instability analysis of incompressibleopencavity flows · instability analysis of...
TRANSCRIPT
ESCUELA TECNICA SUPERIOR DE INGENIEROS AERONAUTICOS
UNIVERSIDAD POLITECNICA DE MADRID
Doctoral Thesis
Instability Analysis of Incompressible Open Cavity Flows
by
Fernando Meseguer Garrido
Aeronautical Engineer
Advisors
Eusebio Valero Sanchez and Javier de Vicente Buendıa
Madrid, October 2014
iii
Dedicado a mis padres
iv
Abstract
The problem of the flow over an open cavity has been studied in depth in the literature,
both for being an interesting academical problem and due to the multitude of industrial
applications, like the landing gear of aircraft, or the water deposit of firefighter airplanes.
The different types of instabilities appearing in this flow studied in the literature are two:
the two-dimensional shear layer modes, and the three-dimensional modes that appear in
the main recirculating vortex inside the cavity.
In this thesis a parametric study in the incompressible limit of the problem is pre-
sented, using the linear stability analysis known as BiGlobal. This approximation allows
to obtain the global stability behaviour of the flow, and to capture both the morphological
features and the characteristics of the eigenmodes of the physical problem, whether they
are stable or unstable. The study presented here characterizes with great detail all the
relevant eigenmodes, as well as the hypersurface of instability on the parameter space of
the incompressible problem (Mach equal to zero, and variation of the Reynolds number,
the incoming boundary layer thickness, the length to depth aspect ratio of the cavity and
the spanwise length of the perturbation). The results allow to construct parametric re-
lations between the characteristics of the leading eigenmodes and the parameters of the
problem, like for example the one existing between the critical Reynolds number and its
characteristic length.
The numerical results presented here are compared with those of an experimental
campaign, with the main conclusion of said comparison being that the linear eigenmode are
present in the real saturated flow, albeit with some significant differences in the frequencies
of the experiments and those predicted by the theory. To try to determine the nature of
those differences a three-dimensional direct numerical simulation, analyzed with Dynamic
Mode Decomposition algorithm, was used to describe the process of saturation.
v
vi
Resumen
El problema del flujo sobre una cavidad abierta ha sido estudiado en profundidad en la
literatura, tanto por el interes academico del problema como por sus aplicaciones practicas
en gran variedad de problemas ingenieriles, como puede ser el alojamiento del tren de
aterrizaje de aeronaves, o el deposito de agua de aviones contraincendios. Desde hace
muchos anos se estudian los distintos tipos de inestabilidades asociadas a este problema:
los modos bidimensionales en la capa de cortadura, y los modos tridimensionales en el
torbellino de recirculacion principal dentro de la cavidad.
En esta tesis se presenta un estudio parametrico completo del lımite incompresible
del problema, empleando la herramienta de estabilidad lineal conocida como BiGlobal.
Esta aproximacion permite contemplar la estabilidad global del flujo, y obtener tanto la
forma como las caracterısticas de los modos propios del problema fısico, sean estables o
inestables. El estudio realizado permite caracterizar con gran detalle todos los modos
relevantes, ası como la envolvente de estabilidad en el espacio parametrico del problema
incompresible (Mach nulo, variacion de Reynolds, espesor de capa lımite incidente, relacion
altura/profundidad de la cavidad, y longitud caracterıstica de la perturbacion en la di-
reccion transversal). A la luz de los resultados obtenidos se proponen una serie de rela-
ciones entre los parametros y caracterısticas de los modos principales, como por ejemplo
entre el Reynolds crıtico de un modo, y la longitud caracterıstica del mismo.
Los resultados numericos se contrastan con una campana experimental, siendo la prin-
cipal conclusion de dicha comparacion que los modos lineales estan presentes en el flujo
real saturado, pero que existen diferencias notables en frecuencia entre las predicciones
teoricas y los experimentos. Para intentar determinar la naturaleza de dichas diferencias
se realiza una simulacion numerica directa tridimensional, y se utiliza un algoritmo de
DMD (descomposicion dinamica de modos) para describir el proceso de saturacion.
vii
viii
0 Acknowledgements ix
Acknowledgements
This work would have been impossible without the help and support of many people.
I want to thank all of them, for making this thesis happen.
First of all I want to thank my advisors, Eusebio Valero and Javier de Vicente for
their help and wisdom, and for this chance to start exploring the world of science. In this
regard I would also like to thank many other people for many fruitful discussions on a
great variety of topics that have led to this thesis, like professors Vassilis Theofilis, Leo
Gonzalez, Julio Soria or Jeremy Basley, among many others.
I also want to thank all my colleagues in the Applied Mathematics and Statistics De-
partment, for creating and incredible work environment of which I have the great pleasure
to be a part of. In the last few years I have had the luxury of entering in another wonderful
place, which is the Department of Aerospace Vehicles. I am immensely grateful for the
opportunity that I have been provided to develop a career in university, and the support
of my colleagues in the Design group has been outstanding.
Another aspect of which I am incredibly grateful for is the chance to travel as part
of this PhD, through the UPM funds for short stays, the help of the Applied Mathemat-
ics and Statistics Department or the Marie Curie Grant PIRSES-GA-2009-247651 “FP7-
PEOPLE-IRSES: ICOMASEF Instability and Control of Massively Separated Flows”. I
had the opportunity to meet great people around the globe, in Florida, Hawaii, Norway,
Italy or Australia. I have an extra word of acknowledgements for the many wonderful
friends I found in Australia, for making me feel at home while I could not be farther away
from home.
I also want to thank my friends and family, without whom this long journey would
have never been possible. The years have come and gone, but they have been always made
worthy by the people that surrounded me.
My last and greatest thanks is for my parents. I could not be more proud to be your
son.
x
0 Acknowledgements xi
Agradecimientos
Este trabajo habrıa sido imposible sin la ayuda de mucha gente. Quiero agradecerselo a
todos ellos, por hacer esta tesis posible.
En primer lugar quiero agradecer a mis tutores de tesis, Eusebio Valero y Javier de
Vicente por su ayuda y su sabidurıa, asi como por esta oportunidad de empezar a explorar
el mundo de la ciencia. En este aspecto tambien quiero agradecerle a mucha otra gente
las discusiones sobre temas muy variados que al final han conducido a esta tesis, como
los profesores Vassilis Theofilis, Leo Gonzalez, Julio soria o Jeremy Basley, entre otros
muchos.
Tambien quiero dar las gracias a todos mis companeros en el Departamento de Matematica
Aplicada y Estadıstica, por crear un entrono de trabajo increible del que tengo el inmenso
placer de formar parte. En los ulitmos anos he tenido la suerte de entrar en otro sitio
maravilloso, el Vepartamento de Aeronaves y Vehıculos Espaciales. Estoy inmensamente
agradecido por la oportunidad que se me ha brindado de desarrollar una carrera en la
universidad, y el apoyo de mis companeros en la catedra de Dibujo ha sido extraordinario.
Otro aspecto por el cual estoy increiblemente agradecido es por la oportunidad que he
tenido de viajar durante este doctorado, a traves de las ayudas del programa propio de
becas de la UPM, la asistencia del Departamento de Matematica Aplicada y estadıstica o
la beca Marie Curie PIRSES-GA-2009-247651 “FP7-PEOPLE-IRSES: ICOMASEF Insta-
bility and Control of Massively Separated Flows”. He tenido la oportunidad de conocer a
gente alrededor del mundo, en Florida, Hawaii, Noruega, Italia o Australia. Tengo que dar
las gracias especialmente a los numerosos amigos que conocı en Australia, por hacerme
sentir como si estuviera en casa cuando no podıa estar mas lejos de casa.
Quiero dar las gracias tambien a mis amigos y a mi familia, sin los cuales este largo
camino hubiera sido imposible. Los anos pasan uno tras otro, pero si merecen la pena es
gracias a la gente que me rodea.
Mi ultimo y mayor agradecimiento es para mis padres. No podrıa estar mas orgulloso
de ser vuestro hijo.
xii
Contents
Abstract v
Resumen vii
Acknowledgements ix
Contents xiii
1 Introduction 1
2 Hydrodynamic Stability 11
2.1 Linear Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Non-Modal Instability . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Modal Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 BiGlobal Linear Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Numerical Methods 21
3.1 Short Review of Spectral Methods . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Base Flow Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 BiGlobal instability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Three-dimensional Direct Numerical Simulation . . . . . . . . . . . . . . . . 29
3.5 Dynamic Mode Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 The three-dimensional dynamics 35
xiii
xiv CONTENTS
4.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.1.1 On complex conjugate eigenvalues . . . . . . . . . . . . . . . . . . . 39
4.2 Global instability validation . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.1 Comparison with Bres & Colonius . . . . . . . . . . . . . . . . . . . 44
4.2.2 Comparison with 3D DNS . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Parametric analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.1 Preliminary considerations on the envelope of stability . . . . . . . . 50
4.3.2 Spanwise wavenumber (β) . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3.3 Reynolds number (ReD) . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3.4 Incoming boundary-layer thickness (θ0/D) . . . . . . . . . . . . . . . 57
4.3.5 Aspect ratio (L/D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3.6 Eigenmode morphology . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3.7 Synthesis of the results . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Summary of the chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5 The two-dimensional limit 77
5.1 Convergence of BiGlobal analysis on β → 0 . . . . . . . . . . . . . . . . . . 77
5.2 Shear layer modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2.1 Characteristic frequency . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.2.2 On eigenmode amplification . . . . . . . . . . . . . . . . . . . . . . . 81
6 Experimental campaign 87
6.1 The experimental campaign . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2.1 Validity of eigenmodes in the saturated regime . . . . . . . . . . . . 89
6.2.2 Effects of three-dimensional boundary conditions . . . . . . . . . . . 93
6.2.3 On the symmetry breakings . . . . . . . . . . . . . . . . . . . . . . . 95
CONTENTS xv
6.3 Concluding remarks on the experimental campaign . . . . . . . . . . . . . . 95
7 Preliminary study on saturation 99
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2.2 Computational Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3.1 Regime I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3.2 Regime II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3.3 Regime III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.3.4 Regime IV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.3.5 Regime V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.4 Concluding remarks on the study on saturation . . . . . . . . . . . . . . . . 109
8 Summary and Future Directions 111
xvi CONTENTS
List of Figures
2.1 Transient growth due to non-orthogonal combination of two vectors decay-
ing in time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.1 Schematic description of the two-dimensional open cavity and problem pa-
rameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 Variation with the Reynolds number of the incoming boundary layer thick-
ness at the leading edge of the cavity, θ0/D, for constant boundary layer
thickness at the start of the computational domain θ−1/D = 0.0337. . . . . 37
4.3 Schematic description of the 3-D open cavity and problem parameters. . . 38
4.4 Spanwise velocity component, w, isosurfaces for a stationary eigenmode. . 40
4.5 Spanwise velocity component isosurfaces for a travelling eigenmode for dif-
ferent combinations of α1 and α2. On top, pulsating perturbation. The
others are structures that are right-travelling (middle row) or left-travelling
(bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 Effect of small variation of the flow conditions on the stability. Ampli-
fication vs β, with full symbols corresponding to variations in incoming
boundary layer thickness and empty symbols to variations in ReD. . . . . . 43
4.7 Comparison between Bres and Colonius [20] results (black diamonds) and
present BiGlobal analysis solution (Mode I in red circles, Mode II (bifur-
cated) in blue circles) for the nominal conditions of case A. Amplification
rate as a function of the spanwise length on the left and of the dimensionless
frequency on the right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.8 Amplification versus β for Case A. Mode I in red circles, Mode II (bifur-
cated) in blue circles and Bres and Colonius [20] results in black rhombi. . . 46
xvii
xviii LIST OF FIGURES
4.9 Temporal evolution of the maximum value of the spanwise velocity pertur-
bation obtained by DNS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.10 3D Visualization of spanwise velocity at ReD = 1500 and β = 6: Leading
disturbance obtained using BiGlobal analysis on the left; DNS solution on
the right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.11 3D Visualization of spanwise velocity disturbance obtained at ReD = 2400
using BiGlobal analysis (left) and DNS at t = 400 (right). . . . . . . . . . 49
4.12 Left:variation of the eigenvalues amplification and frequency with β. L/D =
2, Re= 2300, θ−1 = 0.0337, β = 4.4 in red, β = 4.8 in green, β = 5.2 in blue
and β = 5.6 in black. Right: Variation of the curves σ versus β with the
Reynolds number. All cases are for L/D = 2 and θ−1 = 0.0337, Re= 2300
in red, Re= 2400 in green, Re= 2500 in blue and Re= 2600 in black. . . . . 52
4.13 Amplification, σ, versus β and ReD of the three leading modes with positive
amplification in the open cavity flow of aspect ratio L/D = 2 with θ−1 =
0.0337. Mode I in red, Mode II in blue and the third mode in white. . . . . 53
4.14 Neutral stability curves of the three leading modes in the open cavity flow
of aspect ratio L/D = 2 with θ−1 = 0.0337. Mode I in red, Mode II in blue
and the third mode in white. Critical values are cited in table 4.6. . . . . . 54
4.15 Dependence on Reynolds number of the amplification rate, σ, of the leading
eigenmodes in the open cavity flow of aspect ratio L/D = 2 with θ−1 =
0.0337. Mode I in red, Mode II in blue and the third mode in white. . . . 55
4.16 Neutral stability curves of the two leading modes in the open cavity flow of
aspect ratio L/D = 2 with θ−1 = 0.0337 in circles, approximate maxima in
rhombus. Mode I in red, Mode II in blue. . . . . . . . . . . . . . . . . . . . 56
4.17 Neutral curves in ReD vs θ0/D for the critical β. Mode I in red symbols,
Mode II (bifurcated) in blue symbols. L/D = 2 in circles and L/D = 3 in
squares. Highlighted point corresponds with the nose of the most unstable
mode in figure 4.14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.18 Neutral curves for θ−1/D = 0.0337 of Mode I. Aspect ratio varying from
L/D = 1 (higher ReD numbers) to L/D = 3 (lower ReD numbers) with the
values detailed in Table 4.5. . . . . . . . . . . . . . . . . . . . . . . . . . . 58
LIST OF FIGURES xix
4.19 Neutral curves for θ−1/D = 0.0337 of Mode II. Aspect ratio varying from
L/D = 1 (higher ReD numbers) to L/D = 3 (lower ReD numbers) with the
values detailed in Table 4.5. . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.20 Neutral curves for θ−1 = 0.0337. From left to right, and from top to bottom,
L/D = 1.0, L/D = 1.1, L/D = 1.2, L/D = 1.3, L/D = 1.4, L/D = 1.5,
L/D = 2.0 and L/D = 3.0. Mode I in red, Mode II (bifurcated) in blue,
third mode in white and fourth mode in black. . . . . . . . . . . . . . . . . 61
4.21 Neutral curves in the ReD vs L/D for the critical β and θ−1 = 0.0337. First
mode in red, second mode in blue, with circles being low β and rhombi high
β. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.22 From left to right and top to bottom, w velocity component isosurfaces of
Mode I for, L/D = 1.1 (ReD = 3167, β = 12.45), L/D = 1.1 (ReD = 3630,
β = 6.56), L/D = 1.5 (ReD = 1472, β = 5.96), L/D = 2 (ReD = 1150,
β = 5.62) and L/D = 3 (ReD = 865, β = 5.30) respectively. In all cases
the spanwise length shown is Lz = 2π/β = 2D. . . . . . . . . . . . . . . . . 64
4.23 From left to right and top to bottom, w velocity component isosurfaces for
the stationary branch of Mode II for, L/D = 1 (ReD = 3421, β = 13.07),
L/D = 1.3 (ReD = 2069, β = 10.95), L/D = 1.5 (ReD = 1708, β = 10.11),
L/D = 2 (ReD = 1471, β = 9.86) and L/D = 3 (ReD = 1342, β = 10.06)
respectively. In all cases the spanwise length shown is Lz = 2π/β = 2D. . . 65
4.24 From left to right and top to bottom, w velocity component isosurfaces for
the travelling branch of Mode II for, L/D = 1.1 (ReD = 3884, β = 6.63),
L/D = 1.3 (ReD = 2608, β = 5.63), L/D = 1.5 (ReD = 1853, β = 4.84),
L/D = 2 (ReD = 1523, β = 4.45) and L/D = 3 (ReD = 1638, β = 4.61)
respectively. In all cases the spanwise length shown is Lz = 2π/β = 2D. . . 66
4.25 From left to right and top to bottom, w velocity component isosurfaces for
the third mode, L/D = 1.2 (ReD = 3476, β = 11.61), L/D = 2 (ReD =
2207, β = 10.34) and the fourth mode of L/D = 3 (ReD = 1442, β = 2.92)
respectively. In all cases the spanwise length shown is Lz = 2π/β = 2D. . . 67
4.26 From left to right and top to bottom, u velocity component isosurfaces for
Mode I. Time with respect of the final time t/tf = [0; 0.25; 0.5; 0.625; 0.75; 0.875; 0.94; 1]
In all cases the spanwise length shown is Lz = 2π/β = 2D. . . . . . . . . . 68
xx LIST OF FIGURES
4.27 From left to right and top to bottom, v velocity component isosurfaces for
Mode I. Time with respect of the final time t/tf = [0; 0.25; 0.5; 0.625; 0.75; 0.875; 0.94; 1]
In all cases the spanwise length shown is Lz = 2π/β = 2D. . . . . . . . . . 69
4.28 From left to right and top to bottom, w velocity component isosurfaces for
Mode I. Time with respect of the final time t/tf = [0.75; 0.875; 0.94; 1] In
all cases the spanwise length shown is Lz = 2π/β = 2D. . . . . . . . . . . . 70
4.29 Neutral curves in ReL vs θ0/D for the critical β. Mode I in red symbols,
Mode II (bifurcated) in blue symbols. L/D = 2 in circles and L/D = 3 in
squares. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.30 Qualitative representation of the spanwise velocity component of the low β
lobe of Mode I for L/D = 1.2, 1.5, 2 and 3, from top to bottom and left to
right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.31 Dependence with L/D and LP /D of the critical ReD for the critical β (top)
and said β of maximum amplification (bottom). All data with θ−1/D =
0.0337. Mode I in red symbols, Mode II in blue symbols, with circles for
low β and rhombi high β. In lines, the predicted curves of equation 4.4,
and the hyperbolic law described in § 4.3.7, and in small empty symbols, of
same shape and color, the equivalent LP /D of the same points. . . . . . . . 73
4.32 Strouhal number for the main oscillating modes versus the length-to-depth
aspect ratio of the cavity, both in logarithmic scale. Mode I, Mode II and
the third unstable mode are in red circles, blue circles and empty circles
respectively. The other two, in triangles, are stable modes for the range of
parameters studied. In black rhombi the dominant frequency of Bres and
Colonius [20], and the range of frequencies obtained by Basley et al. [16] on
the crossed line. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1 Evolution of the growth rate of the two least stable eigenvalues on the two
dimensional limit, in a range of β ∼ 10−2 to β ∼ 10−9 for L/D = 2,
ReD = 1500 and the constant δ−1. . . . . . . . . . . . . . . . . . . . . . . . 78
5.2 Velocity profiles of the reconstructed 2D flow with the first shear layer mode.
Streamwise velocity on the top, and wall normal velocity on the bottom. . . 79
LIST OF FIGURES xxi
5.3 Variation of the dimensionless frequency (Strouhal based on cavity length
StL) of the two least stable eigenmodes (circles ◦ and rhombus � respec-
tively) in the 2D limit with the incoming boundary-layer thickness. Results
for L/D = 2 in empty symbols (◦), and L/D = 3 in full symbols(•). In
grey shades, the range of Strouhal numbers obtained by Sarohia [90], and
in red symbols single points of several L/D = 2 works. The 2M2 run of
Rowley et al. [87] in squares (�), the Bres [19] M= 0.3 run as a rhombus
( ), the lower runs of Yamouni et al. [112] in triangles (�). Points from
Basley et al. [16] as circles (©). . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4 Variation of the dimensionless frequency (Strouhal based on cavity length
StL) of the two least stable eigenmodes (circles ◦ and rhombus � respec-
tively) in the 2D limit with the length of the cavity dimensionalized with the
incoming boundary-layer thickness, L/θ0. Results for L/D = 2 in empty
symbols (◦), and L/D = 3 in full symbols(•). In in red symbols single
points of several works: the 2M2 run of Rowley et al. [87] in squares (�),
the Bres [19] M= 0.3 run as a rhombus ( ), the lower runs of Yamouni
et al. [112] in triangles (�) and points from Basley et al. [16] as circles (©). 81
5.5 Streamwise velocity profiles of the first shear layer mode, for ReD = 2400
and θ−1/D = 0.0337. Different length domains, from top to bottom xout =
8, 13, 19 and 21 respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.6 Variation of the dimensionless frequency (Strouhal based on cavity length
StL) and amplification σ of the two least stable eigenmodes with the change
of domain length xout. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.7 Neutral curves in ReD vs θ0 for the two least stable two-dimensional modes,
in circles ◦ and rhombus � respectively. Three-dimensional Mode I and
Mode II in red and blue, as in figure 4.17. Domain length kept constant at
xout = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.1 Sketch of the experimental set-up. Dimensions are given in millimetres. The
laser sheet (at y = −0.1D) is represented in a close-up on the L = 2D –
shaped cavity. High resolution images require three cameras to span the
cavity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
xxii LIST OF FIGURES
6.2 Main flow features at ReD = 2400 for both experimental and linear stability analysis.
(Top) BiGlobal unstable eigenvalues (dots) and qualitative schematic depiction of the
most energetic modes in the experiments (shades) in the β-St plane. (Bottom) Velocity
fields related to the four representative modes highlighted in the upper figure. Left
column corresponds to global Fourier modes from the experimental dataset. Right column
presents the reconstructed flow using BiGlobal analysis. For each mode, streamwise
velocity (top) and spanwise velocity (bottom) are shown. . . . . . . . . . . . . . . . 91
6.3 Streamwise profiles of streamwise velocity U/U0 for case A (ReD = 1500)
on the left and case B (ReD = 2400) on the right. The profile obtained
of the 2D base-flow used by BiGlobal analysis is extracted from the range
−0.12 � y/D � −0.09 (black), to represent the uncertainty on the position
and thickness of the laser-sheet. The profile issued of the 3D mean-flow,
experimentally measured in the zx-plane at y = −0.1D is extracted from
the range −3 � z/D � 3 (blue), to take into account spanwise variations. . 94
7.1 Neutral curves for the L/D = 2 cavity in the ReD vs β plane, and selected
Case B (top). StD vs β map of unstable eigenmodes for Case B, and
selected β by the periodicity conditions of the DNS computations (bottom). 101
7.2 Temporal evolution of the absolute value of the span-wise velocity compo-
nent in the control point. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3 DMD modes on regime II. On the top, situation in the StD vs β plane
(left) of DMD modes A and B (right). On the bottom, BiGlobal mode
corresponding with point A (Mode II for β = 12). . . . . . . . . . . . . . . 104
7.4 Two instantaneous flowfields in region III (top). Composition of the two
linear modes that yields a similar flowfield (bottom). . . . . . . . . . . . . 106
7.5 DMD modes on regime III. Situation in the StD vs β plane of DMD modes
A and C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.6 DMD modes on regime IV. Situation in the StD vs β plane of DMD and
modes A, C, E and D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.7 DMD modes on regime V. Situation in the StD vs β plane of DMD and
modes A, C, E and F. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
List of Tables
4.1 Parameters characterising the inflow Blasius boundary layer and associated
location of the primary, (xC , yC), and secondary, (xc, yc), vortices. . . . . . 37
4.2 Parameters for Case A (ReD = 1500). B&C stands for the nominal condi-
tions of Bres & Colonius 08. ED for the Experimental Data of Chapter 6,
with associated uncertainties. LSA for the values of Linear Stability Anal-
ysis. The values of δ∗0 and θ0 are calculated with an equivalent rig to the
one employed in Chapter 6 for B&C and LSA. . . . . . . . . . . . . . . . . 43
4.3 Parameters of the nominal conditions for Case B (ReD = 2400). . . . . . . 44
4.4 Amplification rate (σ) and frequency (StD) of the leading BiGlobal mode
and the linear part of the DNS, for both cases. . . . . . . . . . . . . . . . . 50
4.5 Range of the parameters studied . . . . . . . . . . . . . . . . . . . . . . . . 50
4.6 Critical parameters of the first three modes for the open cavity flow with
aspect ratio L/D = 2 and θ−1 = 0.0337. . . . . . . . . . . . . . . . . . . . . 52
4.7 Critical parameters of the first three modes for each of the aspect ratios
and for constant δ−1/D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.8 The parameters C1 and C2 of equation (4.4) for the different modes(left)
and the parameters K1 and K2 of equation (4.5) for the different modes
(right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
xxiii
1Introduction
Flow over open cavities is of theoretical and practical interest from both a hydrodynamic
and an aeroacoustic point of view. Two-dimensional cavities, i.e. those in which flow in the
lateral spatial direction is considered to be homogeneous, may be encountered in planar
or axisymmetric form, embedded in incompressible or compressible flow over semi-infinite
(open) or confined (closed) domains. Geometric details of a two-dimensional open cavity
configuration, such as its cross-sectional shape, length-to-depth aspect ratio and the form
of either or both cavity lips, as well as the relative dimension of the cavity compared
with a characteristic length scale of the oncoming flow, make description of open cavity
flow non-unique. A finite spanwise extent of the cavity, i.e. three-dimensionality of the
cavity geometry itself, adds yet another dimension to this multi-parametric problem, as
does turbulence, the latter being essential for the description of most industrially relevant
flows.
