POLITECNICO DI TORINO

SCUOLA DI DOTTORATO
Dottorato in Fluidodinamica – XXVII ciclo

Tesi di Dottorato

Computational fluid dynamics for

aerospace propulsion systems: an
approach based on discontinuous
finite elements

Andrea Ferrero
matricola: 189152

Tutor Coordinatore del corso di dottorato

prof. Francesco Larocca prof. Gaetano Iuso

December 2014
Summary

The purpose of this work is the development of a numerical tool devoted to the
study of the flow field in the components of aerospace propulsion systems. The
goal is to obtain a code which can efficiently deal with both steady and unsteady
problems, even in the presence of complex geometries.
Several physical models have been implemented and tested, starting from Euler
equations up to a three equations RANS model. Numerical results have been com-
pared with experimental data for several real life applications in order to under-
stand the range of applicability of the code. Performance optimization has been
considered with particular care thanks to the participation to two international
Workshops in which the results were compared with other groups from all over the
world.
As far as the numerical aspect is concerned, state-of-art algorithms have been im-
plemented in order to make the tool competitive with respect to existing softwares.
The features of the chosen discretization have been exploited to develop adaptive
algorithms (p, h and hp adaptivity) which can automatically refine the discretiza-
tion. Furthermore, two new algorithms have been developed during the research
activity. In particular, a new technique (Feedback filtering [1]) for shock capturing
in the framework of Discontinuous Galerkin methods has been introduced. It is
based on an adaptive filter and can be efficiently used with explicit time integra-
tion schemes. Furthermore, a new method (Enhance Stability Recovery [2]) for
the computation of diffusive fluxes in Discontinuous Galerkin discretizations has
been developed. It derives from the original recovery approach proposed by van
Leer and Nomura [3] in 2005 but it uses a different recovery basis and a different
approach for the imposition of Dirichlet boundary conditions. The performed nu-
merical comparisons showed that the ESR method has a larger stability limit in
explicit time integration with respect to other existing methods (BR2 [4] and origi-
nal recovery [3]). In conclusion, several well known test cases were studied in order
to evaluate the behavior of the implemented physical models and the performance
of the developed numerical schemes.

iii
Acknowledgements

I’m very grateful to my supervisor, Professor Francesco Larocca for all the time he
dedicated to my doubts and questions.
I would like to thank my family for the support and Francesca for her patience and
comprehension.

v
Contents

Summary iii

Acknowledgements v

I A numerical approach to aerospace propulsion 1

1 Introduction 3
1.1 CFD for aerospace propulsion systems . . . . . . . . . . . . . . . . 3
1.2 Motivations and description of the work . . . . . . . . . . . . . . . 4

2 Physical models 9
2.1 Euler equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Navier-Stokes equations . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Reynolds Averaged Navier-Stokes equations . . . . . . . . . . . . . 11
2.3.1 Spalart-Allmaras model . . . . . . . . . . . . . . . . . . . . 12
2.3.2 Wilcox k − ω model . . . . . . . . . . . . . . . . . . . . . . 14
2.3.3 Laminar Kinetic Energy model . . . . . . . . . . . . . . . . 20
2.4 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 BCs for Euler equations . . . . . . . . . . . . . . . . . . . . 23
2.4.2 BCs for Navier-Stokes equations . . . . . . . . . . . . . . . 23
2.4.3 BCs for Spalart-Allmaras model . . . . . . . . . . . . . . . 24
2.4.4 BCs for Wilcox k − ω model . . . . . . . . . . . . . . . . . 24
2.4.5 BCs for LKE model . . . . . . . . . . . . . . . . . . . . . . 26

II Discretization 27
3 Discontinuous Galerkin spatial discretization 29
3.1 Variational approach . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Basis functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

vii
 3.3 Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Quadrilaterals . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.2 Triangles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Choice of the basis . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Convective fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4 Diffusive fluxes 41
4.1 BR2 method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Original recovery method . . . . . . . . . . . . . . . . . . . . . . . 44
4.3 Enhanced stability recovery method . . . . . . . . . . . . . . . . . 46
4.3.1 Derivation of the method . . . . . . . . . . . . . . . . . . . 46
4.3.2 Extension to 3D . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.3 BCs for ESR . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.4 Dirichlet BC’s in strong form . . . . . . . . . . . . . . . . . 53
4.3.5 Dirichlet BCs in weak form . . . . . . . . . . . . . . . . . . 55
4.3.6 Implementation and computational cost . . . . . . . . . . . 56
4.3.7 Interface integral computational cost . . . . . . . . . . . . . 56
4.3.8 Volume integral computational cost . . . . . . . . . . . . . 57
4.4 Tests and comparisons for the ESR method . . . . . . . . . . . . . 58
4.4.1 1D Poisson equation . . . . . . . . . . . . . . . . . . . . . . 58
4.4.2 1D Convection Diffusion problem . . . . . . . . . . . . . . . 59
4.4.3 2D Heat equation . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4.4 Laminar boundary layer on a flat plate . . . . . . . . . . . . 63
4.5 Considerations on implicit time integration . . . . . . . . . . . . . 65

5 Time discretization 67
5.1 Explicit Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . 68
5.2 Implicit methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 Jacobian evaluation . . . . . . . . . . . . . . . . . . . . . . 70
5.2.2 Iterative solver . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2.3 CFL evolution laws for steady problems . . . . . . . . . . . 71

6 Shock capturing 73
6.1 Feedback filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.1.1 Exponential filter . . . . . . . . . . . . . . . . . . . . . . . . 74
6.1.2 Smoothness indicator . . . . . . . . . . . . . . . . . . . . . 75
6.1.3 Feedback filtering . . . . . . . . . . . . . . . . . . . . . . . . 76
6.1.4 Sod problem . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.1.5 Shock-bubble interaction . . . . . . . . . . . . . . . . . . . 82
6.1.6 Ringleb flow . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.1.7 Transonic NACA0012 . . . . . . . . . . . . . . . . . . . . . 86
6.2 Artificial viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

viii
 6.2.1 Future works . . . . . . . . . . . . . . . . . . . . . . . . . . 90

7 Adaptive algorithms 93
7.1 Sensors and error indicators . . . . . . . . . . . . . . . . . . . . . . 94
7.2 p-adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.2.1 Heat equation and p-adaptivity . . . . . . . . . . . . . . . . 96
7.2.2 Vortex shedding around a circular cylinder . . . . . . . . . 99
7.3 h-adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.3.1 Transonic NACA0012: h-adaptivity . . . . . . . . . . . . . 103
7.4 hp-adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 108

III Numerical results 111

8 Performances 113
8.1 Acceleration for steady problems . . . . . . . . . . . . . . . . . . . 113
8.2 Workshop test problems . . . . . . . . . . . . . . . . . . . . . . . . 116
8.2.1 Subsonic flow in a channel with a smooth bump . . . . . . 116
8.2.2 Transonic Ringleb flow . . . . . . . . . . . . . . . . . . . . . 119
8.2.3 Vortex transport by uniform flow . . . . . . . . . . . . . . . 123
8.2.4 Unsteady viscous flow over tandem airfoils . . . . . . . . . . 128

9 Aerospace propulsion examples and validation 131

9.1 VKI-LS59 turbine cascade . . . . . . . . . . . . . . . . . . . . . . . 131
9.2 T106c turbine cascade . . . . . . . . . . . . . . . . . . . . . . . . . 132

10 Conclusions 141

A Adimensionalization 143

B Normalization of performances 145

C High-order visualizations 147

D Meshes 149

E List of symbols 151

Bibliography 155

ix
List of Tables

4.1 Condition number of the matrix [R]T [R] for the interface of Fig-

ure 4.4 without and with orthonormalization (ON) of the recovery

basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 1D Poisson problem: stability limit (p=1, forward Euler) . . . . . 59
4.3 1D Poisson problem: L2 errors and convergence order (p=1) . . . . 59
4.4 1D Convection-diffusion problem: L2 errors at steady state (p=1) . 60
4.5 L2 -norm of the temperature error for the ESR and the BR2 methods. 62
4.6 Stability limit with explicit time integration for 2D heat diffusion
on unstructured meshes. . . . . . . . . . . . . . . . . . . . . . . . . 62
4.7 Drag coefficient (CD ) for flate plate problem (Re∞ = 106 , M∞ =
0.5 − 0.2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.1 Shock-bubble problem: initial data. . . . . . . . . . . . . . . . . . . 82
6.2 Ringleb flow: entropy accuracy test for feedback filtered DG methods. 86
6.3 Ringleb flow: pressure accuracy test for feedback filtered DG methods. 87
7.1 Average drag coefficient and Strouhal number for cylinder flow (Re∞ =
102 , M∞ = 0.2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2 Experimental results from the literature for cylinder flow (Re∞ = 102 )100

x
List of Figures

2.1 Velocity profile in turbulent boundary layer obtained by the pro-

posed discretization with the Spalart-Allmaras model . . . . . . . . 15
2.2 Skin friction coefficient for turbulent flat plate: comparison between
the implemented Spalart-Allmaras model and correlations from the
literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Skin friction coefficient on flat plate: transition from laminar to
turbulent. Blue=Blasius solution, Light blue=turbulent correlation,
Green= implemented Wilcox k − ω low Re model . . . . . . . . . 19
3.1 Serendipity mappings for quadrilateral elements. . . . . . . . . . . 33
3.2 Mach field for the inviscid subsonic flow in a channel. . . . . . . . 36
3.3 Unstructured mesh with cubic quadrilateral elements. . . . . . . . 37
3.4 Unstructured mesh with cubic triangular elements. . . . . . . . . . 37
3.5 Entropy error vs equivalent length scale for the inviscid channel flow
(p=2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.6 Entropy error vs reconstruction order for Osher and Local Lax-
Friedrichs (LLF) fluxes. . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1 Interface reference frame for constructing the recovery basis. . . . . 48
4.2 Preliminary recovery basis before orthonormalization for the case
pA = pB = 2, in the Enhanced Recovery approach. . . . . . . . . . 48
4.3 Preliminary recovery basis with the original recovery approach for
the case pA = pB = 2. . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4 Interface between two distorted elements for the test of Table 4.1. . 51
4.5 Exact temperature field for the 2D heat diffusion problem. . . . . . 61
4.6 L2 -norm of the temperature error vs the equivalent length scale h. 63
4.7 Wall friction coefficient (cf ) for the flat plate (Re∞ = 106 , M∞ = 0.2) 64
4.8 Adimensionalized L2 -norm of the residual vs computational time [s]. 66
6.1 Adimensionalized L2 -norm of the residual vs computational time [s]. 80
6.2 Adimensionalized L2 -norm of the residual vs computational time [s]. 80
6.3 Detail of the contact surface in the Sod problem (Feedback filtering). 81
6.4 Detail of the shock in the Sod problem (Feedback filtering). . . . . 81
6.5 Domain and initial condition for shock-bubble interaction . . . . . 82

xi
6.6 Numerical Schlieren images for shock-bubble interaction at t = 0.071,
t = 0.21 and t = 0.33 (Feedback filtered DG2 scheme). . . . . . . . 83
6.7 Numerical Schlieren image for shock-bubble interaction at t = 0.18
(Feedback filtered DG2 scheme). . . . . . . . . . . . . . . . . . . . 84
6.8 Ringleb flow: constant speed lines (dashed) and computational do-
main Ω. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.9 Mach field on the NACA0012 airfoil (M∞ = 0.8,α∞ = 1.25o ): feed-
back filtered DG1 scheme. . . . . . . . . . . . . . . . . . . . . . . . 88
6.10 Mach field on the NACA0012 airfoil (M∞ = 0.8,α∞ = 1.25o ): feed-
back filtered DG5 scheme. . . . . . . . . . . . . . . . . . . . . . . . 88
6.11 Numerical Schlieren image for shock-bubble interaction at t = 0.18
(Feedback filtered DG2 scheme). . . . . . . . . . . . . . . . . . . . 89
6.12 Numerical Schlieren image for shock-bubble interaction at t = 0.18
(Feedback filtered DG2 scheme). . . . . . . . . . . . . . . . . . . . 89
7.1 Temperature field for the 2D heat equation problem, with kw = 5. 96
7.2 Example of order distribution with p-adaptivity and the 8x8x2 mesh
(αp = 0.3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.3 L2 -norm of the temperature error vs total number of DOF’s (possi-
ble orders: pe = 1,2,3,4,5). . . . . . . . . . . . . . . . . . . . . . . . 97
7.4 L2 -norm of the temperature error vs total number of DOF’s (possi-
ble orders: pe = 1,3,5) . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.5 Distribution of the reconstruction order with the entropy sensor. . 101
7.6 Distribution of the reconstruction order with the sensor based on
velocity gradients. . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.7 Distribution of the reconstruction order using both the entropy and
the velocity gradient sensors. . . . . . . . . . . . . . . . . . . . . . 102
7.8 Entropy field around a circular cylinder with p-adaptivity (Re∞ =
102 ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.9 Isotropic splitting on unstructured meshes: initial (black) and adapted
(red) meshes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.10 Initial mesh (1060 elements). . . . . . . . . . . . . . . . . . . . . . 106
7.11 Adapted mesh after 5 h-refinements (11806 elements). . . . . . . . 106
7.12 Transonic NACA0012: Mach field on the initial mesh (p=1). . . . 107
7.13 Transonic NACA0012: Mach field on the adapted mesh after 5 h-
refinements (p=1). . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.14 Computational domain for NACA0012 problem. . . . . . . . . . . 108
7.15 Transonic NACA0012: Cd error vs length scale . . . . . . . . . . . 109
7.16 Transonic NACA0012: Cl error vs length scale . . . . . . . . . . . 109
7.17 Transonic NACA0012: Cd error vs work units . . . . . . . . . . . 110
7.18 Transonic NACA0012: Cl error vs work units . . . . . . . . . . . . 110
8.1 Channel with smooth bump: acceleration by p-refinement. . . . . . 115
8.2 Channel with a smooth bump. . . . . . . . . . . . . . . . . . . . . 116

xii
8.3 Channel with smooth bump: entropy error vs equivalent length scale.118
8.4 Channel with smooth bump: entropy error vs work units. . . . . . 118
8.5 Domain for the transonic Ringleb flow. . . . . . . . . . . . . . . . . 120
8.6 Mach field for the transonic Ringleb flow: initial solution. . . . . . 121
8.7 Mach field for the transonic Ringleb flow: spurious shock. . . . . . 121
8.8 Example of structured mesh for the transonic Ringleb flow. . . . . 121
8.9 Transonic Ringleb flow: convergence history. . . . . . . . . . . . . . 122
8.10 Transonic Ringleb flow: entropy error vs equivalent length scale. . 122
8.11 Transonic Ringleb flow: entropy error vs work units. . . . . . . . . 123
8.12 Vortex transport: final velocity distribution on regular mesh. . . . 125
8.13 Vortex transport: error vs equivalent length scale on regular meshes. 125
8.14 Vortex transport: error vs work units on regular meshes. . . . . . . 126
8.15 Vortex transport: example of regular and perturbed mesh. . . . . . 126
8.16 Vortex transport: error vs equivalent length scale on perturbed
meshes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.17 Vortex transport: error vs work units on perturbed meshes. . . . . 127
8.18 Detail of the finest mesh used for the tandem airfoils test. . . . . . 128
8.19 Entropy field for the tandem airfoils test at the end of the simulation
(DG5RK4LS finest mesh) . . . . . . . . . . . . . . . . . . . . . . . 129
8.20 Drag coefficient for the trailing airfoil (finest mesh p=3,4,5). . . . . 130
8.21 Lift coefficient for the trailing airfoil (finest mesh p=3,4,5). . . . . 130
9.1 VKI-LS59 cascade: detail of the mesh. . . . . . . . . . . . . . . . . 132
9.2 VKI-LS59 cascade: Mach field. . . . . . . . . . . . . . . . . . . . . 133
9.3 VKI-LS59 cascade: wall isentropic Mach distribution. . . . . . . . 133
9.4 T106c cascade: example of hybrid mesh. . . . . . . . . . . . . . . . 135
9.5 T106c cascade: Mach at Re2s = 185000, M2s = 0.65 . . . . . . . . 136
9.6 T106c cascade: Mach at Re2s = 80000, M2s = 0.65 . . . . . . . . . 136
9.7 T106c cascade: laminar kinetic energy at Re2s = 185000, M2s = 0.65 137
9.8 T106c cascade: laminar kinetic energy at Re2s = 80000, M2s = 0.65 137
9.9 T106c cascade: turbulent kinetic energy at Re2s = 185000, M2s = 0.65138
9.10 T106c cascade: turbulent kinetic energy at Re2s = 80000, M2s = 0.65138
9.11 T106c cascade: wall isentropic Mach distribution (M2s = 0.65). . . 139
C.1 Curvilinear computational meshes (blue) and visualization meshes
(black) with different subdivisions for triangles and quadrilaterals. 148

xiii
 Part I

A numerical approach to
aerospace propulsion

1
Chapter 1

Introduction

1.1 CFD for aerospace propulsion systems

The mechanics and thermodynamics of fluid flows can be effectively investigated by
both experimental and numerical approaches. The results obtained by experimen-
tal activities are the basis for the knowledge of the phenomena which characterize
the motion of fluids. Starting from experimental results, it is possible to define
mathematical models which try to reproduce the observed phenomena. Usually,
these mathematical models require a numerical solution.
A comparison between the experimental approach and the numerical approach
shows that they are affected by different kind of uncertainties. The results ob-
tained by an experiment are influenced by the uncertainty related to the sensors
and by the assumptions made in order to evaluate quantities which cannot be di-
rectly measured.
On the other hand, the numerical approach introduces two kind of errors: a dis-
cretization error and a modeling error. The discretization error is related to the
numerical solution of the mathematical model and tends to zero when the dis-
cretization is refined. In contrast, the modeling error cannot be avoided and is
independent from the numerical discretization. The modeling error represent the
intrinsic approximation which is hidden in the equations and which can bring to an
unphysical prediction. The effects of the modeling errors become quite evident in
turbulence modeling for RANS equations, in which usually each turbulence model
gives a different result. A possible remedy to this problem is to switch to sim-
pler physical models in which less arbitrary assumptions are included. As far as
turbulent flows are concerned, this can be done by performing Direct Numerical
Simulations in which only the basic conservation principles are assumed and all the
turbulent scales are directly solved. This approach is quite similar to perform a

3
 1 – Introduction

numerical experiment but, unfortunately, its cost becomes prohibitive for the typi-
cal Reynolds numbers which characterize the flows in aerospace propulsion systems
(105 − 107 ).
Starting from these considerations, the experimental approach seems to be more
reliable, especially if turbulent flows are considered. However, the experimental
study of flows related to aerospace propulsion systems can be challenging. This is
due to the fact that they are usually characterized by high speed and very large
jumps in pressure and temperature (e.g. in rockets or jet engines). These fea-
tures make the building of an experimental facility very expensive and complex.
For example, it is difficult to reproduce experimentally the working conditions of
a blade inside a turbomachine because of the unsteady interactions between the
stages. Furthermore the numerical approach gives the full distribution of the state
variables in the domain. This is very useful in several cases. For example, the
numerical approach can be used in conjugated heat transfer problems in order to
obtain directly the solution of the temperature distribution in both the flow and
in the immersed body.
Finally, the numerical approach is interesting in the framework of optimization
processes during the development of an industrial product. Indeed, simulations
can be efficiently used to drive the optimization process because they make it pos-
sible to test several different configurations without the need to build a large set of
prototypes. The experimental approach can then be used as a complementary tool
which allows to validate the numerical results and to quantify the error related to
the numerical predictions.

1.2 Motivations and description of the work

The study of the flows in aerospace propulsion systems is characterized by several
difficulties related to both physical modeling and numerical solution of equations.
These problems have been progressively addressed during the research activity.
Nowadays, the CFD community is divided by a debate about the use of high-order
methods in fluid dynamics. In the literature it is possible to find papers with dif-
ferent views on this topic (see for example [5] and [6]). In this work, high-order
methods have been studied and tested for several problems in the aerospace field.
The results obtained during the research activity show that high-order methods
are more efficient than low-order methods in smooth problems. In other words,
the same error level can be obtained with a lower computational cost if high-order
methods are employed. However, if the solution contains shock waves or other sin-
gularities, then low-order methods (for example classical second order finite volume
methods) can become more efficient. Since the flow field is usually characterized by
both smooth and singular regions, an optimal choice may be to use an adaptive ap-
proach which changes the reconstruction order according to the local flow features.

4
 1.2 – Motivations and description of the work

In the preliminary study performed in this work, this last approach outperforms
both uniform low-order and uniform high-order discretizations (see Chapter 7.4).

In order to explore the possibilities offered by high-order methods, several high-

order finite volume schemes have been implemented during the first six months of
the research activity. In particular, experiments were performed with the WENO
method [7], the compact WENO method [8] and the third order limiter by Cada
and Torrilhon [9]. This preliminary activity gave goods results on cartesian and
regular structured meshes. In particular, the tested finite volume schemes show
excellent shock capturing properties. They are able to describe moving shock
waves with very sharp profiles and negligible oscillations. However, the extension
of these schemes to more complex problems is not trivial. This is mainly due to
the fact that in finite volume schemes the reconstruction is performed through the
use of information from neighbouring elements. When unstructured meshes are
considered the reconstruction process can become involved. Further complications
appear in the presence of non-conforming meshes which can result from h-adaptive
algorithms (see 7.3) or sliding meshes approaches (for example in studies on the
interaction between stator and rotor in turbomachinery flows). These problems
become even more complicate when high-order reconstructions are employed. In
fact, in finite volume schemes the increase in the reconstruction order is related to
an enlargement of the numerical stencil.
For these reasons an alternative approach has been explored, following one of the
current trends in the CFD community. Indeed, the second half of the first year
has been devoted to the implementation of a discontinuous Galerkin method. This
approach brings some simplifications in the previously described situations because
it makes it possible to obtain high-order reconstructions by using only the infor-
mation which is defined inside each element. In this way the stencil remains fixed,
even if high-order reconstructions are chosen. This aspect is not only important
because it greatly simplifies the implementation but also because it allows good
scalability in parallel computations.
During the first year the discontinuous Galerkin approach was initially imple-
mented for a linear scalar one-dimensional convection problem. After that, the
model was extended to the Burger equation and 1D Euler equations. The intro-
duction of non-linear terms in the convective fluxes required the implementation
of a proper shock capturing technique. At the beginning, some experiments were
performed with the Total Variation Bounded in the Mean (TVBM, see [10]) lim-
iter which gave good results mainly for low order reconstructions. After that, a
filter-based approach has been tested. In particular, an adaptive procedure named
Feedback filtering [1] has been developed in this phase. The goal of this procedure
is to obtain a shock capturing method which is as problem independent as possible.
See Section 6.1 for details. In this phase the extension of the code to 2D problems
was also performed.

5
 1 – Introduction

The code developed during the first year was used to participate to the Second
International Workshop on High Order CFD methods [11]. It was a very useful
experience because it makes it possible to compare the developed tool with state of
art codes from all over the world. A first comparison was performed by considering
the error level as a function of the total number of degrees of freedom. Good results
were obtained in this case, especially in the transonic Ringleb flow (8.2.2). A sec-
ond comparison involved the computational time which was properly normalized
in order to take into account the features of the particular hardware used by each
participant. This second comparison showed that the preliminary code was signif-
icantly slower than other existing codes in steady problems. The main reason for
this weakness was the lack of a proper accelerating technique for steady problems.
Indeed, while all the other participants used implicit time integration schemes, the
preliminary code could use only explicit time integration schemes which can be
quite inefficient for steady problems. The workshop experience influenced several
of the following choices.
During the first year, only inviscid compressible flows were considered. This made
it possible to develop a solid numerical background for the following work and to
focus the attention on the problems related to shock capturing.

During the second year, diffusive fluxes were introduced in the model. Heat
equation, Navier-Stokes equations and Reynolds Averaged Navier-Stokes equations
were studied in this phase. The work was mainly related to the development of
a new method for the discretization of the diffusive terms. The method, named
Enhanced Stability Recovery [2], is related to the original recovery method intro-
duced by van Leer and Nomura [3] in 2005. The original recovery method shows
an elegant way to compute diffusive fluxes at the interface between two adjacent
elements. It is based on a procedure in which a smooth differentiable recovery
solution is obtained starting from the discontinuous solution of the two elements.
The main idea is to find a recovery solution which is the weak sense indistinguish-
able from the original discontinuous solution. The ESR method is based on the
same apporach but is uses a different recovery basis and a different procedure for
the imposition of Dirichlet boundary conditions. The performed numerical com-
parisons showed that the ESR method has a larger stability limit in explicit time
integration with respect to the original recovery method but it introduces a larger
discretization error (see Section 4.3).
As far as turbulence modeling is concerned, the Spalart-Allmaras one-equation
model has been implemented during the second year. The model showed good
results on turbomachinery flows characterized by high Reynolds number and fully
turbulent behavior (see Section 9.1). Furthermore it appears particularly suitable
for the use with implicit time integration schemes because of the low stiffness of
its source terms (if compared to Wilcox k − ω model).
However, the weaknesses of this model appear when it is used in very low Reynolds

6
 1.2 – Motivations and description of the work

number transitional flows.

In order to investigate also low Reynolds number flows, other two turbulence
models have been implemented during the third year. In particular, the several
versions of the Wilcox k − ω model (1988 [12], 1998 [13], 2006 [14]) have been
implemented and tested. This phase was very useful to gain experience in tran-
sitional flows. Indeed, several sensitivity studies have been performed in order to
understand the effects of boundary conditions and calibration terms in turbulence
modeling. The attention was focused to a particular test case in the turboma-
chinery field: the transitional flow in the T106c turbine cascade. The numerical
experiments performed on this airfoil suggested the introduction of another turbu-
lence model, focused on transitional flows. In particular a three equation model
based on the k − ω model and the laminar kinetic energy concept has been imple-
mented and tested, following the approach of [15].
The stiffness related to the source terms of these turbulence models makes the
explicit time integration of the equations quite expensive. For this reason an im-
plicit time integration scheme has been implemented. This new feature improves
significantly the performances of the code in steady problems, bringing them to a
level comparable to that of other research codes presented at the Workshop.
The last months of the third year have been dedicated to the development of
adaptive algorithms. The discontinuous Galerkin approach offers the possibility to
easily perform adaptation in both the element order (p-adaptivity) and the element
size (h-adaptivity). Several experiments were done on both smooth flows and flows
with singularities. The most interesting results have been obtained by means of an
hybrid hp-adaptive approach.

7
8
Chapter 2

Physical models

In a CFD simulation the choice of the underlying physical model is the first funda-
mental step required to obtain results which can reproduce real-world phenomena.
In particular the physical model should be chosen in order to obtain a good com-
promise between the cost and the ability to reproduce the physics. For example,
the industrial design process of an aerospace component is characterized by sev-
eral steps in which the physical model is gradually refined and improved. In the
following, the five physical models implemented in the code are described putting
in evidence the typical flow features which they can describe.
All the models’ equations are adimensionalized following the procedure described
in Appendix A.

2.1 Euler equations

The Euler equations describe the behavior of compressible flows. They show all
the relevant phenomena which are related to the propagation of nonlinear waves.
Indeed, the solution of Euler equations can shown expansion and compression fans,
shock waves or contact surfaces.
Consider a 2D domain Ω. The equations can be written in conservative form as:

∂ρ ∂
+ (ρuj ) = 0 (2.1)
∂t ∂xj

∂ ∂ ∂P
(ρui ) + (ρui uj ) = − (2.2)
∂t ∂xj ∂xi

∂E ∂
+ (uj (E + P )) = 0 (2.3)
∂t ∂xj

9
 2 – Physical models

Here ρ, P, E, u, v denote density, pressure, total energy per unit volume and
cartesian components of velocity normalized to conventional reference values (see
Appendix A). The set of equations is closed by the introduction of the following
equation of state:
p 1
E= + ρ(u2 + v 2 ) (2.4)
γ−1 2
in which the ratio of specific heats γ has been set equal to 1.4 for all the considered
test cases.

2.2 Navier-Stokes equations

The compressible Navier-Stokes equations describe the behavior of compressible
flows in the presence of diffusive effects. The equations can be written in the
following form:

∂ρ ∂
+ (ρuj ) = 0 (2.5)
∂t ∂xj
∂ ∂ ∂P 1 ∂τij
(ρui ) + (ρui uj ) = − + (2.6)
∂t ∂xj ∂xi Reref ∂xj
∂E ∂ 1 ∂
+ (uj (E + P )) = [ui τij − qj ] (2.7)
∂t ∂xj Reref ∂xj
The heat flux is described through the linear Fourier law:

γ µ ∂T
qi = − (2.8)
(γ − 1) P r ∂xi

where the heat conductivity λ has been expressed in terms of the dynamic vis-
cosity µ and the Prandtl number P r. In this work the heat conductivity and the
dynamic viscosity are assumed constant. However, it is possible to easily imple-
ment the Sutherland’s law which describes the dependency of dynamic viscosity
from temperature. The components of the shear-stress tensor τij are computed ac-
cording to the Boussinesq approach and assuming the Stoke’s hypothesis for bulk
viscosity:
1 ∂uk
 
τij = 2µ Sij − δij (2.9)
3 ∂xk
where Sij is the mean strain rate tensor:

1 ∂ui
 
∂uj
Sij = + (2.10)
2 ∂xj ∂xi

10
 2.3 – Reynolds Averaged Navier-Stokes equations

2.3 Reynolds Averaged Navier-Stokes equations

The Navier-Stokes equations can be used to predict the behavior of turbulence
flows with Direct Numerical Simulations in which the 3D equations are integrated
on a sufficiently fine mesh. In this way, all the scales of turbulence are solved and
their effect on the solution is directly taken in account. However, this approach
becomes prohibitively expensive for high Reynolds number flows. For this reason,
several alternatives approaches have been introduced in order to reduce the re-
quired computation effort. The key idea of these methods is that only the larger
scales of turbulence have to be solved while the effects of the smaller scales can be
described by a proper closure model.
For example in Large Eddy Simulations the governing equations are obtained by
the application of a filter to the Navier-Stokes equations. The obtained filtered
equations contain new terms which represents the effects of the scales removed by
the filter. Usually the cut-off length of the filter is assumed to be proportional
to the size of the mesh elements and so the filtered flow field can be seen as the
restriction of the flow field to those components which can be resolved by the mesh.
The practical consequence of this approach is that the new terms (subgrid terms)
which appear in the governing equation are a function of the mesh size. When the
mesh size tends to zero the subgrid terms also tend to zero and so the Navier-Stokes
equations are obtained as an asymptotic limit.
An alternative approach is to perform a phase average (or a time average for steady
problems) of the Navier-Stokes equations in order to obtain governing equations
which describe the evolution of the average properties of the flow. This is the main
idea on which Reynolds Averaged Navier-Stokes (RANS) equations are based. In
particular, the new terms which appear after the averaging procedure are expressed
as a function of the averaged quantities. This is the weak point of the RANS ap-
proach: the closure model tries to reproduce the effects of scales which are not
resolved, using the information available in the averaged flow field. This approach
requires some strong hypothesis on the behavior of the flow which are not usually
satisfied in real problems. For example, several turbulence models assume that the
turbulent energy cascade is in equilibrium or they assume isotropic properties for
the turbulent structures.
From an engineering perspective, this approach can be acceptable because in prac-
tical problems there is often the need to know only some averaged or integral
properties of the flow field (for example the lift and drag coefficients of a body).
Furthermore, the cost of the RANS approach is several orders of magnitude lower
than a corresponding DNS computation. This difference grows quickly when the
Reynold number of the flow is increased. Is is also interesting to note that the
RANS approach can be used to perform 2D simulations in problems which show a
direction in which the averaged quantities are constant (for example turbine cas-
cades). This makes it possible to perform very cheap estimates of the properties of

11
 2 – Physical models

these flow fields. In contrast, the LES and DNS approaches require usually a 3D
simulation in order to describe properly the dynamics of vorticity, even if there is
a direction in which the averaged fields are constant.
Because of their low computational cost RANS models are widely used. In this
work, three models have been implemented and compared on several flow fields.
As it can be expected, each model works better than others in some problems and
so a general rule cannot be found. In the literature there is a trend in proposing
new models based on a growing number of equations. The goal is that the use
of more equations can reduce the number of calibration coefficients in the model.
This is not always true.
Since this work is devoted to the study of compressible flows, all the implemented
models are based on a Favre average of the Navier-Stokes equations. In the fol-
lowing the full system of conservation equations is reported for the different mod-
els. This makes it possible to put in evidence some approximations which can be
done during the coupling of the turbulence model with the original conservation
equations. For example, some models involve the definition of a turbulent kinetic
energy variable and this require some modification in the energy equation: the
total energy can be redefined in order to include also the turbulent kinetic energy
or, as it has been done in this work, an additive source term can be introduced in
the original energy equation. In contrast, there are models (like for example the
Spalart-Allmaras model) which do not involve the definition on a turbulent kinetic
energy variable. In these cases the energy equation remains unaltered, with the
exception of the inclusion of turbulent terms in the diffusive fluxes. As far as tur-
bulent heat fluxes are concerned they are computed thanks to the introduction of
a turbulent Prandtl number P rt , which has been set P rt = 0.9 in all simulations.

