FACULTY OF ENGINEERING

Department of Mechanical Engineering

Development of an efficient
Navier-Stokes/LES solver on
unstructured grids for
high-order accurate schemes
Thesis submitted in fulfillment of the requirements for the
award of the degree of Doctor in de Ingenieurswetenschappen
(Doctor in Engineering) by

Matteo Parsani
November 2010
Advisor:

Prof. Dr. Ir. Chris Lacor

Abstract
Researchers are attempting to tackle problems which were considered too
ambitious just a few years ago. Multidisciplinary analysis and design
(MAD), computational aeroacoustics (CAA), large eddy simulation (LES)
and direct numerical simulation (DNS) of turbulence are examples of what
is being attempted today. Improvements to the efficiency of these solutions
are necessary due to the complexity of such problems.
In the field of the computational fluid dynamics (CFD), the use of higherorder accurate spatial discretizations for unstructured grids offers a possible avenue for improving the predictive simulation capabilities for many
modern applications. This is due to the fact that higher-order methods exhibit a faster asymptotic convergence rate in the discretization error than
lower (second)-order accurate finite volume (FV) and finite difference (FD)
methods. The expectation is that an efficient higher-order discretization
may provide an alternate path for achieving high accuracy in a flow with
a wide disparity of length scales at reduced cost, by avoiding the use of
excessive grid resolution.
Although the formulation of compact discretization strategies for higherorder methods such as discontinuous Galerkin (DG), spectral volume (SV)
and spectral difference (SD) methods are now fairly well understood, the
development of techniques for efficiently solving the discrete equations
arising from these methods has generally been lagging. This is partly due
to the complex structure of the discrete equations originating from fairly
sophisticated discretization strategies, as well as the current application
of higher-order methods to problems where simple explicit time-stepping
schemes are thought to be adequate solution mechanisms such as acoustic phenomena. Therefore, the development of optimal, or near optimal
solution strategies for higher-order discretizations, including steady-state
solutions methodologies, and implicit time integration strategies, remains
i

then one of the key determining factors in devising higher-order methods.

The main goal of the present PhD research is to develop an efficient NavierStokes/LES solver on unstructured grids for high-order accurate spatial
discretizations, and build up the necessary know-how to make a high-order
accurate solver for industrial purposes. In order to achieve that, the present
research has been carried out in two parts. In the first part, two implicit time integration schemes, namely backward Euler (BE) scheme and
second-order backward difference formula (BDF2), are coupled with a nonlinear lower-upper Gauss-Seidel (LU-SGS) algorithm for efficiently solving
the discrete equations arising from the spatial discretization with a SV or a
SD method. The non-linear LU-SGS algorithm with the BE scheme is evaluated both with analysis and computation for both spatial operators and
steady flow problems. In addition, the capabilities and the advantages of
the SD method in combination with the implicit time integration/algebraic
solver technique is demonstrated by solving several unsteady reference
test cases. Good agreement between the present results and reference solutions is found, demonstrating the potential benefits of high-order accurate spatial methods.
In the second part, the SD method coupled with a LES approach is investigated. The wall-adapted local eddy-viscosity (WALE) model is chosen
as a subgrid-scale model and a new idea is presented for the definition of
the filter width in the closure of the LES equations. The method is successfully applied to compute two- and three-dimensional turbulent cases. Good
agreement between the present numerical results and reference solutions
is observed, showing the capability and the quality of the new coupling
approach.

ii

Jury members
President

Prof. Hugo SOL

Vrije Universiteit Brussel

Vice-president

Prof. Rik PINTELON

Vrije Universiteit Brussel

Secretary

Prof. Patrick GUILLAUME

Vrije Universiteit Brussel

Internal members

Prof. Gert DESMET

Vrije Universiteit Brussel

External members

Prof. Wim DESMET

Katholieke Universiteit Leuven

Prof. Eli TURKEL

Tel Aviv University

Advisor

Prof. Chris LACOR

Vrije Universiteit Brussel

vi

Contents
1 Introduction

2 Literature survey
2.1 Spectral volume method . . . . . . .
2.2 Spectral difference method . . . . . .
2.3 Time integration schemes . . . . . .
2.4 Geometric and p-multigrid methods

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

7
8
9
10
11

3 Governing equations
3.1 Compressible Navier-Stokes equations . . . . . . . .
3.1.1 Newtonian fluid . . . . . . . . . . . . . . . . .
3.1.2 Thermal conductivity . . . . . . . . . . . . . .
3.1.3 Thermodynamic properties: ideal gas model
3.1.4 Formulation in Cartesian space . . . . . . . .
3.1.5 Dimensionless numbers . . . . . . . . . . . .
3.2 Large eddy simulation . . . . . . . . . . . . . . . . .
3.2.1 Formulation in Cartesian space . . . . . . . .
3.2.2 The wall-adapted local eddy-viscosity model
3.3 Boundary conditions . . . . . . . . . . . . . . . . . .
3.3.1 Far field . . . . . . . . . . . . . . . . . . . . . .
3.3.2 Inlet mass density and velocity . . . . . . . .
3.3.3 Pressure outlet . . . . . . . . . . . . . . . . .
3.3.4 Solid wall . . . . . . . . . . . . . . . . . . . . .
3.4 Aerodynamic coefficients . . . . . . . . . . . . . . . .
3.5 Linear convection equation . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

13
14
16
17
17
20
23
24
26
28
30
30
32
33
33
36
37

4 Spatial discretization
4.1 Spectral volume method . . . . . . . . . . . . . . . . . . . . .
4.1.1 Discretization of convective term . . . . . . . . . . . .
4.1.2 SV basis polynomials . . . . . . . . . . . . . . . . . . .

