Dissertation V4 Embedded Gedreht

High Order Discontinuous Galerkin Methods for
the Simulation of Multiscale Problems
A thesis accepted by the Faculty

of Aerospace Engineering and Geodesy of the University of Stuttgart
in partial fulfillment of the requirements for the degree of
Doctor of Engineering Sciences (Dr.-Ing.)
by
Andrea D. Beck
born in Trier
Main referee: Prof. Dr. Claus-Dieter Munz

Co referee: Prof. Dr. Rupert Klein
Date of defence: February 25, 2015
Institute of Aerodynamics and Gas Dynamics

University of Stuttgart
2015
Für Arno, Christina, Helga und Peter.
Preface
This thesis was developed during my work as an academic employee at the Institute of
Aerodynamics and Gas Dynamics (IAG) of the University of Stuttgart.
I wish to thank my doctoral advisor, Prof. Dr. Claus-Dieter Munz, for his guidance and
support during the last five years. In particular, I am grateful for his open and approach-
able leadership style, which makes the Numerics Research Group both successful and a
great place to work as well. Furthermore, I would like to thank Prof. Dr. Gregor Gassner
for our joint work and all the fruitful discussions. Many thanks also to my co-referee
Prof. Dr. Rupert Klein, whose interesting comments both during the “MetStröm” project
and my defence have always challenged me to consider new aspects of my research.
A big thank you also goes to all my colleagues at the institute, especially to all past
and present members of the Numerics Research Group. In particular, I would like to
thank Dr. Andreas Birkefeld and Dr. Christoph Altmann for their support and mentoring
during my initial time here. Also, a special thank you goes to Dr. Florian Hindenlang
and Thomas Bolemann, with whom I had many fruitful and interesting discussions and
who both impressed me with their work ethics and teamwork. Moreover, I would like
to thank Hannes Frank and David Flad for sharing their enthusiasm and ideas with me –
and their sense of humor.
This work was supported by the Deutsche Forschungsgemeinschaft, through the Schwer-
punktprogramm 1276 “MetStröm”. I am very grateful for this support, and also for the
opportunity to work with colleagues from different scientific disciplines.
Last but certainly not least I would like to thank my family, Arno, Christina, Helga and
Peter for their unwavering support.
Stuttgart, May 1, 2015
Andrea Beck
v
Contents
Preface v
Contents vii
Symbols and Abbreviations xi
Kurzfassung xv
Abstract xvii
1. Introduction 1
1.1. Multiscale Problems: Simulation Strategies . . . . . . . . . . . . . . . 3
1.2. Multiscale Problems: Modeling . . . . . . . . . . . . . . . . . . . . . 4
1.3. Multiscale Problems: Numerics . . . . . . . . . . . . . . . . . . . . . 5
1.4. Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5. Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2. Turbulence Simulation 9
2.1. Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1. Phenomenology of Turbulent Flows . . . . . . . . . . . . . . . 9
2.1.2. Governing Equations . . . . . . . . . . . . . . . . . . . . . . 19
2.2. A Model Turbulence Problem: The Taylor-Green Vortex . . . . . . . . 23
2.3. Numerical Simulation Strategies for Multiscale Flows . . . . . . . . . 26
2.3.1. DNS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.2. RANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.3. Large Eddy Simulation . . . . . . . . . . . . . . . . . . . . . 27
2.4. Aspects of Large Eddy Simulation . . . . . . . . . . . . . . . . . . . 31
2.4.1. General Formalism . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.2. The “Perfect” LES model . . . . . . . . . . . . . . . . . . . . 39
2.4.3. Modeling Strategies . . . . . . . . . . . . . . . . . . . . . . . 47
vii
Contents
3. Numerics for Scale-Resolving Simulations 55

3.1. The Points per Wavelength Paradigm . . . . . . . . . . . . . . . . . . 55
3.2. Global Spectral Methods . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.1. A Pseudo-spectral Code for the Compressible Navier-Stokes Equa-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3. Spectral Element Methods . . . . . . . . . . . . . . . . . . . . . . . . 71
3.3.1. Discontinuous Galerkin Methods . . . . . . . . . . . . . . . . 72
3.3.2. Discontinuous Galerkin Spectral Element Collocation Method . 73
3.4. Efficiency of DGSEM for Scale-Resolving Simulations . . . . . . . . . 82
3.4.1. Operation Count . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4.2. Parallel Performance . . . . . . . . . . . . . . . . . . . . . . 84
3.4.3. Comparison with Other Codes . . . . . . . . . . . . . . . . . 85
3.5. DNS with DG Methods . . . . . . . . . . . . . . . . . . . . . . . . . 89
4. Numerics for Under-Resolved Simulations 97

4.1. Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.1.1. Accuracy for Under-Resolved Problems . . . . . . . . . . . . 99
4.1.2. Stability for Under-Resolved Problems . . . . . . . . . . . . . 102
4.2. Aliasing Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2.1. Source and Effect . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2.2. De-Aliasing Strategies . . . . . . . . . . . . . . . . . . . . . 114
4.3. Stable Simulation of Under-Resolved Turbulent Flows . . . . . . . . . 120
4.3.1. Taylor-Green Vortex . . . . . . . . . . . . . . . . . . . . . . . 120
4.3.2. Flow over a Cylinder at ReD = 3900 . . . . . . . . . . . . . 133
4.4. Interaction of Subgrid Scale Model and Discretization . . . . . . . . . 139
4.4.1. DNS and LES Setup . . . . . . . . . . . . . . . . . . . . . . . 141
4.4.2. Stability and Accuracy Investigations . . . . . . . . . . . . . . 142
4.4.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.5. Flux Functions in Under-resolved Flows . . . . . . . . . . . . . . . . 151
4.5.1. Flux Choices . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.5.2. Influence of the Numerical Fluxes . . . . . . . . . . . . . . . 154
4.6. LES with DG Methods . . . . . . . . . . . . . . . . . . . . . . . . . 160
4.6.1. State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . 160
4.6.2. Perspectives for High Order LES . . . . . . . . . . . . . . . . 161
5. Conclusion and Prospects 163

5.1. Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
A. Initial Conditions for Homogeneous Isotropic Turbulence 169
viii
Contents
B. Flow over a Cylinder at ReD = 3900 - Extended Results 171
C. Taylor-Green Vortex - Extended Results 179

C.1. Taylor-Green Vortex at Re=800, N=7 . . . . . . . . . . . . . . . . . . 179
C.2. Taylor-Green Vortex at Re=1600, N=7 . . . . . . . . . . . . . . . . . 181
C.3. Results of the 1st International Workshop on High Order Methods . . . 182
D. Central Finite Differences of Arbitrary Order 187
E. Derivation of the Vorticity Equation 189
Bibliography 191
List of Tables 211
List of Figures 213
Lebenslauf 223
ix
Symbols and Abbreviations
Symbols
Cs Smagorinsky model constant

E, E(k) Kinetic Energy (at scale associated with wave number k)
√
I Imaginary unit, I = −1
k, ~k Wavenumber, wavenumber vector
L Large(st) scale / macroscale
M Polynomial degree of the ansatz for the fluxes in DG
N Polynomial degree of the ansatz for the solution in DG
NGrid Number of 1D grid points
nP P W Number of grid points / DOF required to accurately resolve
a scale of size η
TW Wall clock time
∆ Filter width
ǫ Dissipation rate of kinetic energy
η Small(est) scale, Kolmogorov microscale
ν Kinematic viscosity
xi
Abbreviations
1D, 2D, 3D one, two, three dimensions / -dimensional

ALDM Approximate Local Deconvolution Method
BR1 First Method of Bassi and Rebay for the viscous fluxes
BR2 Second Method of Bassi and Rebay for the viscous fluxes
cFD Compact Finite Difference method / discretization
CPR Correction Procedure via Reconstruction
CPU-h Computing Time in core hours
DG Discontinuous Galerkin
DGSEM Discontinuous Galerkin Spectral Element Method
DES Detached Eddy Simulation
DNS Direct Numerical Simulation
DOF Degree(s) of Freedom
DRP Dispersion-Relation-Preserving
FCT Flux Correction Transport
FD Finite Difference method / discretization
FE Finite Element method / discretization
FFT Fast Fourier Transform
FV Finite Volume method / discretization
HIT Homogeneous Isotropic Turbulence
HLRS High Performance Computing Center Stuttgart
HPC High Performance Computing
iLES LES with implicit closure
IP Interior Penality Method
JSC Jülich Supercomputing Centre
LDG Local DG
LAD Local Adaptive De-aliasing
LES Large Eddy Simulation
LF Lax-Friedrichs flux function
LLF Rusanov or Local Lax-Friedrichs flux function
Ma Mach number
MILES Monotonically Integrated Large Eddy Simulation
MPI Message-Passing-Interface (Parallelization)
NSE Navier-Stokes equations
xii
ODE Ordinary Differential Equation

PDE Partial Differential Equation
PDG Projection Discontinuous Galerkin
Pr Prandtl number
PID Performance Index
PPM Piecewise Parabolic Method
PPW Points Per Wavelength
PS Pseudo-spectral method
RANS Reynolds-Averaged Navier-Stokes equations
Re Reynolds number
RK(n) nth-order Runge Kutta
Roe Roe’s approximate Riemann solver
SGS Subgrid Scale
SVV Spectral Vanishing Viscosity
TGV Taylor-Green vortex
TKE Turbulent Kinetic Energy
TVD/B Total Variation Diminishing / Bounded
URANS Unsteady Reynolds-Averaged Navier-Stokes equations
VLES Very Large Eddy Simulation
VMS Variational Multiscale Method
WALE Wall-Adapted Local Eddy-Viscosity Model
WENO Weighted Essentially Non-Oscillatory
WU Work Unit(s)
xiii
Kurzfassung
Diese Arbeit stellt einen Beitrag zur effizienten, stabilen und genauen numerischen Si-
mulation von hydrodynamischen nicht-linearen Multiskalenproblemen mit räumlichen
Diskretisierungsverfahren hoher Ordnung dar. Aufgrund der großen Bandbreite an
räumlichen und zeitlichen Skalen erfordern diese Art von Problemen nicht nur hochge-
naue und effiziente numerische Verfahren, sondern auch deren angepaßte Umsetzung
auf Höchstleistungsrechnern. Trotz der Weiterentwicklungen auf dem Gebiet der Hard-
ware und Algorithmen bleibt eine vollständige Auflösung aller Skalen typischerweise
unerreichbar. Daher muss eine approximative Lösung mit deutlich verringerter Zahl an
Unbekannten gesucht werden, die aber in den wesentlichen Größen mit der Lösung des
Ursprungsproblems übereinstimmt. Diese Lösung läßt sich aus der Formulierung eines
reduzierten Multiskalenproblems berechnen, ergänzt durch einen Modellierungsansatz
für die nicht-aufgelösten Skalen und deren Rückwirkung auf die nieder-dimensionale
Lösung. Dieser Ansatz ist nur dann zielführend, wenn die Grobstruktur der Lösung
sich durch die reduzierten Unbekannten gut approximieren läßt und wenn der nicht-
aufgelöste Teil der Skalen universelles Verhalten zeigt, dass sich generell modellieren
lässt.
Turbulente Fluidströmungen erfüllen typischerweise diese Voraussetzungen und sind
somit ein Beispiel für diese Art von Multiskalenproblemen. In dieser Arbeit werden zwei
Verfahren für die Lösung der kompressiblen Navier-Stokes Gleichungen vorgestellt: Ein
eigenentwickelter Fourier Pseudo-Spektral-Löser, und ein mitentwickelter Simulations-
code basierend auf der Discontinuous Galerkin Spectral Element Method (DGSEM).
Beide Diskretisierungsverfahren sind aufgrund ihres spektralen Charakters sehr gut zur
Lösung von Multiskalenproblemen geeignet, da sie sehr geringe Approximationsfehler
über einen weiten Skalenbereich aufweisen und somit einen hohe Lösungsqualität pro
investiertem Freiheitsgrad besitzen. Da DGSEM auf der Variationsformulierung der
Evolutionsgleichungen beruht, läßt sich das Rechengebiet element-basiert diskretisieren,
was wiederum zur sehr guten Parallelisierbarkeit von DGSEM beiträgt und die Gene-
rierung von flexiblen, unstrukturierten Rechengittern erlaubt. Aufgrund dieser Eigen-
schaften eignet sich daher dieses Verfahren sehr gut für die Direkte Numerische Simu-
lation (DNS) von Turbulenz und macht es – wie in dieser Arbeit gezeigt – im Vergleich
zu anderen Diskretisierungsverfahren zumindest gleichwertig und oft überlegen.
Die positiven Approximationseigenschaften von DGSEM lassen sich auch auf unter-
xv
Kurzfassung
aufgelöste Probleme im Rahmen einer Large Eddy Simulation (LES) übertragen, bei
denen eine reduzierte Problemformulierung gelöst wird. Je nach Diskretisierung der
nicht-linearen Terme in den Erhaltungsgleichungen entsteht dabei jedoch ein selbstver-
stärkender Fehler, der zu einer Instabilität des Verfahrens führen kann. Die Ursachen
und die Auswirkungen dieses sogenannten Aliasing-Fehlers werden in dieser Arbeit un-
tersucht. Es werden Strategien zu seiner Vermeidung oder Kontrolle vorgestellt und
in den Simulationscode implementiert. Ein Vergleich dieser Ansätze zeigt, dass nur
durch exakte Integration der nicht-linearen Terme die ursprünglichen Verfahrenseigen-
schaften, die diese Diskretisierung für Multiskalenprobleme effizient machen, erhalten
werden können. Mit Hilfe dieser Strategie kann gezeigt werden, dass DGSEM Verfahren
hoher Ordnung im Vergleich zu anderen LES-Formulierungen basierend auf Verfahren
niedriger Ordnung pro eingesetztem Freiheitsgrad eine höhere Genauigkeit für turbu-
lente Strömungen bei mittleren Reynoldszahlen liefern.
Die Erweiterung auf Probleme mit höheren Reynolds bringt die Notwendigkeit einer
Modellierung des steigenden Abbruchsfehlers mit sich. Dazu werden zwei Strategien
vorgestellt: Im Rahmen einer impliziten LES-Schließung lassen sich die numerischen
Flußfunktionen an den Zellgrenzen anpassen, so dass ihre Dissipationsterme als Fein-
strukturmodell dienen. Wird im Rahmen einer expliziten Schließung ein zusätzlicher
dissipativer Modellterm in die Gleichungen implementiert, so stellt sich die Frage, ob
der explizite De-Aliasing-Mechanismus weiter notwendig bleibt, oder ob diese Aufgabe
vom Schließungsmodell mit übernommen werden kann. Zur Beantwortung dieser Frage
werden die Untersuchungen zum De-Aliasing, kombiniert mit einem expliziten Fein-
strukturmodell, wiederholt, und die Interaktion von Diskretisierung und Modell bewer-
tet. Es zeigt sich, dass nur durch vollständiges polynomiales De-Aliasing eine Entkop-
plung von Modell und Numerik und damit eine hohe Lösungsqualität erreicht werden
kann.
Basierend auf den Ergebnissen dieser Arbeit konnte damit eine konsistente Strategie
für die stabile und genaue Simulation von turbulenten Strömungen in voll-aufgelösten
und unter-aufgelösten Situationen mit DGSEM Verfahren hoher Ordnung etabliert wer-
den. In der Zukunft sollten basierend auf dem spektralen Charakter von DGSEM und
dem damit verbundenen hohen Auflösungsvermögen pro Element implizite und explizite
LES-Schließungen entwickelt werden.
xvi
Abstract
This work provides a contribution to the accurate, stable and efficient numerical simu-
lation of hydrodynamic non-linear multiscale problems with high order discretizations.
Due to their wide range of spatial and temporal scale, these types of problems demand
not only highly accurate and efficient numerical discretization schemes, but also care-
ful code design with regards to supercomputing architectures. Still, as a rule, even for
the most sophisticated algorithms and hardware, a full resolution of all occurring scales
remains infeasible. Thus, an approximate solution with drastically reduced number of
degrees of freedom is sought, which retains the most important characteristics of the full
solution. This solution is obtained by solving a truncated multiscale problem, supple-
mented by a suitable modeling strategy for the omitted scales and their interaction with
the truncated solution. This approach is only meaningful if the resolvable scales deter-
mine the mean solution features accurately, and if the non-resolved scales show some
form of universality behavior, which allows the derivation of meaningful models.
Hydrodynamic turbulence is one example of these types of problems. In this work, two
frameworks for the numerical solution of the compressible Navier-Stokes equations are
presented: A self-developed Fourier pseudo-spectral solver, and a co-developed frame-
work based on the Discontinuous Galerkin Spectral Element Method (DGSEM). Both
discretization schemes are highly efficient for the resolution of multiscale problems as
they – due to their spectral character – exhibit very low approximation errors over a wide
range of scales, and thus return a very high resolution capability per invested degree of
freedom. Since DGSEM is based on the variational form of the governing equations, it
allows an element-based discretization of the computational domain, which in turn leads
to superior parallelization and the possibility for flexible, unstructured meshes. These
features make it attractive for the full resolution of turbulence in a Direct Numerical
Simulation (DNS) approach and – as demonstrated in this work – highly competitive
when compared to other discretization strategies.
These favorable discretization properties carry over into the under-resolved situation
(Large Eddy Simulation, LES), where a lower-dimensional version of the problem is
solved numerically. However, depending on the discretization of the scale-producing
mechanism, its truncation can introduce a self-feeding error into the solution, that can
lead to a global instability. The source and effects of these aliasing errors are investigated
in this work. Strategies for countering or avoiding it are presented, and the code frame-
xvii
Abstract
work is extended accordingly. These strategies are compared and evaluated, showing
that only the exact quadrature of the non-linear terms recovers the favorable approxima-
tion properties and thus the efficiency of the spectral approach. With this discretization
strategy, it is shown that high order DGSEM can outperform established, lower-order
LES formulations in terms of accuracy per invested degree of freedom for challenging
test cases at moderate Reynolds number turbulence.
Extension to higher Reynolds numbers necessitates the introduction of some form of
closure for the un-resolved scales, due to the increase in the truncation error. Aspects
of two modeling approaches are discussed: An implicit modeling strategy for DGSEM
can be based on the modification of the dissipation introduced by the inter-cell fluxes.
The addition of an explicit modeling term which provides a subgrid dissipation mech-
anism raises the question whether de-aliasing remains essential in that situation. The
de-aliasing strategy is revisited, and its interactions with an explicit closure model are
examined. It is shown that only through a proper de-aliasing mechanism, the superior
scale-resolving capabilities of the scheme can be recovered, and that a decoupling of
explicit model and numerics is imperative.
Through these investigations, a consistent strategy for stable and accurate DNS and LES
of turbulent flows with high order DGSEM has been established. As an outlook, further
research strategies into LES modeling should take full advantage of the spectral char-
acter of DGSEM, and the associated scale range resolved within each element can be
exploited in both an implicit as well as explicit closure approach.
xviii
1. Introduction
The overwhelming majority of all processes in nature and science involve interacting
mechanisms on multiple spatial, temporal or frequency scales [2, 3, 61]. While the term
scale defies a strict definition, it is usually associated with an observable feature of the
problem that shows a wide range of variation in a characteristic quantity. Often, these
variations span orders of magnitude in all associated dimensions, giving the problem
its chaotic appearance, when regarded from the “macroscale” level. The macroscale is
usually defined as the largest occurring or observable scale of the problem, while the
smallest features are commonly referred to as “microscales”. Depending on the physical
processes and the scale separation involved, the macro- and microscale regimes may be
ruled by different governing sets of equations, but even when the same mathematical
description of the underlying problem remains valid across all scales, a different set of
effects may dominate across the considered scale bandwidth.
Figure 1.1 shows an example of a multiscale problem of geophysical fluid dynamics:
Two temperature bubbles interact due to buoyancy forces in a stratified medium, form-
ing a range of temporal and spatial scales.
What almost all of the multiscale problems have in common and what makes them chal-
lenging from an analytical and numerical point of view, besides the fact that the involved
scales can differ in size by many orders of magnitude, are three issues: i) interactions be-
tween scales occur across the full spectrum, ii) these interactions are typically non-linear
and iii) the mutual importance of physical effects and associated boundary conditions is
a function of the position within the scale spectrum. Fluid turbulence is a classical ex-
ample of this behavior. In fact, the development of one of its most famous statements
follows along the thoughts delineated by i) to iii): In deriving relations for the scales of
turbulent flows, Kolmogorov made use of conjectures (called the “Kolmogorov’s simi-
larity hypotheses” [116, 117]) which touched on these aspects of multiscale problems:
By assuming a balance of large scale non-linear energy production and small scale linear
dissipation, a dominance of viscous actions and a trend towards universality at the small
scales and dimensional analysis he arrived at the well-known Kolmogorov’s k−5/3 law
(actually first published by Obukhov [150]), which gives a relationship for the universal
shape of the kinetic energy spectrum in the inertial subrange region:
E(k) ∼ ǫ2/3 k−5/3 , (1.1)
1
1. Introduction
Figure 1.1.: Temporal evolution (from left to right) of two temperature bubbles in a
stably stratified medium. Shown are isocontours of temperature difference
∆θ = θ − θbackground .
where E(k) denotes the kinetic energy at wave number k, and ǫ is the rate of kinetic
energy dissipation. Another illuminating result of his work is an estimate for the ratio of
the largest and smallest length scales L and η in turbulence:
3/4
L VL
∼ = Re3/4 , (1.2)
η ν
where V denotes the large scale velocity, ν the kinematic viscosity and Re the associated
Reynolds number. This relation can also be interpreted as an estimate for the information
content (or the number of degrees of freedom, DOF) of the problem. An extension to
three dimensions shows that the total number of spatial information bits Btotal in a
turbulent flow scales as
Btotal ∼ Re9/4 , (1.3)
without taking temporal scales into account. This estimate demonstrates the overwhelm-
ing complexity and information density of this multiscale problem and raises an impor-
tant question: How can these types of problems be analyzed, when the complex non-
linearities restrict analytical treatment to only very simplified “toy problems”, and when
the shear amount of degrees of freedom makes an experimental and numerical investi-
gation infeasible?
Interestingly, the same ideas that led Kolmogorov through his derivation process, can
help to answer this question: The scale separation itself and with it the separation of
dominant effects and boundary condition influence can be exploited to devise suitable
strategies for tackling multiscale problems.
2
1.1. Multiscale Problems: Simulation Strategies
1.1. Multiscale Problems: Simulation Strategies

While we are mainly interested in multiscale problems occurring in turbulent fluid flow
in this work, the general principles outlined in this section can be transferred to other
types as well: For the numerical simulation of multiscale problems, two general strate-
gies exist. Either a laboratory-type experiment with well-defined initial and boundary
conditions for a single representative scale or a very limited bandwidth of scales can be
conducted, which allows the detailed observation of the local processes and interactions,
or the mean (temporal and spatial) effects at the large scales can be observed, while
neglecting or approximating (modeling) small scale effects. The first approach usually
aims at understanding the basic building blocks (i.e. the physical mechanisms) on the
small scale level, taking advantage of their universality and decoupling from large scale
boundary conditions and deriving suitable models for their effects (e.g. [193]). The sec-
ond approach is motivated by the observation that the mean statistics of the flow are often
mainly governed by large scale effects, which are strongly dependent on the boundary
conditions of the flow. Since the small scales are more universal, there is a chance that
finding suitable models for their interaction with the large, anisotropic scales is feasi-
ble [168].
Although these approaches might seem very different, they are complementary in their
exploitation of the universality and of the dominance of different physical effects at op-
posite ends of the scale range.
While the idea of devising models on the small scale level and coupling them to large
scale dynamics is attractive, a major issue is the loss of universality of the building
block, either through a singular event or through insufficient scale separation. Insuffi-
cient scale separation can occur for two reasons: i) The problem itself does not support
this approach, i.e. there is no scale range with universal physical behavior, or no domi-
nating physical effect can be identified. In that case, if no suitable model can be defined,
the only remaining strategy would be to restrict any analysis to problems with low di-
mensionality. ii) An induced insufficient scale separation can occur when the small
scale model is applied in a scale region that does not support the underlying universal-
ity assumption. One example of an event that violates the universality assumption in
turbulent flows is the phenomenon of backscattering, where the main direction of non-
linear interaction is not from the anisotropic large scales to the universal small scales,
but reversed [156]. In such a case, the small scale dynamics result in the generation of
large scale features. This directly contradicts the model assumptions, and can render this
building block approach obsolete.
In summary, the main questions that arise when devising strategies for multiscale prob-
lems are: Is there universality or at least a dominating effect at a given scale level? Can
we model this regime, and how do we couple building blocks at different scale levels?
3
1. Introduction
Where do we have to give up this building block model? And how can singular, non-
typical evens be treated?
In the end, the current state of the art can be summed up in a practical paradigm: resolve
as much as possible of the scales which are non-universal or lead to singular behavior,
and model everything assumed to be universal or strongly dominated by well-understood
physical processes. In fluid dynamics, this approach is commonly referred to as the
Large Eddy Simulation (LES) method. It should be noted that it consists of at least two
important subtasks: The accurate numerical simulation of the non-universal scales and
the modeling of the small scale effects.
1.2. Multiscale Problems: Modeling
The aim of the modeling approach is to reduce the complexity of the multiscale prob-
lem to a point where it becomes tractable by numerical simulation. Ideally, this should
be achieved by modeling the small scale dynamics in such a way that the large scale
properties remain unaltered. Issues shared by all modeling methods arise from the gen-
eral question of modeling physical systems. The identification of “model-able” regimes
with a universal character has been discussed above. The inclusion of singular events
that break this universality like backscattering requires special treatment, for example
through a stochastic approach [64, 137, 138]. Another important aspect is to ensure
that the model fulfills the transformation invariances and symmetries of the governing
equations as well as realizability constraints [187], and regulates itself in well-resolved
scenarios.
Two main approaches exist in the development of these models: i) “Building block”
models are derived either from an analysis of the continuous dynamics or from numer-
ically obtained correlations of high-resolution, small scale simulations. These models
are usually based purely on physical and mathematical reasoning and are independent
from the numerical strategies for the large scale computations. This approach results in
additional terms or equations describing the model which then complements the original
problem formulation. It is called explicit modeling or “explicit LES”. ii) Altering the
discretization properties of the large scale simulation schemes in such a way that the
discretization itself performs the regularization of the problem is referred to as “implicit
LES”. An important drawback of this strategy is the dependence of the modeling terms
on the type of the discretization and additionally on the local resolution, which makes
an a priori evaluation of these models very difficult. Explicit models, on the other hand,
while based on physical reasoning, rely on error-free resolution of the large scales. This
exactness can usually not be guaranteed, thereby resulting in an unwanted interaction of
model and numerical errors.
4
1.3. Multiscale Problems: Numerics
1.3. Multiscale Problems: Numerics

As discussed above, small scale modeling and large scale numerical simulation are two
interacting fields. Either, this interaction is sought and exploited in an implicit modeling
approach, or it is to be minimized for explicit small scale models. In both cases, the se-
lection and design of the numerical scheme for the large scale simulation plays a crucial
role.
The most basic issue that arises when solving multiscale problems numerically is that of
finite digital precision and rounding errors. Since the scales involved can span many or-
ders of magnitude, the representation of these scales can lead to conditioning problems.
In addition, non-linear problems exhibit a high sensitivity to slight changes in initial and
boundary conditions (or run-time round-off errors), which can affect the instantaneous
realization of the turbulent flow, but not its statistics [181]. While the effects of finite
precision are indeed a concern, the range of physical scales present is usually restricted
by the computational cost of the simulation and does not exploit the full digital resolu-
tion available. Therefore, typically also a scale separation between the smallest resolved
scales and the round-off errors exists, which prevents these errors from dominating the
numerical solution.
An estimate for the range of scales occurring in a problem has already been given in
Equation (1.2). Assuming that the spatial discretization scheme requires nP P W grid
points or degrees of freedom per wavelength (PPW) resolve the smallest length scale η,
the total number of grid points N3D in three dimensions can be estimated as [176]
3
L
N3D = ∼ n3P P W Re9/4 . (1.4)
η/nP P W
A refined analysis even found a more stringent requirement of N3D ∼ n3P P W Re37/14
[44]. Considering not only the spatial degrees of freedom, but also the the fact that the
characteristic time scale of the dissipation scales is directly proportional to η, the total
computational cost in terms of spatial and temporal degrees of freedom Ntotal becomes
Ntotal ∼ n4P P W Re3 . (1.5)
This relationship highlights two important issues: i) The computational effort of resolv-
ing all scales can quickly become unbearable and ii) the scale-resolving capabilities of
the discretization dramatically influence the overall cost, i.e. the total effort is a function
of both the physics (through Re) and of the numerical capabilities (through nP P W ).
When resolving only a fraction of the scales in a LES approach, the demands stemming
from the physics can be lowered to N3D ∼ Re9/5 [42] or N3D ∼ Re13/7 [44] for
wall-resolved LES, making the influence of the numerics on the total simulation cost
5
1. Introduction
even stronger. Thus, a fundamental demand for efficient numerical simulation of all or
a subset of the scales of turbulent motion can be formulated as: The number of degrees
of freedom or grid points required to accurately resolve the smallest occurring relevant
scale must be minimized.
In addition to efficiency, which is directly related to the behavior of the approximation
errors of a given discretization, stability of the simulation is also an issue of concern.
Stability can always become an issue when physical processes are not correctly repre-
sented in a discrete setting, leading to a violation of associated symmetries (e.g. the
skew-symmetry of the convective term), constitutive equations (e.g. positivity of density
and pressure) or governing laws (e.g. inconsistent LES closure). When resolving all
scales of the multiscale problem adequately, stability becomes purely a matter of numer-
ics (if the underlying continuous problem is well-posed and stable), i.e. any instabilities
are due to inappropriate numerical methods and can be treated as such. When resolving
only a partial range of scales of the continuous problem and modeling the rest in a LES
setting, overall stability has three contributors: The discretization and the model, as well
as their complex interactions. The question which of these two should provide the sta-
bility is an open issue of research, some LES methods rely on the model, e.g. [142, 158],
others on the scheme itself, e.g. [6, 94, 180].
Directly related to the question of stability is the control over the interaction of model
and discretization errors. Model errors include the incorrect prediction of physical be-
havior (a “non-perfect” model) as well as the representation of the model in a discrete
setting. Discretization errors result in an inaccurate resolution of the resolved scales,
and therefore in imperfect “input parameters” for the model itself. For implicit LES,
where the discretization itself generates the modeling terms, scale-resolving efficiency,
stability and model behavior are completely interdependent.
1.4. Objective
The objective of this work is to provide a contribution to the accurate, efficient and sta-
ble numerical simulation of turbulent flows, in particular in under-resolved situations.
As discussed before, the efficiency of a discretization in terms of invested degrees of
freedom can become a limiting factor in the simulation of multiscale problems. We
demonstrate the effectiveness of high order Discontinuous Galerkin discretizations for
both simulation strategies outlined above: For high resolution, large scale DNS on a
“building block” level as well as for under-resolved simulations combined with a mod-
eling approach. In the latter case, we focus on the approximation properties over the
full scale range and the treatment of non-linear terms, which both influence the stabil-
6
1.5. Outline
ity of the discretization operator. We present and evaluate de-aliasing options and show
that the favorable scale-resolving capabilities of the schemes can only be recover through
exact integration of the non-linear terms. For moderate Reynolds number flows, we com-
pare de-aliased high order Discontinuous Galerkin schemes without additional closure
models against low order formulations, both with explicit and implicit LES models, and
show the increased accuracy in terms of invested degrees of freedom for our approach.
In addition, we present two strategies for the extension to higher Reynolds number flows,
where the effects of the truncation error become more pronounced: An implicit modeling
through the surface flux functions, and the inclusion of an explicit LES closure model.
For this second strategy, we investigate the interaction of the discretization choice of the
non-linear term, the approximation properties of the scheme and the model, and we con-
clude that polynomial de-aliasing is essential, not only from a stability point of view, but
also to exploit the accuracy of the high order discretization and to develop and evaluate
explicit closure models. Finally, we propose a strategy for extending the capabilities of
current DG LES methods, both for explicit and implicit modeling approaches, based on
our findings.
1.5. Outline
In this work, we will present and discuss numerical schemes for the approximation of
fully resolved and under-resolved multiscale problems, focusing on the example of hy-
drodynamic turbulence as described by the Navier-Stokes equations. We will briefly
discuss the features of turbulence that make it a challenging problem from both a nu-
merical and a fluid mechanics standpoint in Section 2.1, followed the description of a
canonical test case in Section 2.2 and by a discussion of numerical strategies for its sim-
ulation in Section 2.3. Focusing on a Large Eddy Simulation strategy, we will derive a
general formalism for the classification of the different LES approaches in Section 2.4,
and discuss the perfect and actual closure strategies.
In Chapter 3, we focus on numerical schemes for efficient resolution of multiscale prob-
lems. The importance of the “points per wavelength” concept will be demonstrated in
detail in Section 3.1, followed by the discussion of two families of methods suitable for
scale-resolving numerics: The global spectral methods in Section 3.2 and the spectral
element methods in Section 3.3. We will present in detail two representative schemes
belonging to each category, and the associated computational frameworks developed in
this work. In particular, we will focus on the details and properties of the Discontinuous
Galerkin Spectral Element method and its efficiency for scale-resolving computations in
Section 3.4, followed by examples of scale-resolving simulations with this framework in
7
1. Introduction
Section 3.5.
Extending the discussion from well-resolved to under-resolved problems in Chapter 4,
we will focus on the additional numerical and modeling challenges that arise in these
situations in Sections 4.1 and 4.2. The discretization of the scale-producing term re-
quires special attention, as it can cause instabilities. Different countering strategies for
under-resolved flows will be discussed and compared in Section 4.3. In Section 4.4, we
investigate the interaction of numerics and explicit modeling, and highlight the impor-
tance of neutrally stable numerics for LES. Finally, we briefly demonstrate the influence
of the flux functions in a LES setting in Section 4.5 and discuss perspectives and state
of the art for LES with high order methods in Section 4.6. We conclude with a summary
and an outlook on research strategies in Section 5.
8
2. Turbulence Simulation
In this chapter, we briefly describe the most prominent features of turbulent flows and
the general numerical solution strategies, with a special focus on aspects of Large Eddy
Simulations.
2.1. Turbulence
Turbulence is a state of fluid motion characterized by non-linear interactions over a wide
range of scales. It is the results of an instability mechanism of a steady base flow that
amplifies successive instability modes and finally leads to turbulence [67]. The onset
of this transition process and the resulting range of turbulent scales is governed by the
Reynolds number Re = VνL of the flow (see Equation(1.2)), a similarity parameter
relating the non-linear convective effects and the linear diffusive actions. An exact, all-
encompassing definition of turbulence is difficult to state due to the many aspects it
would have to include, and thus, the question “what is turbulence ?” is usually answered
by listing its most important features.
2.1.1. Phenomenology of Turbulent Flows

Following the discussion presented in [57,167,192] and in lieu of a definition, turbulence
is characterized by the following features:
• Non-linearity and scale range
Turbulence is characterized by a wide range of spatial and temporal scales, usually
associated with coherent structures called “eddies” with a characteristic length (or
wave number as its inverse) and a turn-over frequency. The largest scales or eddies
are constrained by the boundary conditions and mean flow conditions and are thus
non-universal and anisotropic, while the smallest scales become (with increas-
ing Reynolds number) isotropic and universal. The kinetic energy is extracted
from the mean flow and fed into smaller eddies, which in turn interact with each
other and the mean flow. While these interactions form scales across the whole
scale range, the majority of the resulting transfer of kinetic energy is “downwards”
(called forward scatter), i.e. produces higher wave number eddies [166]. The vis-
cosity of the fluid is the limiting agent for this scale generation, as the dissipative
9
effects on regions of high curvature (i.e. small, high-frequency eddies of sizes on

the order of η) overcome the scale production mechanisms.
The non-linearity of the convective term in the governing equation of fluid flow
is responsible for the generation of this scale cascade. Non-linearity implies an
interaction of different scales, i.e. a strong coupling of the degrees of freedom of
the system. While linear systems may also possess an arbitrary high number of
degrees of freedom, these are not coupled and thus do not interact.
The simplest model of non-linear interaction is described by the viscous Burgers’
equation for a scalar u = u(x, t) ∈ R, given by
∂u 1 ∂u2 1 ∂2u
+ − = f (x, t),
∂t 2 ∂x Re ∂x2
(2.1)
with u = u(x, t), x ∈ [0, 2π], t ∈ [0, ∞+ ).
Here, Re = ν1 is the Reynolds number based on the kinematic viscosity ν > 0

and f a forcing term. Equation (2.1) contains a quadratic non-linearity in the
convective term and a linear viscous term, which also arise in the incompress-
ible Navier-Stokes equations (plus the pressure contribution). Therefore, Burgers’
equation often serves as a model for the full Navier-Stokes system, as its forced
version with special f (x, t) 6= 0 can exhibit chaotic behavior resembling certain
aspects of fully three-dimensional turbulence [79, 139, 145].
Considering an initial solution at time t0 to Equation (2.1) of the form u(x, t0 ) =
û sin(kx) with an initial amplitude û and a single wave number k, the total flux
at t = t0 becomes
∂u ∂u 1 ∂2u
= −u +
∂t t=t0 ∂x t=t0 Re ∂x2 t=t0
1 ûk2 (2.2)
= − û2 k sin(2kx) − sin(kx).
2 Re
This simple model equation shows that the initial flux will introduce a term with
wave number 2k into the time evolution of the solution, i.e. higher wave numbers
will be generated through this term. This is the root of the scale producing mech-
anism that generates the energy cascade. The second term on the right hand side
of Equation (2.2), stemming from the viscous flux, does not introduce additional
higher frequency waves, but dampens the initial wave. Note that the magnitude
of the damping is a function of k2 , i.e. high wave numbers will be affected more
10
2.1. Turbulence
strongly by this term. In general, Equation (2.2) shows that an initial state contain-
ing wave numbers k1 and k2 will produce – through the quadratic non-linearity
– waves with 2k1 , 2k2 and k1 + k2 , which will interact over time and thus suc-
cessively generate waves over the full wave number range, until their production
is balanced by viscous effects. The relation of k to Re determines which effect
dominates in the current situation. For k ≪ Re, the second term on the right
hand side of Equation (2.2) will become negligible, and the flux will be governed
by the inviscid contributions. In the other extreme k ≫ Re, the dissipative flux
will dominate. Thus, this simple scalar model equation demonstrates the basic
mechanism of scale production and destruction in the cascade.
Another important aspect of turbulence and the non-linear terms can be derived
from Equation (2.1). By defining the kinetic energy E(t) of the solution u(x, t)
as Z
1 2π
E(t) = u(x, t)2 dx, (2.3)
2 0
an evolution equation for this energy can be derived from the unforced viscous
Burgers’ equation by multiplication with u and integration over the spatial domain
x ∈ [0, 2π]:
Z 2π
∂u 1 ∂u2 ∂2u
0= u + u − νu 2 dx
0 ∂t 2 ∂x ∂x
Z 2π
1 ∂u 2
1 ∂u 3
∂ ∂u ∂u ∂u
= + −ν (u ) − dx
0 2 ∂t 3 ∂x ∂x ∂x ∂x ∂x
2π 2π Z 2π (2.4)
dE 1 1 ∂u2 ∂u ∂u
= + u3 − ν +ν dx
dt 3 0 2 ∂x 0 0 ∂x ∂x
2π Z 2π
dE ∂u ∂u
⇔ = Fbc −ν dx,
dt 0 0 ∂x ∂x
where we have collected the boundary fluxes in Fbc . It is important to note that
the non-linear term does not have a volume contribution to the kinetic energy, only
a flux across the domain boundaries, while the viscous term contributes through
a surface flux as well as a volume term. If we consider Fbc to be zero (for ex-
ample for a periodic problem), the only contributing factor to the kinetic energy
balance would be the viscous term, i.e. the kinetic energy would decrease with the
rate prescribed by the viscosity ν and the integral over the product of the velocity
gradients. Therefore, the non-linear term is only responsible for the distribution
of kinetic energy across the scales, but it is conservative, while the linear dif-
fusion term does not cause scale interaction, but serves as a sink for the kinetic
11
-2
10 0.02
10-3
0.01
-4
10 k
-5/3
-5
10
t=60
E(k)
E(t)
10-6 ~t-1.25
-7 t=95
10
-1.3
~t
10-8 t=250
-9
10
-10
10 0 1 2 0 1 2
10 10 10 10 10 10
Wavenumber k t
Figure 2.1.: Decaying homogeneous isotropic turbulence. Computations with pseudo-

spectral code Spex described in Section 3.2.1.
Left: Temporal evolution of the spectrum of kinetic energy E(k). Dashed
line denotes Kolmogorov’s k−5/3 law.
Right: Temporal evolution of total kinetic energy E. Dashed lines denote
decay laws published by [53] and [102].
energy. These two effects found in the Burgers’ model equation are also present
in the full Navier-Stokes equations. Their balance (expressed through the flow
Reynolds number) determines the magnitude of the scale cascade, i.e. the scales
at which the viscous effects overcome the scale production mechanism. Figure 2.1
left depicts the spectrum of kinetic energy of decaying isotropic turbulence of the
three-dimensional Navier-Stokes equations. While the scaling laws are different
from those for Burgers’ equation, the region dominated by inertial forces (follow-
ing Kolmogorov’s k−5/3 law in the so-called inertial subrange) and the dissipation
range with its rapid energy decay are clearly visible.
It should be noted that while non-linearity is essential for the production of turbu-
lence, it is not sufficient. Non-linear interaction may lead to multiscale phenom-
ena that show organized, large scale behavior, for example the shocks forming in
the unforced Burgers’ equations or the solitons of the Korteveg-de Vries equa-
tions [197, 205].
12
2.1. Turbulence
• Spatio-temporal irregularity and loss of predictability

The scale production and dissipation mechanisms highlighted by the Burgers’
equation can be translated directly to the three-dimensional Navier-Stokes equa-
tions. However, due to the three-dimensionality and the long-mode pressure cou-
pling, the resulting structures are a lot more intricate and complex. Turbulent
fields appear – although they are still solutions to a set of deterministic equations
– highly irregular and chaotic. It is important to note that this randomness is
not dependent on randomized forcing or boundary conditions, as opposed to the
Burgers’ equation, which only produces chaotic solutions under stochastic forc-
ing. Instead, hydrodynamic turbulence (and the Navier-Stokes system) possess
an intrinsic “self-stochastization” property [197], in which chaotic motions can
occur from smooth initial and boundary conditions. The root of this behavior is
the extreme sensitivity to minuscule changes in the current state or the boundary
conditions, a feature also observed in experiments [72].
Equation (2.5) describes a simple system of ordinary, non-linear differential equa-
tions for a spatial position ~x(t) = [x(t), y(t), z(t)], originally proposed by Lorenz
as a model for particle movement in atmospheric convection [132]:
dx
= a(y − x)
dt
dy
= x(b − z) − y (2.5)
dt
dz
= xy − c .
dt
Depending on the choice of the scalars (a, b, c) ∈ R+ , its numerical solution
exhibits a highly irregular, chaotic behavior. Figure 2.2 shows the solution tra-
jectories of Equation (2.5) for two slightly different initial starting points and the
temporal development of their x-components. Up to about t = 40, both solu-
tions are indistinguishable, but deviate strongly after this time. The x-component
plot clearly shows the chaotic behavior. In the solution space visualizations, the
trajectories begin to deviate, although their general, long term behavior is stable
and governed by a structure called the “Lorenz attractor”. The circular markers in
Figure 2.2 top denote the position in the solution space after the same time tend
for both initial conditions. Although the distance between both starting positions
was minimal, their current state differs greatly, visualizing the high sensitivity of
the system to small deviations.
These properties of the Lorenz system have made it a model problem for deter-
ministic non-linear equations that show inherent chaotic behavior. Its behavior as
outlined above can be translated to the turbulence described by the Navier-Stokes
13
45 45
40 40
35 35
30 30
25 25
Z
Z
20 20
15 15
10 10
5 5
0 0
−10 25 −15 30
−5 20 −10 20
15 −5
0 10 10
5 5 0 0
0 5
10 −5 10 −10
−10
15 −15 15 −20
Y Y
X X
15
10
5
x(t)
−5
−10
−15
0 50 100 150
t
Figure 2.2.: Solution of the Lorenz System (Equation (2.5)) with a = 3, b = 25, c = 1.
Blue lines: initial condition ~x0 = [1, 1, 2]; Black lines: initial condition
~x0 = [1.00001, 1, 2].
Top: Solution trajectories up to tend = 10 (left) and tend = 90 (right).
Green and red markers show the particle position of blue and black trajec-
tories at tend .
Bottom: Temporal evolution of x-components of particle positions.
equations: When subjected to slight changes in the initial and boundary condi-
tions, large deviations in the resulting local solution can occur. However, the long
term, mean solution is governed by so-called “structurally stable attractors” [90],
which are insensitive to these small deviations. This combination of predictability
and unpredictability results in the fact that while single realizations of turbulence
(both numerical and experimental) can differ significantly in their local, small
scale behavior, their statistical properties remain constant against slight perturba-
tions. Turbulence therefore remains statistically stable on the large scales (which
allows the prediction of integral quantities like drag and lift), but becomes unpre-
dictable and chaotic on the small scales.
14
2.1. Turbulence
• Three-dimensionality and vorticity

Turbulence is characterized by regions of highly three-dimensional, coherent struc-
tures undergoing rotational motion. The concept of vorticity as the curl of the
velocity gradient ω ~ ≡ ∇×~ u is used to quantify and visualize these structures.
Vorticity is defined as the vector aligned with the local axis of fluid rotation, with
the sign determined by the local right hand system. The length of the vector cor-
responds directly to the strength of the vortex, i.e. to the radial velocity gradient.
Within a turbulent field, regions of high vorticity can be observed and the over-
all initial vorticity is not conserved, i.e. three-dimensional turbulence contains
a mechanism of producing vorticity. This mechanism can be identified by taking
the curl of the momentum equation of the incompressible Navier-Stokes equations
(full derivation in Appendix E):

∂~
u ∇p
∇× + (~
u · ∇)~
u=− + ν∇2 ~
u , (2.6)
∂t ρ
which leads to an evolution equation for the vorticity:
D~
ω ∂~
ω
= + (~
u · ∇)~
ω = (~ u + ν∇2 ω
ω · ∇)~ ~. (2.7)
Dt ∂t
Equation (2.7) shows that from the non-linear term of the momentum equation,
the term (~ω · ∇)~ u on the right hand side evolves, which acts as a source term
for vorticity if a velocity shear exists. This term is usually referred to as “vortex
stretching”, where the actual stretching is due to the aligned component and a
turning or bending of the vortex is due to the cross components, e.g. for the first
component of ω ~
2
Dω1 ∂u1 ∂u1 ∂u1 ∂ ω1 ∂ 2 ω1 ∂ 2 ω1
= ω1 + ω2 + ω3 +ν + + . (2.8)
Dt ∂x ∂x2 ∂x3 ∂x21 ∂x22 ∂x23
| {z 1} | {z } | {z }
Stretching Turning Diffusion
For the vortex stretching mechanism to be non-zero, two conditions must be

met: i) a velocity shear in the flow must exist, and ii) the flow must be three-
dimensional, i.e. the vorticity vector must not be perpendicular to the plane of
shear. This is always the case in two-dimensional flows, and thus, a production
mechanism for vorticity is absent in these flows. Due to the absence of this mech-
anism, the term “two-dimensional turbulence” is often disputed, as this flow lacks
the basic mechanism of generating vorticity (without external forcing) [121]. In
fact, in two dimensions, only two vortex interactions occur:
15
Figure 2.3.: Vortex interactions in two dimensions, shown are contours of vorticity mag-
nitude. From left to right: initial random vorticity field at t = 0, resulting
fields at t = 100, 500, 10000.
The creation of a dipole or the merging of two vortices into a single vortical struc-
ture at a larger scale. This behavior leads to the so-called “inverse cascade” of
two-dimensional turbulence, where kinetic energy is transferred to larger struc-
tures instead of smaller structures as in three-dimensional turbulence. Since this
upward transfer always occurs when the rotational axes of two vortices are parallel
(or have parallel components), this effect is also present in three-dimensional tur-
bulence, where the alignment of the vortices loses directionality and anisotropy of
the large scale boundary conditions and becomes chaotic at the small scales. This
interaction with its upward energy transfer is the physical cause of the backscat-
tering effect. It should be noted that even in three-dimensional flows with a pre-
ferred directional alignment, e.g. due to external forces like gravity, a quasi-
twodimensional turbulence with dominating upward transfer can develop [178,
195].
Figure 2.3 demonstrates the vortex interactions in two dimensions that lead to the
generation of large scale structures. Starting from an initial field with random
velocity fluctuations about a mean, the backscattering mechanism continuously
merges the small scale structures into successively larger vortices.
Returning to a discussion about the vortex stretching mechanism that increases

vorticity locally in three dimensions (Equation (2.8)), each vortical structure can
be assigned an angular momentum, defined as 21 Θ ω ~ , where the Θ is the (scalar or
tensor) moment of inertia associated with the mass of fluid contributing to the vor-
tex. In an inviscid flow, there are no viscous torques acting as retarding moments,
16
2.1. Turbulence
Figure 2.4.: Interaction of two mono-scale vortices: Development of vortical structures

and scale cascade. Shown are λ2 = −0.001 isocontours, colored by nor-
malized helicity. Temporal evolution from upper left to lower right.
17
so the angular momentum is conserved according to
DΘ ω
~
=0
Dt (2.9)
D~
ω DΘ
⇔Θ = −~
ω .
Dt Dt
As D~Dt
ω
increases through the vortex stretching mechanism in Equation (2.7), it
must be balanced by a decrease in the moment of inertia on the right hand side of
Equation (2.9), thus, higher vorticity results in a lower moment of inertia. For a
fully incompressible flow, a decrease in the moment of inertia can only be accom-
plished by a thinning of the vortex tube, along with a lengthening along its axis to
conserve the volume. Thus, the conservation of rotational momentum causes the
elongation and thinning of the structures undergoing the stretching mechanism,
and so explains the generation of small scale, high frequencies vortices. This
process is self-amplifying, and is only stopped when the small scale structures
succumb to the dissipation due to their self-induced high strain rates.
Figure 2.4 visualizes these three-dimensional vortex interaction mechanisms and
the development of a full scale cascade from two large scale structures.
• High diffusion and dissipation

Turbulent flows are both highly diffusive and dissipative. Due to their chaotic
small scale motion, interchange of momentum, energy and passively transported
objects is highly enhanced. This leads to a perceived increase in “turbulent vis-
cosity” and “turbulent diffusivity” which can be orders of magnitude stronger than
their molecular counterparts. However, these “turbulent” terms are not (like the
molecular viscosity and diffusivity) properties of the fluid, but of the local turbu-
lence statistics of the flow. The resulting increase in strain due to the non-linear
interaction leads to a rapid dissipation of kinetic energy into heat, and thus turbu-
lent flows require a form of energy input (large scale gradients, buoyancy, forcing)
to sustain themselves. This transformation of energy from an ordered (large scale,
kinetic) to an unordered (chaotic, thermal) state causes the highly enhanced losses
encountered in turbulent flows, e.g. over bluff bodies or in internal flows.
• Non-locality
Quoting Tsinober [197], non-locality is “probably one of the main reasons the
problem of turbulence is so difficult.” The term “non-locality” refers to two issues:
i) The non-locality of the pressure field with respect to the velocity field. In the
2 ∂ui ∂uj
incompressible limit, the pressure equation ∂∂xpi = − ∂x j ∂xi
becomes elliptic,
18
2.1. Turbulence
i.e. the pressure becomes non-locally dependent on the velocity field in the whole
domain. It is this elliptic nature of the pressure, called the pressure-velocity cou-
pling, that is associated with long-range forces and thus the non-locality in time
that differentiates the Burgers’ equation model turbulence from the Navier-Stokes
turbulence. ii) By the same argument, the velocity field at each space point is
defined by the global vorticity field, i.e. the large scales (represented by the ve-
locity) and their associated small scales (the vorticity) are strongly coupled. This
coupling is bidirectional in its non-locality [197], i.e. the backward reaction of
the small onto the large scales also occurs in a non-local manner.
2.1.2. Governing Equations