From a theoretical point of view, progress in the understanding of the complex open
cavity flow dynamics is made by addressing the relatively simple two-dimensional rectangu-
lar cavity configuration at moderate Reynolds numbers. In this context, a steady laminar
two-dimensional boundary-layer on a flat surface encountering an open cavity separates
as a free shear-layer from the upstream cavity corner. Two independent Reynolds num-
bers characterise this flow, one based on the incoming boundary-layer flow properties and
a second based on the cavity dimensions. The dependence on the incoming boundary
layer thickness can be studied as a length parameter, keeping the effect of the velocity
on the problem only on the Reynolds number based on cavity depth. At the relatively
low Reynolds number values of interest in the present thesis framework, at which two-
dimensional flow remains laminar, and depending on the cavity aspect ratio, the free
shear-layer either impinges upon the downstream cavity corner in a steady or unsteady
manner, or curves toward the cavity floor to form a closed recirculation bubble. In ei-
ther case, a new boundary layer forms on the downstream cavity wall, starting from the
downstream cavity corner, which may itself be steady or unsteady.
1
2
Flow instabilities in an open cavity have been investigated in the past from the point
of view of understanding the basic physical mechanisms involved, but also motivated
by the multitude of industrial applications in which this configuration arises, such as
open roofs on automobiles, landing gear in aircraft or weapon bays. Theoretical and
numerical expediency have limited most analyses of open cavity flows to strictly two spatial
dimensions, or three-dimensional domains in which spanwise periodicity is assumed. Fully
non-periodic three-dimensional analysis still has a prohibitive computational cost.
The majority of early theoretical work on the open cavity has revolved around the
two-dimensional flow/acoustic resonance producing self sustained oscillations in the shear
layer (Powell [79], Rockwell [81], Rockwell and Naudascher [84]). In compressible flow,
small amplitude oscillations in the shear layer impinging at the downstream edge of the
cavity generate pressure disturbances that travel upstream and reinforce the shear layer
oscillations. This feedback mechanism results in vortex-shedding at the leading edge, which
locks onto a characteristic frequency that follows the semi-empirical formula described by
Rossiter [86], established for compressible flows.
Subsequent works postulated that in the incompressible regime this feedback can be
considered as instantaneous, and in the experimental work of Sarohia [90] two peaks of
frequency that match the prediction of the Rossiter mechanism were found. Later Gharib
and Roshko [40] observed in their work on axisymmetric cavities that, as the incoming
boundary-layer thickness decreased in relation to the depth of the cavity, there was a
substantial change in behaviour of the cavity oscillations, characterised by a large scale
shedding from the cavity leading edge; the authors termed this finding the wake mode.
Similar transitions have been observed by Pereira and Sousa [76] in two-dimensional in-
compressible Direct Numerical Simulations (DNS) with laminar incoming boundary layer.
Sipp and Lebedev [97] studied the bifurcation of the shear layer modes in the incompress-
ible confined cavity of aspect ratio L/D = 1 on supercritical flow conditions, using weakly
non-linear analysis. Following that work, Barbagallo et al. [10] and Sipp [96] delved into
the closed and open-loop control problem, respectively, at those same conditions.
Still in the incompressible regime, the self-sustained oscillations follow the primarily
two-dimensional geometry of the shear layer. In that regime, Colonius et al. [25] showed the
first two-dimensional amplitude function in a two-dimensional open cavity, demonstrating
a Tollmien−Schlichting wave emanating on the downstream wall and being connected with
instability inside the open cavity (see also Supplemental Appendix 4 of Theofilis [102]).
The work of Rowley et al. [87] provided further insight regarding the onset of shear-layer
oscillations from a steady flow and their nonlinear interactions, developing a criterion to
1 Introduction 3
predict the onset of disturbances from the steady flow. Later, Theofilis and Colonius
[103, 104] revisited the open cavity and, applying the residuals algorithm (Gomez et al.
[42], Gomez et al. [43], Theofilis [100]) recovered the same Tollmien−Schlichting eigenmode
as well as an acoustic mode connecting hydrodynamic perturbations inside the cavity
with pressure fluctuations having their origin at the downstream cavity corner. Recently,
Yamouni et al. [112] have given a detailed description of the effect of compressibility on the
interaction of the acoustic feedback and resonance using global instability analysis in a two-
dimensional cavity at a particular set of flow conditions and one Reynolds number. The
authors confirmed that the shear layer modes correspond to the beginning of branches of
those global modes whose frequencies evolve with Mach number as described by Rossiter’s
semi-empirical formula, as is supported by previous findings by Sarohia [90].
The thorough study of the two-dimensional self sustained oscillations led to the dis-
covery of three dimensional structures in the flow, with frequencies far smaller than the
ones of the Rossiter modes. Several studies reported these lower frequencies, mainly as a
modulation of the two-dimensional shear layer modes (Koseff and Street [62], Neary and
Stephanoff [71], Rockwell and Knisely [82]). The hypothesis that these modes were the
result of non-linear interaction between Rossiter modes was rebutted by Cattafesta III
et al. [22] and Kegerise et al. [57].
The aforementioned work by Theofilis and Colonius [104] also contained some of the
first linear computations of three dimensional instabilities inside the cavity. The authors
used a two dimensional base flow multidomain computation, and applying the appropriate
boundary conditions to the upper side on the cavity subdomain they showed a structure
consisting of a perturbation that coils around the main recirculating vortex. That mon-
odomain approach in the eigenvalue problem solution requires a numerical simplification
on the boundary condition, and has the problem of being unable to reproduce the interac-
tions of the cavity flow with the shear layer and the external flow. The first non-conforming
spectral multidomain approach for the numerical solution for the two-dimensional BiGlobal
eigenvalue problem, in this case for the lid driven cavity was presented in de Vicente et al.
[28], paving the road for future multidomain approaches in the open cavity.
The first full three-dimensional global instability analysis of compressible flow over
a rectangular open cavity was performed by Bres and Colonius [20]. In said work the
authors establish that the low frequency mode corresponds to three-dimensional struc-
tures associated to the main recirculation vortex inside the cavity, and they found that
these three-dimensional perturbations were dominant for some flow conditions, and that
they are independent from the shear layer modes. These three-dimensional structures are
4
sometimes known in the literature as centrifugal modes, and may qualitatively be related
to instability modes seen in the lid-driven cavity (Albensoeder et al. [4], Gonzalez et al.
[46], Theofilis [99]). Bres and Colonius [20] also confirmed the previously observed modu-
lation of the two-dimensional shear layer modes by the lower frequencies of the centrifugal
modes. Further discussion on the nature of the interaction between the two types of modes
can be found in Basley et al. [17]
It is interesting to note that such spanwise dynamics arise in the recirculating flow
in shear-driven as well as lid-driven cavities (Albensoeder et al. [5], de Vicente et al.
[29], Gonzalez et al. [46]), that is, regardless of the shear-layer oscillations.
There also exist in the literature numerous works on flow visualisation and charac-
terisation of these three-dimensional open cavity instabilities. Recently, Faure et al. [36]
aimed to understand the three-dimensional flow morphology, and characterise experimen-
tally the three-dimensional structures in the open cavity, confirming that the dynamical
structures are independent of shear layer instabilities. Basley et al. [15] used time-resolved
particle image velocimetry (PIV) to extract the spatial distribution of the most charac-
teristic frequencies in the incompressible open cavity with two different aspect ratios, and
also identified once again the presence of the aforementioned shear layer modes in the
incompressible regime. Alizard et al. [6] focused on the global instability analysis of open
flows using a domain decomposition matrix-free method. One of the benchmark problems
the latter authors employed is the square cavity, where, at a single Reynolds number, the
three-dimensional spanwise periodic leading perturbations were identified.
Despite these previous works a complete study on three-dimensional incompressible
cavity flow is still missing. Recently, the work of de Vicente et al. [31] presented a combined
theoretical, numerical and experimental work on global instability of the open cavity in
incompressible regime. It also includes validation and verification of the instability analysis
results, the focus being on the three-dimensional structures, dominant in the studied flow
regime. This work is partially presented here, as well as the totality of Meseguer-Garrido
et al. [66]. The aim of those publications is to present an in-depth extension of the
theoretical results presented in de Vicente [27], revisiting incompressible flow over the
spanwise periodic rectangular cavity. A systematic global (BiGlobal, using the terminology
of Theofilis et al. [105]) instability analysis was performed, taking into account the effect
of parameters that, no doubt on account of the cost of this type of analysis, have either
been studied incompletely or been altogether ignored in the literature. The incoming
boundary-layer thickness, the Reynolds number based on cavity depth and the spanwise
wavenumber are varied independently, for a range of length-to-depth aspect ratios, until
1 Introduction 5
the most interesting part of the spectrum containing all amplified global eigenmodes is
revealed. Also, these studies aim to bring more insights into the early stages of the three-
dimensional dynamics associated with centrifugal effects around the main recirculation
inside the cavity, in a flow regime in which the shear-layer modes are not dominant.
The unstable three-dimensional perturbations are identified and characterised from both
numerical and experimental approaches.
Moreover, the fruitful combination of linear analysis and experiments leads to a deeper
knowledge of the characteristics and evolution of these modes from their onset to their
final presence in the nonlinear saturated state. In order to delve more deeply on that
evolution from the linear regime to the saturated state, three-dimensional direct numerical
simulations were performed, and analyzed using a dynamic mode decomposition technique,
as presented in Vinha et al. [110].
Data-sequences of snapshots collected from numerical simulations or experiments can
be used to approximate the inherent fluid flow into dynamic modes, allowing then the
identification of the relevant coherent structures in the flow. The most commonly used
data-based techniques so far are the Fourier Transform analysis, the Proper Orthogonal
Decomposition (POD) and, more recently, the Dynamic Mode Decomposition (DMD).
The first approach is particularly efficient when dealing with periodic sampled data-fields.
However, it loses accuracy when dealing with more complex and time-dependent fluid
flows. With the use of POD one can identify the relevant structures in the flow ranked
by their energy content but, since the different POD modes are orthogonal in space, their
temporal behavior is characterized by the presence of multiple frequencies. For a detailed
discussion about the use of Fourier Transform analysis, POD, and alternative data-based
decomposition techniques on the identification of coherent structure in the fluid flow, the
reader is directed to Mezic [67] and Bagheri [8].
The DMD allows the extraction of spatial modal structures from a flow field, where
each identified dynamic mode is associated to a single and unique frequency, consequence of
the orthogonalization in time of the decomposition. The present technique is based on the
Koopman analysis of a dynamical system by Rowley et al. [88], aiming to approximate the
Koopman modes and eigenvalues of a linear infinite dimensional operator that describes
that system, even if its dynamic behavior is nonlinear. In this case, DMD retrieves the
structures of a linear tangent approximation to the underlying fluid flow (Schmid [92]).
Nonetheless, for a system behaving linearly, the extracted DMD modes are expected to
match the global stability modes. Moreover Rowley et al. [88] analytically demonstrated
that performing a DMD over periodic solutions leads to the same decomposition obtained
6
via a discrete temporal Fourier Transform. Contrary to the POD, the DMD does not rank
the extracted coherent structures in terms of energy content. However, their amplitudes
provide a feedback about the individual contribution of a specific mode to the original
data-set (Schmid [93]), providing the DMD the possibility of obtaining models of lower
complexity (Jovanovic et al. [55]), as it already happens with the POD.
The DMD is a recent decomposition technique that is becoming more and more attrac-
tive in post-processing of numerical and experimental results, mainly due to its easiness of
implementation, efficient data analysis, inherent low computational cost, and possibility
of application to large data-sets or to sub-domains of a flow region. Some implementations
of this tool for cavity problems can be found in the literature, for example in Gomez et al.
[42] or in Ferrer et al. [37]. This method has also demonstrated superior performance over
other traditional data-based decomposition techniques for oscillatory dominated problems
(Schmid [93]) and for fluid flows presenting strong peaks in the spectrum (Mezic [67]).
Nonetheless, DMD still has some relevant limitations, as recognized in Bagheri [8]. Ac-
cording to this reference, there is yet no validation between Koopman and DMD modes
for chaotic and noisy high-Reynolds number flow-data and, based on the work of Duke
et al. [33], the Dynamic Decomposition can be sensitive to the presence of noise in the
data-field. Furthermore, Muld et al. [70] observed no particular differences between the
POD and the DMD modes in a flow characterized by a broad frequency spectrum and no
dominant spectral peaks. Finally, the DMD may not guarantee the best approximation of
the flow field, opening new ways for improved variants of the original algorithm based on
optimization techniques, like the ones proposed by Jovanovic et al. [55] and Chen et al.
[23].
Regarding the nature of the BiGlobal analysis, from a numerical point of view solutions
have been obtained in the last decades using a variety of approaches. From the point of
view of solution methodology, straightforward serial (see Theofilis [101] for a review) or
parallel (Rodrıguez and Theofilis [85]) implementations of a subspace iteration variant, can
be considered, or employment of the widely used (Bagheri et al. [9], Bres and Colonius
[20], Crouch et al. [26], Fietier and Deville [38], Janke and Balakumar [54], Wintergerste
and Kleiser [111]) implementation of the Implicitly Restarted Arnoldi Method (IRAM)
in the ARPACK (Lehoucq et al. [64]) software package for the solution of large scale
eigenvalue problems. On the other hand, from the point of view of matrix formation,
existing methodologies for the solution of the BiGlobal eigenvalue problem can be classified
in two categories, one in which the matrix is formed, stored and processed using dense
(serial or parallel) linear algebra technology, and another known as matrix-free/time-
1 Introduction 7
stepper algorithms (Abdessemed et al. [2], Bagheri et al. [9], Barkley et al. [13], Sherwin
and Blackburn [95]).
Regarding spatial discretization methods of the BiGlobal eigenvalue problem, early ap-
plications studied simple two-dimensional domains in which the numerical discretization
techniques employed were straightforward extensions of those used in the solution of classic
one-dimensional linear stability eigenvalue problems, using spectral methods (Pierrehum-
bert [77], Tatsumi and Yoshimura [98], Zebib [113]). At the same time, finite-element
methods were also used for the solution of the BiGlobal eigenvalue problem by Jackson
[52] and Morzynksi and Thiele [69]. Although the low order of those methods limits the
accuracy of the solution, they are not restricted to the single-domain two-dimensional
grids employed in the early spectral analyses. This lack of accuracy can be critical in the
case or problems with sharp gradients, as is the case of the BiGlobal eigenmodes at high
Reynolds numbers, leading to the use of unstructured meshes of ever-increasing density in
order to achieve convergence (Gonzalez et al. [44]), trading off the efficiency of a high-order
method in favor of the flexibility offered by the unstructured mesh discretization.
The use of high-order accurate, flexible and efficient numerical methods in order to
solve the BiGlobal EVP is needed. Such an approach has been introduced in the seminal
work of three-dimensional instability in the wake of a circular cylinder by Barkley and Hen-
derson [12], Henderson and Barkley [49] in the form of spectral-element discretization on
structured meshes. The first application of a spectral/hp−element method (Karniadakis
and Sherwin [56]) to the study of a global instability problem on unstructured meshes was
that of Theofilis et al. [106], who recovered instability in the wake of a NACA0012 airfoil
as the leading BiGlobal eigenmode of the steady wake flow. While the aforementioned
spectral/hp−element analyses utilized time-stepping concepts (Tuckerman and Barkley
[109]), Gonzalez et al. [45] have discussed matrix formation and storage as an alternative
technique for the solution of the same problem in the context of spectral/hp−element
discretization. The method here used was explained in depth in de Vicente [27], as the
codes and numerical tools developed there are the basis for all the stability computations
presented here.
8
The organization of the present thesis is as follows:
Chapter 2 contains a brief explanation of the linear instability theory, focused on
the temporal stability problems here analyzed. The nature of the BiGlobal instability
analysis is explained, resulting on the analysis of spanwise periodic three-dimensional
global disturbances that evolve over a two-dimensional steady base flow.
Chapter 3 is a compendium of the numerical methods employed. Section 3.2 explains
the two-dimensional spectral element method used to integrate the laminar Navier-Stokes
equations in order to obtain the base flows, Section 3.3 details the BiGlobal eigenvalue
problem solver, Section 3.4 the three-dimensional dicontinuous galerking spectral element
code, and Section 3.5 the dynamic mode decomposition tool.
Chapter 4 describes the problem at hand and puts in context previous findings. First,
Section 4.2 contains the extensive validation of the tool with the literature, and with
the three-dimensional DNS code. Results of parametric analyses are shown in Section 4.3,
where the critical Reynolds numbers and the leading perturbations at all conditions studied
are identified. A synthesis of three-dimensional global instability analysis results obtained
herein provides analytical expressions for the relation between critical Reynolds number
and both incoming boundary-layer thickness and cavity aspect ratio, which are given in
Section 4.3.7. The connection between the characteristic frequency of the leading global
eigenmodes and the cavity aspect ratio is revealed from these results.
Chapter 5 contains the results regarding the two-dimensional instabilities. Although
the shear layer modes are not the focus of this thesis, some interesting results are shown,
as well as the concordance of the results with those present in the literature.
Chapter 6 explains the main results of the experimental campaign presented in de
Vicente et al. [31]. Results from both linear stability analysis and experiments enable the
onset of centrifugal instabilities inside the cavity to be related to the final saturated state
of the flow. To that end, Section 6.2 discusses the relationship between global Fourier
modes from the experiments with respect to global eigenmodes. Section 6.3 summarizes
the most significant conclusions obtained.
Chapter 7 is a preliminary study on the saturation phenomena, using the three-
dimensional DNS computations, and comparing the DMD analysis of the snapshots ob-
tained with the linear analysis results.
1 Introduction 9
Most of the material in the present thesis, as well as additional specific details have
been presented in the following list of publications:
• Journal Articles
◦ de Vicente, J. & Basley, J. & Meseguer-Garrido, F. & Soria, J. & Theofilis,
V. “Three-dimensional instabilities over a rectangular open cavity: from linear
stability analysis to experimentation”, J. Fluid Mech. (2014), vol. 748, pp.
189-220.
◦ Meseguer-Garrido, F. & de Vicente, J. & Valero, E. & Theofilis, V. “On lin-
ear instability mechanisms in incompressible open cavity flow”, J. Fluid Mech.
(2014), vol. 752, pp. 219-236.
◦ Vinha, N. & Meseguer-Garrido, F. & de Vicente, J. & Valero, E. “A Dynamic
Mode Decomposition of the saturation process in the open cavity flow”, sub-
mitted to Special Issue of Aerospace Science and Technology on ”Massively
Separated Flows”.
• Conference Papers
◦ Gonzalez L. & Theofilis, V. & Meseguer-Garrido, F. “Applications of High
Order Methods to Vortex Instability Calculations”, International Conference on
Spectral and High Order Methods (ICOSAHOM), 22-26 June 2009, Trondheim,
Norway.
◦ Meseguer-Garrido, F. & de Vicente, J. & Valero, E. & Theofilis, V. “Effect of
Aspect Ratio on the Three-Dimensional Global Instability Analysis of Incom-
pressible Open Cavity Flows”, 6th AIAA Theoretical Fluid Mechanics Confer-
ence, 26-30 June 2011, Honolulu, Hawaii.
◦ Meseguer-Garrido, F. & de Vicente, J. & Valero, E. “Three-dimensional analysis
of incompressible flow over an open cavity using direct numerical simulation:
from linear to saturated regime”,Instability and Control of Massively Separated
Flows, 5980, June 5-8 2013, Prato, Italy.
2Hydrodynamic Stability
Contents
2.1 Linear Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Non-Modal Instability . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Modal Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2.1 Local Instability . . . . . . . . . . . . . . . . . . . . . . 16
2.1.2.2 Non-Local Instability . . . . . . . . . . . . . . . . . . . 17
2.2 BiGlobal Linear Theory . . . . . . . . . . . . . . . . . . . . . . . 17
2.1 Linear Instability
Linear stability theory concerns the evolution of small disturbances superposed on a
steady- or time dependent basic state. A solution, called “base flow”, is said to be stable
if an infinitely small variation will alter only by an infinitely small quantity the basic flow
at some future time.
||U(x, 0)− U∗(x, 0)|| < δ → ||U(x, t)− U∗(x, t)|| < ε (2.1)
Otherwise the solution is said to be unstable. The solution is said asymptotically stable
if it converges to the original solution when it is perturbed:
||U(x, t)− U∗(x, t)|| → 0 ,as, t → ∞ (2.2)
Linear stability analysis in the framework of the solutions of the incompressible Navier-
Stokes equations is related with the decomposition of any flow quantity q in a basic steady
flow q plus a disturbance component q.
q = q+ εq (2.3)
where the flow state is defined by the velocity components in each direction and pressure
q(x, t) = (u(x, t), P (x, t)), with u(x, t) = (u(x, t), v(x, t), w(x, t)) and x = (x, y, z) is the
11
12 2.1 Linear Instability
vector comprising the spatial coordinates. The steady and perturbed states are defined the
same way. Introducing this decomposition into the dimensionless incompressible Navier-
Stokes equations:
∂u
∂t+ (u · ∇)u =
1
Re∇2u−∇P (2.4)
∇ · u = 0
yields,
∂u
∂t+
∂u
∂t+(u ·∇)u+(u ·∇)u+(u ·∇)u+(u ·∇)u =
1
Re∇2u+
1
Re∇2u−∇P −∇P (2.5)
∇ · u+∇ · u = 0
Stability concerns the evolution of small disturbances. The initial decomposition (2.3)
has been fed into the Navier-Stokes equations. O(1) terms, corresponding to the basic
flow, have been canceled while O(ε2) term, the non-linear (u · ∇)u, is neglected. The
perturbed problem is stated only in O(ε) terms.
This gives the linearized problem:
∂u
∂t+ (u · ∇)u+ (u · ∇)u =
1
Re∇2u−∇P (2.6)
∇ · u = 0
From this point, two different analyses derive. Non-modal analysis derives from re-
writing (2.6) as:∂q
∂t= LNSE (Re, q) q (2.7)
where LNSE (Re, q) is the linear operator depending on the Reynolds number and the
base flow. Stability is governed by the properties of the operator. On the other hand
modal analysis, based on classical linear theory, establishes that if the basic flow is steady,
the coefficients of the linearized problem are independent of t, therefore by separation of
variables, the general solution is a linear superposition of normal modes, each of the form:
q(x, t) = q(x)exp(λt) (2.8)
2 Hydrodynamic Stability 13
2.1.1 Non-Modal Instability
The non-modal approach takes the so called state space equation (2.6) as a starting point.
The linearized Navier-Stokes equations can be written as,
dq
dt= LNSE (Re, q) q (2.9)
where LNSE is the former linear operator. This operator is non-normal (it does not
commute with its adjoint) and consequently it is not necessarily diagonalizable. This
implies that the solution, total sum of modes, can experience an initial non-linear growth
(transient growth), although every individual mode decay with time.
This growth in the early stages could provoke that the energy of the disturbance
increases, eventually reaching a critical value leading to transition. The theoretically
predicted modal decay may be overruled depending on the initial perturbation. Linked
to this is the concept of optimal disturbances, i.e., which initial state cause the maximum
energy growth within a time period (Butler and Farrell [21], Farrell and Ioannou [34],
Schmid and Henningson [91], Trefethen et al. [108]).
This linear behavior can be explained from a numerical point of view. Once the linear
operator has been spatially discretized (in this work using spectral collocation) the system
(2.9) behaves like a linear system with constant coefficients in time where solution can be
expressed as
q = exp(A(Re, q)t)q0 (2.10)
where A is the m-dimensional matrix discretizing the spatial differential operator in the
m-points of the computational domain. Stability of the solution (2.10) only depends on
the fundamental matrix Φ(t) = exp(A(Re, q)t), particularly on the eigenvalues of matrix
A(Re, q).
In the general case of n different eigenvalues λi each of them with algebraic multiplicity
ki, associated to each eigenvalue it is possible to find a set of linearly independent solutions
of the kind
Qij = exp(λit)
j−1∑l=0
tl
l!vj−l j = 1, . . . , ki. (2.11)
where vj is a sequence of generalized eigenvectors for the eigenvalue λi.
The set Qij i = 1, . . . n conforms a system of fundamental solutions associated to the
basis of generalized eigenfunctions. These fundamental solutions are the columns of the
basis of the fundamental matrix Φ(t). So each solution of (2.10) can be expressed like
14 2.1 Linear Instability
Figure 2.1: Transient growth due to non-orthogonal combination of two vectors decaying in time.
q =
n∑i=1
ki∑j=1
Qijq0(l) l = 1, . . . ,m. (2.12)
The most significant implication comes from the fact that every solution is formed
by two parts: one exponential and other polynomial. The exponential factor is clearly
dominant when time becomes large, however initially when t is small enough, polynomial
behavior could dominate, and, depending on the initial solution q0, perturbations could
grow in a nonlinear way.
Other source of initial linear growth comes from the non-orthogonality of the eigenfunc-
tions of (2.9) even when LNSE is diagonalizable and all the eigenvalues have negative real
part. Solutions of (2.10) can be expressed as a linear combination of the eigenmodes:
q =
n∑i=1
q0(i)exp(λit) (2.13)
Although the eigenmodes are characterized by decreasing length, due to the nature of
their associated eigenvalues, the solution vector may grow in the early stages as seen in
2.1 taken from Schmid and Henningson [91].
2.1.2 Modal Instability
This work is focused in the modal study of instability of open cavity flows. As mentioned
this analysis is based on the decomposition of the flow quantities in two parts, (i) the
steady or time periodic basic state upon which (ii) small disturbances are permitted to
develop.
2 Hydrodynamic Stability 15
In the particular case of time periodic basic state a similar assumption can be made
using Floquet theory, which seeks the eigenvalues of a time periodic operator. These eigen-
values, also known as Floquet multipliers, determine the development of small-amplitude
perturbations during one period of evolution. In this work only steady states have been
considered. To deal with time periodic basic states in the instability analysis of incom-
pressible flows Barkley and Henderson [12] or more recently Karniadakis and Sherwin [56]
and Abdessemed [1] works could shed light over this problem.