2.3.1 Spalart-Allmaras model

The Spalart-Allmaras (SA) model [16] is one of the most used RANS models.
This is due to its good behavior on fully turbulent high Reynolds number flows.
Furthermore, it requires the introduction of only one transport equation for the
modified turbulent viscosity ν̃. Since its introduction, several versions have been
proposed. A comprehensive review can be found in [17]. Spalart et al. [18] report
that confusion exists in the literature over the formulation of the SA model for
compressible flows. In this work the version of the model described in [18] has be
implemented. It can be obtained by combining the mass conservation equation
with the incompressible version of the SA model. The complete set of equations
is:

∂ρ ∂
+ (ρuj ) = 0 (2.11)
∂t ∂xj

12
 2.3 – Reynolds Averaged Navier-Stokes equations

∂ ∂ ∂P ∂ τ̂ij
(ρui ) + (ρui uj ) = − + (2.12)
∂t ∂xj ∂xi ∂xj

∂E ∂ ∂
+ (uj (E + P )) = [ui τ̂ij − qj ] (2.13)
∂t ∂xj ∂xj

∂ ∂   1 ∂  ∂ ν̃

cb2 ∂ ν̃ ∂ ν̃
(ρν̃) + (ρuj ν̃) = ρ P̂ − D̂ + ρ(ν + ν̃) + ρ +
∂t ∂xj σ ∂xj ∂xj σReref ∂xk ∂xk
1 ∂ρ ∂ ν̃
−
σReref ∂xk ∂xk
(2.14)

The heat flux includes the turbulent contribution which is modeled by the Reynolds
analogy and is expressed through the introduction of the turbulent Prandtl number
P rt :
 
γ µ µt ∂T
qi = − + (2.15)
(γ − 1) P rReref P rt ∂xi
The turbulent stress tensor τ̄ij is:

1 ∂uk
 
τ̄ij = 2µt Sij − δij (2.16)
3 ∂xk

The total (laminar+turbulent) stress tensor is:

2µ 1 ∂uk
 
τ̂ij = Sij − δij + τ̄ij (2.17)
Reref 3 ∂xk
The turbulent viscosity µt is defined as:

µt = ρν̃fv1 (2.18)

The damping function fv1 is given by:

χ3 ν̃
fv1 = χ= (2.19)
χ3 + Cvi
3 ν

The production and destruction terms P̂ and D̂ can be expressed as:

 2
1 ν̃
P̂ = cb1 (1 − ft2 )S̃ ν̃ D̂ = cw1 fw (2.20)
Reref d

13
 2 – Physical models

Here d is the distance from the closest wall while S̃ is the modified vorticity:
ν̃ χ
S̄ = fv2 fv2 = 1 − (2.21)
κ2 d2 1 + χfv1

S + S̄ if S̄ ≥ −cv2 S
S̃ = S (c2v2 S+cv3 S̄ ) (2.22)
S + if S̄ < −cv2 S
(cv3 −2cv2)S−S̄

where S is the magnitude of vorticity. In Eq. 2.22 the modified vorticity has been
limited following the approach proposed in [18]. In that work Spalart et al. observe
that the modified vorticity S̃ should always be positive and greater than 0.3S in
physically relevant situations. In the discretized equations it is possible for S̃ to
become zero or negative due to the fact that fv2 is itself negative over a range of
χ. Negative values of S̃ would introduce problems in the other correlations of the
model.
The function fw is:
1/6
1 + c6w3
  
ν̃
fw = g g = r + cw2 (r6 − r) r = min , rlim (2.23)
g 6 + c6w3 S̃κ2 d2
In the original SA model a trip term and a laminar suppression term are introduced
in the ν̃ equation. This makes it possible to force the transition of the flow in a
particular point specified by the user. Since in most of the applications the location
of the transition point is not known and the model is considered not appropriate
for transitional flows by Rumsey and Spalart [19], the trip and laminar suppression
terms have not been included in the previous equations.

The constants of the model are cb1 = 0.1355, σ = 2/3, cb2 = 0.622, κ =
0.41, cw1 = cb1 /κ2 , cw2 = 0.3, cw3 = 2, cv1 = 7.1, ct1 = 1, ct2 = 2, ct3 = 1.2, ct4 =
0.5, rlim = 10, cv2 = 0.7, cv3 = 0.9.

Some constants of the model (like for example κ) come from the constants of
the theoretical boundary layer law. Indeed, the tests performed on the velocity
profile for a turbulent boundary layer on a flat plate show good agreement with
the boundary layer law (see Figure 2.1). In Figure 2.2 the computed skin friction
coefficient for the turbulent flow on a flat plate is reported and compared with two
correlations from the literature.

2.3.2 Wilcox k − ω model

In this work several versions of the Wilcox k-ω model have been considered (1988
[12], 1998 [13], 2006 [14]). They have been implemented following the non-standard

14
 2.3 – Reynolds Averaged Navier-Stokes equations

Figure 2.1. Velocity profile in turbulent boundary layer obtained by the proposed
discretization with the Spalart-Allmaras model

Figure 2.2. Skin friction coefficient for turbulent flat plate: comparison between
the implemented Spalart-Allmaras model and correlations from the literature

15
 2 – Physical models

approach proposed by Bassi et al. [20]. In this way the equations are solved for the
logarithm of ω (ω̃ = log(ω)) instead of ω itself: this makes the discretization more
stable. Furthermore, both k and ω̃ are limited on the basis of physical consider-
ations. This helps in reducing the stiffness of the discretized equations which can
lead to unphysical values during strong transients. This problem is particularly
significant when the equations are integrated in time by an implicit scheme with a
very large time step.
In the implemented approach the conservative variable E includes only the internal
energy and the kinetic energy of the mean flow. The energy is intended per unit
volume. The conservation of the total energy (internal energy+kinetic energy of
the mean flow+turbulent kinetic energy) is guaranteed by the presence of coupling
source terms in the energy equation and in the k equation.
Wilcox proposed two versions of his models, named high Reynolds and low Reynolds.
The names refer to a local Reynols number (ReT ) obtained by the turbulence vari-
ables. While in the high Re version some coefficients are constant, in the low Re
version these coefficients become functions of ReT . The correlations of the low Re
model tend to the high Re model for high ReT . Since some applications considered
in this work involve low Reynolds number transitional flows, the low Re versions
of the models have been considered.
In the following the set of equations for the low Re k-ω(1998) model is reported in
the form implemented in the code:
∂ρ ∂
+ (ρuj ) = 0 (2.24)
∂t ∂xj
∂ ∂ ∂P ∂ τ̂ij
(ρui ) + (ρui uj ) = − + (2.25)
∂t ∂xj ∂xi ∂xj

∂E ∂ ∂ ∂ui
+ (uj (E + P )) = [ui τ̂ij − qj ] − τij + β ∗ ρk̄eω̃r (2.26)
∂t ∂xj ∂xj ∂xj
  
∂ ∂ ∂ui ∂ µ ∂k
(ρk) + (ρuj k) = τij − β ∗ ρk̄eω̃r + + σk µ̄t (2.27)
∂t ∂xj ∂xj ∂xj Reref ∂xj
 
∂ ∂ α ∂ui µ ∂ ω̃ ∂ ω̃
(ρω̃) + (ρuj ω̃) = τij − βρeω̃r + + σω µ̄t +
∂t ∂xj k̄ ∂xj Reref ∂xk ∂xk
   (2.28)
∂ µ ∂ ω̃
+ + σω µ̄t
∂xj Reref ∂xj
The heat flux includes also the turbulent contribution:
 
γ µ µ̄t ∂T
qi = − + (2.29)
(γ − 1) P rReref P rt ∂xi

16
 2.3 – Reynolds Averaged Navier-Stokes equations

The turbulent stress tensor τ̄ij is:

1 ∂uk 2
 
τ̄ij = 2µ̄t Sij − δij − ρk̄δij (2.30)
3 ∂xk 3

The total (laminar+turbulent) stress tensor is:

2µ 1 ∂uk
 
τ̂ij = Sij − δij + τ̄ij (2.31)
Reref 3 ∂xk
The turbulent viscosity µ̄t is defined as:

α∗ ρk̄
µ̄t = (2.32)
eω̃r
The model constant, the auxiliary functions and the low Re correlations are:

σk = 0.5 σω = 0.5 β0∗ = 0.09 β0 = 9/125

α0 = 1/9 α0∗ = β0 /3 Rk = 6 Rω = 2.61

α0 + ReT /Rω α0∗ + ReT /Rk

 
α = 13/25 (α∗ )−1 α∗ = (2.33)
1 + ReT /Rω 1 + ReT /Rk
ρk
ReT = Reref χω = ∇k · ∇ω̃/e2ω̃ (2.34)
µeω̃

β = β0 f β β ∗ = β0∗ fβ ∗ (2.35)

(
1 + 70χω 1 if χk < 0
fβ = fβ ∗ = (2.36)
1 + 80χω (1 + 680χk )/(1 + 400χk ) if χk ≥ 0
2 2

Ωij Ωjk Ωki 1

χω = χk = ∇k · ∇ω̃ (2.37)
β0∗ e3ω̃ e3ω̃
where Ω is the vorticity. Note that the term χω is identically zero for 2D flows.
For simplicity the correlations and the constants for the 1988 and 2006 versions
are not reported. They can be found in [12] and [14].
In the previous equations the limited turbulent viscosity k̄ appears in the source
terms and in the eddy viscosity. It is defined as

k̄ = max(0, k) (2.38)

17
 2 – Physical models

Notice that k̄ has been limited exactly to zero (and not to an arbitrary small value)
because, after the appropriate substitutions, no term in the previous equations is
divided by k̄.
Furthermore the source terms and the eddy viscosity are computed by using the
variable ω̃r instead of ω̃. The variable ω̃r can be obtained from ω̃ by the impo-
sition of some realizability conditions. In particular, the positivity of the normal
turbulent stresses 2.39 and the Schwarz inequality for the shear turbulent stresses
2.40 are imposed, following the approach of Bassi et al.[20]:
00 2
ρui ≥ 0 (2.39)

 2
00 2 00 2
(2.40)
00 00
ρui uj ≤ ρui · ρuj

These requirements are satisfied thanks to the definition of a lower bound ω̃r0 for
ω̃:
ω̃r = max(ω̃, ω̃r0 ) (2.41)
The lower bound ω̃r0 can be obtained by substituting the modeled turbulent
stresses (Eq.2.30)in Eq.2.39 and Eq. 2.40:

2 1 ∂uk
 
ρk̄ − 2µ̄t Sii − ≥0 i = 1,2 (2.42)
3 3 ∂xk

2 1 ∂uk 2 1 ∂uk
     
2
(−2µ̄t Sij ) ≤ ρk̄ − 2µ̄t Sii − ρk̄ − 2µ̄t Sjj −
3 3 ∂xk 3 3 ∂xk (2.43)
i = 1,2 i=
/ j

Recalling the definition of the turbulent viscosity (Eq. 2.32) it is possible to obtain
ω̃
the following set of realizability inequalities expressed in terms of the unknown αe ∗ :

eω̃ 1 ∂uk
 
− 3 Sii − ≥0 i = 1,2 (2.44)
α∗ 3 ∂xk

2
eω̃ 2 ∂uk
    ω̃ 
e
− 3 Sii + Sjj − +
α ∗ 3 ∂xk α∗
(2.45)
1 ∂uk 1 ∂uk
   
− 9 Sij − Sii −
2
Sjj − ≥0 i = 1,2 i=
/ j
3 ∂xk 3 ∂xk

In the following, the maximum value between the zeros of Eqs. 2.44 and 2.45
ω̃ ω̃
(the unknown is αe ∗ ) will be denoted as â. If αe ∗ > â then Eqs. 2.44 and 2.45 are

18
 2.3 – Reynolds Averaged Navier-Stokes equations

both satisfied. The lower bound ω̃r0 that guarantees realizable turbulent stresses
can be obtained by:

eω̃0
= â (2.46)
α∗
Since the low Re version of the k − ω model is considered here, α∗ depends on
the local turbulent Reynolds number Rek according to Eq. 2.33. In conclusion,
ω̃r0 can be obtained by the solution of the following second degree equation:
 
ω̃r0 2 ρk̄Reref âρk̄Reref
+ ∗
− âα0 eω̃r0 − =0 (2.47)


e
µRk µRk

In order to show an example of the behavior of the Wilcox k − ω low Re model

in the presence of transition the plot of Figure 2.3 is reported. It shows the skin
friction coefficient distribution on a flat plate obtained by the present implementa-
tion. The obtained results show how the behavior change from the laminar Blasius
law (blue line) to the turbulent correlation (light blue). The transition point is
strongly influenced by the freestream turbulence level and by the wall roughness
which determines the boundary condition for ω.

Figure 2.3. Skin friction coefficient on flat plate: transition from laminar to tur-
bulent. Blue=Blasius solution, Light blue=turbulent correlation, Green= imple-
mented Wilcox k − ω low Re model

19
 2 – Physical models

2.3.3 Laminar Kinetic Energy model

The study of transitional flows by RANS models is very difficult because the hy-
pothesis on which the RANS approach is based are usually not satisfied in tran-
sitional flows. This is the typical field in which LES and DNS methods find their
natural application. This is due to the fact that the prediction of transition be-
comes important for low Reynolds number flows and the computational cost of LES
or DNS can be acceptable when the Reynolds number is sufficiently low. However,
several works have been proposed in the literature also on the study of transitional
flows with RANS models.
In particular, there are several transition models which are based on the inter-
mittency concept. They use an additive transport equation for the intermittency
variable (usually named γ) which can usually assume values in the range 0-1. The
intermittency variable is used as a weighting factor to control the intensity of the
production term in classic RANS model. This approach requires the introduction
of empirical correlations which have to be calibrated on the particular flow which is
under study. An important work in this field was made by Langtry and Menter [21]
who introduced the γ − Reθ model. In particular their model has the remarkable
feature of being truly local. Babajee [22] performed a detailed analysis on the use
of the γ − Reθ + SST model for the low pressure turbine cascade T106c, which has
been studied also in this work (Chapter 9.2). In particular, Babajee [22] showed
that it is very difficult to obtain results in good accordance with the experimental
data when the γ − Reθ model is used for very low Reynolds numbers.
In the last years, several transition models based on an alternative idea have been
developed. In particular, the laminar kinetic energy (LKE) concept seems to be
promising. This approach was initially introduced by Mayle and Schulz [23] in
order to take into account the pretransitional rise of fluctuations in transitional
boundary layers and their subsequent breakdown to turbulence. They introduced
an additional equation in order to describe the evolution of the energy which is
associated to these pretransitional fluctuations. This idea has been exploited by
Walters and Leylek [24] and Lardeau et al. [25] to develop a model for natural and
by-pass transition. The concept has been used also by Pacciani et al. [15] for the
study of transitional flows in low pressure gas turbines. In particular, they have
focused the attention on the transition induced by laminar separation. In their
approach, a three equations RANS model has been developed and tested with a
second order finite volume solver. Their model is based on the low Re Wilcox
(1998) k-ω model plus an additive equation for the laminar kinetic energy. The
model contains coupling terms between the laminar kinetic energy equation and the
turbulent kinetic energy equation which are responsible for the transition process
and the exchange of energy from the pretransitional fluctuations to the turbulent
field.
In this work the LKE model [26] has been implemented and tested in the DG

20
 2.3 – Reynolds Averaged Navier-Stokes equations

framework. The numerical experiments performed on a low pressure turbine cas-

cade (see Section 9.2) showed that the LKE approach works well if used together
with the 1998 version of the Wilcox k − ω model. Indeed, this version of the k − ω
model underpredicts the turbulent kinetic energy and so the LKE equation intro-
duces the missing contribution. In contrast, the 1988 and 2006 versions of the
k − ω model seem to overpredict the turbulent kinetic energy and so the additional
contribution given by the LKE equation degrades further the agreement with the
experimental data.
In the following the equations of the implemented LKE model are reported:
∂ρ ∂
+ (ρuj ) = 0 (2.48)
∂t ∂xj
∂ ∂ ∂P ∂ τ̂ij
(ρui ) + (ρui uj ) = − + (2.49)
∂t ∂xj ∂xi ∂xj

∂E ∂ ∂ ∂ui µ kL
+ (uj (E + P )) = [ui τ̂ij − qj ] − τij + β ∗ ρk̄eω̃r − ρνL S 2 + 2
∂t ∂xj ∂xj ∂xj Reref d2
(2.50)

  
∂ ∂ ∂ui ∂ µ ∂k
(ρk) + (ρuj k) = τij − β ∗ ρk̄eω̃r + + σk µ̄t + ρR
∂t ∂xj ∂xj ∂xj Reref ∂xj
(2.51)

 
∂ ∂ α ∂ui µ ∂ ω̃ ∂ ω̃
(ρω̃) + (ρuj ω̃) = τij − βρeω̃r + + σω µ̄t +
∂t ∂xj k̄ ∂xj Reref ∂xk ∂xk
   (2.52)
∂ µ ∂ ω̃
+ + σω µ̄t
∂xj Reref ∂xj

 
∂ ∂ ∂ µ ∂kL µ kL
(ρkL ) + (ρuj kL ) = + ρνL S 2 − 2 − ρR (2.53)
∂t ∂xj ∂xj Reref ∂xj Reref d2

Here d is the distance from the closest wall.

It can be seen from Eqs. 2.51 and 2.53 that in the LKE approach a coupling term
(ρR) is added in the k and kL equations. This term is responsible for the transition
from laminar to turbulent flow: when this term is activated there is a transfer of
energy from the fluctuating laminar field to the turbulent field.
The production term in the kL equation depends on the "laminar eddy viscosity"
νL : p
νL = C1 f1 kL δΩ (2.54)

21
 2 – Physical models

The constant C1 has to be deduced from calibration. In [15] it is set as C1 = 0.01.

In the present work the value C1 = 0.009 is chosen.
The function f1 is
 p 
f1 (Tu ) = max 0.8,2 tanh Tu /4.5 (2.55)

where Tu is the turbulence level. It is defined as Tu = 100 (2/3)k/q 2 where q is

p

the velocity magnitude.

The variable δΩ is the vorticity thickness. It is the most problematic variable in
this model because it is non local. Indeed, Pacciani et al. [15] suggest that it can
be computed as
 −1
q∞ ∂q
δΩ = (2.56)
2 ∂y maxy

where q∞ is the "freestream" velocity magnitude and ∂y ∂q

is the derivative of the
velocity magnitude with respect to the wall normal coordinate y. The value of q∞
can be estimated by the local velocity value at the edge of the boundary layer.
The transfer term R is modeled as:

R = C2 f2 ωβ ∗ kL (2.57)

The damping function f2 is

f2 = 1 − e−Ψ/C3 Ψ = max(0, Ry − C4 ) (2.58)

The constant C2 regulates the intensity of the energy transfer from kL to k and
is set as C2 = 0.3. The constant C4 represents a threshold for the activation of
transition and
√ it is set as C4 = 10. This threshold is compared with the variable
Ry = Reref kd/ν which can be seen as a local Reynolds number based on the
turbulent kinetic energy k and the wall distance d. Finally, the constant C3 is set
as C3 = 8.
Heat fluxes and shear stresses are computed following the same approach used for
the k − ω model.

2.4 Boundary conditions

The evolution of the computed flow field is regulated by the imposed boundary
conditions. For this reason, it is fundamental to take care of the physical phenom-
ena which are responsible for the propagation of the informations in the fluid. This
is particular evident in Euler equations in which the wave propagation effects are
described by the characteristics signals.
In the presence of diffusive effects further conditions have to be imposed in order

22
 2.4 – Boundary conditions

to compute the viscous fluxes (Fv and Gv ).

As far as RANS equations are concerned, the choice of the proper boundary condi-
tions can be quite difficult because often there are not experimental data available
for the variables on which the turbulence model works. In this case, it is necessary
to introduce empirical correlations which link the available boundary information
to the variables requested by the turbulence model.

2.4.1 BCs for Euler equations

The Euler equations are dominated by the presence of wave propagation effects.
It is possible to find characteristic lines in the space-time along which signals are
transported. In order not to violate physics, BCs have to be imposed coherently
with the propagation of these signals. In particular, the number of BCs which can
be imposed is equal to the number of signals which enter in the computational
domain.
Inlet and outlet boundaries are treated by defining a ghost state with some infor-
mation from the computational domain and some from the external environment,
according to the number of characteristic signals which enter in the domain. In this
work, total temperature, total pressure and flow direction are imposed at subsonic
inlets. All the data are imposed at supersonic inlets. Pressure is the only quantity
imposed at subsonic outlets. Alternatively, the incoming signals can be directly
imposed in subsonic inlets or outlets in order to obtain a non-reflecting boundary
condition.
If the boundary is a solid wall then the tangency condition is enforced by defining a
ghost state characterized by a velocity vector specular with respect to the internal
velocity.
The previous considerations hold for the Euler equations but can also be used
for the computation of boundary convective fluxes (F and G) in Navier-Stokes or
RANS equations.

2.4.2 BCs for Navier-Stokes equations

The presence of diffusive terms in the Navier-Stokes equations requires the intro-
duction of additive BCs with respect to the Euler equations. In particular, in the
presence of solid walls the no-slip condition can be imposed on the momentum
variables.
As far as the energy equation is concerned, it is possible to impose Dirichlet, Neu-
mann or Robin BCs depending on the properties of the wall.

23
 2 – Physical models

2.4.3 BCs for Spalart-Allmaras model

The Spalart-Allmaras equation works on the modified turbulent viscosity ν̃ which
is transported by the flow field. At solid walls an homogeneous Dirichlet boundary
condition can be imposed because the velocity fluctuations are zero here and so also
the turbulent viscosity has to be zero. However, the situation is more complex at
the inlet where the incoming turbulent viscosity has to be specified. The turbulent
eddy viscosity is a quantity which cannot be directly measured in an experiment.
It can be obtained from empirical correlations based on the turbulence intensity
and the turbulence characteristic length scale.
However, Rumsey and Spalart [19] recommend to use this model only for fully
turbulent flows and not for low Reynolds transitional flows. In particular, the
recommended inlet turbulent viscosity for fully turbulent flows is 3ν ≤ ν̃ ≤ 5ν [17]
where ν is the molecular kinematic viscosity.

2.4.4 BCs for Wilcox k − ω model

At solid walls the turbulent kinetic energy k has to be set equal zero because there
are no turbulent fluctuations here.
In contrast, the wall value for the specific dissipation ω is not zero. A possible
approach to find the proper wall boundary condition for ω is the so-called rough-
wall method proposed by Wilcox [14]. This approach makes it possible to describe
both smooth and rough walls. In the case of rough walls, it requires the definition
of a non-dimensional equivalent sand-roughness height kr+ = kr uτ /ν, where uτ is
the friction velocity defined as uτ = τw /ρw . Then the boundary condition is:
p

u2τ
ωw = Sr (2.59)
ν
where the correlation Sr is
(
(50/kr+ )2 if kr+ < 25
Sr = (2.60)
100/kr+ if kr+ ≥ 25

Wilcox proposed to use the same approach also for smooth walls with the so-called
"slightly-rough-wall" boundary condition. In practice he suggested to substitute
the perfectly smooth wall surface with an hydraulically smooth surface. In this
way the boundary condition becomes:
ν
ωw = 2500 2 (2.61)
kr
where, according to Wilcox, kr should be low enough to guarantee that kr+ < 5,
i.e., it should ensure that the surface is hydraulically smooth with roughness peaks
lying within the viscous sublayer.

24
 2.4 – Boundary conditions

An alternative approach for smooth walls was proposed by Menter [27] with
the following relation which links the wall specific dissipation (ωw ) to the distance
of the first grid point off the wall (y1 ):
6ν
ωw = 10 (2.62)
βy12
The condition described by Eq.2.62 means √ setting at the wall the analytical solution
computed at y = αM y1 where αM = 1/ 10.
A combination between Eq.2.61 and Eq.2.62 gives a linear relation between the
prescribed wall roughness (kr ) and the distance of the grid point nearest to the
wall (y1 ): r
kr = Cw y1 Cw = 2500αM 2 β0 = 1.77 (2.63)
6
where the constant Cw depends on the numerical scheme and the reconstruction
order. Hellsten [28] proposed to optimize the factor 10 of Menter’s formula by
means of an accurate near-wall numerical study of the ω solution and by comparing
skin friction distributions of flat plate flows computed on differently refined grids.
The value of the factor proposed by Hellsten is 1.25 instead of 10. With this value
the constant of Eq. 2.63 becomes Cw = 2.5. Another different optimal value
(Cw = 0.3) has been found by Bassi et al. [20].
Bassi et al. [29] performed a study in order to understand how the term αM
depends on the reconstruction order pe in the framework of discontinuous Galerkin
schemes. They proposed the following boundary condition:
6ν
ωw = Reref (2.64)
β(hαM )2
where Ppe 1
αM = e− n=1 n (2.65)
They suggested to set h equal to the distance between the wall and the centroid
of the element.
In this work, the boundary condition described by Eq. 2.64 has always been used
in the presence of smooth walls.

As far as the inlet is concerned, both k and ω have to be assigned. The inlet
turbulence intensity is usually known from experimental data and so kin can be
easily assigned as a function of the inlet turbulence level Tu :
2
3 2

Tu
k = qin (2.66)
2 100
where qin in the inlet velocity magnitude.
On the contrary, the specific dissipation rate ωin is usually not directly measured

25
 2 – Physical models

and so it should be obtained indirectly. For example, in this work it is obtained

starting from the turbulent kinetic energy kin and the characteristic turbulence
length scale lt : √
kin
ωin = (2.67)
lt
At the outlet the values of k and ω are extrapolated from the domain.

2.4.5 BCs for LKE model

In the LKE model the same boundary conditions described in 2.4.4 can be used
for k and ω. As far as kL is concerned, it can be set equal to zero at solid walls.
In [15] the recommended value for kL at the inlet is kLin = kin . At the outlet kL
is extrapolated from the interior of the domain.

26
 Part II

Discretization

27
Chapter 3

Discontinuous Galerkin
spatial discretization

The Discontinuous Galerkin (DG) method was introduced first by Hill and Reed in
1973 [30] for the solution of the transport equation in neutronics. Nowadays they
are widely used for the numerical solution of conservation laws thanks to some of
their features.
Indeed, DG methods seem to join the main advantages of finite volumes (FV)
methods and finite elements (FE) methods. From FV methods they have inherited
several numerical fluxes which make it possible to define fluxes based on upwind
considerations: when hyperbolic partial differential equations are considered, this
approach obtains good results since it takes into account physical wave propagation
phenomena.
On the other hand, the strategy used to obtain high accuracy in the space dis-
cretization is very different in FV and DG methods. In FV methods the only
information that is known inside a cell is the average value of the conservative
variables and so high-order reconstructions can be obtained only through interpo-
lation of data coming from neighbouring cells. In contrast, DG methods obtain
high-order reconstructions introducing several degrees of freedom (DOFs) inside the
element. This alternative approach greatly simplify reconstruction since all the re-
quired information is already inside the element. Furthermore, all the integrals in
the governing equations can be easily calculated by mapping the generic element in
the physical space to a reference element in the computational space. This aspect
makes DG methods very similar to FE methods and so all the knowledge on ele-
ment mapping developed for FE methods can also be used for DG methods. This
topic is of considerable interest with regard to the proper treatment of boundary
conditions. Indeed high-order schemes require a careful representation of solid wall

29
 3 – Discontinuous Galerkin spatial discretization

boundaries in order to preserve solution accuracy as was shown by Bassi and Rebay
in 1997 [31]. For this reason curvilinear elements have to be introduced at solid
walls.
Furthermore, the local and discontinuous nature of the DG reconstruction makes
it possible to easily introduce adaptive strategies as will be shown in Chapter 7.