In the following section, the governing equations of motion for both compressible and
incompressible fluids are collected.
2.1.2.1. Compressible Navier-Stokes Equations

The temporal and spatial evolution of a viscous, compressible fluid is governed by the
conservation statements for mass, momentum and energy. In conservative form this set
of partial differential equations (PDE) for a Newtonian fluid is given by
∂ρ ∂ (ρuj )
+ = 0,
∂t ∂xj
∂ (ρui ) ∂ (ρui uj + pδij ) ∂σij
+ = , (2.10)
∂t ∂xj ∂xj
∂ (ρe) ∂ [(ρe + p) uj ] ∂qj ∂ (σij ui )
+ =− + .
∂t ∂xj ∂xj ∂xj
Here, the Einstein summation convention applies, δij denotes the Kronecker delta func-
tion and i, j = 1, 2, 3. The conservative variables of mass, momentum and energy are
U = [ρ, ρu1 , ρu2 , ρu3 , ρe], where ρ denotes the density, ui the i-th component of the
velocity vector and the total energy ρe is given by
1
ρe = ρ( ui ui + cv T ). (2.11)
2
Herein, cv constitutes the specific heat at constant volume and T the temperature. Using
the equation of a perfect gas, this defines the pressure p and the adiabatic exponent γ as
cp
p = ρRT, γ= , (2.12)
cv
19
with R = cp − cv as the specific gas constant computed from cv and the specific heat
at constant pressure cp . The viscous stress tensor σij is a function of the viscosity µ
(which itself is dependent on temperature) and the velocity vector
σij = µ(T )Sij , (2.13)
where the rate of strain tensor is given by
∂ui ∂uj ∂uk

Sij = + − λδij . (2.14)
∂xj ∂xi ∂xk
The bulk viscosity coefficient λ is commonly chosen to be 32 , which removes the trace
from Sij . The remaining unknown in Equation (2.10) is the definition of the heat flux
vector qj as
∂T cp µ
qj = −k , with k = , (2.15)
∂xj Pr
where P r denotes the Prandtl number of the fluid. In compact form, Equation (2.10) can
be recast as
∂U
+ ∇ · F c (U ) = ∇ · F v (U, ∇U ), (2.16)
∂t
where F c denotes the matrix containing the three Euler flux vectors from the left hand
side of Equation (2.10) and F v the matrix containing the viscous contributions from the
right hand side.
2.1.2.2. Incompressible Navier-Stokes Equations
For a fluid of constant density, Equation (2.10) reduces to
∂ui
= 0,
∂xi
(2.17)
∂ui ∂ (ui uj ) 1 ∂p µ ∂ 2 ui
+ + = ,
∂t ∂xj ρ ∂xi ρ ∂x2j
with the associated Poisson equation for pressure derived by taking the divergence of the
momentum equation and inserting the continuity condition

∂2p ∂ui ∂uj ∂ ∂ui uj
=− =− . (2.18)
∂xi ∂xj ∂xi ∂xj ∂xi
20
2.1. Turbulence
By taking the scalar product of the momentum equation in Equation (2.17), taking advan-
tage of the continuity condition and integrating over a periodic domain Ω, the evolution
equation for the contained kinetic energy E(t) = 12 ui ui can be derived as
Z
dE(t) µ 1 ∂ui ∂ui
+2 dx = 0. (2.19)
dt ρ Ω Ω ∂xj ∂xj
Note that the non-linear term does not contribute to the energy budget, as it just shifts
the kinetic energy between scales. Instead, the only mechanism governing the evolution
of the kinetic energy is the viscous dissipation.
2.1.2.3. Incompressible Navier-Stokes Equations in Spectral Space

Most of the computational and theoretical research on the basic turbulence dynamics
focuses on the abstraction of incompressible, homogeneous, isotropic turbulence in pe-
riodic domains [16, 155]. For these situations, a spectral description of turbulence is
highly appropriate, as it allows the examination of turbulent properties as a function of
wave number and of the resulting interactions, and as it forms the basis of the most com-
mon numerical schemes in basic turbulence research.
For a periodic velocity field in a cubic domain ~x ∈ [0, L]3 , with L “large” compared to
largest length scale in turbulence, each component of the velocity can be represented by
a three-dimensional Fourier series as
X ~
ui (~x, t) = ûi (~k, t)eI k·~x , (2.20)
~
k
where I denotes the imaginary unit and the summation occurs over all wave vectors
~k. We drop the temporal dependency in the following for clarity. Defining the Fourier
transform as Z LZ LZ L
1 ~
F~k {f (~x)} = 3 f (~x)e−I k·~x dx3 , (2.21)
L 0 0 0
the expansion coefficients thus become
Z L Z L Z L
1 ~
ûi (~k) = F~k {ui (~x)} = 3 ui (~x)e−I k·~x dx3 . (2.22)
L 0 0 0
Transforming the continuity condition of Equation (2.17) yields

∂ui
F~k = Iki ûi = I~k~
û = 0, (2.23)
∂xi
21
indicating that both vectors are orthogonal in wave space. Applying the transform term
by term to the momentum equation in Equation (2.17) results in:

∂ui ∂ ûi
F~k = ,
∂t ∂t

µ ∂ 2 ui µ
F~k = − k2 ûi ,
ρ ∂x2j ρ
(2.24)
1 ∂p ki
F~k = I p̂,
ρ ∂xi ρ

∂ (ui uj )
F~k = Ĝi (~k, t),
∂xj
with k2 = ~k · ~k. Following Pope [167], the pressure term cancels with −Ĝ|| , the compo-
nent of Ĝ in the direction of the wave vector. The remaining perpendicular component
can be expressed as:

~ ∂ (uj uk )
Ĝ⊥j (k, t) =F~ k = Ikk F~k {uj uk }
∂xk
  
 X ~ ′ X ~ ′′

=Ikk F~k  ûj (~k )e
′ I k ·~
x  ûk (~k )e
′′ I k ·~
x 
 
~
k′ ~
k′′
XX n ~ ′ ~ ′′ o
=Ikk ûj (~k′ )ûk (~k′′ ) F~k eI(k +k )·~x (2.25)
~
k′ ~
k′′
XX
=Ikk ûj (~k′ )ûk (~k′′ ) δ~k,~k′ +~k′′
~
k′ ~
k′′
X
=Ikk ûj (~k′ )ûk (~k − ~k′ ).
~
k′
Combining Equations (2.24) and (2.25) results in the spectral evolution equation for the
modal coefficients of one component of the velocity field in wave space
X
∂ µ
+ k2 ûj (~k, t) = −IPjk (~k)kl ûj (~k′ , t)ûk (~k − ~k′ , t), (2.26)
∂t ρ
~
k′
| {z }
Tj (~
k,t)
with the operator P that projects onto the plane orthogonal to ~k as

ki kj
Pjk (~k) = δjk − 2 , (2.27)
k
22
2.2. A Model Turbulence Problem: The Taylor-Green Vortex
and Tj (~k, t) called the non-linear transfer operator. Equation (2.26) thus describes the
incompressible Navier-Stokes equations in spectral space. The terms on the left hand
side only contain modal content at wave vector ~k, i.e. the temporal derivative and the
viscous terms do not couple different scales. This coupling is introduced by the non-
linear term on the right hand side, as discussed in Section 2.1.1. It contains velocity
modes at wave vectors k~′ and k~′′ with k~′ + k~′′ = ~k, i.e. only a subset of the total
occurring wave vectors [30]. The triads formed by the three interacting wave vectors
are separated into local and non-local interactions, depending on their relative location
in spectral space [63]. In physical space, local interactions are those among structures
of comparable length scale, while non-local contributions stem from the interactions of
different-sized eddies.
The evolution equation for the kinetic energy of a wave vector E(~k, t) can be derived
from Equation (2.26) by multiplication with the complex conjugate û∗j (~k, t) and 21 , lead-
ing to

d µ
+ 2 k2 E(~k, t) = T̂ ∗ (~k, t), (2.28)
dt ρ
with
P the modified transfer term T̂ ∗ . When summing over all wave modes ~k, the sum
∗
k T̂
~ is equal to zero, i.e. this term does not contribute to the overall balance of
kinetic energy. Thus, the change of the total kinetic energy is only governed by the
viscous dissipation, while the non-linear term governs its distributing among the modes
(see Sections 2.1.1 and 2.1.2.2).

The three-dimensional, viscous Taylor-Green vortex flow (TGV) is a classical example
of periodic, decaying turbulence. It has been analyzed extensively both theoretically and
numerically to understand transition and turbulence and to evaluate numerical schemes.
Its inviscid version is conjectured to develop a vorticity singularity in finite time, but the
theoretical and numerical evidence where deemed inconclusive [183]. We describe its
features here briefly, as it will be used as a benchmark problem in later chapters.
Originally proposed by [190] as a mechanism of producing a scale cascade for the in-
compressible Navier-Stokes equations, the kinetic energy of the √ TGV is initially con-
tained in 8 Fourier modes on the wavenumber shell with radius 3. Starting from these
laminar initial conditions, the ordered large scale structures interact, break down and
undergo transition into a turbulent field with a wide range of temporal and spatial vor-
tical scales. Beyond a Reynolds number of ReCrit ≈ 1000, a turbulent kinetic energy
5
spectrum with the expected k− 3 slope in the inertial subrange develops and a nearly
23
isotropic velocity field is found beyond t = 7s [73]. Following Brachet [29], the ve-
locity and length scales of this problem are of order one, so the associated Reynolds
number is defined as Re = ν1 , with ν being the kinematic viscosity. The dissipation
rate of kinetic energy peaks at around t = 9s, independent of the Reynolds number (for
Re > ReCrit ). Since no kinetic energy production at the large scales exists due to the
absence of a mean shear or forcing term, the vortical structures finally decay and the
energy cascade subsides.
The initial flow conditions in a triple periodic box of length 2π are given by
ρ = ρ0 = 1,
u1 = V0 sin(x1 )cos(x2 )cos(x3 ),
u2 = −V0 cos(x1 )sin(x2 )cos(x3 ),
u3 = 0,
ρ0 V02
p = p0 + (cos(2x1 ) + cos(2x2 )) (cos(2x3 ) + 2) , (2.29)
16
where ρ, ui with i = 1...3 and p denote the density, the components of the velocity
vector and the pressure, respectively and the velocity scale V0 typically set to 1. The
relationship for pressure is obtained from the solenoidal velocity field by solving Equa-
tion (2.18). Fig. 2.5 shows the development of this flow at Re = 5000 from an initial
mono-scale state to a full multiscale problem with final dissipative decay. As noted
by [29], the initial development of the flow is governed by near-inviscid roll-up and
stretching of the anisotropic large scale vortical structures. At about t = 3s, an onset of
instability is observed, which leads to the successive breakdown of the coherent struc-
tures. This breakdown is complete after about t = 7s, when the velocity field becomes
nearly isotropic. Due to lack of large scale kinetic energy production, the flow decays
self-similarly for t → ∞.
Figure 2.6 left depicts the spectrum of the kinetic energy after the transitional phase. The
spectrum clearly shows the multiscale nature of this flow as well as the characteristic
inertial subrange region connecting the large and small scales. The plot in Figure 2.6
right emphasizes the non-stationary behavior of this flow by depicting the evolution of
the kinetic energy dissipation through the development of the scale cascade, from the
laminar initial state through the transition to turbulence and the final viscous decay. The
total kinetic energy and the associated energy dissipation rate are given by
Z2π
ZZ Z2π
ZZ
1 2ν ~ u : ∇~
~ u d~x.
k := ~v · ~
u d~
u, ǫ := ∇~ (2.30)
16π 3 8π 3
0 0
24
Figure 2.5.: Temporal evolution of the Taylor-Green vortex, from upper left to lower
right at time t = 0.4s, 1.4s, 2.7s, 5.9s, 8.9s and 15.5s Shown are contours
of vorticity, colored by relative helicity.
Due to its easily reproducible initial and boundary conditions yet complex physical be-
havior, the Taylor-Green vortex flow is a prototypical case the study of turbulence. It is
a widely used benchmark for both DNS and LES simulations (see e.g. [28, 60, 69, 73]).
Orszag [154] sums up the rationale for the prominence of the TGV in the numerical
community:
“In addition to the fundamental fluid dynamical interest in the devel-
opment of the Taylor-Green vortex, the flow is a most convenient one on
which to debug and perform tests of sophisticated three-dimensional nu-
merical hydrodynamics simulation codes.”
25
0.014
10-1
0.012
-3
10
0.010
-5
0.008
10
-dk/dt
E(k)
0.006
10-7
0.004
-9
10 0.002
0.000
100 101 102 0 5 10 15 20
Wavenumber k Time
Figure 2.6.: Integral quantities of the Taylor-Green vortex flow.

Left: Spectrum of kinetic energy for Re = 5000 at t = 9.0, dashed line:
k−5/3 slope.
Right: Temporal evolution of the rate of kinetic energy dissipation for Re =
1600.
2.3. Numerical Simulation Strategies for Multiscale Flows

Due to the complexities of the multiscale nature outlined in Section 2.1, the simulation
of turbulent flows poses some unique challenges, not only from a numerical, but even
from a conceptual standpoint. In this section, we briefly give an overview of the classical
strategies and continue with a deeper discussion of LES in Section 2.4.
2.3.1. DNS
The Direct Numerical Simulation (DNS) aims at finding an exact solution to the full
incompressible or compressible Navier-Stokes equations in physical or spectral space
(Equations (2.10), (2.17), (2.26)). An often overlooked implicit premise of this approach
is the validity of the continuum assumption at the smallest scale level, where large local
Knudsen numbers can occur [88]. However, this assumption is often made, and conclu-
sive evidence to the contrary is still lacking.
By design, in DNS, no modeling error is introduced, however, approximation errors and
the sensitivity to perturbations discussed in Section 2.1.1 may influence the solution.
Clear advantages of a DNS strategy are the availability of the full spatial and temporal
26
information of the solution, while the numerical costs restrict its application to flows of
low complexity. The major advantage from a simulation perspective is the (approximate)
independence from the underlying numerical scheme, i.e. if the resolution requirements
dictated by the physics (the Reynolds number) and the numerics (the requirements to
resolve the smallest occurring scales) are met, the choice of the numerical discretization
becomes arbitrary. Thus, from a numerical standpoint, a DNS is driven by the search for
the most efficient and accurate numerical solvers.
2.3.2. RANS
Whereas DNS resolves all the scales of turbulent motion, the RANS (Reynolds-Averaged
Navier Stokes) approach is situated at the opposite end of the resolution spectrum. This
approach abandons the attempt to resolve the scales of motion, but assumes the motion
to be random about a mean and thus adopt a statistical treatment. Reynolds [173] de-
composed the flow field into a mean Φi and fluctuating component φ′ i of a variable φi
as
Z
1 T
Φi = lim φi (t) dt,
T →∞ T 0 (2.31)
φ ′ i = φ i − Φi ,
which is valid as long as a wide scale separation between the averaging period T and
the time scale of the variation of the mean exists. Applying this averaging operation to
the incompressible Navier-Stokes equations yields their RANS formulation (details and
derivation found e.g. in [204]):
∂Ui
=0
∂xi
(2.32)
∂Ui ∂Ui 1 ∂P 1
+ Uj + = σij − ρ u′ i u′ j ,
∂t ∂xj ρ ∂xi ρ
where capital letters denote the mean component and the overline denotes the explicit
application of the averaging operation in Equation (2.31). σij thus denotes the stress
tensor of the mean flow quantities, and u′ i u′ j the unclosed Reynolds turbulent stresses.
Finding a suitable closure, i.e. a model, for this term is the main challenge of the RANS
approach.
2.3.3. Large Eddy Simulation

The idea of filtering (or averaging) the high amount of information contained in a mul-
tiscale flow to reduce complexity and thus make it computationally tractable can be
27
translated from the temporal domain (RANS) to the spatial domain (LES). Formally, the
LES formulation the Navier-Stokes equations can derived by applying a spatial filter op-
erator to the equations, leading to an evolution equation for the filtered quantities with
an additional unclosed (subgrid) term. Details will be given in Section 2.4. The rationale
for solving the spatially filtered equations instead of the full equations is three-pronged:
1. Computational necessity: As noted in Section 1.3, the space-time scale resolution

of the numerical solution of turbulence must be at least as fine as the continu-
ous problem, i.e. the smallest dynamically active scale must be resolved with
the appropriate numerical accuracy. Since this becomes arbitrarily costly, such a
DNS resolution remains restricted to low Reynolds numbers [105]. In LES, ei-
ther an explicitly or implicitly enforced filter width ∆ is introduced, which fixes
the smallest occurring scale and thus gives direct control over the computational
costs.
2. Structural stability of the filtered solution: The concept of a solution trajectory

and an attractor in phase space has been introduced in Section 2.1.1, where sin-
gle realizations of a solution are highly sensitive to small scale perturbations, but
the general structure of the solution is invariant to them. While this is generally
true for the turbulence of the Navier-Stokes equations, it can be shown that the
truncated (or filtered) Navier-Stokes equations have attractors that are comparable
in their large scale structural layout to those of the continuous Navier-Stokes sys-
tem [191]. In other words, the DNS solution in phase space and the LES solution
in the corresponding truncated space share a number of properties. How large this
overlap is depends strongly on the filter width ∆, the subgrid modeling and the
numerical realization of the LES approach. This commonality between the two
solutions and their attractors is the reason why the average and higher moment
properties LES solutions are often in good agreement with their DNS counter-
parts, and thus, a numerical simulation based on the filtered equations can give
meaningful predictions [182].
3. Universality of the model region: The scale separation through filtering intro-
duces a need for the incorporation of the subgrid effects back into the filtered
formulation. A modeling approach for these small scale properties depends on
the assumption of universality of these scales, i.e. their independence from large
scale boundary conditions. While this small scale “memory loss” and approach to
isotropy is supported by the behavior of homogeneous, isotropic, laboratory-type
experiments [16], it immediately loses validity in the presence of boundary lay-
ers, where the imposed boundary conditions influence the full range of scales. An
28
additional underlying assumption for the classical modeling approach is the cas-
cade of kinetic energy, in which the energy content of the scales decreases rapidly
with increasing wave number. Thus, only a very small percentage of large scale
structures carry the overall dynamics of the flow and thus need to be resolved for
a good approximation to the full solution.
Figure 2.7 depicts a model spectrum based on Kolmogorov’s understanding of
the turbulence cascade. While this view is only strictly valid for homogeneous,
isotropic turbulence, it is often applied qualitatively to other high Reynolds num-
ber situations as well, where the isotropy of the small scales due to their large
separation from the boundary conditions is assumed. Loosely, the wave num-
ber range covered by the turbulent scales of motion can be separated into three
regions:
I) The large scale region, where the scales are of a size typically determined
by the boundary conditions. These scales introduce kinetic energy into the
scale cascade and act as a low-frequency feeding of the system, typically
through a mean shear or large scale vortical motion. Dissipative effects are
negligible in this range. As these scales are strongly anisotropic and non-
universal, they need to be resolved accurately in a LES approach.
II) In the inertial subrange, the spectrum of the kinetic energy is solely depen-
dent on the dissipation rate ǫ at the small scales (a function of the Reynolds
number) and the local wave radius k, following Kolmogorov’s law [116,
117]:
2 5
E(k) ∼ ǫ 3 k− 3 . (2.33)
In this region, the dissipation time scale is much larger than the time scale of
vortex interaction, and thus, energy is only transferred between scales, but
not dissipated. It should be noted that the redistribution is on average down-
scale, but backscatter exists. Equation (2.33) stresses that the dynamics of
the inertial subrange are not dependent on the large scale feeding mecha-
nism or the boundary conditions, but universal. Thus, for LES, the modeling
approach requires the resolution of scales into the inertial subrange, where
non-universal effects subside.
III) With increasing wavenumber, the viscous effects overcome the vortex inter-
actions, thereby imposing a lower limit on the smallest occurring scale η.
The high frequency fluctuations lead to isotropy and independence from
boundary conditions in this dissipation region. It is important to note that
the scale separation between regions I) and III) increases with the Reynolds
number according to Equation (1.2), and thus the assumption of isotropic,
29
10-1
10-2
-5/3
k
-3
10
-4
10
-5
E(k)
10
-6
10
10-7
10-8 I II III
-9
10
-10
10 0 1 2
10 10 10
Wavenumber k
Figure 2.7.: Spectrum of kinetic energy of isotropic homogeneous turbulence (Taylor-

Green Vortex at Re = 5000, t = 11s). I: Anisotropic large scale energy
production, II: Inertial subrange region, III: Small scale isotropic dissipa-
tion range.
universal small scales amenable to modeling becomes more valid as well.

Therefore, two features of the dissipation region support the modeling ap-
proach of the unresolved scales: a) Their (approximate) independence from
large scale anisotropies and b) their dominant, viscous action, which turns
organized, directed kinetic energy into isotropic, chaotic internal energy.
One counterexample of this general picture occurs in reacting flows, where
molecular mixing and reaction kinetics occur within this region, and thus
render it non-isotropic and non-universal [140].
In summary, the rationale behind LES combines the need for reduced computational cost
with physical properties that support a resolution of anisotropic large scale structures and
the modeling of universal small scale features.
It should be noted that while LES can result in a significant reduction in complexity,
the range of anisotropic scales can still make it very challenging in terms of computa-
30
2.4. Aspects of Large Eddy Simulation
tional effort. Various alternative, but related methods like URANS (unsteady RANS) and
VLES (Very Large Eddy Simulations) as well hybrids like DES (Detached Eddy Simu-
lation) exist, which all aim at further reducing the computational costs. An overview is
given in for example in [177].

In the previous section, we have given a brief introduction into the rationale behind LES.
In the following section, we will provide further details on the method and on some of
its important aspects. More general overviews and detailed discussion can be found in
e.g. [81, 90, 127, 177].
2.4.1. General Formalism

As stated above, the basis of the LES approach is to solve the conservation equations
governing the multiscale problem in a suitably filtered form, with reduced spatial and
temporal degrees of freedom. In the following section, we will derive the LES formula-
tion and its different subversions in an abstract manner.
Generally, we can formulate any multiscale problem as
P (u) = 0, (2.34)
where u is the (full) solution to the multiscale problem and P (.) is the operator, e.g. the
Navier-Stokes operator. Equation (2.34) can thus represent any of the conservation laws
described in Section 2.1.2, with u representing the vector of conserved variables. Each
LES method is by definition a reduction in degrees of freedom of the solution, so the
first step is to find a lower-dimensional representation of u in terms of limited degrees
of freedom. When thus discretizing this multiscale problem, we first approximate the
solution u (by choosing a computational grid or by limiting the number of degrees of
freedom)
u ≈ ū. (2.35)
In this first step, a representation of the solution on the computational grid with reduced
number of degrees of freedom is sought. Naturally, this automatically results in a trun-
cation error, which describes the differences between the full and reduced-dimension
solution ǫtrunc = u − ū. This operation (denoted by the bar) can be interpreted and
implemented either as a projection or more generally as a filter step. The filter type and
the filter resolution depends on the choice of the grid and on the choice of how to rep-
resent the approximation, e.g. in terms of local or global polynomials or discrete point
31
values. While the solution is now represented on the grid, the operator itself is naturally
also subject to a discretization:
P (.) ≈ Ph (.), (2.36)
which depends strongly on the numerical method (e.g. finite difference (FD), finite
volume (FV), finite element (FE)) and the numerical model parameters (e.g. order of
the method, stencil choice, numerical flux function). In turn, the approximation of the
operator usually introduces an error, which will manifest itself in an approximate, filtered
solution ūh . Thus, in practice, in an LES, one solves the problem
Ph (ūh ) = 0, (2.37)
instead of the multiscale problem (2.34). The natural question to ask is how good the so-
lution ūh approximates that of the multiscale problem. This question can be formulated
in two different ways:
1. How good does ūh approximate u?
2. How good does ūh approximate ū?
While both questions are valid, they focus on different aspects of the solution of Equa-
tion (2.37). Of course one is typically interested in the answer to question one, i.e. how
the simulation result ūh including all assumptions and approximations compares to the
solution u of the continuous problem. However, as we are dealing with multiscale prob-
lems and with their severely under-resolved discretizations, the comparison of u with ūh
might be unfair due to the large truncation error caused by the inherent grid induced fil-
tering. As stated above, the first step in a discretization u ≈ ū causes a truncation error,
which significantly reduces the number of scales represented on the grid. Therefore, it is
reasonable to ask how the numerical approximation ūh compares to the truncation of the
multiscale solution ū on the same grid, i.e. question two. It should be noted here that the
initial selection of a grid (or of the number of degrees of freedom one can afford to com-
pute) thus plays a primary role in what to expect from an LES and that even an error-free
numerical discretization cannot reproduce the continuous solution with limited degrees
of freedom - at least not in all quantities of interest.
To further discuss question two and to derive the general LES strategies in terms of
filtering and modeling approaches, it is possible to introduce a general problem formu-
lation for the filtered solution ū by (analytically) filtering the whole multiscale problem
itself. This is considered the traditional approach to the derivation of the LES formula-
tion [177]. Starting from the continuous multiscale problem in Equation (2.34) again,
32
f and arrive at
we define a mathematical filtering operation with a general filter by (.)
]
P (u) = 0,
P (e
u) = P (e
u) − P](u) . (2.38)
| {z }
:=S
It should be noted that this filtering process is at the root of each LES formulation, re-
gardless whether it is executed explicitly or implicitly. It is this step that reduces the
number of degrees of freedom of the problem and yields a mathematical description of
the reduced multiscale problem. A general expression for this one-dimensional spatial
filter operation is given by
Z +∞
g = 1
u(x) G
x−ξ
u(x) dξ, (2.39)
∆ −∞ ∆
where the convolution kernel G is defined by the filter used, and is associated with the fil-
ter width ∆. Multidimensional filtering can be achieved in a tensor-product dimension-
by-dimension manner. To separate this filter from the filter introduced by the compu-
tational grid introduced in Equation (2.35), we indicate this explicit filtering process by
f Compared to the original multiscale problem (2.34), the filtered problem (or the LES
(.).
formulation) contains the additional term
S := P (e ]
u) − P (u), (2.40)
which describes exactly the effect of subfilter scales on the filtered solution u
e. Note that
the solution in terms of ue requires the modification of the original problem by the addi-
tion of the source term S. It is this term, or more precisely the effects of this term, that
are the cause and target for a subgrid multiscale model such as e.g. turbulence models.
The complexity of the modeling task arises from the fact that the term S depends on the
full multiscale solution u via the term P](u) and thus cannot be determined solely by the
filtered solution u
e. In other words, while the reduction of the complexity of the problem
seems to have been achieved by replacing P ] (u) with P (e
u), the generated source term
re-introduces the dependence on the full solution u again. The question of why such
an approach is then beneficial at all lies in the basic assumptions of LES described in
Section 2.3.3: By choosing a sufficiently wide filter that filters out only the universal
scales and retains the anisotropic scales in the resolved solution u e, the resulting source
term S becomes universal and mainly dissipative in nature, which allows its replacement
by suitable models.
In fluid dynamics, the dependence of the source term on the solution itself is often re-
ferred to as the closure problem (of turbulence), as an exact closure of this unknown term
33
is impossible without access to the full solution u. Three basic strategies exist for the
closure and filtering strategies of these multiscale problems. To incorporate the effects
of a closure formulation into the general formalism, we introduce a suitable model for
the subgrid source terms into Equation (2.34):
P (u) + M (u) = 0, (2.41)
where we have formally added the operator M (u) to denote any type of model as a
function of the fully resolved u. In particular, to recover the original multiscale problem
Equation 2.34, M (u) = 0 must hold, i.e. in a fully resolved setting, the effect of the
model must be zero. We introduce this term here to show how the model term appears
in the further derivation of the LES formulation.
We now repeat the formal filtering of the whole problem to derive an expression for the
filtered solution u
e:
]
P ^
(u) + M (u) = 0,
(2.42)
]
P (u) + P (e
u) − P (e ^
u) + M (u) + M (e
u) − M (e
u) = 0.
Rearranging yields the LES formulation for Equation (2.41) as
P (e
u) + M (e
u) = S,
(2.43)
]
S = −P (u) + P (e ^
u) − M (u) + M (e
u).
The source term S re-appears as an additional term that incorporates the effects of the
subfilter solution u onto the filtered part u
e. In comparison with the source term in Equa-
tion (2.40), a model contribution M (e u) occurs, which is generally non-zero, even for a
consistently defined model.
Following the discussion leading to Equation (2.37), the representation of Equation (2.41)
in a numerical scheme is of the form
Ph (f
uh ) + Mh (f
uh ) = 0, (2.44)
where the solution u is first filtered by an explicit filter u

e defined in Equation (2.39).
As discussed above, the introduction of a numerical grid generates an additional filter u e,
acting on the analytically filtered solution. This discretization filter function inherits its
properties from the choice of the discretization, and can represent a FD, FV, FE or any
other type of consistent discretization strategy. As an example, the FV discretization can
be interpreted as a box filter in physical space, where the filter width is given by the cell
size and the convolution reduces to an averaging operation.
34
The numerical scheme itself then introduces an approximation error in the representa-
tion of the operators P and M , i.e. the double-filtered solution u
e is advanced in time
by inexact operators Ph and Mh , which in turn leads to an approximate solution u
fh , the
final result of any computed LES solution.
To put Equation (2.44) into the context of the LES-formulation of the filtered problem
in Equation (2.43), we can manipulate Equation (2.44) by addition and subtraction of
suitable terms:
uh ) + Mh (f
Ph (f uh ) = 0 ⇔
P (f
uh ) + M (f
uh ) = P (f
uh ) + M (f
uh ) − Ph (f
uh ) − Mh (f
uh ) ⇔
P (f
uh ) + M (f
uh ) = P (f
uh ) + M (f
uh ) − Ph (f
uh ) − Mh (f
uh ) (2.45)
]
−P ^
(u) − M ]
(u) + P ^
(u) + M (u).
Rearranging the last expression in Equation (2.45) according to the terms in Equa-
tion (2.43) yields
uh ) + M (f
P (f uh ) = S,
]
S = −P (u) + P (f ^
uh ) − M (u) + M (f
uh ) −Ph (f
uh ) − Mh (f ]
uh ) + P ^
(u) + M (u) .
| {z }
!=0
(2.46)
Comparing this expression with the LES-formulation in Equation (2.43), i.e. comparing
the analytically filtered continuous LES formulation with the actual discretized version,
we note that we recover the LES equation in terms of u fh if the last four terms on the
right hand side cancel out. In that case, Equation (2.46) and Equation (2.43) (repeated
here for clarity) read as:
P (e
u) + M (e ]
u) = −P (u) + P (e ^
u) − M (u) + M (e
u),
(2.47)
P (f
uh ) + M (f ]
uh ) = −P (u) + P (f ^
uh ) − M (u) + M (f
uh ).
Thus, the discretized and grid-filtered solution reverts to the true solution of the under-
lying LES problem, i.e.
u
fh = ue. (2.48)
35
This is however only exact if the marked terms in Equation (2.46) cancel out. A closer
look at this general error term reveals that it contains some interesting features. In partic-
ular, it allows a distinction between the three different LES strategies in terms of filtering
and modeling that exist. To describe these strategies, we start by rewriting this term as
]
P uh ) = Mh (f
(u) − Ph (f uh ), (2.49)
^
where we have left off the term M (u), as by initial definition, for a consistent model,
M (u) is zero. It should however be noted that a number of commonly used models,
among them the classical Smagorinsky model, indeed violate this condition [186]. From
this equation, we can now recover the features of the different LES strategies.
2.4.1.1. Implicit Filter and Implicit Modeling
A natural choice is to choose the filter of the multiscale problem (.)f identical to the filter
induced by the computational grid (.), i.e. not to perform an explicit filtering step in
addition to the discretization. This is referred to as implicit filtering, as this is implicitly
part of the discretization process as discussed above. Thus, the discretization combines
the effect of these two filter operations into a single one. Consequently, all filters in the
error formulation Equation (2.49) are represented by the single filter function (.). In
addition, a choice is made to introduce no explicit modeling term M , so we can drop the
terms relating to the modeling. This reduces Equation (2.49) to
P (u) − Ph (uh ) = 0, (2.50)
and analogously the solution equation (Equation (2.48)) to
uh = u. (2.51)
Thus, the requirement for this type of LES (and the source of its error) is to construct a
discretization (consisting of both the implicit filtering/projection (.) and the discretiza-
tion properties of Ph ) that models the effect of the subgrid scales P (u), i.e. the modeling
is achieved by tuning the numerical properties and therefore the approximation errors.
One feature of this LES strategy is that for h → 0, both the discretization error in terms
of h and the filter width go to zero (not necessarily at the same rate), so that the filtering
and the operator can be commuted and the original LES formulation converges to the
unfiltered original multiscale problem (Equation (2.34)), and thus
lim uh = u. (2.52)
h→0
36
This means that the solution of this approach converges to the DNS solution of the unfil-
tered problem, which takes full advantage of the available degrees of freedom but makes
the establishment of grid convergence in a LES resolution regime impossible, as the un-
derlying problem is defined by the grid itself.
Thus, the modeling approach in this strategy consists of constructing the numerical
method in such a way that it models the effect of the subgrid scales. As this model-
ing approach is implicitly given by the numerical approximation, this strategy is referred
to as implicit filtering with implicit modeling of the subgrid scales. While the formula-
tion for this strategy was directly derived from the error Equation (2.49), an alternative
route is to approach directly from the general multiscale formulation, again noting the
implicit filtering and the omission of the model (Equation (2.43)):
P (u) = P (u) − P (u). (2.53)
Comparing with the discretized multiscale problem (2.37) gives
P (u) = P (u) − P (u),

(2.54)
P (uh ) = P (uh ) − Ph (uh ),
which again results in the conclusion that the term P (u) − Ph (ūh ) has to vanish for
uh = u, as shown above.
2.4.1.2. Implicit Filter and Explicit Modeling

It should be noted that the implicit modeling approach burdens the discretization scheme
with two tasks: Being suitable for the correct representation of the resolved parts of the
scales, and having a meaningful approximation error for modeling of subgrid terms. If
these two tasks are to be separated, an explicit modeling term M can be introduced.
Then, Equation (2.49) reduces to
P (u) − Ph (uh ) = Mh (uh ). (2.55)
From this expression, it is clear that the discretization and the modeling now together
have to balance the subgrid term P (u), i.e. that there is a strong interaction of numerics
and model. This can be seen by considering an error-free discretization, i.e. the discrete
operators are equal to the continuous ones. Thus, dropping the subscript h results in
P (u) − P (u) = M (u), (2.56)
which highlights the fact that only for this idealized case, the model and the scheme
would indeed be independent from each other. This interaction of explicit model and
37
numerics can be troublesome, as the derivation of explicit closures is usually done inde-
pendent from any numerical scheme, i.e. for an “ideal” discretization. In practical LES,
the resolution of the underlying problem is by definition coarse, and if the model does
not perfectly smooth it on the given grid, interactions of numerics and model become
unavoidable.
Another interesting aspect of this approach is that grid refinement (and thus a vanishing
filter) does not necessarily lead to the DNS result, as the effect of the model for the full
solution u comes into play:
P (u) − P (u) = M (u) (2.57)
If the model is chosen is such a way that M (u) 6= 0, i.e. if the model also acts in fully
resolved regions, then a refined discretization h → 0 will not lead to the DNS solution.
Instead, a source term of the form M (u) will pollute the original problem. One example
of such a model that does not return zero contributions in the fully resolved case but
introduces unwanted dissipation is the classical Smagorinsky model [186]. This should
be kept in mind, as the most commonly used form of “engineering LES” is indeed an
implicitly filtered approach with the classical Smagorinsky model.
2.4.1.3. Explicit Filtering with Explicit SGS Model

For this most general case, an explicit filter step is used for scale separation, and the
influence of the subgrid terms is closed again via an explicit modeling term. Thus, the
error term remains as derived in Equation (2.49):
]
P (u) − Ph (f
uh ) = Mh (f
uh ), (2.58)
In this case, if the discretization parameter is driven to a value well below the grid cut-off
scale h ≪ ∆, the error term becomes
]
P (u) − P (e
u) = M (e
u), (2.59)
which is exactly the source term S in Equation (2.42). Thus, for h ≪ ∆, the solu-
tion does not converge to the DNS solution u but to the explicitly filtered solution ue.
This is the major advantage of explicit filtering, as the decoupling of the model and
the numerical scheme (no dependence on the grid filter or the discretization remains in
Equation (2.59)!) allows an independent development and evaluation of both.
In the next section, we will demonstrate this strategy and show the closure with a “per-
fect” LES model.
38
2.4.2. The “Perfect” LES model

In a “perfect” LES model, by definition, the model term M (e u) based on the filtered
scales perfectly matches the difference between the filtered operator and the operator
applied to the filtered solution, i.e. Equation (2.59). The existence of such a model
would thus lead to the “perfect” LES solution. Clearly, no general model that fulfills
these properties exists. However, an a posteriori generation of such a model from the
full space-time information of a DNS is possible by defining a suitable filter. By an-
alyzing this perfect closure, both the physics of scale dynamics and the coupling and
interaction of the exact model and the discretization can be investigated [58, 146]. In
this section, we will derive the perfect model equation based on the concepts from Sec-
tion 2.4.1 and show that by explicit filtering, the solution becomes independent of the
discretization.
2.4.2.1. Problem Formulation

The one-dimensional viscous Burgers’ equation (Equation (2.1)) serves again as a sim-
plified model for the non-linear wave interactions and small scale dissipation that char-
acterize turbulent flows. The equation includes all the relevant terms for scale production
and removal and reads in advection form as
ut + u ux − νuxx = 0. (2.60)
In the following, we will refer to Equation (2.60) as the “u”- or DNS-problem. Follow-
ing the explicit filtering approach to generate a perfect source term independent from
numerics, we define the filtered solution as (assuming that the temporal differentiation
and the filtering commute)
Z∞
u
e(x, t) = G(x, y, ∆) u(y, t) dy, (2.61)
−∞
where G(x, y, ∆) is the filter kernel and ∆ a general filter width. Applying this operator
to (2.60) yields the filtered Burgers’ equation
u
et + ug
ux − ν u
g xx = 0. (2.62)
However, this formulation does not lend itself to a straight forward discretization due
to the non-linear term ug
ux , as its evaluation would entail knowledge of the unfiltered
quantity u ux (corresponding to P] (u) in Equation (2.59)). Since only filtered quantities
39
are available (those in terms of u

e), we derive the standard LES formulation by rewriting
Equation (2.62) as
u
et + ueuex − ν u
exx = S(x, t), (2.63)
with
S(x, t) = (e
uuex − ug
ux ) − ν (e
uxx − u
g xx ) , (2.64)
where we did not assume commutation of spatial differentiation and filtering. The source
term S(x, t) represents the effect of the subgrid scales u′ := u(x, t) − ue(x, t) onto the
filtered solution u
e. It is this expression that cannot be computed exactly during a coarse
grid simulation and hence has to be closed by an appropriate model. Note that this source
term exactly matches the formal definition in Equation (2.38).
2.4.2.2. The Perfect Closure

With the assumption that an exact solution to Equation (2.60) exists, the source term in
Equation (2.64) can be evaluated exactly for a given filter definition. This requires of
course the availability of the full temporal and spatial solution u. Assuming that such
a source is available, we can define the following new problem for v(x, t) (henceforth
called the “v-problem”) as
vt + v vx − νvxx = S(x, t). (2.65)
The v-problem shares the same structure as the original LES formulation in Equa-
tion (2.63), so for an exact solution of the v-problem (Equation (2.65)), the solution
v is identical to the filtered solution of the u-problem (Equation (2.60)):
v(x, t) ≡ u
e(x, t). (2.66)
Importantly, a full resolution of the v-problem requires significantly less effort (depend-
ing only on the filter width, i.e. the number of scales in v) than the solution of the
u-problem, as v is smoother than u due to the inherent regularization through filtering.
In this sense, the source term (2.64) constitutes the perfect closure, as the exact solution
of the associated v-problem leads to the exact reproduction of the large-scale solution u e.
In particular, the only parameter in the definition of the source term is the selection of a
suitable filter, i.e. no assumptions about the numerical scheme have to be made.
2.4.2.3. LES with Perfect Model

To demonstrate this approach and its advantages, we apply it to the viscous Burgers’
equation with sinusoidal initial and periodic boundary conditions:
u(x, 0) = sin(2πx) + 0.5 for x ∈ [0, 1] and ν = 0.003, (2.67)
40
where the exact solution of this problem develops a range of at most K scales (wavenum-
bers) limited by the viscosity ν. Note that the resulting single shock is not “turbulent”,
but it is a suitable model problem with a scale range produced by non-linear interactions
and viscous small scale dissipation. For the filter definition, we choose the most natu-
ral formulation: a sharp spectral cut-off filter with a cut-off wavenumber Kc << K,
such that waves with wavenumber k > Kc are considered unresolved scales. Thus u e
(and consequently v) are bandwidth-limited to Kc frequencies. From the source term
definition in Equation (2.64), it is also obvious that the maximum range of scales of S
is 2 Kc due to the non-linear term u euex . Consequently, a resolution of at most at most
2 Kc is required to solve the v-problem exactly, in contrast to K scales for the unfiltered
u-problem. This “saving potential” in terms of required degrees of freedom for an exact
computation of u e = v is
2Kc
, with Kc << K, (2.68)
K
which is one of the three main rationales for the Large Eddy Simulation approach de-
noted in Section 2.3.3.
Following the method outlined later in Section 3.2, a pseudo-spectral Fourier-based one-
dimensional DNS solver without de-aliasing with a 5-stage 4th order Runge Kutta time
integrator is used to generate the “exact” DNS solution of the u-problem. The total num-
ber of grid points was chosen as Nu = 1024, resulting in a resolvable scale band up to
K = 512 due to the Nyquist criterion. For an end time of t = 0.25, this resulted in 7865
time steps based on the advective time step criterion and also ensuring a DNS resolution
in time. Figure 2.8 shows the DNS solution u and the filtered solution u e at t = 0.25.
For the spectral filter, a cutoff wavenumber of Kc = 16 was chosen. The right part of
the figure shows the corresponding energy spectra at time t = 0.25, revealing that the
DNS resolves all scales down to machine precision and the position of the cut-off within
the self-similar scale range.
The corresponding source term is computed from the DNS solution for each Runge-
Kutta stage and stored as a space-time matrix of the source S = S(x, t) for the definition
the v-problem. Figure 2.9 shows the energy spectra of this source for different times t,
confirming the observation that the source consists of at most 2Kc = 32 waves as
required by the filter definition. As a side note, the temporal development of the source
term is consistent with the development of the shock structure, as it shows the injection
of energy into the higher modes and the near equi-distribution among these modes. With
this source term available, the v-problem is now completely defined. Since it is thus
independent from the original problem, it can be solved by any discretization.
In a first step, we re-use the same pseudo-spectral framework. Since the source contains
32 modes, Nv = 64 grid points are needed for the pseudo-spectral method to correctly
represent the source. To focus on the spatial aspects only, we choose the same time step
41
1.75
10-2
1.25 -7
10
0.75 10-12
Energy
u
-17
10
0.25
10-22
DNS
-0.25 Filtered Solution
-27
10
DNS
Filtered Solution
-0.75 10-32 0
0 0.2 0.4 0.6 0.8 1 1 2
10 10 10
x k
Figure 2.8.: Moving shock solution of the viscous Burgers’ equation.

Left: DNS and filtered solution at t = 0.25. Right: Corresponding energy
spectra.
1
10
-4
10
-9
10
Source Spectrum
10-14 Time
-19
10
10-24
-29
10
0 5 10 15 20 25 30 35 40
k
Figure 2.9.: Temporal evolution of the energy spectrum of the source S(x, t).
size for the discretization of the v-problem as for the DNS of the u-problem. This is of
course not a general necessity, as the space-time information could easily be interpolated
onto the time grid of the v-problem. For the spatial discretization of the v-problem with
grid points Nv < Nu , the source term is spectrally projected onto the grid with Nv
42
points.
Table 2.1 shows the convergence of the solution of the v-problem towards the filtered
DNS solution ue. As expected, the L∞ -norm of the difference of v and ue reaches machine
zero (with some roundoff errors) when Nv ≥ 64, confirming that the exact solution
of the v-problem is equal to the filtered DNS solution of the u-problem. Figure 2.10
shows a zoomed-in plot of the different approximate v solutions and Figure 2.11 shows
the exact solution of the v-problem with Nv = 64 at t = 0.25. Thus, these findings
corroborate the claim that large scale simulations with suitable model can (theoretically)
be both accurate and stable at a fraction of the computational costs.
Nv 32 48 60 64 72
L∞ 2.82E-1 1.94E-2 1.30E-3 1.03E-14 1.62E-14
Table 2.1.: L∞ -errors with respect to u

e for different resolutions Nv of the v-problem at
time t = 0.25 with the pseudo-spectral method.
1.65
1.55 Filtered DNS

v-Solution (N=32)
v-Solution (N=48)
v-Solution (N=64)
1.45
u,v
1.35
1.25
1.15
0.35 0.4 0.45 0.5 0.55 0.6
x
Figure 2.10.: Comparison of different approximate solutions to the v-problem at time

t = 0.25.
It is worth noting that a computation of the u-problem with only Nu = 64 grid points
and without any source term, as plotted in Figure 2.12, results in an erroneous solution
which is even unstable due to the aliasing errors for this coarse resolution, furthermore
43
1.75
10-2
1.25 -7
10
0.75 10-12
Energy
v
-17
10
0.25
10-22
-0.25
-27
10
-0.75 10-32 0
0 0.2 0.4 0.6 0.8 1 1 2
10 10 10
x k
Figure 2.11.: Solution of the v-problem with the pseudo-spectral method.

Left: Solution v computed with Nv = 64 grid points at t = 0.25.
Right: Corresponding energy spectrum.
confirming the regularizing effect of the model (source) term.
Nv 64 96 128 256
L∞ FD O2 5.92E-1 1.66E-1 7.85E-2 1.56E-2

EOC - 3.1 2.6 2.3
L∞ FD O4 4.05E-1 3.58E-2 4.00E-3 2.41E-4

EOC - 6.0 7.6 4.0
Table 2.2.: L∞ -errors with respect to u

e for different resolutions Nv of the v-problem
at time t = 0.25 with the FD method.
In a second step we take advantage of the fact that the v-problem is an independent PDE,
and thus we are free to select a discretization strategy of our choice to find the solution,
a great advantage of explicit filtering which decouples the problem formulation (and
thus modeling) from the numerics. To demonstrate this, we discretize the v-problem
by central FD schemes of second and fourth order. Table 2.2 shows the results of h-
convergence for these simulations. We compute the L∞ norm of the difference of the
FD approximation vF D and the filtered DNS solution u e. The results clearly indicate
44
1.75 10
0
1.25 10
-1
0.75 10
-2
Energy
u
0.25 10
-3
-0.25 10-4
-0.75 10-5
0 0.2 0.4 0.6 0.8 1
10 20 30 40
x k
Figure 2.12.: Coarse solution of the u-problem with Nu = 64.

Left: Coarse solution u at t = 0.25.
Right: Plot of the corresponding energy spectrum.
that the design order of convergence is achieved. Thus we have a convergence towards
the solution of the v-problem, which means a grid convergence towards the coarse scale
(LES) solution. Furthermore, we can observe that the high order approximation is more
efficient in resolving the v-problem, yielding an overall better accuracy on the same grid.
Compared to the global method in Table 2.1, the FD approximations give a much larger
error for the same number of degrees of freedom, i.e. the faithful resolution of the scales
in the filtered problem requires considerably more degrees of freedom for low order than
for high order methods.
For completeness, Figure 2.13 shows the results of the fourth order FD solution with 128
grid points and the corresponding spectrum. Note that for this resolution, the Nyquist
wave number is 64, while the agreement between the DNS and FD spectra diverge at
about k = 18. Also, we note that coarsely resolved simulations based on central finite
difference scheme without the source terms are prone to aliasing and yield oscillatory
and even unstable solutions in these cases, again showing the regularization through the
source term.
Thus, the numerical solution of the regularized v-problem can be achieved by any suit-
able discretization method. The numerical solution will converge to the exact solution of
the underlying v-problem for a sufficiently high resolution, keeping in mind that differ-
ent numerical formulations can differ greatly in terms of approximation quality and thus
number of degrees of freedom for an essentially fully resolved DNS of this v-problem.
45
1.75
10-2
1.25 -7
10
0.75 10-12
Energy
v,u
-17
10
0.25
10-22
Filtered DNS
-0.25 O4 FD (128 cells)
DNS
-27
10 O4 FD (128 cells)
-0.75 10-32 0
0 0.2 0.4 0.6 0.8 1 1 2
10 10 10
x k
Figure 2.13.: Solution of the v-problem with the FD O4 method.

