (5.11a)
k=−∞
The Fourier coefficients uk (z) are calculated from the time-dependent solution data via a
fast Fourier transform, as in chapter 2.
The parallel flow approximation and radial homogenisation in this scenario amounts
to replacing r by R in (5.9) as per the local definition of R. Due to the subtleties in the
global behaviour discussed in chapter 4 for the rotating disk, this approximation would
necessarily require rigorous justification in any future studies carried out on this flow
configuration. For the purposes of the following presentation, we will assume that this is
a valid approximation to make, and present the derivations of the corresponding stability
equations. We make no attempt here to justify this step, and leave the reader simply
with a remark that the implications of such an approximation are unknown at this time.
Therefore, applying the parallel flow approximation and replacing r by R, the secondary
variable ξθ and the convective terms N become
2
Nr = F 0 w − Guθ + Hξθ − Gξz (5.12a)
R R
2
Nθ = G0 w + Gur − Hξr + F ξz (5.12b)
R R
Nz = Gξr − G0 uθ − F 0 ur − F ξθ (5.12c)
∂Nz 1
= iα(Gξr − G0 uθ − F 0 ur − F ξθ ) + (Gξr − G0 uθ − F 0 ur − F ξθ ) (5.12d)
∂r R
and Z ∞
ξz = ξr + iαξr + iβξθ dz (5.13)
0 R
while ur and uθ are still given by (2.50). Subsequently, this problem reduces to a set of
2K + 1 dispersion relations of the form
K
X
Dm {µ, α R, β}eimτ = 0 (5.14)
m=−K
where K is a truncation parameter, chosen so that the eigenvalues calculated are inde-
pendent of the choice of K. It is anticipated that the number of harmonics necessary to
177
fully resolve the eigenvalue solutions would be of the order of that for the Stokes layer,
namely O(αR), which may cause computational difficulties on implementation of this
method. As for the periodically modulated case, there are two distinct types of analysis
which can be conducted, namely temporal and spatial. The temporal approach consists of
specifying α ∈ R and calculating µ from (3.13) while the spatial approach specifies µ ∈ R
and calculates α from (3.13). As yet, no such study has been conducted, and is reserved
for future work.
5.1.2 Hydrodynamic Voltammetry at Rocking Disk Electrodes
The primary motivation for the modulation of the rotation rate in the rotating disk
boundary layer was that of methods for laminar flow control on swept wings. We utilised
the commonly applied approximation that the rotating disk exhibits similar instability
mechanisms to the flow over a swept wing, and applied constraints based on this appli-
cation. These constraints amounted to a small deviation in the rotation rate from the
steady case, and as such no mind was paid to any configurations with large oscillatory
parts. Moving away from this application allows us the freedom to explore the other
end of the spectrum and in particular torsional oscillations. Mathematical curiosity is a
valid motivation for the study of such a system, as it is a fundamental flow configuration
for which no previous hydrodynamic stability studies have been conducted. Indeed, it is
arguably the simplest three-dimensional oscillatory flow configuration and its study is in
natural progression from that of the two-dimensional Stokes layer. However, during the
course of this work, an explicit application of such a configuration was discovered in the
field of electrochemical engineering. This application is as a result of Cummings et. al.’s
[22] modification to the more classically studied rotating disk electrode, for which we will
provide an introductory overview.
Voltammetry is the name given to a class of methods used in analytical chemistry and
various industrial processes. In voltammetry, information about an analyte is obtained
by measuring the current at a working electrode as the potential is varied in a controlled
way. This current can be easily influenced or disturbed by convective behaviour of ions
in the solution, and any fluctuations in the current measured at the electrode can have
detrimental effects on the reproducibility and reliability of results. Thus, the solution
movement is critical to the experiment, and mass transfer at the electrode surface is often
178
controlled by forced convection of the analyte. There are several methods commonly used
for such a forced convection, the most important of which for our applications is that of
the rotating disk electrode.
The steady-state laminar flow properties of the rotating disk electrode are widely
known in the chemical engineering literature, see [24, 35, 16, 61], and Lingwood [43] refer-
ences Chin and Litt [18] for their study of the instability properties of the steady system
from an electrochemical engineering perspective. The modification to the rotating elec-
trode proposed by Cummings et al. [22, 23] and Ahn et al. [1, 2] utilises a rocking motion
for the disk instead, which the relevant authors claim is a novel but easier experimental
setup which gives comparable results. However, the hydrodynamic stability properties of
such a system are as yet mathematically poorly understood, and Marken et. al.1 have
only empirically determined the steady state solution and turbulence breakdown proper-
ties. Therefore, a fundamental mathematical study of the configuration, in tandem with
experimental data from the chemical engineering research, must be conducted to properly
understand the stability and transition properties. Such a study has great relevance to
both fields, would be the first systematic mathematical study of this three-dimensional
fundamental configuration and would enable the chemical engineering applications to
proceed in an optimal way, providing a priori assumptions about the flow configuration.
5.2 The Rotating Disk Boundary Layer with Surface

Roughness
As alluded to in section 3.1, other stabilising techniques for the steady rotating disk
boundary layer have been studied by many authors such as designed surface roughness,
see Garrett et al. [36], Cooper et al. [21] and wall compliance, see Cooper and Carpenter
[19, 20]. Both of these techniques show stabilisation of the type I instability for stationary
disturbances, and Garrett et al. [36] use this as an advertisement for further study of
laminar flow control techniques using designed surface roughness on swept wings. In
all of these aforementioned studies, some form of modelling is carried out whether it be
of the surface or of the flow. This modelling results in an averaging procedure which
essentially removes first-order effects and potentially interesting contributions therein.
1
personal communication
179
The temporally periodic configuration discussed in the main body of this thesis requires
no such modelling, and as such retains the full range of information from the basic state.
Garrett et al. [36] and Cooper et al. [21] describe theoretical results investigating effects
of distributed surface roughness on the convective stability of the rotating disk boundary
layer using two distinct approaches to the modelling of the disk surface. It has now been
firmly established that contrary to the classic belief, the interaction of boundary layer flow
with the right sort of roughness, as discussed by Carpenter [17], can result in energetically
beneficial, drag-reducing effects.
The two approaches considered by Garrett et al. [36] are the so-called MW and YHP
models, attributed to Miklavčič and Wang [52] and Yoon et al. [84], and consist of different
ways of modelling the surface. The MW model for roughness replaces the usual no-slip
boundary conditions with partial-slip conditions at the disk surface. As described in
Garrett et al. [36], this is achieved by introducing slip factors in Newton’s law of viscosity
for the azimuthal and radial velocity component. Selecting different slip factors for each
component enables modelling roughness distributed in the radial and azimuthal directions.
The authors refer to the case where both slip factors are equal as isotropic roughness,
whereas different values for the slip factors represent anisotropic roughness.
The YHP approach models roughness by directly imposing a particular surface profile
as a function of the radial position and, since it assumes rotational symmetry, can only
model anisotropic roughness. A major advantage of the YHP approach over the MW
model is that a specific geometric roughness height can be defined explicitly in terms of
the amplitude and the wavelength of the surface profile.
The following sections will outline the methods of Garrett et al. [36] and present
results reproduced using the methods of chapter 2 which are consistent with the previous
literature. We begin our discussion with an overview of the MW model before giving a
brief account of the YHP model and some insights for future directions. These following
sections are not intended to be novel, but will serve as a useful overview of this currently
highly active field of research and allude to potentials for future collaborative work.
5.2.1 The Partial Slip Model (MW)
In 1921, von Kármán formulated the exact similarity solution to the Navier-Stokes equa-
tions for the steady laminar flow over a rotating disk of infinite radius and section 2.1.1
180
gave a detailed presentation of the solution methods we have employed in this work. A
brief recap of the fundamental equations will be given here in the interest of concision,
but the reader is referred to section 2.1.1 for further details. Following von Kármán and
the nomenclature employed throughout this thesis, let U ∗ , V ∗ and W ∗ denote the dimen-
sional radial, azimuthal and axial similarity velocities respectively and r∗ and z ∗ denote
the dimensional radial position and wall-normal height. Let also ν ∗ denote the kinematic
viscosity and Ω∗0 the angular velocity of the disk. For the results presented in this section,
in contrast to the majority of this thesis, Ω∗0 will be taken to be constant.
Under von Kármán’s [80] similarity scaling, we normalise the velocity fields on a local
q
∗
scale r∗ Ω∗0 and the lengths on the boundary layer thickness δ ∗ = Ων ∗ to get
U∗ V∗ W∗
F (z) = , G(z) = , H(z) = (5.15)
r∗ Ω∗0 r∗ Ω∗0 δ ∗ Ω∗0
which gives the system of ordinary differential equations governing the base flow as
F 2 − G2 + F 0 H − F 00 = 0 (5.16a)
2F G + G0 H − G00 = 0 (5.16b)
H 0 = −2F (5.16c)
This local scaling gives a Reynolds number R, associated with the steady rotation of the
disk as
rL∗ Ω∗0 δ ∗ rL∗
R= = = rL (5.17)
ν∗ δ∗
where we identify rL as a local, non-dimensional radial position. The partial slip model
for roughness is imposed by a boundary condition alteration of the radial and azimuthal
components of the base flow to give
F (0) = λF 0 (0), G(0) = G0 (0) H(0) = 0 (5.18a)
F (ζ) → 0, G(ζ) → −1 as ζ → 0 (5.18b)
Non-zero λ corresponds to radially isotropic roughness while non-zero corresponds to