The eigenvalues and their corresponding eigenmodes can be found by substituting (2.8)
into (2.6). That leads to a generalized eigenvalue problem (EVP) of the form
Aq = λBq (2.14)
Numerical discretization of the spatial directions and imposition of the appropriate
boundary conditions complete the matrix formulation. Using this approach, the stability
of the basic flow is determined by the eigenvalues of (2.14). If Re(λ) < 0 for all the
eigenvalues the basic flow is asymptotically stable and unstable if Re(λ) > 0 for at least
one eigenvalue.
Different stability analysis derive from the dimensional decomposition of the fluid vari-
ables in the two mentioned states. In it most general form separability of time and spatial
directions in the governing equations leads to the following decomposition of fluid vari-
ables:
q(x, t) = q(x) + εq(x, t) (2.15)
where q = (u, P )T = (u, v, w, P )T , q = (u, P )T = (u, v, w, P )T and x=(x,y,z). Once
the decomposition (2.15) has been substituted into the governing equations one may write
q(x, t) = q(x)exp(−i ωt) (2.16)
This stability analysis, where physical space is three-dimensional and the time-periodic
disturbances are inhomogeneous in all the three directions, is named TriGlobal linear
instability analysis (Theofilis [101]). However the three-dimensional eigenvalue problem
resulting from this decomposition is extremely expensive, and presently only within the
possibilities of the most advanced supercomputing facilities at relevant Reynolds numbers.
16 2.1 Linear Instability
A wide range of simplifications to the flow instability problem are possible in order
to make it solvable. From the simplest parallel-flow assumption to the BiGlobal insta-
bility analysis, main tool of this work, complexity increases for both the basic flow and
disturbances.
2.1.2.1 Local Instability
The greatest level of simplification comes from considering homogeneous basic flows in
the x and z directions, i.e. neglecting the dependence of the basic flow q on x and
z. Consequently velocity component v is also neglected. These assumptions, known as
parallel-flow assumptions, permit the decomposition
q(x, y, z, t) = q(y) + εq(y)exp{i (αx+ βz − ωt)} (2.17)
where α and β are the wavenumbers in the spatial x and z directions, and ω stands for
the frequency in the temporal instability analysis framework. Gaster’s relation permits
the transformation between the temporal and spatial problems (Gaster [39]).
The substitution into the incompressible flow equations leads to a system of equations
for the amplitude functions of the disturbance velocity components and the disturbance
pressure. Pressure can be eliminated resulting an equation for the normal velocity. To
complete the flow field description another equation is included for the normal vorticity.[1
Re
(d2
dy2− (
α2 + β2))2
− iαu
(d2
dy2− (
α2 + β2))
+ iαd2u
dy2
]v =
(d2
dy2− α2
)ωv
(2.18)[iαu− 1
Re
(d2
dy2− (
α2 + β2))2
]η + iβ
du
dyv = iωη (2.19)
with the boundary conditions v = dvdy = η = 0 at solid walls and in the free stream, where
η =∂u
∂z− ∂w
∂x(2.20)
is the normal vorticity.
Equation (2.18) for the normal velocity is the well known Orr-Sommerfed equation
(Drazin and Reid [32]). In the adopted formulation the equation constitutes an one-
dimensional eigenvalue problem to recover the frequencies, eigenvalues of the temporal sta-
bility problem, and the associated eigenmodes. Meanwhile (2.19), is the Squire equation,
an eigenvalue problem for normal vorticity fed with the solutions of the Orr-Sommerfeld
equation.
2 Hydrodynamic Stability 17
2.1.2.2 Non-Local Instability
The Orr-Sommerfeld equation provides the exact eigenmodes for some flows like pressure
driven Poiseuille channel flow or the shear driven Couette flow. Nevertheless, it is now
known that the predictions of Orr-Sommerfed equation in a variety of flows, like Poiseuille
and Couette flows are wrong; for physically relevant results to be obtained in such prob-
lems, the non-modal approach discussed earlier must be adopted.
Quasi-parallel flows like Blasius boundary layers requires the modeling of the distur-
bance quantities being spatially inhomogeneous to recover proper solutions.
To deal with non-parallel boundary layers flows, an extension of the parallel theory
is the Parabolized Stability Equation (PSE) (Bertolotti et al. [18], Herbert [50]). In this
analysis, analogously to the parallel classic linear theory, one spatial direction is resolved.
By contrast, the basic flow is permitted to grow in one or both remaining spatial directions.
The flow decomposition yields
q(x, y, z, t) = q(y) + εq(y)exp(i (
∫ x
x0
α(ξ)dξ + βz − ωt)) (2.21)
When the ansatz (2.21) is replaced in the Navier Stokes and continuity equation a
system of equations for the disturbance is set. A normalization condition is needed for
closure in order to ensure that the streamwise variation of the amplitude function q remains
small. Further information on this kind of instability analysis can be found in Herbert
[51].
2.2 BiGlobal Linear Theory
Between the simplified parallel-flow assumption and the extremely costly TriGlobal insta-
bility analysis, BiGlobal instability analysis represents a intermediate and affordable step
in the instability analysis. The main difference with the parallel flow assumption is that,
here, three-dimensional space comprises an inhomogeneous two-dimensional domain which
is extended periodically in z characterized by a wavelength Lz. The BiGlobal framework
involves the substitution of a decomposition of any of the independent flow variables, e.g.
the three velocity components, temperature and pressure q(x, y, z, t) = (u, v, w, θ, P )T into
the (incompressible or compressible) equations of motion. All quantities are considered
to be composed of an O(1) steady two-dimensional basic state and small-amplitude O(ε)
unsteady three-dimensional perturbations
18 2.2 BiGlobal Linear Theory
q(x, y, z, t) = q(x, y) + εq(x, y, z, t) (2.22)
The separability of temporal and spatial derivatives in (2.6) permits introduction of an
explicit temporal dependence of the disturbance quantities into these equations. Spatially,
disturbances q(x, y, z, t) are considered three dimensional but only two directions are in-
homogeneous (x, y), while in the third spanwise direction z, disturbances are permitted
to assume a harmonic expansion. These assumptions lead to the ansatz:
q(x, y, z, t) = q(x, y)ei (βz−ωt) (2.23)
for the determination of the complex eigenvalue:
ω = 2πStDU∞
D+ iσ. (2.24)
In the temporal framework used, the Strouhal number (St) represents a dimension-
less frequency and σ is the amplification/damping rate of the disturbance sought, while
barred and hatted quantities denote basic and disturbance flow quantities, respectively.
The two-dimensional domain is extended periodically in the homogeneous z-direction with
the periodicity length Lz = 2π/β associated with the real wavenumber β. The complex
eigenvalues and the corresponding eigenmodes q(x, y), are recovered when the decompo-
sition 2.23 is substituted in the linearized equations 2.6:
[L − (Dxu)] u− (Dyu) v −DxP = −i ωu, (2.25)
− (Dxv) u+ [L − (Dyv)] v −DyP = −i ωv, (2.26)
− (Dxw) u− (Dyw) v + Lw − i βP = −i ωw, (2.27)
Dxu+Dyv + i βw = 0. (2.28)
where Dx = ∂/∂x, Dy = ∂/∂y and
L =1
Re
(∂2
∂x2+
∂2
∂y2− β2
)− u
∂
∂x− v
∂
∂y− iβw. (2.29)
An important simplification comes from the absence of the z-component in the velocity
profile of the basic flow. Homogeneity in the spanwise direction (β real) and redefinitions:
i w → w i ω → ω (2.30)
2 Hydrodynamic Stability 19
result in the following generalized real nonsymmetric partial derivative eigenvalue prob-
lem:
[L∈ − (Dxu)] u− (Dyu) v −DxP = ωu, (2.31)
− (Dxv) u+ [L∈ − (Dyv)] v −DyP = ωv, (2.32)
L∈w − βP = ωw, (2.33)
Dxu+Dyv − βw = 0. (2.34)
where
L∈ =1
ReD
(∂2
∂x2+
∂2
∂y2− β2
)− u
∂
∂x− v
∂
∂y. (2.35)
So the EVP (2.14) can be re-written in the BiGlobal instability analysis framework as
Aq = ωBq (2.36)
There are some implications due to this formulation. The first one, mainly numerical,
is the reduction in the storage requirements of the real EVP compared with the original
complex problem. Halving the storage required for the spatial discretization results in
the ability to address flow instability at higher resolutions, what is essential in case of
higher Reynolds numbers or strong gradients in the flow. The second one is merely an
interpretative issue. Real part of the eigenvalue ω corresponds to the growth rate, σ, while
the imaginary part is related to the oscillation frequency. If the growth rate is positive the
amplitude of the eigenmode will grow exponentially in time and the basic flow is unstable.
The problem is completed with the imposition of boundary conditions for the distur-
bance quantities. At solid walls no-slip boundary conditions are imposed for the velocity
components while homogeneous Neumann condition has been found to be the best for
the disturbance pressure. Boundary conditions for the disturbance pressure do not exist
physically, but compatibility is ensured with the choice of normal first derivative equal
to zero. The imposition of appropriate boundary conditions for open boundaries is not a
straightforward problem, especially when flow leaves the computational domain. In Chap-
ter 3, devoted to the numerical methods the election of appropriate boundary conditions
will be discussed.
3Numerical Methods
Contents
3.1 Short Review of Spectral Methods . . . . . . . . . . . . . . . . . 21
3.2 Base Flow Calculations . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 BiGlobal instability analysis . . . . . . . . . . . . . . . . . . . . 28
3.4 Three-dimensional Direct Numerical Simulation . . . . . . . . . 29
3.5 Dynamic Mode Decomposition . . . . . . . . . . . . . . . . . . . 32
The main objective in this Chapter is to detail the different numerical tools employed
to obtain the results presented in subsequent chapters. In order to do that a tool is needed
to obtain approximated numerical solutions for the instability analysis of the Navier Stokes
equations in a cavity. To perform this analysis an eigenvalue problem (EVP) has to be
solved. The EVP requires a prior steady solution of the two-dimensional Navier Stokes
equations (Base flow DNS). To find accurate solutions of these problems, the spatial
integration is performed via a spectral Chebyshev multi-domain collocation method. In
order to validate and compare the results of the BiGlobal analysis a three-dimensional
direct numerical simulation (3D DNS) tool was used. Also, to process the data provided
by the three-dimensional DNS a Dynamic Mode Decomposition (DMD) tool was used.
3.1 Short Review of Spectral Methods
Spectral methods are members of the broader category of weighted Residual Methods.
The main distinguishing feature of these methods is the selection of mutually orthogonal
global functions as basis for the series expansion of the solution of the differential equation.
Spectral methods have been extensively used in computational fluid dynamics due to
their accuracy and efficiency in the simulation of fluid flows. These features come from
the mathematical properties of orthogonal basis in a Hilbert space. In said space, an
orthogonal basis is a numerable set of vectors, mutually orthogonal, that are complete in
this space. In the particular case of the Hilbert space L2w([a, b]), set of square integrable
21
22 3.1 Short Review of Spectral Methods
functions in [a, b], completeness of the basis means that any function f(x) ∈ L2w([a, b]) can
be approximated as closed as desired using an expansion in terms of that basis.
f(x) =
∞∑k=1
akϕk(x), (3.1)
where the coefficients ak are obtained projecting f(x) over the elements in the basis. The
projection is defined using an inner product,
(u(x), v(x))w =
∫ b
au(x)v(x)w(x)dx, (3.2)
with u(x), v(x) ∈ L2w([a, b]) and w(x) a weight function. The choice of the weight function
and the characteristics of the mathematical or physical problems determine the orthogonal
basis used.
The most common family of orthogonal polynomials used in spectral methods is the
set of trigonometric functions,
ϕk(x) = eikx, (3.3)
also called Fourier basis. The expansion in Fourier series is the most usual for the ap-
proximation of periodic functions. The fast decay of the k-th coefficient, (faster than any
inverse power of k for smooth function with all its derivatives periodic) implies that the
Fourier series represents the best approximation of the function. This feature is referred as
“spectral accuracy” of the Fourier method. However, the characteristic oscillatory behav-
ior of the Fourier expansion of a function in the neighborhood of a point of discontinuity
(Gibbs phenomenon), makes this approximation non-profitable for non-periodic problems.
In order to avoid this restriction in the boundary the use of some other expansion series
is required.
To obtain other families of orthogonal basis functions, two processes can be used:
First one is the Gram-Schmidt orthogonalization procedure. When this process is
applied to the polynomial system {1, x, x2, x3, ...}, depending on the domain [a, b] over
which the problem is defined and the weight w(x) used to characterize the inner product
(3.2) in L2w([a, b]), one can find:
• Jacobi polynomials. Set of orthogonal polynomials in [−1, 1]. These family corre-
sponds to the choice w(x) = (1− x)α(1 + x)β
J(α,β)k (x) =
1
2k
k∑i=0
(k + α
i
)(k + β
k − i
)(x− 1)i(x+ 1)k−i (3.4)
3 Numerical Methods 23
In this broad class of orthogonal polynomials are included the Legendre polynomials,
when α = β = 0, and the Chebyshev polynomials corresponding to α = β = −1/2.
• Laguerre polynomials. Family of orthogonal polynomials in [0,∞), with the weight
function w(x) = xαe−x
• Hermite polynomials. The domain of definition for this set of functions is (−∞,∞),
while the weights are w(x) = 1
Gram-Schmidt procedure results in a set of orthogonal functions that could be used
as the orthogonal basis for the expansion of any function in L2w([a, b]).
The other approach to systems of orthogonal polynomials comes from fact that each
of the aforementioned families of orthogonal polynomials in [−1, 1] are solution of the
second-order linear differential Sturm-Liouville problem.
{−(pφ′)′ + qφ = λwφ in the interval (1,−1),
plus boundary conditions for φ(3.5)
where
p(x) ∈ C1[−1, 1]; q(x), w(x) ∈ C[−1, 1]
p(x) > 0, w(x) > 0 ∀x ∈ (−1, 1)
The Sturm-Liouville problems consist on determining the values of λ, given p(x) and q(x)
continuous functions and weight w(x), for which non trivial solutions φ(x) exist.
The eigenfunctions of a Sturm-Liouville eigenvalue problem are of special interest in
spectral methods due to some of their properties.
• The eigenvalues of a regular Sturm-Liouville problem satisfy
λ1 < λ2 < · · · < λi < · · · (3.6)
• There is a unique eigenfunction φi(x) corresponding to each eigenvalue λi
• The eigenfunctions are mutually orthogonal using the inner product:
(φi(x), φj(x))w = 0, if i �= j (3.7)
• The eigenfunction φi(x) has exactly i− 1 zeros in (−1, 1)
24 3.1 Short Review of Spectral Methods
These properties of the eigenfunctions of a Sturm-Liouville problem are the reason why
this problem is so important in the spectral methods theory. First and second properties
imply that each eigenvalue has algebraic multiplicity one. This guarantees the linear in-
dependence of the eigenfunctions. Moreover the set of mutually orthogonal eigenfunctions
constitutes a family of orthogonal polynomials and the zeros of these polynomials are
preferred points in the collocation method. Nevertheless the most crucial feature of the
eigenfunctions of the Sturm-Liouville problem is the order in the sequence of eigenvalues
(property one) and its relation with the asymptotic decay of the expansion coefficients of
a function in terms of the orthogonal eigenfunctions. This property is the base under the
so-called spectral accuracy, the small quantity of terms in the expansion needed to achieve
a good approximation.
To link the Sturm-Liouville problem with the Chebyshev polynomials, it is necessary
to make a distinction between regular and singular Sturm-Liouville problems.
When p(x) vanishes at the boundaries the Sturm-Liouville problem is singular. The
only polynomials which can be solution of a singular Sturm-Liouville problem are included
in the general class of Jacobi polynomials.
In the particular case when p(x) = (1 − x2)(1/2), q(x) = 0 and w(x) = (1 − x2)(−1/2),
the singular Sturm-Liouville problem becomes the Chebyshev equation,
(x2 − 1)d2u
dx2+ x
du
dx= λ2u (3.8)
The Chebyshev polynomials of first kind are the eigenfunctions of this problem. (The
Chebyshev polynomials of second kind derive from choosing w(x) = w−1(x))
In this work only Chebyshev polynomials have been used in the study of the instability
of flows in cavities. There are several reasons for this choice, these polynomials maintain
the spectral accuracy property, are suitable for non-periodic problems in finite domains
and their roots and peaks play an important role in the optimal approximation theory. In
what follows, only this kind of polynomials will be considered.
3 Numerical Methods 25
3.2 Base Flow Calculations
BiGlobal instability analysis requires a steady two dimensional basic state in which time
disturbances are superimposed. This base solution is computed by the integration of the
laminar Navier-Stokes equations. Consistency with the methodology chosen for the insta-
bility analysis impelled to use a spectral method in the calculation of the two-dimensional
base flow.
The Navier-Stokes equations governing the motion of a viscous incompressible fluid
are set up using velocity-pressure formulation. Numerical methods utilized for spatial-
and time-discretization are presented and discussed. The Chebyshev collocation method is
applied for the spatial discretization using the multi–domain technique to handle nontrivial
geometries. The time integration technique, a semi-implicit coupled scheme where artificial
compressibility is added in the continuity equation, will be detailed.
Finally initial and boundary conditions required for the different problems solved are
also discussed. The problematic open boundary, where the fluid enters or leave the com-
putational domain, will be treated carefully.
The steady two-dimensional basic flow required for the BiGlobal instability analysis
of the open cavity flow is obtained using a semi-implicit artificial compressibility method
detailed in de Vicente et al. [29].
Using this method, the Navier-Stokes equations governing the motion of a viscous
incompressible fluid become:
∂u
∂t+ (u · ∇)u =
1
ReD∇2u−∇P, (3.9)
ε∂P
∂t+∇ · u = 0, (3.10)
where u represents the velocity vector comprising the two velocity components in the
streamwise and wall normal directions (u, v), P is the pressure and ε is an arbitrarily small
parameter for the artificial compressibility component. Chebyshev spectral collocation
technique has been chosen for the spatial discretization, while time advance has been
performed using a semi-implicit Euler scheme leading to the following system of discretized
equations:
26 3.2 Base Flow Calculations
un+1 − un
dt− 1
ReD∇2un+1 + Pn+1
x = −(ununx + vnuny ),
vn+1 − vn
dt− 1
ReD∇2vn+1 + Pn+1
y = −(unvnx + vnvny ),
εPn+1 − Pn
dt− un+1
x − vn+1y = 0. (3.11)
This semi-implicit numerical scheme treats convection terms in the Navier-Stokes equa-
tion explicitly but pressure and viscous terms implicitly. This approach avoids solving a
nonlinear problem at each time step while preserving the stability properties of implicit
solvers. The choice of an Euler method for time advancing is due to its simple implemen-
tation and its good stability properties when a steady solution is sought.
The main drawback of this kind of method is the loss of accuracy before the steady
state is reached. This is not an issue here, since in the BiGlobal analysis methodology a
steady base solution is sought.
The spectral multi-domain methodology implemented is here fully exploited by de-
composing the computational domain into rectangular sub-domains. The presence of a
laminar boundary layer upstream of the cavity, the shear layer developed from the leading
edge, as well as the boundary layer on the downstream cavity wall all require a fine grid
to be properly resolved. A stretching law permits the concentration of the Chebyshev
collocation points in the most interesting part of the domain and consequently leads to a
better resolution.
The computational domain must be defined so that it is large enough to isolate as
much as possible the numerical effects of imposing boundary conditions for open flows.
In the open cavity flow four different kinds of boundaries appear: wall, inflow, far-field
and outflow. First issue concerns the implementation of appropriate boundary conditions.
For viscous wall-bounded flows no-slip boundary conditions are adequate for velocity, while
homogeneous Neumann conditions for the pressure derivative in the normal direction to
the wall are applied.
The velocity components of the laminar boundary layer at the inflow of the computa-
tional domain are calculated by integrating the Blasius equation. Several Blasius profiles
have been computed and used in order to obtain a range of basic flows at different Reynolds
numbers and with different incoming boundary layer thickness, as required by the subse-
quent instability analysis. In the part of the study in which the boundary layer thickness
was not modified it was fixed to a reference value at the start of the computational domain
3 Numerical Methods 27
(δ−1). This reference value for the incoming Blasius profile was chosen to be the same as
in Rowley et al. [87] and Bres and Colonius [20].
This condition, originally chosen to match the works of the aforementioned authors,
is unconventional, in the sense that the parameter kept constant is not the one that
is supposed to dominate the behaviour of the problem (which would be the boundary
layer thickness at the leading edge of the cavity), but a computational one (the incoming
boundary layer thickness at the start of the computational domain). That means that the
constant incoming boundary layer thickness results correspond in reality not to a plane
of such constant thickness, but to a manifold in the parameter space, as it was described
before. The general effect on the results of this is, nevertheless, small, and in any case the
effect of said incoming boundary layer thickness has also been thoroughly studied, as will
be seen in what follows.
So, the following boundary condition is defined at the inflow boundary:
u = uBlasius, (3.12)
v = vBlasius, (3.13)
while, as in standard boundary layer flows, Dirichlet conditions for velocity field lead
to homogeneous Neumann conditions for pressure in the normal direction. The far-field
boundary denotes the upper artificial limit, far enough from the boundary layer, where the
streamwise component of the velocity vector is equal to a characteristic velocity. Normal
velocity is considered constant in the y−direction.
The choice and implementation of the boundary condition in the downstream limit
of the computational domain has been the most critical issue in the convergence of the
steady basic flow calculation. Several works dealing with this open boundary numerical
treatment have been published since 70’s (Gresho and Sani [47], Kobayashi et al. [58],
Orszag and Israeli [72], Orszag et al. [73]). The boundary condition which has shown the
best performance consists of forcing the second derivatives of the flux velocity components
to vanish in the outflow while keeping the pressure constant with a Dirichlet boundary
condition (Gresho and Sani [47], Sani and Gresho [89]):
∂2u
∂x2= 0,
∂2v
∂x2= 0,
P = P0. (3.14)
Regarding the boundary conditions utilised, on the solid walls the viscous conditions
discussed by (de Vicente et al. [29]) have been applied. At the inflow boundary a Blasius
28 3.3 BiGlobal instability analysis
profile is imposed, corresponding with the appropriate Reynolds number and incoming
boundary layer thickness, while the outflow, xout, and far-field, yfar, boundaries have
been placed at large distances from the cavity in order to minimise numerical effects due
to the imposition of artificial boundary conditions for open flows (de Vicente et al. [30]).
Typical parameters chosen are yfar = 5 and xout = 8.
3.3 BiGlobal instability analysis
As it was explained in section 2.2 linear stability analysis in the BiGlobal framework
involves the substitution of a decomposition of any of the independent flow variables, e.g.
the three velocity components, temperature and pressure q(x, y, z, t) = (u, v, w, θ, P )T into
the (incompressible or compressible) equations of motion. The obtention of the complex
eigenvalue from the generalised real non-symmetric eigenvalue problem in schematic form,
Aq = ωBq. (3.15)
Solutions of (3.15) are sought subject to the following boundary conditions for the
disturbance quantities. At solid walls the no-slip boundary condition u = v = w =
0 is imposed on the velocity components, alongside a compatibility condition for the
disturbance pressure, where an homogeneous Neumann condition is the most adequate.
As it was previously mentioned, boundary conditions for the disturbance pressure do not
exist physically, but compatibility is ensured with the choice of normal first derivative
equal to zero.
∂P
∂n= 0. (3.16)
Inflow boundary is treated in the same way, considering there are no velocity fluctua-
tions across this boundary; homogeneous Neumann boundary conditions are imposed there
for the pressure perturbations. At the outflow boundary Neumann boundary conditions
are applied on all the perturbation variables,
∂u
∂n=
∂v
∂n=
∂w
∂n=
∂P
∂n= 0. (3.17)
Far-field, the top boundary of the computational domain, has been initially treated like
the inflow condition. This choice has been proved to be correct for the leading cavity
modes, whose structure is confined within the cavity. On the other hand, these conditions
3 Numerical Methods 29
strongly affect the growth rate of the two dimensional shear layer modes, as it will be
explained in section 5.
For a detailed discussion on the boundary conditions used in this analysis the inter-
ested reader is referred to the work by de Vicente [27]. Spatial domain is decomposed
using the same multi-domain methodology explained in the base flow section, each do-
main discretized using Chebyshev collocation technique. The resulting two-dimensional
BiGlobal eigenvalue problem is solved by Krylov subspace iteration in two stages. First,
the matrix discretizing the EVP is LU-decomposed (though never formed explicitly); sec-
ond, the decomposition is fed into an Arnoldi iteration with shift and invert methodology
to recover the leading eigenvalues closest to the stability bound.
3.4 Three-dimensional Direct Numerical Simulation
The details of the three-dimensional direct numerical simulation used to validate the
BiGlobal results are as follows.
The compressible laminar Navier-Stokes equations constitute a system of partial dif-
ferential equations which can be shortly written in vector form as:
∂U
∂t+∇ · F(U) = 0, (3.18)
where U represents the vector of conservative variables and F(U) represent the 3D fluxes;
including convective and diffusive in the three coordinate directions.
The computationally-demanding nature of the Navier-Stokes solution, in the stability
analysis context, leads to the selection of high-order numerical schemes for the numerical
discretization of System (3.18). High order methods (Spectral type methods) have been
extensively used in computational fluid dynamics due to their accuracy and efficiency
in the simulation of fluid flows. In particular, these methods are suitable for problems
where high accuracy is required and, hence, are well suited to track the evolution of small
perturbations as in stability analysis.