Left: Solution vF D computed with Nv = 128 grid points at t = 0.25 and
comparison to the DNS results.
Right: Corresponding energy spectra.
This gives a strong indication that high order schemes are well suited for this approach,
since they provide higher efficiency per DOF (or a lower number nP P W ) for resolved
approximations of smooth problems.
Regarding the applicability of this approach to actual turbulence modeling, it is obvious
that the exactness of the source term representation is crucial for the potential of success
of the method. Since an exact computation of the source term is practically infeasible in
almost all cases (and would defy the concept of a LES by relying on the solution of the
DNS), the approximation of the source term by a suitable closure model is essential.
In this section, we have demonstrated the most formal LES concept based on inducing
an external scale separation by filtering, i.e. regularizing the underlying problem inde-
pendent from the discretization choice. This approach clearly decouples the numerical
properties and the model definition, thereby allowing an independent choice of both:
Modeling can be based on physical considerations, while numerics can focus on effi-
cient scale resolution. Since high order discretizations are particularly efficient at the
latter, they are highly attractive baseline schemes for this type of LES.
46
2.4.3. Modeling Strategies

In this section, we will briefly describe typical closure models applicable to the three
formal LES approaches outlined in Section 2.4.1.
2.4.3.1. Implicit Models: MILES
One of the first purely implicit LES methodologies (both the filtering and modeling is
implicit, see Equation (2.50)) is the so-called MILES (Monotonically Integrated Large
Eddy Simulation) approach by Boris et al. [25], with a more recent update given in Grin-
stein et al. [93]. The base of this method are so-called monotone discretizations, mainly
developed for capturing strong compressibility effects (shocks) in a monotone way. A
prominent candidate is the flux correction transport (FCT) method developed by Boris
and Book [94]. In this method, a high order low dissipation approximation is combined
via a non-linear switch with a first order monotone flux in such a way that for well re-
solved cases, the low dissipation flux with its low numerical damping is used, while in
under-resolved cases, the monotone flux becomes dominant to ensure numerical diffu-
sion, guaranteeing an overall monotone approximation. Other monotone methods, such
as TVD FV methods or the PPM method are also used in this context [82]. By applying
such algorithms primarily designed for capturing shocks to under-resolved turbulence
computations, it was found that the non-linear dissipation mechanism inherent in the
FCT method provides a form of subgrid scale viscosity. In this case, the leading error
term of the discretization acts like a subgrid scale model. For an overview of applica-
tions we refer to [25, 93]. It is interesting to note that MILES is often used based on
Euler equations only, without the explicit viscous terms of the Navier-Stokes equations.
Investigations show that such a MILES computation converges towards a high Reynolds
number flow for sufficiently fine grids. This is not surprising, as in the limit of Re → ∞,
the actual form of the dissipation operator becomes less important. Due to the implicit
nature of the filtering and the FV character, such algorithms can be directly applied in
complex geometries. Investigations show that for sufficient fine grids (i.e. guaranteeing
a sufficient resolution of the large scales) MILES results compare well to reference data.
The fundamental disadvantage of the MILES approach lies in the origin of the mono-
tonicity of the method, i.e. the limiter. A limiter relates the grid to the evolution of the
represented solution and is thus strongly dependent on the grid topology. In this sense,
the implicit filtering is a disadvantage, as the filter shape and width is not known and
may vary throughout the grid, causing different subgrid dissipations independent from
the solution. Furthermore, this methodology does not satisfy the Galilean invariance as
a consequence. In summary, such an implicit LES type approach relying on the leading
error term as a subgrid scale model with non-linear limiter based dissipation can never
47
be expected to mimic the missing subgrid scales accurately, but numerical experiments
show that large scale motion can be simulated with reasonable accuracy.
2.4.3.2. Implicit Models: ALDM

The Approximate Local Deconvolution Method (ALDM) was introduced by Hickel [97]
as a form of implicit LES for incompressible flows. Again, the idea is to use the numer-
ical error terms as a replacement for an explicitly added LES model. Based on a FV
discretization, the filter is implicitly given by the grid via the spatial averaging. Decon-
volution refers to the inverse of a convolution filter and is thought of as a recovery of
small scales from the large scales. Hickel differentiates between a “hard” and “soft”
deconvolution, where the first describes the inversion of the filter into the unrepresented
scales (obviously, this problem is ill-posed), and the soft deconvolution refers to the re-
covery of the represented, but not (well) resolved scales. While the soft deconvolution
is feasible, it cannot account for the lack of the non-represented dissipation by pure de-
convolution alone, but requires an additional dissipative mechanism. Still, this approach
allows a certain useful methodology to construct LES models, e.g. Adams et al. [188].
In ALDM, deconvolution is based on a non-linear reconstruction similar to the WENO
(Weighted Essentially Non-Oscillatory) philosophy applied to shock capturing. In the
first step of the reconstruction process, several different biased stencils are constructed.
For each of the reconstruction polynomials, a measure for the oscillations is computed.
In the WENO method, the different reconstruction polynomials are combined according
to the oscillation indicators in such a way that the smoothest polynomial gets the highest
weight. These reconstruction polynomials are used to determine input values for suitable
numerical flux functions which typically have a dissipative nature. As an implicit LES
approach, the ALDM method tries to control the discretization error term by introducing
several free parameters into the numerical flux function as well as into the local WENO
type reconstructions. These parameters control the numerical discretization to a certain
extent and are used to tune the behavior of the method for under-resolved turbulence sim-
ulations. For this, a canonical turbulent benchmark problem, namely the homogeneous
isotropic turbulence (HIT) test case, is used. With an optimization approach based on
a genetic algorithm, the free discretization parameters are tuned in such a way that the
dissipation spectrum for the HIT simulation fits the theory for high Reynolds numbers.
ALDM offers basically the same advantages and disadvantages as the MILES approach.
A difference to MILES is the introduction of the free parameters which can be used to
tune the method. Originally, the ALDM method was developed for incompressible flow.
However with this tuning approach, a new parameter set was derived for compressible
flows [98]. On the one hand, the free parameters are an advantage as they allow the adap-
tation of the method to new physics, on the other hand it is obviously a drawback that
48
the performance of the method relies this heavy on auxiliary parameters. For instance
when parts of the flow problem are nearly incompressible and other compressible, the
set of parameters is difficult to choose. However, results from literature show that this
methodology compares favorable to a lot of standard explicit closures, e.g. [97].
2.4.3.3. Explicit LES Closures

Most of the explicit subgrid scale (SGS) model approaches can be combined with either
explicit filtering of the flow variables or an implicit filtering approach. As already dis-
cussed in the sections above, explicit filtering has a major impact on the performance
of the LES model and is thus not computationally efficient. The advantage of implicit
filtering is the lower computational effort as no additional filter operation is needed and
in theory all grid resolved scales (up to Nyquist) are used to represent the flow field. It
is important to note that with implicit filtering, the subgrid scale model parameters and
coefficients must be tuned depending on the actual used discretization and may even de-
pend on the grid topology. An approach based on explicit filtering is obviously free of
all those limitations as long as the explicit filter width is large enough in comparison to
the grid scale.
2.4.3.4. Explicit Model: Smagorinsky-Lilly

The Smagorinsky-Lilly model is the earliest and most commonly used SGS models
based on the eddy viscosity approach. It was developed and expanded by Smagorin-
sky and Lilly [186], and used by Deardorff in the first published LES [59]. It assumes
the instantaneous equilibrium of energy production at the large scales and its dissipation
at the small scales and gives a simple relation for the turbulent viscosity:
νt = (Cs ∆)2 |S|, (2.69)
where Cs is a tunable model constant, ∆ the filter width and S the large scale strain
rate tensor. The theoretical value for homogeneous isotropic turbulence was found to be
Cs = 0.18, but it has to be adjusted to the flow conditions and the numerical scheme
used. Due to the equilibrium assumption, the model performs poorly outside of isotropic
regions, especially in wall-near regions, where a modification by van Driest [68] is added
to account for the changing length scale. By construction, the model is always dissipa-
tive, even in well resolved and laminar regions and thus allows no backscatter. Due to
its simplicity and early availability, this model has been widely used in many flow situa-
tions (see e.g. [23]), even when its largest drawback - the inconsistency for well-resolved
regions - can play a significant role in the flow physics. Figure 2.14 depicts the influence
of the model on a DNS of laminar flow. Clearly, the model fails to regulate itself in
49
this situation, producing additional dissipation (as a function of its model constant) that
falsifies the solution.
Cs=0.0
Cs=0.1
Cs=0.2
Cs=0.3
0.01 Cs=0.4
-dK/dt
0.005
0.5 1 1.5 2 2.5 3 3.5

Time
Figure 2.14.: Smagorinsky model for laminar flow. Kinetic energy dissipation rate over
time as a function of the model constant Cs .
2.4.3.5. Explicit Model: Dynamic Smagorinsky

The shortcomings of the Smagorinsky model with its a priori fixing of the model constant
led to an improved version by Germano et al. [89], the so-called Dynamic Smagorinsky
model. The idea is to determine the model parameter Cs continuously during the compu-
tation from the smallest resolved scales. This is achieved by applying a (computational)
test filter that is wider than the grid filter-width. The “Germano identity” relates the re-
solved turbulent stresses Lij to the filtered subgrid and sub-testfilter stresses τc
ij and Tij
by
Lij = Tij − τc ij (2.70)
For both stresses, an eddy viscosity model is assumed, where νt is approximated as in the
Smagorinsky model (Equation. (2.69)). However, due to the similarity assumption, the
models for both stresses share the same model constant Cs . Thus, in combination with
50
relationship (2.70), this model constant can then be determined for each time and posi-
tion in space from a comparison of the subgrid and sub-testfilter stresses. This approach
is self-contained and thus parameter-free, and is self-adjusting to the flow situation. It
allows backscattering and requires no additional wall treatment, while remaining rela-
tively computationally efficient. It has been applied in a wide area of applications [55],
however, due to the possibility of negative values of Cs , numerical stability problems
can occur [91].
2.4.3.6. Explicit Model: Spectral Vanishing Viscosity
The concept of Spectral Vanishing Viscosity (SVV) was originally introduced by Tad-
mor [189] to stabilize Fourier-Galerkin discretizations of the shock solution occurring in
an inviscid Burgers’ equation. The basic idea is to add a spectral viscosity term (given
by an amplitude and a specific kernel function) with maximum amplitude at the high-
est wave number, but vanishing amplitude for low wave numbers. This regularization
retains spectral accuracy and satisfies the entropy condition, leading to the occurrence
of the correct weak solution. The application of this concept is particularly easy in a
global Fourier-based setting, where the modal solution representation can be modified
directly. Karamanos and Karniadakis [109] were the first to extend of this concept to
spectral/hp element methods, where they used polynomial filtering to establish the scale
separation. Depending on the filter kernel, the numerical application of the method is
relatively cheap compared to more elaborate modeling. SVV can be viewed as a cross
between monotone methods like TVD and an exponentially accurate, but unstable spec-
tral discretization [112]. It has been applied successfully to a number of turbulent flows,
from channels to bluff body flows [160]. One drawback of the methods remains the
choice of the filter kernel, which can significantly influence the solution [115].
2.4.3.7. Explicit Model: Scale Similarity
Scale similarity models were first introduced by Bardina et al. [10], and are based on
the assumption that the subgrid stresses close to the cut-off can be extrapolated from the
resolved scales near the cut-off. These “largest subgrid scales” can be determined from
the subgrid scales ui ′ = ui − ui by filtering as
ui ′ = ui − ui , (2.71)
so that the subgrid stress tensor τij is then closed as
τij = ui uj − ui uj (2.72)
51
Since this modeling relies on the self-similarity assumption in the inertial subrange, it
is often not dissipative enough, so that is is generally combined with a more dissipative
model like the Smagorinsky model [1]. While the scale-similarity is often used in re-
search with very good results and outperforms the Dynamic Smagorinsky model when
coupled with a suitable eddy viscosity model, its relatively high computational cost (the
cut-off must lie in the inertial subrange) and the additional filtering make it less numeri-
cally efficient.
2.4.3.8. Explicit Model: Interscale Transfer

The interscale transfer model is based on the same principles as the the scale similarity
model, but remedies its shortcomings in terms of low overall dissipation [5]. It is based
on the idea of balancing the small scale production terms that lead to a build-up of energy
near the cut off wavenumber. Three wavenumber bands are identified by progressive
filtering of the velocity field u:
R1 : u1i = ubi ,
R2 : u2i = ui − ubi , (2.73)
R3 : u3i = ui − ui ,
where R3 denotes the unresolved scales in LES. Given these bands, the quadratic non-
linear term can be decomposed into its contributions
!m
∂ui uj ∂ X X pX q
= ui uj , (2.74)
∂xj ∂xj m p q
where the superscripts p and q represent the bands in wavenumber space and m denotes
the region affected by the non-linear interaction, not an exponent. Based on physical
considerations and DNS data, a model consisting of terms from Equation (2.74) is con-
structed, that balances the energy transfer to R2 (R3 itself is absent) and prevents a
buildup. Due to the scale separation idea, this model is a natural candidate for spectral
methods. It has been applied successfully in a Fourier pseudo-spectral solver applied to
channel flows.
2.4.3.9. Explicit Mode: WALE

The wall-adapted local eddy-viscosity model (WALE) by Nicoud and Ducros [147] is
an algebraic model similar to the Smagorinsky model with modifications such that the
SGS viscosity is strongly reduced in wall-bounded laminar flows and vanishes towards
52
the viscous sublayer in turbulent boundary layers. It is therefore capable of producing

the laminar to turbulent transition and furthermore recovers the correct wall-asymptotic
y 3 -variation of the SGS viscosity and needs no wall damping functions. In the WALE
model, the eddy viscosity is modeled as
d d
Sij Sij
µt = ρL2s 5/2 d d 5/4
, (2.75)
(S̄ij S̄ij ) + (Sij Sij )
with Ls given by
Ls = min(κ d, Cs ∆), (2.76)
d
where κ is the von Karman constant and d the wall distance. The tensor Sij is defined
as
d 1 2 2 1 2
Sij = (ḡij + ḡji ) − δkk ḡij , (2.77)
2 3
where
∂ ūi
ḡij = . (2.78)
∂xj
The constant Cw must be tuned for the considered example.
53
3. Numerics for Scale-Resolving Simulations
From the discussion in Chapters 1 and 2, it has become obvious that the numerical sim-
ulation of multiscale flows poses two inter-connected challenges for the discretization:
First, the large scale bandwidth requires an accurate resolution of as many waves as
possible, while second, the necessary modeling of the subgrid terms demands low (in
the case of explicit modeling) or controllable (in the case of implicit modeling) approx-
imation errors. Both requirements are closely connected, as accuracy over the resolved
wave range is a direct consequence of the approximation errors, and these errors vanish
for consistent discretization in well-resolved situations.
In this section, we will focus on aspects of the spatial discretization for DNS-like reso-
lutions, i.e. this implies the assumption that the underlying solution is smooth and that
all dynamically active scales are fully resolved. Due to the higher dimensionality of
space and the direct coupling of the smallest resolvable spatial structure to its temporal
scale, the resolution requirements in space are the main contributor to the computational
cost in terms of degrees of freedom. Also when following the common method-of-
lines approach, the resulting semi-discretized (in space) formulation is linear in time, so
none of the complexities and increased resolution requirements of the non-linear spatial
operator come into play. Thus, the spatial discretization is the determining factor for
scale-resolving numerics.
It should be noted that generally, each LES also contains regions and wave number
ranges that are well-resolved, so the items discussed below for a DNS situation also
transfer to the LES setting.
3.1. The Points per Wavelength Paradigm

As has been discussed in Section 1.1, the number of grid points nP P W required to accu-
rately resolve the smallest occurring scale of a problem has a dominant effect on required
number of DOF and thus on the computational cost. Therefore, minimizing nP P W is of
prime importance for scale-resolving schemes for all types of multiscale problems. In
general, when approximating the spatial operator of differential or the associated inte-
gral conservation laws of the form shown in Section 2.1.2, the main complexity (besides
the treatment of non-linearities) and source of error is the approximation of derivatives
(or the equivalent fluxes over control volume boundaries), which in turn is immediately
55
related to the approximation of the solution itself. Two distinct concepts exist that di-
rectly determine the magnitude of nP P W : A global representation of the solution and
the derivatives, or a local one. The notion of globality here refers to the way in which
spatial information is processed to evaluate the solution and its derivatives: a global
method involves all the degrees of freedom in the domain, e.g. through a global ansatz
function, while a local method is based on a subset of these degrees of freedom, where
these subsets are then coupled in some appropriate way.
Figure 3.1 depicts this concept by showing the information stencils for three different
families of discretization schemes. Without loss of generality, a regular grid is used to
highlight the differences. The first of the plots depicts a global (pseudo-spectral) method.
Here, the solution is approximated by an ansatz that is global itself, i.e. all the available
degrees of freedom contribute to the approximation at each node in the domain. Fur-
thermore, the solution representation and the discretization of the derivative are unique
within the domain, i.e. the approximation errors of the derivatives are isotropic with
respect to space. Note that a truncation error of the solution may still exist, its form is
dependent on the choice of determination of the expansion coefficients. The derivative is
then approximated as the derivative of the global ansatz. For derivatives of smooth func-
tions, the associated error in the L2 norm with NGrid denoting the number of spatial
degrees of freedom can be estimated as [164]
NGrid
1
kespectral k2 ∼ O (3.1)
NGrid
If the grid is refined by increasing NGrid , both the grid spacing and the order of the error
change, thus, the error decreases faster than any finite power of NGrid , leading to “infi-
nite order” or spectral convergence. Thus, the incorporation of all available information
leads to the most efficient approximation in terms of invested DOF.
The second subplot in Figure 3.1 shows a discretization based on the FD method. Here,
the solution and the derivatives are approximated by locally defined polynomials. The
size of the local stencil and the associated weights determine the approximation error as
a function of grid spacing as
p
1
keF D k2 ∼ O , (3.2)
NGrid
where the order p is solely dependent on the choice of the derivative approximation.
Thus, since p is fixed, increasing the number of degrees of freedom through NGrid still
results in lower error, but the convergence rate is now finite and given by p (so-called
geometric convergence). In other words, the behavior of the discretization itself (its ac-
curacy) is no longer governed by the choice of the global NGrid , but by the choice of
56
∂/∂x ∂/∂x ∂/∂x
∂/∂x ∂/∂x
Figure 3.1.: Global vs. local approximation of spatial derivatives. The regular grid is
denoted by circles, the boundary points by squares. Blue lines denote the
solution, red and yellow lines the approximate solution based on the stencil
indicated by the arrows. Vertical black lines indicate a subcell.
p. Note that while the differentiation operator itself is usually chosen as fixed (for inner
points), i.e. isotropic, the representation of the derivative is not unique within the do-
main.
The final subplot represents a subdomain- or element-based discretization type, e.g. con-
tinuous or discontinuous FE methods. The spatial domain is divided into non-overlapping
elements. Within each of these elements, a polynomial ansatz for the solution and the
derivatives is constructed. Depending on the details of the method, these ansatz functions
can be low order functions like the hat function, often called “classical” FE, or higher
order polynomials, often applied in a tensor product manner in higher dimensions, which
leads to the name “spectral FE”. Besides the choice of the basis, another distinguishing
feature is the choice of the coupling of the subdomains. In continuous FE, the coupling
occurs by enforced C 0 continuity at the subdomain interfaces, i.e. the solution represen-
tation is unique there. For discontinuous formulation, the degrees of freedom are purely
element-local, i.e. they are double-valued at the boundaries. The coupling is then usually
57
done by a penalty term added to the strong form of the equation , e.g. [62] or a numerical
flux function for the surface contributions resulting from a weak formulation e.g. [100].
Within each element, the solution representation is unique, but the discretization of the
differential operator (its associated stencil and weights) depends on the location within
the domain, i.e. it is generally a one-side biased approximation. For this type of dis-
cretization, the global order of accuracy is determined by the choice of the local basis
functions. When refining the mesh, h-convergence for a fixed polynomial degree p can
be shown. Increasing the local degree of the ansatz p is akin to a NGrid -refinement in
global spectral methods and thus leads to spectral convergence, if the coupling is done
in a consistent manner.
Of the three different discretization concepts depicted in Figure 3.1, the global approxi-
mation incorporates all the available information into the construction of the operators,
and thus introduces the lowest approximation error and the smallest number nP P W . The
FE and FD methods limit the size of the information stencil by introducing subdomains
or incorporating only certain DOF in the local approximation, which results in a higher
nP P W .
PP N
PP 1 2 3 4 5 6 7 8 9 10 cFD O6
δ P
0.01 9.61 7.60 6.53 5.91 5.49 5.20 4.98 4.80 4.67 4.55 4.22
0.001 15.98 10.86 8.65 7.48 6.75 6.24 5.88 5.60 5.37 5.19 5.76
0.0001 25.62 15.25 11.35 9.38 8.22 7.43 6.87 6.45 6.13 5.86 7.93
0.00001 41.63 21.26 14.80 11.68 9.89 8.76 7.96 7.37 6.94 6.57 10.92
Table 3.1.: nP P Wmin (N, δ) for the Gauss DGSEM and the 6th order compact FD for
a given dispersion error δ and the degree N of the local polynomial ansatz.
Reproduced with permission from [85].
Figure 3.2 and Table 3.1 examine the dispersion error relations for the scalar one-dimensional
transport equation
ut + a ux = 0,
(3.3)
u(x, t) = ei(k x−ω t) ,
with a constant transport velocity a and ω = k a. The equation is discretized by a

Discontinuous Galerkin Spectral Element Collocation Method (DGSEM) as a represen-
tative of the FE-type methods, a central finite difference scheme and a pseudo-spectral
method (as a representative of global methods). For the global method, the dispersion
relation remains exact, i.e. no phase error is introduced up to the Nyquist frequency
58
5.5
DG O2
5 DG O6
DG O10
4.5 FD O2
FD O6
4 FD O10
Exact
3.5
Re(Ω*)
2.5
1.5
0.5
resolved represented
0
0 0.5 1 1.5 2 2.5 3
K*
Figure 3.2.: Dispersion relation for the linear transport Equation (3.3) of DGSEM with
Gauss nodes, central finite differences and global spectral (exact) scheme.
K ∗ = NK+1 is the normalized wavenumber and Ω∗ = NΩ+1 is the cor-
responding modified normalized numerical mode (eigenvalue). For FD,
K ∗ = K and Ω∗ = Ω.
of the grid [76]. This would correspond to a requirement of nP P Wmin = 2, regard-

less of the number of grid points (or expansion coefficients N + 1) in Table 3.1. For
the non-global FD and DGSEM methods, nP P Wmin (for a given dispersion error δ)
is clearly a function of the underlying degree N of the local approximation. For both
methods, the deviation from the exact dispersion relation moves to higher frequencies
and thus results in a reduced numerical cost in terms of nP P W . Note that the overshoot
for DGSEM in the marginally resolved region is characteristic of one-sided approxima-
tions, due to the element-local approximation of the differentiation operator which leads
to a bias towards the cell boundaries. Both Figure 3.2 and Table 3.1 thus support the
conclusion that higher order approximations which incorporate more information be-
come more “global” in their approximation character and thus lead to a more efficient
approximation for smooth problems in terms of nP P W . This concept of efficiency of
the approximation cannot only be expressed in terms of points per wavelength required
to resolve the smallest scale, but – in a measure more relevant to practical applications
59
– in the effective usage of the available degrees of freedom. When planning a numerical
simulation, the most prominent limiting factor in terms of memory, wall clock time and
pre- and post-processing capabilities is the total number of degrees of freedom of the
simulation. With each choice of this number comes a theoretically distinguishable range
of scales, determined by the associated Nyquist number. Depending on the discretiza-
tion, only a fraction of these represented scales can be resolved, which – for a non-global
method – is only a function of the local approximation order (polynomial degree, stencil
size). In Figure 3.3, R defines the ratio of the scales recovered up to a chosen dispersion
error δ = ei by a central FD discretization of arbitrary order p to the theoretical limit
(achievable with global spectral methods). The dispersion relations for these discretiza-
1
0.7 e1
e2
0.9
e3
0.6
0.8
0.7
0.5
0.6
0.4
R
1−R
0.5
0.3 0.4
p=2
0.3 p=4
0.2
p=6
0.2 p=8
p=10
0.1 p=12
0.1
p=64
0 0
10 20 30 40 50 60 500 1000 1500 2000 2500 3000 3500 4000
Number of Gridpoints
p
Figure 3.3.: Left: Ratio R of resolved scales (limit determined by phase error ei ) to
available scales on a NGrid = 1024 grid, ei=1,2,3 ∈ {0.1, 0.01, 0.001}, p:
order of central FD discretization
Right: Percentage of modes lost to dispersion error 1 − R as a function of
p for increasing number of grid points.
tions are given analytically, for details see Appendix D. The left plot of Figure 3.3 depicts
the ratio R on a grid with NGrid = 1024 nodes for three dispersion errors. Naturally, for
the largest dispersion error, the ratio of recoverable modes is highest. Increasing p for a
fixed NGrid here corresponds to extending the associated stencil, so a larger p (although
large stencils beyond p ≥ 10 would be impractical) corresponds to a higher ratio of
recovered modes. Note that in the region p ≤ 10, the large slope of R indicates that here
the gain from using higher order polynomials is most pronounced. The right subplot of
Figure 3.3 shows the percentage of lost modes 1 − R (lost to the dispersion error, i.e. not
contributing to the resolution) for a fixed error, but a changing grid size NGrid , again
as a function of p. While for all p, the amount of lost modes increases linearly with
60
NGrid (for large NGrid ), the general level differs greatly with p. Most importantly, the
fraction of resolvable modes only increases through p-refinement, while it deteriorates
through h-refinement. Thus, in terms of practical applications where the total number
of degrees of freedom is the determining factor, the choice of the discretization directly
corresponds to the ratio of recoverable information from these DOF, i.e. the recovered
scales per invested DOF dramatically increase for higher order approximations. Note
that for a global spectral method with perfect dispersion behavior, 1 − R would be 0,
regardless of NGrid .
While the discussion in the last paragraph is not directly transferable to other discretiza-
tion types, it exemplifies the superiority of high order approximations for the resolution
of multiscale problems, both in terms of accuracy and computational efficiency. It should
be noted that the term computational efficiency has been used rather loosely here to re-
fer to the number of required degrees of freedom to resolve a wave length, but not to
the cost associated with the numerical algorithm for providing the approximation. In
fact, the cost of evaluation a high order discretization is significantly larger in terms of
operation count as that of a low order discretization, mainly due to the fact that more
operations on the contributing DOF are necessary. These operations however typically
scale as a fixed power of N . For example, a matrix-matrix multiplication of two square
matrices of size N × N requires N multiplications and N − 1 additions for each of the
N 2 entries, leading to N 2 (2N − 1) total operations. From Equation (3.1), the error of
N
a global method scales spectrally as N1 , so the gain in accuracy can outweigh the
operation cost. Clearly, operation count is by far not the only factor determining the nu-
merical efficiency of a practical implementation of any method, but it gives an indication
that global or high order methods can be efficient, even in terms of wall-clock times.
Figure 3.4 supports this claim by showing the temporal evolution of the dissipation rate
of the Taylor-Green vortex flow (Section 2.2) for different choices of approximation or-
der N for a DGSEM scheme, compared to a DNS solution. Details of DGSEM will be
discussed later in Section 3.3.2, the important aspect here is that they allow an arbitrary
combination of number of elements and local polynomial ansatz degree N , so that the
total number of DOF can be kept constant while changing the underlying discretization.
The required wall clock time is indicated. All computations were run with the same
DOF load per core to minimize the influence of parallelization. The O(2) scheme with
643 DOF (N = 1, 32 Cells) has a very low wall clock time, but is overly dissipative and
misrepresents the physics. In other words, the very high nP P W requirement (or the very
small resolved wavenumber band) of this scheme is too limiting to capture the essential
physics for the given total number of DOF. On the other hand, the O(16) scheme with
643 DOF (N = 15, 4 Cells) requires significantly more CPU time for the same number
61
0.015
1.6h
DNS Brachet
N=15, 4 Cells (64³ DOF)
N=1, 32 Cells (64³ DOF)
0.0125 N=1, 64 Cells (128³ DOF)
N=1, 128 Cells (256³ DOF)
0.01
Dissipation Rate
32.0h
0.0075
79.2h
568.8h
0.005
0.0025
0
0 2 4 6 8 10
Time
Figure 3.4.: Dissipation rate − dk

dt
of the kinetic energy of the Taylor-Green Vortex at
Re = 800. Results for the DNS data taken from Brachet [29]. The required
wall-clock time for each computation is indicated in [h].
of DOF (also suffering from the stringent time step restriction), but resolves most of
the relevant scales and is in very good quantitative agreement with the DNS result with
3843 DOF. Increasing the number of elements for the O(2) scheme results in conver-
gence towards the DNS result, but the associated increase in wall clock time to achieve
comparable accuracy as the O(16) scheme makes this approach highly ineffective, ren-
dering it about 7 times slower than the high order computation. Note that the required
number of degrees of freedom for this result increases from 643 to 2563 , which corre-
sponds to a factor of 4 in one dimension. This fits well with the nP P W criteria for low
and high order DGSEM listed in Table 3.1, assuming an asymptotic extension to higher
N . In a companion plot, figure 3.5 compares the spectra of kinetic energy for the TGV
at t = 9s. Again, the total amount of DOF is limited to 643 . A DNS solution serves
as a reference. Although each computation has nominally the same spatial resolution,
the computed spectra show a clear convergence towards the DNS result for increasing
N . This is in accordance with the lower PPW requirement of high order discretizations,
thereby demonstrating the better scale resolving capabilities per DOF for high order
methods.
In this section, we have demonstrated the importance of an efficient numerical discretiza-
62
3.2. Global Spectral Methods
0.06
0.05 DNS
0.04 N=15, 4³ Cells (Int Pts=32)
N=7, 8³ Cells (Int Pts=16)
0.03
N=3, 16³ Cells
N=1, 32³ Cells
0.02
0.01
E(k)
5 10 15
k
Figure 3.5.: Taylor-Green Vortex at Re = 1600. Comparison of the spectra of kinetic

energy for N = 1, N = 3, de-aliased N = 7 and de-aliased N = 15
solution. All discretizations with a total of 643 DOF, except DNS.
tion, expressed in terms of points per wavelength nP P W , for the resolution of multiscale
problems. Due to their spectral error behavior, global methods are superior in this as-
pect, but due to their limitations not always flexible enough for practical applications.
FE-type methods or spectral element-type methods with a local high order solution rep-
resentation are thus the best alternative for efficient scale-resolving numerics.
In the next two Sections 3.2 and 3.3, we will discuss these two approaches in detail.
Due to their spectral accuracy as outlined above, approximations to partial differential

equations based on a global methods can be highly attractive for specific equations and
problems, in particular smooth periodic problems that support the concept of global basis
functions. These approaches all share the property that they approximate the solution by
a linear combination of orthogonal trial functions which form a complete basis of global,
63
smooth functions:
N
X
u(x) ≈ uh (x) = ûk φk (x), x 0 ≤ x ≤ x1 , (3.4)
k=0
with uh being the approximation to the function u in terms of the expansion coefficients
ûk and the trial or basis functions φk (x), defined on the whole domain.
The choice of this ansatz is determined by the nature of the problem and the boundedness
of the domain; for periodic problems, Fourier trigonometric polynomials are the most
popular. For non-periodic but bounded problems, Chebyshev or Legendre polynomials
are often the method of choice, while rational Chebyshev functions have also been used
[164]. The choice of the basis is reflected in the first part of the name commonly given
to the method, e.g. “Chebyshev-Galerkin Method”.
The second part of the denomination is determined by the choice of the test function or
the method to determine the expansion coefficients ûk in Equation (3.4):
• Galerkin Methods:
Galerkin methods are the most common representative of a projection approach,
where the residual R(x) = u−uh projected onto a set of test functions ψk chosen
identical to the trial functions φk is forced to vanish, i.e. the first (N + 1) terms
of the spectral representation of R(x) are zero:
Z x1 N
!
X
(R, ψi ) = u− ûk φk (x) ψi dx = 0, ψi = φi , i = 0, ..., N.
x0 k=0
(3.5)
One interpretation of this approach is the “error distribution principle”, which
denotes the fact that the first non-zero contribution to the residual error can be
driven to arbitrary high modes by increasing N [164]. Since the Fourier and
Chebyshev coefficients decrease exponentially fast (and the residual is assumed to
be smooth), the error over the whole domain rapidly decreases with N . The more
common denomination for Equation (3.5) is the variational or weak formulation,
as the residuum is to be minimized in an integral (or weak) sense.
By choosing the same number of basis functions (determining the number of
expansion coefficients) and test functions (determining the number of available
equations), the expansion coefficients ûk can be determined from Equation (3.5)
by exploiting the orthogonality properties of the basis as
Z
1 x1
Continuous Fourier Coefficients: ûk = u ψk dx, a ∈ R, k = 0, ..., N.
a x0
(3.6)
64
The earliest application of spectral Galerkin methods to the solution of PDEs

was reported by Silberman in 1954 [184]. While their accuracy and stability
made them attractive, the extensive cost associated with the evaluation of non-
linear terms reduced their efficiency compared to FD methods [104]. Using again
the inviscid Burgers’ equation (Equation (2.1)) with periodic boundary condi-
tions in a domain x ∈ [−π, π] as an example, a set of FourierPpolynomials
{φk (x) = eIkx , k ∈ [0, ..., N ]} as basis and the ansatz u(x) = N k=0 uk e
Ikx
with associated continuous orthogonality property

Z π Z π
φk φ∗l dx = eIkx e−Ilx dx = 2πδkl , (3.7)
−π −π
inserting the ansatz term by term into the inviscid Burgers’ equation yields:
N
∂u ∂ X
= ûk eIkx ,
∂t ∂t
k=0
N N
1 ∂u 2
1 ∂ X X
= ûk eIkx ûl eIlx
2 ∂x 2 ∂x
k=0 l=0
N N
1 XX (3.8)
= I(k + l)ûk ûl eI(k+l)x ,
2
k=0 l=0
N N N
∂ X 1 XX
⇒ ûk eIkx + I(k + l)ûk ûl eI(k+l)x = 0.
∂t 2
k=0 k=0 l=0
Projecting Equation (3.8) onto Fourier modes 0 ≤ m ≤ N and taking again

advantage of the orthogonality property (Equation (3.7)) results in an expression
for the temporal evolution of the modal expansion coefficients ûm
Z π N X
X N
∂ ûm 1
2π = I(k + l)ûk ûl eI(k+l)x e−Imx dx ⇔
∂t 2 −π k=0 l=0
N N
∂ ûm 1X X
= I(k + l)ûk ûl ⇔ (3.9)
∂t 2
k=0 l=m−k
N
∂ ûm 1 X
= I(k + l)ûk ûl .
∂t 2
l+k=m
65
P
Equation (3.9) shows that the non-linear term results in the double sum ûk ûl
(also called convolution sum), which introduces an operation count that scales as
O(N 2 ). For high N , this becomes prohibitively expensive, and puts the global
Galerkin methods at a clear disadvantage when compared to for example FD with
O(N ) operations for the non-linear term. This lead to the development of collo-
cation or pseudo-spectral methods, primarily by Orszag [152].
• Collocation Methods:
In collocation or pseudo-spectral methods, the test function ψ becomes the Dirac
delta function, i.e. the residual R(x) does not vanish in an integral, but in a
pointwise sense at the so-called N collocation points xj = 2πj N
. The associated
discrete orthogonality property derived from the continuous expression (Equa-
tion (3.7)) by numerical integration with the trapezoidal rule [27] becomes
N −1
(
T
X Ixj (k−l) N if k − l = mN, m = 0, ±1, ±2, ...
φk φl = e =
j=0
0 otherwise.
(3.10)
The general idea of the pseudo-spectral approach is to compute all non-linear

terms in a collocative manner in physical space, but all derivatives in spectral or
modal space. Returning to the inviscid Buger’s equation as an example again,
the coefficients of the interpolating Fourier polynomial of the solution u(x) are
determined by
N −1
1 X
Discrete Fourier coefficients: ũk = u(xj )e−Ikxj ,
aN j=0
(3.11)
N N
k = − , ..., − 1,
2 2
with the associated interpolating polynomial given by
N/2−1
X
Spectral Interpolant: IN (u(x)) = ũk eIkx , IN (u(xj )) = u(xj ).
k=−N/2
(3.12)
Here, IN indicates the interpolation and the tilde denotes that the coefficients are
determined by a numerical approximation of the integral in Equation (3.6), not
the exact integral. Depending on the form of the convective term of the Burgers’
2
equation (advection form u ∂u
∂x
vs. conservative form 12 ∂u
∂x
), the order in which
66
the non-linear terms are evaluated differs, but the basic principle remains the same.
For the conservative form, the spatial derivative of its spectral interpolant at the
collocation node xj becomes
N/2−1
∂u2 (xj ) ∂ X
= IN u(xj )2 = Ikṽk eIkxj ,
∂x ∂x
k=−N/2
(3.13)
N −1
1 X 2
ṽk = u (xj )e−Ikxj .
aN j=0
For the advection form of the non-linear term, this reduces to
∂u(xj ) ∂ ∂
u(xj ) = In u(xj ) u(xj ) = u(xj )IN u(xj )
∂x ∂x ∂x
N/2−1
X (3.14)
= u(xj ) Ikũk eIkxj
k=−N/2
Thus, for both forms, the differentiation is achieved in a spectral manner, but the
evaluation of the non-linear term is done in a collocative manner on the physical
grid. This exploitation of the duality of the physical and spectral representation
gave the method the name pseudospectral. The drawback of this last step is the
introduction of aliasing errors, first noted in [165] (but not labeled as such) and
remedied by filtering the upper half of the modes, which was refined by Orszag
to the 2/3-rule [153]. The origin of this error stems from the difference between
the convolution and following projection (Equation (3.9)) and the collocation ap-
proach for the non-linear term. Returning to Equation (3.11), the inverse discrete
transform to obtain the nodal values at location xj from the discrete Fourier coef-
ficients of two arbitrary functions b and c is given by
N/2−1 N/2−1
X X
b(xj ) = b̃m eImxj , c(xj ) = c̃n eInxj , j = 0, 1, ..., N − 1.
m=−N/2 n=−N/2
(3.15)
Collocation of the product of these two functions is given by
N/2−1 N/2−1
X X
d(xj ) = b(xj )c(xj ) = b̃m c̃n eI(m+n)xj , j = 0, 1, ..., N −1.
m=−N/2 n=−N/2
(3.16)
67
Note that this step already introduces the aliasing error, as the range of occurring
modes is doubled (through m + n), while the number of available expansion coef-
ficients remains N , as the spatial grid only supports N modes. This can be shown
by examining the modal content of c(x). The modal expansion coefficients c̃k are
again given by
N −1
1 X
d˜k = d(xj )e−Ikxj , k = −N/2, ..., N/2 − 1. (3.17)
aN j=0
Inserting Equation (3.16) into (3.17) and taking advantage of the discrete orthog-
onality property (Equation (3.10)) leads to
X X
d˜k = b̃m c̃n + b̃m c̃n (3.18)
m+n=k m+n=k±N
| {z } | {z }
Convolution sum Aliasing error
The resulting modes d˜k contain two contributions: the convolution sum, resulting
from the exact projection of the non-linear term (see Equation (3.9)) and an addi-
tional aliasing error.
The summations occurring in Equations (3.11) (the forward discrete fast Fourier
transform, FFT) and Equations (3.13) and (3.14) (the backward discrete Fourier
transform) can be computed in operations on the order of O(N log(N )) [54],
as opposed to O(N 2 ) for the Galerkin method or a direct differentiation via
a spectral differentiation matrix. For large N , this reduction in computational
cost becomes immense, and it is what makes the pseudo-spectral methods the
method of choice for the direct simulation of isotropic, periodic turbulent flows,
see e.g. [56, 66, 106–108, 123].
3.2.1. A Pseudo-spectral Code for the Compressible Navier-Stokes

Equations
As an example of global spectral methods, a Fourier based pseudo-spectral code for the
three-dimensional compressible Navier-Stokes equations in conservation form named
Spex was developed within this work. It represents the solution (i.e. the conservative
variables) at regular interpolation points in a cubic domain. All flux terms are evaluated
by collocation on the physical grid, and then transferred to wave space via a highly ef-
ficient implementation of the discrete Fourier transform [78]. All spatial derivatives are
then computed in wave space, and a backward transform bring them onto the physical
grid. The parallelization is achieved by a two-dimensional domain composition which
68
gives better parallel scaling on O(103 ) cores than the traditional one-dimensional de-
composition. Following the method of lines, once the spatial operator has been evaluated
at a time t, the resulting system of ordinary differential equations is advanced in time by
an explicit time step with a standard fourth order, five stage Runge-Kutta method [34].
De-aliasing is achieved by a cut-off filter in spectral space according to [152], that re-
moves the inaccurate modes in Equation (3.17). The code was validated by the method
of manufactured solution [174] and by comparison with a validated pseudo-spectral code
for the incompressible Navier-Stokes equations developed at the Université Catholique
de Louvain (UCL) [37]. Figure 3.6 shows the spatial convergence over grid spacing h,
measured against an exact solution enforced by an appropriate source term. The spectral
behavior of the spatial error is clearly visible at h ≈ 0.8. Decreasing h further results in
two separate regions with distinct scaling as ∼ h4 and ∼ h8 , respectively. These slopes
can be attributed to the fact the temporal approximation error dominates over the spatial
one once the region of spectral error decay is reached. As the advection time step scales
linearly with h and a fourth order temporal scheme was used, a slope of ∼ h4 develops.
Once the viscous time step restriction which imposes a scaling of ∆t ∼ h2 dominates,
the resulting error scales as ∼ h8 .
Figure 3.7 compares the results for the spectrum of kinetic energy for the Taylor-Green
-4
10
10-5
10-6 ~h
8
L2 Error
-7
10
-8
10
~h4
10-9
t~h
-10
10 2
t~h
0.4 0.6 0.8 1 1.2
h
18 16 14 12 10 8 6
Gridpoints N
Figure 3.6.: L2 error of density over grid spacing h for the pseudo-spectral code Spex,
obtained with the method of manufactured solutions.
69
vortex flow for Spex, the incompressible pseudo-spectral code developed at UCL and the
DGSEM framework Flexi (see Section 3.3.2). The agreement between the DG results
and Spex is nearly perfect, while a slight deviation exists with regards to the Louvain
code. This can be attributed to two reasons: i) The kinetic energy spectra provided by
the Louvain code has been smoothed in a preprocessing step, while the others have not,
ii) although the computations with the compressible codes (Spex and Flexi) were run at
a low Mach number of M a = 0.1, slight compressibility effects exist in the solution.
The Spex code and associated post-processing framework is mainly used to provide high
resolution reference results for comparison with DGSEM methods, see e.g. Figures 2.1
and 2.3.
Figure 3.7.: Spectrum of kinetic energy for the Taylor-Green vortex at Re = 1600
and t = 9s. Comparison of pseudo-spectral codes Spex and Louvain and
DGSEM code Flexi.
70
3.3. Spectral Element Methods

Spectral element methods are essentially Galerkin methods with a basis of high order
orthogonal polynomials of Legendre or Lagrange type, applied locally to a subset of the
whole computational domain. For this approach, the computational domain of interest is
first subdivided into non-overlapping elements or cells. For each of these elements, a so-
lution representation in terms of a locally defined basis with compact support is sought.
Patera [161] was the first to propose these methods as a combination of the advantages
of global spectral methods and classical FE methods: The accuracy in terms of a low
nP P W requirement and spectral convergence for smooth problems in combination with
the flexibility of an element based approach with non-periodic basis functions, allowing
a flexible domain discretization, non-periodic boundary conditions and efficient domain
decomposition strategies for parallelization. The information interchange between the
non-overlapping elements can be achieved by incorporating the coupling directly into
the local bases, i.e. by keeping the solution at the interfaces unique through C 0 continu-
ous basis functions (continuous spectral elements) or introducing double-valued degrees
of freedom at the interface (discontinuous spectral element or spectral penalty schemes).
Some appropriate penalization of the jump term, either for the weak or the strong for-
mulation of the variational problem, then enforces the coupling of neighboring cells,
e.g. [62, 100].
Due to the well-known stability problems continuous FE formulations share with other
centered approximation for the convection terms in FV and FD methods [92], a number
of remedies to recover stability exists, which generally include some form of parameter-
dependent upwinding and thereby introduce numerical diffusion. Common formulations
include the streamline-upwind Petrov Galerkin [32] and Galerkin Least-Squares [103]
approach, which differ in terms of the element-based weighting function. While these
methods have been applied with success to both incompressible and compressible prob-
lems [124,194], the stabilization introduces additional terms that decrease computational
efficiency and free non-unique parameters that require tuning [39].
As opposed to continuous formulations, discontinuous approximations show better sta-
bility for advection-dominated problems. This is due to the inherent upwinding in their
local stencil (see discussion in Section 3.1), which adds numerical diffusion for stability,
and to the jump occurring at the interfaces in under-resolved situations, which allows
a “weak decoupling” of the local approximations – the solution is allowed to be non-
unique, and coupling is only achieved weakly through a penalty term with associated
additional diffusion. This idea of introducing a non-unique solution that absorbs an
under-resolved approximation into a jump term is at the heart of the FV method, which
has superior stability properties for advection dominated problems and – depending on
the specific formulation – can guarantee stable solutions. Due to this inherent stability
71
for problems with strong gradients for continuous FE, we will focus on a discontinuous
discretization in the following section as an example of spectral element methods.
3.3.1. Discontinuous Galerkin Methods

Discontinuous Galerkin (DG) methods are a hybrid of high order FE methods and FV
methods, which gives them a number of favorable properties for scale-resolving simula-
tions:
• Spectral accuracy for smooth problems when increasing the degree of the local
ansatz (p-refinement), which results in low nP P W requirements
• The possibility to use arbitrarily shaped, unstructured grids to mesh non-trivial

geometries
• Local grid refinement in regions of interest (h-refinement)
• Stability for hyperbolic problems with discontinuities
• Local conservation for each element
• Weak imposition of the boundary conditions
• Efficient parallelization due to minimal coupling to neighbor cells
• Orthogonal hierarchical bases which resolve a wave range within an element,

which can be exploited in multiscale modeling
DG methods have a relatively recent history. They were introduced by Reed and Hill
[171] in 1973 for linear advection problems of neutron transport on triangular meshes
(the name “Discontinuous Galerkin” is not mentioned in their original paper) and ana-
lyzed by Lesaint and Raviart in 1974 [126]. Applications of the methods and widespread
use lay dormant for another two decades until Cockburn and Shu extended this approach
in a series of publications to non-linear conservation laws [45–48] such as e.g. the
compressible gas dynamics. Bassi and Rebay were the first to introduce a mixed fi-
nite element type approach for the discontinuous Galerkin discretization of viscous flow
problems [12] and extended the DG method to the compressible Navier-Stokes equa-
tions [14]. Importantly, they also remarked on the need for consistent high-order bound-
ary discretization to achieve high order accuracy in curved domains [13]. These efforts
were accompanied by a number of researchers focusing on alternative approximations
for the viscous fluxes [8, 9, 86, 131, 162], shock capturing [11, 38, 114], spatial and tem-
poral adaptation options [87, 113] and extension to problems beyond gas dynamics, e.g.
72
solid mechanics [129], magneto-hydrodynamics [201] and biological systems [206]. It

is worth noting that even before Reed and Hill in 1973, Nitsche [148] proposed a non-
conforming finite element method for elliptic problems which is nowadays referred to
as the symmetric interior penalty DG method [7], also available for the compressible
Navier-Stokes equations [71, 95].
As discussed in Section 3.2, the variational formulation of the problem is the basis for
the spectral approximation in a (sub)domain. Beyond the type of subdomains in multi-
dimensions, various choices for the specifics of the DG method exist that each lead to
a different formulation. Among these (spatial) choices are the basis functions with as-
sociated quadratures (e.g. Lagrange or Legendre-type polynomials), the approximation
space spanned by these functions in multi-dimensions (a tensor-product approach or a
full order basis), the weak or strong DG-formulation, the discretization choices for the
inviscid and viscous surface fluxes and the treatment of non-linearities. The temporal
integration introduces another level of possible choices.
Among these different variants, the Discontinuous Galerkin Spectral Element Colloca-
tion Method [120] combined with an explicit time integration scheme has shown to be
highly effective and competitive for scale-resolving simulations (see also Section 3.4).
Details of this method will be discussed in the following section.
3.3.2. Discontinuous Galerkin Spectral Element Collocation Method

In this section, we derive details of the Discontinuous Galerkin Spectral Element Col-
location Method for a system of hyperbolic-parabolic conservation equations, following
Kopriva [120] and Hindenlang et al. [100]. We use the compressible Navier-Stokes
equations in physical space R3 (Equation (2.16)) as an example
∂U ∂ c ∂ c ∂ c
+ F (U ) + G (U ) + H (U )
∂t ∂x ∂y ∂z
(3.19)
∂ v ∂ v ∂ v
− F (U, ∇x U ) − G (U, ∇x U ) − H (U, ∇x U ) = 0
∂x ∂y ∂z
with the vector of conserved variables U , the associated inviscid and viscous physical
fluxes {F c , Gc , H c } and {F v , Gv , H v } and the divergence operator defined in physical
space as ∇x = (∂x ∂y ∂z )T . Introducing the flux vectors F ~ c = (F c , Gc , H c )T and
F~ v = (F v , Gv , H v )T and their combination F ~ =F ~c − F~ v , a compact way of writing
Equation (3.19) becomes
∂U
+ ∇x · F~c (U ) − ∇x · F~v (U, ∇x U ) = 0,
∂t (3.20)
∂U ~ (U, ∇x U ) = 0.
+ ∇x · F
∂t
73
Together with suitable initial and boundary conditions, Equation (3.20) describes a sys-
tem of conservation equation of hyperbolic-parabolic type, that can be now be discretized
by the DGSEM method.
3.3.2.1. Spatial Discretization

In order to solve this system of equations, a discretization of the computational domain
consisting of non-overlapping elements is defined. In the DGSEM method, the type of
elements is restricted to hexahedral cells which support a tensor product basis. The ele-
ments can be connected in a fully unstructured way. Conformity of the element faces is
not required by the method itself, but is assumed for the following derivation.
Each element in the physical domain is mapped to a unit reference element E ∈ [−1, 1]3
~ from reference to
with coordinates (ξ 1 , ξ 2 , ξ 3 )T . The associated mapping function ~x(ξ)
physical space is approximated as a polynomial itself and is then used to calculate the Ja-
~ = det( ∂~x ). Clearly, for the mapping to be defined and invertible, J(ξ)
cobian J(ξ) ~ has
∂ξ~
to be positive everywhere, which can be challenging for non-linear mappings of curved
elements [99].
The resulting element-based mapping is then used to transform Equation (3.20) to refer-
ence space
1 ~ (U, ∇x U ) = Ut + 1 ~ ) − H(U,
~
Ut + ∇ξ · F ∇ξ · (G(U ∇x U )) = 0, (3.21)
~
J(ξ) ~
J(ξ)
~ := ~a1 · (~a2 ×~a3 ) is again the Jacobian of the mapping ~x(ξ),
where J(ξ) ~ calculated from
∂~
x
the covariant basis vectors ~al := ∂ξl . The covariant transformed fluxes are given by
~,
F l := J~al · F l = 1, 2, 3, (3.22)
with the metric terms
J~al := ~ak × ~am (l, k, m) cyclic. (3.23)
The way the metric terms are discretized and implemented is important for the prop-
erties of the resulting method. We refer to Kopriva [119] for a discussion on how this
choice ensures the so-called free-stream preserving property. In Equation (3.21), the
divergence operator in reference space is defined as ∇ξ = (∂ξ1 ∂ξ2 ∂ξ3 )T , and the
~ ) and H(U,
terms G(U ~ ∇x U ) are the Euler and viscous contributions to the transformed
~
fluxes F (U, ∇x U ).
Since the equation for each element is now defined in a common reference frame, we
74
can choose a polynomial solution approximation in the same space, which facilitates the
evaluation of many element-local operators in a pre-processing step. In DGSEM, the so-
lution vector within each element is approximated by a tensor product of 1-D Lagrange
polynomials ℓN of degree N
N
X
~ t) ≈
U (ξ, N ~
Ûijk (t)ψijk (ξ) , N ~
ψijk (ξ) = ℓN 1 N 2 N 3
i (ξ )ℓj (ξ )ℓk (ξ ) , (3.24)
i,j,k=0
where Ûijk (t) are time dependent nodal degrees of freedom and ℓN i (ξ) denotes the stan-
dard Lagrange polynomial of degree N defined by a nodal set {ξi }N i=0 ⊂ [−1; 1]. Fol-
lowing Kopriva [120], the N + 1 Gauss-Legendre quadrature points {ξi }N i=0 are cho-
sen as interpolation nodes. Another possible choice would be Gauss-Lobatto-Legendre
points, leading to a slightly less efficient and accurate scheme due to inexact integration
of the mass matrix [118]. Note that due to the Lagrange property of the basis functions,
we have
ℓN
j (ξi ) = δij , i, j = 0, ..., N, (3.25)
where δij denotes the Kronecker delta function as above. This property of the basis func-
tions, together with the double function of the interpolation nodes as integration points,
makes the DGSEM variant the most efficient DG implementation.
~ reads as
Analogously to the solution vector U , the discrete transformed flux F
M
X
~ ≈
F l (ξ) l
F̂ijk M ~
ψijk (ξ), l = 1, 2, 3 (3.26)
i,j,k=0
l
F̂ijk ~ x U ) |~
= G l (U ) − Hl (U, ∇ (3.27)
ξijk
M ~
with ψijk (ξ) = ℓM 1 M 2 M 3
i (ξ )ℓj (ξ )ℓk (ξ ). Note that the nodal fluxes are evaluated on
M + 1 Gauss-Legendre quadrature points, with M ≥ N . This implementation allows
for polynomial de-aliasing of the non-linear flux terms [111]. The choice of M depends
on the non-linearity of the flux for under-resolved calculations, more information on this
topic is given in Section 4.2. For the classical DGSEM, M = N is chosen, which leads
to a collocation of solution and fluxes on the same nodes.
Now that the domain discretization and the solution and flux approximations are in
place, we can derive the variational formulation of the problem and from it the DGSEM
~ (taken from the
scheme. We start by multiplying Equation (3.21) by a test function φ(ξ)
75
same space as the basis functions) and integrating over the reference element E to arrive
at the variational formulation in reference space
Z
JUt + ∇ξ · F ~ dξ~ = 0.
~ (U, ∇x U ) φ(ξ) (3.28)
E
Using a spatial integration by parts to remove the differentiability requirement from the
flux term and noting that the solution is discontinuous across element interfaces yields
the weak formulation
Z I Z
JUt φ dξ~ + (Gn∗ − Hn∗ ~
) φ ds − F(U, ∇x U ) · ∇ξ φ dξ~ = 0, (3.29)
| {z }
E ∂E ∗
Fn E
where Gn∗ denotes the surface normal numerical flux function for the inviscid terms,
given by Gn∗ := Gn∗ (U + , U − ) and superscripts ± denote the values at the grid cell inter-
face from the neighbor and the local grid cell, respectively. For the inviscid numerical
flux, several well-known flux functions derived for FV formulations are possible, which
ensure consistency and uniqueness of the numerical flux. Within the DG community, the
most commonly applied flux functions are Godunov’s method, the local Lax-Friedrichs
or Rusanov flux and Roe’s approximate Riemann solver. For details on those Riemann
∗
solvers we refer to the textbook by Toro [196] and to Section 4.5. The choice of Hn will
be postponed to Section 3.3.2.3.
3.3.2.2. Strong Form