radially anisotropic roughness. Taking λ = = 0 recovers the smooth case. Since we will
primarily be using this method to compare and contrast against the YHP approach which
models anisotropic roughness, we will set λ = 0 for the remainder of this section.
As explained in section 2.1.1, this system can be solved with a relatively straightfor-
ward application of MATLAB’s bvp4c solver, and evaluated at the Chebyshev collocation
181
points required for the solution procedure. It is worth noting that out-of-the-box, bvp4c
has issues with directly solving the equation on a large physical domain. In fact, it is
numerically unstable over [0, a] for a > 12. This was avoided by introducing an itera-
tive scheme which used the solution over [0, K] as an initial guess for the solution over
[0, K + 1] and proceeded for K ≥ 1. Figure 5.5 shows the base flow profiles for varying de-
grees of anisotropic roughness, achieved by alteration of the parameter in (5.18). These
diagrams agree with those presented by Garrett et al. [36].
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
(a) = 0 (b) = 0.14
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
(c) = 0.57 (d) = 1.18
Figure 5.5: Radial (—), azimuthal (- - -) and vertical (· · · ) base flow profiles for the
rotating disk boundary layer with radially anisotropic surface roughness, imposed by the
partial slip MW model. Various values of are shown, corresponding to varying degrees
of anisotropic roughness.
182
5.2.2 The Surface Geometry Model (YHP)
As discussed in Garrett et al. [36], after scaling lengths on the boundary layer thickness
q
∗
δ ∗ = Ων ∗ and velocities on the local scale r∗ Ω∗ where Ω∗ is the disk rotation rate, we
define a non-dimensional surface roughness function as

2πr
s(r) = Λ cos (5.19)
γ
This particular form of s(r) gives rise to two non-dimensional control parameters, namely
Λ and γ, which may be interpreted as the height and pitch of the roughness respectively.
The authors subsequently define an aspect ratio roughness parameter
Λ
a := (5.20)
γ
as a single parameter describing the surface. Both parameters Λ and γ are expressed in
units of boundary layer thickness as a consequence of the spatial scalings and the standard
von Kármán system are recovered by imposing a = 0.
The base flow equations analogous to the von Kármán ODEs (2.12) may be described
by first applying a coordinate transformation ζ = z − s(r) to give the usual base flow
quantities UB = (U, V, W ) by the transformed variables
u(r, ζ̂) = U (r, z)
v(r, ζ̂) = V (r, z)
w(r, ζ̂) = −s0 (r)U (r, z) + W (r, z)
Scaling the radial dependence in an analogous way to von Kármán and applying the
boundary layer approximation by setting R−1 = rL−1 1 with ζ = Rζ̂, we get
1
f (r, ζ) = u(r, ζ)
r
1
g(r, ζ) = v(r, ζ)
r
h(r, ζ) = w(r, ζ)
which results in a system of equations which determine the base flow given by
∂f ∂h
2f + r + =0 (5.23a)
∂r ∂ζ
s0 s00 2 (1 + g)2