39
40
40
43

vii

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

4.1.3 Discretization of diffusive terms . . . . . . .

4.1.4 Spectral volume partition . . . . . . . . . . .
4.2 Spectral difference method . . . . . . . . . . . . . . .
4.2.1 Discretization of convective term . . . . . . .
4.2.2 SD basis polynomials . . . . . . . . . . . . . .
4.2.3 Discretization of diffusive terms . . . . . . .
4.2.4 Component-wise flux point distribution . . .
4.2.5 Solution and flux points distribution . . . . .
4.2.6 Grid filter width for the subgrid-scale model
4.3 Concluding remarks . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

45
47
52
52
55
56
58
59
62
64

5 Time discretization
5.1 Backward Euler scheme . . . . . . . . . . . . . . . . . . . . .
5.2 Second-order backward difference formula . . . . . . . . . .
5.3 Time step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67
70
74
75

6 Analysis of the non-linear LU-SGS algorithm

6.1 Summary of the methodology . . . . . . . . .
6.2 SV method for triangular cells . . . . . . . . .
6.2.1 Second-order SV method . . . . . . . .
6.2.2 Third-order SV method . . . . . . . . .
6.2.3 Fourth-order SV method . . . . . . . .
6.3 SD method for quadrilateral cells . . . . . . .
6.3.1 Second-order SD method . . . . . . . .
6.3.2 Third-order SD method . . . . . . . . .
6.3.3 Fourth-order SD method . . . . . . . .
6.4 Remarks . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

77
77
82
83
92
94
95
95
96
98
99

7 Application I: spectral volume method

7.1 Two-dimensional laminar steady flow simulations . .
7.1.1 Flow over a circular cylinder . . . . . . . . . . .
7.1.2 Flow over a NACA0012 airfoil . . . . . . . . . .
7.1.3 Flow in a channel with a backward-facing step

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

101
102
104
110
116

8 Application II: spectral difference method

8.1 Steady laminar flow simulations . . . . . . . .
8.1.1 Flow over a NACA0012 airfoil . . . . . .
8.1.2 Flow through a 90 bending square duct
8.2 Unsteady laminar flow simulations . . . . . . .
8.2.1 Flow over an open cavity . . . . . . . . .
8.2.2 Flow past a square cylinder . . . . . . .
8.2.3 Flow past a circular cylinder . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

121
122
122
127
133
133
137
141

viii

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

8.3 Large eddy simulations . . . . . . . . . . . . . . . . . . .

8.3.1 Flow around a NACA0012 airfoil . . . . . . . . .
8.3.2 Flow around a square cylinder at Re = 104 . . .
8.3.3 Flow around a square cylinder at Re = 2.2 104
8.3.4 Flow in a muffler . . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

145
146
149
158
162

.
.
.
.

.
.
.
.

.
.
.
.

169
170
173
173
175

A Time integration methods for space-discretized equations

A.1 Stability of spatial discretizations . . . . . . . . . . . . . . . .
A.2 Stability of time discretizations . . . . . . . . . . . . . . . . .
A.2.1 Forward Euler scheme . . . . . . . . . . . . . . . . . .
A.2.2 Backward Euler scheme . . . . . . . . . . . . . . . . .
A.2.3 Second-order backward difference formula . . . . . .
A.2.4 Higher-order backward difference formulae . . . . . .
A.3 Method for the analysis of the non-linear LU-SGS algorithm
A.3.1 Direct inversion method . . . . . . . . . . . . . . . . .
A.3.2 Non-linear LU-SGS algorithm . . . . . . . . . . . . . .
A.3.3 Eigenvalue spectrum of the amplification matrix . . .

177
178
182
183
185
186
188
190
191
191
195

9 Conclusions and future directions

9.1 Achievements . . . . . . . . . . . . . . . . . . . . . . .
9.2 Future work . . . . . . . . . . . . . . . . . . . . . . . .
9.2.1 Compact high-order accurate spatial methods
9.2.2 Time integration/solution iterative approaches

.
.
.
.

B p-Multigrid
197
B.1 Full approximation scheme . . . . . . . . . . . . . . . . . . . 198
B.2 Transfer operators . . . . . . . . . . . . . . . . . . . . . . . . 199
C Newton-Raphson GMRES solver
201
C.1 Newton-Raphson algorithm . . . . . . . . . . . . . . . . . . . 202
C.2 GMRES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
D ESDIRK schemes

205

ix

Nomenclature
BDF2
BDF3
BDF4
BE
BR2
CAA
CEM
CFD
CFL
CV
DES
DG
DNS
DOF
E-RK
ESDIRK

Second-order backward difference formula

Third-order backward difference formula
Fourth-order backward difference formula
Backward Euler scheme
Second approach of Bassi and Rebay
computational aeroacoustics
Computational electromagnetics
Computational fluid dynamics
Courant-Friedrichs-Lewy
Control volume
Detached eddy simulation
Discontinuous Galerkin
Direct numerical simulation
Degree of freedom
Explicit Runge-Kutta scheme
Explicit-first-stage, single-diagonal-coefficient, diagonally-implicit
Runge-Kutta scheme
FE
Forward Euler scheme
FMG
Full multigrid
FV
Finite Volume
GMRES Generalized minimal residual method
I-RK
Implicit Runge-Kutta scheme
LDG
Local discontinuous Galerkin
LES
Large eddy simulation
LSD
Local spectral difference
LSV
Local spectral volume
LU-SGS Lower-upper symmetric Gauss-Seidel
N-S
Navier-Stokes
RANS
Reynolds averaged Navier-Stokes
RK
Runge-Kutta scheme
xi