Equation (3.29) is the so-called weak DG formulation, since the original conservation
equation appears in a weak sense, as the flux divergence has been shifted to the test
function in the volume integral. To recover the strong form, this volume contribution is
integrated in parts again
Z Z I
− F ~ (U, ∇x U ) · ∇ξ φ dξ~ = ∇ξ · F ~ (U, ∇x U )φ dξ~ − Fni φ ds, (3.30)
E E ∂E
where Fni denotes the normal component of the flux F ~ evaluated at the inner element
boundary, i.e. based on element-local information only to achieve analytical consistency.
Re-inserting Equation (3.30) into the weak formulation (Equation (3.29)) and collecting
terms yields the strong form
Z I
JUt + ∇ξ · F ~ (U, ∇x U ) φ dξ~ + Fn∗ − Fni φ ds = 0. (3.31)
E ∂E
76
The first integral in Equation (3.31) is just the Galerkin projection of the original equa-
tion, and the second term can be interpreted as a penalty term, that relaxes the orthog-
onality constraint on the projection of the residual. In this sense, the DG formulation
in strong form can be interpreted as a spectral penalty method, where the local approxi-
mations are coupled by an appropriate diffusion term that penalizes the jump at the cell
interface, i.e. in some sense the under-resolution of the problem [96]. Since for con-
tinuous FE, the solution is C 0 continuous at the interfaces, this penalty term is always
zero and the stabilization mechanism is missing, which overburdens the global ansatz
for advection dominated problems and leads to stability issues, as discussed in in the
introduction of Section 3.3. In this sense, the discontinuity of the ansatz for DGSEM
leads to a more local approximation, as the weak coupling to neighbors through penalty
terms can at least partially absorb de-stabilization effects of under-resolution.
It should be noted that while Equations (3.31) and (3.29) are analytically equivalent, dif-
ferences can arise through implementation choices. In this work, the weak form is used
exclusively.
3.3.2.3. Second Order Derivatives
∗
Returning to Equation (3.29), the last missing term to be defined is Hn . This term de-
notes the numerical flux function for the viscous term, resulting from the viscous flux
H~ in Equation (3.21). Through this viscous flux, Equation (3.29) depends on the solu-
tion gradient in physical space ∇x U . Thus, an approximation of the solution gradient
∗
is necessary to define Hn . The treatment of the gradient terms in the context of DG
approximations was first tackled by Bassi and Rebay [12, 15], who introduced a mixed
finite element approximation. They showed that a local evaluation of the gradient leads
to instabilities, and that some form of “lifted” gradient, containing information from both
adjacent elements, is needed.
To derive the mixed formulation, the system of governing equations is rewritten as a

~ as an approx-
corresponding system of first order equations with an auxiliary variable S
imation of the lifted gradients
~ − ∇x U = 0,
S
(3.32)
~ (U, S)
Ut + ∇ x · F ~ = 0.
77
Applying the discretization steps outlined above to the auxiliary equation leads to
Z I Z
c = 1, ..., 5 : ~
J~sc φ dξ + uc,n φ ds − uc · ∇ξ φ dξ~ = 0,
~ ∗
E ∂E E
Z I Z (3.33)
JUt φ dξ~ + (Gn∗ − ∗
Hn ) φ ds − ~
F(U, ~ · ∇ξ φ dξ~ = 0,
S)
E ∂E E
with the component uc of the state vector U and its lifting operator ~sc . The numerical
flux of the auxiliary equation is ~u∗c,n , and Hn
∗
= Hn∗ ~ +, S
(U + , U − , S ~ − ) denotes the
numerical flux function for the viscous terms. Following [12], we choose
−
c = 1, ..., 5 : u∗c,n = αvisc u+ c + (1 − αvisc ) uc ~ n, (3.34)

∗ + ~+ − ~−
Hn = αvisc Hn (U , S ) + (1 − αvisc ) Hn (U , S ) , (3.35)
with ~n denoting the outward pointing surface normal. For a parameter of αvisc = 12 ,
this treatment of the viscous fluxes is usually labeled BR1 (first method of Bassi and
Rebay [12]) and is known to be formally unstable for purely elliptic problems. How-
ever, for advection dominated problems, the inviscid numerical flux function is helping
in stabilizing the overall operator and no issues have been observed in any of our com-
putations.
3.3.2.4. Boundary Conditions

Due to the weak coupling of the elements through fluxes in DG methods, it is an obvious
idea to enforce the boundary conditions in the same manner, even for the Dirichlet type.
Instead of prescribing a state U at the boundary, an appropriate right hand side state U +
(akin to a ghost cell state) is used to compute the resulting boundary flux. The ratio-
nale for this approach is to ensure consistency in the approximation of the solution and
the boundary conditions, i.e. to use the same discretization operators for both and thus
avoid stability issues [17]. Collis [51] investigated the effect of weakly versus strongly
imposed Dirichlet conditions for the case of an under-resolved one-dimensional station-
ary boundary layer problem and turbulent channel flows. He found that the solution
quality measured in the L2 and H1 error norms is greatly enhanced when using a weak
application of the boundary conditions, while the L∞ norm slightly favors the strong
form, as this norm reacts directly to the error directly at the boundary, where an exact
fulfillment cannot be guaranteed in a weak formulation, but is enforced when using the
strong imposition.
78
We follow this approach, and we enforce the boundary conditions for Equation (3.33)
weakly through the prescription of the boundary fluxes ~u∗c,n , Gn∗ and Hn
∗
. While Collis
reported improvements using weakly enforced wall boundary conditions, we also have
noted positive effects at freestream and outflow boundaries: reflections stemming from
outgoing acoustic waves are rarely encountered, and are often negligible. Even in aeroa-
coustic computations, where an unpolluted instantaneous pressure field is highly impor-
tant, this treatment was found to be sufficient when using Roe’s approximate Riemann
solver which computes the surface flux according to the local flow characteristics [75].
3.3.2.5. The DGSEM Operator
Equation (3.33) describes the general analytical, semi-discretized form of the DG method
of a hyperbolic-parabolic system of equation with the second order derivatives treated
by the first method of Bassi & Rebay and an arbitrary Euler flux function. A full, very
detailed derivation of the DGSEM operator from Equation (3.33) is given by Hindenlang
et al. in [100], here, just some of the details are repeated to highlight the efficiency of
DGSEM.
As noted in Equations (3.24) and (3.26), the solution and the flux are represented by
tensor products of one-dimensional Lagrange interpolating polynomials, associated with
either one-dimensional Legendre-Gauss or Legendre-Gauss-Lobatto quadrature points.
The Lagrange property of the basis functions on these nodes makes the evaluation of the
basis at these points and thus of the solution and fluxes trivial, as only one nodal degree
of freedom gives a contribution. The evaluation of the inner products is then achieved by
the corresponding quadrature rule, which reverts to a sequence of three one-dimensional
sums along a reference coordinate line, which reduces the number of operations from
O(N + 1)6 for a standard DG formulation to O(N + 1)4 for DGSEM.
We demonstrate this concept by applying the DGSEM formulation to the first volume
integral, containing the time derivative of the degrees of freedom, from Equation (3.29).
First, we insert the ansatz for the solution (Equation (3.24)) into the semi-discrete form
and choose the test function φ from the space of Lagrange polynomials of degree N as
N
ψijk with associated N + 1 Legendre-Gauss nodes {ξi }N i=0
Z Z N
!
~ t φ dξ~ = ~ ∂ X N ~
J(ξ)U J(ξ) Ûrst (t)ψrst (ξ) N
ψijk dξ~ . (3.36)
∂t r,s,t=0
E E
79
The integral over the reference space is now split into the coordinate directions and then
replaced by Legendre-Gauss quadrature with associated weights ω :
Z Z1 Z1 Z1 N
!
∂ X
~ t φ dξ~ =
J(ξ)U ~
J(ξ) N ~
Ûrst (t)ψrst (ξ) N ~
ψijk (ξ)dξ~1 dξ~2 dξ~3
∂t r,s,t=0
E −1 −1 −1
 
N
X N
X
∂  N
= J(ξ~αβγ ) 
 ∂t Ûrst (t) ℓN 1 N 2 N 3  ~
r (ξα ) ℓs (ξβ ) ℓt (ξγ ) ψijk (ξαβγ )ωα ωβ ωγ
α,β,γ=0 r,s,t=0
| {z } | {z } | {z }
=δrα =δsβ =δtγ
N
X ∂
= J(ξ~αβγ ) Ûαβγ (t) ℓN 1 N 2 N 3
i (ξα ) ℓj (ξβ ) ℓk (ξγ ) ωα ωβ ωγ
∂t | {z } | {z } | {z }
α,β,γ=0
=δiα =δjβ =δkγ
∂
= J(ξ~ijk )ωi ωj ωk Ûijk .
| {z } ∂t
pre−compute
(3.37)
Note that the Kronecker delta functions indicate the occurrence of the Lagrange prop-
erty and thus reduce the associated summation to a single evaluation. Also, the Jacobian
of the mapping is treated in a collocation way in this approach, i.e. it is not integrated
exactly if the mapping is beyond bi-linear. In this case, an additional error akin to mass-
lumping for Gauss-Lobatto integration of the mass matrix is introduced.
From an efficiency point of view, Equation (3.37) demonstrates how the three-dimensional
integrals reduce to point-wise evaluations in DGSEM. For each of the (N + 1)3 DOF
Ûijk per element, just a single multiplication with a pre-computed term is necessary, due
to the “folding” of the three-dimensional integral based on the tensor-product structure
and the collocation of interpolation and integration nodes.
For the surface and flux volume integrals in Equation (3.29), a similar reduction in op-
erations can be shown, where the volume integral retains an operation count of (N + 1)
multiplications per DOF, as the occurrence of the derivatives of the basis functions does
not support the Lagrange property. Further details and a full discretization of the operator
can be found in [100]. Following the discussion therein, pre-computed one-dimensional
operators can be defined. For the volume integral, the differentiation matrix and the
80
weighted differentiation operator read
dℓj (ξ)
Dij = ,
dξ
ξ=ξi (3.38)
ωi
D̂ij = − Dji , i, j = 0, ..., N.
ωj
The weighted basis functions are given accordingly by
ℓi
ℓ̂i = , i, j = 0, ..., N, (3.39)
ωj
and ŝ is the surface element, relating the physical to the reference surface. With these
definitions, the semi-discrete form of the DGSEM operator for the weak form becomes
!
N
X
1 +ξ1 1
−Jijk Ûijk = D̂iα F̂αjk + [F ∗ ŝ]jk ℓ̂i (+1) + [F ∗ ŝ]−ξ
jk ℓ̂i (−1) +
t
α=0
 
N
X
2  +ξ2 2
 D̂jβ F̂iβk + [F ∗ ŝ]ik ℓ̂j (+1) + [F ∗ ŝ]−ξ
ik ℓ̂j (−1) +
β=0
!
N
X
3 +ξ3 3
D̂kγ F̂ijγ + [F ∗ ŝ]ij ℓ̂k (+1) + [F ∗ ŝ]−ξ
ij ℓ̂k (−1) .
γ=0
(3.40)
Thus, for DGSEM, the three-dimensional operator essentially collapses to a sequence of
three consecutive one-dimensional operators, which is responsible for the dramatically
decrease in operation count for DGSEM. The volume integral reduces to an evaluation
along a line (effort(N + 1)), and all other operations are point-wise. Thus, especially
for high N , the volume integral is the determining factor in terms of operation count.
3.3.2.6. Temporal Discretization

Both Equations (3.40) and (3.29) describe how to discretize the spatial operator, leaving
the temporal dimension continuous. These semi-discrete forms can be advanced in time
by any suitable method. Since the DGSEM operator itself is very local in the sense
that it only communicates with the direct neighbors through the surface fluxes (and not
across corners) and the local operations are very dense, it is suitable for an explicit
time integration. Also, multiscale problems, particularly in fully resolved situations, are
governed by the temporal scale of their smallest structures, and an implicit treatment
81
with large time steps can lead to the introduction of a sizable approximation error. Thus,
we choose to treat the time dimension explicitly by a fourth order accurate, five stage
Runge-Kutta method of Carpenter and Kennedy [34]. The time step for advection for
DG is of the form
1 ∆x
∆tadv
max ∼ , (3.41)
λadv
max 2N + 1
with ∆x as the grid spacing and λadv max the maximum eigenvalue of the flux Jacobian
matrix for advection, and the viscous time step is given by
2
dif f 1 ∆x
∆tmax ∼ dif f , (3.42)
λmax 2N + 1
where λdif f
max the maximum eigenvalue of the diffusion matrix [50]. This strong depen-
dence on the polynomial degree N is due to the overshoot in the dispersion relation at
high wave numbers, and is one of the drawbacks of DG [40].
3.4. Efficiency of DGSEM for Scale-Resolving Simulations

As discussed in Chapter 1, efficient numerics are mandatory for scale resolving sim-
ulations. In the previous sections, efficiency has been defined in terms of points per
wavelength required by a given numerical discretization to accurately reproduce a scale
of a given size. In Section 3.1, we discussed the importance of spectral accuracy and
high order approximations with spectral character in terms of reducing the nP P W re-
quirement and taking advantage of the invested DOF. In Sections 3.2 and 3.3, global
spectral schemes and spectral element schemes, in particular the DGSEM variant, were
introduced. DG methods offer the potential for high order accuracy and spectral con-
vergence in smooth regions. From the discussion in Section 3.1 and an investigation of
the dissipation and dispersion relations for DGSEM in [85], the factor nP P W in Equa-
tion (1.5) for a global spectral method is 2, while it lies between 4 and 7 for a high order
DGSEM scheme. Lower order schemes such as second order FV schemes or classic FE
schemes typically require at least 15 to 20 PPW for a comparable accuracy [43, 128].
Thus, in terms of pure resolution requirement, high order DGSEM can be considered an
efficient method, since it combines a low nP P W requirement with geometric flexibility.
3.4.1. Operation Count

While this estimate in terms of DOF is useful to gauge the limits of each method, it is
not a suitable measure for overall code performance or wall clock time. In particular, the
82
operation count per degree of freedom for all spatial discretization schemes scales with
some power of the size of their local (or global) stencil, i.e. a global spectral method also
scales “globally” with the number of degrees of freedom, while a low order FD method
will scale with its local stencil size.
For example, for a one-dimensional linear first order transport equation discretized on a
grid with Ngrid nodes, the best scaling for a global method is given by a pseudo-spectral
approach, which reduces the operation count for one evaluation of the spatial operator
2
from ∼ Ngrid to ∼ Ngrid log(Ngrid ) (due to the FFT). For a standard first order FD
discretization, the stencil size is 2, i.e. 2 operations are required per grid point, leading to
an operation count of ∼ 2 Ngrid . Table 3.2 summarizes the estimates for the operation
count for different discretizations of a one-dimensional domain with Ngrid DOF and
approximation order p. Note that for DGSEM, the operation count per element is p + 1,
plus an additional evaluation of a surface flux. The number of elements in DGSEM is es-
timated as total DOF divided by DOF in a single element. This estimate does not take the
doubling of the interface nodes into account, but the general scaling remains the same.
With these simplifications, it can be seen that the DGSEM method scales similarly to a
FD method, and as discussed in Section 3.3.2.5, this extends to multi-dimensions due
to the dimension-by-dimension outline of DGSEM. Of particular importance is the fact
that when the total number of DOF remains fixed, the operation count is only weakly
dependent on p, which makes high order DGSEM with the nlowP P W particularly at-
tractive for scale-resolving simulations. Clearly, counting operations is a very crude
measure of performance, due to a number of reasons: a) Different operation types re-
quire different amount of clock cycles on a CPU, b) memory access and caching effects
can dominate the operation speed depending on the size of the operator, c) hardware
architectures differ in terms of execution speed for specific operations beyond multipli-
cation and d) parallelization efficiency can completely change the picture. Nonetheless,
the operation count indicates whether a discretization is theoretically able to be efficient
Discretization Operation Count
Pseudo-spectral ≈ Ngrid log(Ngrid )

FD O(p) ≈ Ngrid (p + 1)
DGSEM O(p) ≈ Nelems (p + 1) + Nelems , Nelems ≈ Ngrid /(p + 1)
Table 3.2.: Estimated operation count for one-dimensional spatial operators for different
discretizations.
83
when compared to others, thus, it poses some form of upper limit on performance on an
ideal computing system.
3.4.2. Parallel Performance

In the previous sections, we have shown that DGSEM has two properties that are fa-
vorable for efficient multiscale numerics: a) spectral accuracy for a high wave number
resolution per invested DOF and b) an operation count that is comparable to FD schemes,
the simplest and least costly discretization choice in terms of operations. However, an ef-
ficient parallelization and implementation of the code along with optimizations specific
to the hardware architecture is as important for large scale simulations as the scheme
properties themselves.
On massively parallel systems, any operation imbalance will cause a bottleneck. Thus,
load balancing (in terms of work per core, but also in terms of communication vs. com-
puting) is essential. The DGSEM operator has two favorable properties that support
efficient parallelization: a) The operator itself is very local, and elements communicate
only with their direct neighbors, i.e. the waiting time and the communication connec-
tions to other cores are minimal. b) The volume integral is a purely local operation, that
does not require neighbor information. It can be used to hide the communication latency
by creating temporal buffers through a non-blocking communication. The dimension-
by-dimension structure of this operator allows the design of buffer sizes that “fit” the
communication bandwidth. In addition, our MPI parallelization does not introduce ad-
ditional operations (besides the data transfer), for example the surface fluxes will only
be computed by one of the neighboring cores and then send to the partner (with this
operation also being equally distributed between the two processors sharing a face).
More details on parallelization, scaling, load balancing and domain decomposition for
DGSEM can be found in [4,99]. Two of the most important results concerning the paral-
lel performance of DGSEM are repeated here, as they pertain directly to the efficiency of
high order DGSEM for scale-resolving simulations: a) As shown in [99], the computing
time to update a single DOF (i.e. the evaluation of the spatial operator) remains nearly
constant over increasing polynomial degree N . Thus, for our DGSEM implementation,
a “high order” DOF is not substantially more expensive than a “low order” DOF. This
means that for the same number of degrees of freedom, the low and high order DGSEM
schemes will take similar CPU times (neglecting the time step restriction), with the high
order discretization offering much improved spatial resolution. b) Weak scaling is near
optimal [4], allowing simulations with many DOF (on the order of hundreds of millions),
while strong scaling (which allows manageable wall times for large scale problems) can
become super-linear due to caching effects. Figure 3.8 shows strong scaling results on
the BlueGene Q at JSC with a speed-up from 32, 678 to 262, 144 ranks (down to 8 ele-
84
ments with N = 7 per core) of 116%, as the decrease in core load allows the caching of
the local data and thus reduces memory bandwidth issues.
Figure 3.8.: Strong scaling of DGSEM on BlueGene Q “JuQueen” at JSC. 1283 ele-
ments, N = 7 up to 262, 144 ranks (131, 072 cores with 2× hyperthread-
ing). Results from the Porting and Scaling Workshop, 2013.
3.4.3. Comparison with Other Codes

At the “1st International Workshop on High Order Methods in CFD, 2012” a comparison
of various high order schemes from different contributors for a scale-resolving simula-
tion of the Taylor-Green Vortex problem at Re = 1600 (see Section 2.2) was conducted.
Details on the problem definition and on the errors measures for comparison are pub-
lished in [200]. The result from this workshop have been reproduced here with permis-
sion.
Table 3.3 lists the participants of the workshop and the computational methods. Note that
all computations were conducted with DNS schemes, i.e. without additional LES mod-
eling. The contributions included DGSEM, other DG methods (modal, nodal, Recovery-
DG, CPR-DG), FD and DRP-FD (dispersion-error minimized FD) and 3rd and 4th or-
85
Color in Fig. 3.9 Affiliation Authors Method
Red University of Stuttgart Beck DGSEM,RK5

Dark Blue Cenaero Carton, Hillewaert DG/IP,RK4
Green ONERA Chapelier et al. DG/BR2, RK4
Light Blue NASA Glenn Debonis FD,RK4
Light Blue NASA Glenn Debonis DRP,RK4
Purple Iowa State University Wang et al. DG–CPR, RK4
Black ONERA Le Gouez FV/Recon, RK3
Yellow University of Michigan Varadan et al. DG/RDG, RK4
Table 3.3.: Contributors and scheme details for the 1st International High Order Work-
shop, test case 3.5, see Figure 3.9.
der reconstructed FV formulations. The computational effort was judged by comparing

against a benchmark tool on each employed architecture, however, some doubts about
the inter-comparability remain. Figure 3.9 compares the error (computed against a refer-
ence solution with 5123 DOF of a pseudo-spectral solver) over invested DOF (top) and
over computational time in Work Units (WU) (bottom). Starting with the top plot, a few
general trends discussed in Section 3.1 can be observed here in practice: The order of the
discretization and the discretization choices themselves strongly influence the accuracy
and efficiency. For a given accuracy, the higher order formulations are located towards
the left of the plot, indicating a lower requirement in DOF, consistent with their lower
nP P W demands. Also, DG formulations can return comparable or better accuracy per
invested DOF than FD or FV schemes for this problem. In terms of minimizing the
error, high order DGSEM (O(8) and O(10)) outperforms the other implementations.
Note that the DGSEM O(16) results likely suffer from aliasing issues, which explains
their somewhat reduced accuracy. Thus, in terms of invested DOF, DGSEM with N ≥ 4
is very favorable.
The bottom plot in Figure 3.9 repeats the comparison, but in terms of computational
effort instead of invested DOF. This essentially scales the left plot by the Work Unit
per degree of freedom, i.e. by a measure of computational efficiency of the scheme. In
general, the DG formulations lose compared to the FD schemes for reasons discussed in
Section 3.3.2.5, and they move towards higher Work Units, requiring one to two orders
of magnitude more computational effort than FD. Only the DGSEM formulation is ca-
pable of competing with FD in terms of accuracy per cost, and it marks the lower limit
in terms of WU for all DG variants. Note that the DGSEM results are between a factor
86
of 1.1 and 1.2 more expensive than the FD computations, but this is made up by a con-
siderable gain in accuracy. Another interesting issue is the fact that the DGSEM results
are almost on a vertical line in this plot, indicating very little difference in WU for the
different order approximations, making high order discretizations as cheap as low order
ones. A number of factors contribute to this behavior: a) The overall number of DOF for
all DGSEM cases was constant, i.e. the number of elements was adjusted accordingly.
b) Just counting operations according to Table 3.2, the number of operations should be
similar in all cases, even in three dimensions. c) The excellent strong scaling perfor-
mance of the code, as all computations were performed on 4096 cores, i.e. a constant
load per core of 4096 DOF, which is close to the optimum region where performance
is near independent from polynomial degree. Note that changing the polynomial degree
will result in a shift between communication and computation time, which can - in con-
junction with the grid spacing - make up for the time step penalty of high order DGSEM.
In summary, these results indicate that the DGSEM method constitutes the most effec-
tive DG implementation, and that it can offer comparable accuracy and computational
efficiency to high order FD methods, while retaining the advantage of fully unstructured
meshing.
87
Figure 3.9.: Code comparison for DNS of Taylor-Green Vortex at Re = 1600.

Top: Error in computed vs. reference dissipation rate over DOF.
Bottom: Error in computed vs. reference dissipation rate over Work Units
(CPU time).
Results from the 1st International Workshop on High Order Methods
in CFD, 2012, with permission from Koen Hillewaert (Cenaero) and
from [200]. For legend refer to Table 3.3.
88
3.5. DNS with DG Methods

Despite the favorable approximation properties and computational efficiency discussed
in the previous sections, the number of scale-resolving simulations with DG methods
published in literature is still remarkably low. This can be attributed to a number of
reasons: a) The relative newness of the Discontinuous Galerkin method and its exten-
sion to system of evolution equations compared to more established methods that have
been adopted for industrial applications and research, b) the focus on incompressible
turbulence in basic research, where continuous spectral element methods are the estab-
lished codes, c) the perceived increased complexity of DG compared to FD and d) the
sharp time step restriction. Another issue that is hindering the widespread use of high
order methods in basic and industrial research is the generation of high order meshes
for complex, curved geometries, which has up-to-date not been solved in a satisfactory
manner [99, 101].
Collis in 2002 was the first to use a high order DG method (p = 6) for the DNS of
a weakly compressible turbulent channel flows at a low Reynolds number, with about
13 mio DOF for his finest mesh [51]. More recently, Wei et al. [202] extended this
investigations from sub- to supersonic channel flow and moderate Reynolds number, us-
ing an over-integrated 10th order DG scheme with a total of 4.3 mio DOF and found
satisfactory agreement with previously published results. Carton de Wiart et al. investi-
gated fourth order DG methods for DNS of the Taylor-Green Vortex with 3843 DOF and
showed convergence towards a pseudo-spectral reference solution. They also reported a
DNS with 10 mio DOF of a transitional flow over an SD7003 airfoil [35]. Recently in
2014, Renac et al. evaluated DNS with a modal DG scheme for the Taylor-Green Vortex
and a channel flow, as well as a two-dimensional dipole interaction with a wall [41].
They used approximation order up to 8, and at most 56 mio DOF.
In the following, we will briefly outline some exemplary DNS computations with our
DGSEM framework. It should be noted that the simulation code itself is just one building
of the framework. Single block structured meshes with volume curving can be generated
from analytical functions by a stand-alone preprocessor, while for general, multi-block
structured or unstructured meshes, existing grid generators are used. Due to the necessity
of high order curvature information which is usually not supplied by the meshing soft-
ware, curved mappings have to be generated in a preprocessing step. Several strategies
exist, a detailed overview is given in [99]. In addition to mesh format transformation and
curving, the preprocessor strings all elements along a space-filling curve, which results
in a one-dimensional list structure, in which elements that are close in physical space
are also close in terms of list position. This approach allows efficient parallel read-in
and ad-hoc domain decomposition. Within the simulation code, the same sorting is used
to write the results simultaneously in a single file, as each process “knows” its location
89
within the complete list and can fill its slot without conflicts.
The post-processing of large scale data is achieved by a collection of specific tools, all
of which share a common code basis for general tasks and use the same data structure as
the code itself. Visualization of surface and volume data is achieved by super-sampling
the DG polynomials on equidistant grids, derived quantities like λ2 can be computed
within the polynomial space and then visualized as well. A standalone parallelized
three-dimensional FFT tool exists to compute volume spectra. On meshes with at least
one structured dimension, spatial averaging based on the high order accurate quadrature
can be performed. All post-processing tools write data readable standard visualization
software (Tecplot1 , Paraview2 ). A special output format and Python-scripted tool chain
connects the post-processing output to professional rendering software like blender3 for
high quality visualizations.
For the evaluation of turbulent statistics, a time-averaging of conserved and derived
quantities is implemented. The averaging of fluctuation products (like Reynolds stresses)
follows from
Z t+∆t
1
uv = uv dt
∆t t
Z t+∆t
1
= (u + u′ )(v + v ′ )dt
∆t t (3.43)
= u v + uv ′ + vu′ + u′ v ′
= u v + u′ v ′
→ u′ v ′ = uv − u v
where both relations are computed and stored during runtime and assembled in post-
processing. All time averaging is done in independent time slices which can later be used
to perform the averaging over user-specific time intervals. Additionally, record points or
probes can be specified a priori, which record the time series of prescribed quantities for
spectral analysis. They are located in the mesh in a pre-processing step and stored along
the space filling curve together with their pre-evaluated local basis functions for efficient
evaluation during the actual computation.
To facilitate comparisons between different computing architectures and to measure the
performance of our code, we compute the performance index PID for each computation
as
TCP U
PID := [s], (3.44)
nDOF × n∆t
1
www.tecplot.com
2
http://www.paraview.org/
3
www.blender.org
90
where n∆t denotes the number of explicit time steps and the total CPU time TCP U is
computed as the number of computing cores ncores times the simulation wall clock time
TW . The number of spatial degrees of freedom per state vector component nDOF is
given by the product of the number of grid cells ncells and the collocation points per
element (N + 1)3 . While the absolute value of the PID is strongly dependent on the
computing systems, its relative changes within a single system allow us to find the op-
timal configuration in terms of processor load for a given system. Where meaningful,
also informations about the wall time per characteristic time unit is given, as for tur-
bulent simulations with statistics gathering, the length of the averaging period is the
determining factor in overall computational cost. The relevant computational data for
all cases presented below is gathered in Table 3.4. Denoted are details about the spatial
discretization, the number of DOF per conservative variable, the computing cores and
a characteristic time step ∆t divided by a reference time unit T ∗ (which is 1 for both
vortex test cases). Also the wall time per characteristic time unit TW /T ∗ , the number of
time steps, the overall wall clock time and the computing architectures are listed.
PID no. time

Case N no. DOF no. cells cores ∆t/T ∗ Tw
[µs] Tw /T ∗ steps
Sphere 4 2.6 · 106 21, 128 4, 096 1.5·10−4 22.0 90s 1.9·106 7.5h
Vortex 7 57 · 106 110, 592 6, 912 3.8·10−3 11.4 - 2.1·105 1.9h
TGV 11 216 · 106 125, 000 125, 000 6.9·10−5 72.3 - 1.8·105 7h
Cylinder 7 164 · 106 319, 488 4, 992 4.7·10−4 17.1 1.9h 1.5·106 460h
Table 3.4.: Computational cost for DNS simulations with DGSEM. All computations
were run on the “Hermit” Cray XE6 at HLRS, except the TGV case, which
was run on the “Jugene” BlueGene P at JSC.
Among others, the following large-scale DNS simulations of turbulent flow problems
were conducted, demonstrating the suitability and performance of the full tool chain for
demanding multiscale simulations.
• Sphere at Re = 1000: A weakly compressible flow at Mach number M a = 0.3
and moderate Reynolds number based on the diameter Re = 1000 was computed.
Further details can be found in [100]. The cylindrical domain consists of about
20, 000 hexahedral elements, which are curved on the geometry surface. The
polynomial degree was chosen to be N = 4, leading to about 2.5 mio DOF for this
computation. The simulation was run on the Cray XE6 cluster “Hermit” on 4096
cores, with the computational effort for one characteristic time unit T ∗ = D/u∞
of 100 CPU-h, corresponding to a wall clock time per characteristic time unit of
91
1.5 minutes. Note that for this low load of only about 600 DOF per core, the PID
is not optimal.
• Vortex Cascade: As discussed in Section 2.1.1, vortex interaction is the mecha-

nism that produces a turbulent cascade. In this simulation, two large scale, sta-
tionary isentropic vortices of equal strength interact in a periodic domain of size
[20 × 20 × 20]. The Mach number based on the average initial vortex velocity is
set to M = 0.1, while the Reynolds number based on the same velocity scale and
the initial vortex diameter (D = 2) is Re = 1000. The domain is discretized by
483 cubic elements with 83 inner points per cell, leading to 56 · 106 DOF total.
We have computed this flow on 6, 912 cores on the Cray XE6, leading to a DOF
per core load of about 8, 000. The total wall clock time in this configuration for
800 time units was 6, 800s.
Figure 3.11 right shows the development of the spectrum of kinetic energy E(k)
over time, and a visualization of λ2 isocontours of the flow field is depicted in the
left subplot.
• Taylor-Green Vortex at Re = 5000: A DNS of the TGV described in Section 2.2

at a high Reynolds number of Re = 5000 was computed to establish a reference
DNS. Two discretizations were chosen, one leading to 3843 DOF, and the second
one with 6003 , resulting in 210 mio DOF per conservative variable. This is up to
date the largest reported computation with a DG method in literature. Figure 3.12
depicts a visualization of the vortical structures and the spectra of kinetic energy
at the time of maximum dissipation. The cut-off frequencies for both resolutions
and Kolmogorov’s scaling is also indicated. Note that a clear inertial subrange can
be observed, and also that even for 6003 DOF, the resolution in the dissipation re-
gion at high wave numbers is not yet fully converged.
Table 3.4 lists the computational details for the case with 503 elements and N =
11. 125k cores were used on the Bluegene P “Jugene” at JSC. Due to the high res-
olution, the explicit time step is very small, but the problem remained dominated
by the advective time step. Because of the lower clock speed of the Jugene and the
different architecture, the PID on this machine is generally worse than on the Cray
system, however, this is partially due to the file I/O, which is not consistent on this
number of cores. Nonetheless, due to excellent scaling of the framework, the full
simulation was completed in about 7 hours, thereby demonstrating its potential
for massively parallel multiscale simulations.
• Cylinder at ReD = 3900: This flow is a classical test case for LES and DNS. It
was computed at a Reynolds number based on the diameter D of ReD = 3900
and a Mach number M a = 0.1. At the given flow conditions, the boundary layer
92
remains laminar until the separation point, where the shear layer is shed period-
ically. The ensuing transition to turbulence occurs very close to the geometry in
the shear layer, and interacts with the resulting reverse flow region. We have sim-
ulated this flow on a grid with 319, 488 elements, spanning a circular domain of
radius R = 100D and ∆z = 8D, with elements of degree N = 7, leading to
164 mio DOF. The characteristic time is given by T ∗ = D/u∞ , and the charac-
teristic shedding time is TStr ≈ 5 T ∗ . Statistics were gathered over 50 shedding
cycles. Figure 3.13 shows a visualization of the flow features in the cylinder wake.
Due to limitations on the computing system at that time, the number of cores for
this simulation was restricted to 4992, which resulted in a deterioration of the PID
and very long wall clock times per shedding cycle.
In summary, we have presented in this chapter a highly efficient numerical scheme for
scale-resolving simulations and have demonstrated the capabilities of the whole frame-
work, from pre-processing to visualization, to handle massively parallel large-scale sim-
ulations.
Figure 3.10.: Flow over a sphere at Re = 1000 and M a = 0.3. The computation was
run the Cray XE6 “Hermit” at HLRS on 4096 cores. The picture shows
the laminar separation, the vortex street and the transition in the wake
(λ2 = −0.01 criterion, colored by vorticity magnitude). Computation by
F. Hindenlang [100].
93
Figure 3.11.: Vortex Cascade at Re = 1000. Left: λ2 isocontours, colored by helicity.

Right: Temporal development of energy spectrum.
-1
10
-2
10 k-5/3
10-3
10-4
48x8
-5 64x6
10 50x12
E(k)
100x6
-6
10
-7
10
10-8
10-9
-10
10 0 1 2
10 10 10
Wavenumber k
Figure 3.12.: Taylor-Green Vortex at Re = 5000. Computation with 3843 and 6003
DOF on 125, 000 cores on BlueGene P (“Jugene”) at JSC.
Left: Isocontours of vorticity magnitude, colored by helicity. Right: Spec-
trum of kinetic energy for 3843 and 6003 DOF, corresponding cut-off
wavenumbers denoted by dashed lines.
94
Figure 3.13.: Flow over a circular cylinder at M a = 0.1, Re = 3900. Computation

with 164 mio DOF on 4, 992 cores on Cray XE6 “Hermit” at HLRS. λ2 =
−0.001 criterion, colored by vorticity magnitude.
Top: side view. Bottom: Top view of near wake.
95
4. Numerics for Under-Resolved Simulations
In Chapter 3, we have presented a numerical scheme optimally suited for scale-resolving

simulations. It provides high geometrical flexibility due to unstructured meshes, as well
as excellent computational efficiency on High Performance Computing (HPC) platforms.
Due to its spectral character, the “solution quality per invested DOF” is very high com-
pared to other discretizations, which results in fast convergence for well-resolved prob-
lems.
As highlighted in Chapter 1, the full resolution of all scales in a multiscale problem
remains only feasible for a limited set of moderately complex flows. The limits of the
simulation envelope can easily be overstepped by a mere doubling of the Reynolds num-
ber. This is a different interpretation of the “curse of dimensionality” concept, as for
three-dimensional, time-dependent problems, the scales and therefore the resolution re-
quirements are coupled across all dimensions. Thus, full scale-resolving simulations are
the exception, while marginally or under-resolved simulations remain the rule. This in
turn opens two issues, both of which have been discussed in Sections 2.3 and 2.4: i) A
conceptual approach to limit the number of scales or to regularize the problem becomes
necessary. ii) If a perfect regularization is not possible and the closure is imperfect, the
solution of the regularized problem becomes dependent on the discretization of the reg-
ularizing mechanism, of an explicit or implicit closure and of the equations themselves.
In this chapter, we will focus on the LES approach to reduce the complexity of the
problem, and we will discuss some of the aspects of numerics and modeling that be-
come significant in under-resolved situations. We will discuss a strategy for LES with
DGSEM, and present results for high order LES simulations of turbulent flows.
4.1. Challenges
In this section, we will briefly outline the challenges that arise in under-resolved situa-
tions as opposed to well-resolved problems discussed in Chapter 3. As discussed before,
the fundamental challenge when simulating multiscale problems comes from the inher-
ent under-resolution of the physical quantities. Thus, only a fraction of the occurring
scales can be resolved and computed with sufficient accuracy. If we assume that the
problem supports some form of scale structure, which can be distinguished and classi-
fied into waves or scales, we can denote the maximum number of physically occurring
97
scales by K. In other words, the problem exhibits a range of scales of width K. Typi-
cally, only m << K scales can be resolved due to the computational limits
K
X m
X
aj (~x, t) ≈ e
aj (~x, t). (4.1)
j=1 j=1
When discretizing this problem, the approximation steps introduce several sources of
error: The first one stems from the truncation itself
K
X
ǫtrunc = aj (~x, t), (4.2)
j=m+1
which is due to the under-resolution of the problem and depends on two factors: i) the
ratio of m/K and ii) the significance of scales beyond m for the solution. The impor-
tance of the high wavenumber scales depends on the physical problem, and the ratio
m/K is determined by the amount of available computing resources, the time alloted
for the computation and the efficiency of the discretization in terms of nP P W . By defin-
ing the resolution of the computational grid via either the number of grid cells and/or the
value of the local polynomial degree, the maximum representable wavenumber on this
computational grid is limited by Nyquist’s theorem: A wave can only be detected by a
sign change, i.e. at least two points per wavelength are required to distinguish a signal.
Thus, for a given computational grid with total resolution (Nx × Ny × Nz ), the up-
per bound on representable wavenumbers is given the Nyquist criterion per dimension
as Nx /2, Ny /2 and Nz /2. This requirement of 2 PPW comes solely from informa-
tion theory, the only discretization that can resolve waves accurately up to this limit are
spectral Fourier methods, as discussed before, while other methods will incur a higher
truncation error. Still, Nx /2, Ny /2 and Nz /2 is a valuable estimate, as it represents the
theoretical limit of representable information on a given grid. One important aspect of
the truncation error is that it is determined by the local resolution and that it thus limits
the approximation quality a priori. Once the initial projection onto the available grid is
completed, the discretization then interacts with the truncated solution. This interaction
then introduces another source of error, namely the approximative determination of the
grid-resolved scales eaj (~x, t) via discretization operators, i.e. the error
ǫdiscj = |e
aj (~x, t) − aj (~x, t)|, for j = 1, ..., m. (4.3)
Note that for a non-linear problem, these discretization errors interact across the scales
and will eventually influence all resolved scales 1, ..., m, even if ǫdiscj of a scale j is
zero initially. Therefore, control and minimization of the numerical discretization error
of the overall framework is the key to an efficient and reliable simulation. This includes
98
4.1. Challenges
both the phase and the amplitude information (dispersion and dissipation error). Due to
these errors, only those parts of the solution which are resolved by n > nP P W points
per wavelength can be considered as accurately resolved, where nP P W denotes again
the theoretical limit depending on the discretization methods, as discussed in Section 3.1.
It should be reiterated that the error relations can only be determined for linear problems
and are then used as an estimate for the non-linear multiscale simulation. In Section 3.4,
typical values from 4 up to 7 for nP P W for efficient discretizations like high order poly-
nomial based methods where discussed, while low order approximations with inherent
upwinding can become prohibitively inefficient with nP P W ≥ 25. However, due to the
under-resolution, even when using appropriate numerical methods with non-excessive
dispersion and dissipation errors, there is always a range of scales in the approximate
solution that is resolved with less than the optimal points per wavelength and is thus af-
fected by numerical errors. As opposed to the discussion in Chapter 3, the fundamental
difference between well-resolved and marginally resolved situations is the fact that the
discretization is not only acting in the wavenumber region where its errors are minimal
or zero, but across the whole range of scales represented on the grid. Thus, in terms
of discretization accuracy, one should distinguish between resolved scales that carry no
considerable discretization error and the represented scales, which are afflicted by errors.
In Figure 3.2, the approximate range (determined by an arbitrary dispersion error bound)
of those two scale bands has been indicated.
In addition to the errors occurring on the represented scales, the fundamental issue of
non-linear multiscale problems like turbulence is that due to the interaction of the dif-
ferent scales, i.e. the influence of all scales onto each other, the temporal evolution of
the approximative coefficients e aj (~x, t) depends on the missing smaller scales, i.e. the
truncation error from Equation (4.2) describing the subgrid scales. The effect of these
scales not represented on the grid may have a significant influence on the represented
and resolved scales. Thus, an effective numerical discretization must be augmented by a
mechanism that models the physical effect of those missing subgrid scales for an accu-
rate solution of the multiscale problem.
The complexity of the multiscale problem stems from the fact that these three sources
of uncertainty (truncation, discretization and model) cannot be separated and even inter-
act with each other. Therefore, all aspects of a numerical model have to be taken into
account to judge its accuracy and efficiency.
4.1.1. Accuracy for Under-Resolved Problems

Numerical discretizations are typically constructed to minimize the approximation er-
rors across the resolved scales, while keeping the formal design order of accuracy (an
exception would be DRP FD schemes, that aim at extending the boundary between re-
99
solved and represented scales by sacrificing optimal convergence). In other words, they
are commonly designed and analyzed for h → 0, i.e. for a vanishing grid spacing and
perfect resolution. The often used classification of discretizations based on their formal
order of convergence is a statement about the behavior of the leading error term of the
approximation as h → 0, i.e. this concept is only meaningful when considering ’small’
h. In fact, for ’large’ h, a non-monotone error behavior is often observed, which only
asymptotically returns to the design order when the grid is refined. A silent assumption,
often not clearly stated in the high order CFD community, is that one needs sufficient
smoothness of the underlying problem for high order discretizations to be efficient. For
a multiscale problem, the underlying solution is smooth at the fine scale level, but can
often only be coarsely resolved due to the large range of occurring spatial and temporal
scales. This has the consequence of artificial roughness being induced by insufficient
grid resolution. With respect to the discretization parameter, this entails that h is ’large’
and thus the theoretical error behavior considerations for h → 0 are not valid and not
useful to judge the accuracy of high order methods. Statements about the superiority of
these methods thus cannot simply be translated to the under-resolved case, which is the
common case in practical fluid flow simulations and furthermore the rule in multiscale
simulations.
Thus, a ’non-small’ h is exactly the situation that occurs in a LES type approach to
simulate multiscale problems. The only exception to this situation would occur in an
explicitly filtered LES with a perfect subgrid closure, as shown in Section 2.4.2. If both
the filtering operation (also a form of discretization with associated error) and the clo-
sure were perfect in a sense that they would regularize the problem and keep the range
of scales bounded, then the associated grid spacing hdisc could be chosen independent
from the filter width ∆, and hdisc and thereby the associated discretization error could
be chosen arbitrarily small. However, the two basic assumptions in this idea (perfect
filter and closure) cannot be met, unless in specifically constructed test cases, which are
impractical for general use. Thus, in a typical LES setting, the problem is always under-
resolved and the discretization is applied not only in the regime h → 0 (it is of course
for well-resolved parts of the problem). It is important to note that in those situations,
the numerical discretization behaves drastically different to the resolved case and even
unpredictably due to the non-linear nature of the underlying problem. In short, we can
state that for ’large’ discretization parameters h, the error of the approximation is ’large’.
The exact definition of ’large’ is difficult and strongly depends on the actual underlying
problem, but error convergence in agreement with the design order of the scheme cannot
be expected.
A better measure for the accuracy of the numerical discretization is not its behavior
for small h, but rather its behavior when approximating solutions with a wide range of
scales. Analysis methods for non-linear problems are still in their infancy and the only
100
4.1. Challenges
possibility up to this point is to investigate their behavior for linear problems. In spite of
this limitation, it is worth noting that this is still an important analysis, since one basi-
cally measures the accuracy of the derivative operator, which is also used to approximate
the more complex non-linear problems.
6 0
N=10
N=10
-1
N=1-10
5 Exact
-2 N
N=1-10
Exact -3
4 N=1
N
-4
Re(Ω*)
Im(Ω*)
3 -5
-6
2
-7
N=1
-8
1
-9
0 -10
0 0.7854 1.5708
π/2 2.3562 3.141
π 0 0.7854 1.5708
π/2 2.3562 3.141
π
K* K*
10-1 10-1
-2 -2
N=1
10 10
N=1
10-3 10-3
-4 -4
10 10
10-5 10-5
*
Re(Ω )-K
Im(Ω )
*
*
10
-6 N=10 10
-6
N=10
10-7 10-7
-8 -8
10 10
10-9 10-9
-10 -10
10 10
0 0.62832
π/5 1.256
2π/5 0 0.62832
π/5 1.256
2π/5
K* K*
Figure 4.1.: Dispersion (left) and dissipation error (right) for DGSEM with Gauss
nodes. K ∗ = NK+1 is the normalized wavenumber and Ω∗ = NΩ+1 is
the corresponding modified normalized numerical mode (eigenvalue). Bot-
tom row: logarithmic plot with zoom on low wavenumber region. Note that
the error is cut off at 10−10 to avoid numerical noise. Plots with permission
from [85].
Figure 4.1 shows the dispersion and dissipation properties of DGSEM for varying poly-
nomial degree N . Together with Figure 3.2 and Table 3.1, they show two essential
issues: Firstly, for increasing N , the boundary between resolved and represented scales
is pushed to higher wavenumbers, making higher order schemes more accurate over a
101
wider scale range. Secondly, as the resolution per scale decreases, the associated er-
rors become significant, until at K ∗ = π, the wave is no longer represented by the
approximation. This is in agreement with Nyquist’s theorem which states that at least
two points per wavelength are necessary to capture a wave. The reasons for the over-
shoot in the dispersion error, which give DG schemes their restricting explicit time step
limit, have already been discussed before. For the dissipation error, the behavior to-
wards large N moves towards a spectral cut-off filter. Since the strong form of DGSEM
(Equation (3.31)) can be interpreted as a penalized spectral scheme, this behavior is not
surprising, as with increasing N , the relative importance of the penalty dissipation is
reduced, and the scheme becomes essentially a global spectral scheme in character.
Obviously, in a LES setting, all wavenumbers up to the Nyquist number will be part of
the solution. Furthermore, the scales close to the grid cut-off will not be energetically
small as in well-resolved situations, but contain significant dynamics. Thus, in such a
situation, due to the relatively coarse grid, large errors in the discretization often cannot
be avoided. When using the implicit filtering with implicit modeling approach for LES,
it is important to note that these discretization errors are tuned in such a way that they
replace the subgrid scale effects. In this sense, those discretization errors are in prin-
ciple not a problem for this approach, since they are actively taken into account in the
LES approach. On the other hand, the purely explicit filter and explicit modeling ap-
proach is not negatively impacted by this either, as in this case, the filter width is chosen
larger than the grid resolution. Due to the explicit filtering of the solution, one actively
ensures that only well resolved solution components are represented by the numerical
approximation. In this case, the computation of the derivatives of the filtered solution
components has small errors. However, when using the mixed approach, implicit fil-
tering with explicit modeling, the solution has components with wavenumbers down to
the Nyquist limit. The derivative operator working in this wavenumber range will then
be inaccurate. In this case, the mixed approach with implicit filtering and explicit mod-
eling has to deal with an interaction of the explicit turbulence model and the inherent
numerical discretization errors. Due to the non-linearity of the underlying problem, this
interaction is hard to predict and to quantify and can be either positive or negative on the
overall fidelity of the approach. Hence, the mixed approach – although the most com-
monly used approach in LES – is the least suitable LES approach, except if one tunes
the subgrid scale model depending on the actual discretization used.
4.1.2. Stability for Under-Resolved Problems

Stability of a discretization or numerical simulation is influenced by a number of factors,
and the term “stability” is used in many different contexts, i.e. stability of an interpo-
lation with respect to the nodes, temporal integration stability with respect to the time
102
4.1. Challenges
step, stable boundary conditions, etc. While all those issues also have to be captured in
under-resolved situations, a somewhat special stability issue becomes more prevalent for
high order methods. We will discuss it here briefly and give more details in Section 4.2.
An important issue not captured by the linear analysis is the influence of the approxi-
mation of non-linear terms, e.g. the convective terms in the Navier-Stokes equations.
As discussed in Section 2.1.1 and in the discussion about Equation (2.2), the convective
term drives the scale cascade and couples the interaction across the whole wavenumber
range. It effectively sums the wavenumbers of two interacting scales, thereby introduc-
ing higher components. For the Burgers’ or incompressible Navier-Stokes equations, the
non-linearity is quadratic, i.e. a scale just at the grid Nyquist wavenumber NN yq will
produce a scale at 2 NN yq , thereby leaving the range of representable scales. For the
compressible Navier-Stokes equations, where the Euler fluxes are cubic, this situation
becomes worse. By definition, when introducing a cut-off through a LES resolution, this
cut-off will almost always be located in the range where the non-linear term is dominant
(inertial subrange). This scale truncation of the term introduces an additional source of
error (besides the truncation itself), as the higher wavenumbers that are created contain
information that must be evaluated correctly when solving the underlying equation. Cor-
rect evaluation of the represented scales then implies exact evaluation of these subgrid
terms, which is contradictory to the definition of the subgrid itself, i.e. a grid with NN yq
as a cut-off is not capable of resolving scales at 2 NN yq . Still, these higher wavenumber
components must be incorporated into the evaluation of the non-linear terms on the ex-
isting grid. The most basic – and mathematically problematic – way of doing that is by
aliasing the higher frequencies onto the representable grid scales. However, while this is
numerically efficient, it can lead to an unstable operator and thus to a guaranteed unsta-
ble solution. How this problem of aliasing is solved determines greatly the stability of
high order methods in under-resolved situations, more details will follow in Section 4.2.
The issue of aliasing instabilities could also be resolved by an explicitly filtered LES
approach. In this case, the size of the smallest occurring scales and thus the small-
est aliasing wavenumber can be decoupled from the spatial resolution, which is chosen
sufficiently fine such that all sub-filter scales and their non-linear interactions are well-
resolved. However, although such an explicit filtering approach is very beneficial on
paper, the additional work associated with the proper resolution of the smallest eddies
comes with a high computational cost. Another reason that explicitly filtered LES is
less prevalent than its implicit counterpart is the additional mathematical and numerical
complexity introduced through the explicit filter operator, in addition to some open is-
sues such as the commutation properties of the filter at boundary conditions and the filter
isotropy on non-regular grids. For those reasons, practical applications of explicitly fil-
tered LES are relatively scarce [26, 133]. Therefore, we will focus on implicitly filtered
LES, where the discretization itself acts as a filter via the projection of the solution onto
103
the computational approximation space in every time step.