∂f ∂f 02 ∂ f
rf +h + 1+r = 1 + s + (5.23b)
∂r ∂ζ 1 + s02 ∂ζ 2 1 + s02
∂g ∂g ∂ 2g
rf +h = 1 + s02 − 2f (1 + g) (5.23c)
∂r ∂ζ ∂ζ 2
183
This system is complemented by the boundary conditions
f (r, 0) = g(r, 0) = h(r, 0) = 0 (5.24a)
f (r, ζ) → 0, g(r, ζ) → −1 as ζ → 0 (5.24b)
which represent the no-slip and far-field decay conditions at all radial positions in the
rotating reference frame. Note that the usual von Kármán equations (2.12) are recovered
if s(r) = 0, as would be expected.
A notable difference in terms of the base flow equations between the MW and YHP
models is that the YHP method results in a system of partial, rather than ordinary
differential equations. Garrett et al. [36] solve these equations (5.23) using the NAG
routine D03PEF, although we use MATLAB’s bvp5c here. The solvers use an initial
solution given by
f (r, ζ) ∼ rF (ζ), g ∼ rG(ζ), h ∼ H(ζ) (5.25)
at r = 0 to find the velocity profiles at the next increment of r and step forward. This
initial condition is the result of the assumption that the von Kármán equations (5.16)
should be recovered as r → 0. Garrett et al. [36] argue that the flow field arising from the
solution to (5.23) vary at two distinct spatial scales in the radial direction, namely a scale
associated with γ and the similarity scale with r, as per von Kármán. They proceed to
state that since they choose γ . O(10−1 ), it is a reasonable approximation to take a spatial
average of the flow field over any complete cycle in r, thereby leading to a modified von
Kármán mean flow (fˆ(z), ĝ(z), ĥ(z)). This method has the distinct advantage of allowing
previous normal-mode analyses to be achieved without much modification, although the
physical interpretation of such an averaging procedure is unclear and should be explored
further. It is this further exploration that we will discuss during the following exposition,
and allude to potentials for using the numerical simulation techniques of chapter 4 to
calculate the flow response to genuine surface roughness. Figure 5.6 shows the base
flow profiles for varying degrees of anisotropic roughness, achieved by alteration of the
parameter a in (5.20). These diagrams agree with those presented by Garrett et al. [36].
184
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
(a) a = 0 (b) a = 0.1
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
(c) a = 0.2 (d) a = 0.3
Figure 5.6: Averaged radial (—), azimuthal (- - -) and vertical (· · · ) base flow profiles
for the rotating disk boundary layer with radially anisotropic surface roughness, imposed
by the surface geometry YHP model. Various values of a are shown, corresponding to
varying degrees of anisotropic roughness. The quantity a is defined by equation (5.20).
185
5.2.3 Future Directions for Surface Roughness
We conclude this chapter with a brief discussion of the potentials for future work analysing
the stabilisation of the rotating disk boundary layer via imposed surface roughness. Both
approaches taken by Cooper et al. [21] and Garrett et al. [36] and discussed in the pre-
ceding sections involve a model for the surface roughness that culminates in either a
potentially unphysical averaging process or a procedure that inherently neglects first or-
der effects in the system. The numerical methods proposed to deal with the modulated
rotation rate introduced in chapters 2 - 4 require no such modelling, and may therefore
allow first order effects to be considered more cleanly. In terms of the spatial averaging
procedure involved in the YHP model, while required to ensure separability in the radial
direction and enable a normal mode analysis, such an averaging is not required if the inho-
mogeneous simulations discussed in section 4.3 are used instead. As described in section
5.2.2, the YHP model involves a spatial average of the flow field over any complete cycle
in r, thereby leading to a modified von Kármán mean flow (fˆ(z), ĝ(z), ĥ(z)) which may
be fed into the existing normal mode analyses as a modified base flow. The simulations
of section 4.3, by construction, allow for a radial dependence to be incorporated directly
into the base flow, leaving the full radial structure given by the results of equations (5.23)
to be retained throughout the simulations.
Furthermore it is the author’s belief, which we stress at this time is untested con-
jecture, that the simulations may be used to directly impose radially anisotropic surface
roughness without any modification to the base flow or inherent periodic modelling of
the surface. Since the simulations are based on Chebyshev spectral methods in the wall-
normal direction, it is practical and feasible to impose boundary conditions at any vertical
location in space, in a similar manner to that described in the appendices of Davies and
Carpenter [26]. This would allow for a base flow and the subsequent evolution of any
perturbations to be calculated explicitly by the simulations for arbitrary roughness, pro-
vided the roughness extended only along one radial direction and was not azimuthally
localised. Such a study would confirm or otherwise the modelling procedures of Cooper
et al. [21] and Garrett et al. [36] and potentially provide a physically truer demonstration
of the modifications to the stability properties introduced by surface roughness.
186
Chapter 6
Conclusions
The main result of this thesis is to have demonstrated, in various configurations, that
the addition of a modulated rotation rate stabilises the boundary layer formed above a
rotating disk. We have also discussed and described novel solution methods for conduct-
ing linear stability analyses of steady and temporally periodic two and three dimensional
boundary layers, which were presented in the context of application to several canonical
flow configurations. Chapter 1 described the solution methods for the two dimensional
Blasius boundary layer and the Stokes boundary layer formed above an oscillating flat
plate. Chapter 2 introduced a three dimensional boundary layer over a disk rotating at a
steady rate and presented results validating the novel numerical method against previous
work in the literature. Chapters 3 and 4 discussed the effects of the modulated rotation
rate on the stability properties for both stationary disturbances and those resulting from
a radially localised impulsive forcing. Chapter 5 alluded to potential avenues for future
study, including techniques in electrochemistry and the effects of surface roughness. The
solution method employed utilised a velocity-vorticity formulation, first seen in Davies
and Carpenter [26], and while we only presented the three dimensional results for a cylin-
drical polar coordinate system, the corresponding methodology for a Cartesian coordinate
system should be evident.
A key advantage of this approach over traditional primitive variable methods for the
rotating disk problem is that under the velocity-vorticity formulation of Davies and Car-
penter [26], the perturbation equations reduce to three coupled second order equations,
as opposed to six of first order. The perturbation equations comprise of three primary
variables; the perturbations to the wall-normal velocity component and the two vorticity
187
components in the plane of the wall and three secondary variables; the perturbations to
the remaining velocity and vorticity components. The implementation of the Chebyshev-
tau method of Bridges and Morris [14] in tandem with the velocity-vorticity formulation
provides a convenient framework for the linear stability analysis of many flow configu-
rations, and we have presented a format for dealing with the secondary variables and
boundary integral constraints in a novel way.
Our solution method employs a Chebyshev discretisation in the wall-normal direction,
and utilises the Chebyshev-tau method to form systems of integral eigenvalue dispersion
relations which are solved by the MATLAB polynomial eigenvalue solver package polyeig.
The boundary conditions and integral constraints placed on the primary variables are
incorporated into the system by replacing the rows of the matrices that would otherwise
correspond to the arbitrary constants of integration, and the secondary variables are
calculated directly from the primary variables by means of matrix multiplications which
correspond to the integral operators from the application of the Chebyshev-tau method.
Validation has been presented for our solution method against the literature for the
steady rotating disk boundary layer and the semi-infinite Stokes layer above an oscillating
flat plate. As a preliminary investigation, validation for a two dimensional steady config-
uration has also been undertaken against the Blasius boundary layer (see [65]), with good
agreement being found in all cases.
Chapter 3 presents illustrative results for a novel flow configuration, and explores the
stabilising properties of the addition of a modulated rotation rate to the steady rotating
disk boundary layer in the case of disturbances which are stationary with respect to the
disk motion. This configuration is three dimensional and temporally periodic, and the
sections detail the first mathematical application of Floquet theory to the stability of
three dimensional temporally periodic boundary layers. We demonstrated several para-
metric studies of the influence of modulation frequency and angular displacement on the
stabilisation, and validated results using four independent methods; a Floquet dispersion
relation, quasi-steady frozen profile analyses, monochromatic direct numerical simulations
and simulations which retain the radial dependence of the flow. Each independent solu-
tion method was shown to be consistent with the others. Following the presentation of
simulation results, an attempt was made to shed further light on the findings using an
energy analysis. This analysis showed a strong correlation between reduction in one of
188
the energy production terms and the stabilisation behaviour, indicating an area for future
exploration as to the cause of the behaviour. It is worth noting that the reductions in
the energy production term can be substantially larger than other stabilisation techniques
reported in the literature (such as [20, 19, 21, 36]), indicating a much stronger stabilisa-
tion mechanism for our configuration in certain parameter cases. It would, however, be
erroneous to claim that we have understood the true nature of the stabilisation following
this work, and further study must be undertaken before concluding as such.
Chapter 4 presents an exploration of the local absolute instability present in the ro-
tating disk boundary layer, and a discussion of the effects of the modulated rotation rate
on the global behaviour incorporating radial inhomogeneity. This chapter looks at the
evolution of a disturbance in the boundary layer following a localised impulsive forcing at
some fixed radial location, and compares results in the modulated scenario to the steady-
state literature (see [27, 28, 72, 74, 3, 4, 5]). There has been limited previous work on
stabilisation techniques for impulsive forcing in the rotating disk boundary layer, and
this exploration constitutes one of the first such studies. Many other authors, such as
Cooper and Carpenter [20], Cooper et al. [21], Garrett et al. [36] examine only stationary
or fixed-frequency disturbances and do not incorporate the full spatio-temporal distur-
bance structure into their analyses of stabilisation techniques. Cooper and Carpenter [19]
explore the effects of wall compliance on the absolute instability in the flow, but this ref-
erence pre-dates the Davies and Carpenter [27] study on the global behaviour. Therefore,
the author is presently unaware of any studies which incorporate the radial inhomogene-
ity and global disturbance structure into an analysis of stabilisation techniques. Such an
analysis was conducted in the latter stages of chapter 4, where a strong stabilisation was
shown in both radially homogeneous and inhomogeneous cases for impulsive forcing with
periodic modulation of the rotation rate. For certain parameter cases, the local absolute
instability present in the radially homogeneous configuration was stabilised, as shown by
the spatio-temporal contours in figure 4.7. A full parametric study, and analytical ex-
planation of this absolute instability was not presented, and such an analysis is left for
future work. However, from the brief preliminary explorations conducted as part of the
background to this thesis, it is anticipated that similar behaviour to the stationary forcing
will be discovered, in that the range of ϕ ≈ 10 will contribute the strongest stabilisation
while modulations with ϕ >> 20 will have little effect.
189
The final chapter of this thesis described some potentials for future directions be-
ginning with torsional oscillations and applications to certain electrochemical processes.
These avenues for further study were briefly introduced to give a flavour for the potential
applications of our new methods and to indicate the necessity for a fundamental study
of the stability properties of the torsionally oscillating disk; a configuration which is ar-
guably the canonical three-dimensional oscillatory boundary layer and to date has not
been studied from a stability framework. Finally, the chapter concluded with an overview
of the effects of designed surface roughness on the stability properties of the steady ro-
tating disk boundary layer and some parallels between this configuration and the one
with a modulated rotation rate. Two models, see ([21], [36]) were discussed briefly, as
a precursor to future studies involving the numerical methods described throughout this
report. Future work on such a configuration could involve arbitrarily distributed surface
roughness, with the numerical simulations of chapter 4 being utilised to calculate the base
flow and subsequent disturbance development.
190
Bibliography
[1] S. D. Ahn, P. E. Frith, A. C. Fisher, A. M. Bond, and F. Marken. Mass transport

and modulation effects in rocking dual-semi-disc electrode voltammetry. Journal
of Electroanalytical Chemistry, 722-723:78–82, 2014. ISSN 1572-6657. doi: https:
//doi.org/10.1016/j.jelechem.2014.02.018.
[2] S. D. Ahn, K. Somasundaram, H. V. Nguyen, E. Birgersson, J. Yang Lee, X. Gao,

A. C. Fisher, P. E. Frith, and F. Marken. Hydrodynamic voltammetry at a rocking
disc electrode: Theory versus experiment. Electrochimica Acta, 188:837–844, 2016.
ISSN 0013-4686. doi: https://doi.org/10.1016/j.electacta.2015.11.143.
[3] E. Appelquist. Direct numerical simulations of the rotating-disk boundary-layer flow.