3.1 Variational approach

The DG space discretization will be described in the case of a general 2D governing
equation for the variable u:
∂u ∂F ∂G
+ + =Q (x, y) ∈ Ω ⊂ R2 , t ∈ R0+ (3.1)
∂t ∂x ∂y

where F and G are the fluxes (convective+diffusive) and Q is a generic source

term. Assume that the domain Ω is subdivided into a collection of non overlapping
elements Ωe . Consider a generic element Ωe . Consider the functional space Vh
defined by
Vh = {v ∈ L2 (Ω) : v|Ωe ∈ P p (Ωe ) ∀Ωe ∈ Th } (3.2)
where P p (Ωe ) is the space of polynomials of degree up to p defined in Ωe . The space
Vh will be used to approximate the solution inside Ωe . There are no continuity
requirements across the element edges in the space Vh and so the reconstruction
order p can vary from element to element.
Consider a proper basis for Vh which contains Ne basis functions Φi . The numerical
approximation of the solution (uh ) belongs to the space Vh and can be represented
using a linear combination of the basis functions:
Ne
X
uh (x, t) = ũ · Φ = ũi Φi (3.3)
i=1

If uh is substituted in the governing equation (3.1) the residual Rh is obtained:

∂uh ∂Fh ∂Gh

Rh = + + −Q (3.4)
∂t ∂x ∂y
The Galerkin’s approach requires that the residual projection on the Vh space is
set equal to zero:

Z Z Z Z Z
∂uh ∂Fh ∂Gh
Rh νdxdy = νdxdy+ νdxdy+ νdxdy− Qνdxdy = 0 ∀ν ∈ Vh
Ωe Ωe ∂t Ωe ∂x Ωe ∂y Ωe
(3.5)

30
 3.2 – Basis functions

The weak form of the discontinuous Galerkin discretization is derived by using

integration by parts:
Z Z Z Z
∂uh ∂ν ∂ν
νdxdy+ (F̂h nx +Ĝh ny )νds− ( Fh + Gh )dxdy− Qνdxdy = 0
Ωe ∂t ∂Ωe Ωe ∂x ∂y Ωe
(3.6)
where nx and nx are the cartesian components of the outward-pointing unit normal
vector and s is the element boundary curve. The last equation shows that the
integral over the element boundary ∂Ωe requires the knowledge of fluxes (F̂h and
Ĝh ) at the interface between neighbouring elements. This problem is solved by the
introduction of a proper numerical flux. The convective and diffusive contributions
to this flux are evaluated separately. They are described in Chapter 3.5 and 4
respectively.
The fact that Eq. 3.6 has to be valid for ∀v can be imposed by considering Eq.
3.6 for all the Ne functions which define the element basis, i.e. setting v = Φi with
1 ≤ i ≤ Ne . In this way Ne equations can be obtained and used to describe the
time evolution of the Ne solution coefficients inside the element Ωe :

Z Ne Z Z
X ∂ ũi ∂Φj ∂Φj
Φi Φj dxdy + (F̂h nx + Ĝh ny )Φj ds − ( Fh + Gh )dxdy+
Ωe i=1 ∂t ∂Ωe Ωe ∂x ∂y
Z
− QΦj dxdy = 0 1 ≤ j ≤ Ne
Ωe
(3.7)

The system can be rewritten in compact form through the introduction of the
element mass matrix ([M])
Z
[M]ij = Φi Φj dxdy (3.8)
Ωe

Z Z Z
∂ũ ∂Φ ∂Φ
[M] =− (F̂h nx + Ĝh ny )Φds + ( Fh + Gh )dxdy + qΦdxdy
∂t ∂Ωe Ωe ∂x ∂y Ωe
(3.9)

3.2 Basis functions

The choice of the basis functions strongly affects the behavior of the numerical
discretization. Indeed, the time evolution of the numerical solution depends on
the linear system (3.9) in which the mass matrix appears. A bad choice of the
basis functions can lead to an ill-conditioned mass matrix and this can deteriorate

31
 3 – Discontinuous Galerkin spatial discretization

the accuracy of the numerical solution. In the DG framework there are several
possibilities about the choice of the element basis. Both nodal and modal basis
functions can be adopted. In this work, only modal basis have been considered
because of the freedom related to the choice of modal functions. In particular,
this makes it possible to choose a set of basis functions which shows two positive
features: hierarchy and orthonormality. In Chapter 7.2 the advantages related to
the use of a hierarchical basis will be explained.
The second important feature of the chosen bases is the fact that they are or-
thonormal. When the basis is orthonormal the element mass matrix is the identity
and so its condition number is one, independently from the shape of the element.
Furthermore, this choice helps to reduce the computational cost with explicit time
integration because it is not necessary to multiply the right hand side of (3.9) for
the inverse of the mass matrix (which is the identity) at each time step.
In order to obtain an orthonormal element basis the modified Gram-Schmidt pro-
cedure has been implemented, following the approach of Bassi et al. [32]. This
algorithm requires to be initialized with a starting basis. In this work, two differ-
ent starting bases have been implemented for this purpose. The first one is obtained
by a tensor product of Legendre polynomials defined on the reference square. With
this basis, an element with a reconstruction of order p contains Ne = (p+1)2 DOFs
in 2D. In the following of this work this basis will be referred to as BASIS A.
Alternatively, it is possible to define a starting basis from a set of monomials de-
fined in the physical space on a reference frame centered on the element’s center
of mass and aligned to its axis of inertia. This last choice has been proposed by
Bassi et al. [32] for general shaped elements. With this basis, an element with
a reconstruction of order p contains Ne = (p + 1)(p + 2)/2 DOFs in 2D. In the
following of this work this basis will be referred to as BASIS B.

32
 3.3 – Mappings

3.3 Mappings
Integrals are approximated through the use of Gauss quadrature formulas (from
[33]) defined on the transformed reference element. The number of quadrature
points is chosen in order to exactly integrate polynomials of order 2p on the refer-
ence element.

3.3.1 Quadrilaterals
The mapping between the general curvilinear quadrilateral element (implemented
up to quartic elements) and the reference square are computed by the Serendipity
mappings (Figure 3.1) reported by Oñate [34].

Figure 3.1. Serendipity mappings for quadrilateral elements.

The transformation is

Pnnod
x = Ni (ξ, η)xi
Pi=1
nnod
y = i=1 Ni (ξ, η)yi (3.10)

where nnod is the number of nodes which define the geometry of the element
and Ni are the shape functions.

Linear element (nnod = 4)

1
Ni = (1 + ξi ξ)(1 + ηi η) ξi = ±1 ηi = ±1 (3.11)
4
33
 3 – Discontinuous Galerkin spatial discretization

Quadratic element (nnod = 8)

1
Ni = (1 + ξi ξ)(1 + ηi η)(ξi ξ + ηi η − 1) ξi = ±1 ηi = ±1
4
1
Ni = (1 − ξ 2 )(1 + ηi η) ξi = 0 ηi = ±1
2
1
Ni = (1 + ξi ξ)(1 − η 2 ) ξi = ±1 ηi = 0 (3.12)
2
Cubic element (nnod = 12)
1
Ni = (1 + ξi ξ)(1 + ηi η)(9(ξ 2 + η 2 ) − 10) ξi = ±1 ηi = ±1
32
9 1
Ni = (1 − ξ 2 )(1 + ηi η)(1 + 9ξi ξ) ξi = ± ηi = ±1
32 3
9 1
Ni = (1 + ξi ξ)(1 − η 2 )(1 + 9ηi η) ξi = ±1 ηi = ± (3.13)
32 3
Quartic element (nnod = 17)
1
Ni = (1 + ξξi )(1 + ηηi )[4(η 2 − 1)ξξi + 4(η 2 − 1)ηηi + 3ξηξi ηi ] ξi = ±1 ηi = ±1
12
ξξi
Ni = 2(1 + ξξi )(η 2 − 1)(η 2 − ) ξi = ±1 ηi = 0
4
ηη i
Ni = 2(1 + ηηi )(ξ 2 − 1)(ξ 2 − ) ξi = 0 ηi = ±1
4
4 1
Ni = (1 + ξξi )(1 − η 2 )(η 2 + ηηi ) ξi = ±1 ηi = ±
3 2
4 1
Ni = (1 + ηηi )(1 − ξ )(ξ + ξξi )
2 2
ξi = ± ηi = ±1
3 2
Ni = (1 − ξ 2 )(1 − η 2 ) ξi = 0 η i = 0 (3.14)

3.3.2 Triangles
Curvilinear triangular elements (implemented up to cubic elements) are trans-
formed into the reference triangle by standard mappings also from [34]. Consider
a point P inside a triangle with the vertices 1,2 and 3. This defines three sub-
areas A1 , A2 and A3 corresponding to the triangles P 23, P 13 and P 12. The area
coordinates L1 , L2 and L3 are defined as
A1 A2 A3
L1 = L2 = L3 = (3.15)
A A A
where A is the area of the triangle.
As it can be expected only two area coordinates are sufficient to define the position

34
 3.4 – Choice of the basis

of a point inside the triangle. Indeed, the three coordinates are not independent:

L1 + L2 + L3 = 1 (3.16)

The mapping between the area coordinates and the physical coordinates is:
Pnnod
x = i=1 Ni (L1 , L2 , L3 )xi
Pnnod
y = i=1 Ni (L1 , L2 , L3 )yi (3.17)

In the following there are the shape functions for linear, parabolic and cubic ele-
ments.

Linear element (nnod = 3)

Ni = Li 1≤i≤3 (3.18)

Quadratic element (nnod = 6)

N1 = (2L1 − 1)L1 N2 = (2L2 − 1)L2 N3 = (2L3 − 1)L3

N4 = 4L1 L2 N4 = 4L2 L3 N6 = 4L1 L3 (3.19)

Cubic element (nnod = 10)

1 1
N1 = L1 (3L1 − 1)(3L1 − 2) N2 = L2 (3L2 − 1)(3L2 − 2)
2 2
1 9
N3 = L3 (3L3 − 1)(3L3 − 2) N4 = (3L1 − 1)L1 L2
2 2
9 9
N5 = (3L2 − 1)L1 L2 N6 = (3L2 − 1)L2 L3
2 2
9 9
N7 = (3L3 − 1)L2 L3 N8 = (3L3 − 1)L1 L3
2 2
9
N9 = (3L1 − 1)L1 L3 N10 = 27L1 L2 L3
2
(3.20)

The position of the nodes on the reference triangle is described in [34]. Note that
the shape function N9 for cubic triangles reported in [34] contains a typo.

3.4 Choice of the basis

Several tests have been carried out with both BASIS A and BASIS B. From these
experiments, some guidelines have been established. When complex geometries are
studied, triangular elements with the BASIS B can be interesting thanks to their

35
 3 – Discontinuous Galerkin spatial discretization

ability to adapt to the geometry. On the other hand, quadrilateral elements with
the BASIS A can be very efficient in some problems, especially when structured
mesh are employed.
In order to show the effects related to the choice of the basis, some results on
the inviscid flow in a channel with a smooth bump are reported in the following.
More details on this test case can be found in Chapter 8.2.1. Two sequences of
meshes have been used for this experiment: a sequence of unstructured meshes
with quadrilateral elements and a sequence of unstructured meshes with triangular
elements. The elements at wall are cubic in all the meshes. A reconstruction order
p = 2 is assigned in all the elements and the Osher [35] numerical flux is used to
evaluate the convective fluxes. In Figure 3.3 and Figure 3.4 the initial meshes are
reported: finer meshes are obtained by isotropic splitting. In Figure 3.2 the Mach
field on one of the finest meshes is reported. In Figure 3.5 a plot of the L2 entropy
error versus the equivalent length scale is reported. The equivalent length scale h
is defined in Appendix B. Results are proposed for three different discretizations:
quadrilateral elements with BASIS A, quadrilateral elements with BASIS B and
triangular elements with BASIS B. It is clear that the BASIS A is much more
convenient than the BASIS B in the presence of quadrilateral elements.
The comparison between the results obtained with triangles and quadrilaterals
is more difficult, even in this simple test case. Indeed, when the nDOFs is low
the quadrilaterals with the BASIS A are significantly more convenient than the
triangles with the BASIS B. However, this behavior changes when the nDOFs is
increased. In conclusion, these results show that there is not a choice which is
always the best for all the situations. Triangles with BASIS B or quadrilaterals
with BASIS A can be both good choices, depending on the problem which is under
study. The only certain guidelines that can be obtained from this test is that
BASIS A is more efficient than BASIS B when quadrilaterals are employed.

Figure 3.2. Mach field for the inviscid subsonic flow in a channel.

36
 3.4 – Choice of the basis

Figure 3.3. Unstructured mesh with cubic quadrilateral elements.

Figure 3.4. Unstructured mesh with cubic triangular elements.

-3
10

Quad. BASIS A
Quad. BASIS B
-4 Tri. BASIS B
10
Entropy error

10-5

-6
10

0.01 0.02 0.03 0.040.05

h

Figure 3.5. Entropy error vs equivalent length scale for the

inviscid channel flow (p=2)

37
 3 – Discontinuous Galerkin spatial discretization

3.5 Convective fluxes

The convective fluxes represent the transport of conservative variables due to the
inviscid terms of the equations. In this work, they coincide with the fluxes of the
Euler equations. The numerical approximation of the convective fluxes in the DG
framework can be done with one of the several numerical fluxes developed for finite
volume schemes.

Since Euler equations are characterized by wave propagation phenomena, a

large family of upwind numerical fluxes has been developed through the years.
In the code developed for this work, four upwind fluxes have been implemented:
Osher, AUSM+, Roe and rotated RHLL. These fluxes take into account the prop-
agation of characteristic signals and introduce a numerical dissipation which is
controlled by physical phenomena associated to the hyperbolic nature of the Euler
equations.
The Osher flux was proposed by Osher and Solomon [35]. In this work, it has been
implemented following the approach proposed by Pandolfi [36]. This numerical
flux is based on an approximate solution of a Riemann problem at the interface
between two elements. First of all, kinematic variables are projected in the normal
and tangential direction of the interface. Then the Riemann problem is approxi-
mately solved in this frame of reference. The main idea of this method is to treat
the compression waves which can appear in the Riemann problem as isentropic
waves. In this way it is possible to obtain an algebraic solution of the Riemann
problem based on compatibility relations. This solution shows how signals prop-
agate across the interface and how they contribute to the value of the numerical
flux.
The AUSM+ [37] flux is one of the several upgraded versions of the original Advec-
tion Upstream Splitting Method (AUSM) [38]. It is an upwind method in which
the advection and the pressure contributions to the flux are computed separately.
The Roe [39] flux is based on the exact solution of a linearized Riemann problem.
In the original version, this numerical flux gives unphysical expansion shocks. In
order to prevent this behavior, the entropy fix proposed by Harten [?] has been
implemented in this work.
The rotated RHLL [41] numerical flux is obtained by combining together the Roe
flux and the HLL flux with a rotated-Riemann solver approach. The purpose of
this approach is to obtain a numerical flux which is very robust for nonlinear in-
stabilities. Indeed, it does not suffer from the carbuncle problem in the presence of
strong shock waves. More details on the carbuncle problem can be found in [42].
On the other hand, several numerical fluxes based on the central approach have
been used in the framework of DG methods. For example, the local Lax-Friedrichs
([43],[44]) flux is widely used and it has been implemented in this work. The great
advantage of this flux is related to its simplicity. On the other hand, it is one of

38
 3.5 – Convective fluxes

the most dissipative numerical fluxes.

In the following, a comparison between the local Lax-Friedrichs and the Osher
fluxes is reported. Computations were carried out on the inviscid channel problem
(see Chapter 8.2.1) with the mesh of Figure 3.3. The L2 entropy error is reported
as a function of the reconstruction order p in Figure 3.6. It can be seen that the
Osher flux gives an error which is always lower than the error given by the local
Lax-Friedrichs flux. However, when the order is increased, the absolute difference
between the two errors becomes negligible.
This behavior is true for smooth flows. However, when there are discontinuities
in the solution the upwind fluxes can show further advantages. Indeed, Wheatley
et al. [45] compared several numerical fluxes on discontinuous problems with DG
discretizations. They observed that in the presence of shock waves there is a
degradation in the numerical order of accuracy of the DG method and so it behaves
like low order methods. In this situation, the better accuracy of upwind solvers
becomes important.

Osher
LLF
Entropy error

-3
10

-4
10

0 1 2 3 4
p

Figure 3.6. Entropy error vs reconstruction order for Osher and Local
Lax-Friedrichs (LLF) fluxes.

39
40
Chapter 4

Diffusive fluxes

In the DG approach the global solution is not continuous but has jumps at the
interfaces between the elements. At first glance, these methods may seem quite
unsuitable for the discretization of elliptic equations because the discontinuous be-
havior of the numerical solution does not match well with the smoothness of the
solution of an elliptic problem. However, the application of the DG framework to
the solution of elliptic problems has proved to be remarkably successful. Several
approaches have been proposed for the study of elliptic problems in the DG frame-
work, see [46] for a comprehensive review. The first works related to these problems
appeared several years ago. Lions [47] proposed the use of a penalization approach
for the imposition of Dirichlet boundary conditions (BCs) in elliptic problems. His
idea was to substitute the exact Dirichlet BC with a weak BC which reproduces
the same effect of the original BC when a penalization factor tends to infinity.
Nitsche [48] proposed an improved technique which was proven to be consistent.
These ideas were subsequently adopted to impose inter-element continuity in the
context of internal penalization methods (see e.g. the work of Babuska and Zlamal
[49]) and were used to compute the diffusive fluxes exchanged between adjacent el-
ements. Bassi and Rebay [50] proposed a DG method (BR1) for the discretization
of the compressible Navier-Stokes equations which was subsequently generalized
by Cockburn and Shu [51] with the Local Discontinuous Galerkin (LDG) method.
Bassi, Rebay et al. [4] proposed also a second version of their scheme (BR2) in
which a compact stencil is obtained thanks to the use of lifting operators. Similar
motivations drove Peraire and Persson [?] to introduce the Compact Discontinuous
Galerkin (CDG) method which can be seen as a compact evolution of the LDG. A
contribution was also given by Baumann and Oden [52] with a method based on a
non-symmetric bilinear form. A different path was chosen by Gassner, Lorcher and
Munz [53] who proposed a numerical flux for the diffusion terms in finite volume
or DG schemes which is based on an exact solution of the diffusion equation with

41
 4 – Diffusive fluxes

piecewise polynomial initial data.

In 2005 van Leer and Nomura [3] proposed the recovery method (RDG) in which
diffusive fluxes are computed after a recovery procedure which gives a smooth
solution that in the weak sense is indistinguishable from the piecewise continuous
discrete solution. A similar approach was proposed by Dumbser [54] with the
PN PM scheme in which both convective and diffusive fluxes are computed through
a reconstruction of order higher than the local element reconstruction. Luo et
al. [55] presented a work in which they applied a reconstruction method for the
Navier-Stokes equations, following the approach originally introduced by Dumbser
et al. [56] in the context of Euler equations.
Other methods recently proposed exploit the idea of recovery. Huynh [57] per-
formed a stability study for several of these schemes. He also proposed a 1D method
which can be extended to 2D on quadrilateral meshes via tensor products. Borrel
and Ryan [58] proposed the elastoplast method in which the recovery idea is ap-
plied to a small rectangle overlapping the interface between two elements. French
et al. [59] introduced a modified recovery scheme which makes use of a penalty
term. Moreover, in time several versions of the original recovery scheme were pro-
posed by van Leer and co-workers. While the first version was based on a weak
formulation twice partially integrated (RDG2x), a second version based on a weak
formulation (RDG1x), requiring a single partial integration, was subsequently in-
troduced. Lo and van Leer [60] suggested that this second version could be more
suited to the discretization of the non linear Navier-Stokes equations, thanks to
some simplifications in the switching to primitive variables. For this reason, the
scheme proposed in this work will be compared mainly with the RDG1x scheme.
In this brief historical review it is possible to see some common lines which have
driven the evolution of DG methods for elliptic problems. One of these is the
compactness of the discretization. A compact stencil is very useful to simplify
the implementation of codes on unstructured grids, and to improve scalability in
parallel computations. Other important aspects are flexibility and robustness in
dealing with highly distorted meshes and p-adaptive algorithms.
During the research activity a new method (ESR) for the computation of dif-
fusive terms has been developed. It has been derived from the original recovery
method (RDG1x) focusing the attention on high order discretizations of convection-
diffusion problems.
In particular, the ESR method is characterized by a recovery basis which makes it
flexible and allows a robust implementation of p-adaptive algorithms. This basis
has less terms than the original recovery basis and remains well conditioned also in
the presence of highly distorted meshes. Furthermore the proposed discretization
shows a larger stability limit than the RDG1x method and the implementation is
simplified (Section 4.3.6). Moreover, a general procedure for the weak imposition
of Dirichlet BCs in the framework of recovery-based methods (Section 4.3.3) is
proposed and tested in the presence of curvilinear wall elements. The proposed

42
 4.1 – BR2 method

boundary procedure is truly local and acts only on the boundary element. In
contrast, the original procedure [61] proposed for the RDG1x method needs infor-
mation also from the neighbouring elements.
The results obtained with the proposed discretization were compared not only with
the original RGD1x method but also with the widely used BR2 method, which rep-
resents a reference in this field.

4.1 BR2 method

In this Section the BRMPS method introduced by Bassi, Rebay, Mariotti, Pedinotti
and Savini [4] is described. The method is also known in the literature as BR2 since
it can be seen as an evolution of the BR1 method proposed by Bassi and Rebay [50].
However, the BR2 method is compact and consistent for pure diffusion problems
while the BR1 method is not.
The BR2 method has been implemented in the code in order to make a comparison
with the proposed ESR approach.
The main idea of the BR2 approach is to introduce lifting operators which take into
account the influence of the jumps at the interfaces when computing the gradients
required by diffusive fluxes. Here the DG formulation already proposed in Eq.3.9
is reported for a diffusion problem after the application of the BR2 method:

Z
∂ũ  
[M] =− F̂h (uh± , zh |± ± ±
∂e )nx + Ĝh (uh , zh |∂e )ny Φds+
∂t ∂Ωe
Z   Z (4.1)
∂Φ ∂Φ
+ Fh (uh , zh ) + Gh (uh , zh ) dxdy + QΦdxdy
Ωe ∂x ∂y Ωe

For simplicity, a pure diffusion problem is considered and so F and G denote

diffusive fluxes. In the presence of convection-diffusion equations the convective
fluxes are computed separately as described in Section 3.5.

In Eq. 4.1 the dependence of the diffusive fluxes on the element solution (uh )
and on the modified gradients (zh and zh |± ∂e ) is stated. In the BR2 method the
fluxes across the faces of the element (F̂h and Ĝh ) are computed in a central way
as:
−
F (uh+ , zh |+
σ ) + F (uh , zh |σ )
−
F̂h = (4.2)
2
−
σ ) + G(uh , zh |σ )
G(uh+ , zh |+ −
Ĝh = (4.3)
2
where σ is a generic face of the element and the data from the two side of the face
are denoted by ±.

43
 4 – Diffusive fluxes

The modified gradients are computed as:

zh = ∇ h u h + R zh |σ = ∇h uh + ησ rσ (4.4)

The previous equation shows that the gradients are obtained by the gradient of
the element solution (∇h uh ) augmented by the global (R) or local (rσ ) lifting
operators. The coefficient ησ is a penalty parameter which should be chosen larger
than the average number of faces of the elements sharing the interface σ on which
the flux is evaluated.
For each face of the element a local lifting operator rσ : [L2 (σ)]2 → [Pp2 (Th )]2 is
defined by the following integral relation:

τ + + τh−
Z Z
τh · rσ (v)dxdy = − ( h ) · vds ∀τh ∈ [P2k (Th )]2 (4.5)
Ωe σ 2

where Eq. 4.5 is valid for all v ∈ [L2 (σ)]2 .

The global lifting operator R is the sum of all the local lifting operators of the
element: X
R(v) = rσ ((v)) (4.6)
∀σ

The BR2 method in convection-diffusion problems takes advantage of the facts

that the jumps at the interfaces are already computed for the evaluation of the
convective fluxes and so they can be immediately used to compute the local lifting
operators. Furthermore, all the functions which are used can be expressed in terms
of element basis functions and so no additive functions have to be introduced: this
is very useful in problems with deforming domains. In contrast, the ESR method
which will be described in the following requires the definition of recovery functions
which depends on the geometry and so have to be recomputed at each time step
when the mesh changes in time.

4.2 Original recovery method

The recovery method (RDG) proposed by van Leer and Nomura [3] is an elegant
way to obtain the numerical diffusive fluxes.
Consider the interface between two elements ΩA and ΩB . The goal is to re-
cover a smooth differentiable function (ur ) on the union of these two elements
(ΩC = ΩA ∪ ΩB ) in order to compute gradients on the quadrature points along the
interface. The key idea is that the recovered function has to be equivalent in the
weak sense to the original solution on the two elements, namely
Z Z
ur Φn dxdy = uh Φn dxdy 1 ≤ n ≤ NA
ΩA ΩA

44
 4.2 – Original recovery method

Z Z
u Φn dxdy =
r
uh Φn dxdy 1 ≤ n ≤ NB (4.7)
ΩB ΩB

A set of basis functions Ψ with Nr degrees of freedom can be introduced for the
recovered solution:
Nr
X
ur = ũr · Ψ = ũrj (t)Ψj (x, y) (4.8)
j=1

Substituting (4.8) and (3.3) into (4.7) the following two sets of linear equations are
obtained:

[RA ] ũr = [MA ] ũA

[RB ] ũr = [MB ] ũB (4.9)

in which [MA ], [MB ], ũA and ũB are the mass matrices and the vectors of the
degrees of freedom of the two elements A and B, respectively. The recovery matrices
[RA ] and [RB ] are defined as
Z
[RA ]nj = Ψj Φn dxdy 1 ≤ n ≤ NA , 1 ≤ j ≤ Nr
ΩA

Z
[RB ]nj = Ψj Φn dxdy 1 ≤ n ≤ NB , 1 ≤ j ≤ Nr . (4.10)
ΩB

Introducing the interface recovery matrix [R], the set of equations (4.9) can be
expressed in a more compact form:

[RA ] r [MA ] 0
    A
ũ
ũ = [R]ũ =
r
. (4.11)
[RB ] 0 [M B ] ũB

If an orthonormal basis is used, the mass matrix within each element is the identity,
and so the matrix multiplication on the right hand side can be removed,
 A
ũ
[R]ũ =
r
. (4.12)
ũB

From the previous derivation it can be seen that the recovery matrix [R] has
dimension (NA + NB ) × Nr . The most difficult task in the recovery approach is
the choice of the recovery basis Ψ.
Van Leer and Lo [61] considered the case in which NA = NB , and they set
Nr = (NA +NB ). They introduced a provisional recovery basis Ψ which makes the
recovery matrix non singular. The elegance of their idea is related to the next step.

45
 4 – Diffusive fluxes

They used the columns of the inverse recovery matrix to define a final recovery basis
Ψ̂ which is in the weak sense indistinguishable from the elements discontinuous
basis (van Leer and Lo [62]). In this way the degrees of freedom of the two elements
which share an interface can be used to put in evidence both the discontinuous
hyperbolic nature or the continuous parabolic nature of the underlying equations
simply by switching from a basis (Φ) to another (Ψ̂).

NA
X NB
X
ur = j Ψ̂j +
ũA j Ψ̂NA +j .
ũB (4.13)
j=1 j=1

4.3 Enhanced stability recovery method

4.3.1 Derivation of the method
The approach proposed by van Leer, Nomura and van Raalte [61] makes it possible
to enhance the accuracy of the viscous fluxes discretization because the recovered
solution has twice the number of degrees of freedom with respect to the element
solution. This advantage is important in pure diffusion problems but can be lost
in convection-diffusion problems if the convective fluxes are computed using the
element internal reconstruction. There are some approaches in which the informa-
tion from the neighbouring elements is used to improve also the accuracy of the
convective terms, see for example the PNPM method [54] or the approach proposed
by Nourgaliev et al. [63]. One of the main reasons for the current success of DG
methods is the possibility of obtaining high order reconstructions without the need
of data from neighbouring elements. For this reason, a classical approach has been
adopted in this work and so the convective fluxes are computed using the internal
element reconstruction. As a consequence the effects of a higher order approxima-
tion of the diffusion terms are mitigated by the lower order approximation of the
convective terms.
Furthermore, in the original recovery approach it is necessary to choose a provi-
sional recovery basis (Ψ) which makes the recovery matrix ([R]) square and invert-
ible. This is not a trivial task in multidimensional problems, especially when high
order reconstructions are used on irregular unstructured meshes. For example, van
Leer et al. [64] reported that the RDG1x scheme becomes unstable for p > 2, if the
gradient in the volume integral is computed with one of the approaches described
in [60]. In contrast, the discretization proposed in this work was successfully tested
up to p = 6 (see Section 4.4.3).
Van Leer, Nomura and van Raalte [61] chose to use an anisotropic monomial recov-
ery basis in which the reconstruction orders in the direction normal and tangential
to the interface are (2p + 1) and p, respectively. Even if this approach improves
the order of the scheme on cartesian meshes, a standard (p + 1) convergence order

46
 4.3 – Enhanced stability recovery method

is obtained on general unstructured meshes [62].

One of the final goals of this work is the discretization of convection-diffusion
problems in which p-adaptivity strategies can bring important benefits. In this
context, it is not clear how to extend the RDG1x approach (Eq. 4.13 which re-
quires Nr = NA + NB ) to problems in which the reconstruction order changes from
one element to another (NA = / NB ). In the literature there are no applications
of the original recovery approach with p-adaptive algorithms to the best of our
knowledge.
Thus an alternative approach has been chosen in this work in order to obtain a
general and flexible scheme (Enhanced Stability Recovery (ESR) [2]), which can
be easily adopted in the context of p-adaptive algorithms. In particular the con-
straint of a square recovery matrix is removed (i.e. it is set Nr < NA + NB ) and
the over-determined system (4.12) is solved by the least square method:
 A
ũ
[R]T [R]ũr = [R]T (4.14)
ũB
Thanks to a proper choice of the recovery basis (defined in the following) the least
square problem is well conditioned and the matrix [R]T [R] is invertible:
 A  A

−1 ũ ũ
ũr = [R]T [R] [R]T = [ R̃] (4.15)
ũB ũB
In this way it is sufficient to perform a matrix-vector product for each interface
and the recovery coefficients are obtained. Then these coefficients can be multi-
plied by the gradient of the recovery basis functions in order to obtain the diffusive
fluxes at the interface quadrature points (first term on the right hand side of Eq.
(3.9)). If the domain is fixed, the rectangular recovery matrix ([R̃]) can be pre-
computed since it depends only on the mesh. In Section 4.3.6 more details on the
implementation are given.
In the proposed approach, the recovered solution is used only for the evaluation of
the diffusive fluxes at the interface. Indeed, the diffusive fluxes within the volume
integral (second term on the right hand side of Eq. (3.9)) are computed inside
each element, and therefore are given by the product between the element degrees
of freedom and the gradient of the element basis functions.
The most important aspect of the method is the choice of the recovery basis
Ψ. An orthonormal basis obtained by the modified Gram-Schmidt procedure is
used for this purpose. The orthonormalization uses the L2 inner product (f, g) =
f gdxdy in which the integral is evaluated on the union of the two elements.
R
ΩA ∪ΩB
This approach makes it possible to ensure that the condition number of the recov-
ery system remains low.
In this work, the orthonormalization procedure is initialized with a particular start-
ing monomial basis defined according to the following rules, depending on the de-
gree pA and pB of the polynomials in the two elements A and B, defining the

47
 4 – Diffusive fluxes

interface. Three cases will be described depending on the order of the two ele-
ments. Consider a reference frame centered on the mid-point of the interface and
with the axis aligned to the normal and tangential directions (ξ, η) as shown in
Figure 4.1:

η ξ

Figure 4.1. Interface reference frame for constructing the recovery basis.

1
ξ η
ξ2 ξη η2

ξ3 ξ2η ξη2

Figure 4.2. Preliminary recovery basis before orthonormalization for the case
pA = pB = 2, in the Enhanced Recovery approach.

1 ξ ξ ξ3 ξ4 ξ5
η ηξ ηξ2 ηξ3
η2 η2ξ

Figure 4.3. Preliminary recovery basis with the original recovery approach
for the case pA = pB = 2.

• pA = pB = p
First of all we consider the case in which the two elements have the same
order p. The starting recovery basis for the initialization of the orthonormal-
ization procedure is obtained by considering the monomial basis of order p in
the variables (ξ, η) and adding all the monomials with order p + 1 in which
ξ appears. In other words these additive functions consist in the first p + 1
monomials of the (p + 1)th row of the Pascal triangle. The tests done in this
work have shown that these additive functions make the difference. Indeed,
the preliminary experiments showed that if the additive functions are not
used or if only the term ξ p+1 is included, the convergence rate of the scheme
can be compromised.