Besides the issues stemming from improper discretization of non-linear terms, another
source of instability comes from the truncation error itself. Since in LES, the cut-off is
typically chosen far away from the dissipation region, the solution lacks a regularization
mechanism to account for the missing scales or their effects. Also, the lack of repre-
sented scales blocks the scale cascade, as not all partners for the triadic interaction are
represented on the grid, leading to errors introduced into the large scales [63]. There-
fore, if these effects are not modeled properly, instabilities may occur from “insufficient”
physics representation. The mechanisms by which this problem manifests itself are typ-
ically strong gradients and oscillations in the velocity field, i.e. the velocity field appears
“rough”. For incompressible formulations, this can lead to convergence problems for
the elliptic pressure solver, but the underlying equations are usually not violated. For
compressible solvers, this situation is quite different, as strong oscillations will quickly
shift energy from internal to kinetic energy (the total energy is the conserved variable),
causing violations of the constitutive equations, i.e. typically the gas law. Therefore,
instabilities introduced by truncation are caused by a non-perfect closure, not by numer-
ical approximations. Thus, a fundamental problem with under-resolved approximations
of turbulence is that at least two de-stabilizing mechanisms exist. Therefore, the turbu-
lence modeling aspect has to be extended by these stability issues: In a non-optimal LES
approach, the role of the turbulence model is often first to stabilize the computation and
not to model the physical effect of missing scales. Especially when using the mixed LES
approach, i.e. implicit filter and explicit model, one deals with solution components at
Nyquist’s limit and hence with possible aliasing issues. For the purely explicit filtering
and explicit modeling approach aliasing issues can be circumvented when the explicit
filter width is large enough compared to the actual grid size. In the implicit filtering and
implicit modeling approach, the issue of aliasing and stability is actively incorporated
into the discretization and accounted for by tuning the discretization errors to act as the
turbulence model. It is worth pointing out that there are discretizations and formulations
which are aliasing free. Most prominent is the case of the incompressible Navier-Stokes
equations, which can be formulated in the so called skew-symmetric form. Discretiza-
tions based on this skew-symmetric form retain the kinetic energy balance and are hence
free of stability issues [144]. For the compressible Euler equations, such energy-bounded
forms are a current topic of research [83, 172].
In summary, three issues arise in under-resolved discretizations. First, the accuracy of

a discretization strongly depends on the resolution of each scale and deteriorates with
increasing wavenumber. Methods with good dispersion and dissipation behavior over a
large range of scales are favorable for this task, whereas the property ’formal order of
convergence’ is in such cases irrelevant due to the discretization parameter h being large
104
4.2. Aliasing Errors
and not asymptotically small. Second, operating near Nyquist’s limit opens the door to
aliasing related issues which can even affect the stability of the overall discretization.
The impact of this non-linear stability is different for the three different LES approaches
and must be taken into account individually. Lastly, modeling of the truncation error or
its effects must regularize the problem and prevent a violation of the conservation laws,
which could render the truncated problem unstable due.

The term aliasing is often defined as “misrepresentation of high frequency content (be-
yond the grid cut-off) as lower frequencies”. This is a rather general description, which
de-emphasizes the most important aspect of aliasing in terms of multiscale problems
like turbulence – its self-feeding mechanism. Aliasing occurs when a physical non-
linear mechanism is wrongfully represented on a lower-dimensional grid, i.e. aliasing is
the violation of some (physical) governing law by numerical inaccuracy. It is obvious
that for fully resolved simulations, where the grid resolution is sufficient to resolve all
scales and their occurring interactions, aliasing is not an issue. For under-resolved sim-
ulations where the modal content in the higher modes is not negligible, these aliasing
errors accumulate energy in the high modes, and eventually lead to an instability of the
scheme. In particular, if the higher modes become excited and more energetic through
aliasing, they will in turn generate additional high energy aliasing errors. Or, in terms of
turbulent scales, when energy is pushed into the scale range near grid cut-off, the cascade
to smaller scales (i.e. the feeding of the aliasing mechanism) will be enhanced. Thus,
aliasing is directly related to the under-resolution of a multiscale problem, and thereby
to the LES approach. Also, the non-linearity of the physical process is important, as it
leads to a growth of the aliasing error and finally to the blow-up.
In this section, we will give details about the origin and effect of aliasing and on numer-
ical strategies countering it.
4.2.1. Source and Effect

4.2.1.1. Conservation of Kinetic Energy
As stated above, aliasing by misrepresentation of the non-linear term can violate the
flow physics. In Section 2.1.2.3, we discussed that the non-linear term in spectral space,
when summed over all wavenumbers, cancels out, and does not contribute to the kinetic
energy equation (see also Equation (2.19)). We also demonstrated in Equation (3.18)
how a misrepresentation (rather: an under-representation) of non-linear products in a
collocative manner can introduce an error term into the representation that acts on the
105
high wavenumbers. We will repeat this discussion here briefly with a very simple model
for the continuous incompressible Navier-Stokes equations.
Starting from the Equation (2.17), we take the scalar product of the momentum equation
for ui with ui itself, just focusing on the non-linear term:

∂ui uj 1 ∂ui uj ui 1 ∂uj
ui · = + ui ui . (4.4)
∂xj 2 ∂xj 2 ∂xj
Since the divergence condition ui,i = 0 holds, the continuous non-linear term thus con-
serves kinetic energy and only modifies it through fluxes across the domain boundaries.
We now assume some error of the form k(ui ), with k ∈ R, in the continuous term
(as a simplistic model for an error introduced by the differential operator acting on the
non-linearity), this error modifies the representation of ui uj to ≈ ui uj k(ui ). Then, the
contribution of the convective term becomes

∂ui uj k 1 ∂ui uj ui k 1 ∂uj 1 ∂k
ui · = + ui ui k + ui uj ui , (4.5)
∂xj 2 ∂xj 2 ∂xj 2 ∂xj
where the first two terms on the right hand side again do not change the kinetic energy
budget, but a third term that is generally non-zero and has an unknown sign appears.
This term thus acts as a source (or sink) for the kinetic energy, thereby violating the
kinetic energy budget.
While this simple demonstration was based on the continuous equation, the same mech-
anism (the misrepresentation of the non-linear term) can be introduced through the dis-
cretization of this term. Figure 4.2 depicts the influence of the aliasing error on the flow
physics for the Taylor-Green vortex. Focusing on the left subplot first, the evolution
of the mean temperature (as a measure of internal energy and normalized by the initial
value) and the kinetic energy dissipation rate are shown. Since the discretization con-
serves the total energy and all boundary conditions are periodic, the energy just changes
its form between internal and kinetic energy. The black curves denote a discretization
which allows aliasing errors, while the red curve is an aliasing-free solution. More de-
tails on the de-aliasing procedure are given in Section 4.2.2. Note that the transfer of
kinetic to internal energy is enhanced for the aliasing-influenced solution, resulting in a
higher final mean temperature. This is the effect of the increased dissipation rate, which
in turn results from the excitation of the higher wavenumbers through aliasing. The
aliasing errors transfer kinetic energy into the upper third (for an essentially incompress-
ible flow) of the polynomial modes, thereby enhancing vorticity on these scales, which
is then attacked by both physical and numerical dissipation. Therefore, although the
aliasing energy is initially injected into the kinetic energy cascade, it eventually results
in higher dissipative losses, if numerical and resolved dissipation are strong enough to
keep the simulation stable. The right plot in Figure 4.2 shows the temporal evolution of
106
maximum of the vorticity within the domain. Clearly, the aliasing-influenced vorticity
shows more pronounced peaks, resulting from the aliasing energy injected into the high
wavenumber modes. As discussed above, this increase in small scale activity results in
higher energy dissipation. Thus, these plots indicate that aliasing errors indeed act as a
source term within in the kinetic energy equation, locally increasing gradients and high
wavenumber energies, which – if kept bounded – induce a more rapid transfer to internal
energy.
1.0006
Tmean 0.015
50
1.0005
40
1.0004
0.01
Tmean/Tmean
max| |
-dk/dt
0
N=7, M=7
N=7, M=8 30
1.0003 N=7, M=9
N=7, M=10
N=7, M=11
1.0002 20
0.005
N=7, M=7
1.0001 10 N=7, M=8
-dk/dt N=7, M=9
N=7, M=10
N=7, M=11
1 0
0 5 10 15 20 25 6 7 8 9 10 11 12 13 14
Time Time
Figure 4.2.: Taylor-Green Vortex at Re = 1000. Effect of aliasing on flow physics.

Left: Mean temperature and kinetic energy dissipation rate over time.
Right: Maximum vorticity over time.
4.2.1.2. Discretization of the Non-Linear Term
In Section 3.2, we introduced the concept of inexact evaluation of non-linear products

through collocation for the pseudo-spectral methods. The same error is introduced into
a DG method when the inner products of non-linear terms are not evaluated exactly,
i.e. when the underlying projection onto the test functions becomes inexact. It should
be noted that in DGSEM, where the interpolation and integration are collocated on the
same nodes, both operations introduce this form of error, as the Gauss quadrature relies
on the interpolation of the integrand. We will now discuss how the aliasing error occurs
in DGSEM, depending on the discretization choice of the non-linear term.
107
In Section 3.3.2, we introduced the transformed flux function in Equation (3.26) as
M
X
~ ≈
F l (ξ) l
F̂ijk M ~
ψijk (ξ), l = 1, 2, 3, (4.6)
i,j,k=0
l
F̂ijk ~ x U ) |~ ,
= G(U ) − H(U, ∇ (4.7)
ξijk
where the non-linearity of the advection part of the flux F l (U ) is either quadratic for
incompressible flows or cubic in a compressible setting with respect to the primitive
variables. When evaluating the compressible flux in terms of conserved variables, it
even becomes a rational function. Note that in Equation (3.26), we have chosen an
ansatz of degree M for the fluxes, which is independent from the ansatz of degree N
for the solution itself. Since F l is a non-linear product of the solution U and therefore
may have a larger wavenumber range than the solution, a collocation of F l on (N + 1)3
solution points ξ~ijk , which in our implementation double as integration nodes, would
introduce an error in the interpolant (Equation (3.26)) and subsequently in the associated
integration, the volume contribution of Equation (3.29):
Z Z M
X
l ∂ ~ dξ~ ≈ l M ~ ∂ ~ dξ~
F φ(ξ) F̂ijk ψijk (ξ) l φ(ξ)
∂ξ l ∂ξ
E E i,j,k=0
M
X M
X l M ~M ∂ ~
≈ F̂ijk ψijk (ξpqr ) φ(ξ) ωpM ωqM ωrM
p,q,r=0 i,j,k=0
∂ξ l ~ ξ
ξ= ~pqr
M
X l ∂ ~
= F̂pqr φ(ξ) ωpM ωqM ωrM ,
p,q,r=0
∂ξ l ~ ξ
ξ= ~pqr
(4.8)
where we have introduced the approximation of the integral through a tensor product of
one-dimensional Gauss quadratures with (M + 1)3 associated nodes (ξ~pqr M
) and weights
M M M l
(ωp , ωq , ωr ). Since F is at most a polynomial of degree 3N for compressible flows,
there exist values of M for which the quadrature of the right hand side of Equation (4.8)
becomes exact. From the exactness of the quadrature rule, M is found to be 2N for
cubic and 23 N for quadratic non-linear integrands. Consequently, an inexact integration
of these terms, i.e. an insufficient choice of M , results in an approximation error of the
non-linear fluxes: the aliasing error discussed in Section 4.2.1.1.
108
4.2.1.3. Operator Spectrum

We now turn to an investigation of the effect of this inexact integration on the stability
of the DG operator itself, as shown in [19]. To highlight the effects of the choice of M ,
we consider the following linear transport model problem with non-constant advection
velocity in a periodic domain

ut + a(x) u x = 0, x ∈ [−1; 1]. (4.9)
By choosing the advection speed a = a(x) > 0, we can simulate aliasing and can
investigate its effect on the operator spectrum for different number of integration points
M + 1. While it might seem that the linear transport Equation (4.9) is too simplistic
to serve as a model for aliasing effects in the full Navier-Stokes equation, it should
be noted that Equation (4.9) can be interpreted as a linearization of Burgers’ equation
about a given “turbulent” state u0 (x) = a(x). Since the scale-producing mechanism of
Burgers’ equation is similiar to that of the Navier-Stokes equations, we can control the
non-linearity in the convective term and thereby the aliasing effects.
For a single grid cell [−1; 1], periodic boundary conditions and an upwind flux at the
element edges, the DGSEM formulation (3.40) in matrix-vector notation reduces to [85]
∂u
M + S au− + V au = 0, (4.10)
∂t
where u := (u0 (t), ..., uN (t))T is the vector of nodal DOF, au is the collocation vector
of u and a, and ( )− indicates an evaluation from the (periodic) neighbor. The matrices
are given by
Mij = δij ωi ,
Sij = −ℓi (−1)ℓj (1), (4.11)
Vij = ℓi (1)ℓj (1) − Dji ωj ,
with Dji , ℓi and ωi as defined in Section 3.3.2.5. Inverting the diagonal mass matrix and
collecting terms, this can be written as a linear dynamical system for fixed a(x) as
ut = AM u, (4.12)
where A is the spatial operator containing the surface and volume fluxes integrated with
a specific Legendre Gauss integration rule M (and thus M + 1 integration nodes and
integration precision 2M + 1). Note that for this linear system, the solution u remains
stable if all eigenvalues of the operator matrix have zero or negative real parts. We now
take a look at different choices of a(x) and M , and the associated eigenvalue spectra.
109
Constant Advection Velocity

For this simplest case, we choose the advection speed a = 1 = constant. Thus, if the
solution u is a polynomial of degree N , the flux function is also a polynomial of degree
N . Then, N + 1 Gauss points with integration precision 2N + 1 are more than sufficient
to integrate the flux exactly. The corresponding spectrum of the DG operator for N = 15
is plotted in Figure 4.3. As expected, the linear flux can be treated exactly with N + 1
points and thus the spectrum does not change when increasing the number of integration
points to 26. The maximum of the real parts of the eigenvalues is 0, confirming the
stability of the DGSEM operator for this linear advection velocity.
45
N=15, Int Points=16
N=15, Int Points=26
30
15
Im(λj)
-15
-30
-45
-100 -50 0
Re(λj)
Figure 4.3.: Operator spectrum for constant advection speed and N = 15 with either
M + 1 = 16 or 26 integration points, respectively. The maximum of the
real parts of the eigenvalues is 0.
Polynomial Advection Velocity

We now select a polynomial advection speed, as this can still be integrated exactly with
polynomial-based integration rules. We choose
a(x) = 1 + (1 − x2 )15 , (4.13)
a polynomial of degree 30. For the volume inner product, the flux is multiplied by
the spatial derivative of the test function (through the matrix D), resulting in a total
polynomial degree of Ntot = 30 + N + (N − 1) = 2N + 29. For an exact evaluation,
110
45
N=15, Int Points=16
30
15
Im(λj) 0
-15
-30
-45
-100 -50 0
Re(λj)
Figure 4.4.: Operator spectrum for polynomial advection speed and N = 15 with
M + 1 = 16 integration points (original DGSEM scheme). The maximum
of the real parts of the eigenvalues is 0.016.
N + 14 Gauss points are needed. Figure 4.4 shows the spectrum of the resulting DG
operator for N = 15 and M + 1 = 16 integration points. The general shape of the
spectrum is similar to the constant case, as the selected polynomial advection speed lies
between 1 and 2. Note that this discretization, with a collocation of the flux and the
solution on the same grid, is the original form of DGSEM [120]. A closer look at the
maximum of the real parts of the eigenvalues reveals that the value is positive and about
0.016, which means that due to the insufficient integration precision, the DG operator
has become unstable, and would produce an unbounded solution in time, regardless of
the time step. In other words, returning to the discussion in Sections 4.2.1.1 and 4.2.1.2,
the inexact treatment of the convective term has resulted in a violation of the governing
equation already through a faulty discretization operator.
Since the advection speed is a polynomial, increasing the integration precision M should
recover the original stability of the scheme. Figure 4.5 shows the resulting spectrum
as a function of M , and Table 4.1 summarizes the maximum of the real parts of the
eigenvalues of the associated DGSEM operator. Both the table and the right plot of
Figure 4.5 show that by increasing the number of integration points, the real part of the
operator eigenvalues move from the unstable region towards the imaginary axis. For
M = 29 integration points and above, the real parts of the eigenvalues are not positive
anymore, resulting in a stable de-aliased approximation.
111
45
N=15, Int Points=16
N=15, Int Points=17 15
30 N=15, Int Points=18
N=15, Int Points=29
N=15, Int Points=16
Im(λj)
Im(λj)
N=15, Int Points=17
0 0 N=15, Int Points=18
N=15, Int Points=29
N=15, Int Points=39
-15
-30
-15
-45
-100 -50 0 -0.15 -0.1 -0.05 0 0.05
Re(λj) Re(λj)
Figure 4.5.: Operator spectrum for polynomial advection speed and N = 15 with dif-
ferent number of integration points M + 1. The right plot shows a zoomed
view of the imaginary axis.
Sinusoidal Advection Velocity

As a final test, we choose two non-polynomial functions

1 2
a(x) = 2 + sin 2π x + sin 3π x , (4.14a)
2 3
8
!
1 X 14
a(x) = 2+ sin jπ x , (4.14b)
2 j=2
5j
where (4.14b) is just a slight modification of (4.14a) with additional higher frequency
terms and a weaker decay of amplitudes. These sinusoidal functions are chosen to re-
Int Points M + 1 16 17 18 19 29 39
max[Re(λj )] 1.6 · 10−2 1.7 · 10−2 1.4 · 10−4 2.6 · 10−6 0 0

j
Table 4.1.: Maximum values of the real parts of the eigenvalues of the DGSEM opera-
tor (N = 15) spectrum with different number of integration points for the
polynomial advection velocity.
112
semble the fluctuating velocity field characteristic of turbulence more closely than the
previous test cases.
Due to their non-polynomial nature, it is not possible to integrate the resulting flux func-
tion exactly using Gauss-based integration rules. However, by increasing the integration
precision to a sufficiently high degree, the integration error can be driven to machine zero
and thus becomes negligible. The amount of Gauss points required to achieve this elimi-
nation of the error clearly depends on the considered advection speed function, i.e. on its
type (polynomial vs. non-polynomial) and on its spectral energy content. The results for
advection speed (4.14a) and (4.14b) are shown in Figures 4.6 and 4.7, respectively. The
maximum values of the eigenvalue real parts for both functions are shown in Table 4.2.
15
45
N=15, Int Points=16
N=15, Int Points=17
N=15, Int Points=19
N=15, Int Points=16
Im(λj)
Im(λj)
N=15, Int Points=17

0 0 N=15, Int Points=18
N=15, Int Points=19
N=15, Int Points=32
-15
-30
-45
-15
-100 -50 0 -0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1
Re(λj) Re(λj)
Figure 4.6.: Operator spectrum for sinusoidal advection speed, Equation (4.14a), and
N = 15 with different number of integration points M + 1. The right plot
shows a zoomed view of the imaginary axis.
For both sinusoidal functions, the original DGSEM formulation with M = N contains
eigenvalues with positive real parts. Increasing the amount of integration points moves
these unstable eigenvalues towards the imaginary axis and beyond, thereby recovering
stability. For M ≥ 19 for the lower frequency function and M ≥ 21 for the higher
frequency one, the resulting de-aliased operator is stable with non-positive real parts of
the eigenvalues. As expected, the higher frequency advection velocity (4.14b) requires
more integration effort to achieve stability.
113
15
45 N=15, Int Points=18 N=15, Int Points=18
N=15, Int Points=19 N=15, Int Points=19
30 N=15, Int Points=20 N=15, Int Points=20
15
Im(λj)
Im(λj)
0 0
-15
-30
-45
-15
-100 -50 0 0 1 2 3
Re(λj) Re(λj)
Figure 4.7.: Operator spectrum for sinusoidal advection speed, Equation (4.14b), and
N = 15 with different number of integration points M + 1. The right plot
shows a zoomed view of the imaginary axis.
Eqn. Int Points M + 1 16 17 18 19 20 21 22 32
(4.14a) max[Re(λj )] 1·10−1 3·10−2 3·10−2 0 0 0 0 0

j
(4.14b) max[Re(λj )] 4 4·10−3 2.8 0 1.2 0 0 0

j
Table 4.2.: Maximum values of the real parts of the eigenvalues of the DG operator
(N = 15) spectrum with different number of integration points for the sinu-
soidal advection velocity.
4.2.2. De-Aliasing Strategies
As discussed in Section 4.2.1, the approximation of the flux function is an important

discretization choice that determines the stability of the scheme. For the Navier-Stokes
equations, the flux function is a non-linear function of the solution and thus in particu-
lar not a polynomial of degree N anymore. The inaccurate approximation of the inner
products can cause severe aliasing issues for high order methods and influence the solu-
tion quality, even if the overall solution remains bounded due to numerical and resolved
114
dissipation.
In the following, we discuss different strategies for the discretization of the non-linear
term to remove aliasing errors and provide a neutrally stable discretization.
4.2.2.1. DGSEM
Starting with the most basic discretization, the DGSEM scheme introduced in Sec-
tion 3.3.2, we note that this scheme is not a properly de-aliased scheme, but just the
opposite due to the collocation. It is listed here as a first variant for two reasons: i) It
serves as a reference scheme in terms of accuracy, efficiency and stability for the follow-
ing discussion and ii) for low N , the numerical dissipation of the scheme can provide
some robustness against aliasing errors, so that this can be interpreted as an “implicitly
de-aliased” version. For this approach however, stability cannot be shown or guaran-
teed, as the introduced numerical diffusion is dependent on the local resolution and may
or may not be sufficient to prevent aliasing accumulation. It should be noted here that the
aliasing mechanism is self-sustaining and self-inciting: If energy is aliased into the upper
modes, these modes will in turn excite a transfer down the cascade to higher (subgrid)
modes, which will again be deposited onto the represented scales. Thereby, the energy
content in these modes increases constantly. Therefore, if constant energy is supplied
to the cascade, the aliasing mechanism will either cause a blow-up due to the continued
increase in high frequency content, or it must be bounded by some form of de-aliasing.
As shown in Section 3.4, DGSEM is the fastest variant of the DG methods, due to the
collocation of the nodal interpolation approximation and the evaluation of the integrals,
i.e. the same set of interpolation and integration nodes. By choosing the integration rule
in Equation (4.8) as M = N , we use the same ansatz for the fluxes as for the solution,
reducing the effort in the evaluation of the nodal flux values. Furthermore, the quadra-
ture rules are based on the same nodes, making the evaluation of the integrals highly
efficient. The price one pays for this high efficiency is that the approximation of the flux
function with polynomial degree M = N is prone to severe aliasing errors. Note that
this issue is certainly not exclusive to the DGSEM variant. The choice of M = N for the
precision of the quadrature rule is very common in the DG community for both the nodal
DG and the modal DG discretizations and hence all of those variants are vulnerable to
the same aliasing issues which result from this under-integration.
4.2.2.2. Projection DG
As demonstrated in Section 4.2.1.3, inexact integration causes operator instabilities.
Thus, a natural remedy for this problem, first discussed by Kirby and Karniadakis [111],
is to increase the integration precision according to the non-linearity of the flux func-
115
tion. Based on the general idea of de-aliasing by exact projection developed by Orszag
[153], they showed that de-aliasing for polynomial spectral methods (often called over-
integration in this context) is achieved similarly by choosing a sufficient integration pre-
cision with respect to the flux function. In other words, the non-linear functions are
evaluated with sufficient precision to remove the aliasing ambiguities, i.e. on a finer grid
(or with more DOF), and the resulting flux is then incorporated into the spatial opera-
tor on the solution grid in a consistent manner, i.e. by projection. The drawback is the
increase in the computational complexity through the number of quadrature nodes. For
weakly compressible flows at low Mach numbers M a ≤ 0.1, the non-linearity of the
flux function becomes approximately quadratic, and thus the Gauss quadrature rule de-
mands M = 32 N , for higher Mach numbers, M has to be 2N for complete de-aliasing
(disregarding the fact that the fluxes become rational, when computed from polynomial
conservative variables). Another very important aspect when choosing the quadrature
rule is that the transformation of the fluxes onto the reference element incorporates the
element mapping (through the grid metrics and the Jacobian) into the covariant trans-
formed fluxes (see Equations (3.22) and (3.26)). Thus, the actual flux function to be
integrated is no longer solely determined by the complexity of the physical flux, but also
by the polynomial mapping. Depending on the degree of the mapping, a higher integra-
tion precision might be necessary for exact integration; however, a full investigation of
this is still missing. One important distinction should be kept in mind: While insufficient
integration of the mapping will introduce an error into the solution, this aliasing error is
static in the sense that it does not depend on a temporally evolving field like dynamic
solution aliasing: The grid mapping is constant in time, therefore, this error contribution
will not self-feed.
Returning to the integration rule as prescribed by the physical fluxes, the choice of M
clearly governs the overall computational costs, which is for high N – according to the
discussion in Section 3.3.2.5 – mainly a function of the integration effort for the volume
3
integral. Thus, the general behavior observable in actual computations is ∼ M N
, i.e.
3
about 8 in case of M = 2 N and about 3.3 in case of M = 2 N .
It should be noted that while the values for M derived from the exactness of the quadra-
ture rule result in an exact integration and thus an exact projection of the fluxes, our
experience and other published work indicate that a “full” de-aliasing might not always
be necessary. Malm et al. [136] showed that for a continuous spectral element method
for incompressible flow, a de-aliasing strategy with M ∗ modes with N < M ∗ < 32 N
can be sufficient, especially when combined with some form of dissipation like SVV.
They proposed that only the first aliasing mode has to be integrated exactly, as it is the
most energetic one. In our investigations, we found “incomplete” de-aliasing to be suffi-
cient for stabilization of medium Reynolds number flows at moderate spatial resolutions.
116
In particular, we noted that stability could be regained for these situations by choosing
M ∗ = N +1 or M ∗ = N +2 [18]. In combination with the numerical dissipation of the
scheme, incomplete de-aliasing can thus be sufficient, but does not guarantee stability. It
should be remarked (as noted by Kirby and Karniadakis in [111] and shown in Table 4.2),
that an increase in M ∗ does not show a monotone convergence towards stability. Fig-
ure 4.2 shows an example of incomplete de-aliasing. Although the DGSEM version
(N = 7, M = 7) is stable due to the supporting numerical dissipation, de-aliasing with
M = N + 1 = 8 results in a significant solution change, in particular in the maximum
vorticity, indicating that the first aliasing mode is the most important one. Subsequent
increase of M only results in minor changes and a convergence of the observed quanti-
ties. In the following, unless stated otherwise, we will consider the complete polynomial
de-aliasing as given by the nature of the flux function. As in this particular DG discretiza-
tion the inner products are evaluated exactly (with respect to machine precision), i.e. by
projection and not by interpolation, we abbreviate this variant as PDG (projection DG).
Other commonly used names are “over-integrated DG” or “polynomial de-aliasing DG”.
• Remark 1: For well-resolved simulations, like DNS of turbulent flows, the dif-
ference between the PDG and the DGSEM variants decreases to the point where
the DGSEM variant becomes the best choice due to its efficiency. In contrast to
the DNS situation in Chapter 3, here we are interested in the behavior of these
discretization in a LES type setting where the simulation is by definition under-
resolved and the grid size is large, and thus aliasing errors will be important.
• Remark 2: Equation (3.29) highlights another important aspect of the de-aliasing

operation: once the inner products are evaluated exactly through sufficient quadra-
ture, the only remaining parameters in the spatial discretization are the viscous and
∗
inviscid numerical flux choices Hn and Gn∗ . Thus, through exact integration, the
“original” DG scheme and its properties can be recovered and analyzed, without
interference from the choices of discretization. Therefore, this formulation is par-
ticularly attractive for the evaluation of the DG scheme, the effects of LES subgrid
scale models or the choice of the flux functions [20].
• Remark 3: Extension of our DGSEM framework to incorporate this approach

is straightforward by using a sharp modal cut-off filter. The calculation is thus
performed with a polynomial degree M , i.e. the full spatial operator is evaluated
on the fine grid that supports the non-linearity. The approximate solution is then
filtered in every Runge-Kutta stage before evaluating the spatial residual using the
117
filter coefficients (
1 if 0 ≤ i ≤ N ,
σi := (4.15)
0 if N + 1 ≤ i ≤ M ,
which makes certain that only N + 1 modes are used for the actual computation,
but ensures that the inner products are evaluated with the desired amount of inte-
gration points. It is important to note that this exact quadrature does not increase
the total number of DOF available for resolving a flow field, but it only removes
collocation as a source of error.
• Remark 4: For DGSEM, where solution and fluxes are evaluated on a grid sup-
porting N polynomial modes, an aliasing-influenced flux operator will pollute (in
an incompressible situation and in a single operator evaluation) the upper third
of the available modes. Opposed to this, for PDG, due to the expansion of the
flux on the fine M + 1 grid and the subsequent projection back onto the solution
grid N + 1 through polynomial de-aliasing, the solution modes stay error-free
(except for the dissipation and dispersion errors and the effects of the subgrid clo-
sure) up to N . Kirby and Karniadakis [111] showed that polynomial de-aliasing
is not a purely dissipative mechanism, as the magnitudes of modes either decrease
or increase due to the over-integration. Thus, one interpretation of this form of
de-aliasing is that it recovers the full solution modes up to N .
4.2.2.3. Filtering
As discussed above, aliasing in a polynomial based approximation typically causes an

overestimation of higher polynomial modes, resulting in an oscillatory and even unsta-
ble approximation. This high frequency aliasing energy can be removed by dissipative
filter techniques that dampen the magnitude of the affected modes. In combination with
a collocation discretization like DGSEM, this can result in a stable simulation. One ex-
ample of such an approach is an exponential based modal filter, e.g. [96]. To construct a
modal, hierarchical basis from our nodal representation (Equation (3.25)), we introduce
a modal orthogonal Legendre basis {ϕj (ξ)}N j=0 and compute the Vandermonde matrix
V associated with our nodal interpolation points {ξi }N
i=0
Vij := ϕj (ξi ), i, j = 0, ..., N, (4.16)
e and vice versa

which allows us to transform any nodal DOF U to modal DOF U
e
U =V U e = V −1 U.
U (4.17)
118
A one-dimensional filter matrix, which can again be applied in a tensor product ap-
proach, can be pre-computed as
e V −1 ,
F =V F (4.18)
with the modal filter matrix
Feij = δij σi , i, j = 0, ..., N, (4.19)
and the modal filter coefficients 0 ≤ σi ≤ 1 with

(
1 s if 0 ≤ i ≤ Nc − 1,
σi := (4.20)
exp −α Ni+1−N +1−Nc
c
if Nc ≤ i ≤ N ,
where α, s and Nc are the filter parameters. Nc indicates the number of the unaffected
modes, α is typically chosen such that exp (−α) is machine zero and s is an even num-
ber determining the strength of the filter (higher s means weaker filtering in terms of a
sharp cut-off), see e.g. the book of Hesthaven and Warburton [96]. Clearly, this modal
filter is a purely dissipative mechanism. If applied to a DGSEM formulation, it becomes
effective within the modal range up to N , and will reduce the magnitude of the aliasing-
afflicted (upper) modes. While this can prevent a blow-up, two disadvantages occur: i)
The filter acts on the solution modes up to N , i.e. it affects the scale-resolving capa-
bilities of the scheme and ii) stability and accuracy now depend on the additional filter
parameters, which are not known a priori. As opposed to the PDG approach, this method
does not prevent the occurrence of aliasing errors, but removes or controls its effects on
stability, at the cost of resolution quality. That is, this aliasing countermeasure effec-
tively increases the nP P W requirement of the underlying scheme. The main advantage
of this filter approach is its computational efficiency when combined with DGSEM. All
operations remain on the N grid, and the additional filtering of the solution per Runge-
Kutta stage increases the computational time insignificantly (on the order of 2-3% in our
investigations).
Other approaches for removing or avoiding aliasing exist, which are not part of this
work. Two should be briefly mentioned here for completeness: A variant of a PDG
method labeled “line DG” was introduced in [163]. It takes advantage of the dimension-
by-dimension form of the operator and only uses full quadrature precision in the normal
direction of the fluxes, i.e. along the associated coordinate line. Persson was able to
show stability for transitional flow over an airfoil at Re = 20, 000 for a two-dimensional
computation. Since the basic scale producing mechanism and thus aliasing contributor
requires three-dimensionality, stability for realistic turbulence still has to be shown.
A different approach to prevent aliasing is to use split forms of the flux derivative, so
119
called skew-symmetric forms. Skew-symmetry of the non-linear terms implies the con-
servation of kinetic energy [185], a property which can be destroyed by aliasing (see
Section 4.2.1 and [136]). Thus, constructing skew-symmetric discretizations can lead to
kinetic energy stable formulations, which recover the properties of the continuous fluxes
on a discrete level [83, 185].
4.3. Stable Simulation of Under-Resolved Turbulent Flows

In this section, we will apply the de-aliasing strategies presented in Section 4.2.2 to
turbulent flows and discuss their accuracy, stability and efficiency.
4.3.1. Taylor-Green Vortex

The Taylor-Green Vortex problem has been described in Section 2.2 and it has gained
renewed interest as a suitable test model for DNS and LES in the high order community
in recent years, e.g. [33, 60, 200]. We use this flow problem to evaluate the polynomial
de-aliasing and filtering strategies. The discussion in this section follows closely the
publication in [84].
As discussed in Section 4.1, in under-resolved situations, the theoretical approximation
order is not a relevant figure of merit, as the grid spacing remains considerably large.
However, high order methods may still profit from their improved numerical accuracy,
i.e. their lower numerical dispersion and dissipation errors. These benefits come at the
price of reduced inherent damping of instabilities like aliasing errors introduced through
inexact treatment of non-linearities. Figure 4.8 highlights this dilemma for a LES situ-
ation by comparing a DGSEM case for low and high order polynomial degrees, while
keeping the overall number of DOF constant. While the low order method fails to resolve
the physical structures because of excessive dissipation, the high order method follows
the DNS solution closely due to its lower PPW requirement, but succumbs to aliasing
instabilities. In the following, we attempt to avoid or remove this instability mechanism
and to recover a stable high order method.
4.3.1.1. Computational Setup

All our coarse grid computations presented in this section discretize the computational
domain with 43 grid cells with polynomial degree N = 15, resulting in 643 degrees of
freedom. The time step of the computation was approximately 7.7 × 10−4 , resulting
in about 13, 000 time steps for the computation with tend = 10. We found that this
explicit time step was small enough to essentially remove any temporal integration er-
rors. All computations are performed on the maximum number of processors governed
120
0.015 0.015
DNS (Brachet et al) DNS (Brachet et al)
N=1, 32³ cells N=1, 32³ cells
N=15, 4³ cells N=15, 4³ cells
0.01 0.01
-dk/dt
-dk/dt
0.005 0.005
0 0
0 2 4 6 8 10 0 2 4 6 8 10
t t
Figure 4.8.: Plot of the kinetic energy dissipation rates of the Taylor-Green Vortex (Left:
Re = 800, Right: Re = 1600), computed with with 643 DOF using either
a low order method (2nd order, N = 1) or a very high order method (16th
order, N = 15).
by the MPI parallelization (i.e. 64) on the NEC Nehalem cluster at HLRS. Typical wall
clock times for the filtered DGSEM computations were about 10 minutes, and about 1.3
hours for the computations with polynomial de-aliasing. The reference DNS data for the
energy spectra was obtained using 3843 DOF on 512 Nehalem cores. The resolution for
the DNS computations is sufficient to achieve grid converged spectra for wavenumbers
k ≤ 64, i.e. twice the coarse grid Nyquist number.
It should be noted that the resolution of 643 degrees of freedom was chosen to match
the typical resolution for this case published in literature, see e.g. [97]. Especially for
the Re = 800 case, this is a high spatial resolution, much finer than a typical LES set-
ting would require. Still, those computations are often termed LES. For Re = 1600,
Figures 4.12 and 4.13 indicate that the resolution is closer to a realistic LES situation.
4.3.1.2. Stabilization via Filtering
As a first stabilization method, we investigate the stabilizing effect of a modal filter

described in Section 4.2.2.3 for the TGV flow, that is, we combine a DGSEM approach
with M = N with a dissipative mechanism. To examine the effect of the filter type and
shape, we choose two significantly different sets of parameters to cover a broad range
of filter effects: a strong version of the filter (α = 36, s = 16, Nc = 4) and a weak set
121
of values (α = 36, s = 32, Nc = 4). Figure 4.9 depicts the transfer functions for the
polynomial modes for both filters. While we show only those two example filter settings
1
strong filter
weak filter
0.8
0.6
σj
0.4
0.2
0
4 8 12 16
polynomial mode j
Figure 4.9.: Transfer functions for the weak and strong variant of the modal filters for
N = 15.
in this work, we found during our investigations that a broad range of filter parameter
sets leads to stable computations and comparable results for the presented Reynolds
numbers.
Taylor-Green Vortex at Re = 800

Figure 4.10 shows the kinetic energy dissipation rate of our computations and the ref-
erence data, taken from a DNS from [28]. Whereas the unfiltered computation is only
stable up to t ≈ 5 as noted in Figure 4.8, both filtered computations remain stable for
the whole simulation time. Up to t = 4, when the laminar flow features are still well-
resolved, the difference of the computations are negligible and match the DNS result
very closely. Since the polynomial modes affected by the filtering operations are not
filled in this well-resolved region, differences for the two filters only appear at a later
time, when a considerable scale cascade has been established. Examining the quality of
the stabilized computations shows that the strong filter variant introduces too much dis-
sipation, while still following the general trend acceptably. In comparison, the weakly
filtered computation almost coincides with the DNS results.
To further investigate the approximation quality for these filter settings, we compute
the spectral energy distribution for the TGV at the moment of the maximum dissipation
(t = 9), Figure 4.11. In agreement with the behavior for the kinetic energy decay rate,
122
0.015 0.015
N=15, strong filter N=15, strong filter
N=15, weak filter N=15, weak filter
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure 4.10.: Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk dt

for filter-stabilized N = 15 (43 grid cells) computation. The right plot is
zoomed in on the time range of maximum enstrophy production.
the strong filter leads to an inferior approximation of the energy spectrum, even up to the
largest scales, thereby indicating a loss in physics due to over-dissipation. In particular,
the strong filter removes kinetic energy from the high polynomial modes within each
cell, which in turn leads to a blockage of the scale cascade, as described before. This
blockage (due to the missing scales near the cut-off) leads to an overestimation of the
energy in the low modes. The weak filter variant shows very good spectral agreement,
in particular in capturing the low wavenumber range.
The right subplot of Figure 4.11 depicts the contributions to the total kinetic energy dis-
sipation rate. According to Equation (2.19), for a periodic problem, the magnitude of
the kinetic energy dissipation rate − dk
dt
equals the viscous dissipation action (see also
Equation (2.30)). For a DNS situation, these must balance exactly. However, in an
under-resolved setting, both terms can differ significantly. They can be computed in-
dependently from the solution field, and their difference is an indicator for the amount
of numerical dissipation (induced by the approximation error and the filtering). In the
right pane of Figure 4.11, the molecular dissipation ǫ created by the resolved scales and
the numerical dissipation introduced by the spatial discretization are depicted. Since the
strong filter dissipates more resolved scales, the ratio of molecular to numerical dissi-
pation is lower compared to the weak filter, thereby indicating a higher loss of physical
information, as also indicated by the wavenumber spectrum on the left.
123
0.05 0.015
0.04 DNS DNS (Brachet et al)
0.03
0.02
-dk/dt
0.01
0.01
molec.
Diss.
E(k)
0.005
num.
Diss.
0
4 8 12 16 0 2 4 6 8 10
k t
Figure 4.11.: Taylor-Green Vortex at Re = 800. Left: Kinetic energy spectra at t = 9.

Right: Kinetic energy dissipation rate − dk
dt
, the resolved (molecular) dissi-
pation rate and the numerical dissipation rate for filter-stabilized N = 15
(43 grid cells) computation.

By keeping the resolution (643 DOF as in the Re = 800 case) and the filter parameters
fixed but doubling the Reynolds number, we reduce the amount of physical dissipation
and thus the inherent stabilization mechanism through kinetic energy drain. At the same
time, a higher Reynolds number also entails smaller Kolmogorov scales, thereby increas-
ing the degree of under-resolution of the problem towards a more realistic LES situation.
The corresponding results of our filtered coarse resolution computation and the reference
data for the kinetic energy dissipation rate are shown in Figure 4.12. The same general
observations as for the Re = 800 case can also be found for this more difficult flow.
Both filters stabilize the computation, and the weak filter in particular gives results very
close to the DNS solution. For the strong filter, the general shape of the dissipation rate
is still in acceptable agreement, but a shift of the location of the maximum dissipation
rate to earlier times is noticeable. This shift can be found when lowering the Reynolds
number of the problem [28], indicating that the filter effects now dominate over the re-
solved dissipation and mimic a more viscous flow. The plot of the energy spectrum in
Figure 4.13 supports the observations from the Re = 800, with the strong filter results
noticeably deteriorating, even for low wavenumbers. The right subplot in Figure 4.13
shows the effect of the increased Reynolds number on both filters. Since the scale range
of the turbulent flow has increased but the spatial resolution has remained fixed, the frac-
124
0.015 0.015
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure 4.12.: Taylor-Green Vortex at Re = 1600. Kinetic energy dissipation rate − dk dt

0.06 0.015
0.05
DNS DNS (Brachet et al)
0.04 N=15, strong filter N=15, strong filter
0.03 N=15, weak filter N=15, weak filter
0.02
-dk/dt
0.01
0.01
E(k)
molec.
Diss.
0.005
num.
Diss.
0
5 10 15 0 2 4 6 8 10
k t

dt
, the resolved (molecular) dissi-
pation rate and the numerical dissipation rate for filter-stabilized N = 15
(4 grid cells) computation.
125
tion of numerical dissipation to resolved dissipation is higher compared to the Re = 800

test case. For the weak filter, a considerable amount of the scales is still resolved, but for
the strong filter, the overall dissipation rate is strongly influenced by the amount of the
numerical dissipation, which contributes about as much to the overall dissipation rate
as the resolved part. Thus, as obvious from the shift in dissipation rate maximum, the
structure of the numerical dissipation begins to govern the flow physics. This effect will
continue to grow in severity with larger Reynolds numbers and thus the burden on the
“physics capturing” capabilities of the numerical dissipation and the filter will increase.
We have demonstrated in this section the stabilization of DGSEM via a filtering approach
in under-resolved turbulence. Next, we repeat our investigations for the computationally
more costly, but parameter-free and numerically more accurate PDG approach.
4.3.1.3. Stabilization via Polynomial De-Aliasing

In this section, we examine the effects of polynomial de-aliasing (PDG formulation)
for de-aliasing of high order computations with respect to their approximation accuracy.
The advantage of this approach compared to the previous filtering technique lies in the
fact that as long as we capture the non-linearity of the flux function, i.e. use a sufficient
number of quadrature points for the evaluation of the inner products, the stabilization
is parameter-free. Since the TGV is an essentially incompressible problem, a choice of
M = 3/2N would be justified, however, to rule out any weak compressibility effects on
stability, we choose M = 29 and M = 31, i.e. the number of corresponding quadrature
points becomes 30 and 32.

Again, we first investigate the accuracy of this approximation for the test case with the
lower Reynolds number. The kinetic energy dissipation rates are plotted alongside the
reference data in Figure 4.14. Both de-aliased computations are essentially identical,
i.e. no influence of the integration precision is visible, supporting our assumption that
compressibility effects are negligible. Both results match the DNS data almost perfectly,
except for a slight deviation near the maximum of the dissipation rate, where the small
scale action is most prominent. The plot of the spectral energy distribution (Figure 4.15)
corroborates this observation. The match between the DNS spectrum and the coarse
resolution result with only 643 DOF is excellent. The underlying reason for this near
perfect spectral agreement between the PDG scheme and the reference data can be de-
duced from the right pane of Figure 4.15. The de-aliasing mechanism does not introduce
dissipation into the polynomial modes up to N like the filtering approach, but instead
just removes the aliasing error term from them. Thus, the full modes remain available
for the resolution of vortical structures, and the nP P W requirement is not altered by the
126
0.015 0.015
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure 4.14.: Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk

dt
for PDG N = 15 (4 grid cells) computation. The right plot is zoomed in
on the time range of maximum enstrophy production.
0.05 0.015
0.04 DNS DNS (Brachet et al)
0.03
0.02
-dk/dt
0.01 molec.
0.01
Diss.
E(k)
0.005
num.
Diss.
0
4 8 12 16 0 2 4 6 8 10
k t

dt
, the resolved (molecular) dis-
sipation rate and the numerical dissipation rate for PDG N = 15 (4 grid
cells) computation.
127
de-aliasing procedure. This allows us to take advantage of the full potential of this high
order approximation and capture most of the relevant scales possible with only 643 DOF.
Note that the grid Nyquist wavenumber for 64 DOF per direction is 32. Considering a
polynomial approximation, which requires (in the limit) nP P W = π, and around 4 for a
realistic N , up to 16 scales should be resolved with good accuracy by our approach. The
energy spectrum on the left shows an almost perfect agreement up to this wavenumber,
indicating that the associated scales are nearly perfectly resolved. Since those scales
provide a molecular dissipation rate which is already close to the total energy loss, the
overall agreement with the DNS is excellent, and the effect of the numerical dissipation
is negligible in this case. Thus, this example clearly demonstrates that the polynomial
de-aliasing recovers the full resolution capabilities of the scheme and supports a very
low PPW requirement.

Following the reasoning from before, we increase the difficulty of the problem by keep-
ing the resolution fixed, but again doubling the Reynolds number. Figure 4.16 shows
the energy dissipation rates for these simulations. Again, both simulations with different
integration points are basically identical and show very good agreement with the DNS
reference data. The higher scale range of this test case is evident, as the maximum of the
dissipation rate is underestimated due to lack of small scale representation of the flow by
the 643 DOF discretization.
Figure 4.17 shows the detailed view of the energy spectrum, which again follows the
DNS reference closely for both choices of M . In the right hand plot, the ratio of re-
solved to numerical dissipation is still substantial, although – as expected – smaller than
for the Re = 800 case. In comparison with the filtering approach in Figure 4.13 how-
ever, the superior accuracy of PDG in terms of resolution capabilities is evident.
Comparison of De-Aliasing Strategies

We now focus on a comparison of the two de-aliasing techniques that have been exam-
ined: The simple filtering approach and a more complex polynomial de-aliasing method.
Figure 4.18 compares the results of these two high order schemes. In the left plot, we
present the total energy decay rate together with the contribution of the numerical dis-
sipation. Although the dissipation rate matches the DNS reference data well for both
simulations and shows almost negligible quantitative difference, the numerical dissipa-
tion curves reveal the conceptual difference due to the stabilization methods. The purely
dissipative nature of the filter, acting on the solution modes N , manifests itself in a
higher fraction of numerical dissipation compared to the PDG solution. Thus, polyno-
128
0.015 0.015
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure 4.16.: Taylor-Green Vortex at Re = 1600. Kinetic energy dissipation rate − dk

dt
0.06 0.015
0.05
DNS DNS (Brachet et al)
0.04 N=15, Int Points=30 N=15, Int Points=20
0.03 N=15, Int Points=32 N=15, Int Points=30 -dk/dt
N=15, Int Points=32
0.02 molec.
Diss.
0.01
0.01
E(k)
0.005
num.
Diss.
0
5 10 15 0 2 4 6 8 10
k t

dt
, the resolved (molecular) dis-
sipation rate and the numerical dissipation rate for PDG N = 15 (43 grid
cells) computation.
129
mial de-aliasing captures more of the scales present on the 643 DOF grid, which is also
confirmed by the right part of the plot where we compare the energy spectra of the two
computations.
Figure 4.19 allows for a more qualitative evaluation of the approximation properties of
the two approaches. It shows the λ2 contours of the Taylor-Green Vortex at Re = 800
for 4 different discretizations: i) A DNS reference solution with 3843 DOF that shows
no grid influence and smooth contours, indicating a fully resolved field (the lower right),
ii) a O(2) computation with 323 elements, resulting in 643 (lower left), iii) the solu-
tion with the weak filtering from Section 4.3.1.2 (upper left) and iv) the PDG solution
from Section 4.3.1.3 (upper right). The difference between the second and 16th order
computation is remarkable, clearly, the inherent numerical dissipation of the low order
scheme smoothens all but the largest features of the flow field. Both high order results
are in very good agreement with the reference DNS, in particular, most of the large scales
structures are in very close agreement with the DNS. Note that the PDG solution shows
sharper features than the filtered version and is visually slightly closer to the reference;
presumably, this is an effect of the filter dissipation that smoothens the gradients, leading
to a more “washed-out” impression. In terms of computational efficiency, the superior
0.015 0.05
0.04 DNS
DNS (Brachet et al)
0.03
0.02
-dk/dt
0.01
0.01
E(k)
0.005
num.
Diss.
0
0 2 4 6 8 10 4 8 12 16
t k
Figure 4.18.: Taylor-Green Vortex at Re = 800. Comparison of kinetic energy dissipa-

tion and numerical dissipation (left) and energy spectra (right) for N = 15
(43 grid cells) computations with stabilization via polynomial de-aliasing
vs. filtering.
accuracy of the PDG solution comes at an increase in computational cost. In case of

the compressible Navier-Stokes equations, where the number of integration points and
130
Figure 4.19.: Taylor-Green Vortex (Re = 800). Isocontours of λ2 = −1.5 at t = 8.5.