PhD thesis, KTH Royal Institute of Technology, 2014.
[4] E. Appelquist, P. Schlatter, P. H. Alfredsson, and R. J. Lingwood. Global linear

instability of the rotating-disk flow investigated through simulations. Journal of
Fluid Mechanics, 765:612–631, 2 2015. ISSN 1469-7645. doi: 10.1017/jfm.2015.2.
[5] E. Appelquist, S. Imayama, P. H. Alfredsson, P. Schlatter, and R.J. Lingwood. Linear

disturbances in the rotating-disk flow: A comparison between results from simula-
tions, experiments and theory. European Journal of Mechanics - B/Fluids, 55, Part 1:
170–181, 2016. ISSN 0997-7546. doi: http://dx.doi.org/10.1016/j.euromechflu.2015.
09.010.
[6] E. Appelquist, P. Schlatter, P. H. Alfredsson, and R. J. Lingwood. Transition to

turbulence in the rotating-disk boundary-layer flow with stationary vortices. Journal
of Fluid Mechanics, 836:43–71, 2018. doi: 10.1017/jfm.2017.771.
[7] P. Balakumar and M. R. Malik. Traveling disturbances in rotating-disk flow. Theo-

retical and Computational Fluid Dynamics, 2(3):125–137, 1990. ISSN 0935-4964.
191
[8] A. J. Bard and L. R. Faulkner. Electrochemical Methods: Fundamentals and Appli-
cations. John Wiley and Sons, 2000. ISBN 9780470452530.
[9] D. J. Benney. The flow induced by a disk oscillating in its own plane. Journal of
Fluid Mechanics, 18(3):385–391, 1964. doi: 10.1017/S0022112064000283.
[10] P. J. Blennerhassett and A. P. Bassom. The linear stability of high-frequency oscil-

latory flow in a channel. Journal of Fluid Mechanics, 556:1–25, 2006.
[11] P. J. Blennerhassett and Andrew P. Bassom. The linear stability of flat stokes layers.
Journal of Fluid Mechanics, 464:393–410, 2002.
[12] John P. Boyd. Spectral methods in fluid dynamics. SIAM Review, 30(4):666–668,
1988. doi: 10.1137/1030157.
[13] J.P. Boyd. Chebyshev and Fourier Spectral Methods: Second Revised Edition. Dover
Books on Mathematics. Dover Publications, 2001. ISBN 9780486411835.
[14] T. J. Bridges and P. J. Morris. Differential eigenvalue problems in which the param-
eter appears nonlinearly. Journal of Computational Physics, 55:437–460, 1984.
[15] R. J. Briggs. Electron-stream Interaction with Plasmas. Cambridge: MIT, 1964.
[16] S. Bruckenstein and B. Miller. Unraveling reactions with rotating electrodes. Ac-
counts of Chemical Research, 10(2):54–61, 1977. doi: 10.1021/ar50110a004.
[17] P. Carpenter. The right sort of roughness. Nature, 388:713–714, 1997. doi: 10.1038/
41870.
[18] D. T. Chin and M. Litt. An electrochemical study of flow instability on a

rotating disk. Journal of Fluid Mechanics, 54:613–625, 1972. doi: 10.1017/
S0022112072000904.
[19] A. J. Cooper and P. W. Carpenter. The stability of rotating-disc boundary-layer flow

over a compliant wall. part 2. absolute instability. Journal of Fluid Mechanics, 350:
261–270, 11 1997.
[20] A. J. Cooper and P. W. Carpenter. The stability of rotating-disc boundary-layer flow

over a compliant wall. part 1. type i and ii instabilities. Journal of Fluid Mechanics,
350:231–259, 1997. doi: 10.1017/S0022112097006976.
192
[21] A. J. Cooper, J. H. Harris, S. J. Garrett, M. Özkan, and P. J. Thomas. The effect
of anisotropic and isotropic roughness on the convective stability of the rotating disk
boundary layer. Physics of Fluids, 27(1):1–16, 2015.
[22] C. Y. Cummings, P. E. Frith, G. Zoppi, I. Forbes, K. D. Rogers, D. W. Lane, and

F. Marken. Rocking disc electro-deposition of copper films on mo/mose2 substrates.
Thin Solid Films, 519(21):7458 – 7463, 2011. ISSN 0040-6090. doi: https://doi.org/
10.1016/j.tsf.2010.12.107. Proceedings of the EMRS 2010 Spring Meeting Symposium
M: Thin Film Chalcogenide Photovoltaic Materials.
[23] C. Y. Cummings, G. Zoppi, I. Forbes, D. Colombara, L. M. Peter, and F. Marken.

Rocking disc electro-deposition of cuin alloys, selenisation, and pinhole effect min-
imisation in cise solar absorber layers. Electrochimica Acta, 79:141 – 147, 2012. ISSN
0013-4686. doi: https://doi.org/10.1016/j.electacta.2012.06.095.
[24] F. Dalton. Ecs classics: Historical origins of the rotating ring-disk electrode. The
Electrochemical Society Interface, 25(3):50–59, 2016. doi: 10.1149/2.F03163if.
[25] J. Daou. Manchester university stability theory. http://www.maths.manchester.

ac.uk/~jd/MATH45132/MATH45132.html, 2016.
[26] C. Davies and P. W. Carpenter. A novel velocity-vorticity formulation of the navier-

stokes equations with applications to boundary layer disturbance evolution. Journal
of Computational Physics, 172(1):119 – 165, 2001.
[27] C. Davies and P. W. Carpenter. Global behaviour corresponding to the absolute

instability of the rotating-disc boundary layer. Journal of Fluid Mechanics, Vol.48:
287–329, June 2003.
[28] C. Davies, C. Thomas, and P.W. Carpenter. Global stability of the rotating-disk
boundary layer. Journal of Engineering Mathematics, 57(3):219–236, 2007. ISSN
0022-0833. doi: 10.1007/s10665-006-9112-8.
[29] S. H. Davis. The stability of time-periodic flows. Annual Review of Fluid Mechanics,
8(1):57–74, 1976.
193
[30] M. R. Dhanak, A. Kumar, and C. L. Streett. Effect of Suction on the Stability of
Flow on a Rotating Disk, pages 151–167. Springer New York, New York, NY, 1992.
ISBN 978-1-4612-2956-8. doi: 10.1007/978-1-4612-2956-8 16.
[31] K. I. Dickson, C. T. Kelley, I. C. F. Ipsen, and I. G. Kevrekidis. Condition estimates

for pseudo-arclength continuation. SIAM Journal on Numerical Analysis, 45(1):263–
276, 2007.
[32] P.G. Drazin. Introduction to Hydrodynamic Stability. Cambridge Texts in Applied

Mathematics. Cambridge University Press, 2002. ISBN 9780521009652.
[33] H. F. Fasel, U. Rist, and U. Konzelmann. Numerical investigation of the three-

dimensional development in boundary-layer transition. AIAA Journal, 28(1):29–37,
2015/05/23 1990. doi: 10.2514/3.10349.
[34] R. Feynman. Turbulence theory gets a bit choppy. http://www.usatoday.com/

tech/science/columnist/vergano/2006-09-10-turbulence-x.htm, 2006.
[35] A. Frumkin, L. Nekrasov, B. Levich, and Ju. Ivanov. Die anwendung der rotierenden
scheibenelektrode mit einem ringe zur untersuchung von zwischenprodukten elektro-
chemischer reaktionen. Journal of Electroanalytical Chemistry (1959), 1(1):84 – 90,
1959. ISSN 0368-1874. doi: https://doi.org/10.1016/0022-0728(59)80012-7.
[36] S. J. Garrett, A. J. Cooper, J. H. Harris, M. Özkan, A. Segalini, and P. J. Thomas.

On the stability of von Kármán rotating-disk boundary layers with radial anisotropic
surface roughness. Physics of Fluids, 28(1), 2016.
[37] M. Gaster. On the generation of spatially growing waves in a boundary layer.

Journal of Fluid Mechanics, 22:433–441, 7 1965. ISSN 1469-7645. doi: 10.1017/
S0022112065000873.
[38] N. Gregory, J. T. Stuart, and W. S. Walker. On the stability of three-dimensional