In this context, a Spectral Discontinuous Galerkin method is used in this work to solve
equation (3.18). To do that, the original domain is divided into non-overlapping hexahedral
sub-domains, Ek, such that Ω =∑
k Ek. Inside each sub-domain, a polynomial of degree
N is used to approximate the unknowns and the fluxes, U,F, thus:
30 3.4 Three-dimensional Direct Numerical Simulation
UN =
N∑i,j,k=0
Ui,j,kΦi,j,k, FN =
N∑i,j,k=0
F(Ui,j,k)Φi,j,k, (3.19)
where
Φi,j,k = Li(x)Lj(y)Lk(z),
is the tensor product of the Lagrange interpolant in the nodes i, j, k and Ui,j,k is the value
of the unknown in each computational node. In this work, the nodes in each direction
follow a Gauss-Legendre distribution, and the basis functions L(i,j,k) are taken as the
Lagrange interpolant at these nodes.
Reconsidering equation (3.18), we obtain, at an element level, the following discretized
equation:
∂UN
∂t+∇ · FN = 0. (3.20)
The Discontinuous Galerkin-Spectral Element method (DG-SEM, Kopriva [59]) makes
use of the Galerkin weak form of the equations and a discontinuous treatment of the
interfaces and boundaries. Thus, equation (3.20) is multiplied by a test function (the
same function as the basis for the Galerkin method) and integrated in the computational
space, then the error is forced to be orthogonal at each test function Φi,j,k in a mesh
element Ek, yielding:
(∂UN
∂t,Φi,j,k
)Ek
+ (∇ · FN ,Φi,j,k)Ek= 0, (i, j, k) = 0...N,
with (a, b)Ek=
∫Ek
ab defining an inner product (typically the L2 inner product). After
integrating by parts we obtain:
(∂UN
∂t,Φi,j,k
)Ek
− (∇Φi,j,k, ·FN )Ek+
∫∂Ek
Φi,j,kFN · n �dS = 0, (3.21)
where the third term (the surface integral) extends over the boundary ∂Ek of the com-
putational element Ek, with external pointing normal n. This boundary may lie at the
interface between two elements or at a physical boundary conditions, and, in both cases,
the treatment is similar. Note that all integrals in equation (3.21) can be numerically
evaluated using Gauss quadrature.
To obtain a solution over the complete discretized computational domain (Ω =∑
k Ek)
it is necessary to sum all the element contributions:
3 Numerical Methods 31
(∂UN
∂t,Φi,j,k
)Ω
− (∇Φi,j,k, ·FN )Ω +∑γ∈Γ
∫γΦi,j,kF
∗(n,UL,UR) �dS = 0, (3.22)
where Γ denotes the set of internal edges in the mesh Ω. In addition, note that we
have replaced FN by F∗(n,UL,UR) in the surface integral. F∗(n,UL,UR) represents
the numerical flux between two consecutive elements in the mesh (Left and Right). This
numerical flux arises from the discontinuous Galerkin setting where we consider that each
element is disconnected from the next and hence contains a complete set of degrees of
freedom to represent a polynomial of order N.
Taking into account the decomposition of the unknown (equation 3.19) and the or-
thogonality of the Lagrange basis in the Gauss nodes, the following expression is finally
obtained for the integrals of equation (3.22).
∂Ui,j,k
∂t+DxF
1i,j,k +DyF
2i,j,k +DzF
3i,j,k = 0, (i, j, k) = 0...N. (3.23)
The discrete divergence (second term of the previous equation) is obtained after the
numerical integration of the second and third terms of equation (3.22). Gauss quadrature
is used to evaluate these integrals, giving:
DxF1i,j,k = F1∗(x, yj , zk)
Li(x)wi
∣∣∣x=1
x=0−
N∑m=0
F1m,j,kdi,m,
DyF2i,j,k = F2∗(xi, y, zk)
Lj(y)wj
∣∣∣y=1
y=0−
N∑m=0
F2i,m,kdj,m,
DzF3i,j,k = F3∗(xi, yj , z)
Lk(z)wk
∣∣∣z=1
z=0−
N∑m=0
F3i,j,mdk,m,
(3.24)
with:
dm,n = L′m(sn)wn
wm. (3.25)
In the previous expression, wn are the Gauss integration weights in dimension x, y, or z,
L′m(sn) is the derivative of the Lagrange interpolant evaluate in the node sn, and F∗ are
the interface fluxes. These fluxes can be differentiated into viscous or inviscid.
Computation of inviscid fluxes requires taking into account the left and right values
of the unknowns at each interface. Let us note that by taking the average value of the
unknowns U at the interface (equivalent to a central scheme) provides a numerically
unstable scheme when the convective terms dominate and is only recommended at very
low Reynolds numbers. For larger Reynolds numbers, an upwinding scheme should be used
32 3.5 Dynamic Mode Decomposition
instead. The most common way to introduce upwinding in the scheme is by solving the
equivalent Riemann problem at the interface. For the particular case of Euler equations
(or inviscid NS) different Riemann solver has already been developed. In this work, a
standard Roe Riemann solver has been used in the computations Toro [107].
The viscous fluxes require discretization for elliptic type equations. A simple approach
consists of averaging the right and left viscous fluxes at the interface, but this solution
has been proof numerically unstable for implicit schemes. A more general framework
for derivation and analysis of discontinuous Galerkin methods for elliptic equations (e.g.
Interior Penalty, Local Discontinuous Galerkin, Bassi-Rebay) was derived in Arnold et al.
[7]. Additional details on implementation, methodology and numerical validation of the
employed tool can be found in Jacobs et al. [53], Kopriva [59, 60].
3.5 Dynamic Mode Decomposition
The Dynamic Mode Decomposition (DMD) is a recent data-based technique, introduced
by Schmid Schmid [92], that follows the Koopman analysis of a dynamical system to
find the relevant spatial modes that evolve in a flow field, as previously discussed on the
introductory section (for additional details the reader is oriented to Rowley et al. [88] and
Mezic [67]). It is classified as a data-based technique because the only input required by
this post-processing method is a set of data snapshots, coming from numerical simulations
or experimental measurements. These flow field snapshots have to be collected with a
constant sampling frequency, dictated by the Nyquist criterion. Therefore, in order to
avoid aliasing and a diverged decomposition, the data must be sampled at least at twice
the highest frequency of the dynamic modes to be captured from the analyzed flow field.
In the present work, the sampling frequency is determined by the time each snapshot
has been saved, i.e. 10 non-dimensional time units. Therefore, the decomposition is only
able to capture features oscillating with a maximum StD of 0.05. A snapshot matrix can
then be constructed containing the selected N snapshots, temporally ordered and equally
spaced by the aforementioned constant sampling time:
V N1 = (v1, v2, v3, ..., vN ) (3.26)
The matrix V N1 may be composed of one or all variables of the flow field. If one con-
siders only one variable, the computational burden of the DMD is substantially reduced.
Nonetheless, to describe more accurately the whole system dynamics as much flow field
3 Numerical Methods 33
variables as possible are needed, given that the basis of the dynamic modes are common
to all of them Richecoeur et al. [80]. In the present chapter, it was verified that perform-
ing the DMD with the three velocity components of the flow or only with the spanwise
velocity component did not produce relevant differences on the obtained dynamic modes.
The results presented and discussed in the following section are the ones obtained from
the DMD using only a single variable (z-velocity component).
The fundamental idea of the DMD is to extract the dynamic characteristics of the linear
operator A that approximates the dynamical process between two consecutive snapshots,
as follows:
A(v1, v2, ..., vN−1) = (v2, v3, ..., vN )
AV N−11 = V N
2
(3.27)
For a sufficiently long sequence of snapshots, after performing a Singular Value Decom-
position (SVD) on the snapshot matrix V N−11 = UΣWH and projecting matrix A onto a
basis spanned by the POD modes U , we find a matrix S that can describe the dynamics
of the unknown matrix A, being computed as follows:
S = UHAU = UHV N2 WΣ−1 (3.28)
The following step consists on solving the eigenvalue problem Syi = μiyi. The dynamic
modes can finally be recovered using the following expression:
φi = Uyi (3.29)
From the phase and magnitude of the eigenvalues of S we retrieve the frequencies and
growth rates of the dynamic modes, respectively. For a complete description of the original
DMD algorithm, see Schmid [92].
4The three-dimensional dynamics
Contents
4.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.1.1 On complex conjugate eigenvalues . . . . . . . . . . . . . . . . . 39
4.2 Global instability validation . . . . . . . . . . . . . . . . . . . . . 42
4.2.1 Comparison with Bres & Colonius . . . . . . . . . . . . . . . . . 44
4.2.2 Comparison with 3D DNS . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Parametric analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.1 Preliminary considerations on the envelope of stability . . . . . . 50
4.3.2 Spanwise wavenumber (β) . . . . . . . . . . . . . . . . . . . . . . 53
4.3.3 Reynolds number (ReD) . . . . . . . . . . . . . . . . . . . . . . 55
4.3.4 Incoming boundary-layer thickness (θ0/D) . . . . . . . . . . . . . 57
4.3.5 Aspect ratio (L/D) . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3.6 Eigenmode morphology . . . . . . . . . . . . . . . . . . . . . . . 63
4.3.7 Synthesis of the results . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.7.1 Incoming boundary-layer thickness and Reynolds number 71
4.3.7.2 Parametric dependence on aspect ratio . . . . . . . . . 72
4.4 Summary of the chapter . . . . . . . . . . . . . . . . . . . . . . . 76
4.1 Problem description
A modal point of view is followed and BiGlobal instability analysis has been used to
analyse the flow over an open cavity. Lengths have been scaled with the depth of the
cavity, D, and two independent Reynolds numbers are used to characterise the flow, the
one based on the cavity depth, ReD = u∞D/ν, and the other based on the boundary layer
momentum thickness at the upstream cavity lip (see figure 4.1), Reθ0 = u∞θ0/ν. In the
vast majority of the cases, and unless otherwise stated the Reynolds is based on cavity
depth, and the value of incoming boundary-layer thickness (δ/D) have been chosen to
characterise the flow. The subscript on δ denotes the position in the streamwise direction
where the boundary-layer thickness is measured. So, δ0/D corresponds to the leading
35
36 4.1 Problem description
edge of the cavity while δ−1/D is the beginning of the computational domain (x−1),. The
distance between the inflow boundary and the upstream cavity lip, x0 − x−1 = D, was
always kept constant. As it was explained in the previous chapter, changing the inflow
condition allowed to control both the Reynolds number and the incoming boundary layer
thickness, without needing to change the domain.
L
INFL
OW
OU
TFLO
W
FARFIELD
WALL
-1
D
0x-1x
Figure 4.1: Schematic description of the two-dimensional open cavity and problem parameters.
Some of the linear analysis results presented in this work have been obtained fixing
the incoming displacement thickness of the boundary layer at the inflow boundary to
δ∗−1 = 0.25 (corresponding to momentum thickness θ−1 = 0.0337). The original reason for
this procedure was to reduce the degrees of freedom in the comparison with experimental
work and other authors’ numerical results. It would have been a better idea from a results
point of view to have a constant θ0, given that this parameter is the characteristic length
of the incoming boundary layer thickness that dominates the behavior, instead of using
a derivated length. On the other hand, the use of the boundary layer thickness at the
entry of the domain resulted in an ease of computation of the base flow results for the
cases of constant δ. Table 4.1 summarises the steady flow results obtained at different
Reynolds numbers, with constant incoming boundary layer thickness at the entry of the
4 The three-dimensional dynamics 37
domain; those at ReD = 1500 correspond to run ”2M01” in the work of reference by Bres
and Colonius [20] and will be discussed further in section 4.2. The change of θ0 with
the Reynolds numbers, when the value of θ−1 = 0.0337 is kept constant, can be seen in
figure 4.2. As it can be appreciated, the variation of boundary layer thickness is small,
especially for high Reynolds numbers.
Table 4.1: Parameters characterising the inflow Blasius boundary layer and associated location
of the primary, (xC , yC), and secondary, (xc, yc), vortices.
ReD x0 δ∗0 θ0 Reθ0 (xC , yC) (xc, yc)
1100 3.845 0.1017 0.0392 43.19 ( 1.408 , -0.412 ) ( 1.381 , -0.594 )
1300 4.363 0.0996 0.0384 50.01 ( 1.409 , -0.417 ) ( 1.388 , -0.566 )
1500 4.880 0.0981 0.0378 56.81 ( 1.410 , -0.420 ) ( 1.393 , -0.536 )
1700 5.397 0.0969 0.0374 63.61 ( 1.412 , -0.422 ) ( 1.397 , -0.515 )
1900 5.915 0.0960 0.0370 70.39 ( 1.414 , -0.425 ) ( 1.398 , -0.498 )
2400 7.207 0.0943 0.0363 87.33 ( 1.418 , -0.429 ) ( 1.400 , -0.473 )
0.02
0.025
0.03
0.035
0.04
0.045
1000 1200 1400 1600 1800 2000 2200 2400 2600 2800 3000ReD
0 /D
Figure 4.2: Variation with the Reynolds number of the incoming boundary layer thickness at
the leading edge of the cavity, θ0/D, for constant boundary layer thickness at the start of the
computational domain θ−1/D = 0.0337.
38 4.1 Problem description
Figure 4.3: Schematic description of the 3-D open cavity and problem parameters.
In the complete parametric study presented in this chapter the effect of the variation
of the incoming boundary layer thickness is also studied.
Regarding the geometrical parameters, the two considered are the length-to-depth as-
pect ratio (L/D) and the wavelength in the spanwise direction (Lz) with the corresponding
wavenumber (β = 2π/Lz). A schematic description of the three dimensional configura-
tion is depicted in figure 4.3. The use of β as a parameter on the BiGlobal analysis was
explained in section 3.3. From a geometrical point of view β is more a characteristic of
the eigenmodes, like the frequency, in the sense that the eigenmodes have characteristics
lengths of maximum amplification, corresponding with the appropriate β, the same way
that they have a characteristic frequency. The nature of the analysis used here is such
as that β is a parameter of the problem, the value of β analyzed can be changed, and
performing a sweep on it the one of maximum amplification, and the limits of instability
of each eigenmode, can be determined. There are two main three-dimensional eigenmodes
for the open cavity problem, referred in what follows as Mode I (depicted in red in the
figures) and Mode II (in blue), or bifurcated mode. Mode I corresponds to a pair of com-
plex conjugate eigenvalues, fact that is usually referred in the literature as a travelling
mode. Mode II is a bifurcated mode, as it suffers a bifurcation along the β parameter. For
higher β the mode is stationary, while for lower β it also corresponds to a pair of complex
conjugate eigenvalues. The nature and behaviour of those two modes will be thoroughly
documented in this chapter.
4 The three-dimensional dynamics 39
4.1.1 On complex conjugate eigenvalues
The basis for the BiGlobal instability analysis was explained in section 3.3. If one would
decompose the flow into a sum of the base flow, plus a series of eigenmodes, the resulting
expression would be:
q(x, y, z, t) = q(x, y) +
n∑i=1
αiqi(x, y)ei βiz−i 2π
StDiU∞D
t+σit, (4.1)
where αi is the initial amplitude of each eigenmode, which is supposed to be very small
in the linear regime. In a real flow, each of those amplitudes would be a function of the
initial conditions, random noise in the flow, etc., but due to the exponential nature of the
grow of the eigenmodes, after a short period of time all the eigenmodes except the one of
greater amplification σ would be negligible. That mode would be, in this case, either part
of Mode I or Mode II, with the appropriate β of maximum amplification. In the case of a
real three-dimensional flow with non-periodic spanwise boundary conditions that β could
be selected somehow.
If the leading mode for a random flow condition is the stationary branch of Mode II,
then a good approximation of the flow in the growth regime would be:
q(x, y, z, t) = q(x, y) + αq(x, y)ei βz+σt, (4.2)
as in that case the Strouhal of the eigenmode is zero. In that case, there is only one
parameter for the composition, α, and given that there is no normalization conditions for
the eigenvector q(x, y) that constant can be set to one, and just use the time to describe
the growth of the mode. These stationary mode have a morphology that is characteristic,
since the function in the spanwise direction is simply periodic. When representing the
eigenmodes the isosurfaces of the w velocity component are represented. The spanwise
velocity component, w, is zero in the base flow (w = 0), so what is shown is only the
perturbation (w = w), whether it is a single eigenmode, or a pair of complex conjugates.
As an example, in figure 4.4 the stationary branch of Mode II is shown.
The other possible case is that the leading eigenmode is Mode I, that is, a mode with
a complex conjugate eigenvalue. The flow approximation would be, then:
q(x, y, z, t) = q(x, y) + α1q1(x, y)ei βz−i 2π
StDiU∞D
t+σt + α2q2(x, y)ei βz+i 2π
StDiU∞D
t+σt.
(4.3)
40 4.1 Problem description
Figure 4.4: Spanwise velocity component, w, isosurfaces for a stationary eigenmode.
In this case the growthrate σ of the two eigenvalues are exactly the same, and there are
two relevant amplifications α1 and α2. There is no condition on the linear analysis that
establishes a value or a relationship between both amplifications. The problem, as stated,
has an additional degree of freedom, the way the two complex conjugate eigenvalues are
combined. In the case of a real flow, that degree of freedom would be determined by the
boundary or initial conditions. This kind of perturbations are usually called travelling
waves in the literature, and here it can be seen why, as the problem results in two waves,
one travelling left, and the other travelling right. Thus, different values of the constants
lead to different looking modes. In some cases, the combination of both eigenmodes forms
a standing pulsating perturbation, as the interference pattern of both waves cancels the
movement. In other cases, the combination of modes leads to structures that entwine,
and travel left or right, and the values of said combination affect the morphology of the
structure. Thus, when representing an eigenmode with complex conjugate eigenvalues
a value for the α-constants has to be chosen, to show a characteristic depiction of the
eigenmode. In figure 4.5 examples of different combinations of α are shown.
4 The three-dimensional dynamics 41
Figure 4.5: Spanwise velocity component isosurfaces for a travelling eigenmode for different
combinations of α1 and α2. On top, pulsating perturbation. The others are structures that are
right-travelling (middle row) or left-travelling (bottom).
42 4.2 Global instability validation
4.2 Global instability validation
Before addressing the stability analysis of the open cavity, the methodology employed has
been extensively validated in several test cases. Among these tests, the analyses performed
for confined cavity flow with cross-sectional geometries of increasing complexity de Vicente
et al. [29] have been a decisive tool for evaluating the domain decomposition methodology
and the accuracy of the resulting solution depending on different conditions, such as the
state of convergence of the basic flow, number of points in each computational domain,
or the number of vectors in the Krylov subspace base used in the iterative procedure for
recovering the leading eigenvalues of (2.36). Extensive validation of the instability analysis
algorithm has been performed. Previous works (de Vicente [27], de Vicente et al. [29])
have validated the domain decomposition methodology and the accuracy of the provided
solution.
The aim of this section is, then, to compare, as well as confront, the present results
with previous studies of rectangular cavity flows and with additional DNS computations
in order to check the validity of the BiGlobal linear analysis. Some of the results of the
BiGlobal analysis are briefly discussed here, but they will be studied in depth in section 4.3.
Two main cases are considered here, and also in Chapters 6 and 7.
The first case under consideration is Case A: ReD = 1500, whose details are provided
in table 4.2. The nominal conditions are those of Bres and Colonius [20]. The experi-
mental data (see Chapter 6) correspond to slightly different values for this case, with the
associated uncertainties shown in the table. The values of the linear analysis correspond
to a parametric sweep around the nominal conditions of this case (of both the experiments
and Bres and Colonius [20]), to take into account the uncertainties on the experimental
campaign. Unless otherwise stated the following comparisons have been made at nominal
conditions for each case.
Uncertainties on the flow conditions on experiments are critical when comparing results
with the linear analysis, as small variations of Reynolds number, or incoming boundary
layer thickness can have a huge effect on the stability of the flow. To illustrate that
effect, figure 4.6 shows the amplification rates of the leading modes in the critical range
of β for different ReD and θ0 in order to quantify the effect of varying each parameter
independently. Full symbols have been obtained keeping ReD constant (ReD=1500) while
θ0 varies from θ0 = 0.0379 to θ0 = 0.0475. Empty symbols, on the other hand, describe
the change in the amplification when ReD changes and θ0 is fixed. Being able to measure
4 The three-dimensional dynamics 43
Table 4.2: Parameters for Case A (ReD = 1500). B&C stands for the nominal conditions of
Bres & Colonius 08. ED for the Experimental Data of Chapter 6, with associated uncertainties.
LSA for the values of Linear Stability Analysis. The values of δ∗0 and θ0 are calculated with an
equivalent rig to the one employed in Chapter 6 for B&C and LSA.
δ∗0 (mm) θ0 (mm) θ0 =θ0D Reθ0 ReD
B & C 4.90 1.89 0.0378 56.8 1500
ED 5.59 ± 0.28 2.16 ± 0.11 0.0432 ± 0.0022 65 ± 4.9 1500 ± 43
LSA 4.42 - 8.87 1.71 - 3.43 0.0341 - 0.0685 51.2 - 102.9 1450 - 1550
4 6 8 10 12-0.01
-0.005
0
0.005
0.01
Re = 1500 = 0.0379Re = 1550 = 0.0432Re = 1500 = 0.0432Re = 1450 = 0.0432Re = 1500 = 0.0475
D
D
D
D
D
Figure 4.6: Effect of small variation of the flow conditions on the stability. Amplification vs
β, with full symbols corresponding to variations in incoming boundary layer thickness and empty
symbols to variations in ReD.
the sensitivity in the response to these two parameters is useful when comparing numerical
solutions with experiments. As it can be seen, variations of the parameter of the same
order of magnitude as the experimental uncertainties can have a big impact on the stability
of the flow. The reference value: ReD = 1500 and θ0 = 0.0432 (triangles) corresponds
44 4.2 Global instability validation
to the nominal parameters of the first experimental case. Full circles, on the other hand,
corresponds to the numerical computations at ReD = 1500 and θ0 = 0.0379. An important
remark must be made at this point, the second mode to become unstable shows a big
dependence on the aforementioned flow parameters, this effect, in addition to the rather
wide margin of uncertainty in the experimental flow conditions, could affect the comparison
between both approaches, numerical and experimental.
The parametric space in the BiGlobal analysis of this case has been defined in a wide
range (see table 4.2), to both comprise the experimental conditions and also to determine
the critical instability parameters. As it has been already stated, a small variation on ReD
was considered to deal with the uncertainty in the experimental flow conditions (figure 4.6).
The nominal values of Case B are presented in table 4.3. In this case, there is no
additional information in the literature to compare with, so all information related to this
case and the experimental campaign is presented in Chapter 6. Also, in this case the values
presented for the Linear Stability Analysis are just the nominal ones, as no study of small
variation of parameters was performed, equivalent the one in figure 4.6, and the value of
the incoming boundary layer thickness used for the nominal case is slightly different, as
can be seen in the table. That effect can be nevertheless considered in light of the results
presented in the rest of this chapter, and the difference with the experimental results due
to incoming boundary layer thickness should be negligible, given that the flow regime is
unstable, and very far away from the neutral curve.
Table 4.3: Parameters of the nominal conditions for Case B (ReD = 2400).
θ0 =θ0D Reθ0 ReD
Experimental Data 0.0340 ± 0.0015 81 ± 5.0 2400 ± 51
Linear Stability Analysis 0.036 84.6 2400
4.2.1 Comparison with Bres & Colonius
The constant value of θ−1 chosen allows a straightforward connection with the results
provided by Bres and Colonius [20]. This work is to-date the most complete account
of compressible flows instability over open cavities. In run 2M01 the authors detail the
stability characteristics of compressible flow for an open cavity with aspect ratio L/D = 2
with Mach number M= 0.1, ReD = 1500 and Reθ0 = 56.81. The low Mach number permits
4 The three-dimensional dynamics 45
the comparison with the incompressible limit performed here. The (single) most unstable
mode obtained by Bres and Colonius [20] lies inside the unstable region, almost along the
line of maximal amplification predicted by the present analysis.
As it can be seen in figure 4.7 the four points provided in that work, corresponding
to the three leading modes, match quite well both in frequency and amplification with
the results here presented. In figure 4.7 the results are presented in the same way as it
Figure 4.7: Comparison between Bres and Colonius [20] results (black diamonds) and present
BiGlobal analysis solution (Mode I in red circles, Mode II (bifurcated) in blue circles) for the
nominal conditions of case A. Amplification rate as a function of the spanwise length on the left
and of the dimensionless frequency on the right.
was in Bres and Colonius [20], using the spanwise length of the perturbation (as always
dimensionalized with the depth of the cavity), Lz to describe it. The authors identified
three different modes as the most unstable or least stable for different ranges of β = 2π/Lz,
one of them being indeed unstable. The results disclosed in the present work, using a finer
discretization on the β parameter, show that, for the chosen flow parameters, there is
another unstable mode, this one stationary, for β 10. The mode can be identified in
figure 4.7 (left) where a peak in the unstable region appears at Lz 0.62. Its stationary
nature is certified in figure 4.7 (right) where the peak arises when ωi = 0. It is also
important to address that two of the modes identified by Bres and Colonius [20] correspond
with the stationary and travelling parts of the bifurcated mode, as it will be explained
with greater detail in the next sections, as the rest of this chapter delves more deeply in
the nature of those eigenvalues.
46 4.2 Global instability validation
In figure 4.8 the same results are plotted, this time using the parameter β, which is
the one employed in the rest of this thesis.
Figure 4.8: Amplification versus β for Case A. Mode I in red circles, Mode II (bifurcated) in
blue circles and Bres and Colonius [20] results in black rhombi.