48
 4.3 – Enhanced stability recovery method

Since the order of the recovery reconstruction is reduced with respect to the
original recovery approach, it is easier to control the condition number of the
recovery matrix. Moreover, this choice is related to the improvement in the
stability limit observed in the numerical experiments presented below.
An example of the proposed starting basis in the case p = 2 is shown in
Figure 4.2. For the sake of comparison, Figure 4.3 reports the provisional
basis for p = 2 obtained according to the original recovery approach.

• |pA − pB | = 1
The starting recovery basis is obtained by considering the monomial basis
of order p̃ = max(pA , pB ) and adding all the monomials of order p̃ + 1 in
which ξ appears. It is possible to use the maximum order between pA and
pB if there is enough information in the union of the two elements to obtain
a recovered solution with this accuracy, i.e. Nr ≤ NA + NB . This condition
is satisfied if min(pA , pB ) ≥ 1. Thus, the p-adaptive strategy avoids the use
of elements with p = 0.

• |pA − pB | > 1
It could happen that a p-adaptive algorithm introduces a jump of two or
more orders between neighbouring elements. In this case it is impossible
to choose the recovery basis with p̃ = max(pA , pB ). Indeed, in the case
|pA − pB | > 1, the recovery degrees of freedom would be more than the
sum of the degrees of freedom in the two elements if p̃ = max(pA , pB ).
For this reason, the order of the recovery is locally reduced, choosing p̃ =
max(pA , pB ) − 1, i.e. at the interface between elements with a different
accuracy, the algorithm chooses an intermediate value for the degree of the
polynomial being reconstructed. This reflects the fact that the adaptive
algorithm assigns a jump in the accuracy across the interface, so it seems
reasonable to choose an intermediate state at the interface. Thus, in this case,
the starting basis is built from a monomial basis of order p̃ = max(pA , pB )−1
plus all the monomials of order p̃ + 1 in which ξ appears. In the performed
tests it was found that the presence of strong jumps in the order distribution
across the domain should be prevented whenever possible (see Section 7.2.1).
For this reason, the condition |pA − pB | > 1 should be avoided during the
simulations by choosing a more regular order distribution. However, if the
condition is verified the algorithm is able to manage this case, as described
above.

Therefore, the starting recovery basis has always Nr = (p̃ + 1)(p̃ + 2)/2 + (p̃ + 1)
degrees of freedom, where p̃ is equal to p if a uniform order distribution is used.
Otherwise p̃ can be set equal to max(pA , pB ) or max(pA , pB ) − 1, depending on

49
 4 – Diffusive fluxes

the jump in the order.

As far as numerical quadrature is concerned, the integral on the interface is com-
puted with a suitable quadrature formula in order to satisfy the accuracy require-
ments of both elements. In particular, if the two elements require MA and MB
points for the integrals on their edges, the formula with M = max(MA , MB ) is
adopted.

Finally, a comment on the use of piecewise constant reconstructions. When

pA = pB = 0 the recovery reconstruction has only two degrees of freedom: a con-
stant term and a linear term related to the direction normal to the interface. From
this information, only the normal derivative can be obtained from the recovery
solution. If the heat equation is considered this information is enough to compute
the fluxes across the interface. On the other hand, if a more complex system of
equations is considered (i.e. Navier-Stokes), not only the normal derivative but
also the tangential derivative has to be computed in order to evaluate the fluxes.
For this reason, the presence of elements with p = 0 is avoided in the domain. Note
that the DG method is usually chosen for its high accuracy. Thus, the occurrence
of piecewise constant reconstructions is expected to be unlikely in most applica-
tions.

It is important to emphasize the use of the Gram-Schmidt orthonormalization

not only on the element basis but also on the recovery basis. This approach makes
the method robust and able to deal with highly distorted meshes. For example the
effect of the recovery basis orthonormalization on the 2-norm condition number
of the matrix [R]T [R] which has to be inverted in Eq. 4.15 is reported in Ta-

ble 4.1. The data refer to the interface between the two distorted elements shown
in Figure 4.4 for the case pA = pB . Orthonormalization is applied on the element
basis functions in both cases, but only in the second case it is also applied to the
recovery basis. These results suggest that the use of orthonormalization on the
proposed recovery basis becomes fundamental when high-order reconstructions are
employed on irregular meshes. The condition number of the matrix was estimated
with the free software package GNU Octave [65].
It can be expected that the use of the orthonormalization procedure on the recov-
ery basis can be useful also for other recovery-based schemes, since they all involve
the definition of a recovery procedure which is influenced by the behavior of the
recovery basis.

50
 4.3 – Enhanced stability recovery method

Figure 4.4. Interface between two distorted elements for the test of Table 4.1.

Scheme Cond. number without ON Cond. number with ON

DG1 4.61E+02 1.97E+00
DG2 1.05E+04 3.87E+00
DG3 2.52E+05 5.36E+00
DG4 6.25E+06 6.62E+00
DG5 1.58E+08 7.87E+00
DG6 5.03E+09 9.09E+00


Table 4.1. Condition number of the matrix [R]T [R] for the interface of Fig-
ure 4.4 without and with orthonormalization (ON) of the recovery basis

51
 4 – Diffusive fluxes

4.3.2 Extension to 3D
In this Section some considerations on the extension of the method to 3D are re-
ported. In particular it is necessary to define how to build the recovery basis in 3D.
Consider a reference frame with the origin in the geometric center of the interface.
Its axis are aligned with the normal and tangential directions (ξ, η, ζ).
First of all, the case pA = pB = p̃ is studied. The starting recovery basis can be
obtained by considering the monomial basis of order p̃ and adding all the monomi-
als with order p̃ + 1 in which ξ appears. With this approach the size of the recovery
basis is Nr = (p̃ + 1)(p̃ + 2)(p̃ + 3)/6 + (p̃ + 1)(p̃ + 2)/2.
The cases |pA − pB | = 1 and |pA − pB | > 1 can be managed in a way similar to the
2D approach.
The same problems described for p = 0 in the 2D context are present also in 3D.
Another important issue which can become important in 3D is the memory re-
quirement. In Section 4.3.6 there are suggestions on implementation details which
can help in saving memory. Future work will be devoted to the assessment of the
method in 3D.

4.3.3 BCs for ESR

In this Section the procedure for the imposition of boundary condition for diffu-
sive equations is described in the framework of the Enhanced Recovery Method.
Neumann BCs are simply satisfied by setting the exact value of the fluxes on the
boundary integral. Dirichlet BCs require further considerations.

In the framework of discontinuous Galerkin methods it is possible to impose

Dirichlet BCs in both strong or weak form [66]. The strong approach consists in
forcing the solution reconstruction to exactly satisfy a prescribed condition on the
boundary. The weak approach is based on the idea that the error in the enforce-
ment of the BC’s and the error in the interior solution can be of the same size.
In other words, the accuracy required in the interior of the domain is the same as
the accuracy imposed at the boundary: in this approach therefore there is an error
between the value prescribed and the solution reconstructed on the boundary, but
this error vanishes in the limit of an infinite mesh refinement, with the same rate
of the error in the interior of the domain.

From a practical point of view, it can be shown that the use of strong Dirichlet
BCs can introduce stability limitations and spurious oscillations in the numerical
solution (see for example [67]). Furthermore several discretization methods allow
the weak imposition of Dirichlet BCs in a natural way. For example, in the BR2
scheme this is simply done through the definition of a proper ghost state in all
the boundary quadrature points. The extension to this approach in recovery-based

52
 4.3 – Enhanced stability recovery method

schemes could be non trivial. Indeed, the recovery procedure is applied to the
union of two elements and so a ghost element would be required in this case, not
just a local ghost state. The definition of the proper values in the ghost element
can be obtained through symmetry considerations between the boundary element
and the ghost element. For example Borrel and Ryan use this approach in their
paper [58] on the elastoplast method. The problem is that this procedure may
become ambiguous in the presence of curvilinear elements, because in this case the
boundary curve is not a symmetry axis between the boundary element and the
ghost element.
Another path was followed by van Leer, Nomura and van Raalte [61]. They pro-
posed to impose the BCs on the recovery solution. They substitute the missing
information in the boundary element with data from one or more interior elements.
They observed that this procedure may reduce the stability range of the method.
Furthermore this approach is not trivial when unstructured meshes are used, be-
cause there are multiple possible choices.

Two of the main reasons for the success of DG methods are the compactness
of the reconstruction and the simplification in the treatment of BC’s when high
order reconstructions are employed. In order to keep these positive features, an
alternative approach has been developed in this work which makes it possible to
impose both strong or weak Dirichlet BC’s in recovery-based DG methods, without
the need to use ghost elements or data from other interior elements.

4.3.4 Dirichlet BC’s in strong form

In the proposed approach boundary conditions are imposed on the element internal
reconstruction, not on the recovery solution.
The strong imposition of Dirichlet BC’s requires that the numerical solution at
the quadrature points of the boundary elements is equal to the prescribed values.
In other words, the degrees of freedom inside the element have to be altered in
order to satisfy the BC’s. Consider for example a discontinuous Galerkin scheme
with Ne degrees of freedom inside the element and M quadrature points on the
edge (M < Ne ). The imposition of the BC’s on the quadrature points gives M
equations for each element on the boundary:
Ne
X
w = uw (xq ) =
um (ũi + ∆ũi )Φi (xqm ) 1≤m≤M (4.16)
m

i=1

where umw represents the prescribed wall value in the quadrature point m.
If the procedure is forced to be conservative the mean value has to be preserved
(∆ũ1 = 0). In this way the system (4.16) has Ne − 1 unknowns which describe
the perturbation of the elements degrees of freedom (∆ũi , 2 ≤ i ≤ Ne ). In general

53
 4 – Diffusive fluxes

the set of equations (4.16) is under-determined and so more information has to

be imposed. In this work it has been chosen to minimize the L2 -norm of the
perturbation ∆ũ:

Ne
X
min(|∆ũ|2 ) = min( ∆ũ2i ). (4.17)
i=2

The minimization problem can be solved using a Lagrange multiplier, µ. In this

approach, equation (4.17) is the goal and equations (4.16) are the constraints:

"N M Ne
!#
e
∂ X X X
∆ũ2i + µm um
w − (ũi + ∆ũi )Φi (xq )
m
= 0. (4.18)
∂∆ũi i=2 m=1 i=1

In other words it is necessary to look for the smallest perturbation which is able to
satisfy the BC’s. The set of equations (4.18) and (4.16) defines the linear system
(4.19) in which there are (Ne − 1 + M ) equations and (Ne − 1 + M ) unknowns:
the perturbations of the element degrees of freedom (with the exception of ∆ũ1
which is set to zero and is already known) plus the Lagrange multipliers. The
boundary matrix [B] and the boundary vector {W} are explicitly written in (4.20)
and (4.21):

∆u2 
 

.. 

 
.

 


 

  ∆uNe  

 
B = W (4.19)
 µ1 
.. 

 

. 

 

 
 
µM
 

2 0 ··· 0 −Φ2 (xq1 ) ··· −Φ2 (xqM )

 
.. .. ..
 0 2 0 . . .
 
··· 
.. .. .. ..
 
. 0 . 0 . .
 
   ··· 
B = 0 0 2 −ΦNe (xq1 ) −ΦNe (xqM ) (4.20)
 
··· ···
 Φ2 (x1 ) ΦNe (xq1 ) 0 0
 
q ··· ··· ··· 
.. .. .. ..
 
 
 . ··· ··· . . ··· . 
Φ2 (xqM ) ··· ··· ΦNe (xqM ) 0 ··· 0

54
 4.3 – Enhanced stability recovery method

0
 
..

 

 
.

 


 

0

 

W = (4.21)

PNe .
 uw1 − i=1 ũi Φi (xq1 ) 
..

 

.

 


 

 PNe 
uwM − i=1 ũi Φi (xq )
M
 

The solution of this system gives the perturbations which have to be applied to
the element. In particular, the Lagrange multipliers are not needed and they can
be eliminated. Thus it is possible to define a reduced boundary matrix [B̂] such
that:  1 PNe
 ∆u2   uw − i=1 ũi Φi (xq ) 
1
  
   
.. ..
  
.  = B̂ . = B̂ Ŵ . (4.22)
∆uNe
 
 M PNe 
uw − i=1 ũi Φi (xq )
M
  

The reduced boundary matrix [B̂] can be obtained by the first Ne − 1 rows and the
last M columns of the inverse of [B]. At this point the gradient required by the
boundary fluxes can be computed with the product between the perturbed degrees
of freedom and the gradient of the element basis functions.

4.3.5 Dirichlet BCs in weak form

In the previous paragraph a procedure for the strong imposition of Dirichlet BC’s
has been presented. The element’s degrees of freedom are perturbed in such a
way that the element internal reconstruction exactly satisfies the condition on the
boundary. After that, the gradient of the element solution is computed at the
boundary quadrature points, yielding the boundary fluxes.
A similar procedure is adopted also for the weak imposition of the BC’s. The
only difference is that, after the fluxes computation, the perturbation previously
applied to the element is removed. In this way the BC’s are weakly imposed in the
sense that they determine the value of the boundary fluxes, but there is not a direct
alteration of the element’s degrees of freedom. The procedure can be summarized
as follows

• Compute the element’s perturbation ∆ũ with Eq. (4.22).

• Apply the perturbations to ũ, ũ := ũ + ∆ũ.

PNe
• Compute the boundary gradients ∇u = i=1 ũi ∇Φi .

• Compute the boundary fluxes F̂h (∇u)

55
 4 – Diffusive fluxes

• Remove the perturbations ũ := ũ − ∆ũ.

All numerical results obtained with the ESR method which are presented in the
following were computed with this weak enforcement of BC’s.

4.3.6 Implementation and computational cost

In this Section the implementation of the method is briefly described in order to
identify the main contributions to the computational cost. In particular some con-
siderations are made in order to compare the proposed approach with the RDG1x
recovery method. Both schemes are compact in the sense that the computation
of viscous fluxes requires the knowledge of information only from the first neigh-
bours of the element. For simplicity a scalar 2D problem in which all the elements
have the same order (p) is considered. In this case Ne = (p + 1)(p + 2)/2 and
Nr = (p + 1)(p + 2)/2 + (p + 1). The domain is supposed to be fixed, so that the
mesh does not change in time. The cost required by the computation of the inte-
grals appearing in Eq. (3.9) can be decomposed in different contributions, analyzed
below.

4.3.7 Interface integral computational cost

The analysis starts from the computational cost due to the evaluation of the dif-
fusive fluxes on an interface with M quadrature points.
If the ESR method is used the following operations are required for each interface
quadrature point:
n Ao n Ao
∇u = [Dr ]ũr = [Dr ][R̃] ũũB = [Dr R̃] ũũB .
The matrix with the derivatives of the recovery basis functions [Dr ] and the recov-
ery matrix [R̃] can be computed and multiplied at the beginning of the simulation
since they depend only on the mesh. The cost for each interface is proportional to
M × 2 × 2Ne .

In the RDG1x scheme the recovery coefficients ũr coincide with the degrees of
freedom inside the two elements and so the gradient can be obtained as
n Ao
∇u = [Dr ] ũũB .

Thus, the cost for each interface is proportional to M ×2×2Ne also in this case.

56
 4.3 – Enhanced stability recovery method

4.3.8 Volume integral computational cost

The computational cost of the volume integral will be evaluated assuming that L
quadrature points are used.
The cost for the ESR method is related to the following operations:

• ∇u = [D]ũ
In each quadrature point the fluxes are computed using the internal element
solution. The cost is proportional to L × 2 × Ne . The matrix with the
derivatives of the element basis functions [D] can be precomputed since it
depends only on the mesh.

The cost for the RDG1x method is related to the following operations:
PNF
• ∇u = N1F i=1 [Dr ]i ũir
In each quadrature point the gradient is computed as an average between
all the gradients obtained by the several available recovery solutions (one for
each face of the element). In this case the cost is proportional to L × NF ×
2 × 2Ne where NF is the number of faces of the element. Lo and van Leer
[60] proposed also another approach which improves accuracy but is more
involved. In practice they use the recovery solution only for evaluating the
derivative in the coordinate normal to the interface and, since there are more
directions than dimensions, they introduce a least square approach. This
implies that, in each internal quadrature point, a least square problem has
to be solved. The cost in this case is greater than in the previous approach
and the implementation is significantly more involved.

The previous considerations suggest that the main difference between the cost
of the ESR method and the RDG1x method is related to the computation of the
volume integral. Indeed, in the described implementation the cost per interface
is the same for the two methods even if they use recovery bases with a different
size. However, the RDG1x method requires more work than the ESR method in
the volume integral, because it uses all the several recovery functions defined in
the element (one for each face). In contrast, the ESR method is based on a very
simple evaluation of the volume integral since it directly uses the gradient of the
element solution. From a practical point of view, the CPU time per iteration re-
quired by the two methods might be similar but it is strongly influenced by the
implementation (cache effects can significantly alter the speed) and by possible
memory limitations which can forbid the storage of precomputed matrices.
Since the test cases considered in this work are 2D, there are no memory limita-
tions, and all quantities that can be precomputed will be stored in memory at the
beginning of each computation. However, memory requirements can become an
important issue in 3D problems. In order to reduce memory usage it is useful to

57
 4 – Diffusive fluxes

avoid the storage of the interface recovery matrix ([R̃]). Indeed, it is more conve-
nient to pre-compute and store the product between the matrix of the derivatives
of the recovery basis functions ( [Dr ]) and the recovery matrix ([R̃]), for each
quadrature point.
In the example described in this Section a scalar problem is studied. If a system of
equations is considered (i.e. Navier-Stokes) the recovery approach gives the gradi-
ents of the conservative variables. In order to compute the fluxes it is necessary to
switch to the gradients of the primitive variables. This can be done by computing
the values of the conservative recovered variables at the interface. In this case too,
it is possible to save memory, avoiding the storage of the recovery matrix and stor-
ing the product between the matrix containing the recovery functions, evaluated
at the interface, and the recovery matrix.

The previous analysis considers only a problem with a fixed grid. If deforming
meshes are employed, the recovery functions and the related matrices have to be
re-computed during the simulation. This additional cost can be significant for
both the RDG1x and the ESR methods. In this kind of problems recovery-based
methods might be less efficient than other existing approaches, like, for example,
the BR2 method.

4.4 Tests and comparisons for the ESR method

4.4.1 1D Poisson equation
In order to evaluate the behavior of the proposed discretization a 1D Poisson prob-
lem was studied with a time-marching approach. The problem has the following
exact solution which also defines Dirichlet BC’s:

u(x) = 1 − x + sin(2πx) 0 ≤ x ≤ 1. (4.23)

Van Leer et al. [64] reported a comparison between several schemes (with p = 1 and
forward Euler) for this test case. In Table 4.2 the dimensionless time step ν∆t/∆x2
obtained for the ESR method are reported and compared with the stability limit
for the BR2 and then RDG1x methods from [64]. Table 4.3 shows the L2 error on
the temperature and the convergence order obtained in the present implementation
for RDG1x, BR2 (ηBR2 = 1) and the proposed ESR scheme. All results shown in
the table, were obtained with p = 1 and forward Euler time integration.
Particular attention was dedicated to the BC’s procedures originally proposed by
the different authors. In particular for the RDG1x scheme, BC’s are imposed on
the recovery solution, exploiting information from the boundary and the first two
elements close to it, as in [61]. For the BR2 scheme a proper ghost state computed
as a function of the BC’s is defined. For the ESR method the weak minimization

58
 4.4 – Tests and comparisons for the ESR method

procedure proposed in Section 4.3.5 is used. The performed numerical experiments

show that the ESR method has a significantly larger stability limit than the other
schemes. On the other hand RDG1x gives excellent results and is fourth order
accurate (2(p + 1)) while the other two schemes are second order accurate (p + 1).
This advantage of the RDG1x method is lost on multidimensional unstructured
meshes in which its convergence order reduces to p + 1 [61]. Even if the ESR
method obtains the expected convergence order, it can be seen that its error is
60% larger than the error obtained by BR2 when p = 1. On the other hand, the
time step for ESR can be chosen 65% larger than it is possible in BR2. From
this first comparison, it can be seen that in pure diffusion problems with cartesian
meshes the RDG1x method is the most accurate scheme between the three tested.
Scheme ν∆t/∆x2
RDG1x (from [64]) 0.08
BR2 (from [64]) 0.10
ESR 0.165

Table 4.2. 1D Poisson problem: stability limit (p=1, forward Euler)

Mesh RDG1x L2 error Order BR2 L2 error Order ESR L2 error Order
8 8.90E-04 - 3.06E-02 - 5.48E-02 -
16 4.24E-05 4.39 8.67E-03 1.82 1.44E-02 1.93
32 2.18E-06 4.28 2.25E-03 1.95 3.69E-03 1.96
64 1.21E-07 4.17 5.66E-04 1.99 9.35E-04 1.98

Table 4.3. 1D Poisson problem: L2 errors and convergence order (p=1)

4.4.2 1D Convection Diffusion problem

The final goal of this work is the application of the proposed discretization in
convection-diffusion problems. For this reason the following 1D model problem
was studied with a time marching approach:
∂u ∂u ∂ 2 u
+ 10 − =0 0≤x≤1 u(0) = 1, u(1) = 0. (4.24)
∂t ∂x ∂x2
Simulations were carried out until the residual reached machine precision. Then,
the converged numerical solution was compared with the exact steady state solution

exp(10x) − exp(10)
u(x) = . (4.25)
1 − exp(10)

59
 4 – Diffusive fluxes

Several computations were performed with the RDG1x, the BR2 and the ESR
methods (p = 1) and the first order forward Euler time integration scheme. Three
non uniform meshes with 5, 10 and 20 elements were used. The meshes are clustered
in order to reproduce the aspect of the typical grids used in Navier-Stokes boundary
layer simulations. The size of the element near the boundary x = 1 is ∆x =
0.08, 0.04 and 0.02 for the different meshes. In Table 4.4 the L2 -errors are reported
for the three schemes. While in the previous test case the RDG1x scheme gave
errors which are several orders of magnitude lower than the other schemes, now
the errors for the different methods are significantly closer. The ratio between the
error of the ESR and the RDG1x methods is equal to approximately 3-4 for the
different meshes. Furthermore the stability limit of the discretization is dominated
by the diffusive terms and so the ESR method can benefit from its double time
step with respect to the RDG1x (the data in Table 4.2 still hold). These results
are limited to a scalar linear problem with p = 1 but they suggest that when
convective effects appear in the equations, the global errors obtained by the ESR
and the RDG1x methods become closer because the bottle neck is represented by
the convective discretization. This difference could become even smaller in high
Reynolds number industrial flows, in which the convective terms are particularly
strong and the meshes are often highly distorted.

Mesh RDG1x L2 error BR2 L2 error ESR L2 error

5 5.58E-03 1.39E-02 1.77E-02
10 1.58E-03 4.25E-03 6.08E-03
20 4.25E-04 1.56E-03 1.82E-03

Table 4.4. 1D Convection-diffusion problem: L2 errors at steady state (p=1)

4.4.3 2D Heat equation

In this Section the behavior of the proposed scheme is evaluated in 2D for several
accuracies, including some very high order schemes. The linear 2D heat equation
was solved in a square domain (0,1) × (0,1) with unstructured triangular meshes.
The meshes were obtained starting from uniform cartesian grids in which quadri-
laterals were split in triangles. The following Dirichlet BC’s were prescribed:

T = 0 on x = 1, y = 0 and y = 1

(4.26)
T = sin(kw πy) on x = 0.
For the ESR method, BC’s were imposed with the weak approach, based on
the minimization procedure of Section 4.3.5. The parameter kw is an integer which
represents the number of semi-waves in the y-direction. In this test it was set to

60
 4.4 – Tests and comparisons for the ESR method

kw = 1. The numerical solution was initialized with a uniform field T0 = 1 and

computations were carried out until the residuals reached machine precision. The
L2 norm of the temperature error was computed through a comparison between
the obtained steady state results and the following exact analytical solution, which
is reported in Figure 4.5:
sinh(kw π(1 − x)) sin(kw πy)
T = . (4.27)
sinh(kw π)

Figure 4.5. Exact temperature field for the 2D heat diffusion problem.

The test was also performed with the widely used BR2 scheme in order to
compare the behavior of the two methods. The stabilization parameter associated
with this last scheme was set to ηBR2 = 4 ( ηBR2 > NF according to [32]). The
BC’s for the BR2 scheme are weakly imposed by means of proper ghost values on
the boundary quadrature points.
The mesh refinement analysis is reported in Table 4.5. It can be seen that even
with few elements, both schemes reach quickly the theoretical convergence order.
In Figure 4.6 a direct comparison of the two methods is presented. The error is
given as a function of the equivalent length scale (h), which depends on the number
of total degrees of freedom (nDOF s): h = (nDOF s)−0.5 . The plot shows that the
error for the ESR method is greater than the error for the BR2 method but, when
the order is increased, the two schemes give comparable errors.
These numerical experiments were performed with a time marching approach in
which explicit integration schemes were employed. The stability limits for the ESR
method and the BR2 method are reported in Table 4.6. It is interesting to see that
the allowed time step for the ESR method is approximately 4 times larger than
the allowed time step for the BR2 method when p = 1. The ratio tends to grow
when the order is increased and it is near 8 for p = 6. This behavior, together with
the fact that for high orders the two schemes give comparable errors, suggest that

61
 4 – Diffusive fluxes

Scheme Mesh Error ESR Order ESR Error BR2 Order BR2
DG1 2 × (2 × 2) 7.33E-02 - 2.98E-02 -
2 × (4 × 4) 2.02E-02 1.86 1.23E-02 1.28
2 × (8 × 8) 4.92E-03 2.04 3.88E-03 1.66
DG2 2 × (2 × 2) 2.15E-02 - 8.79E-03 -
2 × (4 × 4) 2.00E-03 3.43 9.77E-04 3.17
2 × (8 × 8) 2.47E-04 3.02 1.14E-04 3.10
DG3 2 × (2 × 2) 1.75E-03 - 1.08E-03 -
2 × (4 × 4) 9.17E-05 4.25 7.38E-05 3.87
2 × (8 × 8) 5.32E-06 4.11 4.79E-06 3.95
DG4 2 × (2 × 2) 1.45E-04 - 1.23E-04 -
2 × (4 × 4) 4.90E-06 4.89 4.25E-06 4.86
2 × (8 × 8) 1.45E-07 5.08 1.41E-07 4.91
DG5 2 × (2 × 2) 1.51E-05 - 1.24E-05 -
2 × (4 × 4) 2.96E-07 5.67 2.12E-07 5.87
2 × (8 × 8) 4.53E-09 6.03 3.44E-09 5.95
DG6 2 × (2 × 2) 1.25E-06 - 8.82E-07 -
2 × (4 × 4) 9.74E-09 7.00 7.60E-09 6.86
2 × (8 × 8) 7.79E-11 6.97 6.15E-11 6.95

Table 4.5. L2 -norm of the temperature error for the ESR and the BR2 methods.

Scheme CFLmax ESR CFLmax BR2

DG1-RK2TVD 6.2E-02 1.4E-02
DG2-RK3TVD 3.9E-02 8.5E-03
DG3-RK4LS 5.6E-02 8.6E-03
DG4-RK4LS 3.5E-02 5.0E-03
DG5-RK4LS 2.4E-02 3.2E-03
DG6-RK4LS 1.6E-02 2.1E-03

Table 4.6. Stability limit with explicit time integration for 2D heat diffu-
sion on unstructured meshes.

the ESR method becomes competitive for explicit time integration of high order
reconstructions. On the other hand, the fact that the BR2 method gives a lower
error for all the reconstruction orders considered here, makes it a general purpose
scheme, with a wider applicability than the ESR method.
It is very hard to compare the efficiencies of the different schemes. A possible
approach might be to evaluate the total computational cost required to obtain a
certain error level. However, the CPU time obtained is strongly related to the

62
 4.4 – Tests and comparisons for the ESR method

particular implementation chosen, and to the nature of the problem under study.
Indeed, the presence of memory limitations (typically in 3D) or deforming grids
can dramatically increase the computational cost, since several quantities cannot be
stored but have to be re-computed during the simulation. Furthermore, the choice
of the time integration scheme can enhance or cancel some benefits of the different
methods. For example, in steady state problems, implicit integration schemes are
commonly used. Their large stability domain can reduce the advantages observed
in the explicit time integration with the ESR method.

10-1

-3
10

DG1-ESR
-5 DG2-ESR
10
Error

DG3-ESR
DG4-ESR
DG5-ESR
DG6-ESR
10-7 DG1-BR2
DG2-BR2
DG3-BR2
-9 DG4-BR2
10 DG5-BR2
DG6-BR2

-11
10
0.05 0.1 0.15 0.2
h

Figure 4.6. L2 -norm of the temperature error vs the equivalent length scale h.

4.4.4 Laminar boundary layer on a flat plate

This test case was proposed at the First, Second and Third International Workshop
on High-Order CFD Methods ([68], [11] and [69]). The flow is characterized by
a Reynolds number based on the plate length equal to Re∞ = 106 . The angle of
attack is 0◦ and the free-stream Mach number is M∞ = 0.5. The plate is adiabatic
and has unit length.
The purpose of this test simulations is to evaluate the behavior of the ESR
method in the presence of highly clustered meshes. In particular one of the meshes

63
 4 – Diffusive fluxes

proposed by Bassi and Colombo was used. It can be downloaded from the Work-
shops website (file a1-125-2s.msh, [11]). This grid contains 560 quadrilateral ele-
ments. The distance between the inlet and the leading edge is 1.25. The distance
between the freestream and the plate is 2. The size of the elements near the wall
is 3.75 · 10−4 in the normal direction. The most stretched element in the domain
has a length/thickness ratio equal to 877.
A p-refinement study was performed on this mesh in order to evaluate the
convergence of the solution. The simulations make it possible to compare the
results with those presented by several groups at the workshop, which are reported
in [68]. In Figure 4.7 the wall friction coefficient distribution is shown for several
orders of accuracy. Also, Blasius’ incompressible solution is reported. The effect of
p-refinement can be easily seen in the region near the leading edge. We performed
a second set of simulations with a lower free-stream Mach number (M∞ = 0.2)
in order to understand the influence of compressibility. In Table 4.7 the drag
coefficient (CD ) is reported as a function of the total number of degrees of freedom
per equation (nDOF s/eq). Even when the computations were performed using
one of the meshes with less elements between those proposed at the Workshop,
the p-refinement quickly obtained a converged solution. The ESR method behaves
well also in the presence of highly clustered meshes with very stretched elements.

10-1
Blasius
DG1
DG2
DG3
DG4
DG5
-2
10
cf

-3
10

-5 -4 -3 -2 -1 0
10 10 10 10 10 10
x

Figure 4.7. Wall friction coefficient (cf ) for the flat plate (Re∞ = 106 , M∞ = 0.2)

64
 4.5 – Considerations on implicit time integration

Scheme nDOFs/eq CD (M∞ = 0.5) CD (M∞ = 0.2)

DG1-RK2TVD 1680 1.277 1.282
DG2-RK3TVD 3360 1.306 1.324
DG3-RK4LS 5600 1.308 1.327
DG4-RK4LS 8400 1.307 1.327
DG5-RK4LS 11760 1.307 1.327
Blasius - 1.328 1.328

Table 4.7. Drag coefficient (CD ) for flate plate problem

(Re∞ = 106 , M∞ = 0.5 − 0.2).