boundary layers with application to the flow due to a rotating disk. Philosophical
Transactions of the Royal Society of London A: Mathematical, Physical and Engi-
neering Sciences, 248(943):155–199, 1955. doi: 10.1098/rsta.1955.0013.
[39] P. Hall. The linear stability of flat stokes layers. Proceedings of the Royal Society of
London. A. Mathematical and Physical Sciences, 359(1697):151–166, 02 1978.
194
[40] J. J. Healey. Model for unstable global modes in the rotating-disk boundary layer.
Journal of Fluid Mechanics, 663:148–159, 11 2010. ISSN 1469-7645. doi: 10.1017/
S0022112010003836.
[41] D.W. Jordan and P. Smith. Nonlinear Ordinary Differential Equations: An Intro-
duction to Dynamical Systems. Oxford applied and engineering mathematics. Oxford
University Press, 1999. ISBN 9780198565628.
[42] R. E. Kelly and Alison M. Cheers. On the stability of oscillating plane couette flow.
The Quarterly Journal of Mechanics and Applied Mathematics, 23(1):127–136, 1970.
doi: 10.1093/qjmam/23.1.127.
[43] R. J. Lingwood. Absolute instability of the boundary layer on a rotating disk. Journal
of Fluid Mechanics, 299:17–33, 9 1995.
[44] R. J. Lingwood. On the effects of suction and injection on the absolute instability of
the rotating-disk boundary layer. Physics of Fluids (1994-present), 9(5):1317–1328,
1997. doi: http://dx.doi.org/10.1063/1.869246.
[45] R. J. Lingwood. On the impulse response for swept boundary-layer flows. Journal
of Fluid Mechanics, 344:317–334, 1997. doi: 10.1017/S0022112097006149.
[46] R. J. Lingwood and P. H. Alfredsson. Instabilities of the von kármán boundary layer.
Applied Mechanics Reviews, 67(3), 05 2015.
[47] R.J. Lingwood. Stability and transition of the boundary layer on a rotating disk. PhD
thesis, Cambridge University, 1995.
[48] J. Luo and X. Wu. On the linear instability of a finite stokes layer: Instantaneous
versus floquet modes. Physics of Fluids, 22(5):054106, 2010. doi: http://dx.doi.org/
10.1063/1.3422004.
[49] L. M. Mack. Linear stability theory and the problem of supersonic boundary- layer
transition. AIAA Journal, 13(3):278–289, 2015/05/23 1975. doi: 10.2514/3.49693.
[50] M. R. Malik. The neutral curve for stationary disturbances in rotating-disk flow.
Journal of Fluid Mechanics, 164:275–287, 3 1986.
195
[51] M. R. Malik and P. Balakumar. Non-parallel stability of rotating disk flow using pse.
In M.Y. Hussaini, A. Kumar, and C.L. Streett, editors, Instability, Transition, and
Turbulence, ICASE NASA LaRC Series, pages 168–180. Springer New York, 1992.
[52] M. Miklavčič and C. Y. Wang. The flow due to a rough rotating disk. Zeitschrift für
angewandte Mathematik und Physik ZAMP, 55(2):235–246, Mar 2004. ISSN 1420-
9039. doi: 10.1007/s00033-003-2096-6.
[53] The Numerical Algorithms Group (NAG). The nag fortran library. www.nag.com,
n.d.
[54] Netlib. Fftpack, n.d.
[55] William M’F. Orr. The stability or instability of the steady motions of a perfect
liquid and of a viscous liquid. part i: A perfect liquid. Proceedings of the Royal Irish
Academy. Section A: Mathematical and Physical Sciences, 27:9–68, 01 1907. doi:
10.2307/20490590.
[56] H. Othman and T. C. Corke. Experimental investigation of absolute instability of a

rotating-disk boundary layer. Journal of Fluid Mechanics, 565:63–94, 10 2006. ISSN
1469-7645. doi: 10.1017/S0022112006001546.
[57] B. Pier. Finite-amplitude crossflow vortices, secondary instability and transition in

the rotating-disk boundary layer. Journal of Fluid Mechanics, 487:315–343, 6 2003.
ISSN 1469-7645. doi: 10.1017/S0022112003004981.
[58] L. Prandtl. Bemerkungen über die entstehung der turbulenz. ZAMM - Journal of
Applied Mathematics and Mechanics / Zeitschrift für Angewandte Mathematik und
Mechanik, 1(6):431–436, 1921. ISSN 1521-4001. doi: 10.1002/zamm.19210010602.
[59] A. Ramage. Linear disturbance evolution in the semi-infinite Stokes layer and related
flows. PhD thesis, Cardiff University, 2017.
[60] G. Bhaskar Reddy, S. Sreenadh, R. Hemadri Reddy, and A. Kavitha. Flow of a jeffrey
fluid between torsionally oscillating disks. Ain Shams Engineering Journal, 6(1):355
– 362, 2015. ISSN 2090-4479. doi: https://doi.org/10.1016/j.asej.2014.09.004.
196
[61] Pine Research. Rotating electrode theory. https://www.pineresearch.com/
shop/knowledgebase/pine-rotating-electrode-theory/#rotating_disk_
electrode_rde_theory, 2016.
[62] O. Reynolds. An experimental investigation of the circumstances which determine

whether the motion of water shall be direct or sinuous, and of the law of resistance
in parallel channels. Philosophical Transactions of the Royal Society of London, 174:
935–982, 01 1883. doi: 10.2307/109431.
[63] S. Rosenblat. Torsional oscillations of a plane in a viscous fluid. Journal of Fluid

Mechanics, 6(2):206–220, 1959. doi: 10.1017/S002211205900057X.
[64] Oleg S. Ryzhov and E. D. Terent’ev. Streamwise absolute instability of a three-

dimensional boundary layer at high reynolds numbers. Journal of Fluid Mechanics,
373:111–153, 1998. doi: 10.1017/S0022112098002456.
[65] P. J. Schmid and D. S. Henningson. Stability and transition in shear flows. Applied
mathematical sciences. Springer, New York, 2001. ISBN 0-387-98985-4.
[66] D. T. Schwartz, T. J. Rehg, P. Stroeve, and B. G. Higgins. Fluctuating flow with

mass transfer induced by a rotating disk electrode with a superimposed time periodic
modulation. Physics of Fluids A: Fluid Dynamics, 2(2):167–177, 1990. doi: 10.1063/
1.857766.
[67] T. W. Secomb and S. Rosenblat. Torsional oscillations of a non-newtonian fluid with

a free surface. Journal of Fluid Mechanics, 93(pt 4):767–780, 1 1979. ISSN 0022-1120.
[68] A. Sommerfeld. Ein beitrag zur hydrodynamischen erklärung der turbulenten

flüssigkeitsbewegungen. Proc. 4th Intl. Congr. Math., III:116–124, 1908.
[69] A. C. Srivastava. Torsional oscillations of an infinite plate in second-order fluids.

Journal of Fluid Mechanics, 17(2):171–181, 1963. doi: 10.1017/S0022112063001221.
[70] M. J. Taylor and N. Peake. The long-time behaviour of incompressible swept-wing

boundary layers subject to impulsive forcing. Journal of Fluid Mechanics, 355:
359–381, 1998. doi: 10.1017/S002211209700788X.
197
[71] C. Thomas. Numerical simulations of disturbance development in rotating boundary
layers. PhD thesis, Cardiff University, 2007.
[72] C. Thomas and C. Davies. The effects of mass transfer on the global stability of the
rotating-disk boundary layer. Journal of Fluid Mechanics, 663:401–433, 2010.
[73] C. Thomas and C. Davies. Global stability of the rotating-disc boundary layer with
an axial magnetic field. Journal of Fluid Mechanics, 724:510–526, 6 2013. ISSN
1469-7645. doi: 10.1017/jfm.2013.162.
[74] C. Thomas and C. Davies. Global stability of the rotating-disc boundary layer with
an axial magnetic field. Journal of Fluid Mechanics, 724:510–526, 6 2013. ISSN
1469-7645. doi: 10.1017/jfm.2013.162.
[75] C Thomas and C. Davies. On the impulse response and global instability development
of the infinite rotating-disc boundary layer. Submitted to Journal of Fluid Mechanics,
2018.
[76] C. Thomas, A. P. Bassom, P. J. Blennerhassett, and C. Davies. The linear stability

of oscillatory poiseuille flow in channels and pipes. Proceedings of the Royal Society
of London A: Mathematical, Physical and Engineering Sciences, 2011.
[77] W. Tollmien. Über die entstehung der turbulenz. 1. mitteilung. Nachrichten von der
Gesellschaft der Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse,
1929:21–44, 1928.
[78] L. N. Trefethen et al. Chebfun Version 4.2. The Chebfun Development Team, 2011.
http://www.chebfun.org/.
[79] L.N. Trefethen. Spectral Methods in MATLAB. Software, Environments, and Tools.
Society for Industrial and Applied Mathematics, 2000. ISBN 9780898714654.
[80] T. von Kármán. Uber laminaire und turbulente reibung. Zeitschrift für angewandte
Mathematik und Mechanik, 1:233–52, 1921.
[81] C. H. Von Kerczek. The instability of oscillatory plane poiseuille flow. Jour-
nal of Fluid Mechanics, 116:91–114, 3 1982. ISSN 1469-7645. doi: 10.1017/
S002211208200038X.
198
[82] D. J. Wise and P. Ricco. Turbulent drag reduction through oscillating discs. Journal
of Fluid Mechanics, 746:536–564, 2014. doi: 10.1017/jfm.2014.122.
[83] J. R. Womersley. Method for the calculation of velocity, rate of flow and viscous drag
in arteries when the pressure gradient is known. The Journal of Physiology, 127(3):
553–563, 1955.
[84] M. S. Yoon, J. M. Hyun, and J. S. Park. Flow and heat transfer over a rotating disk
with surface roughness. International Journal of Heat and Fluid Flow, 28(2):262 –
267, 2007. ISSN 0142-727X. doi: https://doi.org/10.1016/j.ijheatfluidflow.2006.04.
008.
199
Appendices
200
Appendix A
Convergence of Eigenvalue Routines
A.1 Blasius Boundary Layer