The Strouhal number corresponding to the dominant oscillation frequency, in run
2M01, is St= 0.025. The spanwise wavelength of leading eigenmode is also reported and
its value is β = 1. In the present analysis, the amplification rate of the leading eigenmode
predicted is σ = 0.00981 while the frequency is ω = 0.1623, leading to a Strouhal number of
St= 0.0259. For visualisation purposes the isosurfaces of the spanwise velocity component
(w) are shown in figure 4.10 (left), in which the three dimensional spanwise disturbance
is reconstructed periodically attending to the characteristic wavenumber β = 6, showing
two characteristic lengths in the domain Lz = 2.1.
4 The three-dimensional dynamics 47
4.2.2 Comparison with 3D DNS
Another numerical comparison in the cases A and B has been done using the three-
dimensional non-steady DNS explained in section 3.4. The comparison is expected to be
exact in the linear regime of the DNS.
The DNS computation has been designed acording to the information provided in
table 4.2 and table 4.3. A Mach number equal to 0.1 has been selected, minimizing the
compressibility effects. Periodic boundary conditions in the spanwise direction have been
used, fixing a periodicity length of Lz = 1.05 corresponding to a real spanwise wavenumber
β = 6.
After a long transient, due to the small amplification rate of the leading disturbance,
a periodic behavior with exponential growth is reached. In figure 4.9 the L∞-norm of
the spanwise velocity component of the perturbed flow is monitored as a function of non-
dimensional time. The exponential growth is obtained by a linear least squares technique
fitting the maximum of each period, while a Fast Fourier Transform (FFT) is accomplished
to extract the frequency. As happened in the comparison with the Bres & Colonius
work there is an excellent agreement between DNS and linear analysis results using the
BiGloblal methodology when the wavenumber in the spanwise direction is fixed β = 6. The
amplification rate of the leading eigenmode predicted σ = 0.00981 matches with the one
obtained by DNS σ = 0.00978 while frequencies show an analogous conduct: ωi = 0.1623
vs ωi = 0.1634 computed from DNS results. This results, as well as the ones corresponding
with the next case can be seen in table 4.4.
Finally in figure 4.10 the isosurfaces of the spanwise velocity component (w) of the lead-
ing travelling disturbance are shown for both the DNS computation and the BiGlobal anal-
ysis. The three dimensional spanwise velocity disturbance, left picture, is reconstructed
periodically attending to the characteristic wavenumber β = 6, showing two characteristic
lengths in the domain Lz = 2.1. DNS solution, right figure, reveals a qualitative similar
appearance.
These structures are in good agreement with those obtained in previous analyses, not
only that of Bres and Colonius [20] but also the earlier analysis of Theofilis and Colonius
[104] (in which only the domain inside the cavity was analysed) and that of Theofilis and
Colonius [103], who employed the residuals algorithm from Theofilis [100], all such results
being presumably related with the wake mode instability.
48 4.2 Global instability validation
non-dimensional time
log
(m
ax|
|)
400 600 800 1000
10-11
10-10
10-9
w
Figure 4.9: Temporal evolution of the maximum value of the spanwise velocity perturbation
obtained by DNS.
Figure 4.10: 3D Visualization of spanwise velocity at ReD = 1500 and β = 6: Leading disturbance
obtained using BiGlobal analysis on the left; DNS solution on the right.
Second case, Case B: ReD = 2400, exhibits a more challenging behaviour. The first
drawback to consider comes from the fact that the base flow employed in the computations
is artificially stabilised due to the nature of the numerical method used. The effects of
4 The three-dimensional dynamics 49
using this kind of mean flow in the stability analysis have been previously studied in
Barkley [11] for wake flows and in Piot et al. [78], where a good agreement between the
simulations and linear stability analysis results was observed.
BiGlobal results confirm those findings, two unstable two-dimensional perturbations
are recovered: a mode with frequency StI 0.257 and other with frequency StI 0.42.
However, in the parametric space studied, the flow exhibits three-dimensional disturbances
with higher amplification rates, more relevant in the framework of this work.
According to linear analysis results, as it will be explained in the next section, in the
range of β between β 2 and β 18, the predicted numerical solution is a combination
of three unstable modes. The most unstable one, see figure 4.15, is one of the stationary
(ωi = 0) branches in which Mode II splits. The amplification rate of this disturbance
reaches its maximum σ = 0.044 for β 11.7.
Figure 4.11 depicts the isosurfaces of the three dimensional reconstruction of the span-
wise velocity component (w) of the leading growing disturbance obtained by linear analysis.
Two characteristic length are shown in the domain Lz = 1.05. The structures presented
here are morphologically similar to the ones seen in Faure et al. [36]. In that work, the
aspect ratio was fixed to L/D = 1, but the most unstable structure was the same bifur-
cated mode of β = 12 (see figure 4.23). Once more, there is great agreement between the
BiGlobal results and the linear part of the DNS (see table 4.4). Additional results on the
non-linear part of the DNS computations can be seen in Chapter 7.
Figure 4.11: 3D Visualization of spanwise velocity disturbance obtained at ReD = 2400 using
BiGlobal analysis (left) and DNS at t = 400 (right).
50 4.3 Parametric analysis
Table 4.4: Amplification rate (σ) and frequency (StD) of the leading BiGlobal mode and the
linear part of the DNS, for both cases.
Case Method StD σ
Case A BiGlobal 0.0258 0.00981
Case A DNS 0.0260 0.00978
Case B BiGlobal 0 0.0439
Case B DNS 0 0.0440
4.3 Parametric analysis
The parametric dependence of the stability features of the flow is now discussed in a three-
dimensional framework (β �= 0). Besides the geometric parameter aspect ratio L/D, the
wavelength in the spanwise direction, β, is also examined. In addition, the effect of the
incoming boundary-layer thickness and the Reynolds number has been investigated in this
section. Table 4.5 contains the ranges of the different parameters examined.
Table 4.5: Range of the parameters studied
L/D 1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 2 and 3
ReD ∼ 800 to 4600
β ∼ 0 to 22
δ0/D ∼ 0.05 to 0.3
4.3.1 Preliminary considerations on the envelope of stability
The separation between the stable and unstable regimes in the parametric space is a
hypersurface. This manifold marks the change of stability of the critical eigenvalue at each
point of the parameter space. Owing to its multi-dimensional nature, the representation
of that hypersurface is rather complex, so in order to characterise the stability boundaries
neutral curves are presented as cuts through the hypersurface with parameter planes,
namely ReD vs β, or ReD vs θ0/D. In the next sections, the effect of Reynolds number
and spanwise length on the stability features of the eigenmodes is presented. Subsequently
the consequence of varying the incoming boundary-layer thickness θ0/D and aspect ratio
4 The three-dimensional dynamics 51
L/D is analysed. To conclude this chapter a synthesis of the results, including the relevant
parametric dependencies, is presented.
The most significant result obtained using the BiGlobal analysis is the identification
of the critical parameters related to the destabilisation of the flow over the open cavity.
To illustrate this behaviour is useful to map the different regions in the parametric space
depending on the nature of the flow. Not only that, within the unstable region of the pa-
rameter space, linear analysis can give information about which modes are more unstable,
and what are their characteristics.
The determination of these maps, also known as neutral curves, is done through solving
iteratively the eigenvalue problem, changing the relevant parameters, appropriately. The
effect of the different parameters will be explained in detail in what follows, but here
the analysis start with the variation of the spanwise wavenumber, β. In the left part of
figure 4.12 the relation between the frequency, StD, and the amplification rate, σ, of the
leading modes is shown for different β values, while all the other parameters (Reynolds
number, incoming boundary layer thickness and length do depth aspect ratio) are kept
constant. As can be seen the two critical eigenvalues for β = 4.4 (red) become more
stable as β increases, while the other two leading eigenvalues become unstable. In order
to determine the neutral curves, β has been changed in small steps, ranging from Δβ = 0.2
in zones where no interesting phenomena occurred to Δβ = 0.001 when trying to describe
the bifurcations in detail or to determine the values for which the change in stability
occurs.
The representation of the variation of amplification or frequency with β leads to graph-
ics like figure 4.7 and figure 4.8 shown in section 4.2.1. Changing other parameter, like,
for example, the Reynolds number, produces changes in those curves, making them more
unstable, or more stable, as can be seen in the right part of figure 4.12 (or previously
shown in figure 4.6). The black line in σ = 0 of figure 4.12 constitutes the limit of sta-
bility, where an eigenvalue changes its nature from stable to unstable. Or, looking at a
three-dimensional representation of the eigenvalues in amplification, σ, versus β and ReD
(as can be seen in figure 4.13) for a given L/D and θ0 the plane of σ = 0 marks that limit,
in this case the L/D = 2 neutral curves.
The neutral curves corresponding to the first three cavity eigenmodes are presented
in figure 4.14 for the L/D = 2 case. As it was previously mentioned, the value of the
momentum thickness of the incoming boundary layer at the inflow of the computational
domain is fixed θ−1 = 0.0337, while in the upstream lip of the cavity Reθ0 varies from
52 4.3 Parametric analysis
Figure 4.12: Left:variation of the eigenvalues amplification and frequency with β. L/D = 2,
Re= 2300, θ−1 = 0.0337, β = 4.4 in red, β = 4.8 in green, β = 5.2 in blue and β = 5.6 in black.
Right: Variation of the curves σ versus β with the Reynolds number. All cases are for L/D = 2
and θ−1 = 0.0337, Re= 2300 in red, Re= 2400 in green, Re= 2500 in blue and Re= 2600 in black.
Reθ0 = 43.19 to Reθ0 = 87.33 when ReD increases from the very stable ReD = 1100 to
the unstable ReD = 2900. The fine parametric scan employed (ΔReD 1, Δβ 0.01
in the nose of the modes) permits a precise identification of the three-dimensional critical
conditions for global stability analysis of this case, which are presented in table 4.6.
Table 4.6: Critical parameters of the first three modes for the open cavity flow with aspect ratio
L/D = 2 and θ−1 = 0.0337.
Mode Recrit βcrit
I 1149 5.62
II (a) 1471 9.86
II (b) 1522 4.45
Third 2207 10.34
Thus, the use of representations like figure 4.14 provides additional interesting infor-
mation related to the global stability of the cavity flow. It can be stated that below
ReD = 1149 the basic flow is three-dimensional globally stable for that L/D and θ−1. As
ReD increases, unstable perturbations, initially confined inside the cavity, begin to appear.
The characteristics of these leading eigenmodes: I, II and third in table 4.6, are explained
in detail in the next section. More information can be extracted from the data contained
in figure 4.13. Looking at the maximal amplification of the different eigenmodes, and how
4 The three-dimensional dynamics 53
Figure 4.13: Amplification, σ, versus β and ReD of the three leading modes with positive am-
plification in the open cavity flow of aspect ratio L/D = 2 with θ−1 = 0.0337. Mode I in red,
Mode II in blue and the third mode in white.
they change with the Reynolds number allows to identify which eigenmode is the lead-
ing disturbance for different flow conditions. This can be seen in figure 4.15 where those
maxima are represented for the first three modes.
4.3.2 Spanwise wavenumber (β)
In an idealised spanwise infinite cavity the first structures to amplify in a linear context
would be the ones that correspond to the most unstable for that given point in the param-
eter space. The spanwise wavenumber, β, would characterise the instability modes instead
of being a parameter of the problem. When dealing with three-dimensional centrifugal
instabilities two characteristic spanwise lengths for the periodic instabilities are found to
be inherently more unstable. Perturbations that are of the size of the cavity depth, β 6,
54 4.3 Parametric analysis
Figure 4.14: Neutral stability curves of the three leading modes in the open cavity flow of aspect
ratio L/D = 2 with θ−1 = 0.0337. Mode I in red, Mode II in blue and the third mode in white.
Critical values are cited in table 4.6.
and also perturbations with a spanwise length of about one half of the cavity depth, or
slightly bigger β 10 − 12. In addition, the two-dimensional instabilities, detailed in
Chapter 5, also tend to appear as a third main inherently unstable area in the parame-
ter space. In figure 4.16 the values of maximum amplification on the Reynolds versus β
appear superimposed to the neutral curves already shown on figure 4.14 (the precision of
the values on β is only of 0.2, as they were obtained with automatic post-processing of
the data, and correspond to zones in with the grid variation of β is coarse, without use of
interpolation).
4 The three-dimensional dynamics 55
Figure 4.15: Dependence on Reynolds number of the amplification rate, σ, of the leading eigen-
modes in the open cavity flow of aspect ratio L/D = 2 with θ−1 = 0.0337. Mode I in red, Mode II
in blue and the third mode in white.
4.3.3 Reynolds number (ReD)
The effect of Reynolds number on the flow stability has also been studied, and partially
explained in section 4.3.1. In general, as ReD increases so does the amplification/damping
rate (denominated σ) for all the eigenvalues, as increasing ReD means increasing the
velocity gradient at the shear layer, which is one of the main mechanisms for introducing
energy into the small perturbations.
The neutral curves for L/D = 2 were shown in figure 4.14. There are three main
unstable modes. The first one to become unstable (Mode I) for this configuration, in red
circles in the figures, is the mode associated with a pair of complex conjugated eigenvalues,
known in the literature as a travelling mode. The second mode, in blue circles, undergoes
a bifurcation with the variation of β. For low wavenumbers, there are two complex conju-
gated eigenvalues (so it is also a travelling mode) but when β grows those two eigenvalues
56 4.3 Parametric analysis
Figure 4.16: Neutral stability curves of the two leading modes in the open cavity flow of aspect
ratio L/D = 2 with θ−1 = 0.0337 in circles, approximate maxima in rhombus. Mode I in red,
Mode II in blue.
collapse and become two real eigenvalues, one of which gets stabilised, and the other one
gets destabilised. This disturbance will be named Mode II, or Bifurcated Mode in what
follows. The third unstable eigenmode shown, in empty circles, is also a travelling distur-
bance. That third mode is of little relevance and it is only present for relatively high ReD
configurations, in which the validity of the linear approximation can be questionable.
As it was shown in figure 4.15, the different modes change their amplification with
the Reynolds numbers at different rate, initially Mode I is dominant but as the Reynolds
number increases the unstable stationary branch of Mode II becomes dominant.
4 The three-dimensional dynamics 57
4.3.4 Incoming boundary-layer thickness (θ0/D)
To perform the study on the effect of the incoming boundary layer thickness the critical
parameter ReD and β of these modes have been tracked with the change of θ0. In figure 4.17
the behaviour of the critical ReD when θ0 varies from θ0 = 0.0317 to θ0 = 0.0665 is shown.
The critical wavenumber β remains almost constant (β = 5.62 for the leading and β = 9.86
for the stationary mode) independently of the boundary layer thickness, this effect is the
same when the flow parameter modified is ReD. To clarify the meaning of this figure, the
particular point (ReD, θ0) = (1149, 0.039) has been highlighted. This value corresponds to
the nose of the most unstable mode in figure 4.14. The parametric sweep in the ReD vs
θ0/D space is presented for Mode I and Mode II for L/D = 2, and Mode I for L/D = 3.
Both increasing ReD and decreasing boundary-layer thickness leads to larger gradients,
and eventually to destabilisation of the flow. The effect of decreasing the boundary-layer
thickness is qualitatively similar to increasing ReD for the three-dimensional instabilities,
while it has a much more pronounced effect in the two-dimensional shear layer modes.
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
700 1200 1700 2200 2700ReD
0 /D
Figure 4.17: Neutral curves in ReD vs θ0/D for the critical β. Mode I in red symbols, Mode II
(bifurcated) in blue symbols. L/D = 2 in circles and L/D = 3 in squares. Highlighted point
corresponds with the nose of the most unstable mode in figure 4.14.
58 4.3 Parametric analysis
4.3.5 Aspect ratio (L/D)
The effect of the aspect ratio on the stability is studied next and the dependence between
the size of the perturbations and the critical Reynolds number based on D is documented
(see Section 4.3.7). While L D the size of that perturbation is constrained by the
length of the cavity. Once L/D is big enough the vortex over which the perturbations
develop is not bounded by the cavity length. That means that the size of the perturbation,
denoted LP /D, is determined by the size of the recirculating vortex, not directly by L/D.
In addition, the different behaviour of the different modes with L/D leads to significant
differences in the neutral curves for each aspect ratio, as can be seen in figures 4.18, 4.19
and 4.20. Figure 4.18 shows just the evolution of Mode I with the L/D parameter, and
figure 4.19 does the same for Mode II, while figure 4.20 shows the neutral curves for each
of the L/D configurations studied.
Figure 4.18: Neutral curves for θ−1/D = 0.0337 of Mode I. Aspect ratio varying from L/D = 1
(higher ReD numbers) to L/D = 3 (lower ReD numbers) with the values detailed in Table 4.5.
4 The three-dimensional dynamics 59
Mode I presents a lobe with β ∼ 12 for low L/D (figure 4.18). In the case of L/D = 1
this is the second mode to become unstable (figure 4.20). When the aspect ratio increases
up to L/D = 1.1 a new lobe of that mode becomes unstable in the range of Reynolds
numbers studied, with a maximum amplification on the range of β ∼ 6. For the L/D = 1.2
cavity there is a change in the leading instability of this mode, as the low β lobe is the
first one to become unstable. Increasing the aspect ratio even more to L/D = 1.3 leads to
the main shift in stability, as Mode I becomes the first one to become unstable, while the
second lobe of that mode, now on β ∼ 11 starts to vanish (figure 4.20. By L/D = 1.4 the
mode comprises a single lobe with a maximum in the range of β ∼ 6, and a slight lump
of increased instability around β ∼ 11. As L/D continues to increase, to 1.5, 2 and 3 the
basic shape of the instability curves of the mode remains unmodified and only a change
in the rate of variation of the critical ReD can be appreciated, as it will be discussed in
the next section.
Figure 4.19: Neutral curves for θ−1/D = 0.0337 of Mode II. Aspect ratio varying from L/D = 1
(higher ReD numbers) to L/D = 3 (lower ReD numbers) with the values detailed in Table 4.5.
60 4.3 Parametric analysis
Mode II (figure 4.19) has an even more interesting behaviour, being a bifurcated mode.
At lower β a complex conjugate pair of eigenvalues is found in the analysis. Around β ∼ 9
a bifurcation occurs, and two real modes appear, one becoming more unstable (σ in-
creases), and the other one more stable, as σ decreases. So for L/D = 1 the most unstable
eigenmode corresponds to the real branch of the bifurcated mode, with a maximum am-
plification around β ∼ 12, i.e. a characteristic periodicity length around D/2. In the case
of this mode the most unstable part is always the real branch of high β. The difference
in critical Reynolds number between the two branches decreases as L/D increases. This
mode is the critical one for L/D = 1.2 and lower, as was discussed above when explaining
the features of Mode I. At L/D = 1.3 and L/D = 2 there is a small increase of instability
in the zone of β ∼ 3 and an small lobe is formed. In the case of L/D = 1.4 and L/D = 1.5
there is a turning point in that region, but a complete lobe is not formed. That change of
behaviour is interesting, because the trend is not continuous with the change of the L/D
parameter. Also, the only instance of critical ReD increasing as L/D increases can be seen
in the travelling branch of this mode, for L/D = 3.
Lastly, in figure 4.20 two other unstable modes can be seen in the range of parameters
studied. The third mode corresponds with a complex conjugate pair of eigenvalues, and it
can be seen between L/D = 1.2 and L/D = 2 with a maximum amplification at β ∼ 11.
This mode is more complex in structure, but it is relatively not very important for the
physical behavior of the flow. By the time this mode starts to become unstable the other
two modes have been developing out of the linear regime for any real flow, with saturation
or non-linear interaction of the modes, and even more, the relatively high ReD at which
it appears may imply that there are two dimensional instabilities present already, as it
will be explained in the two dimensional section of this paper. The last mode present is
a particularity of L/D = 3, being strikingly different, and appearing for β ∼ 3, where no
previous disturbances have been founded for lower cavity aspect ratios.
The variation of the critical Reynolds number of the two different lobes (when present)
of Modes I and II with L/D for the constant δ−1/D is shown in figure 4.21.
A summary of the main characteristics of the critical values when a change in stability
occurs for constant δ−1/D is shown in table 4.7. This table serves as a reference for the
rest of the figures, as it is impossible to provide numerical values for the whole problem
(i.e., when the variation of incoming boundary layer thickness is taken into account).
4 The three-dimensional dynamics 61
Figure 4.20: Neutral curves for θ−1 = 0.0337. From left to right, and from top to bottom,
L/D = 1.0, L/D = 1.1, L/D = 1.2, L/D = 1.3, L/D = 1.4, L/D = 1.5, L/D = 2.0 and
L/D = 3.0. Mode I in red, Mode II (bifurcated) in blue, third mode in white and fourth mode in
black.
62 4.3 Parametric analysis
0.81
1.21.41.61.82
2.22.42.62.83
800 1200 1600 2000 2400 2800 3200 3600 400ReD
L/D
Figure 4.21: Neutral curves in the ReD vs L/D for the critical β and θ−1 = 0.0337. First mode
in red, second mode in blue, with circles being low β and rhombi high β.
Table 4.7: Critical parameters of the first three modes for each of the aspect ratios and for
constant δ−1/D.
First Mode Second Mode Third Mode
L/D ReD β ReD β ReD β ReD β ReD β
1.0 - - 3937 13.47 - - 3421 13.07 - -
1.1 3630 6.56 3167 12.45 3884 6.63 2807 12.22 - -
1.2 2565 6.38 2659 11.60 3205 6.16 2383 11.54 3476 11.61
1.3 1985 6.23 2289 10.66 2608 5.63 2069 10.95 3071 11.07
1.4 1659 6.09 - - 2147 5.13 1849 10.45 2863 10.88
1.5 1472 5.96 - - 1853 4.84 1708 10.11 2724 10.83
2.0 1150 5.62 - - 1523 4.45 1471 9.86 2207 10.34
3.0 865 5.30 - - - - 1342 10.06 - -
4 The three-dimensional dynamics 63
4.3.6 Eigenmode morphology
For visualization purposes the structures corresponding with the w velocity component of
Mode I for different L/D can be seen in figure 4.22. It is interesting to notice the difference
in the characteristic lengths of the structures in the spanwise direction, corresponding
with the two different β lobes that appear for lower L/D configurations (L/D = 1.1 in the
figure). As it was previously said, the high beta lobe disappears for L/D > 1.3. Also, it is
remarkable how the structures are constrained by the left wall of the cavity. In the most
extreme case, for L/D = 3, the mode is located only in the downstream part of the cavity,
and does not occupy it in its entirety, which has an important effect in the stability, as it
will be explained in the next section.
In figure 4.23 the w velocity component for the stationary branch of Mode II with
different aspect ratios L/D are shown. As it happened with Mode I, for different L/D
the mode shows different morphology, as the upstream lobe position is affected by the
presence of the left wall of the cavity. In this case the lobe is straight for L/D > 1.5,
instead of being bent down, like it was in all the other cases. Several other interesting
morphological variations on the structure surrounding the main recirculating vortex can
also be observed, as the complexity increases with L/D.
In figure 4.24 the structures of the bifurcated mode at the other side of the bifurcation,
the low β branch, are shown. These structures correspond with the travelling part of the
mode. The aspect ratios are in the same position in figure 4.24 as they were in figure 4.23,
to make easier the comparison between both (except of the top left one, corresponding to
L/D = 1.1 instead of L/D = 1, given that in the range of Reynolds number studied there
is no low β lobe for L/D = 1). A very remarkable change in behavior occurs in this case,
as the mode for larger L/D extends to the secondary recirculating vortex, occupying the
whole cavity, even for L/D = 3. That anomaly might explain why in that case the critical
Reynolds number of that branch is higher for L/D = 3 than for L/D = 2. In the next
section the relationship between the length of the perturbation and the critical Reynolds
number is explained, and in this case the radical change of shape of the eigenmode affects
such relation.
Finally, the last, and less relevant, eigenmodes are shown in figure 4.25. The fourth
mode, on top, has a high β, and shows a more complex coiling than Mode I. The fourth
mode, only present in L/D = 3 shows a radically different structure, unlike any of the
other modes, not only occupying the whole cavity, but having a maximum amplification
in the range of β 3, lower than any other mode.
64 4.3 Parametric analysis
Figure 4.22: From left to right and top to bottom, w velocity component isosurfaces of Mode I
for, L/D = 1.1 (ReD = 3167, β = 12.45), L/D = 1.1 (ReD = 3630, β = 6.56), L/D = 1.5
(ReD = 1472, β = 5.96), L/D = 2 (ReD = 1150, β = 5.62) and L/D = 3 (ReD = 865, β = 5.30)
respectively. In all cases the spanwise length shown is Lz = 2π/β = 2D.
4 The three-dimensional dynamics 65
Figure 4.23: From left to right and top to bottom, w velocity component isosurfaces for the
stationary branch of Mode II for, L/D = 1 (ReD = 3421, β = 13.07), L/D = 1.3 (ReD = 2069,
β = 10.95), L/D = 1.5 (ReD = 1708, β = 10.11), L/D = 2 (ReD = 1471, β = 9.86) and L/D = 3
(ReD = 1342, β = 10.06) respectively. In all cases the spanwise length shown is Lz = 2π/β = 2D.
66 4.3 Parametric analysis
Figure 4.24: From left to right and top to bottom, w velocity component isosurfaces for the
travelling branch of Mode II for, L/D = 1.1 (ReD = 3884, β = 6.63), L/D = 1.3 (ReD = 2608,
β = 5.63), L/D = 1.5 (ReD = 1853, β = 4.84), L/D = 2 (ReD = 1523, β = 4.45) and L/D = 3
(ReD = 1638, β = 4.61) respectively. In all cases the spanwise length shown is Lz = 2π/β = 2D.
4 The three-dimensional dynamics 67
Figure 4.25: From left to right and top to bottom, w velocity component isosurfaces for the third
mode, L/D = 1.2 (ReD = 3476, β = 11.61), L/D = 2 (ReD = 2207, β = 10.34) and the fourth
mode of L/D = 3 (ReD = 1442, β = 2.92) respectively. In all cases the spanwise length shown is
Lz = 2π/β = 2D.