SD
SD-LES
SGS
SSP
SV
TVD
WALE

Spectral difference
SD method coupled with LES approach
Symmetric Gauss-Seidel
Strong stability preserving
Spectral volume
Total variation diminishing
Wall-adapted local eddy-viscosity

Subscripts
c
|
cc
gho
I
int
L
nb
R
R

Reference quantity
Boundary quantity
Current cell
Ghost value
Imaginary part of a complex number
Internal value
Left cell
Neighbouring cells
Real part of a complex number
Right cell

Symbols

~1k
~1n~

3
4

3
4

m
W
0

Wm

~
1 , 2 , 3
~
j
B
g
~ L(R)
W

4
jm

Cross section
Unit vector in direction of the wave vector
Unit normal to a face in a cell-mapped coordinate system
Damping factor in averaging operator for diff. term treatment
DOF of a 3rd-order SV partition/SD flux point distribution
DOF of a 4th-order SV partition
Bias in the averaging operator for the diffusive term treatment
DOF of a 3rd-order SV partition
DOF of a 4rd-order SV partition
in terms of eigenvectors V
Coefficients of expansion of W
m
Coefficients of initial solution expansion in terms of eigenvectors
Vm
T
Vector of general Cartesian coordinates [1 , 2 , 3 ] , m
General coordinates in Cartesian space, m
Volume of CV j in the mapped coordinate system ~
~ 1, m
Length of B
Polynomial used in the BR2 lifting operator definition for SD
Grid filter width for LES
DOF of a 4rd-order SV partition
Kronecker delta function
xii

3
4
~

~ ~

T
T+1,0
T1,0
T0,+1
T0,1
T0,0
T

Gd
GQD
f,1
GTf,1R
GQD
f,m
R
GTf,m

GQD
SGS,1
R
GTSGS,1
GQD
SGS,m
R
GTSGS,m

m
m

mach
~
1 , 2 , 3
~lf
~s
j

DOF of a 3rd-order SV partition

DOF of a 4rd-order SV partition
Lifting operator used with the BR2 diff. term treatment
Divergence operator in mapped coordinate system
Averaged gradient approximations on a face
Riemann flux upwinding parameter
Polynomial approximation of the conserved variable gradients
2D advection speed vector orientation angle
Courant-Friedrichs-Lewy (CFL) number, dimensionless
Spatial discretization matrix for linear advection equation
Spatial discretization matrix for linear advection equation
Spatial discretization matrix for linear advection equation
Spatial discretization matrix for linear advection equation
Spatial discretization matrix for linear advection equation
Spatial discretization matrix for linear advection equation
Spatial discretization matrix for linear advection equation
Parameter of the BDF2 with variable time step
Amplification matrix for direct inversion method
Amplification matrix for 1-st forward GS sweep for quadrilateral
cells
Amplification matrix for 1-st forward GS sweep for triangular
cells
Amplification matrix for m-th forward GS sweep for quadrilateral cells
Amplification matrix for m-th forward GS sweep for triangular
cells
Amplification matrix for 1-st SGS sweep for quadrilateral cells
Amplification matrix for 1-st SGS sweep for triangular cells
Amplification matrix for m-th SGS sweep for quadrilateral cells
Amplification matrix for m-th SGS sweep for triangular cells
2D wave vector orientation angle
Eigenvalue of matrix T
Eigenvalue of matrix T
Small constant used for the calculation of Jacobians
Machine zero
T
Vector of mapped coordinates [1 , 2 , 3 ]
Mapped coordinates
Mapped coordinates of flux point l
Mapped coordinates of solution point j
Module of advection speed, ms1
xiii

CD
CL
CP
Cf
~1
B
~ 1
B
~2
B
~
B
2

w
vars
m
~
|m|
~
~u
|~u|
dim
M
Pr
P rt
Re

~
Fi
~

fC,i

Drag coefficient, dimensionless

Lift coefficient, dimensionless
Pressure coefficient, dimensionless
Skin friction coefficient, dimensionless
Vector defining a 2D generating pattern, m
Dimensionless vector defining a 2D generating pattern
Vector defining a 2D generating pattern, m
Dimensionless vector defining a 2D generating pattern
Mass density, kg m3
Conserved variables
Number of the scalar conserved variables, vars = 4 in 2D, vars =
5 in 3D
Momentum, kg m2 s1
Module of the momentum, kg m2 s1
Velocity vector, m s1
Module of the velocity vector, m s1
Dimensions of the physical space, dim = 2 in 2D, dim = 3 in 3D
Mach number
Prandtl number, dimensionless
Turbulent Prandtl number, dimensionless
Reynolds number
Computational domain
Boundary of the computational domain
~
Polynomial approximation of fi
1 -components of convective flux vectors in mapped coordinate
system