Section 1.4.3 describes the application of the novel numerical method discussed throughout
chapter 1 to a two-dimensional steady flow configuration; the Blasius boundary layer. The
perturbation variables are expanded in terms of Chebyshev polynomials as follows
X
f= fn T2n−1
n
and truncated at some appropriate level. Convergence tests were carried out to empirically
determine the appropriate truncation level when solving the dispersion relation
D(R, α, ω) = 0 (A.1)
in both the temporal (specifying ω as real and solving for α) and spatial (specifying α as
real and solving for ω) settings. Tables A.1-A.2 show the convergence of the most unstable
mode in the temporal and spatial spectra of the Blasius dispersion relation (A.1).
201
α R N ω
4 0.034967 - 0.009858i
12 0.072992 - 0.0032771i
20 0.073044 - 0.0033146i
0.2 500 24 0.073039 - 0.0033168i
48 0.073039 - 0.0033173i
96 0.073039 - 0.0033172i
128 0.073039 - 0.0033172i
4 0.059196 - 0.001624i
12 0.044011 - 0.0071315i
20 0.053099 - 0.0092124i
0.2 10000 24 0.049906 - 0.014714i
48 0.054488 - 0.014431i
96 0.054489 - 0.014431i
128 0.054489 - 0.014431i
Table A.1: Convergence with Chebyshev discretisation order N of the most unstable mode
in the temporal spectra of the Blasius dispersion relation (A.1).
202
ω R N α
4 0.25674 - 0.0050924i
12 0.22695 + 0.0043486i
20 0.22678 + 0.0045134i
0.085 500 24 0.22678 + 0.0045133i
48 0.22677 + 0.0045137i
96 0.22677 + 0.0045137i
128 0.22677 + 0.0045137i
4 0.26979 + 0.036377i
12 0.2811 - 0.00080097i
20 0.27628 + 0.0081857i
0.085 10000 24 0.2705 + 0.015632i
48 0.27381 + 0.012845i
96 0.27381 + 0.012845i
128 0.27381 + 0.012845i
in the spatial spectra of the Blasius dispersion relation (A.1).
A.2 Stokes Boundary Layer

Section 1.4.4 describes the novel numerical method discussed throughout chapter 1, ap-
plied using Floquet theory to a two-dimensional temporally periodic flow configuration;
the Stokes oscillatory boundary layer. The perturbation variables are expanded in terms
of Chebyshev polynomials as follows
X
f= fn T2n−1
n
determine the appropriate truncation level when solving the system of equations that make
up the dispersion relation
Lk (α, R)fˆk = −µI2 fˆk (A.2)
203
in the temporal (specifying µ as real and solving for α) setting. There are two discretisa-
tion orders in this problem that should be considered when assuring convergence, namely
those of the Chebyshev discretisation and the number of harmonics. Table 1.2 in the main
text illustrates the convergence with the number of harmonics while table A.3 shows the
convergence with Chebyshev discretisation order N of the most unstable mode in the
temporal spectra of the Stokes dispersion relation (A.2). The spatial analysis was not
possible in this setting due to memory constraints and the number of harmonics required.
α R N µr
4 0.97922
12 -0.052847
20 -0.044284
0.3 700 24 -0.045992
48 -0.045159
96 -0.045016
128 -0.045016
4 1.1681
12 -0.052528
20 0.11162
0.3 800 24 0.1487
48 0.081821
96 0.081745
128 0.081745
in the temporal spectra of the Stokes dispersion relation (A.2). In each case, the number
of harmonics was chosen so as to be greater than the number required by Blennerhassett
and Bassom [11] for convergence.
A.3 Steady Rotating Disk Boundary Layer

Section 2.2.1 describes the application of the novel numerical method discussed throughout
chapter 2 to a three-dimensional steady flow configuration; the rotating disk boundary
204
layer. The perturbation variables are expanded in terms of Chebyshev polynomials as
follows
X
f= fn T2n−1
n
determine the appropriate truncation level when solving the dispersion relation
D(R, α, n, ω) = 0 (A.3)
for the temporal (specifying ω and n as real and solving for α) setting and for station-
ary (ω = 0) disturbances with fixed n in the spatial setting. Tables A.4-A.5 show the
convergence of the most unstable mode of the steady rotating disk dispersion relation
(A.3).
205
α n R N ω
4 -0.060424 + 0.0040383i
12 -0.014563 + 0.011932i
20 -0.013946 + 0.0030149i
0.3 32 350 24 -0.013171 + 0.0022687i
48 -0.013364 + 0.0020526i
96 -0.013364 + 0.0020526i
128 -0.013364 + 0.0020526i
4 -0.042808 + 0.0090325i
12 -0.011335 - 0.00091879i
20 -0.010526 - 0.0029679i
0.2-0.12i 32 400 24 -0.010689 - 0.0032092i
48 -0.01068 - 0.0032368i
96 -0.010682 - 0.003239i
128 -0.010682 - 0.003239i
4 -0.053142 + 0.064032i
12 -0.044739 + 0.018807i
20 -0.035369 + 0.0046288i
0.3 68 500 24 -0.033816 + 0.0046894i
48 -0.03341 + 0.0046586i
96 -0.03341 + 0.0046586i
128 -0.03341 + 0.0046586i
in the temporal spectra of the steady rotating disk dispersion relation (A.3).
206
ω n R N α
4 0.041031 - 0.1251i
12 0.21696 + 0.036043i
20 0.42371 + 0.10186i
0 32 350 24 0.45287 - 0.023098
48 0.46309 - 0.015994i
96 0.46309 - 0.015993i
128 0.46309 - 0.015993i
4 0.040614 - 0.1425i
12 0.21541 + 0.073515i
20 0.39671 + 0.073663i
0 68 550 24 0.59856 + 0.021235i
48 0.60954 - 0.008059i
96 0.60955 - 0.0080429i
128 0.60955 - 0.0080429i
4 0.039256 - 0.12277i
12 0.19899 + 0.053982i
20 0.27861 + 0.28042i
-0.01023 32 300 24 0.43467 - 0.007954i
48 0.44191 - 0.0020801i
96 0.44191 - 0.0020794i
128 0.44191 - 0.0020794i
in the spatial spectra for stationary disturbances in the steady rotating disk dispersion
relation (A.3).
207
A.4 Periodically Modulated Rotating Disk Bound-
ary Layer
Section 3.1 describes the novel numerical method discussed throughout chapter 3, ap-
plied using Floquet theory to the three dimensional periodically modulated rotating disk
boundary layer on which the bulk of this thesis is based. The perturbation variables are
expanded in terms of Chebyshev polynomials as follows
X
f= fn T2n−1
n
and truncated at some appropriate level NC . Convergence tests were carried out to em-
pirically determine the appropriate truncation level when solving the system of equations
that make up the dispersion relation
∞
X
Dm {µ, α R, n}eimτ̃ = 0 (A.4)
m=−∞
for the temporal (specifying ω and n as real and solving for α) setting and for stationary
(ω = 0) disturbances with fixed n in the spatial setting.
There are three discretisation orders in this problem that should be considered when
assuring convergence, two of which are the Chebyshev discretisation order, NC , and the
number of harmonics, NF . The third discretisation order to be taken into account is the
number of terms taken in the Fourier expansion of the base flow UM (z, τ̃ )
UM (z, τ̃ ) := UB (z, τ̃ ) − US (z) (A.5)

NB
X
= uk (z)eikτ̃ (A.6)
k=−NB
as given by equation (3.11) in the main text. Tables 3.2-3.3 in the main text illustrate
the convergence with the number of harmonics and terms in the base flow expansion
while table A.6 shows the convergence with Chebyshev discretisation order N of the
most unstable mode in the temporal spectra of the periodically modulated rotating disk
dispersion relation (A.4). The results for the spatial analysis are similar and so are not
presented here in the interest of concision.
208
R α Uw ϕ NC µ
48 0.004229 - 0.011160i
64 0.004229 - 0.011160i
500 0.3 0.2 5 96 0.004229 - 0.011160i
128 0.004229 - 0.011160i
48 0.004206 - 0.020812i
64 0.004206 - 0.020812i
500 0.3 0.2 10 96 0.004206 - 0.000811i
128 0.004206 - 0.000811i
48 0.004497 - 0.000840i
64 0.004497 - 0.000840i
500 0.3 0.2 15 96 0.004497 - 0.000840i
128 0.004497 - 0.000840i
Table A.6: Convergence with Chebyshev discretisation order NC of the most unstable
mode in the temporal spectra for stationary disturbances in the periodically modulated
rotating disk dispersion relation (A.4).
209
Appendix B
Algorithms and Numerical Schemes
B.1 Numerical Solution of the Base Flow Equations