All computations with 643 DOF except DNS. From upper left to lower
right: i) 16th order filtered computation, ii) 16th order computation with
polynomial de-aliasing, iii) 2nd order computation, iv) DNS reference
result with 3843 DOF.
thus the volume operator cost is doubled, the CPU time increases by a factor of eight
compared to the filtered computations. On the other hand, the filter depends on a set of
parameters, whose optimal values are not known a priori, while the PDG approach is
131
parameter-free.
It should be noted that both approaches are limited in their extension to higher Reynolds
number flows. For the filtered method, as the discrepancy between resolved and subgrid
scales increases, the burden on the filter operator rises immensely, shifting its primary
function from a de-aliasing mechanism to a subgrid closure. In other words, the filter
function should no longer be motivated by the suppressing of aliasing, but by provid-
ing physically sound dissipation to account for the subgrid effects. This approach then
becomes an LES with an added dissipation mechanism (it is a philosophical question
whether the filter is more like an explicit closure model or an implicit modification of
the discretization itself). The PDG approach on the other hand does not have a mecha-
nism of introducing dissipation; instead, the inherent approximation errors must act as a
subgrid closure. If the Reynolds number is too large and the resolution too low, this en-
ergy drain will not be sufficient to suppress oscillations, which will eventually become
large enough to violate the physical constraints of the equations. In this way, PDG is
akin to implicit LES closure models.
Comparison with LES Methods

Figure 4.20 displays the results of PDG computations with fixed resolution (643 DOF)
for increasing Reynolds number and compares them with other LES methods. DNS data
from Brachet and Fauconnier [28, 73] obtained with 2563 and 8643 nodes serves as a
benchmark for the quality of the solutions. We compare the PDG results with both an
explicit and an implicit LES closure: the standard dynamic Smagorinsky formulation
and the ALDM (see Section 2.4.3), both results published by Hickel and based on an at
most 4th order FV WENO formulation [97]. Since Hickel used 643 DOF for both his
LES as we do in our computations, a direct comparison of the quality of the results per
invested DOF is possible. It is worth noting that our PDG results are at least comparable
with respect to accuracy to more sophisticated state of the art LES results and show a
considerably closer agreement for the higher Reynolds numbers. Note that while our
result matches the DNS data for Re = 800 almost perfectly, the effect of the fixed
resolution limit is evident from the Re = 1600 and Re = 3000 cases, especially in the
situation of the dominance of the smallest scales (i.e. maximum of the dissipation rate
at t ≈ 9). The missing effects which cannot be captured on the fixed grid therefore must
necessarily be generated with either implicit or explicit subgrid models, if this approach
is to be extended to higher Reynolds number regimes.
132
DNS (Brachet et al, Fauconnier et al) Re = 3000

N=15, Int Points=32
0.015 Dyn. Smagorinsky (Hickel) Re = 1600
ALDM (Hickel)
Re = 800
0.01
-dk/dt
0.005
0
t
Figure 4.20.: Comparison of LES methods: Kinetic energy dissipation rate for Re ∈
{800, 1600, 3000} for PDG computations with N = 15 compared to
DNS and LES reference data.
4.3.2. Flow over a Cylinder at ReD = 3900

In Section 4.3.1.3, we have demonstrated that with proper de-aliasing, non-linear stabil-
ity of DGSEM can be achieved in under-resolved situations. Additionally, exact quadra-
ture of the non-linear terms recovers the scale-resolving capabilities of the high order
discretization, and thus makes this approach attractive for multiscale problems. In the
following section, we show results from [18], where we applied the PDG strategy with
a high order ansatz to the subsonic flow over a circular cylinder at ReD = 3900, a
classical benchmark problem for LES.
4.3.2.1. Computational Setup

The subcritical flow over a cylinder with circular cross section at ReD = 3900, where
D is the diameter, is a well-established test case for the evaluation of implicit and ex-
plicit LES modeling approaches and has been extensively investigated by a large num-
ber of researchers [23, 77, 80, 122, 135, 141, 157]. The boundary layer on the upstream
face of the geometry remains laminar until it sheds periodically, forming a trailing shear
layer. Within this layer, transition to turbulence occurs, which interacts with the turbulent
133
back-flow region in the wake of the body. Due to its rather simple geometry, initial and
boundary conditions yet complex physics, this flow is very popular among researchers to
assess the accuracy of LES methods for incompressible or weakly compressible flows.
A considerable spread in literature exists, often attributed to the sensitivity to the span-
wise domain extensions [135]. Recent DNS results by Lehmkuhl et al. [125] show the
occurrence of very low frequency oscillations within the wake region, which might also
contribute to the partially conflicting published results. In terms of LES, the most com-
monly reported strategy consists of an implicitly filtered approach with an explicitly
added classical Smagorinsky model. Although the shortcomings of this model in lami-
nar and transitional flow situations are known, many publications report the results for
this type of subgrid scale closure, e.g. [23, 31, 80, 122, 134, 157].
Table 4.3 collects the integral quantities of LES and DNS simulations of this flow from
literature.
Author CpBase Str CD Lr /D Scheme LES DOF
Kravchenko et al. [122] -0.94 0.210 1.04 1.35 B-Spline SEM Smag. 1-2 M
Blackburn et al. [23] -0.93 0.218 1.01 1.63 GL-SEM Smag. 1.5 M
Meyer et al. [141] -0.92 0.210 1.05 1.38 FV ALDM 6.0 M
Fröhlich et al. [80] -1.03 0.216 1.08 1.09 FV Smag. 1.4 M
Ouvrard et al. [157] -0.85 0.218 0.99 1.54 FV/FE Smag. 1.5 M
Ouvrard et al. [157] -0.81 0.226 0.93 1.68 FV/FE-VMS Smag 1.5 M
Franke et al. [77] -0.85 0.209 0.98 1.64 FV Yoshizawa 1.2 M
Ma et al. (I) [135] -0.96 0.203 0.96 1.12 h/p-FEM DNS 24 M
Fourier spanw.
Ma et al. (II) [135] -0.84 0.219 0.96 1.59 h/p-FEM DNS 12 M
Fourier spanw.
Current: N = 11 -1.01 0.206 1.10 1.13 PDG - 0.7 M
Current: N = 7 -0.91 0.208 1.00 1.31 PDG - 1.6 M
Table 4.3.: Integral quantities and simulation parameters for ReD = 3900 cylinder
flow, reproduced from [18]. CpBase : pressure coefficient at the downstream
position x = D/2, y = 0, Str: Strouhal number of the lift coefficient
fLif t D/U∞ , CD : drag coefficient, Lr : length of separation bubble.
Spatial Discretization and Statistics Gathering

While our framework supports fully unstructured hexahedral meshes, we choose to dis-
cretize this simple geometry in a structured manner with a fully volume-curved mesh, i.e.
the mapping ~x(ξ)~ from reference to physical space is a polynomial of arbitrary degree,
134
which can be chosen to match the polynomial ansatz for the unknown flow variables to
ensure a consistent geometry discretization. The physical domain is circular with radius
r = 100D in the plane perpendicular to the geometry axis, and extruded along 4D in
spanwise direction. In axial and tangential direction (x − y plane), we choose an expo-
nential stretching in the radial and a Gaussian stretching in the circumferential direction
to cluster the grid points towards the geometry and in the wake.
Referring to the published resolution data collected in Table 4.3 as a reference, we limit
Figure 4.21.: Time- and spanwise-averaged streamwise velocity component and grid.
Left: N = 11 grid with 8 × 8 × 6 elements (circumferential, radial and
spanwise), right: N = 7 grid with 16 × 16 × 12 elements.
the number of degrees of freedom of our computations to a maximum of 1.6 mio. Fol-
lowing our discussion from Section 4.3.1, we choose two high order discretization ap-
proaches to take full advantage of their low nP P W requirements: first, an N = 7 setup,
leading to (N + 1)3 = 512 degrees of freedom per element and a total of 1.57 M DOF
on the associated grid. For an exact integration, we choose M = 11 in accordance
with the integration rule for quadratic non-linear functions, as we set the Mach number
M a = 0.1 for an essentially incompressible flow. The second discretization consists of
an ansatz of N = 11 polynomials with M = 17, giving a total of only 0.66 million DOF
on the associated grid, far below the number of DOF usually reported for this case. A
zoomed-in view of both grids and the resulting time- and spanwise-averaged streamwise
velocity distribution is shown in Figure 4.21. We use the local Lax-Friedrichs flux for
135
the convective fluxes, and Bassi and Rebay’s first method BR1 for the viscous contribu-
tions. The wall boundary conditions are set to isothermal, with a stagnation temperature
of 1.002 T∞ . Periodic boundary conditions are applied in the spanwise direction, while
the remaining ones are weakly enforced Dirichlet free stream conditions. The collection
of the time averaged data was started after the establishment of stable vortex shedding
at 100 T ∗ , with the convective time unit T ∗ defined as D/U∞ (and U∞ denoting the
freestream velocity) and continued for 80 shedding cycles.
4.3.2.2. Results for PDG

Table 4.3 lists the results for the integral flow quantities of the PDG computations. In
general, the computed integral quantities agree well with the published data, while the
match for the N = 11 case is weaker, due to the very low total number of degrees of
freedom for this case and the insufficient wall resolution. Still, even for this severely
under-resolved situation, the results are in reasonable agreement.
Figure 4.22 gives an impression of the importance of the de-aliasing procedure. The left
subplot depicts the instantaneous u-component of the flow velocity. Here, the vortex
street is clearly visible. In the middle plot, the value of a resolution indicator is plot-
ted. This indicator evaluates the slope of the norm of the hierarchical modal degrees of
freedom of the x-momentum in each cell. Yellow and red colors denote a high energy
content in the higher modes, i.e. a likely occurrence of aliasing. This is in agreement
with the physics of this flow, as strong non-linearities in the wake drive the scale cascade.
Note that this cell-wise indicator does not have a high spatial resolution, as the associ-
ated grid cells are large for the high order method. Still, the wake structure can also be
found in the indicator values. The last subplot shows the associated aliasing energy of
the (incompressible) u2 -flux. It is computed as the difference between the projection by
proper quadrature and collocation. 2-h waves appear as a checkerboard pattern in each
cell, indicating the excitation of the highest modes through the aliasing. Also note that
the wake structure is clearly visible in the aliasing error. Comparing first order statistics,
Figure 4.23 shows the time- and spanwise-averaged streamwise velocities hui, normal-
ized by the freestream velocity U∞ , at different locations downstream of the geometry.
We compare the PDG results with experimentally determined values from Parnaudeau
et al. [159], and two LES computations: the implicit approach through an ALDM used
in conjunction with a low order Finite Volume formulation presented in [141] and a high
order B-Spline based Galerkin projection method with explicitly added Smagorinsky
model [122]. The evolution of the wake profile is of particular interest, as the mixing
and recovery of the velocity deficit is governed by viscous action, and therefore by the
combination of resolved and numerical viscosity. A high viscosity generally leads to a
fast interchange of momentum through shear and thus to a less pronounced but overall
136
Figure 4.22.: Instantaneous flow field in the cylinder region. Left: u-velocity, middle:
Modal resolution indicator, right: Difference between projection and col-
location of u2 .
longer wake.
For the N = 7 case, the agreement of the PDG results for the streamwise velocity is
very good for all locations in the wake. In particular, a very close agreement with the
data from the continuous Galerkin approach can be observed. For the N = 11 case,
the match with the published data is again very good in the near wake, but deteriorates
downstream. Together with the data in Table 4.3 which indicates a higher drag coeffi-
cient and reduced length of the separation bubble, this behavior is characteristic of too
little dissipation. The notably reduced spatial resolution and the additional reduction
in numerical dissipation from an N = 7 to an N = 11 approximation result in this
compact, high drag wake. This motivates the addition of subgrid scale dissipation for
higher Reynolds numbers or reduced spatial resolution. More comparisons can be found
in [18], combining the results from Table 4.3 and the detailed discussion therein shows
that the PDG approach again (as in Section 4.3.1) outperforms classical LES approaches
combined with low order schemes at comparable number of DOF.
Table 4.4 summarizes the computational costs in terms of the metrics defined in Sec-
tion 3.5 and the specific time T ∗ . Although the PDG approach incurs a penalty on
efficiency due to the de-aliasing procedure, the wall clock time per T ∗ is on the order
of minutes, which makes a full LES with extensive statistics gathering possible in very
short wall clock times.
In addition to the cylinder flow, computations of a confined periodic hill flow at Reh =
2800 and the transitional flow over a SD7003 airfoil at Rec = 60, 000, both established
benchmark problems for LES, can also be found in [18]. Also for those cases, our PDG
approach achieves equal or better match to experimental and high resolution DNS data
137
1.5 1.5
1 x/d=0.58 1 x/d=0.58
0.5 0.5
0 x/d=1.06 0 x/d=1.06
-0.5 -0.5
<u>/u
<u>/u
-1 x/d=1.54 -1 x/d=1.54
-1.5 -1.5
-2 x/d=2.02 -2 x/d=2.02
-2.5 -2.5
-3 -3
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
y/D y/D
Figure 4.23.: Mean streamwise velocity at different downstream locations in the wake
of the circular cylinder flow at ReD = 3900. Left: N = 11 computation,
Right: N = 7 computation. Solid lines: present DG results, dashed lines:
computational LES results from [141], open squares: computational LES
results from [122], filled circles: experimental results from [159].
than other (typically low order) implicit and explicit LES results from literature, at the
same number of degrees of freedom. It is worth noting that for these two problems,
“incomplete” de-aliasing in combination with a moderate degree N was sufficient for
stability.
Thus, summarizing Section 4.3, we have shown the accuracy and efficiency, both in
terms of invested DOF as well as computing times, of de-aliased DGSEM for under-
resolved turbulent flows. Countering or avoiding aliasing effects has been achieved in
two ways: an additional computationally cheap filter, that fights the aliasing effects on
the solution modes N , thereby increasing the PPW requirement, and a polynomial de-
Case no. DOF no. cells cores ∆t/T ∗ PID [µs] TCP U /T ∗ [h] TW all /T ∗ [min]
N =7 1.572 M 3072 3072 1.3 · 10−3 13 135 2.9

N = 11 0.664 M 384 384 1.7 · 10−3 19 105 16.9
Table 4.4.: Computational cost for the cylinder flow computations.
138
4.4. Interaction of Subgrid Scale Model and Discretization
aliasing on a finer quadrature grid that increases computational cost but retains the orig-
inal scale-resolving capabilities. In terms of accuracy, high order PDG can outperform
other LES strategies for moderate Reynolds numbers.
So far, the numerical strategies presented in Section 4.3, in particular PDG, cannot
strictly be classified as either explicit or implicit LES, since neither an explicitly added
dissipation is introduced, nor is the discretization error adjusted to account for subgrid
effects. Still, the results presented in this section and the comparison with other methods
clearly show the potential of this approach for moderate Reynolds numbers. However,
as discussed above, the lack of additional dissipation limits the stability of PDG, not in
terms of aliasing effects, but in terms of oscillations introduced by the truncation of the
full solution. This induced roughness will eventually lead to a violation of the physical
constraints, e.g. unphysical flow properties.
Figure 4.24 shows the TGV problem for a fixed spatial resolution of 83 elements with
N = 7 and increasing Reynolds number. All computations, unless stated otherwise, use
M + 1 = 12 integration points for the non-linear terms, following the assumption of
essential incompressibility of this flow. For this PDG discretization, the solution remains
stable up to Re = 3000, indicating that the scale resolving capabilities of the scheme to-
gether with the remaining numerical dissipation are sufficient to regularize the solution.
However, examining the solutions for Re = 2000 and Re = 3000 more closely reveals
unphysical oscillations in this integral quantity, which are indicators of the increasing
truncation error. For Re = 4000, this truncation error becomes large enough to cause
a violation of the physical constraints and an unphysical sudden jump in kinetic energy,
denoted by the sharp drop of the dissipation rate at around t = 9. To rule out compress-
ibility effects in the non-linear terms causing aliasing issues, the computation is repeated
with the quadrature rule for cubic non-linearity, but no change in solution is visible, and
the crash occurs at the same time. Increasing the integration precision further results in
the same behavior. Thus, for this configuration, the PDG computation becomes unsta-
ble not due to aliasing, but due to the large truncation error with respect to the missing
scales, which causes the solution on the grid to become oscillatory. Adding an explicit
subgrid closure (a Smagorinksy model) that introduces a dissipative mechanism stabi-
lizes this solution and allows the stable computation of higher Reynolds number flows.
Thus, in general, an extension of the high order approach – regardless of DGSEM or

PDG type – to higher Reynolds number flows will require some form of additional
dissipative mechanism, either explicitly or implicitly. We have shown in Section 4.3,
139
Figure 4.24.: Kinetic energy dissipation rate for the Taylor-Green Vortex with increasing
Reynolds number: 83 Elements, N = 7. M + 1 denotes the number of
quadrature points for the non-linear fluxes, Cs indicates the Smagorinsky
model constant and is 0.0 unless stated otherwise.
that for the under-resolved computations at moderate Reynolds numbers, PDG can pro-
vide stable and accurate results, but requires the computationally expensive polynomial
de-aliasing. We have also shown that a stabilization by a dissipative filtering is also
possible. Since higher Reynolds numbers enforce the need for an additional dissipative
mechanism, a valid question is whether explicit de-aliasing remains necessary, or if the
closure model itself can also account for this. In this section, we investigate this question
by examining the interaction of the discretization scheme, in particular the treatment of
the non-linear terms, and a basic closure. We will focus on an explicit closure approach,
and use the classical Smagorinksy model due to its prevalence in literature for this case.
Thus, our investigations do not aim at explicitly evaluating or optimizing the Smagorin-
sky model parameter for the given discretization choices, but at demonstrating the gen-
eral interaction of eddy-viscosity type subgrid closure and the treatment of non-linear
terms in a high order DG setting.
Depending on whether the baseline scheme has a de-aliasing mechanism, an explicit
140
subgrid scale model typically needs to achieve two goals: guaranteeing numerical sta-
bility, which is not trivial due to the under-resolution, and at the same time modeling
of missing SGS physics. To answer the question posed above, we are investigating two
discretization configurations: A de-aliased PDG scheme, which is aliasing-free and thus
requires no stabilization from the closure model. Thus, we can run this scheme with or
without the model (for moderate Reynolds numbers, as discussed above, as long as the
truncation error does not cause instabilities) and the choice of the model parameters is
decoupled from the numerics. The second configuration choice is the collocated stan-
dard DGSEM, with an added SGS model. In this setting, the closure model has to fulfill
the double role of aliasing suppression and physics modeling. When comparing the two
approaches in terms of accuracy in a LES setting, we first have to establish the stability
of the simulation, and then gauge their accuracy.
As a test case, we have chosen again the cylinder flow described in Section 4.3.2 with
slight modifications.
4.4.1. DNS and LES Setup

As discussed in Section 4.3.2.1, a considerable spread exists in the results reported in
literature, both for experimental [65, 149, 151, 159] and DNS results [65, 70, 135]. Due
to this uncertainty in the reported results and the lack of published reference data for
the chosen Mach number of M a = 0.3 (based on the freestream velocity U∞ and the
freestream speed of sound c∞ ), we have computed a reference DNS for this case, using
our framework presented in Section 3.5.
Figure 4.25 shows a detail of the computational grid for this DNS computation. The
unstructured grid consists of 327680 hexahedral elements, spanning a cylindrical com-
putational domain of radius rcyl = 100 D and spanwise extension ∆z = 8. To account
for the surface curvature, the geometry is represented through a mapping with Ngeo = 4
within the first cell layer on the cylinder. All other mappings are trilinear. The solu-
tion within each element is approximated by a tensor product ansatz of degree N = 4,
resulting in a total of 40.96 mio degrees of freedom.
The Reynolds and Mach number are set to 3900 and 0.3, respectively, and we enforce pe-
riodic boundary conditions on the spanwise domain faces. On the outer domain bound-
aries, we choose weakly enforced Dirichlet freestream conditions for density ρ∞ = 1.0
and the components of the velocity vector U∞ = [1.0, 0, 0]. The freestream pressure
is adjusted to reproduce M a = 0.3. Together with an adiabatic exponent γ = 1.4,
this Mach number yields a stagnation temperature of 1.018 T∞ , which we use as a wall
temperature for the isothermal cylinder wall boundary condition. The collection of the
time-averaged and fluctuating data was started after the establishment of stable vortex
shedding at 100 T ∗ , with the convective time unit T ∗ defined as D/|U∞ | and continued
141
-2
-4
-2 0 2 4
x/D
Figure 4.25.: Grid for the DNS computation, detailed view of cylinder vicinity.
for 100 shedding cycles, characterized by the Strouhal period TSt .

For reasons of comparability, the setup of the LES computations is identical to that of
the DNS described above. In particular, all initial and boundary conditions, averaging
periods and physical domain sizes are identical. We choose an upper limit on the num-
ber of degrees of freedom of 4.6 mio, and discretize the domain with three consecutively
refined grids in a structured, volume-curved manner. Since the total number of DOF is
fixed, the coarse mesh corresponds to the highest local ansatz degree N . Figure 4.26
shows the three LES grids and gives mesh details.
4.4.2. Stability and Accuracy Investigations

We will first focus on the stability aspects of PDG and DGSEM, and relegate the closure
model to a pure stabilization mechanism. Since the flow under consideration is only
weakly compressible at M a = 0.3, the character of the flux function remains quadratic
142
4 4 4
2 2 2
y/D
y/D
y/D
0 0 0
-2 -2 -2
-4 -4 -4
-2 0 2 4 -2 0 2 4 -2 0 2 4
x/D x/D x/D
Figure 4.26.: From left to right: Grids I, II and III for the LES computations, detailed
view. Grid details: I: 16 × 16 × 32 elements (circumferential, radial and
spanwise direction), N = 7, II: 22 × 22 × 44, N = 5, III: 32 × 32 ×
64, N = 3.
instead of cubic, and we choose M = 32 N for PDG in all computations. It should

be noted that this approximation no longer guarantees complete avoidance of aliasing
errors, but we did not encounter any stability issues in our computations. For DGSEM,
additional stabilization through the subgrid scale model is needed if the aliasing error
cannot be bounded by the numerical dissipation of the scheme. We have chosen the
’minimal’ Smagorinksy model constant Cs that resulted in a stable simulation.
Table 4.5 collects the results of our investigations for the three different grids and the as-
sociated N . The DGSEM computations (with M = N ) are denoted by a trailing capital
letter following the polynomial degree (e.g. N5-A), while the PDG computations (with
M = 23 N ) are denoted by a trailing numeral. We define a stable simulation as one that
does not crash for 100 shedding cycles after the initial transient phase.
A detailed discussion of the results can be found in [21]; here, we just focus on the most
important details. As expected, the PDG simulations are stable without an additional
viscosity introduced by the closure. The low order DGSEM computation (N3-A) re-
mains stable even for Cs = 0, due to inherent high dissipation of the N = 3 scheme,
while the medium to high order DGSEM formulation require Cs > 0. Since the numeri-
cal dissipation of DGSEM decreases towards a spectral cut-off behavior with increasing
polynomial degree, it is not surprising that the minimal Cs parameter also increases with
N for N5-A and N7-A.
Two important trends can be extracted from the data in Table 4.5. First, the match
to the DNS reference data improves significantly with increasing N , regardless of the
143
Case Cs min CpBase Str CD CLrms Lr /D θsep [◦ ] θre [◦ ]
DNS - -0.979 0.204 1.082 0.246 1.160 87.2 116.3

N7-1 0.0 -0.995 0.206 1.085 0.252 1.165 81.8 117.1
N7-A 0.16 -0.919 0.200 1.045 0.164 1.413 87.9 124.8
N5-1 0.0 -0.975 0.208 1.072 0.232 1.217 86.0 117.2
N5-A 0.10 -0.878 0.203 1.016 0.127 1.560 87.6 132.7
N3-1 0.0 -0.847 0.216 0.927 0.106 1.569 90.5 123.4
N3-A 0.0 -0.767 0.236 0.800 0.059 1.655 96.8 128.0
Table 4.5.: Simulation details for the cylinder flow with minimal Cs . CpBase : pressure
coefficient at the downstream position x = D/2, y = 0, Str: Strouhal
number of the lift coefficient fLif t D/U∞ , CD : drag coefficient, CL rms :
RMS value of the lift coefficient, Lr : length of separation bubble, θsep , θre :
separation and reattachment location.
choice of PDG or DGSEM. This can be attributed to the general accuracy of high or-
der discretizations in terms of the nP P W requirement, as discussed before. The second
important observation supported by this data is a visible shift towards the DNS refer-
ence data from DGSEM to PDG, with the N7-1 being in very good agreement with the
DNS. All DGSEM results display features of overly dissipative behavior, with case N3-
A showing the worst agreement with the DNS data. As noted above, the N3 cases are
stable for Cs min = 0, however, even here the de-aliasing shows a significant effect on
the accuracy of the solution, as all mean values listed in Table 4.5 shift towards the DNS
data from DGSEM to PDG. Especially noticeable is the change in the predicted Strouhal
frequency, which is governed by the large scale motion. This improvement can be at-
tributed to the increase in accurately represented scales for PDG.
Figure 4.27 depicts the time- and spanwise-averaged streamlines in the wake region. The
DNS result is given as a reference, and the subplots contrast the PDG solution (a, c, e)
with their DGSEM counterparts (b, d, f). For high N (a, b), the effect of the de-aliasing
by subgrid model on the near-wake structure is very pronounced. In particular, N7-1 (a)
closely predicts the secondary separation bubble at the downstream face, which is absent
for the DGSEM result (N7-A (b)). Also, the length of the back-flow region is signifi-
cantly overestimated for this case, due to the strong influence of the subgrid viscosity.
For the medium N case (c, d), the same behavior, although less pronounced, can be ob-
served. In both cases, the back-flow region is elongated compared to the N7 cases, due
to the higher numerical dissipation of the medium N . Although the Cs min required for
144
1.5 1.5
1 1
a
0.5 0.5
y/D
y/D
0 0
-0.5 -0.5
b
-1 -1
-1.5 -1.5
-1 -0.5 0 0.5 1 1.5 2 -1 -0.5 0 0.5 1 1.5 2
x/D x/D
1.5 1.5
1 1
c e
0.5 0.5
y/D
y/D
0 0
-0.5 -0.5
d f
-1 -1
-1.5 -1.5
-1 -0.5 0 0.5 1 1.5 2 -1 -0.5 0 0.5 1 1.5 2
x/D x/D
Figure 4.27.: Time- and spanwise-averaged streamlines. Upper left plot: DNS. Rest: a)
N7-1, b) N7-A, c) N5-1, d) N5-A, e) N3-1, f) N3-A
stabilization is lower, this increased numerical dissipation and the associated loss in ac-
curately resolved waves results in an overall increased dissipation. For low N (e, f) with
Cs min = 0, this numerical dissipation is the governing factor on the solution accuracy.
Nonetheless, it is interesting to note that the PDG solution (e) shows a better resolution
of the region close to the secondary separation bubble and a slightly shorter back-flow
region, a result that emphasizes proper de-aliasing even for low order schemes.
Figure 4.28 compares the velocity fluctuations hu′ u′ i at different positions in the wake,
which contribute strongly to the turbulent kinetic energy and are thus a measure of the
145
quality of the turbulence representation. The general match to the DNS data is greatly
improved for PDG, with only N3-1 showing a significant difference to the reference data.
For the DGSEM results, the agreement in the near wake region is poor for all cases, as
the scale resolution of the scheme is strongly diminished by the aliasing suppression.
Further downstream, where the flow physics are less dominated by non-linear small
scale effects and more by the mixing through the shear layer gradients, the agreement
improves. It is again worth noting that the low order PDG (N3-1) shows a significant
improvement over the associated DGSEM results (N3-A). In support of these findings,
0.2 0.2
x/D=0.58 x/D=0.58
0 0
-0.2 -0.2
x/D=1.06 x/D=1.06
<u’u’>/U
<u’u’>/U
-0.4 -0.4
x/D=1.54 x/D=1.54
-0.6 -0.6
-0.8 -0.8
x/D=2.02 x/D=2.02
-1 0 1 -1 0 1
y/D y/D
Figure 4.28.: Time- and spanwise averaged velocity fluctuations hu′ u′ i in the cylinder
wake. Line with squares: DNS results;
Left: DGSEM according to Table 4.5. Dashed line: N7-A; dotted line:
N5-A; dash-dotted line: N3-A;
right: PDG according to Table 4.5. Dashed line: N7-1; dotted line: N5-1;
dash-dotted line: N3-1.
Figure 4.29 depicts the one-dimensional spectra of the streamwise velocity component,
recorded at x/D = 3.0, y/D = 0.25 and averaged over 50 equispaced spanwise sta-
tions. All frequencies are normalized by the Strouhal frequency of the DNS reference
result. The DNS reference spectrum is shown for comparison, the dashed line denotes
Kolmogorov’s k−5/3 law. About 100,000 samples were collected over a dimension-
less time interval T U∞ /D ≈ 500 (corresponding to about 100 shedding cycles). The
146
samples were then interpolated on an equidistant temporal grid, using 4th order interpo-
lation, on which a FFT was then conducted. The spectra have been shifted vertically for
easier distinction.
As a general observation, the dominant shedding frequency and its first two harmonics
Figure 4.29.: One-dimensional spectra of the streamwise velocity component in the

wake at x/D = 3.0, y/D = 0.25. All frequencies normalized by the
Strouhal frequency of the reference DNS (see Table 4.5). a) N7-1, b) N7-
A, c) N5-1, d) N5-A, e) N3-1, f) N3-A; dashed line denotes Kolmogorov’s
k−5/3 law.
are captured in all cases. The associated frequencies are over-predicted for cases N3-1
and N3-A (e, f), which causes the shift to higher frequencies for the chosen normaliza-
tion. Also, an inertial subrange closely following the k−5/3 slope is observable for cases
N7-1, N7-A, N5-1 and N5-A (a, b, c, d), with the best agreement to the DNS for N7-1.
For N3-1 and N3-A (e, f), the extent of the inertial range is significantly reduced. Most
interesting is the influence of PDG versus DGSEM on the high frequency range. For
high N (a,b), a significant improvement in scale resolution can be recovered for PDG, as
the N7-1 spectrum follows the DNS spectrum more closely up to a significantly higher
frequency. Although this trend is also observed for medium N (c, d), the gain is less
pronounced, and it is completely covered by noise for low N (e, f). The reason for this
147
behavior is the interaction of the de-aliasing and the dissipative error of the discretiza-
tion as a function of N : Following the 3/2 rule, a single evaluation of the non-linear
inner product in DGSEM introduces an aliasing error into the upper 1/3 of the polyno-
mial modes. The error of course then interacts non-linearly with the solution and finally
manifests itself across all wavenumbers. Therefore, de-aliasing allows the recovery of
this 1/3 of the polynomial modes for the accurate representation of the solution. As
this upper third is less affected by the approximation errors for high N , the effects of
polynomial de-aliasing become more pronounced.
So far, we have only considered results for a minimal Cs , dictated by the aliasing stabi-
lization for DGSEM. In the following, we extend this discussion by choosing the model
constant for both PDG and DGSEM to find the best approach for this LES computation.
To this end, we have conducted extensive numerical experiments for varying N and Cs ,
both for PDG and DGSEM. Table B.1 in Appendix B summarizes the details for all com-
putations. While the choice of Cs is arbitrary for PDG, its minimum value for DGSEM
is given by the stability discussion above.
Summarizing the results from the investigations and Table B.1, a clear trend can be iden-
tified: For high N , increasing the model parameter for both PDG and DGSEM has the
same overall effects: A decrease in the base pressure, drag coefficient and lift fluctua-
tions, along with a lengthening of the wake region, all indicative of an increase in vis-
cous actions in the flow. Thus, for both discretization options, any choice of Cs leads to
a weaker agreement with the DNS reference. In the PDG case, the basic de-aliased case
(N7-1) is already very close to the DNS result due to the moderate Reynolds number and
high resolution capabilities of the scheme, and additional viscosity is detrimental. For
DGSEM, the lower limit of Cs (N7-A) is already too high from a modeling standpoint,
i.e. the suppression of aliasing requires more damping than the truncation error. Thus,
the stabilization of the aliasing errors dominates any physically motivated considerations
for selecting this model parameter, i.e. the collocative discretization suffers a penalty in
terms of accuracy compared to PDG.
For the medium N case, the behavior is very similar to the one described above. It should
be noted that the artificial dissipation introduced by the lower degree N becomes more
pronounced, resulting in a loss in resolved scales and in turn also in less severe aliasing
issues, allowing a wider range of possible model parameters for DGSEM. The optimal
Cs for best match with the DNS reference data for DGSEM is not Cs min as for N = 7,
but slightly higher, although the improvement is not very pronounced (N5-B). Since the
loss in scale resolution for the N = 5 scheme compared to the N = 7 scheme results
in a higher truncation error and thus in a stronger cascade blockage (energy cannot be
transferred downrange due to lack of small scale interaction partners), a build-up of en-
ergy occurs in the large scales, i.e. the large scale energy increase leads to an oscillatory
148
behavior. For high Reynolds numbers or low resolutions, the numerical dissipation of
the scheme may not be sufficient to remove the excess energy. Introducing additional
viscosity on these resolved scales can drain this energy and regularize the truncated so-
lution. This interplay of numerical and model dissipation is dependent on the properties
of the scheme, the resolution, the model and the flow, which explains why the optimal
choice of Cs depends on all of those factors [143]. Thus, although counter-intuitive, a
discretization with a stronger artificial dissipation will benefit from a higher value for
Cs . In general, according to Table B.1, PDG results show better agreement with the
DNS than the DGSEM. It is remarkable that the optimal PDG results for N = 5 (N5-1)
shows better agreement with the reference data than the DGSEM N = 7 (N7-A) solu-
tion, indicating again that the excessive damping to counter aliasing essentially negates
the scale resolving advantages of the high order discretization.
For the low order cases N = 3, the choice of Cs is free, as the numerical dissipation of
the low order keeps the aliasing effects bounded. Due to the low scale resolving capa-
bilities and high truncation error, these computations gain from an increase in Cs , with
the PDG variants outperforming the DGSEM variants. Even for the same Cs (compare
e.g. N3-D and N3-4), the PDG results are in better agreement with the DNS, showing
the positive effect of the de-aliasing on the resolved scales.
Figure 4.30 gives a visual impression of the velocity fluctuations hu′ u′ i in the wake
region, for the PDG results with and without Smagorinksy model and DGSEM with
Smagorinksy model. Shown are the results for the “best” model constant with respect
to the DNS, see Table B.1. Note that the PDG results always show better small scale
resolution, and that high order discretizations are superior to low order ones. Also,
while stabilization of collocation methods with a Smagorinsky closure is possible, the
associated loss in accuracy effectively increases the nP P W requirement. This becomes
obvious when comparing the N = 7 DGSEM (last plot, second row, N7-A) with the
N = 5 PDG (first plot, third row, N5-1). The PDG solution, although theoretically with
a higher nP P W requirement, resolves more scales than its DGSEM counterpart. More
details on the results and a fuller discussion can be found in [21].
4.4.3. Discussion
For high Reynolds numbers, some form of dissipative mechanism becomes unavoidable
for LES, and the addition of an explicit closure model is the most common approach. For
high order methods, the choice of the treatment of the non-linear fluxes plays a major
role in solution accuracy and stability and in particular interacts with the closure model.
Not only are subgrid scale models evaluated in terms of the flow solution and are thus di-
rectly influenced by approximation and aliasing errors, they also act on the same range of
resolved scales. Polynomial de-aliasing removes a non-linear source of error that influ-
149
Figure 4.30.: Time- and spanwise-averaged velocity fluctuations hu′ u′ i in the wake
2
region, normalized by U∞ . Top row: DNS reference. First column:
PDG results, second column: PDG with Smagorinsky model, last column:
DGSEM with Smagorinsky model; second row: N7-1, N7-2, N7-A, third
row: N5-1, N5-2, N5-B, fourth row: N3-1, N3-4, N3-D.
ences this interaction, and thus makes a more meaningful evaluation of closure models
possible. Particularly for high polynomial degrees with low numerical errors, this ap-
proach gives an (approximate) decoupling of numerics and model influences and allows
an independent selection of both, hence leading to superior solution accuracy compared
to the collocation variant. For LES with low order polynomial degrees, the numerical
damping may prevent the aliasing effects from causing stability problems, but their ef-
fect on solution accuracy in both “no model” LES and explicit subgrid scale model LES
can be pronounced.
Returning to the question posed at the beginning of this section, whether de-aliasing is
necessary and beneficial when an explicit dissipation mechanism is needed anyway to
counter the truncation error, a three-fold answer can be given:
150
4.5. Flux Functions in Under-resolved Flows
i) As summarized above, we have shown the advantages in terms of accuracy for a de-
aliased approach. While our investigations cannot be generalized (to other SGS closure
strategies), they indicate that the interaction of aliasing and closure can be very strong,
non-linear and hard to control a priori, even for low order schemes. Thus, removing
the source of instability and error and providing a baseline “neutrally stable” scheme is
important.
ii) Both the development and the evaluation of physically motivated subgrid scale clo-
sures require numerical schemes with minimal interaction of approximation errors and
the modeling terms. High order DG methods with their inherent low dissipation and
dispersion properties are attractive candidates for this approach, however, without de-
aliasing, these favorable properties cannot be recovered for efficient LES simulations.
iii) While the first two answers are strongly in favor of proper de-aliasing, one possible
restriction has to be made. The considered test case was at a moderate Reynolds number,
and the spatial resolution, while coarse, did capture most of the relevant scales, at least
for high N . If the gap between resolution and Kolmogorov scales is increased however,
the magnitude of the truncation errors can become larger than the aliasing error, i.e. the
necessary dissipation to account for the missing scales can become “large” enough to
make de-aliasing by closure attractive, as it comes “for free”. This however depends on
the discretization, the resolution and the closure, and has to be investigated further.
While we have focused on a step towards high order DG for LES with explicit modeling
in the previous section, we now investigate a discretization aspect that could become
important in an implicit LES approach: the choice of the numerical flux functions. For
the Discontinuous Galerkin formulation in Equation (3.29), if all inner products are eval-
uated with the appropriate integration precision (the PDG approach), the discretization
itself becomes “analytical” in the sense that it is no longer dependent on the geomet-
rical features of the discretization choices like basis functions, element shapes or node
types. The only remaining free parameters are the numerical inviscid and viscous flux
functions Gn∗ and Hn ∗
. Therefore, their choice determines the consistency, accuracy and
stability of the analytical DG method. The task of the flux function is to couple two in-
terfaces with non-unique solutions (and thus non-unique normal fluxes) by determining
a suitable unique flux. One method of achieving this is the concept of defining this flux
as an (approximate) solution to an initial value problem with constant states (a Riemann
problem), an idea developed in the FV method community [196]. While an exact so-
lution to the Riemann problem exists, this so-called Godunov flux function is often too
costly to evaluate. Instead, approximations of various degree of physical consistency
151
exist, leading to a large class of flux functions for the convective and diffusive fluxes.
4.5.1. Flux Choices

4.5.1.1. Inviscid Fluxes
For the DG method, the choice of the flux function for the Euler equations has been
investigated by Qui and Qui et al. [169, 170] for smooth and shock-dominated prob-
lems. They found that for smooth, well-resolved problems, the flux functions differ only
marginally in terms of accuracy and thus the most computationally effective flux is a
good candidate. This finding is not surprising, as the number of flux points in a three-
dimensional problem scales quadratically with a given reference length of an element,
while the volume fluxes scale cubically. Or, considering the strong form of the DG
formulation, the influence of the fluxes is just through a surface penalty, which is very
small for well-resolved problems. Hence, for higher N and well resolved situations, the
influence of the flux functions becomes negligible. Qiu et al. found that for problems
with discontinuities, the flux functions with higher physical approximation quality (more
waves are considered) clearly outperform more dissipative variants, also a result that can
be expected from the experiences of the FV community. Kesserwani et al. [110] also
examined these aspects for the dam-break problem of the Shallow Water Equations with
similar findings, while Wheatley et al. [203] performed similar numerical experiments
with DG for the Magneto-Hydrodynamic (MHD) equations.
The focus of our investigations here is shifted towards the behavior of these flux func-
tions in an under-resolved turbulent setting, which differs significantly from the smooth
problems or those with strong, but localized gradients investigated by the researchers
mentioned above. In our investigations, we focus on two representatives for the convec-
tive fluxes, namely the Local Lax-Friedrichs flux (LLF) and Roe’s approximate Riemann-
solver. Our choice was governed by the difference in dissipation introduced by these
fluxes and their widespread use in the DG community.
Lax-Friedrichs Flux Function

The Lax-Friedrichs (LF) flux and its local variant (LLF or Rusanov flux) are the simplest
flux functions, disregarding all but the fastest wave and thus introducing the highest
amount of numerical viscosity, see e.g. [196]. Due to its simplicity, robustness and
computational efficiency, the LLF is widely used by the DG community. The convective
numerical flux in (3.29) is approximated as
1 + 1
Gn∗ = − β λadv max U − U− + G~n (U + ) + G~n (U − ) , (4.21)
|2 {z } |2 {z }
Penalty Central flux
152
where β is a real number which allows control over the amount of numerical viscosity
by penalizing the jump in the solution at the interface (with β = 1 being the classical
LF definition), the superscripts ± denote the values from the neighbor and local element
and λadv
max corresponds to the maximum eigenvalue of the Euler flux matrix as
λadv
max := max (|~
v | + c). (4.22)
U + ,U −
√
Here, c denotes the speed of sound waves computed as c := κRT and ~v is the velocity
vector. For the local LF variant considered in this work, the value of λadv
max is computed
from the local flow field.
Roe’s Approximate Riemann Solver

Another favorite flux function in the DG community (see e.g. [198]) is the approximate
~
dG
Riemann solver due to Roe, see e.g. [196], where the exact Euler flux Jacobian A = dU
is replaced by a linearization Ã about an average (Roe) state. The underlying system
becomes linear with constant coefficients, i.e. instead of the exact Riemann problem, an
approximation is generated, which is then solved exactly. The numerical flux is approx-
imated as
1 X
m
Gn∗ = − β ~ (i) + 1 G~n (U + ) + G~n (U − ) ,
α̃i |λ̃i |K (4.23)
2 i=1 |2 {z }
| {z } Central flux
Penalty
where the ˜ denotes the evaluation at the Roe state, m stands for number of eigenvalues
λi (U + , U − ) of Ã, α̃i (U + , U − ) denote the wave strengths and K~(i) (U + , U − ) are the
corresponding right eigenvectors. There are two different approaches to finding the in-
termediate state and from there the wave strengths and eigenvectors which are detailed
in [196], in our approach, we use the classical Roe formulation. We have again intro-
duced the parameter β as in Equation (4.21), which allows control over the amount of
numerical viscosity.
Note that the structure of both fluxes is similar: a dissipation-free central average of the
fluxes, stabilized by a penalty term that introduces some for of numerical diffusion based
on the height of the jump and on eigenvalues that describe the strength and direction of
information interchange.
4.5.1.2. Viscous Fluxes

For the viscous fluxes, a large choice of formulations exists which, like the convective
fluxes, lead to different stability and accuracy properties. An overview of the available
153
options for the Laplace equation is e.g. given in [9]. The parameter αvisc in Equa-
tion (3.35) allows a switch between two commonly used variants:
• αvisc = 12 leads to the first variant of Bassi and Rebay [12] by using the arith-
metic mean for both fluxes. This flux is stable for parabolic problems, while it is
known that it becomes unstable for purely elliptic cases.
• αvisc = 0 or αvisc = 1 lead to the local DG variant (LDG) by Cockburn and

Shu [49].
4.5.2. Influence of the Numerical Fluxes

In this section, we briefly report the influence of the numerical choice of the numerical
flux function on the TGV and cylinder computations in a LES setting, as described in
Sections 4.3.1 and 4.3.2. As both problems are still advection dominated at the given
resolution, we focus on the inviscid fluxes.
4.5.2.1. Taylor-Green Vortex

Following the setup in Section 4.3.1, we have fixed the total number of DOF to 643 for
all our LES simulations, but have again computed different combinations of grid cells
and degree N . This means that for the cases N = 15, N = 7, N = 3 and N = 1, we
use 43 , 83 , 163 and 323 grid cells, respectively. In Figure 4.31, we show the comparison
of kinetic energy dissipation rate for the two different convective fluxes described in
Section 4.5.1.1. As expected, the polynomial degree has a strong impact on the influence
of the flux function on the solution. Although all computations have nominally the
same spatial resolution, the difference between the choices is very low for high N and
increases subsequently for lower N . Comparing all results from Figure 4.31, we see a
clear progression from high order to low order in flux function impact. As the number
of cell interfaces increases for a given overall resolution and the approximative strength
of the local polynomial ansatz decreases, the influence of the flux function becomes
more pronounced. Furthermore, it can be observed that, as expected, the kinetic energy
dissipation is lower for the Roe flux function and thus closer to the DNS results. In
contrast to the convective fluxes, the impact of the viscous flux function seems negligible
in this advection-dominated case. For the low order variants, the numerical dissipation is
totally governed by the dissipation mechanism of the Euler fluxes. Only for the case of
very high polynomial degree, N = 15, where the influence of the Euler flux is low, we
observe a measurable impact of the viscous flux choice. But as the results in Figure 4.32
clearly show, the influence on the overall dissipation behavior even in this case remains
negligible.
154
Figure 4.31.: Kinetic energy dissipation rate for the Taylor-Green Vortex at Re = 1600:
Comparison of convective numerical flux function influence for different
polynomial degrees N with same total no. of DOF (643 ). Square symbol
denotes the reference DNS solution [28], dashed line denotes the LLF flux
result, the solid line denotes the Roe flux result. Plot reproduced from [20].
As expected, in an under-resolved situation, where the inter-cell jump terms become non-
negligible, the choice of the flux functions can play a role in solution quality. As long as
the spatial resolution in terms of DOF is high enough, for a high order discretization, the
scale-resolving capabilities will keep the jumps small, but this mechanism will clearly
fail for higher Reynolds number. Thus, modeling the subgrid effects in an implicit LES
approach through an appropriate flux function is worth investigating. To demonstrate
155
Figure 4.32.: Kinetic energy dissipation rate for the Taylor-Green Vortex at Re = 3000:
Comparison of LDG flux and BR1 flux with DNS, 43 elements, polyno-
mial degree N = 15. Right: Detailed view. Plot reproduced from [20].
Figure 4.33.: Optimization of the penalty parameters α1 and α2 for the kinetic energy
dissipation rate. Shown is the L2 deviation between DNS and LES results
as a function of α1 and α2 .
the possible benefits and effectiveness of this approach, we have modified the parameter
β in Equations (4.21) and (4.23) that penalizes the jump term and thereby regulates
156
the numerical dissipation. This is a simple ad-hoc version of discretization tuning, but
we found a significant increase in solution quality in terms of the energy dissipation
rate. Figure 4.34 left shows the results for the Re = 800 simulation with the classical
(β = 1.0) Roe flux and a modification (β = 0.025) with lowered dissipation. The
overall agreement with the DNS results improves significantly for the modified version,
and the physical structure of the dissipation rate reappears in the solution. In the right
subplot of Figure 4.34, the modification has been extended to a linear combination of
the penalty terms, yielding a combined flux function with free parameters α1 and α2 :
+ Xm
1 − 1 ~ (i) + 1 G~n (U + ) + G~n (U − ) .
Gn∗ = − α1 λadv max U − U − α 2 α̃i |λ̃i |K
2 2 i=1
2
(4.24)
Optimization of α1 and α2 for various resolution and Reynolds numbers showed a po-
tential of implicit LES modeling through the modification of the penalty dissipation.
The kinetic energy dissipation of the DNS served as the target function, and the L2 error
between this reference and the computed dissipation rate was optimized with a gradi-
ent based algorithm in parameter space. Figure 4.34 right shows the result of such an
optimization. While a clear improvement over the standard flux formulations exists, sta-
bility issues in the laminar region were encountered. In addition, tuning the flux function
for the full range of laminar, transitional, non-decaying and decaying turbulence is not
optimal, as it should be conducted for the different physical regimes separately. Thus,
while this preliminary study shows the importance of the flux function in under-resolved
turbulence and the possibility of implicit modeling through it, the investigations should
be repeated with a more efficient optimization algorithm and for homogeneous isotropic
decaying turbulence.
4.5.2.2. Cylinder at ReD = 3900
To examine whether the choice of the inviscid flux functions influences the cylinder
results from Section 4.3.2 and thereby acts as an implicit LES model, we have repeated
the computations for the Roe flux function. Table 4.6 compares the results for the integral
values for the originally used LLF and the Roe flux function, and Figure 4.35 compares
the pressure coefficient and crossflow velocity in the wake.
While the effects on the integral quantities are small (except for the ratio Lr /D), the
wake structure reveals a considerable difference, with the Roe flux being in better agree-
ment with the experimental results. In general, this flux formulation generates a longer
wake, indicative of a higher numerical viscosity (see discussion in Section 4.3.2).
157
Figure 4.34.: Kinetic energy dissipation rate for the Taylor-Green Vortex: Left: Re =
800: Comparison of classical Roe flux and modified Roe flux with DNS.
Right: Re = 1600: Comparison of modified combined flux and standard
fluxes with DNS.
In this section, we have demonstrated that the choice of the flux functions can play an
important role, as the induced dissipation acts as an implicit LES model. Depending on
the choice of the polynomial ansatz degree N , the character of DG can be closer to a
local, cell-based scheme (a first order DG with N = 0 is analytically identical to a first
order FV scheme) or a quasi-global, spectral scheme coupled weakly by penalty terms.
Thus, in particular for low order DG, the choice of the flux function is important in an
under-resolved setting.
We have shown preliminary results for the choice of the inviscid and viscous fluxes and
have demonstrated that adjusting the dissipation introduced by the flux function can act
as a simple LES closure. In the future, a more thorough analysis of the influence of the
flux functions, their behavior in a LES situation and implicit modeling will be necessary.
158
Case CpBase Str CD Lr /D Scheme LES DOF
LLF: N = 11 -1.01 0.206 1.10 1.13 PDG - 0.7 M

Roe: N = 11 -1.00 0.212 1.09 1.26 PDG - 0.7 M
LLF: N =7 -0.91 0.208 1.00 1.31 PDG - 1.6 M
Roe: N =7 -0.90 0.208 1.02 1.50 PDG - 1.6 M
Table 4.6.: Integral quantities and simulation parameters for ReD = 3900 cylinder flow,
as a function of the Euler flux function. CpBase : pressure coefficient at
the downstream position x = D/2, y = 0, Str: Strouhal number of the
lift coefficient fLif t D/u∞ , CD : drag coefficient, Lr : length of separation
bubble.
1 0
x/d=0.58
0.8 DG N=7 LF
DG N=11 LF
0.6 DG N=7 Roe
DG N=11 Roe -0.5
0.4 Ma et al. DNS
Norberg, RE=4020, Exp.
0.2 x/d=1.06
-1
<v>/u
0
Cp
-0.2
-0.4 -1.5 x/d=1.54
-0.6
-0.8
-2
-1
x/d=2.02
-1.2
0 50 100 150 -2.5
-3 -2 -1 0 1 2 3
y/D
Figure 4.35.: Comparison of Cylinder LES results for Roe and Rusanov-flux for com-
putation described in Section 4.3.2. Left: Pressure coefficient along ge-
ometry. Filled circles denote experimental data for Re = 5000 of Son &
Hanratty for vorticity and for Re = 4020 of Norberg for pressure coeffi-
cient (taken from [141]). Open squares denote DNS results [135]. Right:
Mean crossflow velocity at different downstream locations in the wake of
the circular cylinder flow at ReD = 3900, N = 11 computation. Solid
black lines: DGSEM with Rusanov flux. Solid blue lines: DGSEM with
Roe flux. Dashed lines: computational LES results from [141]. Open
squares: computational LES results from [122]. Filled circles: experi-
mental results from [159].
159
4.6. LES with DG Methods

In this section, we will briefly summarize the current state of the art regarding LES
with DG methods and then outline a more consistent strategy for current application and
future research based on our results from the previous sections.
4.6.1. State of the Art

Along with the first DNS with DG, Collis [51] in 2002 was among the first to use DG
schemes in a LES setting. He applied a Smagorinsky eddy viscosity model to the small
scales via the Variational Multiscale (VMS) method [52]. Sengupta et al. [179] com-
bined the element-wise filtering approach of Blackburn and Schmidt [24] with a DG
spectral element method (DGSEM) and performed LES with a dynamic Smagorinsky
model. More recently, Uranga et al. [198] used an N = 3 DG formulation for the simu-
lation of a transitional flow over a SD7003 airfoil, taking advantage of the numerical dis-
sipation to account for the damping action of the unresolved scales in an implicit LES ap-
proach. A similar investigation was carried out by Carton de Wiart and Hillewaert [35].
The accuracy of DG methods for high Reynolds number vortical flows for N = 3 was
investigated in [36]. For higher orders (N = 7, ..., 15), Gassner and Beck [84] and the
discussion in Section 4.3 showed that de-aliased DG schemes applied to under-resolved
simulations of isotropic turbulence compare very well with low order FV schemes with
explicit or implicit subgrid scale models. In this case, the number of degrees of freedom
was chosen as in the FV approach, and no additional dissipation was added explicitly or
implicitly to the DG formulation. Instead, the high resolution capabilities of high order
discretizations resolve a much larger bandwidth of turbulent scales, thereby allowing the
viscous action of the Navier-Stokes operator to remove most of the energy without the
need for an additional model. In terms of computational efficiency, it was shown in [200]
that high order DG schemes can compete with FD and FV formulations in a LES setting
of this flow problem. In [18], this approach was recently extended to transitional and
turbulent wall-bounded flows, demonstrating again that for moderate Reynolds numbers
and typical LES resolutions, no additional modeling is required, when de-aliased high
order DG formulations are used. Also, the importance of de-aliasing and its interac-
tion with an explicit subgrid model was investigated in [21]. Van der Boss et al. [199]
conducted a similar analysis for homogeneous isotropic turbulence and found that the
numerical dissipation of the scheme and that of an explicit Smagorinksy model inter-
act strongly, and that the optimal model constant strongly depends on the order of the
scheme. In particular, they found that an implicit LES approach gave inferior results than
one with an added Smagorinksy dissipation, yet, their investigations were limited to low
order DG (N ≤ 2). As discussed in Section 4.4.2, the low order DG discretizations
160
4.6. LES with DG Methods
lack the scale-resolving capabilities of their high order counterparts and need additional
dissipation to account for the lack of resolved, dissipative scales.
4.6.2. Perspectives for High Order LES