~f
C,i
i,l

()
()
q
~
~f
~fC
~fD
f~
f~m
~
f~et
R
~
F ~
1n

Convective flux vector projected in mapped coordinate system

Averaged filter width of the SD face flux point l in cell i, for LES
Favre filtered quantity
Spatially filtered quantity
Heat flux vector, P a m s1
Flux vector
Convective flux vector
Diffusive flux vector
Mass density flux vector, kg m2 s1
Momentum flux tensor, P a
Total energy flux vector, P a m s1
Riemann flux through a face with unit normal ~1n
Amplification factor of a full discretization in space and time

xiv

Gi

Polynomial approximation of gi

gC,i

2 -components of convective flux vectors in mapped coordinate

system
Characteristic gas constant. It is the specific heat capacities
ratio, for the ideal gas model
Specific gas constant defined by the ratio of the universal gas
constant and the molar mass of the gas
Local length scale associated to a face, m

R
hf
~

Hi

Polynomial approximation of hi

hC,i

3 -components of convective flux vectors in mapped coordinate

system
Imaginary unit number, square root of 1
Filter width of the SD internal flux point l in cell i, for LES
Dimensionless wave number
Wave number, m1
Basis polynomial associated to solution point j
Basis polynomial associated to flux point l
SV basis polynomial with index j in cell i
Number of flux points in a cell
Number of solution variables in a cell
Number of solution variables in a generating pattern
Number of stage for a Runge-Kutta scheme
Nabla/del operator, m1
Partial derivative with respect to time, s1
Position vector in a general coordinate system, m
Reference surface of an object, m2
Set of complex number
Set of real number
Strouhal number, dimensionless

I
i,l
K
k
Lsj
Lfl
i,j
L
Nf
Ns
N s,GP
NRK
~

~r
Ac
C
R
St
~~
G
~~
~~
S
~
I~

k sgs
~~ sgs
~qsgs

Stress tensor, P a
Deviatoric stress tensor. It becomes the viscous stress tensor
when is set equal to the pressure P , P a
Symmetric velocity gradient tensor, s1
Isotropic stress tensor characterized by the scalar
~~
Scalar number characterizing the isotropic stress tensor I.
It is
set equal to the pressure P , P a
Subgrid-scale kinetic energy, J kg 1
Subgrid-scale stress tensor, P a
Subgrid-scale heat flux vector, P a m s1
xv

S
E

t
Nmol
P
cP
cv
s
e
v
c
T
V
t
hi
et
Vm

W
~
wi
Wi,j
i
W
 
Wi ~

i,j
W
z

Conductivity coefficient, J m1 K 1
Dilatation viscosity coefficient, kg m1 s1
Dynamic viscosity coefficient, kg m1 s1
Electro-chemical potential, J mol1
Entropy, J K 1
Internal energy, J
Kinematic viscosity coefficient, m2 s1
Turbulent kinematic viscosity coefficient (eddy-viscosity), m2 s1
Number of moles in a single-component fluid
Pressure, P a
Specific heat capacity at constant pressure, J kg 1 K 1
Specific heat capacity at constant volume, J kg 1 K 1
Specific entropy, J kg 1 K 1
Specific internal energy, J kg 1
Specific volume, m3 kg 1
Speed of sound, m s1
Temperature, K
Volume , m3
Time, s
Time averaged quantity
Specific total energy, J kg 1
Eigenvector of matrix T
Complex amplitude of the numerical spatial Fourier wave
Averaged conserved variables, used for diffusive term treatment
Conserved variables in mapped coordinate system in cell i
SD solution variable, solution at solution point j in cell i
Polynomial of degree p + 1, for gradient computation with SD
Solution polynomial in cell i
SV solution variable, averaged solution in CV j in cell i
A complex number

Superscripts

Latest available solution

0
Initial quantity
LES
Quantity related to large eddy simulation (LES)
sgs
Subgrid-scale quantity
T
Transpose

xvi

Chapter 1

Introduction
Basic forms of computational fluid dynamics (CFD) were established in
1960 with the research at the Courant Institute on hyperbolic systems of
conservation laws which led to the development of the first second-order
accurate dissipative methods by Lax and Wendroff. This class of methods
forms the basis of the efficient explicit method developed in 1969 for the
Navier-Stokes equations by MacCormack. At the Douglas Aircraft Company, the aerodynamic research group, led by A. M. O. Smith, developed
the first panel method for three-dimensional, linear, potential flows, and
the first implementation of Kellers box method for turbulent boundary
layer flows. CFD began developing quickly in the 1980s as the advances
of computing and processor technology gave scientists the means to solve
complex fluid dynamics problems. This has enabled progress at many
fronts, including numerical algorithms for the Euler and Navier-Stokes
equations, grid generation and adaptation, turbulence modeling, flow visualization as well as the basic understanding of complex phenomena and
better design quality of technological applications. Nowadays, CFD has
reached a high level of sophistication, and at the same time its range of
applications is broadening, including diverse topics such as aerodynamics, aeroacoustics, combustion, global atmospheric modeling, oceanography etc. In some circumstances and especially in the early stages of the
design process, experiments might be too expensive or even impossible to
perform. Therefore, modeling is the only reasonable way to get answers
and to study a range of parameters for optimal design. In addition, CFD
is used routinely to complement wind tunnel tests for the final design of
compressors, turbines, pumps as well as complete aircrafts.