We will begin our discussion with the numerical solution procedure for the periodically
modulated rotating disk base flow equations (2.20a), and will build to describing the
algorithm used for the full simulations with radial dependence procedure described in
section 3.5. From equations (2.20a), we have
ZZ ZZ ZZ ZZ ZZ
3 2 l+1 2 1
− D F dη = F l dη − l−1
F dη + RlF dη
2∆τ ∆τ 2∆τ
ZZ ZZ ZZ ZZ ZZ
3 2 l+1 2 1
− D G dη = Gl dη − l−1
G dη + RlG dη
2∆τ ∆τ 2∆τ
subject to the boundary conditions

√
Rs ϕ ϕ
F (1, τ ) = H(1, τ ) = 0, G(1, τ ) = 1 + cos τ
R R
F → 0, G → 0 as η → 0
as described in section 3.5. We can write this for simplicity as
F l+1 = L−1 P1 F l − P2 F l−1 + RlF

(B.1a)
Gl+1 = L−1 P1 Gl − P2 Gl−1 + RlG

(B.1b)
and solved with a predictor-corrector algorithm, using an algorithm akin to the following
diagram:
210
Initialise (F, G, H) = (Fvk , Gvk , Hvk )
Initialise L−1 and set boundary conditions
Start temporal march
Calculate RF and RG
1
3RlF − Rl−1
Fp = P1 F l − P2 F l−1 +

F
Predictor step: 2
1
Gp = P1 Gl − P2 Gl−1 + 3RlG − Rl−1

G
2
Set boundary conditions and multiply L−1 (FP , GP )
Loop Solve continuity equation for H
Calculate RF and RG using (FP , GP , H)
1
F = P1 FPl − P2 FPl−1 +
RlF + Rl−1

F
Corrector step: 2
l−1 1
l
RlG + Rl−1

G = P1 GP − P2 GP + G
2
Set boundary conditions and multiply L−1 (F, G)
End temporal march: reset flow fields
211
B.2 Monochromatic Numerical Simulations
The algorithm for the solution of the equations which appear in section 2.2.4 when con-
ducting monochromatic direct numerical simulations is very similar to that presented in
section B.1, although we will outline it here for clarity. The equations (2.59) are given by

∂ξr ∂Nθ 2 1 ˆ 2 1 2iβ
+ iβNz − − (ξθ + iαw) = ∇ − 2 ξr − ξθ
∂t ∂z R R r R

∂ξθ ∂Nr ∂Nz 2 1 ˆ 2 1 2iβ
+ − + (ξr − iβw) = ∇ − 2 ξθ + ξr
∂t ∂z ∂r R R r R

ˆ 2 1 ∂ξr ∂(rξθ )
∇w= −
r ∂θ ∂r
which we integrate and rearrange to get

ZZ ZZ ZZ ZZ ZZ
3 1 2 l+1 2 1
− D ξr dη = ξrl dη − l−1
ξr dη + Rlr dη
2∆τ 2R ∆τ 2∆τ
ZZ ZZ ZZ ZZ ZZ
3 1 2 l+1 2 1 l−1
− D ξθ dη = ξθl dη − ξθ dη + Rlθ dη
2∆τ 2R ∆τ 2∆τ
where
∂Nθl 1 ∂2

2 l 1 iα 1 2iβ l
Rlr = ξθ + iαwl + − iβNzl + −α2 + − β2 + − ξl − ξ
R ∂z R R 2 ∂z 2 R2 r R θ
∂Nzl ∂Nrl 1 ∂2

2 l 1 iα 1 2iβ l
Rlθ = − ξr − iβwl + − + −α2 + − β2 + − ξl + ξ
R ∂r ∂z R R 2 ∂z 2 R2 θ R r
together with the continuity equation
∂2 ξl

iα
2
−α + − β + 2 wl = iβξθl − iαξrl − r
2
R ∂z R
As with section B.1, we may write this succinctly as
ξrl+1 = L−1 P1 ξrl − P2 ξrl−1 + Rlr

ξθl+1 = L−1 P1 ξθl − P2 ξθl−1 + Rlθ

which gives a similar equation set to (B.1). The algorithm may be explained by the
following diagram:
212
Initialise (ξr , ξθ , w) = (0, 0, 0)
Initialise L−1 and set boundary conditions
Start temporal march
Calculate temporally localised impulsive forcing
Calculate Rr and Rθ
1
ξ˜r = P1 ξrl − P2 ξrl−1 + 3Rlr − Rl−1

r
Predictor step: 2
1
ξ˜θ = P1 ξθl − P2 ξθl−1 + 3Rlθ − Rl−1

θ
2
Set boundary conditions and multiply L−1 (ξ˜r , ξ˜θ )
Loop
Solve Poisson equation for w
Calculate Rr and Rθ using (ξrP , ξθP , w)
1
ξr = P1 ξ˜rl − P2 ξ˜rl−1 + Rlr + Rl−1

r
Corrector step: 2
1
ξθ = P1 ξ˜θl − P2 ξ˜θl−1 + Rlθ + Rl−1

θ
2
Set boundary conditions and multiply L−1 (FP , GP )
End temporal march: reset flow fields
213
B.3 Numerical Simulations with Radial Dependence
The algorithm for the simulations which include the radial dependence of the perturba-
tions is similar to that described in section B.2, although has the distinct difference that
there are radial derivatives to be calculated and a radial domain to be taken into account.
This leads to the need for radial inflow and outflow conditions to be imposed in the al-
gorithm and highly accurate finite difference schemes to calculate the radial derivatives.
The solution to the Poisson equation also becomes trickier with this modification as it
cannot now be simply solved by a matrix inversion across the boundary layer. The inflow
and outflow conditions are described in section B.3.1 while the finite difference schemes
are described in B.3.2. Finally, this section concludes with a discussion of the Poisson
solver routine in section B.3.3. The equations to be solved are given by (2.36) but are
stated here again for completeness:

∂ξr 1 ∂Nz ∂Nθ 2 ∂w 1 2 1 2 ∂ξθ
+ − − ξθ + = ∇ − 2 ξr − (B.2a)
∂t r ∂θ ∂z R ∂r R r r2 ∂θ

∂ξθ ∂Nr ∂Nz 2 1 ∂w 1 2 1 2 ∂ξr
+ − + ξr − = ∇ − 2 ξθ + (B.2b)
∂t ∂z ∂r R r ∂θ R r r2 ∂θ

2 1 ∂ξr ∂(rξθ )
∇w= − (B.2c)
r ∂θ ∂r
where
N = (Nr , Nθ , Nz ) = (∇ × UB ) × u + ξ × UB
and
∂2 1 ∂ 1 ∂2 ∂2
∇2 = 2 + + 2 2+ 2
B B
∂r r ∂r
B
r ∂θ ∂z
B
∂UrB

B 1 ∂U z ∂U θ ∂U r ∂U z 1 ∂ B
∇×U = − er + − eθ + rUθ − ez
r ∂θ ∂z ∂z ∂r r ∂r ∂θ
are the usual Laplacian and curl operators in cylindrical polar coordinates.
B.3.1 Inflow and Outflow Conditions
A thorough examination of the inflow and outflow conditions required in the steady rotat-
ing disk configuration has been undertaken for this problem by several authors including
Davies and Carpenter [27] and Thomas [71]. Following these authors, all perturbation
quantities are set to zero at the inflow boundary, which is placed sufficiently far away from
214
the domain of interest to not influence the results. The outflow condition is somewhat
trickier, and for the purposes of the current investigation we impose a wavelike condition
on the second derivative of the normal velocity, similar to Fasel et al. [33], namely
∂ 2w
= −α2 w
∂w2
for some radial wavenumber α. In any case, the outflow boundary is placed sufficiently
far away from the initial forcing so as not to impact the results. These conditions have
been explored by Davies and Carpenter [27] and Thomas [71] in detail, so we proceed
with confidence in their application to the current problem in question.
B.3.2 Finite Difference Discretisation Schemes
The finite difference schemes used in the radial direction to calculate r-derivatives of the
perturbation quantities are similar to those found in Fasel et al. [33] and are described
below for completeness. As discussed in the main text, the schemes are centered and
non-compact.
First Derivative
f j+1 − f j−1

∂f
=
∂r {j=2} 2∆r
11f − 18f j−1 + 9f j−2 − 2f j−3
j

∂f
=
∂r {j=N } 6∆r
(f j−2 − f j+2 ) + 8(f j+1 − f j−1 )