The current state of the art in LES with DG methods is the use of moderately high
approximation orders in an implicit “no model” LES approach for moderate Reynolds
numbers. Some publications combining DG with an explicit LES model (typically a
version of the standard Smagorinsky model) exist, but the baseline scheme in these in-
vestigations is of low to moderate order, thereby not taking advantage of the spectral
character of high order DG.
In general, the solution quality of an under-resolved computation is bounded by the
available number of degrees of freedom, i.e. the theoretically smallest resolvable scale
(Nyquist theorem). The properties of the numerical discretization determine the quality
of these degrees of freedom, i.e. the fraction of the complete information represented
on this grid. High order discretizations allow the recovery of a significant part of the
theoretically available physical information, thereby limiting the “numerical truncation
error” (as opposed to the grid or DOF truncation error) compared to low order schemes.
The associated reduction in numerical dissipation comes at a cost: De-aliasing becomes
necessary not only for stability, but also for accuracy. Since the use of proper polyno-
mial de-aliasing prevents numerical instabilities through wrongfully represented energy,
the only issue remaining is the lack of subgrid scales and their associated dissipation,
as a direct consequence of under-resolution. Thus a mechanism needs to be present that
accounts for the energy transferred to high wavenumbers, or in other words, mimics the
dissipation of the unresolved scales, while leaving large scales unaffected. In a classical
LES setting, this role is played by the (explicitly added or implicitly introduced) subgrid
modeling terms. In high order discretizations, a large bandwidth of scales can be re-
solved within each cell. These scales can carry a non-negligible physical dissipation, and
thereby provide a significant part of the physical regularization through the viscous op-
erator in the Navier-Stokes equations. The remaining numerical dissipation acting near
the cut-off wavenumber can effectively serve as a source for the non-resolvable cascaded
energy. Thus, a simple principle applies for high order DG for moderate Reynolds num-
bers and typical LES resolution: The accurate resolution of the dissipative mechanisms
is the best available “model”.
Clearly, this approach reaches its limits when the scale range between the smallest phys-
ical scale and cut-off scale increases, i.e. when the truncation error becomes larger, or
when the small scale physics are no longer dominantly dissipative. In that situation,
additional dissipation (or closures modeling other subgrid effects) becomes necessary.
For both implicit and explicit modeling approaches, high order DG is particularly well-
161
suited:
Implicit modeling implies the manipulation of the approximation errors to serve as a
subgrid model. In DG, the simplest way to influence the discretization error is through
local p-adaptation and through the choice of the flux function. Both mechanisms allow
direct control over the numerical dissipation without additional modeling efforts. An in-
teresting approach, directly related to the high wavenumber range resolved within each
element, is the implicit LES modeling through controlled, adaptive de-aliasing. As dis-
cussed in Section 4.2, the aliasing errors enhance the fluctuation energy in the resolved
scales near the cut-off, thereby also enhancing the viscous dissipation through those
scales. Since the mechanism is self-feeding, de-aliasing becomes necessary to avoid an
excessive build-up for stability. However, to ensure stability, this de-aliasing does not
have to become active at every time step. Instead, a locally and temporally adaptive
de-aliasing mechanism that relies on intelligent indicators can allow a certain amount of
aliasing energy to accumulate and then dissipate, thereby enhancing the overall dissipa-
tion and extending the Reynolds number range compared to a “no model” LES.
Still, without an additional explicit source of dissipation or a filtering operation, the im-
plicit strategy outlined above will not be suitable for very high Reynolds numbers. Thus,
some form of explicit modeling in combination with high order DG will be necessary.
Here again, the high local resolution within a cell can be exploited in three ways: i)
The local spectral information content can be used to generate sensitive indicators and
analyze the flow field during the computation to select appropriate closures. ii) The very
low numerical error makes the combination with explicit models based on physical con-
siderations very attractive, as an (approximate) decoupling of discretization and model
becomes feasible. iii) More elaborate models, based not only on the solution and its
first order derivatives become meaningful, as the high order discretization provides local
spatial derivatives with acceptable accuracy.
In summary, for moderate Reynolds numbers, an implicit LES approach with a de-
aliased high order DG discretization is highly accurate and efficient. For higher Reynolds
numbers, the most important feature of DG is the local spectral resolution, as it provides
nearly error-free numerics and highly accurate information about the local scale cascade,
which can be exploited both in an explicit or implicit closure approach.
162
5. Conclusion and Prospects
Multiscale problems are ubiquitous in nature and science due to the fact that all but the
most basic laws of physics contain non-linearities. What makes them difficult to tackle
from both an analytical as well as a numerical point of view is not only the resulting large
bandwidth of scales, but also the shift in relative importance of the dominating physical
effects along the supported scale range. In this work, we have focused on hydrodynamic
turbulence as a prominent representative of these types of problems. Although for typical
Mach and Reynolds numbers the governing equations remain valid for the whole scale
range (other multiscale problems may require different sets of descriptive equations at
both ends of their scale range), both their mathematical and physical understanding is
still incomplete due to high complexities and sensitivity to boundary and initial condi-
tions. We have started by discussing the prominent features of these flows, with a special
focus on two issues that become particularly important when resolving only a truncated
subset of the full scales: i) The non-linearities that establish a self-feeding scale cascade
in three dimensions and regulate the transfer of energy between the scales and ii) the sta-
bility of the large scale solution against deviations, which make the LES methodology
meaningful despite the chaotic and highly sensitive small scale behavior.
From this discussion of the physical groundwork, we have outlined the rationale for
LES, and derived a general problem description that incorporates the different strategies
for filtering and modeling. Next, we have defined what a “perfect” LES would entail:
a regularization of the multiscale problem with a significant reduction of its degrees of
freedom, combined with an exact closure for the subgrid terms. Both processes can
only be approximatively realized for realistic turbulence problems, nonetheless, we have
demonstrated that for a simplified problem, this approach leads to the exact solution of
the LES problem, independent from the numerical details. These details however be-
come very important in the practical design of discretization schemes for the numerical
simulation of these problems, both for fully resolved and under-resolved cases.
For three-dimensional turbulence, the properties of the spatial discretization operator de-
termine strongly the efficiency and accuracy of the numerical solution. In particular, the
computational effort for a full resolution of all dynamically active scales is directly pro-
portional (per dimension) to the number of degrees of freedom nP P W required by the
discretization to accurately resolve this scale. Thus, minimizing this quantity has a dras-
tic effect on the accuracy and efficiency per invested degree of freedom, and determines
163
the limit of practically achievable DNS resolutions on HPC systems. We have shown that
the magnitude of nP P W is directly related to the domain of dependence of the discretiza-
tion operators, and that efficiency in terms of the ratio of resolved to represented modes
(on a given grid) can only be achieved by methods with a spectral (or high order) charac-
ter. As an example of global spectral methods, we have discussed Fourier-Galerkin and
Fourier-pseudo-spectral methods, and have presented a Fourier-based pseudo-spectral
solver for the three-dimensional compressible Navier-Stokes equations developed within
this work. Due to the limitations of the global ansatz in terms of geometrical flexibility,
the field of application of these methods is limited to basic research on the “building
block” level. Discontinuous Galerkin methods combine the idea of incorporating all
available information in a domain into the construction of the approximation operators
(resulting in spectral error convergence) with the geometrical flexibility of coupling do-
mains weakly through an appropriate penalty term. In addition, by being element-based,
these hybrid methods avoid the parallelization bottleneck encountered by global me-
thods.
In this work, we have presented a DGSEM framework for the massively parallel nu-
merical simulation of multiscale problems. We have discussed details of the derivation
and implementation, which make this variant of DG highly effective, both in terms of
invested degrees of freedom as well as parallel execution speed. We have compared it
for a DNS of vortical flows against other codes, and found our implementation to be
highly efficient and competitive, clearly distancing other DG methods and challenging
FD schemes. In addition, among the DNS cases reported in this work, the computation
of the Taylor-Green vortex with 216 mio DOF (per conservative variable) on 125, 000
cores is up to date the largest reported DG computation of turbulent flows, both in terms
of the number of unknowns as well as number of computing units, thereby demonstrat-
ing the capabilities of the full framework for Petascale computing.
Despite the steady increase in computing power, DNS of multiscale flows will remain
the exception, while the simulation of truncated flow fields will remain the norm. This
truncation (either through a mathematical filter procedure or through the projection onto
a finite set of DOF, e.g. the grid) imposes two additional challenges, which we have dis-
cussed in detail: i) the truncation operation itself can only in very specific circumstances
(e.g. explicit filtering with a commutative filter on a regular, periodic grid) return a
consistently regularized truncated problem, and ii) the truncation induces the need for a
closure of the unrepresented scales. Since the explicit filtering is numerically very ex-
pensive and difficult on non-regular grids, this approach is seldom followed in practice,
instead, a grid-filtered implicit LES is the norm. In this situation, the discretization op-
erators act on the grid representation of the full solution of the multiscale problem, i.e.
on a “rough” solution containing wavenumbers beyond the resolution capabilities of the
discretization. This entails the introduction of approximation errors into the solution,
164
and as the grid spacing h becomes “large” (with respect to the occurring scales), the
concept of “order of error convergence” of a scheme becomes less meaningful. Instead,
the dispersion and dissipation properties of a scheme over the full wave range become
important. We have discussed these issues and shown that high order DG schemes ben-
efit greatly from their low nP P W requirement, even in under-resolved settings.
Another issue introduced by the truncation onto a limited set of DOF is the occurrence
of aliasing errors through the scale-producing terms. We have investigated the source
of these errors, both physically and numerically, as well as their effects on accuracy of
the solution representation and on the operator stability. Several strategies for avoid-
ing or suppressing these instabilities exist. We have focused on two approaches: i) a
polynomial de-aliasing, an idea translated from the Fourier-pseudo-spectral methods to
element-based spectral methods, where the non-linearities are evaluated on a sufficiently
fine “subgrid” and then accurately projected back onto the computational grid, and ii) a
stabilization through a dissipative filtering, that acts directly on the modes represented on
the computational grid. We have discussed the implementation of both mechanisms into
our framework, and have evaluated them for under-resolved turbulent flows at moderate
Reynolds numbers: the Taylor-Green Vortex case and the flow over a circular cylinder
at ReD = 3900. In both settings, we have compared against published LES results
with approximately the same number of DOF and DNS results. For both cases, we have
found that properly de-aliased high order DGSEM without additional LES closure re-
sults in better solution accuracy compared to low order methods with implicit or explicit
LES modeling. This is due to the superior scale-resolving capabilities of the high order
formulations and the moderate Reynolds number. In particular for the Taylor-Green vor-
tex flow, we have shown that while the stabilization through filtering may be effective
from a numerical point of view, it destroys the favorable nP P W requirement of high
order methods. Thus, our “no model” LES or under-resolved DNS approach based on
de-aliased high order DGSEM, is highly attractive for LES at moderate Reynolds num-
bers.
An extension of this approach to higher Reynolds numbers will necessitate some form
of either explicit or implicit LES modeling. As a first step towards this goal, we have in-
vestigated the interaction of aliasing errors and a basic explicit SGS model. Since these
closure models are essentially dissipative in nature, the question whether an explicit
de-aliasing strategy remains necessary or whether the closure can perform both tasks
(physical model and aliasing control) opens. We have provided an answer by combining
an either aliasing-afflicted or de-aliased discretization with a Smagorinsky model. Our
results clearly underline the need for proper de-aliasing to provide a “neutrally stable”
discretization, both in terms of solution accuracy as well as subgrid model evaluation.
165
5.1. Future Work

Based on the results and discussions presented in this work, a number of interesting
directions for further research should be explored:
Flux Functions in Under-Resolved Turbulence

In Section 4.5.2, we have briefly touched on the influence of the numerical flux func-
tions (with a focus on the Euler fluxes) for under-resolved turbulence. It is obvious from
the strong form of the DG formulation, that for well-resolved situations, the jump at
the cell interface and with it the influence of the flux functions diminishes, but if the
jump becomes non-negligible, the choice of the flux functions will play a role in solu-
tion accuracy. For high order formulations, where the number of volume DOF is much
larger than that of the surface nodes, this effect might not be as pronounced as for low
order schemes, but still can be significant, depending on the degree of under-resolution.
In addition, the choice of Gauss vs. Gauss-Lobatto nodes in combination with the flux
functions should also be investigated, as they differ in terms of how the surface flux is
incorporated into the volume solution.
With regards to implicit LES modeling, we have shown a very simple ad-hoc modeling
approach through the linear combination of penalty terms. This investigation should be
extended, combined with other modeling terms and a more efficient optimization frame-
work and repeated for a homogeneous isotropic turbulence and a turbulent boundary
layer.
Large Scale Simulations

What makes our DGSEM framework stand out among the other codes for the simulation
of multiscale flows are three features: i) As opposed to Fourier spectral or FD methods,
fully unstructured meshes are incorporated naturally in this element-based approach, al-
lowing the discretization of complex geometries. ii) Advection-dominated problems like
vortical flows do not require additional stabilization like in continuous FE, in addition,
the discontinuous approximation space makes efficient parallel scaling easier to achieve.
iii) Accuracy can easily be increased through p-adaptation, without losses in numerical
efficiency (like for high order FV schemes). Thus, DGSEM is essentially a hybrid, com-
bining the favorable properties from other discretization schemes into a single package.
In addition to the scheme itself, it has been shown in Section 3.5 that our complete frame-
work from pre- to post-processing is set up and ready to handle large scale computations
at Petascale level. In future work, this framework should be used to push the limits of
large scale computations of turbulent flows, both in an industrial setting in complex do-
mains, as well as to establish benchmark computations of compressible turbulence for
166
5.1. Future Work
basic research on the next generation of supercomputers.
LES Closure for High Order DG

As has been discussed in Section 4.6, the current state of the art for LES with DG meth-
ods is either a “no model” implicit LES approach or the incorporation of classical ex-
plicit SGS models, both with a baseline scheme of moderate order. While these two
approaches have produced acceptable results for moderately complex flows, they both
do not exploit the main advantage of a spectral element method: The existence of a so-
lution that contains a significant range of accurately resolved scales within in element.
This property can be exploited for both implicit and explicit LES closures strategies:
i) The low approximation errors of high order DG make the combination with an ex-
plicit model particularly attractive, as model and numerics can be considered essentially
decoupled. Thus, the interaction between the model and the discretization is minimal.
This is particularly important for scale-based models, i.e. those derived from physical
considerations of the scale cascade (e.g. the Scale Similarity model or Interscale Trans-
fer model discussed in Section 2.4.3), since the numerical representation of the scales
close to the grid cut-off retains acceptable accuracy. In addition, high order closure ap-
proaches developed within the context of global spectral methods can now directly be
transferred to DGSEM [22, 130].
ii) The de-aliasing strategy through exact integration of the non-linear fluxes and its
implementation through a modal filtering approach has been discussed in detail in Sec-
tion 4.2.2. It should be noted that this approach itself should not be considered as an
implicit LES closure, as it merely removes a source of error, but does not exploit or ma-
nipulate it as a SGS model. Nonetheless, the polynomial de-aliasing procedure gives an
idea for a novel SGS closure for high order methods, that is outlined here briefly: From
a purely numerical point of view, the expansion of the flux function on a finer grid (with
associated quadrature and interpolation points M > N ) serves as a way of evaluating
its non-linearity exactly. From a modeling point of view, this fine grid also serves as a
form of subgrid, i.e. an extension of the computational grid beyond N (or the associated
Nyquist wavenumber). On this subgrid, the non-linear terms produce fluctuating scales
that act as a partner in the triadic interactions and allow the deposition of energy from
the larger scales. Due to the truncation error, this energy will accumulate and eventu-
ally lead to a crash, if not removed ahead of time. Based on physical considerations,
slope indicators can be constructed, that monitor this build up and trigger a de-aliasing
step to clear the subgrid scales. This operation can be conducted element-wise. In other
words, this locally adaptive de-aliasing mechanism (LAD) exploits the fine integration
grid as an energy sink, where it allows the occurrence of subgrid scales in a decon-
volution sense. If the deconvolved scales become energetically charged, the energy is
167
10-1
E(k)
-2
10
DNS
PDG
LAD
-3
10
10 20 30 40
k
Figure 5.1.: Kinetic energy spectra of HIT at Reλ = 82. DNS (solid line) computed
with Spex on 5123 nodes. Dashed line denotes PDG result, solid line with
markers denotes LAD result, both with 723 DOF (N = 11). Data taken
from [74].
removed by a projection onto the solution grid. Thereby, the de-aliasing procedure can
be modified towards an implicit LES closure.
Figure 5.1 shows the potential of this approach for decaying isotropic turbulence. The
PDG solution is given for comparison, as it denotes the fixed de-aliasing in every oper-
ator evaluation, while the LAD scheme only performs a de-aliasing step when triggered
by the indicators. It can be seen that PDG method suffer from a build-up of energy in the
low modes, due to the cascade blockage. This blockage is reduced in LAD, resulting in
an improvement agreement with the DNS results for all wave numbers. In a related idea,
a linear combination of a de-aliased and an aliasing-influenced fluxes could provide a
similar energy drain mechanism. Thus, both of these ideas takes full advantage of the
spectrality and locality of DGSEM, and should be investigated and improved in future
research.
168
A. Initial Conditions for Homogeneous Isotropic
Turbulence
Following Rogallo [175], an initial incompressible velocity field with zero mean and a
prescribed energy spectrum for the simulation of homogeneous isotropic turbulence can
be generated from the following equations for the Fourier coefficients at wave vector
~k = (k1 , k2 , k3 )T for the three components of velocity:
αkk2 + βk1 k3
ũ1 (~k) = p ,
k k12 + k22
βk2 k3 − αkk1
ũ2 (~k) = p ,
k k12 + k22 (A.1)
p
β( k12 + k22 )
ũ3 (~k) = − ,
k
where k denotes the magnitude of ~k, α and β are given by Equation (A.2) and the
sign error for ũ3 in the original publication has been corrected. The incompressibility
condition in wave space has been incorporated into the determination of the components
in Equation (A.1), and the energy spectrum is prescribed through the magnitude of the
random coefficients α and β as
1
E(k) 2 Iθ1
α= e cos φ
4πk2
1 (A.2)
E(k) 2 Iθ2
β= e sin φ,
4πk2
with θ1 , θ2 and φ uniformly distributed random numbers on the interval (0, 2π) and
E(k) the desired energy spectrum. In our implementation, each coefficient is scaled by
the number of wavevectors assigned to its shell, so that the total energy per shell can
be described directly. Figure A.1 shows the coefficients α and β for a box spectrum.
Note the hollow sphere structure in wave space, and that red colors indicate a higher
magnitude, due to the decreasing number of wave vectors on the inner shells. It should
169
A. Initial Conditions for Homogeneous Isotropic Turbulence
Figure A.1.: Coefficients α (left) and β (right) in wave space for a box spectrum E(k) =
const for 4 ≤ k ≤ 12, E(k) = 0 otherwise.
be noted that Rogallo’s procedure yields a velocity field that isotropic in the x-y plane
in physical space, but anisotropic in the y-z plane. Figure A.2 shows the scatter plots of
these correlations.
Figure A.2.: Scatter plot of u-v and v-w correlations from Rogallo’s procedure (Equa-
tions (A.1) and (A.2)).
From Equation (A.1), the resulting modal velocity coefficients can be transferred to
physical space with a FFT. The Poisson equation for pressure (Equation (2.18)) can
then be solved in a pseudo-spectral manner, where the product ui uj is evaluated in a
collocation step in physical space, and the differentiation is conducted in spectral space.
170
B. Flow over a Cylinder at ReD = 3900 -
Extended Results
In this appendix, we provide additional results for the investigations reported in Sec-
tion 4.4.
x/D=1.06 x/D=1.06
0 0
-0.2 -0.2
x/D=1.54 x/D=1.54
<u’v’>/U
<u’v’>/U
-0.4 -0.4
x/D=2.02 x/D=2.02
-0.6 -0.6
-0.8 -0.8
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
y/D y/D
Figure B.1.: Time- and spanwise averaged velocity fluctuations hu′ v ′ i in the cylinder
wake. Line with squares: DNS results;
left: DGSEM according to Table 4.5. Dashed line: N7-A; dotted line: N5-
A; dash-dotted line: N3-A.
right: PDG according to Table 4.5. Dashed line: N7-1; dotted line: N5-1;
dash-dotted line: N3-1.
171
B. Flow over a Cylinder at ReD = 3900 - Extended Results
Figure B.2.: One-dimensional spectra of the crossflow velocity component in the wake
at x/D = 3.0, y/D = 0.25. All frequencies normalized by the Strouhal
frequency of the reference DNS (see Table 4.5). a) N7-1, b) N7-A, c) N5-1,
d) N5-A, e) N3-1, f) N3-A; dashed line denotes Kolmogorov’s k−5/3 law.
172
Figure B.3.: Time- and spanwise-averaged velocity fluctuations hu′ v ′ i in the wake re-
gion, normalized by square of the free-stream velocity U∞ . Top row: DNS
result. Second to fourth row: results for N7, N5 and N3 cases, with PDG
in the left and DGSEM in the right column.
173
Figure B.4.: Time- and spanwise-averaged crossflow velocity fluctuations hv ′ v ′ i in the

wake region, normalized by square of the free-stream velocity U∞ . Top
row: DNS result. Second to fourth row: results for N7, N5 and N3 cases,
with PDG in the left and DGSEM in the right column.
174
Table B.1.: Simulation parameters for ReD = 3900 cylinder LES computations.
Case Cs CpBase Str CD CLrms Lr /D θsep [◦ ] θre [◦ ]
DNS - -0.979 0.204 1.082 0.246 1.160 87.2 116.3

N7-1 0.0 -0.995 0.206 1.085 0.252 1.165 81.8 117.1
N7-2 0.05 -0.938 0.206 1.050 0.190 1.329 89.8 121.4
N7-3 0.10 -0.916 0.204 1.043 0.167 1.426 87.5 126.0
N7-4 0.20 -0.867 0.198 1.013 0.118 1.626 88.9 136.7
N7-A 0.16 -0.919 0.200 1.045 0.164 1.413 87.9 124.8
N7-B 0.18 -0.893 0.200 1.030 0.143 1.506 88.7 125.6
N7-C 0.20 -0.883 0.199 1.023 0.136 1.555 89.4 128.9
N5-1 0.0 -0.975 0.208 1.072 0.232 1.217 86.0 117.2
N5-2 0.05 -0.935 0.206 1.053 0.195 1.350 87.4 121.2
N5-3 0.10 -0.895 0.205 1.026 0.145 1.499 87.1 131.1
N5-4 0.20 -0.906 0.200 1.037 0.156 1.474 88.6 125.2
N5-A 0.10 -0.878 0.203 1.016 0.127 1.560 87.6 132.7
N5-B 0.15 -0.886 0.202 1.022 0.136 1.541 88.0 131.8
N5-C 0.16 -0.878 0.200 1.018 0.130 1.570 88.1 130.8
N5-D 0.20 -0.871 0.202 1.013 0.127 1.607 88.5 133.0
N3-1 0.0 -0.847 0.216 0.927 0.106 1.569 90.5 123.4
N3-2 0.05 -0.862 0.208 0.973 0.110 1.626 87.2 124.7
N3-3 0.10 -0.875 0.202 0.999 0.122 1.598 86.7 126.6
N3-4 0.20 -0.883 0.202 1.016 0.140 1.569 90.1 128.2
N3-A 0.0 -0.767 0.236 0.800 0.059 1.655 96.8 128.0
N3-B 0.05 -0.813 0.212 0.902 0.084 1.731 90.5 125.3
N3-C 0.10 -0.854 0.200 0.981 0.105 1.683 86.7 131.6
N3-D 0.20 -0.876 0.200 1.011 0.127 1.588 90.6 132.1
CpBase : pressure coefficient at the downstream position x = D/2, y = 0, Str: Strouhal number of the lift coefficient
175
fLif t D/u∞ , CD : drag coefficient, CL rms : RMS value of the lift coefficient, Lr : length of separation bubble, θsep , θre :
separation and reattachment location.
176
Figure B.5.: Time- and spanwise-averaged turbulent kinetic energy in the wake region, normalized by free-stream
kinetic energy. First row: DNS reference results, first column: PDG results, second column: PDG with
Smagorinsky model, last column: DGSEM with Smagorinsky model; Second row: N7-1, N7-2, N7-A,
third row: N5-1, N5-2, N5-B, fourth row: N3-1, N3-4, N3-D.
Figure B.6.: Time- and spanwise-averaged velocity magnitude in the wake region, normalized by free-stream kinetic
energy. First row: DNS reference results, first column: PDG results, second column: PDG with Smagorin-
177
sky model, last column: DGSEM with Smagorinsky model; Second row: N7-1, N7-2, N7-A, third row:
N5-1, N5-2, N5-B, fourth row: N3-1, N3-4, N3-D.
1 1
0.8 0.8
0.6 0.6
y/D
y/D
0.4 0.4
c c
0.2 0.2
DNS a b d DNS a b d
0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
x/D x/D
(a) PDG: a) N7-1, b) N7-2, c) N7-3, d) N7-4 (b) DGSEM: a) N7-1, b) N7-A, c) N7-B, d) N7-C
1 1
0.8 0.8
0.6 0.6
y/D
y/D
0.4 0.4
b
c
0.2 0.2
DNS a b d DNS a c d
0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
x/D x/D
(c) PDG: a) N5-1, b) N5-2, c) N5-3, d) N5-4 (d) DGSEM: a) N5-1, b) N5-A, c) N5-B, d) N5-C
0.4 0.4
a b
y/D
y/D
c b
0.2 0.2
d
DNS DNS c
d a
0 0
1 1.5 2 1 1.5 2
x/D x/D
(e) PDG: a) N3-1, b) N3-2, c) N3-3, d) N3-4 (f) DGSEM: a) N3-A, b) N3-B, c) N3-C, d) N3-D
Figure B.7.: Isocontours of time- and spanwise-averaged streamwise velocity: u = 0.

Shaded area denotes the DNS reference solution.
178
C. Taylor-Green Vortex - Extended Results
In this appendix, we provide the companion results for filter-stabilized DGSEM and
PDG of O(8), i.e. N = 7 to the discussion in Section 4.3.1. All compuations were
again performed with 643 , i.e. on a grid with 83 elements. The findings discussed
before also hold for the N = 7 computations, although the superiority of PDG over
filter-stabilized DGSEM is less pronounced, due to the increased nP P W requirement of
a N = 7 discretization.
C.1. Taylor-Green Vortex at Re=800, N=7
0.015 0.015

ALDM (Hickel) ALDM (Hickel)
N=7 N=7
N=7, Strong Filter N=7, Strong Filter
N=7, Weak Filter N=7, Weak Filter
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure C.1.: Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk dt

0.015 0.015
N=7 N=7
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure C.2.: Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk

dt
0.015 0.015

N=7, Int Points=9 N=7, Strong Filter
N=7, Int Points=12 N=7, Weak Filter
N=7, Int Points=16
-dk/dt
-dk/dt
0.01 molec. 0.01
Diss. molec.
Diss.
0.005 0.005
num.
num. Diss.
Diss.
0 0
0 2 4 6 8 10 0 2 4 6 8 10
t t
Figure C.3.: Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate, re-
solved dissipation rate and numerical disspation rate for PDG N = 7 (left)
and filter-stabilized DGSEM N = 7 computation (right).
180
0.015 0.015

N=7 N=7
N=7, Strong Filter N=7, Strong Filter
N=7, Weak Filter N=7, Weak Filter
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure C.4.: Taylor-Green Vortex at Re = 1600. Kinetic energy dissipation rate − dk dt

0.015 0.015
N=7 N=7
-dk/dt
0.01
-dk/dt
0.01
0.005
0 0.005
0 2 4 6 8 10 5 6 7 8 9 10
t t
Figure C.5.: Taylor-Green Vortex at Re = 1600. Kinetic energy dissipation rate − dk

dt
181
0.015 0.015

N=7, Int Points=9 N=7, Strong Filter -dk/dt
N=7, Int Points=12 -dk/dt N=7, Weak Filter
N=7, Int Points=16
molec.
Diss.
0.01 0.01
molec.
Diss.
0.005 0.005
num.
Diss.
num.
Diss.
0 0
0 2 4 6 8 10 0 2 4 6 8 10
t t
Figure C.6.: Taylor-Green Vortex at Re = 1600. Kinetic energy dissipation rate, re-
solved dissipation rate and numerical disspation rate for PDG N = 7 (left)
and filter-stabilized DGSEM N = 7 computation (right).
C.3. Results of the 1st International Workshop on High Order

Methods
In the following, we provide additional results from the “1st International Workshop on
High Order Methods” discussed in Section 3.4.3 and published in [200]. Comparison
plots are reproduced with friendly permission of Dr. Koen Hillewaert, Cenaero.
For the compressible formulation of the TGV, the total dissipation rate of kinetic energy
is given by
Z Z Z
dE µ µv 1
ǫ=− =2 S d : S d dΩ + (∇ · ~v )2 dΩ − p∇ · ~v dΩ, (C.1)
dt ρ Ω ρ Ω ρ Ω
| {z } | {z } | {z }
ǫ1 =2νE ǫ2 ≈0 ǫ3 ≈0
where S d denotes the deviatoric part of the stress tensor and µv the bulk viscosity, which
is assumed to be zero in this case. The results were compared by computing the follow-
182
C.3. Results of the 1st International Workshop on High Order Methods
ing metrics with respect to a reference solution, denoted by ( )ref
dE dE ref
∆ǫ1 = − t
,
dt dt
ref
∆ǫ2 = 2νE − 2νE t
,
(C.2)
dE
∆ǫ3 = + 2νE t
.
dt
Figure C.7.: Taylor-Green Vortex at Re = 1600: Comparison of the resolved (left) and
total dissipation rate (right) for the contributions to the high order work-
shop with highest order per participant. Refer to Table 3.3 for the legend.
183
Figure C.8.: Taylor-Green Vortex at Re = 1600: Comparison of error measures ac-

cording to Equation C.2 for the contributions to the high order workshop.
For a legend refer to Table 3.3.
184
C.3. Results of the 1st International Workshop on High Order Methods
Figure C.9 shows the vortical structures for DGSEM with 2563 DOF.
Figure C.9.: Taylor-Green Vortex at Re = 1600: Visualization of vortex detection cri-

terion λ2 = −1.5 for N = 1 (top), N = 3 (bottom left) and N = 15
(bottom right) computation (2563 DOF in each case).
185
D. Central Finite Differences of Arbitrary Order
In [76], Fornberg introduces a shorthand notation for standard FD approximations of a

function u(x):
du(x)
D (True derivative): D u(x) =
dx
I (Identity operator): I u(x) = u(x)
E (Translation operator): E u(x) = u(x + h) (D.1)
D+ (Forward difference): D+ u(x) = (E − I)/h u(x)
D− (Backward difference): D− u(x) = (I − E −1 )/h u(x)
D0 (Central difference): D0 u(x) = (E − E −1 )/2h u(x)
With this notation, a second order central approximation becomes

!
du(x) u(x + h) − u(u(x − h)
≈
dx 2h
1
1 X
D≈ β1,i E i , with β1,−1 = −1, β1,0 = 0, β1,1 = 1, (D.2)
2h i=−1
or
D ≈ α 0 D0 with α0 = 1.
Similarly, a fourth order approximation can be written as
2
1 X
D≈ β2,i E i ,
2h i=−2
1 4 (D.3)
with β2,−2 = −β2,2 = β2,−1 = −β2,1 = − , β2,0 = 0, or
6 3
1
D ≈ D0 (α0 I − α1 h2 D+ D− ) with α0 = 1, α1 = .
6
187
D. Central Finite Differences of Arbitrary Order
For an arbitrary order of accuracy 2p, this extends to

p
1 X 2(p!)2 (−1)i+1
D≈ βp,i E i , with β0,p = 0, βp,i6=0 =
2h i=−p i(p + i)!(p − i)!
or (D.4)
p−1
X (i!)2
D ≈ D0 (−1)i αi (h2 D+ D− )i with αi = .
i=0
(2i + 1)!
With relationship (D.4), an analytical derivation of the dispersion error of central FD

discretizations is possible: Considering a linear transport equation for u and an initial
solution of the form u(x, t0 ) = eIωx with wave mode ω on a regular grid with periodic
boundary conditions, the true derivative is given by
D u(x, t0 ) = DeIωx = IωeIωx . (D.5)
Its numerical approximation can be computed from (D.4), for example, a second order
central approximation gives:
eIω(x+h) − eIω(x−h) sin hω Iωx

D0 u(x, t0 ) = =I e . (D.6)
2h h
Thus, the phase error between the true derivative and its approximation becomes
(D − D0 )eIωx sin hω Iωx

=ω− e . (D.7)
IeIωx h
From Equation (D.4), Fornberg showed that for a general approximation of order 2p,
this dispersion error becomes
p−1 2i
sin ωh X 2i ωh
ω− αi 2 sin . (D.8)
h i=0
2
188
E. Derivation of the Vorticity Equation
Incompressible Navier Stokes equations in Vectorial Form:
∇ · u = 0, divergence-free condition on the velocity field (E.1)

∂u ∇p
+ (u · ∇)u = − + ν∇2 u vectorial momentum equation. (E.2)
∂t ρ
Vorticity is defined as the vector aligned with the local axis of fluid rotation, with the
sign determined by the local right hand system. The length of the vector corresponds to
the strength of the vortex, i.e. directly to the radial velocity gradient.
ω ≡∇×u (E.3)
Taking the curl of Equation (E.2) term-by-term yields:

• First term LHS:
∂u ∂(∇ × u) ∂ω
∇× = =
∂t ∂t ∂t
• Second term RHS:
∇ × ν∇2 u = ν∇2 ω
• First term RHS:

∇p
∇× = 0,
ρ
since the gradient of a scalar is curl-free.
• Non-linear term LHS:

∇ × (u · ∇)u,
with vector relation
∇(u · u) =(u · ∇) u + (u · ∇) u + u × (∇ × u) + u × (∇ × u)
(E.4)
=2 (u · ∇) u + 2 u × (∇ × u),
189
E. Derivation of the Vorticity Equation
and ω ≡ ∇ × u becomes
1
∇ × (u · ∇) u = ∇ × ∇( u2 ) − ∇ × (u × ω). (E.5)
2
Again, the curl of the gradient of scalar field is zero, so the first term on the RHS
vanishes. Expanding Equation (E.5) yields
∇ × (u · ∇) u = (u · ∇) ω + (∇ · u) ω − (∇ · ω) u − (ω · ∇) u. (E.6)
Applying the incompressibility condition and noting that the divergence of the
curl is zero leaves
∇ × (u · ∇) u = (u · ∇) ω − (ω · ∇) u. (E.7)
The Vorticity Equation

Collecting the terms derived above yields the evolution equation for vorticity
Dω ∂ω
= + (u · ∇) ω = (ω · ∇) u + ν ∇2 ω (E.8)
Dt ∂t
Remarks
• Equation (E.8) supports Kelvin’s theorem: In the absence of viscosity (ν = 0)
then if no vorticity exists in the flow field initially (ωt=0 = 0), none will be
created.
• The first term on RHS is a source for the vorticity and a mechanism of increasing
the vorticity if shear exists in the mean flow. It is commonly called the vortex-
stretching term.
• The first term on the RHS vanishes in 2D, so no production mechanism for vor-
ticity exists.
190
Bibliography
[1] Large Eddy Simulation and Related Techniques, Theory and Application. VKI
Lecture Series. Von Karman Institute for Fluid Dynamics, 2010.
[2] V. I. Abramenko. Fractal multi-scale nature of solar/stellar magnetic fields. In

Kosovichev A. G., de Gouveia Dal Pino E., and Yan Y., editors, IAU Symposium,
volume 294 of IAU Symposium, pages 289–300, July 2013.
[3] P. Abry, R. Baraniuk, P. Flandrin, R. Riedi, and D. Veitch. Multiscale nature of

network traffic. Signal Processing Magazine, IEEE, 19(3):28–46, May 2002.
[4] Ch. Altmann, A. Beck, F. Hindenlang, M. Staudenmaier, G. Gassner, and C.-D.

Munz. An efficient high performance parallelization of a discontinuous galerkin
spectral element method. In Rainer Keller, David Kramer, and Jan-Philipp Weiss,
editors, Facing the Multicore-Challenge III, volume 7686 of Lecture Notes in
Computer Science, pages 37–47. Springer Berlin Heidelberg, 2013.
[5] B. W. Anderson and J. A. Domaradzki. A subgrid-scale model for large-eddy sim-

ulation based on the physics of interscale energy transfer in turbulence. Physics
of Fluids (1994-present), 24(6), 2012.
[6] A. Aprovitola and F. M. Denaro. On the application of congruent upwind

discretizations for large eddy simulations. Journal of Computational Physics,
194(1):329 – 343, 2004.
[7] D. Arnold. An Interior Penalty Finite Element Method with Discontinuous Ele-
ments. PhD thesis, The University of Chicago, 1979.
[8] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini. Discontinuous Galerkin

methods for elliptic problems. In B. Cockburn, G. Karniadakis, and C. W. Shu,
editors, Discontinuous Galerkin Methods. Lecture Notes in Computational Sci-
ence and Engineering, pages 89–101. Springer, 2000.
[9] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini. Unified analysis of

discontinuous Galerkin methods for elliptic problems. SIAM J. Numer. Anal.,
39(5):1749–1779, 2002.
191
Bibliography
[10] J. Bardina, J. Ferziger, and W. Reynolds. Improved subgrid-scale models for

large-eddy simulation. In Fluid Dynamics and Co-located Conferences. American
Institute of Aeronautics and Astronautics, 1980.
[11] G. E. Barter and D. L. Darmofal. Shock capturing with PDE-based artificial

viscosity for DGFEM: Part 1. Formulation. Journal of Computational Physics,
229(5):1810 – 1827, 2010.
[12] F. Bassi and S. Rebay. A high order accurate discontinuous finite element method
for the numerical solution of the compressible Navier–Stokes equations. Journal
of Computational Physics, 131:267–279, 1997.
[13] F. Bassi and S. Rebay. A high-order discontinuous Galerkin finite element method
solution of the 2D Euler equations. Journal of Computational Physics, 138:251–
285, 1997.
[14] F. Bassi and S. Rebay. Numerical evaluation of two discontinuous Galerkin meth-
ods for the compressible Navier–Stokes equations. International Journal for Nu-
merical Methods in Fluids, 40:197–207, 2002.
[15] F. Bassi, S. Rebay, G. Mariotti, S. Pedinotti, and M. Savini. A high-order accu-

rate discontinuous finite element method for inviscid and viscous turbomachin-
ery flows. In R. Decuypere and G. Dibelius, editors, Proceedings of 2nd Euro-
pean Conference on Turbomachinery, Fluid and Thermodynamics, pages 99–108,
Technologisch Instituut, Antwerpen, Belgium, 1997.
[16] G. K. Batchelor. The Theory of Homogeneous Turbulence. Cambridge University

Press, United Kingdom, 1959.
[17] Y. Bazilevs and T. J. R. Hughes. Weak imposition of Dirichlet boundary condi-

tions in fluid mechanics. Computers & Fluids, 36(1):12 – 26, 2007.
[18] A. Beck, Th. Bolemann, D. Flad, H. Frank, G. Gassner, F. Hindenlang, and C.-D.
Munz. High-order discontinuous galerkin spectral element methods for transi-
tional and turbulent flow simulations. International Journal for Numerical Meth-
ods in Fluids, 76(8):522–548, 2014.
[19] A. Beck, G. Gassner, and C.-D. Munz. High order and underresolution. In R. An-
sorge, H. Bijl, A. Meister, and Th. Sonar, editors, Recent Developments in the
Numerics of Nonlinear Hyperbolic Conservation. Springer, 2013.
192
Bibliography
[20] A. Beck, G. Gassner, and C.-D. Munz. On the effect of flux functions in dis-
continuous Galerkin simulations of underresolved turbulence. In Mejdi Azaı̈ez,
Henda El Fekih, and Jan S. Hesthaven, editors, Spectral and High Order Meth-
ods for Partial Differential Equations - ICOSAHOM 2012, volume 95 of Lecture
Notes in Computational Science and Engineering, pages 145–155. Springer In-
ternational Publishing, 2014.
[21] A. Beck, G. Gassner, C. Tonhäuser, and C.-D. Munz. The influence of poly-
nomial de-aliasing on subgrid scalemodels. submitted to Flow, Turbulence and
Combustion, 2014.
[22] L. C. Berselli and T. Iliescu. A higher-order subfilter-scale model for large eddy
simulation. Journal of Computational and Applied Mathematics, 159(2):411 –
430, 2003.
[23] H. M. Blackburn and S. Schmidt. Large eddy simulation of flow past a circular
cylinder. In Proceedings of 14th Australasian Fluid Mechanics Conference, 2001.
[24] H. M. Blackburn and S. Schmidt. Spectral element filtering techniques for large
eddy simulation with dynamic estimation. Journal of Computational Physics,
186(2):610–629, 2003.
[25] J. P. Boris, F. F. Grinstein, E. S. Oran, and R. L. Kolbe. New insights into large
eddy simulation. Fluid Dynamics Research, 10(4–6):199 – 228, 1992.
[26] S. T. Bose, P. Moin, and D. You. Grid-independent large-eddy simulation using

explicit filtering. In Proceedings of the 2008 Center for Turbulence Research
Summer Program, 2008.
[27] J. P. Boyd. Chebyshev and Fourier Spectral Methods. Dover Publications, USA,
2001.
[28] M. E. Brachet. Direct simulation of three-dimensional turbulence in the Taylor–

Green vortex. Fluid Dynamics Research, 8(1-4):1 – 8, 1991.
[29] M. E. Brachet, D. I. Meiron, S. A. Orszag, B. G. Nickel, R. H. Morf, and

U. Frisch. Small-scale structure of the Taylor–Green vortex. Journal of Fluid
Mechanics, 130:411–452, 4 1983.
[30] J. G. Brasseur and Ch. H. Wei. Interscale dynamics and local isotropy in high
Reynolds number turbulence within triadic interactions. Physics of Fluids, 6(2),
1994.
193
Bibliography
[31] M. Breuer. Large eddy simulation of the subcritical flow past a circular cylinder:
numerical and modeling aspects. International Journal for Numerical Methods in
Fluids, 28(9):1281–1302, 1998.
[32] A. N. Brooks and T. J. R. Hughes. Streamline Upwind/Petrov-Galerkin formu-

lations for convection dominated flows with particular emphasis on the incom-
pressible Navier-Stokes equations. Comput. Methods Appl. Mech. Eng., pages
199–259, September 1990.
[33] J. R. Bull and A. Jameson. Simulation of the compressible Taylor Green vor-
tex using high-order flux reconstruction schemes. In AIAA Aviation. American
[34] M. Carpenter and C. Kennedy. Fourth-order 2N-storage Runge-Kutta schemes.

Technical Report NASA TM 109111, 1994.
[35] C. Carton de Wiart and K. Hillewaert. DNS and ILES of transitional flows around
a SD7003 using a high order discontinuous Galerkin method. In Seventh Interna-
tional Conference on Computational Fluid Dynamics (ICCFD7), 2012.
[36] C. Carton de Wiart, K. Hillewaert, M. Duponcheel, and G. Winckelmans. Assess-

ment of a discontinuous Galerkin method for the simulation of vortical flows at
high Reynolds number. International Journal for Numerical Methods in Fluids,
2013.
[37] C. Carton de Wiart, K. Hillewaert, M. Duponcheel, and G. Winckelmans. As-

sessment of a discontinuous galerkin method for the simulation of vortical flows
at high reynolds number. International Journal for Numerical Methods in Fluids,
74(7):469–493, 2014.
[38] E. Casoni, J. Peraire, and A. Huerta. One-dimensional shock-capturing for high-

order discontinuous Galerkin methods. International Journal for Numerical
Methods in Fluids, 71(6):737–755, 2013.
[39] L. Catabriga, A. L. G. A. Coutinho, and T. E. Tezduyar. Compressible flow

SUPG parameters computed from element matrices. Communications in Numer-
ical Methods in Engineering, 21(9):465–476, 2005.
[40] N. Chalmers, L. Krivodonova, and R. Qin. Relaxing the CFL number of the
discontinuous Galerkin method. submitted to SIAM JSC.
194
Bibliography
[41] J.-B. Chapelier, M. de la Llave Plata, F. Renac, and E. Lamballais. Evaluation

of a high-order discontinuous Galerkin method for the DNS of turbulent flows.
Computers & Fluids, 95(0):210 – 226, 2014.
[42] D. K. Chapman. Computational aerodynamics development and outlook. AIAA

Journal, 17(12):1293–1313, Dec 1979.
[43] A. Chatterjee and R. S. Myong. Efficient implementation of higher-order finite

volume time-domain method for electrically large scatterers. Progress In Electro-
magnetics Research B, 2009.
[44] H. Choi and P. Moin. Grid-point requirements for large eddy simulation: Chap-
man’s estimates revisited. Physics of Fluids (1994-present), 24(1):–, 2012.
[45] B. Cockburn, S. Hou, and C.-W. Shu. The Runge-Kutta local projection discon-
tinuous Galerkin finite element method for conservation laws IV: The multidi-
mensional case. Math. Comput., 54:545–581, 1990.
[46] B. Cockburn, S. Y. Lin, and C.-W. Shu. TVB Runge-Kutta local projection dis-
continuous Galerkin finite element method for conservation laws III: One dimen-
sional systems. Journal of Computational Physics, 84:90–113, 1989.
[47] B. Cockburn and C.-W. Shu. TVB Runge-Kutta local projection discontinuous
Galerkin finite element method for conservation laws II: General framework.
Math. Comput., 52:411–435, 1989.
[48] B. Cockburn and C.-W. Shu. The Runge-Kutta local projection p1 -discontinuous
Galerkin method for scalar conservation laws. M2 AN, 25:337–361, 1991.
[49] B. Cockburn and C.-W. Shu. The local discontinuous Galerkin method for time-
dependent convection diffusion systems. SIAM Journal on Numerical Analysis,
35:2440–2463, 1998.
[50] B. Cockburn and C.-W. Shu. Runge–Kutta discontinuous galerkin methods for
convection-dominated problems. Journal of Scientific Computing, 16(3):173–
261, 2001.
[51] S. S. Collis. Discontinuous Galerkin methods for turbulence simulation. In Pro-

ceedings of the 2002 Center for Turbulence Research Summer Program, pages
155–167, 2002.
195
Bibliography
[52] S. S. Collis. The DG/VMS method for unified turbulence simulation. In 32nd
AIAA Fluid Dynamics Conference and Exhibit, 2002.
[53] G. Comte-Bellot and S. Corrsin. The use of a contraction to improve the isotropy
of grid-generated turbulence. Journal of Fluid Mechanics, 25:657–682, 8 1966.
[54] J. W. Cooley and J. W. Tukey. An algorithm for the machine calculation of com-
plex Fourier series. Mathematicas of Computation, 19:297–301, 1965.
[55] G.-H. Cottet and O. V. Vasilyev. Comparison of dynamic Smagorinsky and

anisotropic subgrid-scale models. In Proceedings of the 1998 Center for Tur-
bulence Research Summer Program, 1998.
[56] R. B. Dahlburg and J. M. Picone. Pseudospectral simulation of compressible

magnetohydrodynamic turbulence. Computer Methods in Applied Mechanics and
Engineering, 80(1–3):409 – 416, 1990.
[57] P. A. Davidson. Turbulence. Oxford University Press, United Kingdom, 2004.
[58] G. De Stefano and O. V. Vasilyev. “Perfect” modeling framework for dynamic

SGS model testing in large eddy simulation. Theoretical and Computational Fluid
Dynamics, 18(1):27–41, 2004.
[59] J. W. Deardorff. A numerical study of three-dimensional turbulent channel flow

at large Reynolds numbers. Journal of Fluid Mechanics, 1970.
[60] J. DeBonis. Solutions of the Taylor-Green vortex problem using high-resolution

explicit finite difference methods. In Aerospace Sciences Meetings. American
[61] T. S. Deisboeck, Z. Wang, P. Macklin, and V. Cristini. Multiscale cancer model-

ing. Annu Rev Biomed Eng, 13:127–155, Aug 2011.
[62] P. J. Diamessis, J. A. Domaradzki, and J. S. Hesthaven. A spectral multido-

main penalty method model for the simulation of high Reynolds number local-
ized incompressible stratified turbulence. Journal of Computational Physics,
202(1):298–322, 2005.
[63] J. A. Domaradzki. Nonlocal triad interactions and the dissipation range of

isotropic turbulence. Physics of Fluids A: Fluid Dynamics (1989-1993), 4(9),
1992.
196
Bibliography
[64] J. A. Domaradzki and E. M. Saiki. Backscatter models for large-eddy simulations.

Theoretical and Computational Fluid Dynamics, 9(2):75–83, 1997.
[65] S. Dong, G. E. Karniadakis, A. Ekmekci, and D. Rockwell. A combined direct nu-

merical simulation–particle image velocimetry study of the turbulent near wake.
Journal of Fluid Mechanics, 569:185–207, 2006.
[66] D. A. Donzis, J. D. Gibbon, A. Gupta, R. M. Kerr, R. Pandit, and D. Vincenzi.

Vorticity moments in four numerical simulations of the 3D Navier-Stokes equa-
tions. Journal of Fluid Mechanics, Volume 732:316–331, October 2013.
[67] P. G. Drazin and W. H. Reid. Hydrodynamic Stability. Cambridge University

Press, United Kingdom, 2004.
[68] E. R. Van Driest. On turbulent flow near a wall. Journal of Aeronautical Sciences,
1956.
[69] D. Drikakis, C. Fureby, F. F. Grinstein, and D. Youngs. Simulation of transition

and turbulence decay in the Taylor–Green vortex. Journal of Turbulence, 2007.
[70] M. Dröge. Cartesian Grid Methods for Turbulent Flow Simulation in Complex
Geometries. Dissertation, University of Groningen, Netherlands, 2007.
[71] M. Drosson and K. Hillewaert. On the stability of the symmetric interior penalty
method for the Spalart-Allmaras turbulence model. J. Comput. Appl. Math.,
246:122–135, 2013.
[72] H. Faisst and B. Eckhardt. Sensitive dependence on initial conditions in transition

to turbulence in pipe flow. Journal of Fluid Mechanics, 504:343–352, 4 2004.
[73] D. Fauconnier. Development of a Dynamic Finite Difference Method for Large-

Eddy Simulation. PhD thesis, University of Ghent, Belgium, 2008.
[74] D. Flad, A. Beck, G. Gassner, and C.-D. Munz. Locally filtered Large Eddy
Simulation. in preparation.
[75] D. Flad, A. Beck, G. Gassner, and C.-D. Munz. A discontinuous Galerkin spectral
element method for the direct numerical simulation of aeroacoustics. In AIAA
Aviation. American Institute of Aeronautics and Astronautics, 2014.
[76] B. Fornberg. The pseudospectral method; comparisons with finite differences for
the elastic wave equation. Geophysics, 52(4):483–501, 1987.
197
Bibliography
[77] J. Franke and W. Frank. Large eddy simulation of the flow past a circular cylinder
at ReD = 3900. Journal of Wind Engineering and Industrial Aerodynamics,
90(10):1191 – 1206, 2002.
[78] M. Frigo and S. G. Johnson. The design and implementation of FFTW3. Proceed-
ings of the IEEE, 93(2):216–231, 2005. Special issue on “Program Generation,
Optimization, and Platform Adaptation”.
[79] U. Frisch and J. Bec. Burgulence. In M. Lesieur, A. Yaglom, and F. David, ed-
itors, New trends in turbulence Turbulence: nouveaux aspects, volume 74 of Les
Houches - Ecole d’Ete de Physique Theorique, pages 341–383. Springer Berlin
Heidelberg, 2001.
[80] J. Fröhlich, W. Rodi, Ph. Kessler, S. Parpais, J.P. Bertoglio, and D. Laurence.
Large eddy simulation of flow around circular cylinders on structured and un-
structured grids. In Ernst Heinrich Hirschel, editor, Numerical Flow Simulation
I, volume 66 of Notes on Numerical Fluid Mechanics (NNFM), pages 319–338.
Springer Berlin Heidelberg, 1998.
[81] E. Garnier, N. Adams, and P. Sagaut. Large Eddy Simulation for Compressible
Flows. Springer Heidelberg, Germany, 2009.
[82] E. Garnier, M. Mossi, P. Sagaut, P. Comte, and M. Deville. On the use of shock-
capturing schemes for large-eddy simulation. Journal of Computational Physics,
153(2):273 – 311, 1999.
[83] G. Gassner. A kinetic energy preserving nodal discontinuous galerkin spectral

element method. International Journal for Numerical Methods in Fluids, 2014.
[84] G. Gassner and A. Beck. On the accuracy of high-order discretizations for under-
resolved turbulence simulations. Theoretical and Computational Fluid Dynamics,
27(3-4):221–237, 2013.
[85] G. Gassner and D. A. Kopriva. A comparison of the dispersion and dissipation

errors of Gauss and Gauss-Lobatto discontinuous Galerkin spectral element meth-
ods. SIAM Journal of Scientific Computing, 33(5):2560–2579, October 2011.
[86] G. Gassner, F. Lörcher, and C.-D. Munz. A contribution to the construction of

diffusion fluxes for finite volume and discontinuous Galerkin schemes. Journal
of Computational Physics, 224(2):1049–1063, 2007.
198
Bibliography
[87] G. Gassner, F. Lörcher, and C.-D. Munz. A discontinuous Galerkin scheme based
on a space-time expansion. II. Viscous flow equations in multi dimensions. Jour-
nal of Scientific Computing, 34(3):260–286, 2008.
[88] G. Gassner, M. Torrilhon, A. Beck, S. Knechtel, and Th. Bolemann. Compari-

son of Navier-Stokes-Fourier equation and Grad’s moment equation solutions for
turbulence. In NIC Series Volume 47. Forschungszentrum Jülich GmbH, 2014.
[89] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamic subgrid-scale

eddy viscosity model. Physics of Fluids A: Fluid Dynamics, 3(7):1760–1765,
1991.
[90] B. J. Geurts. Elements of Direct and Large Eddy Simulation. R. T. Edwards Inc.,
USA, 2003.
[91] S. Ghosal, T. S. Lund, P. Moin, and K. Akselvoll. A dynamic localization model

for large-eddy simulation of turbulent flows. Journal of Fluid Mechanics, 286.
[92] Ph. M. Gresho and R. L. Lee. Don’t suppress the wiggles—they’re telling you
something! Computers & Fluids, 9(2):223 – 253, 1981.
[93] F. F. Grinstein and C. Fureby. Recent progress on MILES for high Reynolds
number flows. J. Fluids Eng., 124:848–861, 2002.
[94] F. F. Grinstein and C. Fureby. On monotonically integrated large eddy simula-

tion of turbulent flows based on FCT algorithms. In Dmitri Kuzmin, Rainald
Löhner, and Stefan Turek, editors, Flux-Corrected Transport, Scientific Compu-
tation, pages 79–104. Springer Berlin Heidelberg, 2005.
[95] R. Hartmann and P. Houston. Symmetric interior penalty DG methods for the
compressible Navier–Stokes equations I: Method formulation. Int. J. Num. Anal.
Model., 3(1):1–20, 2006.
[96] J. S. Hesthaven and T. Warburton. Nodal Discontinuous Galerkin Methods: Al-

gorithms, Analysis, and Applications. Springer Verlag, New York, 2008.
[97] S. Hickel. Implicit Turbulence Modeling for Large-Eddy Simulation. Dissertation,

Technische Universität München, Munich, Germany, 2008.
[98] S. Hickel and J. Larsson. An adaptive local deconvolution model for compressible
turbulence. In Proceedings of the 2008 Center for Turbulence Research Summer
Program, 2008.
199
Bibliography
[99] F. Hindenlang. Mesh Curving Techniques for High Order Parallel Simulations on
Unstructured Meshes. PhD thesis, University of Stuttgart, Germany, 2014.
[100] F. Hindenlang, G. Gassner, Ch. Altmann, A. Beck, M. Staudenmaier, and C.-D.