All the figures presented until now correspond to the w velocity component. In order
to visualize the process of mode growth figure 4.26, figure 4.27 and figure 4.28 show the
evolution of the three velocity components of the composition of the base flow plus the
travelling disturbance that is the peak of Mode I. Each figure represents one of the velocity
components at different times during the temporal evolution in the linear regime. In the
case of figure 4.26, showing the streamwise velocity component, the isosurfaces are shown
upside down to the usual, so the structure in the main recirculation vortex is not covered
by the ones of the developing boundary layer.
68 4.3 Parametric analysis
Figure 4.26: From left to right and top to bottom, u velocity component isosurfaces for Mode I.
Time with respect of the final time t/tf = [0; 0.25; 0.5; 0.625; 0.75; 0.875; 0.94; 1] In all cases the
spanwise length shown is Lz = 2π/β = 2D.
In the first two figures the recirculation vortex can be clearly seen, and also how the
perturbation starts to grow around it, varying its shape, and finally coiling around it. In
figure 4.28 the evolution of the spanwise velocity isosurfaces is shown. Since the times
shown are the same as in the last two figures, only the last four frames are shown, since
the value of w is too small to have any isosurface at t/tf < 0.6. These last three figures
allow the reader to have an idea of how does the eigenmode growth look like.
4 The three-dimensional dynamics 69
Figure 4.27: From left to right and top to bottom, v velocity component isosurfaces for Mode I.
Time with respect of the final time t/tf = [0; 0.25; 0.5; 0.625; 0.75; 0.875; 0.94; 1] In all cases the
spanwise length shown is Lz = 2π/β = 2D.
70 4.3 Parametric analysis
Figure 4.28: From left to right and top to bottom, w velocity component isosurfaces for Mode I.
Time with respect of the final time t/tf = [0.75; 0.875; 0.94; 1] In all cases the spanwise length
shown is Lz = 2π/β = 2D.
4 The three-dimensional dynamics 71
4.3.7 Synthesis of the results
In this section a description of relations between parameters of the flow and its global
instability characteristics is presented. An empirical law is provided when possible.
4.3.7.1 Incoming boundary-layer thickness and Reynolds number
In figure 4.17 it can be appreciated that for sufficiently thick boundary layers there seems to
exist a linear relation between the critical ReD and the incoming boundary-layer thickness
for the modes studied. Mode I of L/D = 3 presents a saturation regime for thin boundary
layers in which that linear behaviour does not occur. The slope of the linear part of the
curve depends on the length of which the Reynolds numbers is based. Basing Re on L
instead of D causes the neutral curves of Mode I for both L/D = 2 and L/D = 3 pictured
in figure 4.17 to have almost the same slope in the higher Reynolds range, as can be seen
in figure 4.29.
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
1700 2200 2700 3200 3700 4200 4700 5200ReL
0 /D
Figure 4.29: Neutral curves in ReL vs θ0/D for the critical β. Mode I in red symbols, Mode II
(bifurcated) in blue symbols. L/D = 2 in circles and L/D = 3 in squares.
72 4.3 Parametric analysis
4.3.7.2 Parametric dependence on aspect ratio
The results of figure 4.21 indicate that an exponential fit can be constructed between the
critical ReD of the modes and the inverse of the aspect ratio, that is, D/L, for aspect
ratios lower than L/D = 1.5. That exponential law can be seen in figure 4.31 (top).
To explain that change in behaviour for higher aspect ratios, figure 4.30 contains a cut
of the spanwise velocity component of the low β lobe of Mode I for L/D = 1.2, L/D = 1.5,
L/D = 2 and L/D = 3, similar to what could be seen in figure 4.22. When L/D increases
sufficiently the size of the perturbation stops being constrained by the upstream wall
and it is more appropriate to use the characteristic length of the perturbation LP /D on
the flowstream direction in the aforementioned exponential law, as seen in figure 4.31
(top). That perturbation length can be defined, to use as reference, as the horizontal
distance between the 50% of the leftmost peak of the spanwise velocity component of the
perturbation and the downstream wall of the cavity.
ReDcrit = C1eC2D/LP . (4.4)
The parameters C1 and C2 for the two lobes of each of the first and second modes can
be found in table 4.8 (left).
Figure 4.30: Qualitative representation of the spanwise velocity component of the low β lobe of
Mode I for L/D = 1.2, 1.5, 2 and 3, from top to bottom and left to right.
4 The three-dimensional dynamics 73
800
1300
1800
2300
2800
3300
3800
1 1.5 2 2.5 3
ReD
L/D
4
5
6
7
8
9
10
11
12
13
14
1 1.5 2 2.5 3L/D; L P /D
Figure 4.31: Dependence with L/D and LP /D of the critical ReD for the critical β (top) and
said β of maximum amplification (bottom). All data with θ−1/D = 0.0337. Mode I in red symbols,
Mode II in blue symbols, with circles for low β and rhombi high β. In lines, the predicted curves
of equation 4.4, and the hyperbolic law described in § 4.3.7, and in small empty symbols, of same
shape and color, the equivalent LP /D of the same points.
74 4.3 Parametric analysis
Similar behaviour occurs in the case of the β of maximum amplification. While the
modes are in the range in which LP /D L/D there is a linear relation between L/D
and Lz (that is, an hyperbolic relation with β), as can be seen in figure 4.31 (bottom).
Out of that range, substituting the characteristic length for LP /D, yields that the curve
is being followed for Mode I. In the case of Mode II the behaviour is as expected for the
low β branch, except for the L/D = 3 point, that having a complete different morphology
does not follow it. The other branch has the same behaviour, while the morphology of
the modes is similar (see figure 4.23), the law is followed, but when the leftmost structure
uncoils, the behavior is different. The results presented lead to the conclusion that a
proportionality relationship exists between the depth to width ratio and the length of
maximum amplification 2π/β, for the critical eigenmodes, so as L/D, and consequently
LP /D increases, the length of the perturbation in the spanwise direction tends to increase
in similar proportion. These laws of behavior are valid as long as the morphological shape
of the eigenmodes is similar, once the variation of L/D produces significant differences the
expressions are no longer valid.
Another observation is the apparent relation between the frequency of the eigenmodes,
expressed as the dimensionless Strouhal number, St, and the cavity aspect ratio L/D. For
a given aspect ratio there is little variation of the Strouhal number of the modes with
the Reynolds number or the incoming boundary-layer thickness, so each mode appears to
be associated with a characteristic frequency. Figure 4.32 is a plot of the dimensionless
frequency of the different eigenmodes versus the aspect ratio. The frequency of some
stable eigenmodes is also considered. Using the results of the present analysis a correlation
between the two parameters can be constructed, which follows a potential law described
in equation 4.5.
St = K1(L/D)K2 (4.5)
The parameters K1 and K2 for the different modes can be found in table 4.8 (right).
Very few results can be found in the literature on these three dimensional structures with
low Strouhal. Nevertheless, the recent work of Basley et al. [16] presents an experimental
study on the open cavity flow in which a broad band of frequencies corresponding with
three dimensional structures can be found. The range of those frequencies match those
found here, as do the results presented in Bres and Colonius [20].
4 The three-dimensional dynamics 75
Table 4.8: The parameters C1 and C2 of equation (4.4) for the different modes(left) and the
parameters K1 and K2 of equation (4.5) for the different modes (right).
Mode C1 C2 Mode K1 K2
Mode I low β branch 112.1 3.79 Mode II 0.0069 0.696
Mode I high β branch 376.0 2.35 Mode I 0.0167 0.551
Mode II low β branch 234.5 3.11 Mode III 0.0340 0.635
Mode II high β branch 411.7 2.11 Stable Mode 0.0545 0.609
Stable Mode 0.0854 0.645
Figure 4.32: Strouhal number for the main oscillating modes versus the length-to-depth aspect
ratio of the cavity, both in logarithmic scale. Mode I, Mode II and the third unstable mode are
in red circles, blue circles and empty circles respectively. The other two, in triangles, are stable
modes for the range of parameters studied. In black rhombi the dominant frequency of Bres and
Colonius [20], and the range of frequencies obtained by Basley et al. [16] on the crossed line.
76 4.4 Summary of the chapter
4.4 Summary of the chapter
The onset of three-dimensional instability of the incompressible flow over spanwise homo-
geneous flow in a rectangular cavity has been described. For a wide range of the parameters
involved in the problem, the features of the dominant perturbations have been studied, al-
lowing the extraction of relations between the different parameters and a description of the
neutral hypersurface, through cuts with several planes of the parameter space. Two main
three-dimensional modes appear in the majority of the cases studied: a bifurcated mode,
noted Mode II, with both a travelling and a stationary (and more unstable) lobe, as well
as a purely travelling mode, noted Mode I. The stability behaviour of those modes with
variations of all the parameters has been reported, allowing predictions of the stability of
any configuration sufficiently close to critical conditions using only geometrical features
of the cavity. In particular, the relations between the characteristic Strouhal number of
the different eigenmodes, their critical Reynolds number and the L/D ratio have been
characterised. Such information should be useful not only in order to predict instabilities
of a given open cavity, but also as basis for reduced-order-models of global flow instability
and subsequent flow control, both theoretical and experimental.
5The two-dimensional limit
Contents
5.1 Convergence of BiGlobal analysis on β → 0 . . . . . . . . . . . . 77
5.2 Shear layer modes . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2.1 Characteristic frequency . . . . . . . . . . . . . . . . . . . . . . . 80
5.2.2 On eigenmode amplification . . . . . . . . . . . . . . . . . . . . . 81
5.1 Convergence of BiGlobal analysis on β → 0
The two-dimensional analysis of the flow over a rectangular cavity has been widely studied
in the literature. For the sake of completeness, a two-dimensional BiGlobal (β → 0)
analysis of the stability of the laminar two-dimensional incompressible cavity flow is first
discussed.
The first concern is with the numerical stability of the code when β → 0. As it can
be seen in de Vicente [27] the problem becomes singular for β = 0, and relatively low β
lead to ill conditioned but solvable problems. For these cases the behaviour of the two
dimensional structures can be assimilated to the one of structures with high enough Lz.
In figure 5.1 the evolution of the growth rate of the two least stable eigenmodes on
the 2D limit is shown for a test case. It can be seen that in the range of β ∼ 0.01 to
β ∼ 10−6 the eigenvalues do not change too much, while for β smaller than that numerous
spurious eigenvalues start to appear, and the two eigenvalues jump to a new value. It is
unlikely that there is something in the physics of the problem that changes at Lz/D ∼ 106,
given that β ∼ 10−6 is the range in which the eigenvalues related to the ill conditioning
of the problem grow. Consequently it is safe to assume that the eigenvalues on β = 0.01
sufficiently describe the two dimensional instabilities. All the computations presented in
the following 2D section are made with said β = 0.01.
77
78 5.2 Shear layer modes
Figure 5.1: Evolution of the growth rate of the two least stable eigenvalues on the two dimensional
limit, in a range of β ∼ 10−2 to β ∼ 10−9 for L/D = 2, ReD = 1500 and the constant δ−1.
5.2 Shear layer modes
The relevance and behaviour of the shear layer modes has been explained in Chapter 1.
This kind of structures consist on a feedback loop between the shear layer disturbances
and the acoustic disturbances in the impinging jet. Small perturbations in the shear layer
force periodical movements up and down (Kelvin-Helmholtz instabilities), which in turn
excites acoustic noise to develop from the impinging edge, where vortex shedding can
occur. These acoustic disturbances, which in the incompressible regime can be consid-
ered as instantaneous, travel upstream and excite the shear layer oscillations, closing the
feedback loop. These modes, extensively documented in Gloerfelt [41], follow Rossiter
semi-empirical formula,
Stn =n− γ
M+ 1κ
, n = 1, 2, 3, ..., (5.1)
5 The two-dimensional limit 79
where M is the Mach number, and γ and κ are empirical constants, given that the
characteristic length of the cavity is related to the appropriate resonance frequency. In
figure 5.2 the velocity profiles of the reconstruction of the total flow consisting of the base
flow plus the first of those shear layer modes is shown for a L/D = 2, and ReD = 2400 test
case. The shedding of vortices due to the impingement of the shear layer on the trailing
edge can be seen in the figure, as well as the resulting perturbation of the recirculating
vortex.
Figure 5.2: Velocity profiles of the reconstructed 2D flow with the first shear layer mode. Stream-
wise velocity on the top, and wall normal velocity on the bottom.
80 5.2 Shear layer modes
5.2.1 Characteristic frequency
In figure 5.3 the characteristic frequency of the least stable 2D eigenmodes in the range
of ReD and δ0/D studied (shown in table 4.5) is presented for L/D = 2 and L/D = 3.
0
0.02
0.04
0.06
0.08
0.4 0.5 0.6 0.7 0.8 0.9 1StL
0 /D
Figure 5.3: Variation of the dimensionless frequency (Strouhal based on cavity length StL) of
the two least stable eigenmodes (circles ◦ and rhombus � respectively) in the 2D limit with the
incoming boundary-layer thickness. Results for L/D = 2 in empty symbols (◦), and L/D = 3 in
full symbols(•). In grey shades, the range of Strouhal numbers obtained by Sarohia [90], and in
red symbols single points of several L/D = 2 works. The 2M2 run of Rowley et al. [87] in squares
(�), the Bres [19] M= 0.3 run as a rhombus ( ), the lower runs of Yamouni et al. [112] in triangles
(�). Points from Basley et al. [16] as circles (©).
As it can be seen, the frequency of the two shear layer modes found in this analysis
matches the range predicted in the literature (Basley et al. [16], Bres [19], Rowley et al.
[87], Sarohia [90], Yamouni et al. [112]), as well as the values of the three-dimensional DNS
computations made, as it was explained in section 4.2. The reduction in the Strouhal num-
ber of the modes with the increase of the incoming boundary-layer thickness reported here
was already observed in Yamouni et al. [112]. Given the relevance of L in the behavior
of this eigenmodes it is more usual in the literature to show the results of the two dimen-
sional shear layer modes as a function of the length of the cavity divided by the incoming
boundary layer thickness, L/θ0. In figure 5.4 the same data can be seen in that form, for
5 The two-dimensional limit 81
ease of comparison with other works. Note that using this dimensionalization does not
lead to a total collapse of the data of all L/D.
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
40 50 60 70 80 90 100 110
StL
L/ 0
Figure 5.4: Variation of the dimensionless frequency (Strouhal based on cavity length StL) of
the two least stable eigenmodes (circles ◦ and rhombus � respectively) in the 2D limit with the
length of the cavity dimensionalized with the incoming boundary-layer thickness, L/θ0. Results
for L/D = 2 in empty symbols (◦), and L/D = 3 in full symbols(•). In in red symbols single
points of several works: the 2M2 run of Rowley et al. [87] in squares (�), the Bres [19] M= 0.3 run
as a rhombus ( ), the lower runs of Yamouni et al. [112] in triangles (�) and points from Basley
et al. [16] as circles (©).
5.2.2 On eigenmode amplification
Regarding the amplification of those modes, the behaviour is consistent with previous
results in the literature: when the shear layer becomes thicker the Kelvin-Helmholtz in-
stability dampens until the layer becomes stable, as has been observed experimentally
(Gharib and Roshko [40], Rockwell and Knisely [83]). Nevertheless, an issue arises when
considering the stability of these shear layer modes. Since the structure that can be seen
in figure 5.2 is convected downstream, it is partially responsible for the global stability of
the mode, which is dependant on the resolution of the mode wake.
82 5.2 Shear layer modes
Figure 5.5: Streamwise velocity profiles of the first shear layer mode, for ReD = 2400 and
θ−1/D = 0.0337. Different length domains, from top to bottom xout = 8, 13, 19 and 21 respectively.
5 The two-dimensional limit 83
These modes are spatially unstable, that is, after the cavity there is a growing wake,
corresponding with the shedding of vortices from the cavity, and then that wake sometimes
dissipate, depending on the Reynolds number. The aforementioned spacial instability is
reflected on the temporal instability analysis through a special sensibility to the domain
and the boundary conditions. Our studies show that the variation of the frequencies of
the mode is small when the length of the domain is changed (less than 5%). However, the
growth rate varies greatly, to the point of affecting the stable or unstable nature of the
mode. Figure 5.5 shows the streamwise velocity of the leading eigenmode with varying
domain length, and figure 5.6 shows the variation of the amplification of the eigenmodes
as the domain length changes from xout = 8 to xout = 21.
Figure 5.6: Variation of the dimensionless frequency (Strouhal based on cavity length StL) and
amplification σ of the two least stable eigenmodes with the change of domain length xout.
84 5.2 Shear layer modes
As it can be seen in figure 5.6, the Strouhal of the two leading two-dimensional eigen-
values suffers little variation as the domain changes, while the amplification changes dras-
tically. The behavior with the change of domain, and the change of flow parameters is
consistent. If the mode is sufficiently resolved (i.e., the domain is not excessively short),
for a given domain length increasing the incoming boundary layer thickness makes the
mode more unstable, as expected, and so on, independently of the length of the domain.
Nevertheless, if the domain is kept constant through the parametric sweep the behavior
of the modes is consistent, and a neutral curve, similar to the one made for the 3D modes
can be made, but once the domain changes so does the growth rate, rending those neutral
curves moot. In any case, the general effect of the parameters on the stability behavior
can be seen on that neutral curves, and as such are represented in figure 5.7.
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
1100 1600 2100 2600ReD
0 /D
Figure 5.7: Neutral curves in ReD vs θ0 for the two least stable two-dimensional modes, in circles
◦ and rhombus � respectively. Three-dimensional Mode I and Mode II in red and blue, as in
figure 4.17. Domain length kept constant at xout = 8.
The two dimensional stability results presented in that figure corresponding with runs
in which the computational domain was kept constant in extension, number of points
and boundary conditions (and the last domain goes up to xout = 8). It can be seen
that the behavior here is not linear, with a lobe of instability appearing for the second
Rossiter mode on a certain range of incoming boundary layer thickness. The first mode
seems to behave similarly, but it was impossible to map the shape of the lobe due to it
being sufficiently into the unstable zone of the second mode to make impossible to get
5 The two-dimensional limit 85
stable 2D DNS computations. Additional experiments have been made, using several
2D DNS codes, and the limit in which those shear layer modes start to grow has been
found to vary significantly with the computational domain, matching the variation on the
obtained eigenvalues. This sensitivity on the conditions explains why it was possible to
make predictions on the behavior of 2D modes using a tool that requires a 2D stationary
base flow. As explained in Section 3.2 the DNS used to compute the base flow employs an
artificial compressibility method, with certain boundary conditions, that allow the flow to
remain 2D stable longer than the prediction made in this case.
6Experimental campaign
Contents
6.1 The experimental campaign . . . . . . . . . . . . . . . . . . . . . 87
6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2.1 Validity of eigenmodes in the saturated regime . . . . . . . . . . 89
6.2.2 Effects of three-dimensional boundary conditions . . . . . . . . . 93
6.2.3 On the symmetry breakings . . . . . . . . . . . . . . . . . . . . . 95
6.3 Concluding remarks on the experimental campaign . . . . . . . 95
6.1 The experimental campaign
The experiments were conducted in a recirculating water tunnel at the Laboratory for Tur-
bulence Research for Aerospace & Combustion (LTRAC). All data collection and analysis
on the experimental campaign was made by Dr. Jeremy Basley, in collaboration with the
other authors of the resulting publication, de Vicente et al. [31], mainly Dr. Javier de
Vicente and Dr. Julio Soria, as well as the author of this text.
Regarding the water tunnel, the test section is 5 m long of cross-section 500×500 mm2
and turbulence intensity in the core region is less than 0.35 %, as shown in Parker et al. [74].
The reader may refer to Basley [14] for details on the experimental campaign. This chapter
uses material extracted from de Vicente et al. [31], published in JFM earlier this year. The
focus is here on the results and analyses directly related to the work of the author. The
reader may refer to the original and complete article for an exhaustive description, and for
the main results on the saturated flow. So, In what follows the main results related with
the relation between the linear analysis and the experimental campaign are reproduced.
The experimental set-up is sketched in Figure 6.1. The test plate is mounted vertically
in the middle of the test section. The 50 mm deep, D, 100 mm long, L, cavity spans
the water tunnel and is located 6.34 D from the leading edge of the plate. The distance
from flat plate water-tunnel walls is nominally F = 225 mm, such that F/D = 4.5. The
87
88 6.1 The experimental campaign
θ0 �x
�y
�U
D
L
Figure 6.1: Sketch of the experimental set-up. Dimensions are given in millimetres. The laser
sheet (at y = −0.1D) is represented in a close-up on the L = 2D – shaped cavity. High resolution
images require three cameras to span the cavity.
results presented in the following have been obtained for two mean free-stream velocities
UA = 29.5±0.8 mm/s and UB = 47.0±0.9 mm/s, which correspond to Reynolds numbers
ReD of 1500 and 2400, respectively.
6 Experimental campaign 89
6.2 Discussion
The BiGlobal stability analysis and velocity experimental measurements do not strictly
highlight the same state of the system. Firstly, linear stability analysis (section 2.2)
is concerned with the onset and the nature of the flow instabilities whereas the exper-
iments (section 6.1) deal with the final state of non-linearly saturated dynamics. Fur-
thermore, BiGlobal analysis involves an ideal noiseless two-dimensional basic flow with
periodic spanwise boundary conditions, as opposed to real conditions and confinement of
the experiments, which might change the stability properties of the flow. In addition,
measurements only give access to partial imperfect information, with uncertainties, being
also only a modelization of a real engineering problem to be compared with the complete
three-dimensional structure of the eigenmodes of the flow.
However, from merging those different objects comes a further understanding of the
mechanisms governing the evolution of the system, from the theoretical onset of centrifu-
gal instabilities to the real flow.
6.2.1 Validity of eigenmodes in the saturated regime
The forthcoming discussion mainly relies on figure 6.2, which presents a side-by-side de-
scription of case B (ReD = 2400, θ0 = 0.0340) from both points of view: linear stability
analysis and experimental measurements. Nominal case B presents a greater variety of
linearly unstable modes (figure 4.14) than case A, for which the control parameters are less
critical (ReD = 1500, θ0 = 0.0432). Indeed, case B reveals richer dynamics since it leads
to a flow being a combination of a greater number of different structures. As a result,
the features discussed hereinafter, regarding the intrinsic instabilities in the permanent
regime, apply similarly to case A in a simpler manner.
In the entire figure 6.2 squared letters and dotted annotations in red refer to results
predicted using BiGlobal analysis while circled blue letters and shaded regions symbolise
results from the experimental campaign. Series of dots in 6.2(top) denote the branches of
eigenvalues corresponding to growing disturbances, represented in the β-St plane. Four
of these eigenvalues are identified with letters (A, B, C and D) inside a square. Letter
D corresponds to β = 11.8, the wavenumber of maximum amplification for branch I,
90 6.2 Discussion
while letter A to β = 6.3, the wavenumber associated with the second local maxima of
amplification of the same branch (see figure 4.14 and figure 4.15 for more details).
Letter C points to two different values of β in the most unstable disturbance, Mode II:
β = 12 corresponding to maximum amplification in the stationary branch (II.a) and β =
8.5 just before the bifurcation in the oscillatory branch of the same Mode (II.b). Finally,
letter B indicates another mode of the same branch II.b, but with different properties
(β = 7.8, St = 0.0054).
In the same sub-figure, shades are a qualitative representation of the energy in this
β-St plane according to the space-time Fourier analysis of the experimental data. Dashed
lines, labelled by circled letters A to D, denote four characteristic frequencies of the Fourier
spectrum.
In the bottom part of figure 6.2, the spatial modes associated with the four items A to
D are depicted through their streamwise and spanwise velocity fields. The Fourier modes
(left column) and BiGlobal eigenmodes (right column) are discussed side-by-side. To that
end, velocity profiles have been extracted from a three-dimensional reconstruction of the
eigenmodes in the plane y = −0.1 D (as in the experiments). For the sake of clarity each
BiGlobal mode is only depicted in the area of most resemblance with the experimental
Fourier mode.
In order of increasing Strouhal number, the first modes to be considered are those cor-
responding to steady disturbances (denoted by letter C in figure 6.2). According to linear
analysis there is a range of wavenumbers (9 � β � 19) for which the stationary branch
of Mode II is unstable. In the experiments, wavenumbers for steady structures match the
BiGlobal predictions, but with a tendency to concentrate nearby the endwalls of the rig.
Within the uncertainty of experiments and in real conditions, ”quasi-steady” dynamics
can also be considered with regards to stationary eigenmodes. For instance, the structures
present on the right side of the Fourier mode (C) in figure 6.2(bottom) resemble those pre-
dicted by the linear stability analysis in the stationary branch II.a. The most coherent
and energetic structures correspond to β 12, which is the wavenumber for maximum
amplification according to BiGlobal analysis. Other coherent structures visible near the
left wall in the experiments for the same Fourier mode (C) exhibit a tilted shape, charac-
teristic of travelling waves. These slow-moving structures likely belong to the oscillatory
branch of the same mode (II.b) close to the bifurcation. Indeed, streamwise and spanwise
velocity components of the linearly reconstructed flow present a qualitative morphological
similarity with the experimental data in the region of the cavity where the mode appears.
6 Experimental campaign 91
Figure 6.2: Main flow features at ReD = 2400 for both experimental and linear stability analysis. (Top)
BiGlobal unstable eigenvalues (dots) and qualitative schematic depiction of the most energetic modes in the
experiments (shades) in the β-St plane. (Bottom) Velocity fields related to the four representative modes
highlighted in the upper figure. Left column corresponds to global Fourier modes from the experimental
dataset. Right column presents the reconstructed flow using BiGlobal analysis. For each mode, streamwise
velocity (top) and spanwise velocity (bottom) are shown.