4.5 Considerations on implicit time integration

In the previous paragraph, explicit time integration schemes have been considered
in order to evaluate the stability limits of the different methods. This aspect is
particular important in time dependent problems. However, implicit schemes can
be much more efficient than explicit ones when steady problems are studied. In
this Section we describe how to implement the ESR method with implicit time
integration. Implicit schemes require the evaluation of the jacobian of the system.
If an approach based on the analytical computation of the jacobian is used, the
ESR method shows some interesting features which can be useful during the im-
plementation. A linear heat diffusion problem with a fixed mesh is considered as
example.
In order to evaluate each term of the jacobian it is necessary to perform the deriva-
tive of the residuals with respect to each degree of freedom. This derivative con-
tains contributions from the volume and the boundary integrals which appear in
Eq. (3.9). In the proposed approach, the gradients required by the volume integral
are approximated using the internal reconstruction. This simplifies significantly
the computation of the jacobian. In contrast, both the RDG1x method and the
BR2 method compute these gradients using information also from the interfaces.
As far as the boundary integral is concerned, there are two possible cases. Indeed,
an interface can be inside the domain or it can belong to the domain boundary.
In the first case, the required gradients are computed with the recovery approach.
This is done through matrix-vector multiplications in which all the matrices depend
only on the mesh and not on the solution:

ũA
 
∇u = [D]ũ = [D][R̃]
r
. (4.28)
ũB

The terms in the jacobian related to this contribution can be immediately computed
from the constant terms in the previous relation. On the other hand, if the interface

65
 4 – Diffusive fluxes

belongs to a Dirichlet boundary, the procedure of Section 5 has to be considered.

In particular the gradient on the boundary is computed as:

∇u = [D] (ũ + ∆ũ) = [D]ũ + [D][B̂]Ŵ. (4.29)

Again, the contribution to the jacobian can be directly obtained from the constant
terms which appear in the previous relation.
A performance test was carried out on the 2D heat diffusion problem of Section
4.4.3. In particular an unstructured mesh with 512 triangles is considered. All
the elements are fifth order accurate (p = 4). The solution is initialized with a
uniform field and time integration is performed with both the RK4-LS explicit
scheme and the backward Euler (BE) implicit scheme. In Figure 4.8 the L2 norm
of the residuals is reported as a function of the computational time. As expected,
the plot shows the significant advantage offered by the implicit approach for steady
problems.

2
10
0
10
10-2
-4
10 RK4LS σ=0.035
Residual

-6 BE σ=10
10 BE σ=100
BE σ=1000
-8
10
-10
10
-12
10
-14
10

0 50 100 150
CPU time [s]

Figure 4.8. Adimensionalized L2 -norm of the residual vs computational time [s].

66
Chapter 5

Time discretization

Several time integration methods have been tested during the research activity.
During the first year both explicit multistep and explicit Runge-Kutta algorithms
have been implemented. Multistep methods [70] require the computation of fluxes
only once for each time step but they need to keep in the memory the numerical
solution at several previous steps. Furthermore, their stability limit becomes lower
when the order of accuracy is increased. On the other hand, Runge-Kutta methods
introduce several stages inside each step but they show a larger stability domain
when the order is increased.
Some experiments have also been performed with the ADER scheme [71]: it is
based on a particular procedure which gives a time dependent solution for the
fluxes at the interfaces between elements. The time evolution of the solution is
then obtained by a proper quadrature of the time dependent fluxes.
In order to increase the performances of the code in steady problems the implicit
backward Euler scheme has been implemented.
It is interesting to note that there are current trends in the literature which suggest
the use of implicit time integration schemes also for unsteady problem. A method
which seems to be very promising has been proposed by Nigro et al. [72].
In the following, only the implemented explicit Runge-Kutta methods and the
backward Euler scheme are described because most of this work has been performed
with these methods.
In Chapter 3 the discontinuous Galerkin semidiscrete formulation (Eq. 3.9) has
been derived for a generic element. It describes the evolution in time of the degrees
of freedom of the element. The global set of equations for all the elements in the
domain can be expressed as:

dũ
[M] = −R(ũ) (5.1)
dt
67
 5 – Time discretization

where [M] is the block diagonal global mass matrix and ũ is the global vector of
degrees of freedom. The vector R(ũ) is the global vector of the residuals obtained by
the boundary and volume integrals of Eq. 3.9. In order to simplify the notation, the
different methods will be presented by considering a reference problem described
by an equivalent system of ordinary differential equations:

du
= L(u) (5.2)
dt
In all the simulations a global time stepping strategy is employed. In the
general case in which convection-diffusion equations are considered the diffusion
and convection stability limits are evaluated separately in each element:

He2
∆td = CF L (5.3)
κ(2pe + 1)

He
∆tc = CF L (5.4)
c(2pe + 1)
where CF L, c, κ, He and pe are the stability coefficient, maximum signal speed,
diffusivity coefficient and element characteristic size and order. Then, the minimum
allowable time step is chosen in all the domain and it is used as global time step.
Cockburn [73] reports that a CFL number equal to 1 can be used in Eq. 5.4
for convection problems when a DG reconstruction of order p is integrated by an
explicit Runge-Kutta algorithm with an accuracy order equal to p + 1.

5.1 Explicit Runge-Kutta methods

The explicit Euler method is used for first order accurate time integrations:
Forward Euler
un+1 = un + ∆tL(un ) (5.5)
Two and three stages Total Variation Diminishing Runge-Kutta (TVD RK)
methods proposed by Gottlieb and Shu [74] have been implemented in order to
achieve second and third-order accuracy. These are schemes particularly suited for
problems containing discontinuities since they avoid the introduction of spurious
oscillations connected to the time integration.

Second-order two stages TVD-RK

u(1) = un + ∆tL(un )
1 n 1 (1) 1
un+1 = u + u + ∆tL(u(1) ) (5.6)
2 2 2
68
 5.2 – Implicit methods

Third-order three stages TVD-RK

u(1) = un + ∆tL(un )
3 n 1 (1) 1
u(2) = u + u + ∆tL(u(1) )
4 4 4
1 n 2 (2) 2
un+1 = u + u + ∆tL(u(2) ) (5.7)
3 3 3
Fourth-order accuracy has been obtained with three different Runge-Kutta al-
gorithms: the classical four stages method [75], the Strong Stability Preserving
(SSP) five stages method [76] and the low memory five stages method [77].
Most of the simulations have been performed with the SSP RK4 which is reported
in the following. Notice that the presence of the fifth stage requires a greater com-
putational effort in comparison with the classical fourth-order/four stages scheme,
but this is paid by a larger allowable time step.

Fourth-order five stages SSP-RK

u(1) = un + 0.39175222700392∆tL(un )
u (2)
= 0.44437049406734un + 0.55562950593266u(1) + 0.36841059262959∆tL(u(1) )
u(3) = 0.62010185138540un + 0.37989814861460u(2) + 0.25189177424738∆tL(u(2) )
u(4) = 0.17807995410773un + 0.82192004589227u(3) + 0.54497475021237∆tL(u(3) )
un+1 = 0.00683325884039un + 0.51723167208978u(2) + 0.12759831133288u(3) +
+ 0.34833675773694u(4) + 0.08460416338212∆tL(u(3) ) +
+ 0.22600748319395∆tL(u(4) ) (5.8)

5.2 Implicit methods

In this work the implicit Backward Euler method has been implemented. It can
be obtained by linearizing Eq. 5.1:

ũn+1 − ũn
[M] = − R(ũn ) + [J](ũn+1 − ũn ) (5.9)


∆t
where the matrices [J] and [M] are the jacobian matrix and the global mass matrix.
The global mass matrix is a block diagonal matrix in which each block is obtained
by an element mass matrix. The jacobian matrix is a block sparse matrix which
contains the derivatives of the residual with respect to the degrees of freedom:

∂Ri
Jij = (5.10)
∂ ũj

69
 5 – Time discretization

In PEq. 5.9 ũ represents the global vector of degrees of freedom which has size
nele
M i=1 Ni , where M , nele and Ni are the number of governing equations, the
number of elements and the number of degrees of freedom in the element i, re-
spectively. The equation can be rewritten in order to put in evidence the following
linear system:
1
 
[M] + [J] (ũn+1 − ũn ) = −R(ũn ) (5.11)
∆t

[A](ũn+1 − ũn ) = −R(ũn ) (5.12)

which has to be solved at each time step in order to update the vector of degrees of
freedom. The matrix A resulting from Eq. 5.12 is memorized in the Compressed
Sparse Row (CSR) format [78].

5.2.1 Jacobian evaluation

The terms in the jacobian matrix represents the dependency of the numerical
residuals on the degrees of freedom of the discretization. A possible way to obtain
the jacobian matrix is to perform the analytical differentiation of the discretized
equations. This approach is very efficient and can be easily parallelized. It has been
used in this work for the implicit integration of the unsteady linear heat equation.
However, this approach requires a lot of time for the implementation and the
maintenance of the code. Indeed, if a new physical model or a new numerical flux
is introduced, then the code for the computation of the jacobian has to be rewritten.
In order to simplify this task it is possible to use automatic differentiation tools
which automatically generates the source code for the terms which contribute to
the jacobian.
An alternative approach is based on the numerical approximation of the jacobian
by means of finite differences. This is done by the application of a perturbation 
on each of the degrees of freedom of the numerical solution. The residuals are then
evaluated before and after the application of the perturbation, simply by calling
the same subroutines. This approach is very interesting from the point of view of
code maintenance because it does not require any change in the subroutine which
evaluate the jacobian when a new physical model is introduced. However, there are
several drawbacks. First of all, great care must be taken in parallelization because
it is necessary to avoid the perturbation of two neighbouring elements (and also
the neighbours of the neighbours) at the same time.
Furthermore, the perturbation should be chosen sufficiently small to give a good
approximation of the derivative but not so small to introduce round-off problems.
In this work, the perturbation  for the i − th degree of freedom is chosen according
to the following rule:
 = max(min , rel ui ) (5.13)

70
 5.2 – Implicit methods

where the coefficients min and rel are usually chosen equal to 10−10 and 10−8 . A
similar approach has been proposed by Brown and Saad [79].

5.2.2 Iterative solver

The linear system 5.12 has to be solved at each time step. For this purpose the it-
erative solver GMRES [80] has been employed. In particular, the PARALUTION
[81] library has been used. This library offers several iterative solvers and pre-
conditioners. Furthermore, the library is parallelized by OpenMp directives for
computations on multi-core machines.
In all the simulations the Incomplete Lower Upper preconditioner with zero fill-in
(ILU0) has been employed to improve the convergence of the GMRES algorithm.
Some tests have been performed also with the restarted GMRES algorithm.

5.2.3 CFL evolution laws for steady problems

It is interesting to note that the backward Euler method applied to the unsteady
equations 5.1 tends to the Newton method applied to the steady equations when
∆t → +∞. This means that, for sufficiently high values of the time step, it is
possible to obtain quadratic convergence towards the steady solution. The presence
of the unsteady term in the equations plays the same role of the damping term in
damped Newton methods applied on the steady equations.
The time step cannot be chosen too high at the beginning of the simulation because
unphysical values can be obtained when the transient is too strong. For this reason,
it is necessary to introduce a CFL evolution strategy which represents a compromise
between the robustness and the efficiency of the integration. This approach is called
pseudo-transient continuation strategy.
Several approaches have been implemented in this work.
One of them is a robust method proposed by Bassi et al. [82] in the context of
turbulent problems. It uses the L2 and the L∞ norms of the residual to control
the CFL evolution. Let xL2 and xL∞ be defined as

|Ri |L2 |Ri |L∞

xL2 = max xL∞ = max 1≤i≤M (5.14)
|Ri0 |L2 |Ri0 |L∞

where Ri0 and Ri are the residuals of the i-th equation at the first time step and
at the current iteration. M is the number of governing equations which depends
on the chosen physical model and the dimensionality of the problem.

if x ≤ 1
( CF L
xα ,
0

CF L = CF L
α CF L −CF0 Lexp (1−x) (5.15)
CF Lexp + (CF L0 − CF Lexp )e 0 , if x > 1

71
 5 – Time discretization

where CF L0 , CF Lexp and α are the minimum CFL number, the maximum allow-
able CFL number of explicit schemes and a parameter (usually ≤ 1) which controls
the growing rate, respectively. The chosen CF L number is then limited under a
maximum value (usually CF Lmax = 1010 − 1020 ). The variable x is defined as:
(
min(xL2 ,1) if xL∞ ≤ 1
x= (5.16)
xL ∞ if xL∞ > 1

An alternative approach is the Switched Evolution Relaxation proposed by

Mulder and van Leer [83]. In this approach the CF L number at the step n is
obtained from the CF L number at the previous step according to the following
relation:
|Rn−1 |L2
CF Ln = min(CF Ln−1 , CF Lmax ) (5.17)
|Rn |L2
Finally, the Residual Difference Method by Bucker et al. [84] has been imple-
mented. The algorithm reads as follows:
|Rn−1 | 2 −|Rn | 2
L L
CF L = min(CF L
n n−1
·β |Rn−1 | 2
L , CF Lmax ) (5.18)

where β is a user-defined coefficient (β > 1) which controls the growing rate. Also
a monotonic variant of this algorithm has been considered by setting the exponent
of Eq. 5.18 to zero when |Rn |L2 > |Rn−1 |L2 .

In conclusion, all the previously described strategies have to be carefully cal-

ibrated. Indeed, the total computational time for a steady problem is strongly
influenced by the CF L evolution law. It is easy to change the computational time
by several orders of magnitude simply by choosing a strategy instead of another or
changing one of the parameters which control the growing rate.

72
Chapter 6

Shock capturing

The treatment of discontinuities which can appear in compressible flows requires

great care. Where the solution is discontinuous numerical methods of order greater
than one tend to show spurious oscillations. Several techniques have been proposed
to deal with these problems in the DG framework. A review of them can be found
in the synthesis report of the ADIGMA european project [85].
A widely used approach consists in adding artificial viscosity to the governing equa-
tions. This artificial term should be activated only in the regions in which there
are discontinuities ([86], [82], [89], [90]). The technique provides good results but
if used with explicit time integration schemes can introduce a significant limitation
on the maximum allowable time step and it also produces some oscillations when
used with high-order reconstructions. Furthermore great care must be taken for
the definition of the algorithm which controls the amount of numerical viscosity,
in order to keep it as problem independent as possible. On the other hand, the
artificial viscosity approach works very well with implicit time integration in steady
problems because the artificial term is directly introduced in the jacobian and so
convergence to machine precision can be reached.
A different approach has been derived from shock capturing techniques developed
in the context of FV methods; it consists in limiting the solution in a element
by the use of functions which compare the average value inside the element with
the average values of the neighbouring elements [91]. This approach suffers from
accuracy degradation on local extremes. The problem can be avoided by the in-
troduction of Total Variation Bounded in the Mean limiters [10] which require
the definition of a threshold that is problem dependent. Momentum limiters [92]
have also been proposed: they generalize the classic limiters approach extending
it to high-order reconstructions. Finally the WENO scheme [93] developed for
FV methods has been applied also to DG methods. All the techniques based on

73
 6 – Shock capturing

limiters share the need to know information from neighbouring cells and so they re-
quire special treatment for boundary elements or hanging-nodes meshes. A further
way of dealing with discontinuities is to filter the numerical solution by reducing
high-order modal coefficients, which are responsible for oscillations. This operation
shows results similar to those obtained by the artificial viscosity approach but does
not introduce any additional stability restriction for explicit integration. Further-
more, the filter acts only on the element solution and does not require information
from neighbouring elements. It is possible to apply filtering directly to the solu-
tion modal coefficients ([94]), but also to the time derivatives of these coefficients
[95]. In any case the main problem is to understand how much filtering is needed:
enough to avoid spurious oscillations but not so much as to degrade accuracy. In
order to give a more precise answer to this question, a feedback filtering procedure
has been developed during the research activity. It is described in Section 6.1. The
aim of the proposed approach is to maintain the locality of the algorithm and to
make it as problem independent as possible.
Both limiters and filters work well in time-accurate calculations but they can in-
troduce convergence problems in steady flows.

6.1 Feedback filtering

6.1.1 Exponential filter
In the present work an exponential filter has been used to keep under control
oscillations induce by shock waves. The filter multiplies the solution’s modal co-
efficients according to the approach described by Gottlieb and Hesthaven [96]. In
particular, the filter has been implemented in the explicit version of the code. It
acts as a post-processing step after each stage of the Runge-Kutta time integration
algorithm. In 1D the filtered solution coefficients are obtained in the following way:

(ũn )f iltered = σ(ηn )ũn 0≤n≤p (6.1)

r n
σ = e(−αη )
ηn = (6.2)
p
where n is the degree of the generic basis function and p is the order of the re-
construction inside the element. The extension to 2D is performed in the following
way:

(ũnm )f iltered = σ(ηn )σ(ηm )ũnm 0≤n≤p 0≤m≤p (6.3)

n m
ηn = ηm = (6.4)
p p

74
 6.1 – Feedback filtering

where n and m are the degrees of the basis function in the two coordinate variables.
The parameters r (filter order) and α (filter intensity) have to be chosen carefully
in order to eliminate oscillations while preserving the accuracy of the numerical
solution in smooth regions. In the literature there are several works (see e.g. Hes-
thaven and Warburton [97]) which suggest to set α = −log(eps) where eps is the
machine precision (for double precision α ' 36). Other indications are given by
Boyd [98] who suggests the use of high-order filters in smooth regions and low-order
filters in the proximity of discontinuities.
The numerical experiments performed in this work show that the choice of α =
−log(eps) degrades significantly the solution accuracy, even when high-order fil-
ters are used. These results are in line with the work previously carried out by
Ueckermann and Lermusiaux [99]. They introduced a function which measures the
smoothness of the numerical solution inside each cell: the signal obtained by this
function is compared to a threshold and, if the signal is greater than the threshold,
an exponential filter is applied in the considered cell. In their computations they
set α = −log(0.01) ' 4.6.
In the simulations performed in this work, this kind of approach appears to be not
very robust since the choice of α and the threshold has to be done very carefully.
Indeed, the optimal values for these parameters depend on the particular problem
which is under study and depend also on the reconstruction order. For this reasons
an adaptive procedure has been introduced in order to automatically adjust the
filter intensity.

6.1.2 Smoothness indicator

One of the fundamental components of the proposed procedure is a smoothness
indicator. In particular, the sensor developed by Persson and Peraire [86] for their
shock capturing scheme is used in this work.
This sensor compares the numerical solution inside an element to a truncated
numerical solution, which does not contain higher-order terms. In practice, it
can be shown that where the solution is smooth the modal coefficients decay very
quickly. For this reason the sensor shows weak signals in smooth regions. In
contrast, the high-order coefficients earn importance near discontinuities and so
the sensor gives intense signals where the solution is not smooth. In particular
the following expression is obtained when the sensor is applied to a generic scalar
variable uh inside the element Ωe :

(u − ū)2 dxdy
R
SΩe = RΩe 2 (6.5)
Ωe
u dxdy + ε
In 1D problems the test variable ū is obtained by truncating the reconstruction at
the first N̄e = p modes and is compared with the complete u obtained by all the

75
 6 – Shock capturing

Ne = p + 1 modes.
This is generalized to 2D by setting N̄e = p2 , Ne = (p + 1)2 for BASIS A and
N̄e = p(p + 1)/2 and Ne = (p + 1)(p + 2)/2 for BASIS B.

N̄e
X Ne
X
ū = ũi Φi u= ũi Φi (6.6)
i=1 i=1

In this way the sensor gives a signal which quantifies the smoothness of the
numerical solution in each element. Numerical experiments on discontinuous prob-
lems show that the order of magnitude of the signal changes significantly when
moving from smooth regions to the elements near a discontinuity. For this reason
the logarithm of the smoothness indicator signal (SI) is monitored:

SI = log10 (SΩe + ε) (6.7)

In some cases the denominator of Eq. 6.5 and the logarithm’s argument in
Eq. 6.7 could become equal to zero, depending on which variable is monitored. In
particular the sensor signal can become equal to zero when the truncated solution
is able to exactly describe the physical solution. For example, this happen when
a uniform field is considered. In order to avoid this problem an additive positive
constant  is introduced in Eqs. 6.5 and 6.7 and set equal to 1−10 in all simulations.

6.1.3 Feedback filtering

The key idea of feedback filtering [1] consists in setting a goal for the solution
smoothness (SR = smoothness requirement) and in automatically finding the filter
intensity needed to achieve this goal. This is done through an iterative procedure
which measures the smoothness of the solution (SI). When SI is not satisfactory
a weak filter is applied (for example an exponential filter with αiter = 0.1). These
operations are repeated until the desired smoothness is reached. In other words, in
each element at each stage of the Runge-Kutta algorithm the procedure described
by the following pseudocode is applied:

Compute SI
do while (SI > SR)
Apply weak filter
Compute SI
end do

In this way the number of parameters which have to be calibrated is halved in com-
parison to the approach of Ueckermann and Lermusiaux: it is only necessary to set
a requirement for the solution smoothness and the filter intensity will be adjusted
automatically in order to satisfy it. This is the concept of feedback filtering: the

76
 6.1 – Feedback filtering

filter influences the smoothness of numerical solution but this in turn influences
the filter intensity.
It is useful to say few words about idempotent filters (Kanevsky et al. [100]). If
a filter is applied repeatedly to a solution providing always the same results, inde-
pendently on how many times it is applied, the filter is referred to as idempotent.
In other words an idempotent filter does not show cumulative effects when it is
applied repeatedly. The exponential filter is not idempotent but if it is used inside
the proposed feedback procedure, an idempotent filtering operation is obtained.
Indeed when the procedure is applied for the first time the solution is filtered in
order to satisfy the smoothness requirement: any subsequent application of the
procedure does not change the solution, since it already satisfies the requirement.

The proposed filtering procedure can be implemented in several ways. As far

as the choice of αiter is concerned, it can be observed that an increase in αiter
reduces the number of iterations required to reach the goal and so it reduces the
computational cost. However, αiter should not be chosen too large otherwise the
solution obtained after the last iteration could be over-filtered. In an over-filtered
solution the high-order modes have been strongly reduced and only the first modal
coefficient remains unaltered. Furthermore, high-order DG schemes are usually
employed with relatively coarse meshes since they introduce several DOFs inside
each element and so they require few elements. For this reasons, an over-filtered
solution can appear quite inaccurate as it is close to a first order solution on a
relatively coarse mesh (see Par. 5.1). In all the computations performed in this
work the following parameters were used: αiter = 0.1 and r = 2.
As far as computational cost is concerned, it is important to keep in mind that the
iterative filtering procedure is applied only in those elements in which the smooth-
ness requirement is not satisfied, i.e. in the elements containing a discontinuity.
They usually represent a small fraction of the total number of elements.

Some guidelines about the choice of the smoothness threshold SR can be given.
In their work Persson and Peraire [86] observe that if the solution is based on a
Fourier expansion the sensor SΩe provides a signal which depends on the recon-
struction order p according to the following relation:

g
SΩe = (6.8)
p4

where g is a constant. They assume that this relation holds also for polynomial re-
constructions. They use this result to adjust the artificial viscosity of their scheme.
In the present work the Eq. 6.8 is used to understand how the threshold SR should
depend on the reconstruction order. If the logarithm of Eq. 6.8 is considered the

77
 6 – Shock capturing

following relation is obtained:

SR = log10 (g) − 4 log10 (p) (6.9)

in which the constant g is the only parameter which the user has to adjust.

Some considerations can be done about the practical implementation of this

procedure. There are several variables which can be monitored by the smoothness
indicator in the framework of Euler equations. A computationally efficient choice
could be to monitor only one variable, which should be significant for both shocks
and contact surfaces: for example density could be a good candidate. When in
an element the sensor applied to density provides a signal greater than the thresh-
old then all the conservative variables are filtered. Alternatively each conservative
variable could be monitored applying the filter only to those variables which do not
satisfy the requirement. All the results shown in the following have been obtained
with the first approach, by monitoring density. If the proposed algorithm is used in
the context of the compressible Navier-Stokes equations the sensor variable should
be chosen in order that the filter does not act on boundary layers or wakes. In par-
ticular, the pressure could be a better sensor variable for Navier-Stokes equations
because it helps in preserving the accuracy of the boundary layer discretization.

Results obtained for four different test cases are presented in the following. The
first two are characterized by the presence of discontinuities, both in 1D (Section
6.1.4) and 2D (Section 6.1.5). In the third test case (Section 6.1.6) a smooth flow is
studied and the numerical solution is compared with the existing analytical solution
in order to test whether feedback filtering preserves solution accuracy in smooth
regions. In the fourth test case (Section 6.1.7) the transonic flow field around an
airfoil is studied and results are compared with reference values from the literature.
In order to verify the robustness of the proposed approach the same law is used
for the smoothness requirement SR in all test cases :

SR = −2 − 4 log10 (k) (6.10)

6.1.4 Sod problem

The first test case is the classic one-dimensional Sod shock tube problem. It con-
sists in a Riemann problem in which there are two initial uniform fields A (left) and
B (right) defined by the following conditions: uA = 0, PA = 0.4, ρA = 1, uB = 0,
PB = 0.04, ρB = 0.125. The exact solution is characterized by the propagation
of three waves: an expansion fan, a contact surface and a shock. Between the
waves there are two uniform fields which can be computed by means of an iterative
procedure in which the non-linear Rankine-Hugoniot jump relations are considered
together with the compatibility relations for the expansion wave.

78
 6.1 – Feedback filtering

After the computation of these constant fields it is possible to find the exact solu-
tion inside the expansion fan by the use of compatibility relations. Further details
can be found in Toro [87] and Shapiro [88].

Numerical results obtained by DG0, DG1, DG2 and DG3 schemes are com-
pared in the following. Time integration was performed by Runge-Kutta schemes
described in Section 5.1 in order to obtain the same accuracy order in space and
time. The number of DOFs per equation was kept approximately equal to 100
in all simulations in order to make a fair comparison between different schemes.
For this reason a uniform mesh with 100 elements was used for the DG0 scheme
while coarser meshes were used for higher-order schemes. The numerical solution
is represented by plotting in each element a number of points equal to the number
of DOFs in that element.
First of all, some computations were performed with the exponential filter (α = 36,
r = 2) activated in all the elements. In Figure 6.1 the density field at time t = 0.4
is reported. The results show clearly the effects of over-filtering. The accuracy
of high-order reconstructions is completely lost. Indeed, the DG0 solution on the
finest mesh is more accurate than all the others high-order solutions on the coarser
meshes.
A second set of simulations was performed with the proposed feedback filtering
procedure. Results are reported in Figure 6.2. In this case the global solution
shows that high-order schemes achieve a significant improvement over the DG0
scheme. This is particularly evident in the contact surface and at the edges of the
expansion fan. A detail of the contact surface is presented in Figure 6.4. It can be
seen that results obtained bt DG1, DG2 and DG3 schemes are very close and show
a significant improvement with respect to the DG0 results. However, high-order
schemes produce a weak undershoot near the discontinuity. Figure 6.3 shows a
detail of the captured shock. The filtered high-order schemes introduce a slight
deterioration in the quality of the captured shock with respect to the DG0 results.
This behavior can be compensated by the introduction of an adaptive h-refinement
strategy in order to increase shock resolution.
In conclusion this first test shows how feedback filtered high-order DG methods
can capture discontinuities and keep spurious oscillations under control.

79
 6 – Shock capturing

DG0
0.8 DG1
DG2
DG3
Exact
Density ρ

0.6

0.4

0.2

0 0.2 0.4 0.6 0.8 1

X

Figure 6.1. Adimensionalized L2 -norm of the residual vs computational time [s].

Figure 6.2. Adimensionalized L2 -norm of the residual vs computational time [s].

80
 6.1 – Feedback filtering

Figure 6.3. Detail of the contact surface in the Sod problem (Feedback filtering).

Figure 6.4. Detail of the shock in the Sod problem (Feedback filtering).

81
 6 – Shock capturing

6.1.5 Shock-bubble interaction

The second test case is a two-dimensional problem characterized by the interaction
between a normal shock and a cylindrical contact surface with radius equal to 0.2.
The initial condition (Figure 6.5) shows a field at rest (B) containing a bubble (C)
of hot gas. A normal shock wave coming from the left separates the field B from
the supersonic uniform field A and interacts with the bubble. As a consequence the
bubble’s edge is deformed. After some time it breaks and gives rise to the formation
of vortices. This test case makes it possible to carry out a qualitative analysis of
numerical schemes. Indeed the solution shows vortex structures which come from
the destabilization of the initial contact surface: the greater the numerical dissipa-
tion introduced by the numerical scheme, the less it will be able to capture small
scale structures. As far as the boundary conditions are concerned, the left side is a
supersonic inlet, the right side is a subsonic outlet while the upper and lower sides
are solid walls. Since the problem is symmetric the computational domain was
restricted to the upper part of the physical domain and solid wall boundary condi-
tions were imposed on the symmetry axis. Initial data are summarized in Table 6.1.

Figure 6.5. Domain and initial condition for shock-bubble interaction

Region ρ P u v
A 3.81 × 100 1.00 × 101 2.58 × 100 0.00 × 100
B 1.00 × 100 1.00 × 100 0.00 × 100 0.00 × 100
C 1.00 × 10−1 1.00 × 100 0.00 × 100 0.00 × 100

Table 6.1. Shock-bubble problem: initial data.

The numerical solution is represented by computing the density gradient mag-

nitude and plotting it with a nonlinear gray scale, in order to obtain a numerical

82
 6.1 – Feedback filtering

Schlieren image. Results can be compared with those reported by Cada and Tor-
rilhon [9] and obtained by their third-order finite volume scheme. They use a very
fine cartesian mesh with 1700x500 cells. The solution quality was evaluated by
performing a simulation with the feedback filtered DG2 scheme on a 567x167 ele-
ments grid, in order to maintain approximately the same number of DOFs of the
simulation carried out by Cada and Torrilhon. Results obtained by feedback fil-
tering show the expected behavior: the evolution of the contact surface can clearly
be seen in the sequence reported in Figure 6.6. A comparison between the struc-
tures visible in Figure 6.7 and those obtained by Cada and Torrilhon [9] shows
that the discontinuities captured by the feedback filtered DG scheme have a less
sharp profile than those captured by its finite volume counterpart. This is in line
with the results obtained in the Sod test case, in which the filtered DG2 scheme
captures the shock with a slight less sharp profile than the DG0 scheme. However,
the feedback filtered DG method shows its accuracy in the description of contact
surfaces and smooth regions. This feature permitted a good representation of the
main structures of this flow. In Figure 6.6 it can be seen that the normal shock
wave is followed by two numerical starting waves which appear at the beginning
of the simulation. This numerical phenomenon is common also in finite volume
computations and its origin is explained by Leveque [44].

Figure 6.6. Numerical Schlieren images for shock-bubble interaction at t = 0.071,

t = 0.21 and t = 0.33 (Feedback filtered DG2 scheme).

83
 6 – Shock capturing

Figure 6.7. Numerical Schlieren image for shock-bubble interaction at t = 0.18

(Feedback filtered DG2 scheme).