Munz. Explicit discontinuous Galerkin methods for unsteady problems. Comput-
ers & Fluids, 61(0):86 – 93, 2012.
[101] F. Hindenlang, G. Gassner, T. Bolemann, and C.-D. Munz. Unstructured high or-
der grids and their application in discontinuous Galerkin methods. In Proceedings
of ECCOMAS, 2010.
[102] M.-J. Huang and A. Leonard. Powerlaw decay of homogeneous turbulence at low
Reynolds numbers. Physics of Fluids (1994-present), 6(11), 1994.
[103] T. J. R. Hughes, L. P. Franca, and G. M. Hulbert. A new finite element formu-

lation for computational fluid dynamics: VIII. the Galerkin/least-squares method
for advective-diffusive equations. Computer Methods in Applied Mechanics and
Engineering, 73(2):173 – 189, 1989.
[104] M. Y. Hussaini and T. A. Zang. Spectral methods in fluid-dynamics. Annual

Review of Fluid Mechanics, 19:339–367, 1987.
[105] A. G. Hutton. The emerging role of large eddy simulation in industrial practice:
challenges and opportunities. Philosophical Transactions of the Royal Society A:
Mathematical, Physical and Engineering Sciences, 367(1899):2819–2826, 2009.
[106] T. Ishihara and Y. Kaneda. High resolution DNS of incompressible homoge-

neous forced turbulence - Time dependence of the statistics. In Yukio Kaneda and
Toshiyuki Gotoh, editors, Statistical Theories and Computational Approaches to
Turbulence, pages 177–188. Springer Japan, 2003.
[107] Y. Kaneda and T. Ishihara. Small-scale statistics in high-resolution direct nu-

merical simulation of turbulence. In Martin Oberlack, George Khujadze, Silke
Günther, Tanja Weller, Michael Frewer, Joachim Peinke, and Stephan Barth, edi-
tors, Progress in Turbulence II, volume 109 of Springer Proceedings in Physics,
pages 11–16. Springer Berlin Heidelberg, 2007.
[108] Y. Kaneda, T. Ishihara, M. Yokokawa, K. Itakura, and A. Uno. Energy dissipa-

tion rate and energy spectrum in high resolution direct numerical simulations of
turbulence in a periodic box. Physics of Fluids (1994-present), 15(2), 2003.
200
Bibliography
[109] G-S. Karamanos and G.E. Karniadakis. A spectral vanishing viscosity method for
large-eddy simulations. Journal of Computational Physics, 163(1):22 – 50, 2000.
[110] G. Kesserwani, R. Ghostine, J. Vazquez, A. Ghenaim, and R. Mosé. Riemann
solvers with Runge-Kutta discontinuous Galerkin schemes for the 1D shallow
water equations. Journal of Hydraulic Engineering, 134(2):243–255, 2008.
[111] R. M. Kirby and G. E. Karniadakis. De-aliasing on non-uniform grids: algorithms
and applications. Journal of Computational Physics, 191:249–264, 2003.
[112] R. M. Kirby and S. J. Sherwin. Stabilisation of spectral/hp element methods
through spectral vanishing viscosity: Application to fluid mechanics modelling.
Computer Methods in Applied Mechanics and Engineering, 195(23–24):3128 –
3144, 2006.
[113] C.M. Klaij, J. J. W. van der Vegt, and H. van der Ven. Spacetime discontinu-
ous Galerkin method for the compressible Navier-Stokes equations. Journal of
Computational Physics, 217(2):589–611, 2006.
[114] A. Klöckner, T. Warburton, and J. S. Hesthaven. Viscous shock capturing in a
time-explicit discontinuous Galerkin method. ArXiv e-prints, February 2011.
[115] K. Koal, J. Stiller, and H.M. Blackburn. Adapting the spectral vanishing viscos-
ity method for large-eddy simulations in cylindrical configurations. Journal of
Computational Physics, 231(8):3389 – 3405, 2012.
[116] A. N. Kolmogorov. Dissipation of Energy in the Locally Isotropic Turbulence.
Royal Society of London Proceedings Series A, 434:15–17, July 1991.
[117] A. N. Kolmogorov. The local structure of turbulence in incompressible viscous
fluid for very large Reynolds numbers. Royal Society of London Proceedings
Series A, 434:9–13, July 1991.
[118] D. Kopriva and G. Gassner. On the quadrature and weak form choices in colloca-
tion type discontinuous Galerkin spectral element methods. Journal of Scientific
Computing, 44(2):136–155, 2010.
[119] D. A. Kopriva. Metric identities and the discontinuous spectral element method
on curvilinear meshes. Journal of Scientific Computing, 26(3):301–327, 2006.
[120] David A. Kopriva. Implementing Spectral Methods for Partial Differential Equa-
tions: Algorithms for Scientists and Engineers. Springer Publishing Company
Incorporated, 1st edition, 2009.
201
Bibliography
[121] R. H. Kraichnan and D. Montgomery. Two-dimensional turbulence. Reports on

Progress in Physics, 43(5):547, 1980.
[122] A. G. Kravchenko and P. Moin. Numerical studies of flow over a circular cylinder
at ReD =3900. Physics of Fluids, 12:403–417, February 2000.
[123] A. G. Lamorgese, D. A. Caughey, and S. B. Pope. Direct numerical simulation of

homogeneous turbulence with hyperviscosity. Physics of Fluids (1994-present),
17(1), 2005.
[124] G. J. Le Beau and T. E. Tezduyar. Finite element computation of compressible

flows with the SUPG formulation. In Advances in a Finite Element Analysis in
Fluid Dynamics. American Society of Mechanical Engineering, 2001.
[125] O. Lehmkuhl, I. Rodrı́guez, R. Borrell, and A. Oliva. Low-frequency unsteadiness

in the vortex formation region of a circular cylinder. Physics of Fluids (1994-
present), 25(8), 2013.
[126] P. Lesaint and P.-A. Raviart. On a finite element method for solving the neutron
transport equation. In C.A. deBoor, editor, Mathematical Aspects of Finite El-
ements in Partial Differential Equations., pages 89–145. Academic Press, New
York, 1974.
[127] M. Lesieur, O. Metais, and P. Comte. Large Eddy Simulation of Turbulence.

Cambrigde University Press, United Kingdom, 2005.
[128] R. J. LeVeque. Finite Volume Methods for Hyperbolic Problems. Cambridge

University Press, United Kingdom, 2002.
[129] A. Lew, A. Ten Eyck, and R. Rangarajan. Some applications of discontinuous

Galerkin methods in solid mechanics. In B. Daya Reddy, editor, IUTAM Sympo-
sium on Theoretical, Computational and Modelling Aspects of Inelastic Media,
volume 11 of IUTAM BookSeries, pages 227–236. Springer Netherlands, 2008.
[130] G. Lodato, P. Castonguay, and A. Jameson. Structural LES modeling with high-
order spectral difference schemes. In Proceedings of the 2011 Center for Turbu-
lence Research Summer Program, 2011.
[131] F. Lörcher, G. Gassner, and C.-D. Munz. An explicit discontinuous Galerkin

scheme with local time-stepping for general unsteady diffusion equations. Journal
of Computational Physics, 227(11):5649–5670, 2008.
202
Bibliography
[132] E. N. Lorenz. Deterministic nonperiodic flow. Journal of the Atmospheric Sci-

ences, 20(2):130–141, Mar 1963.
[133] T. S. Lund. The use of explicit filters in large eddy simulation. Computers and
Mathematics with Applications, 46(4):603 – 616, 2003.
[134] D. A. Lysenko, I. S. Ertesvåg, and K. E. Rian. Large-eddy simulation of the flow
over a circular cylinder at Reynolds number 3900 using the openfoam toolbox.
Flow, Turbulence and Combustion, 89(4):491–518, 2012.
[135] X. Ma, G.-S. Karamanos, and G. E. Karniadakis. Dynamics and low-
dimensionality of a turbulent near wake. Journal of Fluid Mechanics, 410:29–65,
5 2000.
[136] J. Malm, Ph. Schlatter, P. F. Fischer, and D. S. Henningson. Stabilization of
the spectral element method in convection dominated flows by recovery of skew-
symmetry. Journal of Scientific Computing, 57(2):254–277, 2013.
[137] L. Marstorp, G. Brethouwer, and A. V. Johansson. A stochastic subgrid model
with application to turbulent flow and scalar mixing. Physics of Fluids (1994-
present), 19(3), 2007.
[138] P. J. Mason and D. J. Thomson. Stochastic backscatter in large-eddy simulations
of boundary layers. Journal of Fluid Mechanics, 242:51–78, September 1992.
[139] J. Mcdonough and R. Bywater. Effects of local large-scale parameters on the
small-scale chaotic solutions to Burgers’ equation. In Fluid Dynamics and Co-
located Conferences. American Institute of Aeronautics and Astronautics, 1985.
[140] S. Menon and S. Srinivasan. Challenges for multiscale large-eddy simulation of
application systems: Gas turbine to scramjet. In SCIDAC 2011.
[141] M. Meyer, S. Hickel, and N.A. Adams. Assessment of implicit large-eddy simu-
lation with a conservative immersed interface method for turbulent cylinder flow.
International Journal of Heat and Fluid Flow, 31(3):368 – 377, 2010.
[142] J. Meyers, B. J. Geurts, and P. Sagaut. A computational error-assessment of cen-
tral finite-volume discretizations in large-eddy simulation using a Smagorinsky
model. Journal of Computational Physics, 227(1):156–173, 2007.
[143] J. Meyers and P. Sagaut. On the model coefficients for the standard and the vari-
ational multi-scale Smagorinsky model. Journal of Fluid Mechanics, 569:387–
319, 2006.
203
Bibliography
[144] Y. Morinishi. Skew-symmetric form of convective terms and fully conservative

finite difference schemes for variable density low-mach number flows. Journal of
Computational Physics, 229(2):276 – 300, 2010.
[145] J. D. Murray. On Burgers’ model equations for turbulence. Journal of Fluid

Mechanics, 59:263–279, 6 1973.
[146] B. T. Nadiga and D. Livescu. Instability of the perfect subgrid model in

implicit-filtering large eddy simulation of geostrophic turbulence. Phys. Rev. E,
75:046303, Apr 2007.
[147] F. Nicoud and F. Ducros. Subgrid-scale stress modelling based on the square of
the velocity gradient tensor. Flow, Turbulence and Combustion, 3:183–200, 1999.
[148] J. A. Nitsche. Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen

bei Verwendung von Teilräumen, die keinen Randbedingungen unterworfen sind.
Abh. Math. Sem. Univ. Hamburg, 36:9–15, 1971.
[149] C. Norberg. An experimental investigation of the flow around a circular cylinder:

influence of aspect ratio. Journal of Fluid Mechanics, 258:287–316, 1994.
[150] A. M. Obukhov. On the distribution of energy in the spectrum of turbulent flow,.

Dokl. Akad. Nauk SSSR, 32:22–24, 1941.
[151] L. Ong and J. Wallace. The velocity field of the turbulent very near wake of a
circular cylinder. Experiments in Fluids, 20(6):441–453, 1996.
[152] S. A. Orszag. Numerical methods for the simulation of turbulence. Physics of

Fluids (1958-1988), 12(12), 1969.
[153] S. A. Orszag. On the elimination of aliasing in finite-difference schemes by

filtering high-wavenumber components. Journal of the Atmospheric Sciences,
28(6):1074–1074, Sep 1971.
[154] S. A. Orszag. Numerical simulation of the Taylor-Green vortex. In R. Glowinski

and J. L. Lions, editors, Computing Methods in Applied Sciences and Engineering
Part 2, volume 11 of Lecture Notes in Computer Science, pages 50–64. Springer
Berlin Heidelberg, 1974.
[155] S. A. Orszag and G. S. Patterson. Numerical simulation of three-dimensional

homogeneous isotropic turbulence. Phys. Rev. Lett., 28:76–79, Jan 1972.
204
Bibliography
[156] S. Ossia and M. Lesieur. Energy backscatter in large-eddy simulations of three-

dimensional incompressible isotropic turbulence. Journal of Turbulence, 2000.
[157] H. Ouvrard, B. Koobus, A. Dervieux, and M. V. Salvetti. Classical and varia-

tional multiscale LES of the flow around a circular cylinder on unstructured grids.
Computers & Fluids, 39(7):1083 – 1094, 2010.
[158] N. Park, J. Y. Yoo, and H. Choi. Discretization errors in large eddy simulation:
on the suitability of centered and upwind-biased compact difference schemes.
Journal of Computational Physics, 198(2):580 – 616, 2004.
[159] Ph. Parnaudeau, J. Carlier, D. Heitz, and E. Lamballais. Experimental and numer-
ical studies of the flow over a circular cylinder at Reynolds number 3900. Physics
of Fluids, 20:85–101, 2008.
[160] R. Pasquetti. Spectral vanishing viscosity method for large-eddy simulation of

turbulent flows. Journal of Scientific Computing, 27(1-3):365–375, June 2006.
[161] A. T. Patera. A spectral element method for fluid dynamics: Laminar flow in a
channel expansion. Journal of Computational Physics, 54(3):468 – 488, 1984.
[162] J. Peraire and P.-O. Persson. The compact discontinuous Galerkin (CDG) method
for elliptic problems. SIAM J. Sci. Comput., 30(4):1806–1824, 2008.
[163] P.-O. Persson. A sparse and high-order accurate line-based discontinuous

Galerkin method for unstructured meshes. Journal of Computational Physics,
233:414–429, January 2013.
[164] R. Peyret. Spectral Methods for Incompressible Viscous Flow. Springer Heidel-
berg, 2002.
[165] N. J. Phillips. An example of non-linear computational instability. In The Atmo-

sphere and the Sea in Motion, pages 501–504. Rockefeller Institute Press, 1959.
[166] U. Piomelli, W. H. Cabot, P. Moin, and S. Lee. Subgrid scale backscatter in

turbulent and transitional flows. Physics of Fluids A: Fluid Dynamics (1989-
1993), 3(7), 1991.
[167] S. B. Pope. Turbulent Flows. Cambridge University Press, 2000.
[168] S. B. Pope. Ten questions concerning the large-eddy simulation of turbulent flows.
New Journal of Physics, 6(1):35, 2004.
205
Bibliography
[169] J. Qiu. A numerical comparison of the Lax-Wendroff discontinuous Galerkin

method based on different numerical fluxes. Journal of Scientific Computing.
[170] J. Qiu, B. C. Khoo, and C.-W. Shu. A numerical study for the performance of the
Runge-Kutta discontinuous Galerkin method based on different numerical fluxes.
Journal of Computational Physics, 212:540 – 565, 2006.
[171] W. H. Reed and T. R. Hill. Triangular mesh methods for the neutron transport
equation. Technical Report LA-UR-73-479, Los Alamos Scientific Laboratory,
1973.
[172] J. Reiss and J. Sesterhenn. Conservative, skew–symmetric discretization of the

compressible navier–stokes equations. In Andreas Dillmann, Gerd Heller, Hans-
Peter Kreplin, Wolfgang Nitsche, and Inken Peltzer, editors, New Results in Nu-
merical and Experimental Fluid Mechanics VIII, volume 121 of Notes on Nu-
merical Fluid Mechanics and Multidisciplinary Design, pages 395–402. Springer
Berlin Heidelberg, 2013.
[173] O. Reynolds. On the dynamical theory of incompressible viscous fluids and the
determination of the criterion. Philosophical Transactions of the Royal Society of
London. (A.), 186:123–164, 1895.
[174] P. J. Roache. Code verification by the method of manufactured solutions. Journal

of Fluids Engineering, 124(1):4–10, Nov 2001.
[175] R. S. Rogallo. Numerical experiments in homogeneous turbulence. Technical

Report NASA TM 81215, 1981.
[176] R. S. Rogallo and P. Moin. Numerical simulation of turbulent flows. Annual

Review of Fluid Mechanics, 16(1):99–137, 1984.
[177] P. Sagaut. Large Eddy Simulation for Incompressible Flows. Springer Heidelberg,
Germany, 2006.
[178] R. Salmon. Lectures on Geophysical Fluid Dynamics. Oxford University Press,

USA, 1998.
[179] K. Sengupta, F. Mashayek, and G.B. Jacobs. Large-eddy simulation using a dis-
continuous Galerkin spectral element method. In 45th AIAA Aerospace Sciences
Meeting and Exhibit, 2007.
206
Bibliography
[180] T. K. Sengupta and M. T. Nair. Upwind schemes and large eddy simulation.
International Journal for Numerical Methods in Fluids, 31(5):879–889, 1999.
[181] J.-M. Senoner, M. Garcia, S. Mendez, G. Staffelbach, O. Vermorel, and

T. Poinsot. Growth of rounding errors and repetitivity of large eddy simulations.
AIAA Journal, 46(7):1773–1781, Jul 2008.
[182] L. Shtilman and J. R. Chasnov. LES versus DNS: A comparative study. In

D. Spinks, editor, Studying Turbulence Using Numerical Simulation Databases,
pages 137–143, November 1992.
[183] Ch.-W. Shu, W.-S. Don, D. Gottlieb, O. Schilling, and L. Jameson. Numerical
convergence study of nearly incompressible, inviscid Taylor-Green vortex flow.
Journal of Scientific Computing, 24(1):1–27, 2005.
[184] I. Silberman. Planetary waves in the atmosphere. Journal of Meteorology,

11(1):27–34, Feb 1954.
[185] B. Sjögreen and H.C. Yee. On skew-symmetric splitting and entropy conservation
schemes for the Euler equations. In Gunilla Kreiss, Per Lötstedt, Axel Målqvist,
and Maya Neytcheva, editors, Numerical Mathematics and Advanced Applica-
tions 2009, pages 817–827. Springer Berlin Heidelberg, 2010.
[186] J. Smagorinsky. General circulation experiments with the primitive equations.

Mon. Wea. Rev., 91:99–164, 1963.
[187] C. G. Speziale. Galilean invariance of subgrid-scale stress models in the large-

eddy simulation of turbulence. Journal of Fluid Mechanics, 156:55–62, 7 1985.
[188] S. Stolz and N. A. Adams. An approximate deconvolution procedure for large-

eddy simulation. Physics of Fluids (1994-present), 11(7), 1999.
[189] E. Tadmor. Convergence of spectral methods for nonlinear conservation laws.

SIAM Journal on Numerical Analysis, 26(1):30–44, Feb 1989.
[190] G. I. Taylor and A. E. Green. Mechanism of the production of small eddies from
large ones. Proceedings of the Royal Society of London. Series A, Mathematical
and Physical Sciences, 158(895):pp. 499–521, 1937.
[191] R. Temam. Approximation of attractors, large eddy simulations and multiscale

methods. Proceedings: Mathematical and Physical Sciences, 434(1890):pp. 23–
39, 1991.
207
Bibliography
[192] H. Tennekes and J. L. Lumley. A First Course in Turbulence. MIT Press, 1972.
[193] S. Tenneti and S. Subramaniam. Particle-resolved direct numerical simulation
for gas-solid flow model development. Annual Review of Fluid Mechanics,
46(1):199–230, 2014.
[194] T. E. Tezduyar and M. Senga. Stabilization and shock-capturing parameters in
SUPG formulation of compressible flows. Computer Methods in Applied Me-
chanics and Engineering, 195(13-16):1621 – 1632, 2006. A Tribute to Thomas
J.R. Hughes on the Occasion of his 60th Birthday.
[195] J. Thuburn, J. Kent, and N. Wood. Cascades, backscatter and conservation in
numerical models of two-dimensional turbulence. Quarterly Journal of the Royal
Meteorological Society, 140(679):626–638, 2014.
[196] E. F. Toro. Riemann Solvers and Numerical Methods for Fluid Dynamics.
Springer, June 1999.
[197] A. Tsinober. An Informal Conceptual Introduction to Turbulence. Springer
Netherlands, 2009.
[198] A. Uranga, P.-O. Persson, M. Drela, and J. Peraire. Implicit large eddy simula-
tion of transition to turbulence at low Reynolds numbers using a discontinuous
Galerkin method. International Journal for Numerical Methods in Engineering,
87(1-5):232–261, 2011.
[199] F. van der Bos and B. J. Geurts. Computational error-analysis of a discon-
tinuous Galerkin discretization applied to large-eddy simulation of homoge-
neous turbulence. Computer Methods in Applied Mechanics and Engineering,
199(13–16):903 – 915, 2010.
[200] Z.J. Wang, K. Fidkowski, R. Abgrall, F. Bassi, D. Caraeni, A. Cary, H. Deconinck,
R. Hartmann, K. Hillewaert, H.T. Huynh, N. Kroll, G. May, P.-O. Persson, B. van
Leer, and M. Visbal. High-order CFD methods: current status and perspective.
International Journal for Numerical Methods in Fluids, 72(8):811–845, 2013.
[201] T. C. Warburton and G. E. Karniadakis. A discontinuous Galerkin method for the
viscous MHD equations. Journal of Computational Physics, 152(2):608 – 641,
1999.
[202] L. Wei and A. Pollard. Direct numerical simulation of compressible turbulent
channel flows using the discontinuous Galerkin method. Computers & Fluids,
47(1):85 – 100, 2011.
208
Bibliography
[203] V. Wheatley, H. Kumar, and P. Huguenot. On the role of Riemann solvers in

discontinuous Galerkin methods for magnetohydrodynamics. Journal of Compu-
tational Physics, 229(3):660–680, 2010.
[204] D. C. Wilcox. Turbulence Modeling for CFD. D C W Industries, 2006.
[205] N. J. Zabusky and C. J. Galvin. Shallow-water waves, the Korteweg-deVries

equation and solitons. Journal of Fluid Mechanics, 47:811–824, 6 1971.
[206] J. Zhu, Y. T. Zhang, S. A. Newman, and M. Alber. Application of discontinu-

ous Galerkin methods for reaction diffusion systems in developmental biology.
Journal of Scientific Computing, 40:391–418, 2009.
209
List of Tables
2.1. L∞ -errors with respect to u

at time t = 0.25 with the pseudo-spectral method. . . . . . . . . . . . 43
2.2. L∞ -errors with respect to u
at time t = 0.25 with the FD method. . . . . . . . . . . . . . . . . . . 44
3.1. nP P Wmin (N, δ) for the Gauss DGSEM and the 6th order compact FD
for a given dispersion error δ and the degree N of the local polynomial
ansatz. Reproduced with permission from [85]. . . . . . . . . . . . . . 58
3.2. Estimated operation count for one-dimensional spatial operators for dif-
ferent discretizations. . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.3. Contributors and scheme details for the 1st International High Order
Workshop, test case 3.5, see Figure 3.9. . . . . . . . . . . . . . . . . . 86
3.4. Computational cost for DNS simulations with DGSEM. All computa-
tions were run on the “Hermit” Cray XE6 at HLRS, except the TGV
case, which was run on the “Jugene” BlueGene P at JSC. . . . . . . . . 91
4.1. Maximum values of the real parts of the eigenvalues of the DGSEM
operator (N = 15) spectrum with different number of integration points
for the polynomial advection velocity. . . . . . . . . . . . . . . . . . . 112
4.2. Maximum values of the real parts of the eigenvalues of the DG operator
(N = 15) spectrum with different number of integration points for the
sinusoidal advection velocity. . . . . . . . . . . . . . . . . . . . . . . 114
4.3. Integral quantities and simulation parameters for ReD = 3900 cylinder
flow, reproduced from [18]. CpBase : pressure coefficient at the down-
stream position x = D/2, y = 0, Str: Strouhal number of the lift
coefficient fLif t D/U∞ , CD : drag coefficient, Lr : length of separation
bubble. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.4. Computational cost for the cylinder flow computations. . . . . . . . . . 138
211
List of Tables
4.5. Simulation details for the cylinder flow with minimal Cs . CpBase : pres-
sure coefficient at the downstream position x = D/2, y = 0, Str:
Strouhal number of the lift coefficient fLif t D/U∞ , CD : drag coeffi-
cient, CL rms : RMS value of the lift coefficient, Lr : length of separation
bubble, θsep , θre : separation and reattachment location. . . . . . . . . 144
4.6. Integral quantities and simulation parameters for ReD = 3900 cylinder
flow, as a function of the Euler flux function. CpBase : pressure coef-
ficient at the downstream position x = D/2, y = 0, Str: Strouhal
number of the lift coefficient fLif t D/u∞ , CD : drag coefficient, Lr :
length of separation bubble. . . . . . . . . . . . . . . . . . . . . . . . 159
B.1. Simulation parameters for ReD = 3900 cylinder LES computations. . 175
212
List of Figures
1.1. Temporal evolution (from left to right) of two temperature bubbles in a

stably stratified medium. Shown are isocontours of temperature differ-
ence ∆θ = θ − θbackground . . . . . . . . . . . . . . . . . . . . . . . 2
2.1. Decaying homogeneous isotropic turbulence. Computations with pseudo-

spectral code Spex described in Section 3.2.1. Left: Temporal evolu-
tion of the spectrum of kinetic energy E(k). Dashed line denotes Kol-
mogorov’s k−5/3 law. Right: Temporal evolution of total kinetic energy
E. Dashed lines denote decay laws published by [53] and [102]. . . . . 12
2.2. Solution of the Lorenz System (Equation (2.5)) with a = 3, b = 25, c =
1. Blue lines: initial condition ~x0 = [1, 1, 2]; Black lines: initial con-
dition ~x0 = [1.00001, 1, 2]. Top: Solution trajectories up to tend = 10
(left) and tend = 90 (right). Green and red markers show the particle po-
sition of blue and black trajectories at tend . Bottom: Temporal evolution
of x-components of particle positions. . . . . . . . . . . . . . . . . . . 14
2.3. Vortex interactions in two dimensions, shown are contours of vorticity
magnitude. From left to right: initial random vorticity field at t = 0,
resulting fields at t = 100, 500, 10000. . . . . . . . . . . . . . . . . . 16
2.4. Interaction of two mono-scale vortices: Development of vortical struc-
tures and scale cascade. Shown are λ2 = −0.001 isocontours, colored
by normalized helicity. Temporal evolution from upper left to lower right. 17
2.5. Temporal evolution of the Taylor-Green vortex, from upper left to lower
right at time t = 0.4s, 1.4s, 2.7s, 5.9s, 8.9s and 15.5s Shown are con-
tours of vorticity, colored by relative helicity. . . . . . . . . . . . . . . 25
2.6. Integral quantities of the Taylor-Green vortex flow. Left: Spectrum of
kinetic energy for Re = 5000 at t = 9.0, dashed line: k−5/3 slope.
Right: Temporal evolution of the rate of kinetic energy dissipation for
Re = 1600. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
213
List of Figures
2.7. Spectrum of kinetic energy of isotropic homogeneous turbulence (Taylor-

Green Vortex at Re = 5000, t = 11s). I: Anisotropic large scale energy
production, II: Inertial subrange region, III: Small scale isotropic dissi-
pation range. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.8. Moving shock solution of the viscous Burgers’ equation. Left: DNS and
filtered solution at t = 0.25. Right: Corresponding energy spectra. . . . 42
2.9. Temporal evolution of the energy spectrum of the source S(x, t). . . . 42
2.10. Comparison of different approximate solutions to the v-problem at time
t = 0.25. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.11. Solution of the v-problem with the pseudo-spectral method. Left: Solu-
tion v computed with Nv = 64 grid points at t = 0.25. Right: Corre-
sponding energy spectrum. . . . . . . . . . . . . . . . . . . . . . . . 44
2.12. Coarse solution of the u-problem with Nu = 64. Left: Coarse solution
u at t = 0.25. Right: Plot of the corresponding energy spectrum. . . . . 45
2.13. Solution of the v-problem with the FD O4 method. Left: Solution vF D
computed with Nv = 128 grid points at t = 0.25 and comparison to the
DNS results. Right: Corresponding energy spectra. . . . . . . . . . . . 46
2.14. Smagorinsky model for laminar flow. Kinetic energy dissipation rate
over time as a function of the model constant Cs . . . . . . . . . . . . . 50
3.1. Global vs. local approximation of spatial derivatives. The regular grid
is denoted by circles, the boundary points by squares. Blue lines denote
the solution, red and yellow lines the approximate solution based on the
stencil indicated by the arrows. Vertical black lines indicate a subcell. . 57
3.2. Dispersion relation for the linear transport Equation (3.3) of DGSEM
with Gauss nodes, central finite differences and global spectral (exact)
scheme. K ∗ = NK+1 is the normalized wavenumber and Ω∗ = NΩ+1
is the corresponding modified normalized numerical mode (eigenvalue).
For FD, K ∗ = K and Ω∗ = Ω. . . . . . . . . . . . . . . . . . . . . . 59
3.3. Left: Ratio R of resolved scales (limit determined by phase error ei ) to
available scales on a NGrid = 1024 grid, ei=1,2,3 ∈ {0.1, 0.01, 0.001},
p: order of central FD discretization Right: Percentage of modes lost to
dispersion error 1 − R as a function of p for increasing number of grid
points. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4. Dissipation rate − dk
dt
of the kinetic energy of the Taylor-Green Vortex
at Re = 800. Results for the DNS data taken from Brachet [29]. The
required wall-clock time for each computation is indicated in [h]. . . . 62
214
List of Figures
3.5. Taylor-Green Vortex at Re = 1600. Comparison of the spectra of ki-

netic energy for N = 1, N = 3, de-aliased N = 7 and de-aliased
N = 15 solution. All discretizations with a total of 643 DOF, except DNS. 63
3.6. L2 error of density over grid spacing h for the pseudo-spectral code
Spex, obtained with the method of manufactured solutions. . . . . . . . 69
3.7. Spectrum of kinetic energy for the Taylor-Green vortex at Re = 1600

and t = 9s. Comparison of pseudo-spectral codes Spex and Louvain and
DGSEM code Flexi. . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.8. Strong scaling of DGSEM on BlueGene Q “JuQueen” at JSC. 1283 el-

ements, N = 7 up to 262, 144 ranks (131, 072 cores with 2× hyper-
threading). Results from the Porting and Scaling Workshop, 2013. . . . 85
3.9. Code comparison for DNS of Taylor-Green Vortex at Re = 1600. Top:

Error in computed vs. reference dissipation rate over DOF. Bottom: Er-
ror in computed vs. reference dissipation rate over Work Units (CPU
time). Results from the 1st International Workshop on High Order Meth-
ods in CFD, 2012, with permission from Koen Hillewaert (Cenaero) and
from [200]. For legend refer to Table 3.3. . . . . . . . . . . . . . . . . 88
3.10. Flow over a sphere at Re = 1000 and M a = 0.3. The computation was
run the Cray XE6 “Hermit” at HLRS on 4096 cores. The picture shows
the laminar separation, the vortex street and the transition in the wake
(λ2 = −0.01 criterion, colored by vorticity magnitude). Computation
by F. Hindenlang [100]. . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.11. Vortex Cascade at Re = 1000. Left: λ2 isocontours, colored by helicity.

Right: Temporal development of energy spectrum. . . . . . . . . . . . 94
3 3
3.12. Taylor-Green Vortex at Re = 5000. Computation with 384 and 600
DOF on 125, 000 cores on BlueGene P (“Jugene”) at JSC. Left: Iso-
contours of vorticity magnitude, colored by helicity. Right: Spectrum of
kinetic energy for 3843 and 6003 DOF, corresponding cut-off wavenum-
bers denoted by dashed lines. . . . . . . . . . . . . . . . . . . . . . . 94
3.13. Flow over a circular cylinder at M a = 0.1, Re = 3900. Computation

with 164 mio DOF on 4, 992 cores on Cray XE6 “Hermit” at HLRS.
λ2 = −0.001 criterion, colored by vorticity magnitude. Top: side view.
Bottom: Top view of near wake. . . . . . . . . . . . . . . . . . . . . . 95
215
List of Figures
4.1. Dispersion (left) and dissipation error (right) for DGSEM with Gauss
nodes. K ∗ = NK+1 is the normalized wavenumber and Ω∗ = NΩ+1
is the corresponding modified normalized numerical mode (eigenvalue).
Bottom row: logarithmic plot with zoom on low wavenumber region.
Note that the error is cut off at 10−10 to avoid numerical noise. Plots
with permission from [85]. . . . . . . . . . . . . . . . . . . . . . . . 101
4.2. Taylor-Green Vortex at Re = 1000. Effect of aliasing on flow physics.
Left: Mean temperature and kinetic energy dissipation rate over time.
Right: Maximum vorticity over time. . . . . . . . . . . . . . . . . . . 107
4.3. Operator spectrum for constant advection speed and N = 15 with either
M + 1 = 16 or 26 integration points, respectively. The maximum of the
real parts of the eigenvalues is 0. . . . . . . . . . . . . . . . . . . . . 110
4.4. Operator spectrum for polynomial advection speed and N = 15 with
M + 1 = 16 integration points (original DGSEM scheme). The maxi-
mum of the real parts of the eigenvalues is 0.016. . . . . . . . . . . . . 111
4.5. Operator spectrum for polynomial advection speed and N = 15 with
different number of integration points M + 1. The right plot shows a
zoomed view of the imaginary axis. . . . . . . . . . . . . . . . . . . . 112
4.6. Operator spectrum for sinusoidal advection speed, Equation (4.14a), and
N = 15 with different number of integration points M + 1. The right
plot shows a zoomed view of the imaginary axis. . . . . . . . . . . . . 113
4.7. Operator spectrum for sinusoidal advection speed, Equation (4.14b), and
N = 15 with different number of integration points M + 1. The right
plot shows a zoomed view of the imaginary axis. . . . . . . . . . . . . 114
4.8. Plot of the kinetic energy dissipation rates of the Taylor-Green Vortex
(Left: Re = 800, Right: Re = 1600), computed with with 643 DOF
using either a low order method (2nd order, N = 1) or a very high order
method (16th order, N = 15). . . . . . . . . . . . . . . . . . . . . . . 121
4.9. Transfer functions for the weak and strong variant of the modal filters
for N = 15. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.10. Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk dt
for filter-stabilized N = 15 (43 grid cells) computation. The right plot
is zoomed in on the time range of maximum enstrophy production. . . . 123
4.11. Taylor-Green Vortex at Re = 800. Left: Kinetic energy spectra at t = 9.
Right: Kinetic energy dissipation rate − dk dt
, the resolved (molecular)
dissipation rate and the numerical dissipation rate for filter-stabilized
N = 15 (43 grid cells) computation. . . . . . . . . . . . . . . . . . . 124
216
List of Figures

zoomed in on the time range of maximum enstrophy production. . . . . 125
4.13. Taylor-Green Vortex at Re = 1600. Left: Kinetic energy spectra at
t = 9. Right: Kinetic energy dissipation rate − dk dt
, the resolved (molec-
ular) dissipation rate and the numerical dissipation rate for filter-stabilized
N = 15 (4 grid cells) computation. . . . . . . . . . . . . . . . . . . . 125
for PDG N = 15 (4 grid cells) computation. The right plot is zoomed
in on the time range of maximum enstrophy production. . . . . . . . . 127
4.15. Taylor-Green Vortex at Re = 800. Left: Kinetic energy spectra at t = 9.
Right: Kinetic energy dissipation rate − dk dt
, the resolved (molecular)
dissipation rate and the numerical dissipation rate for PDG N = 15 (4
grid cells) computation. . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.17. Taylor-Green Vortex at Re = 1600. Left: Kinetic energy spectra at
t = 9. Right: Kinetic energy dissipation rate − dk dt
, the resolved (molec-
ular) dissipation rate and the numerical dissipation rate for PDG N = 15
(43 grid cells) computation. . . . . . . . . . . . . . . . . . . . . . . . 129
4.18. Taylor-Green Vortex at Re = 800. Comparison of kinetic energy dis-
sipation and numerical dissipation (left) and energy spectra (right) for
N = 15 (43 grid cells) computations with stabilization via polynomial
de-aliasing vs. filtering. . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.19. Taylor-Green Vortex (Re = 800). Isocontours of λ2 = −1.5 at t = 8.5.
All computations with 643 DOF except DNS. From upper left to lower
right: i) 16th order filtered computation, ii) 16th order computation with
polynomial de-aliasing, iii) 2nd order computation, iv) DNS reference
result with 3843 DOF. . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.20. Comparison of LES methods: Kinetic energy dissipation rate for Re ∈
{800, 1600, 3000} for PDG computations with N = 15 compared to
DNS and LES reference data. . . . . . . . . . . . . . . . . . . . . . . 133
4.21. Time- and spanwise-averaged streamwise velocity component and grid.
Left: N = 11 grid with 8 × 8 × 6 elements (circumferential, radial and
spanwise), right: N = 7 grid with 16 × 16 × 12 elements. . . . . . . . 135
4.22. Instantaneous flow field in the cylinder region. Left: u-velocity, middle:
Modal resolution indicator, right: Difference between projection and
collocation of u2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
217
List of Figures
4.23. Mean streamwise velocity at different downstream locations in the wake

of the circular cylinder flow at ReD = 3900. Left: N = 11 com-
putation, Right: N = 7 computation. Solid lines: present DG results,
dashed lines: computational LES results from [141], open squares: com-
putational LES results from [122], filled circles: experimental results
from [159]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.24. Kinetic energy dissipation rate for the Taylor-Green Vortex with increas-
ing Reynolds number: 83 Elements, N = 7. M + 1 denotes the
number of quadrature points for the non-linear fluxes, Cs indicates the
Smagorinsky model constant and is 0.0 unless stated otherwise. . . . . 140
4.25. Grid for the DNS computation, detailed view of cylinder vicinity. . . . 142
4.26. From left to right: Grids I, II and III for the LES computations, detailed
view. Grid details: I: 16 × 16 × 32 elements (circumferential, radial and
spanwise direction), N = 7, II: 22 × 22 × 44, N = 5, III: 32 × 32 ×
64, N = 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.27. Time- and spanwise-averaged streamlines. Upper left plot: DNS. Rest:
a) N7-1, b) N7-A, c) N5-1, d) N5-A, e) N3-1, f) N3-A . . . . . . . . . 145
4.28. Time- and spanwise averaged velocity fluctuations hu′ u′ i in the cylin-
der wake. Line with squares: DNS results; Left: DGSEM according to
Table 4.5. Dashed line: N7-A; dotted line: N5-A; dash-dotted line: N3-
A; right: PDG according to Table 4.5. Dashed line: N7-1; dotted line:
N5-1; dash-dotted line: N3-1. . . . . . . . . . . . . . . . . . . . . . . 146
4.29. One-dimensional spectra of the streamwise velocity component in the
Strouhal frequency of the reference DNS (see Table 4.5). a) N7-1, b)
N7-A, c) N5-1, d) N5-A, e) N3-1, f) N3-A; dashed line denotes Kol-
mogorov’s k−5/3 law. . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.30. Time- and spanwise-averaged velocity fluctuations hu′ u′ i in the wake
2
region, normalized by U∞ . Top row: DNS reference. First column:
PDG results, second column: PDG with Smagorinsky model, last col-
umn: DGSEM with Smagorinsky model; second row: N7-1, N7-2, N7-
A, third row: N5-1, N5-2, N5-B, fourth row: N3-1, N3-4, N3-D. . . . . 150
4.31. Kinetic energy dissipation rate for the Taylor-Green Vortex at Re =
1600: Comparison of convective numerical flux function influence for
different polynomial degrees N with same total no. of DOF (643 ).
Square symbol denotes the reference DNS solution [28], dashed line
denotes the LLF flux result, the solid line denotes the Roe flux result.
Plot reproduced from [20]. . . . . . . . . . . . . . . . . . . . . . . . 155
218
List of Figures
4.32. Kinetic energy dissipation rate for the Taylor-Green Vortex at Re =

3000: Comparison of LDG flux and BR1 flux with DNS, 43 elements,
polynomial degree N = 15. Right: Detailed view. Plot reproduced
from [20]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
4.33. Optimization of the penalty parameters α1 and α2 for the kinetic en-
ergy dissipation rate. Shown is the L2 deviation between DNS and LES
results as a function of α1 and α2 . . . . . . . . . . . . . . . . . . . . 156
4.34. Kinetic energy dissipation rate for the Taylor-Green Vortex: Left: Re =
800: Comparison of classical Roe flux and modified Roe flux with DNS.
Right: Re = 1600: Comparison of modified combined flux and standard
fluxes with DNS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.35. Comparison of Cylinder LES results for Roe and Rusanov-flux for com-
putation described in Section 4.3.2. Left: Pressure coefficient along ge-
ometry. Filled circles denote experimental data for Re = 5000 of Son &
Hanratty for vorticity and for Re = 4020 of Norberg for pressure coeffi-
cient (taken from [141]). Open squares denote DNS results [135]. Right:
Mean crossflow velocity at different downstream locations in the wake of
the circular cylinder flow at ReD = 3900, N = 11 computation. Solid
black lines: DGSEM with Rusanov flux. Solid blue lines: DGSEM with
Roe flux. Dashed lines: computational LES results from [141]. Open
squares: computational LES results from [122]. Filled circles: experi-
mental results from [159]. . . . . . . . . . . . . . . . . . . . . . . . . 159
5.1. Kinetic energy spectra of HIT at Reλ = 82. DNS (solid line) computed
with Spex on 5123 nodes. Dashed line denotes PDG result, solid line
with markers denotes LAD result, both with 723 DOF (N = 11). Data
taken from [74]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
A.1. Coefficients α (left) and β (right) in wave space for a box spectrum
E(k) = const for 4 ≤ k ≤ 12, E(k) = 0 otherwise. . . . . . . . . . . 170
A.2. Scatter plot of u-v and v-w correlations from Rogallo’s procedure (Equa-
tions (A.1) and (A.2)). . . . . . . . . . . . . . . . . . . . . . . . . . . 170
B.1. Time- and spanwise averaged velocity fluctuations hu′ v ′ i in the cylin-
der wake. Line with squares: DNS results; left: DGSEM according to
Table 4.5. Dashed line: N7-A; dotted line: N5-A; dash-dotted line: N3-
A. right: PDG according to Table 4.5. Dashed line: N7-1; dotted line:
N5-1; dash-dotted line: N3-1. . . . . . . . . . . . . . . . . . . . . . . 171
219
List of Figures
B.2. One-dimensional spectra of the crossflow velocity component in the

Strouhal frequency of the reference DNS (see Table 4.5). a) N7-1, b)
N7-A, c) N5-1, d) N5-A, e) N3-1, f) N3-A; dashed line denotes Kol-
mogorov’s k−5/3 law. . . . . . . . . . . . . . . . . . . . . . . . . . . 172
B.3. Time- and spanwise-averaged velocity fluctuations hu′ v ′ i in the wake
region, normalized by square of the free-stream velocity U∞ . Top row:
DNS result. Second to fourth row: results for N7, N5 and N3 cases, with
PDG in the left and DGSEM in the right column. . . . . . . . . . . . . 173
′ ′
B.4. Time- and spanwise-averaged crossflow velocity fluctuations hv v i in
the wake region, normalized by square of the free-stream velocity U∞ .
Top row: DNS result. Second to fourth row: results for N7, N5 and N3
cases, with PDG in the left and DGSEM in the right column. . . . . . . 174
B.5. Time- and spanwise-averaged turbulent kinetic energy in the wake re-
gion, normalized by free-stream kinetic energy. First row: DNS ref-
erence results, first column: PDG results, second column: PDG with
Smagorinsky model, last column: DGSEM with Smagorinsky model;
Second row: N7-1, N7-2, N7-A, third row: N5-1, N5-2, N5-B, fourth
row: N3-1, N3-4, N3-D. . . . . . . . . . . . . . . . . . . . . . . . . . 176
B.6. Time- and spanwise-averaged velocity magnitude in the wake region,
normalized by free-stream kinetic energy. First row: DNS reference re-
sults, first column: PDG results, second column: PDG with Smagorin-
sky model, last column: DGSEM with Smagorinsky model; Second row:
N7-1, N7-2, N7-A, third row: N5-1, N5-2, N5-B, fourth row: N3-1, N3-
4, N3-D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
B.7. Isocontours of time- and spanwise-averaged streamwise velocity: u =
0. Shaded area denotes the DNS reference solution. . . . . . . . . . . 178
C.1. Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk dt

C.2. Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate − dk
dt
C.3. Taylor-Green Vortex at Re = 800. Kinetic energy dissipation rate, re-
solved dissipation rate and numerical disspation rate for PDG N = 7
(left) and filter-stabilized DGSEM N = 7 computation (right). . . . . . 180
220
List of Figures

C.6. Taylor-Green Vortex at Re = 1600. Kinetic energy dissipation rate,
resolved dissipation rate and numerical disspation rate for PDG N = 7
(left) and filter-stabilized DGSEM N = 7 computation (right). . . . . . 182
C.7. Taylor-Green Vortex at Re = 1600: Comparison of the resolved (left)
and total dissipation rate (right) for the contributions to the high order
workshop with highest order per participant. Refer to Table 3.3 for the
legend. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
C.8. Taylor-Green Vortex at Re = 1600: Comparison of error measures ac-
cording to Equation C.2 for the contributions to the high order workshop.
For a legend refer to Table 3.3. . . . . . . . . . . . . . . . . . . . . . 184
C.9. Taylor-Green Vortex at Re = 1600: Visualization of vortex detection
criterion λ2 = −1.5 for N = 1 (top), N = 3 (bottom left) and N = 15
(bottom right) computation (2563 DOF in each case). . . . . . . . . . . 185
221
Lebenslauf
05.01.1979 Geboren in Trier
1985 - 1989 Grundschule Wincheringen
1989 - 1998 Angela-Merici Gymnasium der Ursulinen, Trier
1998 Allgemeine Hochschulreife
1998 - 2002 Studium der Luft- und Raumfahrttechnik
an der Universität Stuttgart, Vertiefungsrichtungen:
Strömungslehre, Raumfahrttechnik
2002 - 2004 Studium am Georgia Institute of Technology, USA,
Vertiefungsrichtung Aerodynamics & Fluid Mechanics
2004 Abschluß Master of Science in Aerospace Engineering
2009 - 2015 Wissenschaftliche Mitarbeiterin am
Institut für Aerodynamik und Gasdynamik
Stuttgart, den 1. Mai 2015 Andrea Beck
223

Dissertation V4 Embedded Gedreht

Uploaded by

Copyright:

Available Formats

Dissertation V4 Embedded Gedreht

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dissertation V4 Embedded Gedreht

Uploaded by

Copyright:

Available Formats

High Order Discontinuous Galerkin Methods for

the Simulation of Multiscale Problems

A thesis accepted by the Faculty

Main referee: Prof. Dr. Claus-Dieter Munz

Institute of Aerodynamics and Gas Dynamics

Stuttgart, May 1, 2015

Symbols and Abbreviations xi

3. Numerics for Scale-Resolving Simulations 55

4. Numerics for Under-Resolved Simulations 97

5. Conclusion and Prospects 163

A. Initial Conditions for Homogeneous Isotropic Turbulence 169

B. Flow over a Cylinder at ReD = 3900 - Extended Results 171

C. Taylor-Green Vortex - Extended Results 179

D. Central Finite Differences of Arbitrary Order 187

E. Derivation of the Vorticity Equation 189

List of Tables 211

List of Figures 213

Cs Smagorinsky model constant

1D, 2D, 3D one, two, three dimensions / -dimensional

ODE Ordinary Differential Equation

E(k) ∼ ǫ2/3 k−5/3 , (1.1)

1.1. Multiscale Problems: Simulation Strategies

1.2. Multiscale Problems: Modeling

1.3. Multiscale Problems: Numerics

Ntotal ∼ n4P P W Re3 . (1.5)

2.1.1. Phenomenology of Turbulent Flows

effects on regions of high curvature (i.e. small, high-frequency eddies of sizes on

Here, Re = ν1 is the Reynolds number based on the kinematic viscosity ν > 0

Figure 2.1.: Decaying homogeneous isotropic turbulence. Computations with pseudo-

• Spatio-temporal irregularity and loss of predictability

• Three-dimensionality and vorticity

For the vortex stretching mechanism to be non-zero, two conditions must be

Returning to a discussion about the vortex stretching mechanism that increases

Figure 2.4.: Interaction of two mono-scale vortices: Development of vortical structures

so the angular momentum is conserved according to

• High diffusion and dissipation

2.1.2. Governing Equations

2.1.2.1. Compressible Navier-Stokes Equations

σij = µ(T )Sij , (2.13)

where the rate of strain tensor is given by

∂ui ∂uj ∂uk

2.1.2.2. Incompressible Navier-Stokes Equations

For a fluid of constant density, Equation (2.10) reduces to

2.1.2.3. Incompressible Navier-Stokes Equations in Spectral Space

Transforming the continuity condition of Equation (2.17) yields

with the operator P that projects onto the plane orthogonal to ~k as

2.2. A Model Turbulence Problem: The Taylor-Green Vortex

Figure 2.6.: Integral quantities of the Taylor-Green vortex flow.

2.3. Numerical Simulation Strategies for Multiscale Flows

2.3.3. Large Eddy Simulation

1. Computational necessity: As noted in Section 1.3, the space-time scale resolution

2. Structural stability of the filtered solution: The concept of a solution trajectory

Figure 2.7.: Spectrum of kinetic energy of isotropic homogeneous turbulence (Taylor-

universal small scales amenable to modeling becomes more valid as well.

2.4. Aspects of Large Eddy Simulation

2.4.1. General Formalism

1. How good does ūh approximate u?

2. How good does ūh approximate ū?

P (u) + M (u) = 0, (2.41)