92 6.2 Discussion
However, the coherent structures in the experiments are not dominant in terms of energy,
whereas the corresponding BiGlobal eigenmodes are associated with the largest growth
rates. It is important to point out that the mode with the highest growth rate in linear
analysis does not necessarily have to be the most energetic mode in the saturated regime.
The oscillatory branch II.b remains unstable as the Strouhal number increases, while the
associated spatial wavenumber decreases. This disturbance seems to correspond to a con-
tinuum of modes observed in the experiments. For instance, the dynamics depicted in the
Fourier mode (B) exhibit energetic coherent structures consistent with the eigenmodes (B)
existing in the same range of space-time scales, except for a slight reduction of the Strouhal
numbers. As expected, the velocity fields are morphologically similar to those previously
observed in modes (C), given that it is the same branch with a different wavenumber.
From the BiGlobal analysis, the most linearly unstable oscillatory branch is Mode I, for
both case A (ReD = 1500) and case B (ReD = 2400), corresponding to Strouhal numbers
such that St 0.025. In case B, although the fastest-growing eigenmode is associated
with β = 11.8 and correspond to St = 0.027, a broad range of spanwise waves are actually
unstable (for 4 � β � 18). The reconstruction of the most unstable configuration of this
branch is shown on the right side of figure 6.2(bottom), denoted with letter D. From the
experiments, the frequency band St 0.027 is associated with broad-banded dynamics
involving wavenumbers in the range 5 ≤ |β| ≤ 15. The global Fourier mode (D) seen in
figure 6.2(bottom) exhibits many patterns recalling the spatial structure of various eigen-
modes of branch I. This suggests that the saturated dynamics observed in the experiments
are composed of a continuum of waves deriving from intrinsic instabilities pertaining to
the unstable branch of Mode I.
Finally, the most prominent dynamics revealed by experimental datasets are associated
with Strouhal numbers such that 0.013 ≤St≤ 0.023. In both cases A and B, these domi-
nant features consist of highly coherent right or left travelling waves, corresponding to a
well defined wavelength λ D (|β| 2π). These structures have been encountered as
pairs of counter-propagating waves or as a stand-alone pulsating pattern when and where
two waves overlap. A global Fourier mode of such a travelling wave can be favourably
compared to an eigenmode from the branch I corresponding to the same β = 2π, as seen
with (A) in figure 6.2. In fact, the velocity fields are qualitatively analogous in a wide sec-
tion of the y = −0.1 D plane. However, the Strouhal numbers exhibited by the saturated
dynamics do not correspond to any linearly unstable eigenvalues, neither from branch I
(St 0.025) nor branch II.b (St < 0.010). Such a discrepancy between BiGlobal eigen-
modes and Fourier modes extracted from the real flow in the permanent regime is caused
6 Experimental campaign 93
either by nonlinearities or by the effects of solid boundary conditions on the stability
properties of the base-flow.
6.2.2 Effects of three-dimensional boundary conditions
As opposed to the two-dimensional base-flow around which the linear stability analysis
is performed, the real flow investigated experimentally involves solid boundary conditions
caused by endwalls located at z = ±5D. Such boundary conditions likely lead to the
creation of Bodewadt (Ekman-like) layers of opposite sign near both endwalls. The effect
of endwall layers on the centrifugal instabilities in cavity flows has been observed first
in lid-driven cavity flows (Albensoeder and Kuhlmann [3], Albensoeder et al. [5], Chiang
et al. [24], Koseff and Street [61, 62, 63], Shankar and Deshpande [94]). Endwall layers
are usually modelled as slow-rotating centripetal disks making the junction between the
main recirculation and rigid boundaries. Guermond et al. [48], Migeon et al. [68] notably
demonstrated that Bodewadt layers inject momentum through the centreline of the main
recirculation. In the case of confined flows such as lid-driven cavities, Bodewadt layers
would hence draw the outer edge of the inner-flow from the mid-span region towards the
endwalls. This would imply a spanwise drift of the centrifugal instability vortices, which
coil along the outer region of the main flow, and a consequent increase of the effective
Strouhal number, which is not seen here.
Similar dynamics have been observed by Faure et al. [35, 36] for open cavities of aspect
ratio around L/D ≤ 1.25. For such geometries, the most linearly unstable perturbation
over the two-dimensional base-flow is known to be a stationary disturbance (Bres and
Colonius [20], de Vicente [27], Meseguer-Garrido et al. [66], Pastur et al. [75]). On the
other hand, for larger L/D ratios the more complex geometry of the main recirculation
vortex results in increasing growth rates of oscillatory eigenmodes. Contrary to steady
modes, these unsteady modes derive into intrinsically travelling waves. Effects of endwalls
is hence more difficult to foresee.
A second effect regarding the effect of the walls was described by Shankar and Desh-
pande [94]. The authors observed the discrepancy between the 3-D and 2-D velocity
profiles for increasing Reynolds numbers due to the influence of endwall vortices. These
vortices not only provoke an increase in the spanwise flow but also slow down the main
centrifugal recirculation on the cavity, and that braking increases with Reynolds number.
So, the confinement causes the decrease in the velocity of the centrifugal perturbation,
forcing the stability properties of the base-flow to change. Indeed, Bres and Colonius [20]
94 6.2 Discussion
have asserted that, at the first order, the Strouhal numbers associated with oscillatory
eigenmodes are conditioned by the time required for a perturbation to travel along the
recirculation. From that, a deceleration due to Bodewadt layers could decrease the intrinsic
frequency of the spanwise waves coiling onto the recirculation, despite the influence of
the drift. This effect would become stronger for waves closer to the endwalls. Such an
hypothesis could explain the different Strouhal numbers between BiGlobal eigenmodes and
Fourier modes for the most energetic experimental mode (A) in figure 6.2. In figure 6.3 a
comparison between the streamwise velocity profiles in the experimental y/D = −0.1 plane
for the base flow for the linear analysis and the mean flow in the experiments. Both case
A and B are shown, taking into account the uncertainties of the measures. It can be seen
that, as expected, the experimental values of the velocity are smaller, suggesting a braking
in the main vortex. Also, this reduction of the velocity is greater in the higher Reynolds
case, which is consistent with the findings in Shankar and Deshpande [94]. This means
that the braking phenomenon is a plausible explanation to the reduced Strouhal numbers
reported in the experiments, although other possibilities are explored in Chapter 7.
0 0.5 1 1.5 2
−0.02
0
0.02
0.04
0.06
0.08
0.1
x/D
U
U0
2D Base−flow (BiGlobal analysis)3D mean−flow (experiments)
0 0.5 1 1.5 2
−0.02
0
0.02
0.04
0.06
0.08
0.1
x/D
U
U0
2D Base−flow (BiGlobal analysis)3D mean−flow (experiments)
Figure 6.3: Streamwise profiles of streamwise velocity U/U0 for case A (ReD = 1500) on the left
and case B (ReD = 2400) on the right. The profile obtained of the 2D base-flow used by BiGlobal
analysis is extracted from the range −0.12 � y/D � −0.09 (black), to represent the uncertainty on
the position and thickness of the laser-sheet. The profile issued of the 3D mean-flow, experimentally
measured in the zx-plane at y = −0.1D is extracted from the range −3 � z/D � 3 (blue), to take
into account spanwise variations.
6 Experimental campaign 95
6.2.3 On the symmetry breakings
The nonlinearly saturated flow exhibits asymmetries, which are, by definition, absent of
the (periodic) eigenmodes obtained through BiGlobal analysis. In particular, the question
of why the counter-propagating dominant waves are not symmetrical can arise.
Two plausible explications arise to justify this phenomenon. The symmetry breaking
may be caused by facility-dependent effects or to be inherent to the sensitivity of the flow
to initial conditions.
As always, experimental conditions are characterised within uncertainties. The sources
of possible experimental bias that could lead to spanwise asymmetries are: (a) an imperfect
cavity geometry, or a crooked velocity profile due to (b) water-tunnel design or (c) angular
discrepancy in cavity orientation.
Uncertainties (a) & (b) constitute systematic biases. They are ruled out since asymme-
try changes from one recording to another. On the other hand, (c) is concerned with the
alignment of the rig with z-axis (see de Vicente et al. [31] for a sketch of the setup). That
alignment could vary by about ±1 mm over the span S = 500 mm, corresponding to an
angular discrepancy of ±0.11°. Such an uncertainty of only ±0.1% cannot explain alone
the symmetry breaking. Consequently, an intrinsic sensitivity of the dynamics should
rather be considered.
In fact, the reconstructed flow obtained from stability analysis gives no prevalence to
left-travelling, right-travelling or pulsating structures as it was explained in section 4.1.1.
The smallest variation in the initial conditions hence causes the real flow to break sym-
metries by selecting a particular pattern.
6.3 Concluding remarks on the experimental campaign
The main goal of this section is to use both linear stability analysis and experiments to
cover the evolution of centrifugal instabilities in an open cavity flow from their onset to
their observation within the nonlinearly saturated state. The intrinsic stability properties
of the 2D base-flow were fully investigated through an extensive study of the parameter
space in Chapter 4 and the features changed by nonlinear effects or/and real boundary
conditions are identified and studied.
96 6.3 Concluding remarks on the experimental campaign
Overall, for the ReD = 2400 case the range of wavenumbers corresponding to growing
perturbations is as broad as 2.6 � β � 19, showing that this is a point in the parameter
space far from the stability threshold. The three-dimensional organisation associated with
each family of eigenmodes has been identified and characterised, thus allowing to determine
in advance some of the main agents involved in the real flow with low computational cost.
On the other hand, the experimental investigation of the real flow in the permanent
regime brings more insight into the dynamics that are actually selected by the real flow
beyond the linear transient growth, and once real boundary conditions are set, such as
a noisy incoming flow, lateral walls imposed to the cavity or even the influence of the
wake. Experimental measurements of centrifugal instabilities remain challenging in open
cavities, since those three-dimensional dynamics involve particularly low frequencies and
are greatly perturbed by the normally unstable shear layer above the cavity. Here, time-
resolved high-resolution PIV measurements were performed in a spanwise plane parallel
to the bottom of the cavity. Applying space-extended time-Fourier transform to such
experimental data allowed the identification of the coherent structures associated with
any given Strouhal number. The hypothesis of spanwise-waves has been confirmed by
experimental results, with dynamics in the range of unstable wavenumbers predicted by
linear stability analysis. Most of the eigenmodes were recovered within the real flow, in
spite of different lateral boundary conditions.
It must be noted that eigenmodes are recovered only locally, and that they can be
distorted. Indeed, the saturated dynamics are strongly modulated in amplitude, resulting
in local states, and those states change with time, that is, spanwise waves can also become
more or less dominant in terms of energy at different times. In BiGlobal analysis, travelling
eigenmodes have some degrees of freedom: the composition of the real and imaginary part
of the eigenmodes can form structures that travel right, left, or that pulsate without
shifting. Additional conditions in the real flow, confinement, noise in the upstream flow,
etc. take away that degree of freedom, and a concrete structure is formed. In the case of
several of those modes locally coexisting, they can overlap to produce interferences and
standing waves can appear.
In the saturated flow the fastest growing family of steady eigenmodes (II.a) is fairly
recovered as broad-banded spatial structures associated with the slowest dynamics (for
St→ 0). Similarly, the branch of Mode (I) corresponds to broad-banded dynamics at
0.025 � St � 0.03. On the other hand, spanwise waves associated with Strouhal num-
bers St� 0.01 exhibit narrower ranges of space-scales, which is consistent with eigen-
modes from branch (II.b). The most energetic travelling waves observed in the permanent
6 Experimental campaign 97
regime, though, partially depart from the linear stability results. Those highly coher-
ent waves strikingly resemble the eigenmodes from the unsteady low β branch of Mode
(I). However, they are associated with time-scales such that 0.013 � St � 0.023, lower
than the Strouhal numbers predicted for those structures by the linear stability analy-
sis. The hypothesis presented here to explain those different time-scales relies upon the
modification of the base-flow due to confinement effects. The presence of endwalls could
be responsible for a braking of the main recirculation, leading to slower travelling waves
in the three-dimensional base-flow, relatively to the two-dimensional base-flow, used by
BiGlobal stability analysis.
7Preliminary study on saturation
Contents
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2.2 Computational Setup . . . . . . . . . . . . . . . . . . . . . . . . 102
7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3.1 Regime I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3.2 Regime II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3.3 Regime III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.3.4 Regime IV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.3.5 Regime V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.4 Concluding remarks on the study on saturation . . . . . . . . . 109
7.1 Introduction
Section 6.2.2 elaborated on the possibility that the apparent reduction of the characteristic
frequencies of the most energetic eigenmode from the theoretical value predicted by the
linear analysis is a consequence of the presence of the span-wise walls, which had the
effect of slowing down the main centrifugal recirculation of the cavity, thus reducing the
characteristic Strouhal number of these structures. Other possible explanations for this
phenomenon are the saturation of the flow, or the non-linear interaction between several
unstable eigenmodes.
A preliminary study was presented in Meseguer-Garrido et al. [65], trying to separate
these three effects. A three-dimensional DNS computation was performed for the same
flow parameters of cases A and B of de Vicente et al. [31], as presented in Chapter 6, but
with periodic boundary conditions. Thus, the effect of the end-walls was neglected, and the
restriction on the span-wise wavenumber β limited the number of eigenmodes interacting,
leaving the saturation as the main mechanism present in the study. The main finding of
99
100 7.2 Methodology
that paper is that the reduction of characteristic Strouhal number reported in de Vicente
et al. [31] occurs also in absence of span-wise walls. The work of Meseguer-Garrido et al.
[65] relied on the study of instantaneous flow-fields (snapshots) extracted from a three-
dimensional DNS computation, as well as on the evolution of the flow variables at one
point. A deeper study of the data is however required to fully understand the physics
behind it, as no conclusions could be reached once the flow became complex enough.
In the present chapter, the original DMD algorithm Schmid [92] is applied to the incom-
pressible fluid flow over a rectangular open cavity from the linear to the saturated regime,
as detailed in Meseguer-Garrido et al. [65], and in section 3.5, in order to understand
the evolution of span-wise instabilities of the flow and the interactions between different
dynamic modes. The numerical solutions required to construct the data-sequences of snap-
shots were obtained by means of a three-dimensional non-steady DNS solver, explained
in section 3.4. The most relevant DMD modes and associated oscillation frequencies are
then compared to the ones obtained using linear stability analysis, allowing us to assess
the accuracy of the aforementioned snapshot-based decomposition. The numerical details
of the application of the different methods used for the present investigation are explained
in section 7.2. The BiGlobal and DNS results, and the DMD analysis performed are pre-
sented in section 7.3. Finally, the most significant conclusions obtained are summarized
in section 7.4.
7.2 Methodology
7.2.1 Problem description
A schematic representation of the flow configuration was depicted in figure 4.3. The case
studied here corresponds with the experimental Case B, with ReD = 2400 and θ0/D =
0.036 in a cavity of L/D = 2, as explained before.
As it was explained in Chapters 4 and 6 in the range of parameters close to the limit
of stability, the linear analysis performed shows the presence of two main branches of
unstable eigenmodes, as can be seen in figure 7.1 (top). The mode that becomes unstable
at lower Reynolds number, Mode I, is a travelling disturbance which is more unstable in
the proximity of β 6 and β 12. Mode II, the second to become unstable, is stationary
at higher β, while undergoing a bifurcation at β 9, resulting in a pair of complex
conjugate eigenvalues for values of β lower than that. The range of unstable eigenvalues
7 Preliminary study on saturation 101
of both Mode I and II in the St-β plane, for the chosen parameters in this study, can be
seen in figure 7.1 (bottom).
Figure 7.1: Neutral curves for the L/D = 2 cavity in the ReD vs β plane, and selected Case
B (top). StD vs β map of unstable eigenmodes for Case B, and selected β by the periodicity
conditions of the DNS computations (bottom).
102 7.2 Methodology
7.2.2 Computational Setup
The DNS code employed to compute the flow in the mentioned cases is detailed in sec-
tion 3.4. As it was previously mentioned, the span-wise length of the computational
domain was selected in order to reduce the amount of interaction between the different
modes. By using a Lz = 2π/6 only the modes of spanwise wavenumber, β, multiple of 6
can appear, and those correspond with the β maximum amplification of the linear modes
(β = 6 and β = 12), as can be seen in figure 7.1 (bottom). The amplification of the modes
on β = 18 is very small, so it is unlikely that they appear in the simulation. Also, span-
wise periodic boundary conditions were used to avoid the effect of the span-wise walls, as
it was previously explained.
The chosen flow conditions were the same as in the linear analysis, with the exception
of the Mach number, i.e. M= 0.1. The results obtained in the linear growth regime
with this low Mach number are identical to those obtained with the incompressible linear
analysis, as can be seen in Meseguer-Garrido et al. [65], and in section 4.2.2. A block
structured mesh with 104 hexahedral subdomains was used, each of those subdomains
with (nx, ny, nz) = 15. For the chosen conditions the two-dimensional stationary solution
was extended periodically in the whole domain, and then a random noise of 10−8 was
introduced to kick-start the growth of the linearly unstable modes. The flow variables in
a point in the middle of the cavity were recorded, and snapshots of the whole flowfield
were saved each 10 non-dimensional time units. The L∞-norm of the span-wise velocity
component of the perturbed flow is monitored as a function of non-dimensional time on
the control point to determine the different regimes of the flow. Since the value of said
velocity component is zero in the two dimensional base flow, the whole effect corresponds
to perturbation, allowing to see the modes in greater detail, without dealing with the base
flow.
In the investigated fluid flow over a rectangular open cavity, five instability regions of
span-wise velocity component are identified from the linear to the saturated regime, as
reported in Meseguer-Garrido et al. [65]. Those five regimes can be seen in figure 7.2.
A DMD (see section 3.5 for details on this tool) was thus applied to regions II to V.
Convergence for each regime was accepted when the residual norm of the DMD and the
eigenvalues corresponding to the most unstable dynamic modes became independent of
the number of snapshots comprising the matrix V N1 (see section 3.5).
7 Preliminary study on saturation 103
Figure 7.2: Temporal evolution of the absolute value of the span-wise velocity component in the
control point.
7.3 Results
7.3.1 Regime I
The preliminary study conducted in Meseguer-Garrido et al. [65] shows that the behavior
of the DNS in the linear growth phase (region I of figure 7.2) matches perfectly with the
results predicted by the linear analysis, with the most unstable eigenmode appearing and
growing exponentially with the anticipated σ. Those results can be seen in section 4.2.2.
7.3.2 Regime II
After the linear growth phase there is a saturation of the leading mode (t 500), and
then the stationary mode starts to pulsate with a StD of 0.0098 in regime II. The physical
morphology of the mode does not change, and the oscillations gets dampened with time.
The Strouhal number of these oscillations is the same as the StD of the low β branch
of Mode II (StD = 0.0099), according to linear analysis. So the behavior seems to be
104 7.3 Results
Figure 7.3: DMD modes on regime II. On the top, situation in the StD vs β plane (left) of DMD
modes A and B (right). On the bottom, BiGlobal mode corresponding with point A (Mode II for
β = 12).
that after saturation the stationary mode starts vibrating. That vibration is progressively
dampened, and in this case, with strong selection of Lz the perturbations resonate at
the characteristic frequency of the same branch of said mode, in the low β regime, even
though the structures do not change their characteristic β = 12. This is a very interesting
behavior, as the mode tends to vibrate with a frequency which is somehow natural to
it (although for a different β), even though for this β it is inherently stationary. The
7 Preliminary study on saturation 105
DMD analysis of regime II can be seen in figure 7.3. It shows a strong presence of a
stationary mode (A), with a pulsation of the same structure (B). The morphology of
these structures is qualitatively identical to those of the stationary branch of the Biglobal
analysis (figure 7.3 top right and bottom, respectively).
7.3.3 Regime III
After the oscillations characterizing regime II have almost completely dampened a new
mode starts to grow in regime III (t 1500 to t 2100). Figure 7.4 (top) shows two
instantaneous flowfields belonging to that period of time, compared with the linear com-
position of the low β branch of Mode I and the high β branch of Mode II of linear analysis
(bottom). As it was explained in section 4.1.1 the lack of normalization on the eigenvector
makes so the values of the constants in the composition gives no additional information, in
this case there is four degrees of freedom, corresponding to two pairs of α, and they were
chosen to make that composition as similar as possible to the DNS flow. The oscillating
mode has β = 6 and StD = 0.0194, by determining the frequency on figure 7.2. That
frequency is smaller than the predicted one from linear analysis (StD = 0.025), but the
identified mode is the same Mode I, as can be seen in figure 7.4. As it was previously
discussed, the reduction of the characteristic frequency of this dominant mode of β = 6
was already observed in the previous experimental work of de Vicente et al. [31]. These
observations point to the fact that the saturation process and the non-linear interaction of
the modes is enough to produce the St reduction, without the presence of spanwise walls.
Figure 7.5 shows the results obtained with DMD for the same regime III. In this case
the short period of time and limited sampling frequency available, coupled with the nature
of the growth of the second mode, makes the DMD convergence not as good as in other
cases. In any case, the two main structures that can be appreciated in this regime are the
aforementioned mode A, still stationary and strongly resembling the stationary branch
of BiGlobal Mode II, and a structure pulsating with StD = 0.019 (mode C in figure 7.5)
that bears some similarity with the linear structure of β = 6 of Mode I (shown in the left
part of figure 7.4 bottom). In this case the mere composition of linear modes is closer to
an instantaneous flowfield than the DMD modes, due to, as it was previously mentioned,
poor convergence motivated by the short length of this regime, and the varying nature of
mode C.
106 7.3 Results
Figure 7.4: Two instantaneous flowfields in region III (top). Composition of the two linear modes
that yields a similar flowfield (bottom).
7.3.4 Regime IV
The next step in the DNS evolution, regime IV (t 2100 to t 2900), features the
apparition of a secondary oscillation in counter-phase with the main one, producing a
shifting of the frequency to StD = 0.015. This regime is at first stable in amplification, until
t 2700. In this stage the structures of the isocontours of spanwise velocity component
no longer resemble easily constructable combinations of linear modes. This last frequency
is once more in the range of frequencies of high energy in the experiments of Chapter 6
and de Vicente et al. [31]. The results of the DMD analysis in this regime can be seen in
figure 7.6, and allow the separation of the features of the flow into identifiable structures.
Mode A still resembles the stationary branch of the linear mode, although the mode has
7 Preliminary study on saturation 107
Figure 7.5: DMD modes on regime III. Situation in the StD vs β plane of DMD modes A and C.
Figure 7.6: DMD modes on regime IV. Situation in the StD vs β plane of DMD and modes A,
C, E and D.
varied morphologically. Since the structures coiling around the main recirculating vortex
have started breaking it, the modes are no longer restricted to coil around the exterior
108 7.3 Results
part of the vortex, but fill a greater fraction of the cavity space. A similar phenomenon
occurs to mode C, which still corresponds, as before, with the linear mode in figure 7.4,
but this time oscillating with a frequency of StD = 0.015 (also within the range of StD
of the dominant mode in de Vicente et al. [31]). Other structures appear here, one of
StD = 0.030, and β = 12, which corresponds with the aforementioned counter-phase
oscillation (mode E), and another one in which two different β can be seen (mode D), and
with a frequency in line with the linear mode, StD = 0.025. Since DMD does a separation
by frequency, not by spatial structure, the two structures which oscillate with StD = 0.025,
one with β = 6 and the other with β = 12, get combined into a single DMD mode. It
is reasonable to assume that the linear composition of different structures of the linear
Mode I of different β could produce structures similar to mode D. On the other hand those
two modes could correspond with structures related with the secondary instability of the
saturated flow, fruit of non-linear interaction between the different structures. It seems
relevant that the appearance of a mode of StD = 0.030, which does not appear in the
linear analysis, nor it was present in the experiments conducted in de Vicente et al. [31],
coincides with the reduction of the characteristic frequency of mode C from StD = 0.019
to StD = 0.015, which is exactly half of the one of mode E. There seems to be some
sort of coupling between those modes in frequency, with stable amplitude. That constant
amplitude changes after t 2700, leading to regime V.
7.3.5 Regime V
After the breaking of the amplitude-stable counter-phase oscillation on t 2700 another
change occurs, as the structures trespass the x−y plane in the middle of the computational
domain. This produces a drastic increase in the complexity of the structures present in
the flow, as different packs of waves start travelling left or right, instead of just staying
stationary, or pulsating without changing position. These complex structures character-
ize the behavior of regime V (t 2900 onwards). The possibility of having travelling
structures appears in the linear analysis, as the not travelling ones are just a single com-
bination of the complex conjugate eigenvectors, but once the flow starts developing this
kind of structures, those shift in short periods of time. This might explain the not so good
convergence of the DMD technique for the present regime. The physics of the problem still
select some natural frequencies, so the DMD modes are extracted with constant Strouhal
numbers, but the shape of the modes varies greatly. Nevertheless, the results of the DMD
of regime V can be seen in figure 7.7, and some of the familiar structures are still present,
like the stationary mode A, the dominant mode C, or mode E. The nature of the mode on
7 Preliminary study on saturation 109
Figure 7.7: DMD modes on regime V. Situation in the StD vs β plane of DMD and modes A, C,
E and F.
the StD = 0.021 to StD = 0.025 band experiences the most salient changes, as structures
of all the possible β (0, 6 and 12) merge in a single mode. An example of such is mode F
in figure 7.7.