84
 6.1 – Feedback filtering

6.1.6 Ringleb flow

The Ringleb flow is an inviscid steady isentropic two-dimensional flow. It can be
seen as a shock-less transonic flow around a blunt body. This test case makes it
possible to evaluate the accuracy order of a scheme by a comparison between the
numerical solution and the exact solution, which can be obtained by the hodo-
graph method [101]. In particular, the flow field is characterized by the following
streamlines:
1 1 2
 
J
x= − + (6.11)
2ρ̄ q̄ 2 κ2 2
s  2
1 q̄
y=± 1− (6.12)
κρ̄q̄ κ
where
1
κ= (6.13)
Ψ

1 1 1 1 1 + ā
J= + 3 + 5 − ln (6.14)
ā 3ā 5ā 2 1 − ā

γ−1 2
r
ā = 1− q̄ (6.15)
2
2
ρ̄ = ā γ−1 (6.16)
In the previous relations q̄, ā and ρ̄ are non-dimensional velocity magnitude, speed
of sound and density. Here they are referred to the stagnation values of speed
of sound and density respectively. Notice that in all the rest of this work, the
normalization described in the Appendix A is used.

Figure 6.8. Ringleb flow: constant speed lines (dashed) and computational domain Ω.

85
 6 – Shock capturing

The computational domain Ω is inside the subsonic region and it is shown in

Figure 6.8. It is delimited by two streamlines with Ψ = 1.15 and Ψ = 2, where Ψ is
the streamline constant which appears in Eq. 6.13. The corresponding boundaries
are treated as solid walls. The inlet is chosen on the flow symmetry axis. The
outlet corresponds to the iso-Mach line for which q̄ = 0.35. Curvilinear elements
are used near solid walls and also near the outlet, on which pressure is imposed.
The numerical solution is initialized with a uniform field at rest and then it is
integrated up to steady state. Steady-state entropy (S) and pressure (P) errors in
both norm-2 (L2 ) and norm infinity (L∞ ) are reported in Table 6.2 and Table 6.3 for
the feedback filtered computations. Another set of computations without filtering
gave exactly the same results. This suggests that during the transient the filter
is applied in those elements in which discontinuities appear but when the smooth
steady state field is reached the filter does not have any effect on the numerical
solution and so accuracy is preserved.

Scheme Grid L2 S error Order L2 S L∞ S error Order L∞ S

DG1B2 8x4 4.78E-04 - 1.69E-03 -
16x8 1.14E-04 2.07 4.94E-04 1.77
32x16 2.66E-05 2.10 1.33E-04 1.89
64x32 6.31E-06 2.08 3.48E-05 1.93
DG2B3 8x4 1.41E-04 - 9.00E-04 -
16x8 1.46E-05 3.27 1.52E-04 2.57
32x16 1.43E-06 3.35 2.24E-05 2.76
64x32 1.42E-07 3.23 3.05E-06 2.88
DG3B4 8x4 1.85E-05 - 1.55E-04 -
16x8 1.02E-06 4.18 1.31E-05 3.56
32x16 4.88E-08 4.39 9.58E-07 3.77
64x32 2.25E-09 4.44 6.49E-08 3.88

Table 6.2. Ringleb flow: entropy accuracy test for feedback filtered DG methods.

6.1.7 Transonic NACA0012

This inviscid test case is characterized by a transonic flow field with shock waves
on both the pressure and suction side of the airfoil. The freestream Mach number
is M∞ = 0.8 and the angle of attach is α = 1.25o .
Six simulations were performed with the reconstruction order varying from p = 0
to p = 5 and feedback filtering activated. Time integration was performed by EE
(for p = 0), RK2 (for p = 1), RK3 (for p = 2) and SSPRK (for p >= 3) algorithms
respectively. Computations were carried out on a structured C-mesh and domain

86
 6.1 – Feedback filtering

Scheme Grid L2 p error Order L2 p L∞ p error Order L∞ p

DG1B2 8x4 1.44E-03 - 5.22E-03 -
16x8 3.02E-04 2.25 1.12E-03 2.22
32x16 9.87E-05 1.61 2.86E-04 1.97
64x32 2.58E-05 1.94 8.75E-05 1.71
DG2B3 8x4 1.51E-04 - 3.84E-04 -
16x8 1.45E-05 3.38 4.16E-05 3.21
32x16 1.63E-06 3.15 5.26E-06 2.98
64x32 2.03E-07 3.01 7.77E-07 2.76
DG3B4 8x4 1.67E-05 - 8.98E-05 -
16x8 9.76E-07 4.10 8.23E-06 3.45
32x16 5.81E-08 4.07 7.10E-07 3.54
64x32 3.51E-09 4.05 5.19E-08 3.77

Table 6.3. Ringleb flow: pressure accuracy test for feedback filtered DG methods.

boundaries were fixed at 20 chords from the airfoil. This distance is too small
to obtain results which are independent from the boundary distance. However,
the main purpose of this test is to study the shock capturing properties of the
proposed approach. A more accurate study (with freestream distance greater than
2000 chords) can be found in Sections 7.3.1.
The chosen mesh has 50 quartic curvilinear elements on both the pressure side and
the suction side of the airfoil. In total the mesh contains 200x50 elements.
In Figure 6.9 and Figure 6.10 the Mach field for DG1 and DG5 simulations is
reported. The DG1 results on this very coarse mesh show a significant generation
of spurious entropy at wall which disappears in the DG5 results. On the other
hand, the DG5 results show some oscillations behind the stronger shock. These
oscillations are related to jumps in entropy production where the shock moves from
an element to another. This behavior is probably related to the fact that the filter
removes the higher-order modes in the elements crossed by the shock and so the
solution in this region is nearly piecewise constant. It is interesting to observe that
in the test case of Section 6.1.5 there are strong curvilinear shock waves but there
are not significant oscillations in the region behind the shocks. However, in that
case a very fine grid was employed while in this problem the mesh is relatively
coarse. These considerations suggest that the use of local grid refinement in the
shock region could help in reducing this kind of post-shock oscillations.
In Figure 6.11 and Figure 6.12 the wall pressure coefficient (Cp) distribution for
the DG1 and DG5 computations is reported. For the purpose of comparison, DG5
results obtained by Barter [102] with the artificial viscosity approach of Persson
and Peraire [86] are also plotted. It can be seen that the shock on the pressure

87
 6 – Shock capturing

side appears very similar with the two approaches, since in both cases there is an
undershoot. On the other hand, the shock on the suction side is captured more
sharply by the artificial viscosity approach. The magnitude of the post-shocks
oscillations are comparable to those obtained by other common shock capturing
DG schemes reported in the ADIGMA report [85] for this test case. The lift and
drag coefficients reported in Table 4 converge when the order is increased. They
are in good agreement with the reference values obtained by Vassberg and Jameson
[103] with a second order finite volume method on a very fine grid (4096x4096).

Figure 6.9. Mach field on the NACA0012 airfoil (M∞ = 0.8,α∞ = 1.25o ):
feedback filtered DG1 scheme.

Figure 6.10. Mach field on the NACA0012 airfoil (M∞ = 0.8,α∞ = 1.25o ):
feedback filtered DG5 scheme.

88
 6.1 – Feedback filtering

Figure 6.11. Numerical Schlieren image for shock-bubble interaction at t = 0.18

(Feedback filtered DG2 scheme).

Figure 6.12. Numerical Schlieren image for shock-bubble interaction at t = 0.18

(Feedback filtered DG2 scheme).

89
 6 – Shock capturing

6.2 Artificial viscosity

In this work the artificial viscosity method proposed by Nguyen and Peraire [89]
has been implemented. It can be activated in transonic or supersonic problems
characterized by the presence of shock waves. The method is based on a laplacian
term which is added to the governing equations:
∂u
+ ∇ · (F − νa ∇u) = 0 (6.17)
∂t
where u and F are the vector of the conservative variables and the vector of the
physical fluxes, respectively. The artificial viscosity νa can be computed as:
 
l∇ · q
νa = 0 f (6.18)
a
where 0 , h, q and a are a positive user-defined constant, a characteristic length,
the velocity vector and the speed of sound, respectively. The characteristic lenght
l depends on the element characteristic size h and on the distance from the closest
wall d. It can be obtained by the following relation, which guarantees zero artificial
viscosity at wall:
l = min(h,10d) (6.19)
In the present work, when Eq. 6.19 is evaluated at an internal interface, an average
element characteristic size is computed between the two elements.
The function f is:  
f (x) = α1 log 1 + e(α2 −x)/α1 (6.20)
where the constants are α1 = 0.1 and α2 = −0.5. It can be seen that the previous
relation gives a strong artificial viscosity when the divergence of velocity becomes
negative.

The described shock capturing approach requires the definition of a parameter

0 which controls the maximum amount of artificial viscosity that can be introduced
in the discretization. This parameter is problem dependent. Some preliminary
experiments performed during this work have shown that a bad choice of its value
can significantly alter the results of the simulation, especially in the presence of
boundary layers.
A possible value for inviscid transonic computations (see Chapter 7.3.1) can be
0 = 0.004.

6.2.1 Future works

Some preliminary experiments were done with the implemented artificial viscosity
method in viscous flow problems. In particular, the tests performed on the laminar

90
 6.2 – Artificial viscosity

flow over a flat plate showed that the implemented artificial viscosity can signif-
icantly alter the velocity profile in the boundary layer if the parameter 0 is set
to the value used for the transonic NACA0012 problem (see Chapter 7.3.1). This
means that this artificial viscosity approach is strongly problem dependent.
An alternative approach could be the shock capturing term proposed by Bassi et al.
[82]. This method introduces an artificial viscosity which depends on the inviscid
residuals of the discretization. This approach makes the scheme consistent to the
Euler equations but only asymptotically consistent to Navier-Stokes equations or
RANS equations.
Hartmann [90] proposed an artificial viscosity approach which is based on the full
residuals and so it is consistent also for Navier-Stokes and RANS equations. The
introduction of these shock capturing approaches could bring several benefits.

91
92
Chapter 7

Adaptive algorithms

Fluid dynamics is characterized by problems in which the complexity of the solu-

tion can change significantly from zone to zone. Indeed, there are several structures
(shock waves, boundary layers, wakes, turbulent structures,...) which are localized
in small portions of the domain but which have great influence on the results. In
order to increase the efficiency of the discretization, it is necessary to adapt the
discretization to the physics of the problem. This can be done by concentrating
the degrees of freedom in the regions which require more attention.
This approach is very important in unsteady problems in which the most inter-
esting structures move across the domain. Furthermore, the possibility to auto-
matically adapt the discretization is useful in the framework of simulation-based
optimizations, in which several geometries have to be tested and it is not possible
to manually adapt each mesh.
There are two main families of adaptive algorithms: p-adaptive algorithms and
h-adaptive algorithms. The local nature of the discontinuous Galerkin reconstruc-
tion allows an easy implementation of both strategies.

P-adaptive algorithms exploit a local variation of the reconstruction order (p)

across the domain without any modification to the mesh. In the framework of DG
methods, this can be done without any change in the stencil of the reconstruc-
tion. Furthermore, if a hierarchical basis is employed, the implementation of a
p-adaptive algorithm becomes very simple. Indeed, when an element needs to be
refined, it is sufficient to add more degrees of freedom without the need to change
the ones already known. On the other hand, if an element needs to decrease the
reconstruction order, it is enough to eliminate the degrees of freedom related to
the higher order modes.
If the p-adaptation process is driven by a proper error indicator, then an exponen-
tial convergence of the error can be obtained in sufficiently smooth problems.

93
 7 – Adaptive algorithms

On the other hand, p-adaptive algorithms are not the best approach in the
presence of singular solutions. Indeed, the increase in the reconstruction order
gives small benefits on shock waves or near singular geometries. In these cases the
numerical convergence order is far from the theoretical one. For this reason, it is
more convenient to keep the reconstruction order constant and to reduce the size
of the elements. However, the error convergence rate obtained by this approach is
not exponential but only algebraic.
Finally, it is possible to use mixed approaches in which both p-adaptive and h-
adaptive strategies are employed. Several works are concerned with these algo-
rithms in the framework of discontinuous Galerkin methods. An overview of some
techniques is presented in the book which summarizes the results of the ADIGMA
European project [104].

7.1 Sensors and error indicators

The adaptation process requires a suitable sensor to drive the algorithm. The sen-
sor should be able to detect the main sources of errors and its evaluation should
be as cheap as possible. In particular, the sensors which use quantities already
computed by the discretization scheme are more suitable for unsteady problems in
which the adaption has to be performed at each time step. In this work, several
approaches have been tested.

Some "feature-based" sensors have been implemented as first step. They control
some specific features of the solution which are indicative of the most interesting
regions in the domain. These sensors are very easy to implement but are problem
dependent. They have been used in this work mainly to test the behavior of the
ESR method when the reconstruction order varies from an element to its neigh-
bours. These algorithms are based on a sensor and a control law. The choice of the
sensor depends on the particular set of equations which is considered. For example
it can be related to the intensity of a variable or to its gradient.

A more rigorous approach can be obtained by the choice of an error indicator.

There are two main families of error indicators: adjoint-based and residual-based
error indicators. Adjoint-based error indicators are obtained by the definition of a
dual problem. They identify the sources of error for a given target function. The
aim of this approach is to minimize the error on a particular goal (for example
the lift or drag coefficients of a wing) for a given number of degrees of freedom
employed in the numerical solution. Several works based on this approach can be
found in the literature, like for example [105].

94
 7.2 – p-adaptive algorithm

Residual-based error indicators try to estimate the discretization error on the con-
servative variables and so they target at resolving all flow features, not only those
related to a particular target function. Furthermore, they are very cheap because
they make use of the residual of the governing equations which are already avail-
able. In this work, a residual-based error indicator has been implemented following
the approach of Leicht and Hartmann [106]. In particular, here the error indicator
ηe has been implemented for inviscid flows in the following way:
p
ηe = he |R(uh )|e + he |r∂e (uh )|∂e (7.1)

where he is the characteristic element size while R(uh ) and r∂e (uh ) are the volume
and surface residual defined as:
∂Fh ∂Gh
R(uh ) = − − (7.2)
∂x ∂y

r∂e (uh ) = nx Fh + ny Gh − (F̂h nx + Ĝh ny ) (7.3)

The notation of Eqs. 7.2 and 7.3 refers to Eq. 3.6.

7.2 p-adaptive algorithm

P-adaptive algorithms are very useful in smooth and regular problems. In this
work, they have been tested on the linear heat equation and the compressible
Navier-Stokes equations. The main purpose of these tests is to study the behavior
of the ESR method in problems with an inhomogeneous order distribution. For
this reason, a very simple p-adaptive strategy is used and a sensor based on some
flow features is chosen. A more rigorous approach can be obtained by the use of
an error indicator, similar to Eq. 7.1.
The implemented strategy makes use of a feature-based sensor and a control law.
The control law is responsible for the choice of the element order pe according to
the signal obtained locally by the sensor se . The following very simple control law
has been chosen:
"  α #
se − smin
pe = Int pmin + (pmax − pmin ) (7.4)
smax − smin p
where the function Int rounds its argument to the nearest integer number.
Eq. (7.4) shows that the reconstruction order in a generic element is allowed to
vary between a minimum (pmin ) and a maximum (pmax ). The local value of the
sensor (se ) is compared to the minimum (smin ) and maximum (smax ) values found
in the computational domain at the considered time step. The coefficient αp is a

95
 7 – Adaptive algorithms

positive number which controls the distribution of the order in the domain: the
lower the αp , the larger the number of elements in which high order reconstructions
are employed. The chosen law helps in avoiding strong jumps in the order between
neighbouring elements.

7.2.1 Heat equation and p-adaptivity

Several numerical experiments were performed on the 2D heat equation with the p-
adaptive algorithm and the ESR method [2]. The influence of the parameter αp was
also investigated. The considered test case is similar to the one used before for the
convergence analysis (Section 4.4.3), but now the number of semi-waves is increased
up to 5 (kw = 5) as shown in Figure 7.1. The exact solution is characterized by
regions in which there are large gradients and regions in which the temperature
distribution is close to constant. In a situation like this, the p-adaptive algorithm
can show its advantages with respect to a uniform order discretization. The same
unstructured meshes used in Section 4.4.3 were adopted for these experiments. The
magnitude of the temperature gradient was chosen as sensor variable.

T
y
x

Figure 7.1. Temperature field for the 2D heat equation problem, with kw = 5.

In the plot of Figure 7.3 the L2 -norm of the temperature error is reported as
a function of the total number of degrees of freedom. DG1 and DG5 denote data
obtained by a uniform order of accuracy, namely p = 1 and p = 5. The other data
are obtained by the p-adaptive strategy (with pmin = 1 and pmax = 5). The con-
trol law defined by Eq.7.4 is used for this test. It can be seen that the p-adaptive

96
 7.2 – p-adaptive algorithm

1
p
0.8 5
4
3
0.6 2
1
y

0.4

0.2

0 0.2 0.4 0.6 0.8 1 1.2

x

Figure 7.2. Example of order distribution with p-adaptivity and

the 8x8x2 mesh (αp = 0.3)

-1
10

-2
10

-3
10
Error

-4
10

DG1
-5 DG5
10
p-adapt. (α=0.1)
p-adapt. (α=0.2)
p-adapt. (α=0.3)
-6 p-adapt. (α=0.5)
10

2 3 4
10 10 10
nDOFs
Figure 7.3. L2 -norm of the temperature error vs total number of DOF’s
(possible orders: pe = 1,2,3,4,5).

97
 7 – Adaptive algorithms

-1
10

-2
10

-3
10
Error

-4
10

DG1
-5 DG5
10
p-adapt. (α=0.1)
p-adapt. (α=0.2)
p-adapt. (α=0.3)
-6 p-adapt. (α=0.5)
10

2 3 4
10 10 10
nDOFs
Figure 7.4. L2 -norm of the temperature error vs total number of DOF’s
(possible orders: pe = 1,3,5)

98
 7.2 – p-adaptive algorithm

algorithm is convenient with respect to the uniform order discretization if a suffi-

ciently low value of αp is chosen. The results suggest that the adaptive algorithm
is competitive for 0.1 ≤ αp ≤ 0.2.

Another test was performed in order to study the effects of the proposed dis-
cretization for the case |pA − pB | > 1, which is described in Section ??. In par-
ticular, the control law (Eq.7.4) was modified in such a way that the algorithm
can choose only the following reconstruction orders: pe = 1, pe = 3 or pe = 5. In
this way, when there is a jump in the order across an interface, the jump is always
greater or equal than 2. The results for this somewhat artificial case are reported in
Figure 7.4. A comparison with Figure 7.3 shows clearly that the presence of strong
jumps in the order distribution is not convenient. For this reason, it is important
to use a control law which chooses a smooth distribution of the order and which
discourages the occurrence of the case |pA − pB | > 1.

7.2.2 Vortex shedding around a circular cylinder

In this classical test case, the unsteady flow field around a circular cylinder with
Re∞ = 102 and M∞ = 0.2 is considered. Characteristic far field boundary condi-
tions are prescribed on a circular boundary. The distance of the far field from the
body is equal to 100 cylinder diameters. The solid wall is adiabatic. The domain
is discretized with a structured O-mesh with 19x48 quadrilateral elements. Wall
elements have quartic edges. Convective fluxes are computed by the Osher [35]
numerical flux while diffusive fluxed are computed by the ESR method [2].
Five simulations were performed with schemes ranging from DG1 to DG5 in
order to evaluate the behavior of the proposed discretization in the presence of
unsteady flow fields. Computations were carried out until periodicity was reached.
The Strouhal number (St) related to vortex shedding and the average drag coeffi-
cient were evaluated over a period of 10 cycles. After that, the p-adaptive algorithm
was tested and the order of the elements was allowed to vary between pmin = 1
and pmax = 5 at each time step.
First of all, some experiments with the algorithm’s sensor were performed.
Figure 7.5 and Figure 7.6 show the order distribution in the region near the cylinder
when the magnitude of the velocity gradient or the entropy are used as sensors.
In the first case, it can be seen that the separated flow zone does not have strong
velocity gradients, and as a consequence the algorithm chooses to use low order
reconstructions here. On the other hand, if entropy is used as sensor, the wake
and the separation zone are correctly detected, but low order reconstructions are
selected in front of the cylinder. For this reasons a mixed approach was adopted:
both sensors were used and the order of the element was set to the maximum
between the orders suggested by the two sensors (see Figure 7.7). In all cases the
control law described by Eq. (7.4) with αp = 0.3 was employed.

99
 7 – Adaptive algorithms

Scheme nDOFs/eq Average CD St

DG1-RK2TVD 2736 1.287 0.137
DG2-RK3TVD 5472 1.334 0.161
DG3-RK4LS 9120 1.334 0.163
DG4-RK4LS 13680 1.336 0.163
p-adapt.-RK4LS 5200-5300 1.335 0.163

Table 7.1. Average drag coefficient and Strouhal number for cylinder flow
(Re∞ = 102 , M∞ = 0.2).

Author Average CD St
Norberg [107] - 0.164
Tritton [108] 1.26 0.157-0.164
Wieselsberger [109] 1.43 -
Williamson [110] 1.33 0.160-0.164

Table 7.2. Experimental results from the literature for cylinder flow (Re∞ = 102 )

Table 7.1 reports the results obtained as a function of the total number of
degrees of freedom per equation. Note that in the case of p-adaptivity the reported
number of degrees of freedom refers to the range spanned in a period. The average
drag coefficient and the Strouhal number are in line with the experimental data
available in the literature (see Table 7.2). It can be seen from these results that
the strong clustering of the mesh makes it difficult to resolve the wake in regions
far from the cylinder. The natural remedy for this problem is the use of a hybrid
hp-adaptive algorithm, which will be described in Section 7.4.

100
 7.2 – p-adaptive algorithm

Figure 7.5. Distribution of the reconstruction order with the entropy sensor.

p
4
3
2
1

Figure 7.6. Distribution of the reconstruction order with the sensor based
on velocity gradients.

101
 7 – Adaptive algorithms

p
4
3
2
1

Figure 7.7. Distribution of the reconstruction order using both the entropy and
the velocity gradient sensors.

Figure 7.8. Entropy field around a circular cylinder with p-adaptivity (Re∞ = 102 ).

7.3 h-adaptive algorithm

Several flows in the field of aerospace propulsion are characterized by the presence of
shock waves or other singular features. For this reason, an h-adaptive algorithm has
been developed during the research activity. This algorithm is based on isotropic

102
 7.3 – h-adaptive algorithm

splitting of triangular meshes. When an element requires to be refined, it is split

in four elements and then it is deactivated but kept in the memory. In particular,
a hierarchical data structure has been introduced: each element knows his "sons"
and his "father". In this way, it is possible not only to refine the mesh but also
to re-agglomerate elements. This feature is useful in unsteady problems in which
there are moving shock waves inside the domain.

Figure 7.9. Isotropic splitting on unstructured meshes: initial (black) and

adapted (red) meshes.

The meshes obtained by the application of this algorithm cointain hanging

nodes (see Figure7.9). This is not a problem in the framework of discontinuous
Galerkin methods thanks to the locality of the reconstruction. On the contrary,
finite volume methods or continuous Galerkin finite elements methods require spe-
cial treatments for the reconstruction on non-conforming meshes.

7.3.1 Transonic NACA0012: h-adaptivity

The described h-adaptive algorithm, together with the error indicator of Eq. 7.1,was
used for an inviscid transonic test case on the NACA0012 airfoil. This test case was
proposed at the First, Second and Third Workshops on High-Order CFD Methods
([68], [11], [69]). The computations performed during the research activity were
submitted to the Third Edition of the Workshop (Orlando, 2005).
The flow field is characterized by a transonic inviscid flow on a NACA0012 airfoil.
The freestream conditions are M∞ = 0.8 and α∞ = 1.25o . The original NACA0012
airfoil has a finite trailing edge of 0.252% chords. The following modified version

103
 7 – Adaptive algorithms

with a different x4 coefficient is considered in order to have a trailing edge with

zero thichness:
√
y = ±0.6 0.2969 x − 0.1260x − 0.3516x2 + 0.2843x3 + 0.1036x4 0≤x≤1

(7.5)
The flow is characterized by the presence of two shock waves: a strong shock
wave on the suction side and a weak shock wave on the pressure side. Furthermore,
the sharp trailing edge introduces a singularity in the solution.
In order to efficiently use the backward Euler implicit integration scheme, the arti-
ficial viscosity approach described in Section 6.2 is chosen. The artificial viscosity
parameter is set as 0 = 0.004 in all the simulations for this test case. Convective
fluxes are approximated by the Osher numerical flux [35] while diffusive fluxes are
approximated by the ESR method [2].
The computational domain is reported in Figure 7.14. The freestream boundaries
are at more than 2000 chords from the airfoil, according to the suggestions reported
on the Workshop website [69].
In Figure 7.10 the initial mesh is reported: it is a very coarse mesh with 1060
triangular elements. Wall elements have curvilinear cubic edges. Since this is a
steady problem, the adaptation process refines the initial mesh recursively, without
the need to re-agglomerate elements.
In order to accelerate convergence, the solution is obtained through a sequence of
progressively higher-order reconstructions according to the strategy described in
Section 8.1. The adaptation procedure is performed when the final reconstruction
order is reached. At each refinement step the algorithm splits 20 % of the elements
which are characterized by the largest values of the error indicator. After that, a
further check and splitting is performed in order to avoid too large jumps in the
mesh size distribution. This last step is required by the employed visualization
subroutines (see Appendix C) which are not able to deal with too large jumps in
the mesh. However, it is not required by the numerical discretization and can be
avoided by the use of an updated visualization subroutine.
When a wall element is refined, his sons are still curvilinear cubic elements:
the extra points required to define the wall geometry are obtained by Eq.7.5. In
Figure 7.11 the adapted mesh after 5 refinement steps is reported. The Mach field
on the initial and the adapted meshes is reported in Figure 7.12 and Figure 7.13,
respectively. A comparison between Figure 7.13 and Figure 7.11 shows that the
error indicator identifies correctly the main sources of error in the flow: the leading
edge, the trailing edge, the shocks and the contact surface downstream of the
profile.
The performances of the algorithm are evaluated for p = 1 and p = 2 by
computing the errors on the lift (CL ) and drag (CD ) coefficients. The errors are
evaluated by taking the RWTH values from the second edition of the Workshop
[11] (Cl = 3.52914010−1 ;Cd = 2.27463610−2 ) as reference. In Figure 7.16 and

104
 7.3 – h-adaptive algorithm

Figure 7.15 the errors on the lift and drag coefficients are reported as a function
of the equivalent length scale. It can be seen that the convergence behavior is
quite irregular. This is in line with the results obtained by other groups and is
due to the presence of singularities in the solution. Furthermore, the results show
that, with the setup chosen for this test, there is no advantage in using uniform
high-order reconstructions in the presence of discontinuities. Indeed, the results
for p = 1 and p = 2 are quite close in Figure 7.15 while in Figure 7.16 the p = 1
reconstruction performs better than the p = 2 scheme. A similar behavior can be
seen in Figure 7.15 and Figure 7.16 where the errors are plotted as a function of
the work units. These results suggest the introduction of an hybrid hp-adaptive
strategy, which will be described in the following Section.

Further investigations should be performed in order to clearly understand why

the h-adaptive p = 1 scheme performs better than the h-adaptive p = 2 scheme
in this test case. A possible answer is related to the fact that the extra degrees of
freedom introduced by the higher-order reconstructions increase significantly the
cost of the computation but they introduce a small benefit. Indeed, the global
error is dominated by the presence of the singularities on which the high-order
schemes show an accuracy degradation. Starting from this point, it could be useful
to introduce a very aggressive h-refinement strategy in order to keep the error
related to the singularities sufficiently low: in this way the high-order schemes
would be able to show their accuracy which would not be hided by the errors on
the singularities. If this hypothesis is correct, than the h-adaptive strategy should
become more aggressive when high-order reconstructions are chosen. This would
explain why the proposed h-adaptive strategy (which is very "soft" because at each
refinement the size of the elements is only halved) gives good results for p = 1
but it is not convenient for p = 2. In order to adopt more aggressive h-refinement
strategies it would be possible to remesh the domain after each adaptation instead
of splitting the elements. However, this last approach would be very expensive in
unsteady problems.

105
 7 – Adaptive algorithms

Figure 7.10. Initial mesh (1060 elements).

Figure 7.11. Adapted mesh after 5 h-refinements (11806 elements).

106
 7.3 – h-adaptive algorithm

Figure 7.12. Transonic NACA0012: Mach field on the initial mesh (p=1).

Figure 7.13. Transonic NACA0012: Mach field on the adapted mesh

after 5 h-refinements (p=1).

107
 7 – Adaptive algorithms

2000

1000

0
y

-1000

-2000

-2000 -1000 0 1000 2000 3000 4000

x

Figure 7.14. Computational domain for NACA0012 problem.

7.4 hp-adaptive algorithm

In the previous Sections of this Chapter both p-adaptive and h-adaptive algorithms
are described. Both algorithms are more efficient than uniform (order and mesh re-
finement) discretization in smooth and discontinuous problems, respectively. How-
ever, real problems are characterized by a mix of smooth and singular regions. For
this reason, it is natural to introduce an hybrid approach in which p-refinement is
used in smooth regions and h-refinement is used on singularities.

The previously described test case on the NACA0012 airfoil is considered here
as example. Since the flow is steady and the initial mesh is very coarse then the
algorithm performs only element splitting and there is no need to agglomerate
elements.
The approach chosen for this work can be summarized by the following steps:

• Evaluate the error indicator (Eq. 7.1) in each element

• Select a fraction of the total elements (e.g. 20%) with the largest error

• Perform a smoothness test on the selected elements and choose between h or

p-refinement

The smoothness test is performed by the discontinuity indicator proposed by

Persson and Peraire [86] applied to pressure. The indicator is the same which is

108
 7.4 – hp-adaptive algorithm

used for the feedback filtering algorithm (see Section 6.1). The threshold chosen
for this test case is:
SR = −2.5 − 4 log10 (pe ) (7.6)
where pe is the reconstruction order inside the element. If SI > SR then h-
refinement is chosen. Otherwise, the order of the element in increased to the level
pe + 1.

The results reported in Figure 7.15, Figure 7.16, Figure 7.17 and Figure 7.18
show clearly the advantages of the hp-adaptive algorithm over the h-adaptive al-
gorithm, both in terms of equivalent length scale and work units.

-3
10
|Cd error|

10-4 p=1 h-adapt

p=2 h-adapt
hp-adapt

-5
10

-6
10
0.005 0.01 0.015 0.02
h

Figure 7.15. Transonic NACA0012: Cd error vs length scale

p=1 h-adapt
p=2 h-adapt
hp-adapt

10-3
|Cl error|

-4
10

0.005 0.01 0.015 0.02

h

Figure 7.16. Transonic NACA0012: Cl error vs length scale

109
 7 – Adaptive algorithms

-3
10
p=1 h-adapt
p=2 h-adapt
hp-adapt
|Cd error|

10-4

-5
10

0 1 2
10 10 10
Work units

Figure 7.17. Transonic NACA0012: Cd error vs work units

-2
10
p=1 h-adapt
p=2 h-adapt
hp-adapt
|Cl error|

10-3

-4
10 0 1 2
10 10 10
Work units

Figure 7.18. Transonic NACA0012: Cl error vs work units

110
 Part III

Numerical results

111
Chapter 8

Performances

This Chapter is devoted to the performances of the proposed numerical discretiza-

tion. It is subdivided in two Sections. The first Section is devoted to the techniques
that have been implemented in this work in order to accelerate the convergence for
steady problems. In the second Section there are the results for some test cases
from three Workshop on High-Order CFD Methods ([68], [11], [69]).