CHAPTER 1. INTRODUCTION
In the field of CFD, spatially low-order (first- and second-order) numerical
methods, typically based on finite volume (FV) or finite difference (FD) discretizations, are less accurate than their high-order counterpart, but they
are generally more robust and reliable; as a result, they are routinely employed in commercial flow solver software packages. In fact, for many industrial application problems, these methods are often good choices, when
considering the balance between computational speed, simplicity of coding,
and resolution required. This statement is especially valid if sought after
solutions are piecewise simple (almost linear) with several isolated discontinuities in between, for example, the solution of most Riemann problems. However, in certain modern applications with complicated geometries and complex physics, where the accurate resolution of small scales
is required (e.g. direct numerical simulation (DNS), large eddy simulation
(LES), computational aeroacoustics (CAA), computational electromagnetic
(CEM), turbulent combustion etc.), the solution structures are so complicated and the time of evolution of these structures is so long that it is
impractical to use low-order methods to obtain an acceptable resolution.
Low-order methods are in fact too dissipative to resolve accurately rich
structures in the smooth part of the solutions. Therefore, it should be decided, based on the problem at hand, whether to use a high-order scheme
(order of accuracy >2) or to find a suitable first- or second-order scheme.
In addition, since CFD is increasingly used as an industrial design and
analysis tool, high-order accuracy must be achieved on unstructured grids
which are required for efficient meshing of complex geometries. With classical spatial discretizations, such as FV and FD methods, high-order accuracy can be obtained theoretically for an arbitrary unstructured grid by using high-order polynomial data reconstructions. However, higher than linear reconstructions are rarely used for three dimensions in practice. This
is mainly because of the difficulty in finding valid (non-singular) stencils
and the enormous memory required to store the coefficients used in the reconstruction. For each cell, the stencil is unique for an unstructured grid
and its size increases non-linearly with the order of accuracy. A data reconstruction must be performed at each iteration for each element. This
reconstruction step is the most memory consuming in higher than secondorder schemes. In addition, the stencil size complicates numerical formulations near boundaries, increases the matrix bandwidth and can increase
the communication time required by algorithms for parallel architectures.
All the needs mentioned above have been the driving force for the development of a new class of spatially high-order schemes for unstructured grids,
2

e.g. the discontinuous Galerkin (DG) method, the spectral volume (SV)
method and the spectral difference (SD) method. Such methods approximate the solution by a polynomial of a certain degree in each cell by providing sufficient pieces of independent information, i.e. degrees of freedom
(DOFs), in the cell itself. These schemes use piecewise continuous functions as the solution approximation space. They have a compact stencil,
since only the data local to the cell and the data of its immediate neighboring cells are required for the evaluation of the fluxes. Consequently, they
are easily parallelizable. In this thesis, two spatially compact methods,
namely the SV and SD methods, are used for the spatial discretization of
the fluid dynamics equations.
When high-order schemes are combined with classical solution methods,
such as explicit Runge-Kutta (E-RK) solvers, they suffer from a restrictive Courant-Friedrichs-Lewy condition or CFL condition and hence a relatively slow convergence rate. In addition, the solver should also be able
to deal with the geometrical stiffness imposed by the Navier-Stokes grids
where high-aspect ratios occur near walls. In the case of compressible
solvers there is an additional stiffness when solving for low speed flows
caused by the disparate eigenvalues of the system. Therefore, the development of optimal, or near optimal solution strategies for higher-order discretizations, including steady-state solutions methodologies, and implicit
time integration strategies, remains one of the key determining factors
in devising higher-order methods which are not just competitive but superior to lower-order methods in overall accuracy and efficiency. Implicit
time-integration schemes can be used to deal with these problems. They
can advance the solution with significantly larger time steps compared to
explicit methods. Two implicit time marching schemes, namely the backward Euler (BE) scheme and second-order backward difference formula
(BDF2) with variable time step, are used here for the time discretization.
The BE scheme, which is first-order accurate in time, is used to advance
in (pseudo) time steady flow problems, whereas the BDF2 is used for unsteady flow simulations.
Implicit temporal schemes imply the solution of non-linear algebraic systems. Therefore, if the algorithm for such systems is not efficient, implicit schemes might be more expensive than explicit ones. In this work,
an efficient algebraic solver, namely the non-linear lower-upper symmetric
Gauss-Seidel algorithm (LU-SGS), is used to solve the non-linear algebraic
systems associated to the implicit time discretizations. This algorithm was
proposed a few years ago in combination with the BE scheme for the SD
3