7.4 Concluding remarks on the study on saturation
This work shows the power of the DMD tool to analyze a complex problem. The separation
of structures provided by this technique allows to identify the two most relevant linear
modes well beyond the linear regime of the DNS computation. The description of the
dominant structures in the five different flow regimes allows to better understand how the
process of saturation affects the morphology of the modes, as well as their characteristic
frequencies, without having to deal with the effect of the spanwise walls, unavoidable in an
experimental setup. The reported reduction of the characteristic Strouhal number of the
dominant mode from the value given by the linear analysis is found again here, indicating
that the effect is either a byproduct of the change of the secondary instability, or an effect
of the non-linear interaction of the different modes.
8Summary
and Future Directions
In the present thesis the BiGlobal instability analysis tool has been used to characterize
the onset of instability in the incompressible open cavity flow. Both two-dimensional and
spanwise periodic three-dimensional have been identified and thoroughly described. For
a wide range of the parameters of the problem the features of the leading perturbations
have been studied, allowing to construct expressions that relate the different parameters
and characteristics of the flow, as well as describe the neutral hypersurface through cuts
with several planes on the paremeter space.
Two main three-dimensional modes appear in the majority of the cases studied: a bi-
furcated mode, noted Mode II, with both a travelling and a stationary (and more unstable)
lobe, as well as a purely travelling mode (Mode I). The stability behaviour of those modes
with variations of all the parameters has been reported, allowing to predict the stability
of any configuration sufficiently close to critical conditions using only geometrical features
of the cavity. In particular, the relations between the characteristic Strouhal number of
the different eigenmodes, their critical Reynolds number and the L/D ratio have been
characterised. Such information should useful not only in order to predict instabilities of
a given open cavity, but also as basis for reduced-order-models of global flow instability
and subsequent flow control, both theoretical and experimental.
Also, the use of experiments has allowed to determine the validity BiGlobal tool, linear
in nature, well outside the linear regime, and the features changed by nonlinear effects
or/and real boundary conditions were identified and studied. The study of the dynamics
on the saturated regime was continued through a preliminary study on saturation through
numerical tools, using a three-dimensional DNS to compute a simple periodic case, and
using DMD to extract the temporal behaviour of the flow during saturation. The reported
reduction of the characteristic Strouhal number of the dominant mode from the value given
by the linear analysis is found again here, indicating that the effect is either a byproduct
of the change of the secondary instability, or an effect of the non-linear interaction of the
different modes.
112
Regarding future directions of this work, the study of the dynamics of saturation is
going to be followed, using three-dimensional DNS simulation that are wall bounded in
the spanwise direction, and also periodic with different lengths. Using those different
conditions the true nature of the decrease in Strouhal number should be determined. The
identification of the secondary instabilities is also being studied, first through the use of a
BiGlobal code that allows for a base flow with the three velocity components, and using
an average flow of the three-dimensional saturated flow, and in future works, through the
use of a TriGlobal tool, or at least a BiGlobal with a fourier transformation in the base
flow. Lastly, efforts are being made in regard to the two-dimensional instabilities, and the
sensibility to the domain and boundary conditions.
Bibliography
[1] N. Abdessemed. Stability analysis of flow past a low-pressure turbine blade. PhD
Thesis. Imperial College, 2007.
[2] N. Abdessemed, S. J. Sherwin, and V. Theofilis. Linear instability analysis of low-
pressure turbine flows. J. Fluid Mech., 628:57–83, 2009.
[3] S. Albensoeder and H. C. Kuhlmann. Nonlinear three-dimensional flow in the lid-
driven square cavity. J. Fluid Mech., 569:465–480, 2006.
[4] S. Albensoeder, H. C. Kuhlmann, and H. J. Rath. Three-dimensional centrifugal-
flow instabilities in the lid-driven-cavity problem. Physics of Fluids (1994-present),
13(1):121–135, 2001.
[5] S. Albensoeder, H. C. Kuhlmann, and H. J. Rath. Multiplicity of steady two-
dimensional flows in twosided lid-driven cavities. Theor. Comp. Fluid Dynamics, 14:
223–241, 2001.
[6] F. Alizard, J. C. Robinet, and X. Gloerfelt. A domain decomposition matrix-free
method for global linear stability. Comput. and Fluids, 66:63–84, 2012.
[7] D.N. Arnold, F. Brezzi, B. Cockburn, and L.D. Marini. Unified analysis of discontin-
uous Galerkin methods for elliptic problems. SIAM Journal of Numerical Analysis,
39(5):1749–1779, 2001.
[8] S. Bagheri. Koopman-mode decomposition of the cylinder wake. Journal of Fluid
Mechanics, 726:596–623, 2013.
[9] S. Bagheri, E. Akervik, L. Brandt, and D. S. Henningson. Matrix-free methods for
the stability and control of boundary layers. AIAA J., 47(5):1057–1068, 2009.
[10] A. Barbagallo, D. Sipp, and P. J. Schmid. Closed-loop control of an open cavity flow
using reduced-order models. Journal of Fluid Mechanics, 641:1–50, 11 2009. ISSN
1469-7645.
113
114 BIBLIOGRAPHY
[11] D. Barkley. Linear analysis of the cylinder wake mean flow. Europhysics Letters, 75
(5):750–756, 2006.
[12] D. Barkley and R. D. Henderson. Three-dimensional floquet stability analysis of the
wake of a circular cylinder. J. Fluid Mech., 322:215–241, 1996.
[13] D. Barkley, H. M. Blackburn, and S. J. Sherwin. Direct optimal growth analysis for
timesteppers. Int. J. Numer. Meth. Fluids, 57(9):1435–1458, July 2008.
[14] J. Basley. An Experimental Investigation on Waves and Coherent Structures in a
Three-Dimensional Open Cavity Flow. PhD thesis, Universite Paris-Sud – Monash
University, 2012.
[15] J. Basley, L. R. Pastur, F. Lusseyran, T. M. Faure, and N. Delprat. Experimental
investigation of global structures in an incompressible cavity flow using time-resolved
PIV. Exp. Fluids, 50(4):905–918, April 2011. DOI: 10.1007/s00348-010-0942-9.
[16] J. Basley, L. R. Pastur, N. Delprat, and F. Lusseyran. Space-time aspects of a
three-dimensional multi-modulated open cavity flow. Physics of Fluids, 25(6), 2013.
doi: 10.1063/1.4811692.
[17] J. Basley, L. R. Pastur, F. Lusseyran, J. Soria, and N. Delprat. On the modu-
lating effect of three-dimensional instabilities in open cavity flows. J. Fluid Mech,
Submitted and under revision, 2014.
[18] F. P. Bertolotti, Th. Herbert, and P. R. Spalart. Linear and nonlinear stability of
the Blasius boundary layer. J. Fluid Mech., 242:441–474, 1992.
[19] G. A. Bres. Numerical simulations of three-dimensional instabilities in cavity flows.
PhD thesis, California Institute of Technology, 2007.
[20] G. A. Bres and T. Colonius. Three-dimensional instabilities in compressible flow
over open cavities. J. Fluid Mech., 599:309–339, 2008.
[21] K. Butler and B. F. Farrell. Three-dimensional optimal perturbations in viscous
shear flow. Phys. Fluids, 4(8), 1992.
[22] L. N. Cattafesta III, S. Garg, M. S. Kegerise, and G. S. Jones. Experiments on
compressible flow-induced cavity oscillations. AIAA J., page 2912, 1998.
[23] K. Chen, J. Tu, and C. Rowley. Variants of dynamic mode decomposition: boundary
condition, koopman, and fourier analyses. Journal of Nonlinear Science, 22(6):887–
915, 2011.
BIBLIOGRAPHY 115
[24] TP Chiang, WH Sheu, and RR Hwang. Effects of the reynolds number on the eddy
structure in a lid-driven cavity. Int. J. Numer. Methods Fluids, 26:557–579, 1998.
[25] T. Colonius, C. W. Rowley, and V. Theofilis. Global instabilities and reduced-order
models of cavity flow oscillations. In V. Theofilis, T. Colonius, and A. Seifert, editors,
Proceedings of the AFOSR/EOARD/ERCOFTAC SIG-33: Global Flow Instability
and Control Symposium - I, Sept. 22-26, 2001, Crete, Greece, Sept. 2001. ISBN-13:
978-84-692-6244-3. Fundacion General UPM. ISBN 84-89925-65-8.
[26] J. D. Crouch, A. Garbaruk, and D. Magidov. Predicting the onset of flow unsteadi-
ness based on global instability. J. Comput. Phys., 224:924–940, 2007.
[27] J. de Vicente. Spectral Multi-Domain Method for the Global Instability Analysis of
Complex Cavity Flows. PhD thesis, Universidad Politecnica de Madrid, 2010.
[28] J. de Vicente, E. Valero, L. M. Gonzalez, and V. Theofilis. Spectral multi-domain
methods for biglobal instability analysis of complex flows over open cavity configu-
rations. 34th Fluid Dynamics Conference and Exhibit, Portland, Oregon, June 28 –
July 1 2004. AIAA Paper 2004-2544.
[29] J. de Vicente, D. Rodriguez, V. Theofilis, and E. Valero. Stability analysis in
spanwise-periodic double-sided lid-driven cavity flows with complex cross-sectional
profiles. Comput. and Fluids, 43(1):143–153, 2010.
[30] J. de Vicente, P. Paredes, E. Valero, and V. Theofilis. Wave-like disturbances on the
downstream wall of an open cavity. Number 3754 in 6th AIAA Theoretical Fluid
Mechanics Conference, Honolulu, Hawaii, June 27-30 2011. AIAA.
[31] J. de Vicente, J. Basley, F. Meseguer-Garrido, J. Soria, and V. Theofilis. Three-
dimensional instabilities over a rectangular open cavity: from linear analysis to
experimentation. J. Fluid Mech., 748:189–220, 2014.
[32] P. G. Drazin and W. H. Reid. Hydrodynamic Stability. Cambridge University Press,
1981.
[33] D. Duke, J. Soria, and D. Honnery. An error analysis of the dynamic mode decompo-
sition. Experiments in Fluids, 52(2):529–542, 2012. doi: 10.1007/s00348-011-1235-7.
[34] B. F. Farrell and P. J. Ioannou. Generalized stability theory. Part I: Autonomous
operators. J. Atmos. Sci., 53(14):2025–2040, 1996.
116 BIBLIOGRAPHY
[35] T. M. Faure, P. Adrianos, F. Lusseyran, and L. R. Pastur. Visualizations of the flow
inside an open cavity at medium range reynolds numbers. Experiments in Fluids,
42(2):169–184, 2007.
[36] T. M. Faure, L. R. Pastur, F. Lusseyran, Y. Fraigneau, and D. Bisch. Three-
dimensional centrifugal instabilities development inside a parallelepipedic open cav-
ity of various shape. Experiments in Fluids, 47(3):395–410, 2009.
[37] E. Ferrer, J. DeVicente, and E. Valero. Low cost 3d global instability analysis and
flow sensitivity based on dynamic mode decomposition and high order numerical
tools. International Journal for Numerical Methods in Fluids, 2014.
[38] N. Fietier and M. O. Deville. Time-dependent algorithms for the simulation of vis-
coelastic flows with spectral element methods: applications and stability. J. Comput.
Phys., 186 (1):93–121, 2003.
[39] M. Gaster. A note on the relation between temporally increasing and spatially
increasing disturbances in hydrodynamic instability. J. Fluid Mech., 14:222–224,
1962.
[40] M. Gharib and A. Roshko. The effect of flow oscillations on cavity drag. J. Fluid
Mech., 177:501–530, 1987.
[41] X. Gloerfelt. Aerodynamic noise from wall-bounded flows: Cavity noise, chapter 0.
VKI Lectures, 2009.
[42] F. Gomez, S. L. Clainche, P. Paredes, M. Hermanns, and V. Theofilis. Four decades
of studying global linear instability: Progress and challenges. AIAA Journal, 50(12):
2731–2743, December 2012. doi: 10.2514/1.J051527.
[43] F. Gomez, R. Gomez, and V. Theofilis. On three-dimensional global linear instability
analysis of flows with standard aerodynamics codes. Aerosp. Sci. Techn., 32(1):223–
234, 2014.
[44] L. Gonzalez, V. Theofilis, and R. Gomez-Blanco. Finite element methods for viscous
incompressible biGlobal instability analysis on unstructured meshes. AIAA J., 45
(4):840–854, 2007.
[45] L. Gonzalez, V. Theofilis, and S. J. Sherwin. High-order methods for the numerical
solution of the BiGlobal linear stability eigenvalue problem in complex geometries.
Int. J. Numer. Meth. Fluids, (to appear):DOI:10.1002/fld.2220, 2010.
BIBLIOGRAPHY 117
[46] L. M. Gonzalez, M. Ahmed, J. Kuhnen, H. C. Kuhlmann, and V. Theofilis. Three-
dimensional flow instability in a lid-driven isosceles triangular cavity. Journal of
Fluid Mechanics, 675:369–396, 5 2011. ISSN 1469-7645.
[47] P. M. Gresho and R. L. Sani. On pressure boundary conditions for the incompressible
navier-stokes equations. International Journal for Numerical Methods in Fluids, 7
(10):1111–1145, 1987. ISSN 1097-0363.
[48] J.-L. Guermond, C. Migeon, G. Pineau, and L. Quartapelle. Start-up flows in a
three-dimensional rectangular driven cavity of aspect ratio 1:1:2 at re=1000. J.
Fluid Mech., 450:169–199, 2002.
[49] R. D. Henderson and D. Barkley. Secondary instability of the wake of a circular
cylinder. Phys. Fluids, 8(6):65–112, 1996.
[50] T. Herbert. Secondary instability of boundary layers. Annu. Rev. Fluid Mech., 20:
487–526, 1988.
[51] T. Herbert. Parabolized stability equations. Annual Review of Fluid Mechanics, 29:
245 – 283, 1997.
[52] C. P. Jackson. A finite-element study of the onset of vortex shedding in flow past
variously shaped bodies. J. Fluid Mech., 182:23–45, 1987.
[53] G. B. Jacobs, D. Kopriva, and F. Mashayek. Validation study of a multidomain
spectral code for simulation of turbulent flows. AIAA journal, 43(6):1256–1264,
2005.
[54] E. Janke and P. Balakumar. On the secondary instability of three-dimensional
boundary layers. Theor. Comp. Fluid Dynamics, 14:167–194, 2000.
[55] M. Jovanovic, P. Schmid, and J. Nichols. Sparsity-promoting dynamic mode decom-
position. Physics of Fluids, 26(2):024103 (22 pages), 2014.
[56] G. Em Karniadakis and S. J. Sherwin. Spectral/hp element methods for CFD. OUP,
2005.
[57] M. A. Kegerise, E. F. Spina, S. Garg, and L. N. Cattafesta III. Mode-switching and
nonlinear effects in compressible flow over a cavity. Phys. Fluids, 16:678–687, 2004.
[58] M. H. Kobayashi, J. C. F. Pereira, and J. M. M. Sousa. Comparison of several
open boundary numerical treatments for laminar recirculating flows. Int. J. Numer.
Meth. Fluids, 16(5):403–419, 1993.
118 BIBLIOGRAPHY
[59] D. Kopriva. Implementing Spectral Methods for Partial Differential Equations: Al-
gorithms for Scientists and Engineers. Springer, 2009.
[60] D. A. Kopriva. A staggered-grid multidomain spectral method for the compressible
navier-stokes equations. J. Comput. Phys., 244:142–158, 1998.
[61] J. R. Koseff and R. L. Street. The lid-driven cavity flow: A synthesis of qualitative
and quantitative observations. Trans. ASME: J. Fluids Eng., 106:390–398, 1984.
[62] J. R. Koseff and R. L. Street. On endwall effects in a lid-driven cavity flow. Trans.
ASME: J. Fluids Eng., 106:385–389, 1984.
[63] J. R. Koseff and R. L. Street. Visualization studies of a shear driven three-
dimensional recirculating flow. Trans. ASME: J. Fluids Eng., 106:21–29, 1984.
[64] R. Lehoucq, D. Sorensen, and C. Young. ARPACK Users’ Guide: Solution of Large
Eigenvalue Problems with Implicitly Restarted Arnoldi Method. SIAM Philadelphia,
1998.
[65] F. Meseguer-Garrido, J. de Vicente, and E. Valero. Three-dimensional analysis of
incompressible flow over an open cavity using direct numerical simulation: from
linear to saturated regime. 5980, Prato, Italy, June 5-8 2014. Instability and Control
of Massively Separated Flows.
[66] F. Meseguer-Garrido, J. de Vicente, E. Valero, and V. Theofilis. On linear instability
mechanisms in the incompressible open cavity flow. J. Fluid Mech., 752:219–236,
2014.
[67] I. Mezic. Analysis of fluid flows via spectral properties of the koopman operator.
Annual Review of Fluid Mechanics, 45:357–378, 2013.
[68] C. Migeon, G. Pineau, and A. Texier. Three-dimensionality development inside
standard parallelepipedic lid-driven cavities at re=1000. J. Fluids Struct., 17:717–
738, 2003.
[69] M. Morzynksi and F. Thiele. Numerical stability analysis of flow about a cylinder.
Z. Angew. Math. Mech., 71:T242–T248, 1991.
[70] T. Muld, G. Efraimsson, and D. Henningson. Mode decomposition on surface-
mounted cube. Flow Turbulence Combust, 88(3):279 – 310, 2012.
[71] M. D. Neary and K. D. Stephanoff. Shear-layer-driven transition in a rectangular
cavity. Phys. Fluids, 30(10):2936–2946, June 1987.
BIBLIOGRAPHY 119
[72] S. A. Orszag and M. Israeli. Numerical simulation of viscous incompressible flows.
Annual Review of Fluid Mechanics, 6:281–318, 1974.
[73] S. A. Orszag, M. Israeli, and M. O. Deville. Boundary conditions for incompressible
flows. Journal of Scientific Computing, 1(1):75–111, 1986.
[74] K Parker, K D von Ellenrieder, and J. Soria. Morphology of the forced oscillatory
flow past a finite-span wing at low Reynolds number. J. Fluid Mech., 571:327–357,
2007.
[75] L. R. Pastur, Y. Fraigneau, F. Lusseyran, and J Basley. From linear stability analysis
to three-dimensional organisation in an incompressible open cavity flow. Arxiv, July
2012.
[76] J. C. F. Pereira and J. J. M. M. Sousa. Influence of impingement edge geometry on
cavity flow oscillations. AIAA Journal, 1994.
[77] R. T. Pierrehumbert. A universal shortwave instability of two-dimensional eddies in
an inviscid fluid. Phys Rev Letters, 57:2157–2159, 1986.
[78] E. Piot, G. Casalis, F. Muller, and C. Bailly. Investigation of the pse approach for
subsonic and supersonic hot jets. detailed comparisons with les and linearized euler
equation results. Int. Journal of Aeroacoustics, 5(4):361–393, 2006.
[79] A. Powell. On edge tones and associated phenomena. Acustica, 3, 1953.
[80] F. Richecoeur, L. Hakim, A. Renaud, and L. Zimmer. Dmd algorithms for experi-
mental data processing in combustion. In Proceedings of the Summer Program 2012,
pages 459–468. Center for Turbulence Research, 2012.
[81] D. Rockwell. Prediction of oscillation frequencies for unstable flow past cavities.
Trans. ASME: J. Fluids Eng., pages 294–300, June 1977.
[82] D. Rockwell and C. Knisely. Observations of the threedimensional nature of unstable
flow past a cavity. Phys. Fluids, 23(425), 1980.
[83] D. Rockwell and C. Knisely. The organized nature of flow impingement upon a
corner. Journal of Fluid Mechanics, 93:413–432, 1987. ISSN 1469-7645.
[84] D. Rockwell and E. Naudascher. Self-sustained oscillations of impinging free shear
layers. Ann. Rev. Fluid Mech., 11:67–94, 1979.
120 BIBLIOGRAPHY
[85] D. Rodrıguez and V. Theofilis. Massively parallel numerical solution of the BiGlobal
linear instability eigenvalue problem using dense linear algebra. AIAA J., 47(10):
2449–2459, 2009.
[86] J.E. Rossiter. Wind-tunnel experiments on the flow over rectangular cavities at sub-
sonic and transonic speeds. Aeronautical Research Council Reports and Memoranda,
3438, October 1964.
[87] C. W. Rowley, T. Colonius, and A. J. Basu. On self-sustained oscillations in two-
dimensional compressible flow over rectangular cavities. J. Fluid Mech., 455:315–346,
2002.
[88] C. W. Rowley, I. Mezic, S. Bagheri, P. Schlatter, and D. S. Henningson. Spectral
analysis of nonlinear flows. J. Fluid Mech., 641:115–127, 2009.
[89] R. L. Sani and P. M. Gresho. Resume and remarks on the open boundary con-
dition minisymposium. International Journal for Numerical Methods in Fluids,
18(10):983–1008, 1994. ISSN 1097-0363. doi: 10.1002/fld.1650181006. URL
http://dx.doi.org/10.1002/fld.1650181006.
[90] V. Sarohia. Experimental and analytical investigation of oscillations in flows
over cavities. PhD thesis, California Institute of Technology, 1975. URL
http://resolver.caltech.edu/CaltechETD:etd-05032007-131245.
[91] P. Schmid and D. S. Henningson. Stability and Transition in Shear Flows. Springer,
New York, 2001.
[92] P. J. Schmid. Dynamic mode decomposition of numerical and experimental data.
Journal of Fluid Mechanics, 2010.
[93] P. J. Schmid. Dynamic Mode Decomposition. LS 2014-01. von Karman Institute
Lecture Series on ”Advanced Post-Processing of Experimental and Numerical Data”,
November 2013.
[94] P. N. Shankar and M. D. Deshpande. Fluid mechanics in the driven cavity. Annu.
Rev. Fluid Mech., 32:93–136, 2000.
[95] S. J. Sherwin and H. M. Blackburn. Three-dimensional instabilities of steady and
pulsatile axisymmetric stenotic flows. J. Fluid Mech., 533(297-327), 2005.
[96] D. Sipp. Open-loop control of cavity oscillations with harmonic forcings. Journal of
Fluid Mechanics, 708:439–468, 10 2012. ISSN 1469-7645.
BIBLIOGRAPHY 121
[97] D. Sipp and A. Lebedev. Global stability of base and mean flows: a general approach
and its applications to cylinder and open cavity flows. J. Fluid Mech., 593:333–358,
2007.
[98] V. Tatsumi and T. Yoshimura. Stability of the laminar flow in a rectangular duct.
J. Fluid Mech., 212:437–449, 1990.
[99] V. Theofilis. On steady-state flow solutions and their nonparallel global linear insta-
bility. In Dopazo, C, editor, Advances In Turbulence VIII, pages 35–38, Barcelona,
Spain, 2000. ISBN 84-89925-65-8.
[100] V. Theofilis. Globally-unstable flows in open cavities. 6th AIAA/CEAS Aeroacous-
tics Conference and Exhibit, Maui, HI, June 12-14 2000. AIAA Paper.
[101] V. Theofilis. Advances in global linear instability analysis of nonparallel and three-
dimensional flows. Prog. Aerosp. Sci., 39 (4)(4):249–315, 2003.
[102] V. Theofilis. Global linear instability. Annu. Rev. Fluid Mech., 43:319–352, 2011.
[103] V. Theofilis and T. Colonius. An algorithm for the recovery of 2- and 3-D BiGlobal
instabilities of compressible flow over 2-d open cavities. In AIAA Paper 2003–4143,
33rd Fluid Dynamics Conference and Exhibit, June 23 – 26, 2003, Orlando, FL, 2003.
[104] V. Theofilis and T. Colonius. Three-dimensional instablities of compressible flow
over open cavities: direct solution of the biglobal eigenvalue problem. 34th Fluid
Dynamics Conference and Exhibit, Portland, Oregon, June 28 – July 1, 2004 2004.
AIAA Paper 2004-2544.
[105] V. Theofilis, T. Colonius, and A. Seifert. Procedings afors/eoard/ercoftac sig-
33: Global flow instability control symposium i. Fundacion General Universidad
Politecnica, 2001.
[106] V. Theofilis, D. Barkley, and S. J. Sherwin. Spectral/hp element technology for flow
instability and control. Aero. J., 106:619–625, 2002.
[107] E. F. Toro. Riemann solvers and numerical methods for fluid dynamics : a practical
introduction. Springer, Berlin, New York, 1997.
[108] L. N. Trefethen, A. E. Trefethen, S. C. Reddy, and T. A. Driscoll. Hydrodynamic
stability without eigenvalues. Science, 261(5121):578–584, 1993.
122 BIBLIOGRAPHY
[109] L. S. Tuckerman and D. Barkley. Bifurcation analysis for timesteppers. In E. Doedel
and L.S. Tuckerman, editors, Numerical Methods for Bifurcation Problems and
Large-Scale Dynamical Systems, volume 119, pages 543–466. Springer, New York,
2000.
[110] N. Vinha, F. Meseguer-Garrido, J. de Vicente, and E. V. A dynamic mode decom-
position of the saturation process in the open cavity flow. Aerospace Science and
Technology, 2014. submitted to Special Issue on ”Massively Separated Flows”.
[111] T. Wintergerste and L. Kleiser. Secondary stability analysis of nonlinear crossflow
vortices. In H. Fasel and W. Saric, editors, Proc. of the IUTAM Laminar-Turbulent
Symposium V, pages 583 – 586, Sedona, AZ, USA, 2000.
[112] S. Yamouni, D. Sipp, and L. Jacquin. Interaction between feedback aeroacoustic and
acoustic resonance mechanisms in a cavity flow: a global stability analysis. Journal
of Fluid Mechanics, 717:134–165, 2013. ISSN 1469-7645. doi: 10.1017/jfm.2012.563.
URL http://dx.doi.org/10.1017/jfm.2012.563.
[113] A. Zebib. Stability of viscous flow past a circular cylinder. J. Eng. Math., 21:
155–165, 1987.