8.1 Acceleration for steady problems

One of the lessons learned from CFD Workshops [68],[11],[69] is that the compu-
tational time required to obtain a given error level is strongly affected by both
the chosen numerical methods and the implementation. Indeed, it is quite com-
mon to see different groups which obtain the same error level with very different
computational times (for example by a factor 102 or 103 ). These differences be-
come even more evident when steady test cases are considered. Indeed, the total
computational time required to obtain the converged steady-state solution can be
dramatically reduced by the use of some acceleration techniques.

First of all, implicit time integration schemes can be quite effective in damping
the transient. This is particularly true when the physical model contains equations
with stiff terms, which can represent a severe bottle-neck for explicit time integra-
tion schemes. For example, in Section 4.5 the advantages of the implicit approach
over the explicit RK4 scheme are clearly shown in a heat diffusion problem, as can
be seen in Figure 4.8. Similar advantages can be found in more complex system of
equations, like for example Navier-Stokes equations or RANS equations. However,
in this case the benefits are reduced by the presence of non-linear terms which can
introduce a limit on the maximum allowable CFL number, even with implicit time

113
 8 – Performances

integration schemes. For example, the presence of moving shock waves inside the
domain introduces a severe limitation on the CFL number.
In this work, the robustness of the implicit time integration is increased by the
introduction of some positivity checks on pressure and density: after each time
step, the sign of pressure and density in the quadrature points is checked. If a neg-
ative value is found, then the last update is removed and the solution is brought
back to the previous time step. After that a smaller time step is chosen and the
procedure is repeated. This approach is used together with several CFL evolution
laws, which are described in Section 5.2.3. It is important to keep in mind that a
small variation in the parameters which control the CFL evolution law can change
the order of magnitude of the total computational time. Clearly, a compromise
between speed and robustness has to be chosen.
Finally, it is necessary to remember the main drawback of implicit schemes: the
large memory requirements. Indeed, the jacobian matrix which results from the
implicit discretization is very large and its storage in the memory can be pro-
hibitively expensive in complex 3D problems.

An effective technique which can reduce the computational time in steady prob-
lems is the multigrid method. This approach is commonly used in the framework
of finite volume methods. The solution is advanced in time on several levels of
nested grids: the idea is that the low frequency oscillations which characterized
the transient on the original grid can be be damped more quickly on a coarser
grids, where they appear as high frequency oscillations. The described approach
can be referred to as h-multigrid, where the "h" indicates that the acceleration is
obtained by exploiting different grids.
In the framework of discontinuous Galerkin methods, an alternative approach can
be chosen: the p-multigrid method. The idea is to substitute the different level of
grids with different reconstruction orders. Furthermore, if the solution is described
by a hierarchical basis, it is very easy to reduce and extend the solution from a
level to the next. Several works have been done on this topic (see for example
[111], [112], [113]). However, some authors have found that p-multigrid techniques
applied to DG discretization are not very efficient if the time integration is per-
formed by explicit schemes [112]. For this reason, p-multigrid DG methods are
usually associated with implicit time integration, at least for the coarser level of
the solution which involve less degrees of freedom and so less memory require-
ments. This behavior is in contrast to what happens in the framework of finite
volume methods in which h-multigrid methods are efficiently used together with
explicit RK schemes.
In this work, the multigrid approach has not been implemented yet but a simpler
(but less effective) approach is used. In particular, the hierarchical nature of the
employed modal bases is exploited to perform a sequence of computations in which
the reconstruction order p is increased from p = 0 to the final p. In this way, a

114
 8.1 – Acceleration for steady problems

sequence of progressively more accurate (and more expensive) solution is obtained,

and each solution is initialized by the previous coarser solution.
In Figure 8.1 the evolution of the residuals is reported as a function of the compu-
tational time for the test case of the channel with the smooth bump (see Section
8.2.1). The computations were performed by a RK3 time integration scheme. The
green curve describes the results obtained by using a p = 2 reconstruction from
the beginning of the simulation. The blue curve refers to the results obtained by
a sequence of progressively higher-order reconstructions, from p = 0 to p = 2. The
plot shows that, even in this simple and smooth test case, the technique allows a
reduction in the computational time.

100

-1
10

10-2 DG2+RK3
p refinement
-3
10
Residual ratio

-4
10

10-5

-6
10

10-7

-8
10

10-9

0 50 100 150 200

CPU time

Figure 8.1. Channel with smooth bump: acceleration by p-refinement.

The approach becomes even more convenient in problems in which there are
shock waves. Indeed, the initial transient is usually characterized by the presence
of moving shock waves which limit the maximum CFL number. However, if the
computation is started with a p = 0 reconstruction then an approximate solution in
which the shock waves are at rest is quickly obtained. Then the expensive higher-
order reconstructions are activated and they can benefit of a larger time step since
the shock waves are already close to the final position.

Finally, similar benefits can be obtained by the use of h or hp adaptive al-

gorithms. Indeed, the recursive local refinement makes it possible to obtain a

115
 8 – Performances

sequence of progressively more accurate (and more expensive) solutions, in which

each solution is initialized by the previous coarser solution.

8.2 Workshop test problems

The proposed discretization has been evaluated on several test cases from the three
editions of the Workshop on High-Order CFD Methods ([68],[11],[69]). These tests
are very useful because allow a direct comparison of the accuracy and the efficiency
of the proposed discretization with different approaches proposed by other groups.
The results for the transonic NACA0012 and the laminar flat plate test cases which
were submitted to the Third Workshop [69] are not reported in this Section because
they are already described in Sections 7.3.1 and 4.4.4, respectively.

8.2.1 Subsonic flow in a channel with a smooth bump

This test case was proposed at the First and Second Workshops on High-Order CFD
Methods ([68],[11]). The computations performed during the research activity were
submitted to the Second Edition of the Workshop.
The problem is characterized by an inviscid steady flow in a channel with a smooth
bump. The flow field is subsonic. The geometry of the problem is reported in
Figure 8.2.

Y
Z X

Figure 8.2. Channel with a smooth bump.

The lower and upper boundaries are described by the following laws:
2
y = 0.0625e−25x − 1.5 ≤ x ≤ 1.5 (8.1)

y = 0.8 − 1.5 ≤ x ≤ 1.5 (8.2)

116
 8.2 – Workshop test problems

They are treated as solid walls. The left side is a subsonic inlet in which total
temperature, total pressure and velocity angle are imposed (Tin 0
= 1, Pin0
= 1 and
αin = 0 ). The right side is a subsonic outlet in which the static pressure is im-
o

posed in order to obtain the prescribed exit Mach number (Mexit = 0.5).
The flow field was initialized with a uniform flow field at M = 0.5. The computa-
tions were carried on until the L2-norm of the density residual was reduced by 10
orders of magnitude with respect to the initial value. In particular, the residuals
of the first modal coefficient (constant basis function) were monitored.
The quality of the numerical results was evaluated by the computation of the
L2-norm of the entropy error as a function of the equivalent length scale. The
performances of the code were evaluated by measuring the work units required to
obtain the steady solution (defined as the solution obtained after a reduction of 10
orders of magnitude in the density residuals).

The results submitted by the author to the Second Edition of the Workshop
were obtained by a preliminary version of the code. In particular, that version
could perform only explicit time integration by means of Runge-Kutta schemes.
For this reason, it was very slow in steady problems like the described one.
In order to accelerate the convergence to the steady solution, a sequence of increas-
ing order reconstructions are employed. This approach introduces a small benefit
as is shown in Figure 8.1 of Section 8.1.
These simulations were carried out by using an element basis obtained by a tensor
product of Legendre polynomials. The orthonormalization of the basis was not
available in that preliminary version of the code.
The domain was discretized by a sequence of structured meshes (16x4, 32x8, 64x16,
128x32). The elements at wall had curvilinear edges (up to quartic elements) while
the internal elements were linear.

In Figure 8.3 the L2 entropy error is reported as a function of the equivalent

length scale for several reconstruction orders (p = 1,2,3). The numerical results
show a convergence rate which is very close to the theoretical value (dotted lines).
In Figure 8.4 the L2 entropy error is reported as a function of the work units.
This plot shows that, in this problem, high-order methods are more convenient
than low-order methods. This benefit can be significantly increased by the use of
implicit time integration schemes which can dramatically reduce the convergence
time of high-order discretizations in this steady problem.
The main achievement obtained by this test case is the confirm that the code is
able to obtain the expected convergence rate on smooth problems with curvilinear
meshes.

117
 8 – Performances

Figure 8.3. Channel with smooth bump: entropy error vs equivalent length scale.

Figure 8.4. Channel with smooth bump: entropy error vs work units.

118
 8.2 – Workshop test problems

8.2.2 Transonic Ringleb flow

The Ringleb flow is a transonic inviscid smooth flow. Its analytical solution can be
found by the hodograph method (see Section 6.1.6). This test case was proposed at
the First, Second and Third Workshops on High-Order CFD Methods ([68], [11],
[69]). The computations performed during the research activity were submitted to
the Second Edition of the Workshop.
The domain is reported in Figure 8.5. It is delimited by two streamlines which
are defined by the values κmin = 0.7 and κmax = 1.5 where κ is the streamlines
function already used in Section 6.1.6. The subsonic inlet (red curve) and the sub-
sonic outlet (blue curve) are defined by the isospeed line q̄ = 0.5 where q̄ is defined
according to Section 6.1.6. It is interesting to note the the isospeed lines are cir-
cles and so the inlet and the outlet are curvilinear boundaries. At the inlet, total
temperature, total pressure and flow angle are imposed. At the outlet the static
pressure is imposed in order to obtain the prescribed value for the exit velocity.
The flow field is smooth and it is characterized by a transonic acceleration region
and a subsequent transonic deceleration region.
The numerical solution of this test case is difficult due to the presence of an insta-
bility in the transonic deceleration region. Indeed, if the solution is not sufficiently
accurate in this zone, a spurious shock wave can appear (see Figure 8.7). It is cru-
cial to properly describe the geometry of the boundaries. In the tests performed
during this work, it was found that it is mandatory to describe the wall elements by
high order curvilinear mappings. Furthermore, the coordinates of the wall bound-
ary points have to be defined in double precision. Indeed, the small perturbations
related to a single precision boundary description are sufficient to promote the de-
velopment of the shock. A similar behavior has been found by other participants to
the Workshops. Some groups have found that the shock can be more easily avoided
if the lateral boundaries are described by an exact boundary condition instead of a
solid wall boundary condition. However, in this work only the solid wall boundary
condition is used. This choice is due to the fact that in general problems the exact
solution is not known and so it is better to test the behavior of the discretization
with a more affordable boundary condition.
The instructions for this test case require that the solution is initialized with the
exact solution and then the computations have to be carried on until there is a
drop of 10 orders of magnitude in the density residuals. Since the initial solution
is a projection of the exact solution then the initial residuals are very low: in some
cases it is impossible to reduce the density residuals by a factor equal to 10−10
because the residuals become too close to the machine precision. An example of
the residual history for one of these cases is reported in Figure 8.9. For this reason,
in some configurations a less severe convergence criteria is adopted (reduction of a
factor 10−8 ).
All the computations were performed on structured meshes (18x12, 36x24, 72x48)

119
 8 – Performances

generated by a free code [114] and curved by an in-house tool. The solution inside
each element is described by a tensor product of Legendre polynomials.
The accuracy of the simulations was evaluated by computing the L2-norm of the
entropy error on the steady solution.
In Figure 8.10 the error is reported as a function of the equivalent length scale for
p = 1,2,3 reconstructions. In the plot there is only one point for the p = 1 recon-
struction. This is due to the fact that a shockless solution was obtained only on
the finest mesh (72x48) with p = 1. In contrast, the higher-order reconstructions
are able to give the shockless solution even on the coarsest meshes. The observed
convergence rate for p = 2 and p = 3 is very close to the theoretical value. Further-
more, the error level is good if compared to the results of the other participants.
As far as the performances are concerned, these computations were performed at
the end of the first year and so only explicit time integration schemes were available
in the code. In particular, explicit Runge-Kutta schemes are employed in order to
obtain the same discretization order in space and time. In Figure 8.11 the error
is reported as a function of the work units for the different reconstruction orders.
From this plot it is not possible to understand clearly which is the most conve-
nient reconstruction order. This irregular behavior is probably due to the fact
that different convergence criterias are used when the residuals are too close to the
machine precision. Furthermore, the performances of the higher-order schemes in
this test case can be positively affected by the use of an implicit time integration
scheme.
However, the main achievement of this test case is the confirm that the proposed
discretization is able to describe the shockless transition: this is a quite severe test
according to the literature [115].

Figure 8.5. Domain for the transonic Ringleb flow.

120
 8.2 – Workshop test problems

Figure 8.6. Mach field for the transonic Ringleb flow: initial solution.

Figure 8.7. Mach field for the transonic Ringleb flow: spurious shock.

Figure 8.8. Example of structured mesh for the transonic Ringleb flow.

121
 8 – Performances

Figure 8.9. Transonic Ringleb flow: convergence history.

Figure 8.10. Transonic Ringleb flow: entropy error vs equivalent length scale.

122
 8.2 – Workshop test problems

Figure 8.11. Transonic Ringleb flow: entropy error vs work units.

8.2.3 Vortex transport by uniform flow

This test case was proposed at the First, Second and Third Workshops on High-
Order CFD Methods ([68], [11], [69]). The computations performed during the
research activity were submitted to the Second Edition of the Workshop.
The aim of this problem is to test high-order method’s capability to preserve vor-
ticity in an unsteady inviscid flow. This capability is important in the framework
of Large-Eddy and Detached-Eddy simulations and has to be preserved for a wide
range of Mach number, including very low Mach numbers.
Compressible solvers can give several problems when applied to very low Mach
number flows. Also this aspect is investigated in this test case.
The domain is square ([0,1] × [0,1]) and periodic conditions are imposed on all
boundaries. The flow field is initialized by a uniform field (M∞ = 0.05) super-
posed to a vortical motion described by the following perturbations on the velocity
components and on the temperature:
y − 0.5 −r2 /2
δu = −u∞ β̄ e (8.3)
R

x − 0.5 −r2 /2
δv = u∞ β̄ e (8.4)
R
2
δT = 0.5(u∞ β̄)2 e−r /Cp (8.5)

(x − 0.5)2 + (y − 0.5)2
p
r= (8.6)
R
123
 8 – Performances

where β̄ = 1/50, R = 0.05 and Cp is the constant pressure specific heat.

In the exact solution the superposed vortex is transported without distortions by
the uniform flow.
The simulation has to be carried on until 50 vortex revolutions are computed and
then the L2-norm of the velocity error is computed. Since the flow is periodic the
initial solution can be used as reference solution for the computation of the error.

Also for this test case, the element basis was obtained by a tensor product of
Legendre polynomials without orthonormalization. Time integration is performed
by means of explicit Runge-Kutta algorithms (described in Section 5.1) chosen in
order to match the orders of accuracy in space and in time.
A first set of simulations was performed on cartesian (regular) structured meshes
(24x24, 48x48, 96x96) for p = 1,2,3. In Figure 8.12 the distribution of the y-
component of the velocity is reported at the end of the simulation. The results are
reported for the p = 1 computation on the 64x64 mesh and for the p = 3 compu-
tation on the 32x32 mesh. This two computations require the same total number
of degrees of freedom but the results are significantly different. Indeed, while the
p = 3 reconstruction describes accurately the vortex, the p = 1 reconstruction
shows a large error.
In Figure 8.13 the error is reported as a function of the equivalent length scale. It
can be seen that the p = 1 reconstruction is very far from the asymptotic range and
so its convergence rate is very low. In contrast, the p = 2 and p = 3 reconstructions
reach quickly a convergence rate very close to the theoretical value. These results
show that the higher-order reconstructions are significantly more convenient than
the p = 1 scheme, for this kind of problems.
The same conclusions can be found from the plot of the work units. Indeed, Fig-
ure 8.14 shows that the work units required by the higher-order schemes to obtain
a given level of error are significantly lower than the work units required by the
second order scheme (p = 1).

124
 8.2 – Workshop test problems

0.01
DG1 64x64
DG3 32x32
Exact
0.005

V/U ∞
0

-0.005

-0.01

0.2 0.4 0.6 0.8

X

Figure 8.12. Vortex transport: final velocity distribution on regular mesh.

Figure 8.13. Vortex transport: error vs equivalent length scale on regular meshes.

A second set of simulations was performed on a sequence of three perturbed

meshes obtained the application of random perturbartions on the coordinate of the
nodes. In particular, the maximum perturbation is defined as

δM AX = 0.15h (8.7)

where h is the size of the elements. An example of this process is reported in

Figure 8.15.
In Figure 8.16 and Figure 8.17 the error is reported as function of the equivalent

125
 8 – Performances

Figure 8.14. Vortex transport: error vs work units on regular meshes.

Figure 8.15. Vortex transport: example of regular and perturbed mesh.

length scale and work units for the simulations on the perturbed meshes. The
results do not show significant differences with respect to the results on the regular
meshes: this means that the proposed discretization is not very sensitive to the
quality of the mesh and so it can be interesting for the use on irregular meshes
obtained by general purpose mesh generator for complicate geometries.

126
 8.2 – Workshop test problems

Figure 8.16. Vortex transport: error vs equivalent length scale on perturbed meshes.

Figure 8.17. Vortex transport: error vs work units on perturbed meshes.

127
 8 – Performances

8.2.4 Unsteady viscous flow over tandem airfoils

This 2D test case was proposed at the First and Second edition of the International
Workshop on High-Order CFD Methods ([68], [11]). It is characterized by the
presence of two NACA0012 airfoils in tandem configuration. The leading airfoil is
rotated by 10◦ about its aerodynamic center (25% of the chord) while the trailing
airfoil is translated backward by 1.5 chords. The far field boundary is a circle with
a radius of 100 chords centered on the leading edge of the trailing airfoil. The free-
stream Mach number is M∞ = 0.2 and the Reynolds number based on the chord
of an airfoil is Re∞ = 104 . An initial smooth condition is prescribed according to
the following relations:

v(x, y) = 0 P (x, y) =
 P∞ ρ(x, y) = ρ∞
u∞ d > δ , (8.8)
u(x, y) =
2δ ) d ≤ δ
u∞ sin( πd
where u and v are the cartesian velocity components, d is the distance to the
nearest wall and δ is a constant equal to 5% of the chord length.
Two mixed structured-unstructured grids with 2101 and 8404 elements were
used. Quadrilaterals were used near the airfoils and triangles in the rest of the
domain. Elements with cubic edges were employed at solid walls. Here the size
of the first layer of elements was equal to 0.005 chords in the direction normal to
wall for the finest mesh. Figure 8.18 shows a detail of the finest mesh near the two
bodies.

Figure 8.18. Detail of the finest mesh used for the tandem airfoils test.

The drag coefficient of the trailing airfoil was monitored during the first 20
convective times. Several simulations were performed by changing the reconstruc-
tion order from DG1 to DG5 on both meshes. Time integration was performed by
RK4LS.
The flow field is characterized by the stall of the leading airfoil. The vortices which

128
 8.2 – Workshop test problems

are generated in the separation region strongly interact with the trailing airfoil.
A plot of the entropy field at the end of the simulation is reported in Figure 8.19
for DG5 on the finest mesh. The time evolution of the drag and lift coefficient for
the trailing airfoil is reported in Figure 8.20 and Figure 8.21 for p = 3,4,5 on the
finest mesh.
It is clear from these results that the flow is not periodic, at least in the time
interval considered here. The chaotic aperiodic evolution of the vortices makes it
difficult to perform a deterministic study. The results obtained by the different
simulations tend to overlap for the first part of the computation but after a certain
time they follow a different evolution. It can be seen that when the number of
degrees of freedom is increased (with both mesh refinement or order refinement)
the time interval in which the different solutions overlap becomes more extended.
In particular the plots show that the DG4 method follows well the DG5 solution
up to time = 11. In contrast, the DG3 prediction follows the higher order solutions
only up to time = 5. A similar behavior can be seen in the results reported by
other groups on the website of the First edition of the Workshop [68]. This test
case showed that the proposed discretization is able to deal with strong unsteady
wall-vortex interactions.

Figure 8.19. Entropy field for the tandem airfoils test at the end of the
simulation (DG5RK4LS finest mesh)

129
 8 – Performances

0.2

0.15 DG3RK4LS
DG4RK4LS
DG5RK4LS
0.1

0.05
Cd

-0.05

-0.1

-0.15

-0.2
0 5 10 15 20
Time

Figure 8.20. Drag coefficient for the trailing airfoil (finest mesh p=3,4,5).

1.5 DG3RK4LS
DG4RK4LS
DG5RK4LS
1

0.5
Cl

-0.5

-1

-1.5
0 5 10 15 20
Time

Figure 8.21. Lift coefficient for the trailing airfoil (finest mesh p=3,4,5).

130
Chapter 9

Aerospace propulsion
examples and validation

9.1 VKI-LS59 turbine cascade

The VKI-LS59 turbine-rotor cascade was studied during the research activity in
order to evaluate the behavior of the proposed discretization on high Reynolds
number turbomachinery flows. The experimental data available in the literature
([116], [117]) indicate that the flow is nearly two-dimensional. Indeed, end-wall
effects and aspect ratio influence are practically negligible and so it is acceptable
to perform a 2D simulation.
The simulation refers to a transonic working condition, characterized by the fol-
lowing parameters:

αin = 30o M2is = 1.2 Re2is = 6.63 × 105 (9.1)

The Reynolds number Re2is is based on the blade chord c and the isentropic exit
conditions.
The Spalart-Allmaras model is used in fully-turbulent mode. Indeed, the modified
turbulent viscosity at the inlet is set as ν̃ = 3ν, according to [17].
The computations were carried out with the basis A and a reconstruction order
p = 1 is chosen. Shock capturing was performed by means of the feedback filtering
algorithm proposed in Section 6.1. Time integration was performed by an explicit
RK2 algorithm. A set of refined meshes was employed in order to evaluate the
grid dependence behavior. The final mesh contains 19866 elements. It is an hybrid
mesh in which there is a structured O-type zone near the blade. The surrounding
region is filled with an unstructured quadrilateral mesh, as Figure 9.1 shows. Wall
elements have parabolic edges. The size of the elements near the wall is 5 × 10−4 c

131
 9 – Aerospace propulsion examples and validation

in the normal direction. The distance between the leading edge and the inlet is
equal to one chord. The same distance is adopted between the trailing edge and
the outlet.
In Figure 9.2 the computed Mach field is reported. The results show clearly the
presence of the two shock waves which are generated by the trailing edge. Further-
more, it is possible to see on the suction side the interaction between the boundary
layer and the shock wave which is coming from the next blade. In particular, the
incoming shock wave generates two reflected shock waves due to the presence of a
small separation bubble in the impinging point.
In Figure 9.3 the wall isentropic Mach number distribution is reported and com-
pared with experimental data from [116]. The plot shows that the simulation is
able to detect the main features of the distribution. In particular, the position of
the shock on the suction side is correctly computed.

Figure 9.1. VKI-LS59 cascade: detail of the mesh.

9.2 T106c turbine cascade

Nowadays, several efforts are devoted to the improvement of low pressure turbines
(LPT) in turbofan engines. One of the goals is the reduction of the weight. This
leads to a reduction in the number of blades. The obvious consequence is that the
remaining blades have to be more loaded. For this reason, high-lift and ultra-high-
lift blades have been developed. When a blade is forced to work in such conditions,
in which the turning angle is very high, it is mandatory to optimize the blade’s
shape in order to reduce fluid dynamics losses, otherwise the benefits obtained by
the weight reduction can be obfuscated by a deterioration of the efficiency.
Hourmouziadis [118] reported the working conditions for a LPT for the entire flight

132
 9.2 – T106c turbine cascade

Figure 9.2. VKI-LS59 cascade: Mach field.

1.5

1
Mis

0.5

0
0 0.2 0.4 0.6 0.8 1
x/c x

Figure 9.3. VKI-LS59 cascade: wall isentropic Mach distribution.

133
 9 – Aerospace propulsion examples and validation

envelope of a typical commercial aircraft. This data, which are representative of

the late 1980s design, show that at cruise condition the Reynolds number for the
blades of the LPT is below 200000 while at takeoff is more than 400000. Nowadays,
the LPT Reynolds number at cruise condition can be even below 100000.
This very low Reynolds number introduces several problems related to the aero-
dynamic behavior of the blades. At high Reynolds number the flow is usually
attached and so the required turning angle is obtained. On the contrary, at low
Reynolds number it is possible to observe separation which reduces the efficiency
and the turning angle. The less severe separation, which can occur at intermediate
Reynolds numbers, is characterized by the presence of a separation bubble after
which the flow reattachs. However, for very low Reynolds number the separation
bubble becomes larger and finally an open separation is obtained. This last condi-
tion is very detrimental in terms of efficiency.
For these reasons, several efforts have been done in the last years for the prediction
of these phenomena. The T106c high-lift cascade is representative of this class
of problems. It has been investigated both experimentally and numerically. The
testing working conditions are characterized by low Reynolds numbers and several
turbulence intensity. The upstream turbulence intensity can significantly affect the
separation behavior.
In this work, the T106c cascade was studied by means of the developed discon-
tinuous Galerkin approach applied to the laminar kinetic energy transition model
(see Section 2.3.3). Two working conditions were considered, in which the exit
isentropic Reynolds number is 185000 and 80000, respectively. The isentropic exit
Mach number is 0.65 and the inlet flow angle is 32.7◦ , for both cases. The inlet
turbulence intensity was set as Tu = 0.8%, according to the experimental value
which corresponds to the tests without grid in the wind tunnel. The inlet turbu-
lence length scale was set as lt /cx = 2.5 × 10−3 , where cx is the axial chord. This
value comes from a study on the decay of inlet turbulence [15]. The inlet specific
dissipation ωin was then obtained by kin and lt according to Eq. 2.67.
The experiments have shown that the flow is characterized by a laminar separa-
tion and a subsequent transition to turbulence. After the separation the flow can
reattach o remain separated, depending on the working Reynolds number.
The simulations were performed by p = 1 reconstructions on hybrid meshes. The
element basis A was employed. In Figure 9.4 an example of mesh is reported: it is
obtained by an O-type structured mesh surrounded by an unstructured quadrilat-
eral mesh. The wall elements have parabolic edges. The grid spacing close to wall
is such that the dimensionless wall distance y + is less than one on all the surface
with the exception of some points near the leading edge.
The extension of the structured zone in the wall normal direction has been chosen
by a trial-and-error process, in order to include the shear layer inside the struc-
tured mesh. This is due to the fact that the LKE model requires the evaluation of
the thickness of the shear layer (vorticity thickness) which can be evaluated more

134
 9.2 – T106c turbine cascade

easily on a structured mesh. However, the tests done in this work have showed that
the evaluation of this parameter (which controls the production term in the LKE
equation) is quite problematic. Indeed, if the topology of the grid is changed (from
O to H or C) or the direction of the grid lines in the structured region is altered
then the results can be affected and it is necessary to tune again the calibration
constant C1 which appears in Eq. 2.54. In this work the constant C1 is set to 0.009
while the original value C1 = 0.01 is suggested in [15].
In Figure 9.11 the computed wall isentropic Mach number distribution is reported
and compared with the experimental results from [119]. The simulations are able
to correctly predict how the separation region changes when the Reynolds number
is reduced. In particular, the extension of the plateau in the distribution on the
suction side and the reduction of the maximum Mach number are related to the ex-
tension of the separation. The Mach field, the laminar kinetic energy distribution
and the turbulent kinetic energy distribution are reported in Figure 9.5, 9.6, 9.7,
9.8 , 9.9, 9.10 for M2s = 185000 and M2s = 80000, respectively. Again, the pictures
show clearly the dramatic variation in the separation region when the Reynolds
number is reduced to 80000.

Figure 9.4. T106c cascade: example of hybrid mesh.

135
 9 – Aerospace propulsion examples and validation

Figure 9.5. T106c cascade: Mach at Re2s = 185000, M2s = 0.65

Figure 9.6. T106c cascade: Mach at Re2s = 80000, M2s = 0.65

136
 9.2 – T106c turbine cascade

Figure 9.7. T106c cascade: laminar kinetic energy at Re2s = 185000, M2s = 0.65

Figure 9.8. T106c cascade: laminar kinetic energy at Re2s = 80000, M2s = 0.65

137
 9 – Aerospace propulsion examples and validation

Figure 9.9. T106c cascade: turbulent kinetic energy at Re2s = 185000, M2s = 0.65

Figure 9.10. T106c cascade: turbulent kinetic energy at Re2s = 80000, M2s = 0.65

138
 9.2 – T106c turbine cascade

0.8

0.6

Exp. Re=80000
Mis

Exp. Re=185000
0.4 Num.

0.2

0
0 0.2 0.4 0.6 0.8 1
x /c x

Figure 9.11. T106c cascade: wall isentropic Mach distribution (M2s = 0.65).

139
140
Chapter 10

Conclusions

This thesis describes the main activities performed during the research project.
The reader can found an evolution in the complexity of the reported test cases
across the Chapters. Through all the work it is possible to find a mix between
numerical problems and physical modeling issues. The numerical issues dominated
the first part of the work because initially there was the need to build a solid nu-
merical platform. Several work has been done on the evaluation of the numerical
accuracy and on the performance comparison with other groups. These aspects
have also driven the development of the two proposed algorithms (Feedback Fil-
tering and Enhanced Stability Recovery) which can show both advantages and
drawbacks over other existing methods, depending on the problem.
The developed numerical platform, together with the set of physical models in-
troduced in the second part of the work, can be used to predict the behavior of
industrial flows in the aerospace field.
Several aspects can be improved. First of all, the code is parallelized only for
shared memory systems. The support for distributed memory systems should be
introduced.
As far as the implicit integration scheme is concerned, an analytic evaluation of
the jacobian should be introduced also for Navier-Stokes and RANS equations in
order to improve performances and make easier the parallelization.
Since aerospace propulsion systems are characterized by compressible flows, further
work should be devoted to shock capturing techniques. In particular, the imple-
mentation of the artificial viscosity schemes [82] and [90] would be interesting, like
the implementation of WENO limiters[93] in the DG framework.
Finally, all the simulations performed in this work are 2D and so the next step
is the extension to 3D. This would make it possible to perform LES and DNS
simulations, provided that a massive parallelization is introduced.

141
142
Appendix A

Adimensionalization

All the equations reported in Chapter 2 and implemented in the code are adimen-
sionalized with respect to the following reference values:
• LENGTH Lref
• TEMPERATURE Tref

• PRESSURE Pref
• DENSITY ρref
• SPEED uref =
p
RTref

• TIME tref = Lref /uref

• ENERGY PER MASS UNIT eref = u2ref
• ENTROPY Sref = cv (constant volume heat capacity)
When the previous adimensionalization is applied to equations with diffusive
terms then the following reference Reynolds number appears in the equations:
ρref uref Lref
Reref = (A.1)
µref

Also the data which define all the studied test cases are normalized with respect
to the previous value.