CHAPTER 1. INTRODUCTION
method. It solves the non-linear algebraic system through multiple cellwise symmetric forward and backward Gauss-Seidel sweeps. Because of
the Gauss-Seidel nature of this algorithm, where the latest available solution in the neighboring cells is used to update the solution in a cell, information travels much faster across the domain than with a traditional
explicit solver, where only the solution at the previous time step or stage
is used to update the solution in a cell. Besides its cell-wise implicitness,
the Gauss-Seidel nature is one of the reasons why the non-linear LU-SGS
method is much more efficient than an explicit solver. The coupling of the
non-linear LU-SGS solver with the SV and SD methods and the evaluation
of its performance both with analysis and computation are two of the main
cores of this thesis.
Although spatially high-order accurate numerical schemes guarantee a
better resolution of small scales than low-order ones (the latter require
more grid points to achieve the same level of accuracy), their application
to the simulation of general turbulent flows implies that particular attention has still to be paid to subgrid models. Therefore, since this PhD work
has been accomplished in the framework of the IWT Project SBO 050163
(Simulation and design tools towards the reduction of aerodynamic noise
in confined flows), and its final goal was the the development of an efficient N-S/LES solver for high-order accurate schemes, the SD scheme has
been coupled with the wall-adapted local eddy-viscosity (WALE) model to
perform large eddy simulations (SD-LES). The development and the evaluation of the accuracy and the reliability of the implicit SD-LES solver are
also two main cores of this thesis. It should be noted that the coupling
SV-LES was not treated because no stable three-dimensional SV partition
for third-order accurate scheme seems to exist.
The remainder of this thesis is organized as follows. A survey of the
available literature on the SV and SD methods, time integration/iterative
solution approaches for spatially high-order methods and multigrid algorithms is included in Chapter 2. The governing equations that describe the
flow problems considered in this thesis, namely the compressible NavierStokes equations and the filtered compressible Navier-Stokes equations for
LES, are presented in Chapter 3. The WALE model for the closure of the
subgrid-scale terms is also discussed. Chapter 4 is devoted to the description of the SV and SD methods and the coupling of the latter scheme with
the WALE model through a new definition of the grid filter width. The
non-linear LU-SGS algebraic solver, in combination with the BE scheme
and the BDF2 is presented in Chapter 5. In the same chapter, the lo4

cal time stepping technique is also discussed. A Von Neumann stability

analysis for the LU-SGS algorithm with the BE scheme and for general
SV and SD methods, respectively on triangular and quadrilateral grids,
is described in Chapter 6. In Chapter 7 the main results achieved by applying the SV method and the nonlinear LU-SGS algorithm with the BE
scheme to problems governed by the two-dimensional steady compressible
Navier-Stokes equations are presented. In order to assess the convergence
properties of the non-linear LU-SGS algorithm, the performance of the latter scheme is compared with that of a family of E-RK smoothers. Both time
marching schemes are coupled with a p-multigrid algorithm to accelerate
the convergence to steady state solutions. Chapter 8 deals with the main
results achieved by applying the SD method and the nonlinear LU-SGS
algorithm to problems governed by 2D and 3D compressible Navier-Stokes
and filtered Navier-Stokes equations. The convergence properties of the
non-linear LU-SGS algorithm for steady flow simulations are assessed by
comparing the algorithms performance with that of a Newton-Raphson
GMRES algebraic solver. For the unsteady laminar and turbulent flow
simulations, an extensive study of the accuracy and reliability of the implicit SD and SD-LES solvers is done by means of validation against experimental data and/or reference solutions available in literature. Finally,
in Chapter 9, conclusions from the results presented this thesis are drawn
and an outlook for future work is given.

CHAPTER 1. INTRODUCTION

Chapter 2

Literature survey
In this chapter an overview of the available literature on the spectral volume (SV) and the spectral difference (SD) methods, efficient time marching
and algebraic solvers for spatially high-order methods and multigrid algorithms is given.
Both SV and SD methods are strongly related to the discontinuous Galerkin
method (DG), which was introduced in 1973 by Reed and Hill [140] to solve
the neutron transport equation. The development of the DG method for
hyperbolic conservation laws was pioneered by Cockburn, Shu and their
co-workers in a series of papers on the Runge-Kutta DG (RKDG) method
[37, 39, 41, 43]. Nowadays, the DG method is the most popular and most
developed high-order accurate method for unstructured grid. Bassi and
Rebay demonstrated the capabilities of the DG method by achieving highorder accuracy for the compressible Euler and Navier-Stokes (N-S) equations [15, 16]. For a comprehensive review of the DG history and literature
the interested reader is referred to Cockburn et al. [38].
Recently, Huynh [79, 80], Wang and Gao [52, 184, 185] and Haga et al.
[61] have proposed a new high-order method for unstructured grids. This
method is called lifting collocating penalty (LCP) approach and unifies several of the popular methods including the DG method, the SV method and
the SD method with a technique that does not require the evaluation of
any integrals. Consequently, the evaluation of the residuals is relatively
cheap. If the parameters of the LCP method are chosen such that it is linearly equivalent to the DG method, then, like the DG method, it is linearly
stable on general hybrid grids.
7

CHAPTER 2. LITERATURE SURVEY

2.1 Spectral volume method

The SV method, with applications to one-dimensional (1D) scalar conservation laws, was proposed in 2002 by Wang [183] as an alternative for the
DG method. Further development of the SV method for two-dimensional
(2D) and for non-linear hyperbolic systems, such as the Euler equations,
was then reported in subsequent papers by Wang et al. [186, 187, 191].
The SV method was successfully extended to 2D N-S equations, and threedimensional (3D) Maxwell equations respectively by Sun et al. [164] and
Liu et al. [104]. Chen [32, 33] developed many high-order SV partitions
for simplexes in 2D and 3D with relatively small Lebesgue constants. The
appropriate treatment of curved wall boundaries for all high-order methods, was addressed for the 2D SV method by Wang and Liu [188], by using
a high-order geometric mapping of the SV cells near the boundaries. Comparisons between the SV and DG methods were made by Sun and Wang
[163] and by Zhang and Shu [199]. The SV method was also applied to
solve the 3D Euler and N-S equations by Haga et al. [62] on Japans Earth
Simulator. A positive step towards addressing the issue of stability was
given by Van den Abeele et al. [174, 176], who performed Fourier analysis for both 1D and 2D SV methods, and identified a weak instability
in several SV partitions. New partitions were derived which showed improved stability properties. In addition, Harris et al. [67] developed a
more efficient quadrature free implementation for the SV method, which
was significantly faster than the standard quadrature-based SV method.
Different approaches for the discretization of the diffusive terms in the NS equations with the SV method, based on similar approaches that were
developed for the DG method, were investigated in Kannan and Wang [90].
In 2009, Van den Abeele et al. [175] performed an extensive study of the
variation of the stability properties of the SV method on tetrahedral grids.
The study indicates that probably there is no partition that yields a stable third-order SV scheme for tetrahedral cells. In 2010, Harris and Wang
[66] have presented a constrained minimization approach in the design of
3D third-order SV schemes. Several new partitions were proposed which
have a reduced maximum real part of the Fourier footprint by up to 20%
over the original un-optimized partition proposed by Chen [32]. Numerical simulations have shown that the strength of the instability has been
weakened by about an order of magnitude for some cases by employing the
constrained minimization approach. However, also in this case, no fully
stable third-order SV scheme for tetrahedral cells has been found.
8