∂f
=
∂r {j=3,j=N −2} 12∆r
(f j+3 − f j−3 ) + 9(f j−2 − f j+2 ) + 45(f j+1 − f j−1 )

∂f
=
∂r {j=4,j=N −3} 60∆r
12f j+1 + 65f j − 120f j−1 + 60f j−2 − 20f j−3 + 3f j−4

∂f
=
∂r {j=N −1} 60∆r
3(f j−4 − f j+4 ) + 32(f j+3 − f j−3 ) + 168(f j−2 − f j+2 ) + 672(f j+1 − f j−1 )

∂f
=
∂r {j=5,...,N −4} 840∆r
215
Second Derivative
∂ 2f 2(f j+3 + f j−3 ) − 27(f j+2 + f j−2 ) + 270(f j+1 + f j−1 ) − 490f j

=
∂r2 {j=4,...,N −3} 180(∆r)2
∂ 2f −(f j+2 + f j−2 ) + 16(f j+1 + f j−1 ) − 30f j

=
∂r2 {j=3,j=N −2} 12(∆r)2
2
(f j+1 + f j−1 ) − 2f j

∂ f
=
∂r2 {j=2} (∆r)2
2
∂ f 10f j+1 − 15f j − 4f j−1 + 14f j−2 − 6f j−3 + f j−4
=
∂r2 {j=N −1} 12(∆r)2
2
∂ f
= −α2 f j
∂r2 {j=N }
B.3.3 Poisson Equation
The solution to the Poisson equation is computed by an iterative solver embedded within
the main predictor-corrector algorithm used to compute the solutions to the vorticity
transport equations. In particular, on expanding the Poisson equation in (B.2), we have
2
n2 ∂2

∂ 1 ∂ n ∂ξθ 1
2
+ − 2 + 2 w = ξr − − ξθ
∂r r ∂r r ∂z r ∂r r
Clearly, the radial derivatives are computed in practise by the finite difference schemes
described in B.3.2. However, to illustrate the method and for notational brevity, we will
use a simpler scheme here, namely
f j+1 − f j−1

∂f
=
∂r {j=2,...,N −1} 2∆r
f j − f j−1

∂f
=
∂r {j=N } ∆r
2
∂ f f j+1 − 2f j + f j−1
=
∂r2 {j=2,...,N −1} (∆r)2
2
∂ f
= −α2 f j
∂r2 {j=N }
At the interior points j = 2, . . . , N − 1, we thus write
n2 ∂2
Z Z j+1
+ wj−1 1 wj+1 − wj−1
ZZ ZZ
2 j w
+ 2 − 2 w dz = dz + dz
(∆r)2 rj ∂z (∆r)2 rj ∆r
!
∂ξθj
ZZ
1 j n j
+ + ξθ − ξr dz
∂r rj rj
216
while we have
!
n2 ∂2 1 wj − wj−1 ∂ξθj
ZZ ZZ ZZ
1 n
2 j
α + 2 − 2 w dz = dz + + ξθj − ξrj dz
rj ∂z rj ∆r ∂r rj rj
at the outflow boundary j = N . The perturbation quantities are set to zero at the inflow
boundary which is placed far away from the centre of excitation, as discussed in the
opening to this section. This allows us to write the equation succinctly as
Lwj = P1 wj+1 + P2 wj−1 + Rj
and use an iterative solver to obtain the solution which may be described as follows:
Start iterative loop
Set up L for each radial coordinate point
Calculate: Lwj = P1 wj+1 + P2 wj−1 + Rj
Loop Streamwise march
Set boundary conditions using wall velocity
Multiply L−1 (Lw)
Stop iterative loop
217

2018 - Scott Morgan - Stability of Periodically Modulated Rotating Disk Boundary Layers

Uploaded by

Copyright:

Available Formats

2018 - Scott Morgan - Stability of Periodically Modulated Rotating Disk Boundary Layers

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2018 - Scott Morgan - Stability of Periodically Modulated Rotating Disk Boundary Layers

Uploaded by

Copyright:

Available Formats

Stability of Periodically Modulated

Rotating Disk Boundary Layers

Submitted for the degree of Doctor of Philosophy

Signed . . . . . . . . . . . . . . . . . . . . . . . . (candidate) Date . . . . . . . . . . . .

Signed . . . . . . . . . . . . . . . . . . . . . . . . (candidate) Date . . . . . . . . . . . . . . . . . .

Signed . . . . . . . . . . . . . . . . . . . . . . . . (candidate) Date . . . . . . . . . . . . . . . . . .

Signed . . . . . . . . . . . . . . . . . . . . . . . . (candidate) Date . . . . . . . . . . . . . . . . . .

1 Linear Stability Analysis of Two-Dimensional Boundary Layers using a

2 Local Linear Stability Analysis of the Steady Rotating Disk Boundary

3 Local Stability of the Periodically Modulated Rotating Disk Boundary

4 Global Behaviour and Absolute Instabilities in Periodically Modulated

5 Future Directions 170

A Convergence of Eigenvalue Routines 201

B Algorithms and Numerical Schemes 210

Linear Stability Analysis of

1.1.1 About this Thesis

1.2 The Orr-Sommerfeld Equation

Non-dimensionalisation is carried out by some characteristic velocity scale V and length

U = (U (z), 0, 0), P = constant for z1 ≤ z ≤ z2

u(x, t) = U(z) + u0 (x, t), p(x, t) = P + p0 (x, t)

u0 = û(z)ei(αx+βy−ωt) , p0 = p̂ei(αx+βy−ωt) (1.5)

(D2 − α2 )2 − iR (αU − ω)(D2 − α2 ) − αU 00

φ̂(x, z, t) = φ(z)ei(αx−ωt) (1.7)

The boundary conditions that are imposed on φ are

φ(z1 ) = 0, φ(z2 ) = 0, φ0 (z1 ) = 0, φ0 (z2 ) = 0 (1.8)

1.2.1 Eigenvalue Solutions of the Orr-Sommerfeld Equation

D(ω, α, Re) = 0 (1.9)

1.2.2 Inviscid Theory and Rayleigh’s Equation

1.3 Numerical Solutions of the Orr-Sommerfeld Equa-

Tn (x) = cos(n arccos(x)), x ∈ [−1, 1] (1.11)

Tk+1 (x) = 2xTk (x) − Tk−1 (x) for k ∈ {1, 2, ...}

1.3.1 The Fast Fourier Transform for Chebyshev Series

We begin by introducing Fourier spectral methods in general, before moving towards

and we can interpret the transformed variable F as a decomposition of the physical

while the inverse discrete Fourier transform is given by

Tn (x) = cos(nθ) (1.19)

which is equivalent up to an overall scale factor of 2 to a discrete Fourier transform of

1.3.2 Numerical Solution of the Orr-Sommerfeld Equation

Considering the interpolating polynomial p to an arbitrary appropriate function f defined

where a is an arbitrary constant of integration. This results in a tridiagonal banded

1.4 Linear Stability of Two-Dimensional Parallel Flows

where bi , i = 1, . . . , 4 denote constants of integration. We note that, throughout this

Linearisation can be performed if necessary by dropping out the ξ × u term in N.

Reduction of the Full System to Two Dimensions

∂ ξˆ ∂ ξˆ ∂(ûω) ∂(ŵξ) 1 1 000

where the single secondary variable û is given by

1.4.2 Two-Dimensional Steady Base Flow

where a0 , a1 , b0 and b1 are arbitrary constants of integration.

where f is an arbitrary perturbation variable. The boundary conditions are included

Local Eigenvalue Iteration

Given an appropriate initial guess αk , a cubically convergent method described in Bridges

f (αk ) = Tr[D−1 (αk )D0 (αk )]

1.4.3 Application of Solution Method to the Blasius Boundary

Figure 1.3: Schematic of Blasius boundary layer flow configuration.

Defining a streamfunction ψ such that

and substituting into the boundary layer equations gives

Following the standard derivation, introducing a similarity variable ζ such that

we arrive at the third order ordinary differential equation

Numerical Solution of the Base Flow Equations

u(x, t) = U(z) + u0 (x, t), p(x, t) = P + p0 (x, t)

u(x, τ ) = UB (z, τ ) + u0 (x, τ ), ξ(x, τ ) = Ξ(z, τ ) + ξ 0 (x, τ )

Ω∗ (t∗ ) = Ω∗0 + φ∗ cos(φ∗ t∗ ) (2.3)

F = F0 + ˆf1 + ˆ2 f2 + O(ˆ3 )

G = G0 + ˆg1 + ˆ2 g2 + O(ˆ3 )

H = H0 + ˆh1 + ˆ2 h2 + O(ˆ3 )