143
144
Appendix B

Normalization of
performances

In order to compare the performances of different discretizations it is fundamental

to define a set of proper cost functions. The results reported in this work are
evaluated by two cost functions: the equivalent length scale and the work units.
The equivalent length scale h is defined as:
1
h= √ (B.1)
nDOF s
where nDOF s is the total number of degrees of freedom per equation, which de-
pends on both the number of element and the order of the elements. The square
root is employed because 2D problems are considered. In 3D, the cubic root has
to be used.
The meaning of this cost function is directly related to the memory requirements
of the discretization.

The work units (W U s) are a measure of the total computational time which is
normalized in order to make possible a comparison between simulations performed
on different machines:
∆t
WUs = (B.2)
τref ncore
where ∆t is the wall clock time required by the simulation, τref is the reference time
and ncore is the number of employed cores. The reference time τref is obtained by
the TauBench program which can be downloaded from [68], [11] or [69]. TauBench
has to be run in serial, since the number of cores is already taken in account by
ncore .

145
146
Appendix C

High-order visualizations

The DG discretization introduces several degrees of freedom per equation in each

element. This is in contrast to what is done in finite volume methods, in which
only the average values of the conservative variables are known inside each control
volume. For this reason, the visualizations of high-order DG solutions require
particular care in order to take into account all the information memorized in the
elements.
In this work a simple post-processing approach has been developed. In particular,
the visualizations are done on finer meshes obtained by splitting the computational
meshes. In particular, the larger the number of degrees of freedom in the original
element is, the larger the number of subdivisions of the element will be. The
values of the solution in the additional nodes are computed thanks to the high-
order internal element reconstruction.
In this way, the output file can be read by standard visualization tools developed
for finite volumes or finite elements methods.
In the following, an example of these visualization meshes is reported for both
quadrilateral and triangular meshes. In order to simplify the implementation,
the visualization on triangular elements is performed by handling the triangles as
quadrilateral elements with a collapsed edge.

147
 C – High-order visualizations

Y
Y
Z X
Z X

Figure C.1. Curvilinear computational meshes (blue) and visualization meshes

(black) with different subdivisions for triangles and quadrilaterals.

148
Appendix D

Meshes

The meshes used during the research activity were generated in several ways.
An elliptic mesh generator developed by Professor Larocca at Politecnico di Torino
was employed for H, C or O structured meshes. This approach makes it possible
to obtain grid lines which are normal to the wall.
However, the meshes obtained by this tool are linear. In order to obtain higher-
order curvilinear meshes a specific tool has been developed. The tool reads the
linear meshes in PLOT3D format and then adds the points required to define
the curvilinear elements. Only the elements near the boundary of the domain are
curved. The tools requires a parametric representation of the solid wall. A possible
alternative approach is based on agglomeration of linear meshes [32].

Unstructured meshes were obtained by Gmsh [120]. Both the Delaunay and
the Delaunay for quads algorithms have been successfully employed, for both tri-
angular and quadrilateral meshes.

Some simulations of the flow around airfoils were performed by mixed structured-
unstructured meshes. In particular, it is useful to introduce a structured region
near the airfoil in order to use high-aspect ratio elements in the boundary layer.
Furthermore, the structured nature of the mesh in this zone is useful to simplify the
evaluation of the vorticity thickness if the LKE transition model (Section 2.3.3) is
chosen. On the other hand, unstructured meshes are very useful for the discretiza-
tion of the space far from the airfoil because the local size of the elements can be
chosen with more freedom.
In order to obtain this kind of meshes, a tool has been developed. The tool reads
the geometry of the airfoil and then projects the geometry points far from the
wall in order to define an external boundary for the structured zone. The curves
generated in this way are used to define an input file for Gmsh, which produce the

149
 D – Meshes

final mixed mesh.

Finally, the optimization tool for high-order meshes included in Gmsh is used for
some RANS simulations. Indeed, in high Reynolds number RANS simulation the
elements in the boundary layer have a very high aspect ratio. For this reason, when
the wall edge is curved the element can become singular due to overlapping edges.
The optimization tool included in Gmsh extends the deformation of the element to
several layers avoiding the overlap. In this way the mesh has curvilinear elements
not only at wall but also far from it.

150
Appendix E

List of symbols

In the following the list of the main symbols used through the text is reported.

Model variables
a = Speed of sound
c = Maximum signal speed
cf = Wall friction coefficient
cP = Wall pressure coefficient
E = Total energy (internal+kinematic) per unit volume
F, G = Physical fluxes
k = Turbulent kinetic energy
kL = Laminar kinetic energy
lt = Turbulence length scale
u = Generic conservative variable
Mis = Isentropic Mach number
M∞ = Freestream Mach number
P = Pressure
P ◦ = Total pressure
P r = Prandtl number
P rt = Turbulent Prandtl number
Q = Source term in the heat equation
q = Velocity magnitude
qx , qy = Cartesian components of heat flux
Reref = Reference Reynolds number obtained by adimensionalization
Re∞ = Freestream Reynolds number
S = Entropy
Sij = Mean strain rate tensor
T = Temperature

151
 E – List of symbols

T ◦ = Total temperature
Tu = Turbulence level
u, v = Cartesian component of velocity
uτ = Wall friction velocity
α∞ = Freestream angle of attack
γ = Specific heat ratio
µ = Dinamyc viscosity
ν = Kinematic viscosity
λ = Heat conductivity
τw = Wall shear stress
ω = Specific dissipation rate
ρ = Density
τij = Laminar stress tensor
τ̄ij = Turbulent stress tensor
τ̂ij = Total (laminar+turbulent) stress tensor
cb1 , σ, cb2 , κ, cw1 , cw2 , cw3 , cv1 , ct1 , ct2 , ct3 , ct4 , rlim , cv2 , cv3 = Constants for the Spalart-
Allmaras model
σk , σω , β0∗ , β0 , α0 , α0∗ , Rk = 6, Rω = Constans for the low-Re Wilcox k − ω model
C1 , C2 , C3 , C4 = Constants for the laminar kinetic energy model

Geometry
d = Distance from the closest solid wall
L1 , L2 , L3 = Area coordinates for triangle
nx , ny = Cartesian components of normal unit vector
x, y = Cartesian coordinates in physical space
ξ, η = Cartesian coordinates on the reference quadrilateral and local coordinates
for interface reference frame

h iDiscretization
B̂ = Boundary matrix for Dirichlet BC’s with the ESR method
CF L = Stability coefficient for time integration
[D] = Element derivative matrix
[Dr ] = Recovery derivative matrix
Fh , Gh = Physical fluxes computed as a function of uh
F̂h , Ĝh = Interface numerical fluxes
g = User defined constant in the feedback filtering procedure
H = Element characteristic size
h = Equivalent length scale which depends on nDOF s
[J] = Jacobian matrix for implicit scheme
[M] = Element mass matrix
[M] = Global mass matrix

152
 N = Number of degrees of freedom in an element
NF = Number of faces in an element
nDOF s = Total number of degrees of freedom per equation
p = Reconstruction order in space
R = Element residual of the discretization
R = Global residual of the discretization
[R] = Recovery matrix
R̃ = Interface recovery matrix
r = Filter order in the feedback fitlering procedure
rσ , R = Local and global lifting operators for BR2 method
SΣe = Smoothness sensor for element e
SI = Smoothness indicator in the feedback filtering procedure
SR = Smoothness requirement in the feedback filtering procedure
se = Element sensor for adaptivity
uh = Element solution for the conservative variable u
ur = Recovery solution for the conservative variable u
u = Vector of degrees of freedom
{W} = Boundary vector for Dirichlet BC’s with the ESR method
zh = Modified gradient for BR2 method
αiter = Filter intensity in the feedback filtering procedure
η0 = Penalty parameter for BR2 method
ηe = Residual-based error indicator
Φ = Element vasis function
Ψ = Recovery vasis function

153
154
Bibliography

[1] Ferrero A., Larocca F. Feedback filtering in discontinuous Galerkin methods

for Euler equations, Progress in Computational Fluid Dynamics, In press.
[2] Ferrero A., Larocca F., Puppo G. A robust and adaptive recovery-based dis-
continuous Galerkin method for the numerical solution of convection-diffusion
equations, Int. J. Numer. Meth. Fluids, 77:63-91, 2015.
[3] van Leer B., Nomura S. Discontinuous Galerkin for Diffusion, AIAA paper
2005-5108, 2005.
[4] Bassi F., Rebay S., Mariotti G., Pedinotti S., Savini M. A high-order accu-
rate discontinuous finite element method for inviscid and viscous turbomachin-
ery flows. In: Decuypere R, Dibelius G, editors. 2nd European Conference on
Turbomachinery Fluid Dynamics and Thermodynamics, Technologisch Institut,
Antwerpen, Belgium, 1997, 99-108.
[5] Löhner R. Improved error and work estimates for high-order elements, Int. J.
Numer. Meth. Fluids, 72:1207-1218, 2013
[6] Wang Z.J., Fidkowski K., Abgrall R., Bassi F., Caraeni D., Cary A., Decon-
inck H., Hartmann R., Hillewaert K., Huynh HT., Kroll N., May G., Persson
P.O., van Leer B., Visbal M. High-order CFD methods: current status and
perspective Int. J. Numer. Meth. Fluids, 72: 811-845, 2013.
[7] Liu X.-D., Osher S., Chan T. Weighted essentially nonoscillatory scheme, J.
Compu. Phys., 115:200-212, 1994
[8] Levy D., Puppo G., Russo G. Compact Central WENO Schemes for Multidi-
mensional Conservation Laws , SIAM J. Sci. Comp., 22:656-672, 2000.
[9] Cada M., Torrilhon M. Compact third-order limiter functins for finite volume
methods, J. Comput. Phys., 228:4118-4145, 2009.
[10] Shu C.W., TVB uniformly high-order schemes for conservation laws, Math.
Comput., 49:105-121, 1987.
[11] Second International Workshop on High-Order CFD Methods, Cologne, May
27-28 2013: http://www.dlr.de/as/hiocfd (Accessed January 2015)
[12] Wilcox D.C. Turbulence Modeling for CFD, 1th edition, DCW Industries, Inc.,
La Canada CA, 1993.

155
 Bibliography

[13] Wilcox D.C. Turbulence Modeling for CFD, 2nd edition, DCW Industries,
Inc., La Canada CA, 1998.
[14] Wilcox D.C. Turbulence Modeling for CFD, 3rd edition, DCW Industries, Inc.,
La Canada CA, 2006.
[15] Pacciani R., Marconcini M., Arnone A., Bertini F. An assessment of the lami-
nar kinetic energy concept for the prediction of high-lift, low-Reynolds number
cascade flows, Proc. IMechE Part A: J. of Power and Energy, Vol. 225 pp.
995-1003, 2011.
[16] Spalart P.R., Allmaras S.R. A One-Equation Turbulence Model for Aerody-
namic Flows, Recherche Aerospatiale, No. 1, pp. 5-21, 1994.
[17] NASA Turbulence Modeling Resource, http://turbmodels.larc.nasa.gov/spalart.html
(Accessed January 2015)
[18] Allmaras S.R., Johnson F.T.,and Spalart P.R. Modifications and Clarifications
for the Implementation of the Spalart-Allmaras Turbulence Model, Seventh
International Conference on Computational Fluid Dynamics (ICCFD7), Big
Island, Hawaii, 9-13 July 2012.
[19] Rumsey C.L., Spalart P.R. Turbulence Model Behavior in Low Reynolds Num-
ber Regions of Aerodynamic Flowfields, 38th AIAA Fluid Dynamics Conference
and Exhibit, June 23 - 26, 2008, Seattle, WA, AIAA Paper 2008-4403
[20] Bassi F., Crivellini A., Rebay S., Savini M. Discontinuous Galerkin solution of
the Reynolds-averaged Navier-Stokes and k-omega turbulence model equations,
Computers & Fluids, 34: 507-540, 2005.
[21] Langtry R.B., Menter F.R. Correlation-based transition modeling for unstruc-
tured parallelized computational fluid dynamics codes, AIAA Journal, 47(12),
pp. 2894-2906, 2009.
[22] Babajee J. Detailed numerical characterization of the separation-induced tran-
sition, in- cluding bursting, in a low-pressure turbine environment, Ecole Cen-
trale de Lyon; Institut von Karman de dynamique des uides (Rhode-Saint-
Genese, Belgique), PhD Thesis, 2013.
[23] Mayle R.E., Schulz A., The Path to Predicting Bypass Transition, ASME J.
Turbomach., 119(3), pp. 405-411, 1997.
[24] Walters D.K., Leylek J.H., A New Model for Boundary Layer Transition Using
a Single-Point RANS Approach, ASME J. Turbomach., 126(1), pp. 193-202,
2004,
[25] Lardeau S., Leschziner M.A., Li N., Modelling Bypass Transition With Low-
Reynolds-Number Non-Linear Eddy-Viscosity Closure, Flow,Turbul. Combust.,
73, pp. 49-76, 2004.
[26] Pacciani R., Marconcini M., Fadai-Ghotbi A., Lardeau S., Leschziner M.A.
Calculation of High-Lift Cascades in Low Pressure Turbine Conditions Using
a Three-Equation Model, Journal of Turbomachinery, Vol. 133, 2011.
[27] Menter F.R. Two-equation eddy-viscosity turbulence models for engineering
applications, AIAA Journal, 32(8):1598-605, 1994.

156
 Bibliography

[28] Hellsten A. On the solid-wall boundary condition of x in the k-x-type turbu-

lence models. Technical Report B-50, Helsinky University of Technology, Lab-
oratory of Aerodynamics, 1998.
[29] Bassi F., Botti L., Colombo N., Ghidoni A., Rebay S. Discontinuous Galerkin
for turbulent flows. In Adaptive high-order methods in computational fluid dy-
namics, Z. J. Wang, Ed., vol. 2 of Advances in Computational Fluid Dynamics.
World Scientific, Spring 2011.
[30] Hill T.T. and Reed W.H. Triangular mesh methods for neutron transport
equation, Tech. Rep. LA-UR-73-479, Los Alamos Scientific Laboratory, 1973.
[31] Bassi F., Rebay S. High-order accurate discontinuous finite element solution
of the 2D Euler equations, J. Comput. Phys., 138, 251-285, 1997.
[32] Bassi F., Botti L., Colombo A., Di Pietro D.A., Tesini P. On the flexibility
of agglomeration based physical space discontinuous Galerkin discretizations,
Journal of Computational Physics 2012; 231 : 45-65.
[33] Solin P., Segeth K., Dolezel I. Higher-Order Finite Element Methods, Chap-
man and Hall/CRC Press, 2003.
[34] Oñate E. Structural Analysis with the Finite Element Method: Linear Statics,
Volume 1: Basis and Solids. Springer, 2009.
[35] Osher S., Solomon F. Upwind schemes for hyperbolic systems of conservation
laws, Mathematics of Computation 1982, 38:339.
[36] Pandolfi M. A contribution to the numerical prediction of unsteady flows,
AIAA Journal, 22: 602-610, 1984
[37] Liou MS. A sequel to AUSM: AUSM+, Journal of Computational Physics,
Vol. 129, pp. 364, (1996).
[38] Liou MS. On a new class of flux splitting, Lecture Notes in Physics, Springer-
Verlag, 414, (1993).
[39] Roe P.L. Approximate solvers, parameter vectors and difference schemes, J.
Comput. Phys., 43, 357, (1981).
[40] Harten A., Lax P.D., van Leer B. On upstream differencing and Godunov-type
schemes for hyperbolic conservation laws, SIAM Reviews 25 (1): 35-61, (1983).
[41] Nishikawa H., Kitamura K. Very Simple, Carbuncle-Free, Boundary-Layer Re-
solving, Rotated-Hybrid Riemann Solvers, Journal of Computational Physics,
227, pp. 2560-2581, (2008).
[42] Pandolfi M. and D’Ambrosio D. Numerical instabilities in upwind methods:
analysis and cures for the "Carbuncle Phenomenon". In: Journal of Computa-
tional Physics, vol. 166, pp. 271-301, (2001).
[43] Lax P.D. Weak solutions of nonlinear hyperbolic equations and their numerical
computation, Comm. Pure Appl. Math., 7, 159-193, (1954).
[44] Leveque R.J., Finite Volume Methods for Hyperbolic Problems, Cambridge
Texts in Applied Mathematics, (2002).
[45] Wheatley V., Kumar H., Huguenot P., On the role of Riemann solvers in

157
 Bibliography

Discontinuous Galerkin methods for magnetohydrodynamics, Journal of Com-

putational Physics, Vol. 229, pp. 660-680, (2010).
[46] Arnold D.N., Brezzi F., Cockburn B., Marini D. Unified analysis of discon-
tinuous Galerkin methods for elliptic problems. SIAM Journal on Numerical
Analysis, 39, 5: 1749-1779, (2002).
[47] Lions J.L. Problemes aux limites non homogenes a donees irregulieres: une
methode d’approximation, Numerical Analysis of Partial Differential Equations
(C.I.M.E. 2 Ciclo, Ispra, 1967). Edizioni Cremonese: Rome, 1968; 283-292.
[48] Nitsche J.A. Uber ein Variationsprinzip zur Losung Dirichlet-Problemen bei
Verwendung von Teilraumen, die keinen Randbedingungen unteworfen sind.
Abh. Math. Sem. Univ. Hamburg 36: 9-5, (1971).
[49] Babuska I., Zlamal M. Nonconforming elements in the finite element method
with penalty, SIAM Journal on Numerical Analysis, 10: 863-875, (1973).
[50] Bassi F., Rebay S. 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).
[51] Cockburn B., Shu C.W. The local discontinuous Galerkin method for time-
dependent convection-diffusion systems, SIAM Journal on Numerical Analysis,
35: 2440-2463, (1998).
[52] Baumann C.E., Oden J.T. A discontinuous hp finite element method for
convection-diffusion problems. Comp. Meth. Appl. Mech. Eng., 175: 311-341,
(1991).
[53] Gassner G., Lorcher F., Munz C.D., A contribution to the construction of
diffusion fluxes for finite volume and discontinuous Galerkin schemes, Journal
of Computational Physics, 224: 1049-1063, (2007).
[54] Dumbser M., Zanotti O. Very High Order PNPM Schemes on Unstructured
Meshes for the Resistive Relativistic MHD Equations, Journal of Computa-
tional Physics, 228: 6991-7006, (2009).
[55] Luo H., Luo L., Nourgaliev R., Mousseau V.A. A Reconstructed Discontinuous
Galerkin Method for the Compressible Navier-Stokes Equations on Arbitrary
Grids, Journal of Computational Physics 229: 6961-6978, (2010).
[56] Dumbser M., Balsara D.S., Toro E.F., Munz C.D.. A unified framework for
the construction of one-step finite volume and discontinuous Galerkin schemes
on unstructured meshes. Journal of Computational Physics, 227: 8209-8253,
(2008).
[57] Huynh H.T. A reconstruction approach to high-order schemes including Dis-
continuous Galerkin for diffusion, AIAA Paper 2009-0403, (2009).
[58] Borrel M., Ryan J. The Elastoplast Discontinuous Galerkin (EDG) method
for the Navier-Stokes equations, Journal of Computational Physics, 231: 1-22,
(2012).
[59] French D.A., Galbraithy M.C., Osorio M. Error Analysis of a Modified Dis-
continuous Galerkin Recovery Scheme for Diffusion Problems, AIAA Paper

158
 Bibliography

2010-1071, (2010).
[60] Lo M., van Leer B. Analysis and Implementation of Recovery-Based Discon-
tinuous Galerkin for Diffusion, AIAA Paper 2009-3786, (2009).
[61] van Leer B., Nomura S., van Raalte M. A discontinuous Galerkin Method for
Diffusion Based on Recovery, AIAA Paper 2007-4083, (2007).
[62] van Leer B., Lo M. Unification of Discontinuous Galerkin Methods for Advec-
tion and Diffusion, AIAA Paper 2009-400, (2009).
[63] Nourgaliev R., Theofanous T., Park H., Mousseau V., Knoll D. Direct Nu-
merical simulation of interfacial flows, AIAA Paper 2008-1453, (2008).
[64] van Leer B., Lo M., Gitik R., Nomura S. Adaptive High-order Methods in
Computational Fluid Dynamics. World Scientific: Singapore; 185-201.
[65] Eaton J.W., Bateman D., Hauberg S. GNU Octave version 3.0.1 manual: a
high-level interactive language for numerical computations. CreateSpace Inde-
pendent Publishing Platform, (2009).
[66] Riviére B. Discontinuous Galerkin methods for solving elliptic and parabolic
equations : theory and implementation. SIAM: Philadelphia, 66, (2008).
[67] Bazilevs Y., Hughes T.J.R. Weak imposition of Dirichlet boundary conditions
in fluid mechanics, Computers & Fluids, 36: 12-26, (2007).
[68] First International Workshop on High-Order CFD Methods, Nashville, Jan-
uary 7-8, 2012: http://zjwang.com/hiocfd.html (Accessed January 2015)
[69] Third International Workshop on High-Order CFD Methods, Orlando, Jan-
uary 3-4, 2015: https://www.grc.nasa.gov/hiocfd/ (Accessed January 2015)
[70] Quarteroni A., Sacco R., Salieri F. Matematica numerica, Springer, Milan,
(2008).
[71] Schwartzkopff T., Munz C.D., Toro E.F. ADER: High-order approach for
linear hyperbolic systems in 2D. Journal of Scientific Computing. Vol. 17, pages
231-240, (2002).
[72] Nigro A., Ghidoni A., Rebay S., Bassi F. Modified extended BDF scheme
for the discontinuous Galerkin solution of unsteady compressible flows, Int. J.
Numer. Meth. Fluids, 76:549-574, (2014).
[73] Cockburn B. Advanced Numerical Approximation of Nonlinear Hyperbolic
Equations, An Introduction to the Discontinuous Galerkin spatial discretiza-
tion Galerkin Method for Convection-dominated Problems, Lecture Notes in
Mathematics, Springer, pp. 151-268, (1998).
[74] Gottlieb S., Shu C-W. Total variation diminishing Runge-Kutta schemes,
Math. Comput. 67 (221), 73-85, (1998).
[75] Kaw A., Kalu E., Numerical Methods with Applications (1st ed.), (2008)
[76] Carpenter M.H., Kennedy C. Fourth-order 2N-storage Runge-Kutta schemes,
NASA Report TM 109112, NASA Langley Research Center, (1994).
[77] Spiteri R., Ruuth S., A new class of optimal high-order strong stability pre-
serving time discretization methods, SIAM Journal on Numerical Analysis, Vol.
40, pp. 469-491, (2002).

159
 Bibliography

[78] CSR sparse matrix format. http://netlib.org./linalg/html_templates/node91.html

(Accessed December 2014)
[79] Brown P.N., Saad Y. Hybrid Krylov methods for nonlinear systems od equa-
tions, SIAM J. Sci. and Stat. Comput., 11(3):450-481, (1990).
[80] Saad Y., Schultz M.H. GMRES: A generalized minimal residual algorithm for
solving nonsymmetric linear systems, SIAM J. Sci. Stat. Comput., 7:856-869,
(1986).
[81] Lukarski D., Trost N. PARALUTION Project, http://www.paralution.com/
(Accessed December 2014)
[82] Bassi F., Botti L., Colombo A., Crivellini A., Franchina N., Ghidoni A., Rebay
S. (2010) Very high-order accurate discontinuous Galerkin computation of tran-
sonic turbulent flows on aeronautical configurations, in ADIGMA - A European
Initiative on the Development of Adaptive Higher-Order Variational Methods
for Aerospace Applications, N.Kroll, H. Bieler, H. Deconinck, V. Couaillier,
H. van der Ven, and K. Sorensen, Eds., Vol. 113 of Notes of Numerical Fluid
Mechanics and Multidisciplinary Design, Springer Berlin, pp. 25-38.
[83] Mulder W.A., van Leer B. Experiments with implicit upwind methods for the
Euler equations, Journal of Computational Physics Volume 59, Issue 2, Pages
232-246, (1985).
[84] Bücker H.M., Pollul B., Rasch A. On CFL evolution strategies for implicit
upwind methods in linearized Euler equations, International Journal for Nu-
merical Methods in Fluids, Volume 59, Issue 1, pages 1-18, (2009).
[85] Taube A., Munz C.-D., Synthesis report on shock capturing strategies, in
ADIGMA - A European Initiative on the Development of Adaptive Higher-
Order Variational Methods for Aerospace Applications, N.Kroll, H. Bieler, H.
Deconinck, V. Couaillier, H. van der Ven, and K. Sorensen, Eds., Vol. 113 of
Notes of Numerical Fluid Mechanics and Multidisciplinary Design, Springer
Berlin, pp. 195-207, (2010).
[86] Persson P.-O., Peraire J. Sub-Cell Shock Capturing for Discontinuous Galerkin
Methods, Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A.
[87] Toro, E.F. Riemann Solvers and Numerical Methods for Fluid Dynamics,
Springer, Berlin, (2009).
[88] Shapiro, A. The Dynamics and Thermodynamics of Compressible Fluid Flow,
The Ronald Press Company, New York, (1953).
[89] Nguyen N.C., Peraire J. An Adaptive Shock-Capturing HDG Method for Com-
pressible Flows, AIAA Paper 2011-3060, (2011).
[90] Hartmann R. Higher-order and adaptive discontinuous Galerkin methods with
shock-capturing applied to transonic turbulent delta wing flow, Int. J. Numer.
Meth. Fluids 72, 8, pages 883-894, (2013).
[91] Cockburn B., Hou S., Shu C.W. TVB Runge-Kutta local projection discon-
tinuous Galerkin finite element method for conservation laws IV: The multidi-
mensional case, Math. Comput., 54:545-581, (1990).

160
 Bibliography

[92] Biswas R., Devine K., Flaherty J.E. Parallel adaptive finite element methods
for conservation laws, Applied Numerical Mathematics 14: 255-284, (1994).
[93] Li W., Ren Y-X. The multi-dimensional limiters for discontinuous Galerkin
method on unstructure grids, Seventh International Conference on Computa-
tional Fluid Dynamics, Hawaii, July 9-13, 2012.
[94] Hesthaven J.S., Kirby R.M. Filtering in Legendre spectral methods, Math.
Comput. 77(263):1425-1452, (2008).
[95] Allaneau Y., Jameson A. Connections between the filtered discontinuous
Galerkin method and the flux reconstruction approach to high-order discretiza-
tions, Comput. Methods Appl. Mech. Engrg. 200: 3628-3636, (2011).
[96] Gottlieb D., Hesthaven J.S. Spectral Methods for Hyperbolic Problems, J.
Comput. Appl. Math., 128, 83-131 (2001).
[97] Hesthaven J.S., Warburton T. Nodal Discontinuous Galerkin Methods, Texts
in Applied Mathematics, Springer (2008).
[98] Boyd J.P. The Erfc-Log Filter and the Asymptotics of the Euler and Vandeven
Sequence Accelerations, Houston Journal of Mathematics, (1996).
[99] Ueckermann M.P., Lermusiaux P.F.J. High-order schemes for 2D unsteady
biogeochemical ocean models, Ocean Dynamics, 60, 6, pp 1415-1445, (2010).
[100] Kanevsky A., Carpenter M.H., Hesthaven J.S. Idempotent filtering in spec-
tral and spectral element methods, J. Comput. Phys., 220: 41-58, (2006).
[101] Chiocchia G. Exact solutions to transonic and supersonic flows, Technical
Report AR-211, AGARD, (1985).
[102] Barter G.E. Shock Capturing with PDE-Based Artificial Viscosity for an
Adaptive, Higher-Order Discontinuous Galerkin Finite Element Method, PhD
Thesis, Massachusetts Institute of Technology, Massachusetts, (2008).
[103] Vassberg J.C., Jameson A. In Pursuit of Grid Convergence for Two-
Dimensional Euler Solutions, Journal of Aircraft, Vol. 47, pp. 1152-1166,
(2010).
[104] Kroll N., Bieler H., Deconinck H., Couaillier V., van der Ven H., Sorensen
K. ADIGMA - a European initiative on the development of adaptive higher-
order variational methods for aerospace applications. Notes on Numerical Fluid
Mechanics and Multidisciplinary Design, vol. 113. Springer, 2010.
[105] Hartmann R., Held J., Leicht T. Adjoint-based error estimation and adaptive
mesh refinement for the RANS and k-ω turbulence model equations, J. Comput.
Phys., Vol. 230, No. 11, pp. 4268-4284, (2011).
[106] Leicht T., Hartmann R. Error estimation and anisotropic mesh refinement
for 3d laminar aerodynamic flow simulations, Journal of Computational Physics
229: 7344-7360, (2010).
[107] Norberg C. Fluctuating lift on a circular cylinder: review and new measure-
ments, Journal of Fluids and Structures, 17: 57-96, (2003).
[108] Tritton DJ. Experiments on the flow past a circular cylinder at low reynolds
number, Journal of Fluid Mechanics, 6: 547-567, (1959).

161
 Bibliography

[109] von Wieselsberger C. Neuere Festellungen über die Gesetze des

Flüssigkeits un Luftwiderstands, Phys. Z. 22: 321-328, (1921).
[110] Williamson C.H.K. Oblique and parallel modes of vortex shedding in the
wake of a circular cylinder at low reynolds number, Journal of Fluid Mechanics,
206 : 579-627, (1989).
[111] Bassi F., Colombo A., Franchina N., Ghidoni A., Rebay S. High-order ac-
curate p-multigrid discontinuous Galerkin solution of the RANS and k − ω
turbulence model equations, V European Conference on Computational Fluid
Dynamics ECCOMAS CFD 2010, J.C.F. Pereira and A. Sequeira (Eds), Lis-
bon, Portugal, 14-17 June 2010.
[112] Luo H., Baum J.D., Löhner R. A p-multigrid discontinuous Galerkin method
for the Euler equations on unstructured grids, Journal of Computational
Physics 211(2):767 (2006)
[113] Fidkowski K.J., Oliver T.A., Lu J., Darmofal D.L. p-Multigrid solution of
high-order discontinuous Galerkin discretizations of the compressible Navier-
Stokes equations, Journal of Computational Physics, 207: 92-113, (2005).
[114] Masatsuka K. Mesh generator for Ringleb flow, http://www.cfdbooks.com
(Accessed January 2015).
[115] Wang Z.J., Liu Y. Extension of the spectral volume method to high-order
boundary representation, J. Comput. Phys. 211: 154-178, (2006).
[116] Sieverding C.H. Experimantal data on two transonic turbine blade sections
anc comparisons with various theoretical methods, VKI Report No. LS59, von
Karman Institute, Belgium, (1973).
[117] Kiock R., Lethaus F., Baines N.C., Sieverding C.F. The transonic flow
through a plane turbine cascade as measured in four European wind tunnels,
ASME J. Eng. Gas Turbines Power, 108, 277-285, (1986).
[118] Hourmouziadis J. Aerodynamic Design of Low Pressure Turbines, AGARD
Lecture Series, 167, (1989).
[119] Michalek J., Monaldi M., Arts T. Aerodynamic performance of a very high
lift low pressure turbine airfoil (T106C) at low Reynolds number and high Mach
numer with effect of free stream turbulence intensity, ASME Paper GT2010-
22884, Glasgow, UK, 14-18 June 2010.
[120] Geuzaine C., Remacle J.F. Gmsh: a three-dimensional finite element mesh
generator with built-in pre- and post-processing facilities. International Journal
for Numerical Methods in Engineering, 79: 1309-1331, (2009).

162