2.2. SPECTRAL DIFFERENCE METHOD

2.2 Spectral difference method

The SD method originated in the staggered grid multi-domain spectral
method proposed by Kopriva and Kolias [94] and Kopriva [93]. It was
generalized to simplex elements by Liu et al. [103], who applied the SD
method to 2D scalar conservation laws and the Maxwell equations. Extension of the SD method to the Euler equations was described in Wang
et al. [189], and to the N-S equations in May and Jameson [111] and
Wang et al. [190]. A 3D N-S implementation of the SD method for hexahedral grids was presented by Sun et al. [165]. Different approaches
for the discretization of the diffusive terms in the N-S equations with the
SD method, based on similar approaches that were developed for the DG
method, were investigated by Van den Abeele et al. [178]. Huang et al.
[77] reported an implicit space-time implementation of the SD method for
discontinuity capturing using adaptive polynomials. Huynh [79] showed
that for quadrilateral and hexahedral cells, tensor product flux point distributions based on a 1D flux point distribution consisting of the end points
and the Legendre-Gauss quadrature points, lead to stable schemes for arbitrary order of accuracy. In 2008, Van den Abeele et al. [177] showed an
interesting property of the SD method, namely that it is independent of
the positions of its solution points. Recently, this property has been proved
by Jameson [85]. In the same paper Jameson also has proved that for the
case of one dimensional linear advection the SD method is stable for all orders of accuracy in a norm of Sobolev type, provided that the interior fluxes
collocation points are placed at the zeros of the corresponding LegendreGauss polynomial, as reported by Huynh [79].
The performance of the SD method for turbulent flow simulations was
investigated by Liang and his co-workers [102] and Parsani et al. [129],
with an unresolved DNS type approach. More recently, Parsani et al.
[122, 124, 126] coupled, for the first time, a high-order SD scheme with
the local eddy-viscosity (WALE) model to perform 2D and 3D large eddy
simulations. The 3D LES was performed to simulate the turbulent flow
in a muffler which is also one of the benchmark test cases of the IWT
Project SBO 050163. The coupling of the SD scheme with the WALE
model, through a new definition of the grid filter width for high-order SD
schemes, is discussed in Section 4.2.6. The accuracy and the reliability of
the SD-LES approach are discussed in Section 8.3, where the turbulent
flow simulations are presented.
Application of the SD method in combination with a perfectly matched
9

CHAPTER 2. LITERATURE SURVEY

layer (PML) approach was reported by Zhou and Wang [201], for the simulation of computational aeroacoustics (CAA) benchmark problems. More
recently, Parsani and his co-workers [123, 125] have used the LES-highorder spectral difference method to provide the acoustic sources for aerodynamic sound field simulations with a Ffowcs-Williams Hawkings (FW-H)
code.

2.3 Time integration schemes

High-order spatial operators are usually much stiffer than lower-order
ones. For time accurate problems, the allowable Courant-Friedrichs-Lewy
(CFL) number usually decreases with increasing order of accuracy for explicit schemes [73]. For viscous problems, where cells with high-aspect ratio occur near the walls, resolving the viscous boundary layer with an economical distribution of grid points is severely limited by the time step size
for explicit high-order time marching schemes, and usually not competitive against low-order implicit methods in terms of efficiency, as demonstrated in Venkatakrishnan and Mavriplis [181]. In fact, the computational cost of high-order explicit time integration methods for many steadystate problems is so high that they become less efficient than low-order
implicit methods in terms of the total CPU time, given the same level of
solution error. Many types of implicit algorithms have been successfully
developed for unstructured grid-based solvers in the last two and a half
decades. Non-linear element-Jacobi, as well as linearized element-Jacobi
and Gauss-Seidel schemes were used by e.g. Nastase and Mavriplis [115]
for the DG method. Bassi and Rebay [15] successfully developed a preconditioned Newton-Raphson GMRES solvers for the DG method to solve the
compressible Navier-Stokes equations. In Van den Abeele et al. [178, 179]
a preconditioned Newton-Raphson GMRES approach was applied to SV
and SD methods. A matrix-free Krylov method was applied to DG schemes
by Rasetarinera and Hussaini [139] and to SD schemes by May et al. [110].
The non-linear lower-upper symmetric Gauss-Seidel (LU-SGS), developed
by Chen and Wang [34], was used with SD schemes by Sun et al. [166, 168]
and Van den Abeele et al. [178], with SV schemes by Parsani et al. [128],
Kannan et al. [90] and Haga et al. [63]. Line-implicit solvers were developed by Fidkowski et al. [48] for their DG method.
The present author performed an extensive study of the smoothing properties of the non-linear LU-SGS solver in combination with the backward
Euler scheme for SV schemes. This analysis was reported in Parsani et al.
10

2.4. GEOMETRIC AND P -MULTIGRID METHODS

[127, 130] and discussed in Chapter 6. Recently, the non-linear LU-SGS algorithm with the BDF2 was used to perform 2D and 3D unsteady laminar
and turbulent flow simulations in Parsani et al. [122, 129]. An application of the non-linear LU-SGS solver to implicit Runge-Kutta schemes was
reported by Parsani et al. [128], where explicit-first-stage, single-diagonalcoefficient, diagonally-implicit Runge-Kutta (ESDIRK) schemes were used
to advance the quasi-1D Euler equations in pseudo time. In the latter work
a SV scheme was used for the spatial discretization. A short description of
the non-linear LU-SGS solver in combination with ESDIRK schemes has
been included in Appendix D.

2.4 Geometric and p-multigrid methods

One powerful solution strategy for solving large scale problems in fluid
dynamics is multigrid [24, 25, 60, 172]. The standard multigrid algorithm (geometric multigrid or h-multigrid) has been used very effectively
in computational fluid dynamics (CFD) to accelerate the rate of convergence to steady state. In its implementation, multiple levels of coarse
grids are generated either from the finer grids or independently, with each
coarser grid doubling the mesh size of the next finer mesh in all directions. Jameson and Caughey [86] demonstrated that an Euler solution for
airfoil flows, converged to the level of the truncation error, could be obtained in 3-5 multigrid cycles. However, several factor contribute to the
effectiveness of the multigrid approach. Multigrid methods for hyperbolic
problems depend on two elements to accelerate convergence. One element
is the smoothing of high-frequency components of the solution error. The
choice of an iterative scheme for smoothing is crucial, since multigrid requires a smooth solution error to approximate a fine grid problem on a
coarser grid. In addition, the smoother must be effective on the coarser
grids, since these grids are responsible for removing the low-frequency error modes that cause slow asymptotic convergence of iterative schemes.
The second element for accelerating convergence is the expulsion of errors
on the coarse grids, which occur faster for time-like iterative methods due
to the larger time steps permitted on coarser grids.
The most critical element of a successful multigrid approach is the development of an effective smoother to remove various types of stiffness in the
spatial operator. For example, the stiffness due to low flow speed and grid
anisotropy was known to degrade the performance of multigrid solvers.
Many novel numerical techniques were developed to address these issues,
11

CHAPTER 2. LITERATURE SURVEY

including low speed preconditioning [35, 173], line-implicit solvers [109],
semi-coarsening [113], multigrid cycles with a preconditioned non-linear
LU-SGS smoother [86] and multigrid cycles with a Runge-Kutta/Implicit
scheme [149, 169].
In addition, p-multigrid, or multiorder, solution strategies can be used
in an analogous manner to accelerate convergence of high-order methods
to steady state by recursively iterating on solution approximations of different polynomial order. Therefore, lower-order spatial operators in a pmultigrid approach serve as the coarse grid operators in the h-multigrid
counterpart. Lower-order operators have similar advantages: they have
more numerical damping and allow larger time steps than high-order operators. The p component of this algorithm was proposed by Rnquist and
Patera [146] and analyzed by Maday and Munoz [107] for 1D Galerkin
spectral element discretization of Laplace equation. Helenbrook [68] combined p-multigrid with standard low-order multigrid and applied it to an
unstructured stream-wise-upwind-Petrov-Galerkin (SUPG) discretization
of the incompressible Navier-Stokes equations. In the context of DG approximations, Hemker et al. [71] analyzed block Jacobi smoothing strategies with h-multigrid for 1D diffusion problems. In addition, p-multigrid,
or multiorder, solution strategies have been studied for high-order DG by
Helenbrook et al. [70], Bassi and Rebay [18] and Fidkowski et al. [48],
showing several advantages such as ease of implementation and orderindependent convergence rates. Nastase and Mavriplis [115] used a hpmultigrid approach for the DG method, showing convergence rates which
are independent of both order of accuracy (p) of the discretization and level
of mesh resolution (h). It was also applied to the SV method by Van den
Abeele et al. [174], Parsani et al. [127, 128, 130] and Kannan et al. [90],
and to the SD method by Premasuthan et al. [133] and May et al. [110].
In Liang et al. [101] a p-multigrid method was applied to the spectral
difference method with explicit and implicit smoothers on unstructured
triangular grids.

12

Chapter 3

Governing equations
In this chapter, the governing equations which mathematically describe
all physical flow problems presented in this thesis are reviewed. Firstly,
the complete set of convection-diffusion equations for the macroscopic motion of real fluid substances are presented. These equations correspond to
the compressible Navier-Stokes (N-S) equations. Afterwards, the constitutive relation for Newtonian fluids is introduced. Air, which is the working
fluid in the present work, belongs to this fundamental class of fluids. The
system of equations for Newtonian fluids is then completed by modelling
the thermodynamic properties of the fluid with the ideal gas model. This
model compares well with the thermodynamic behavior of air, for the flow
problems considered in this thesis.
Next, important non-dimensional parameters as Reynolds number, Mach
number and the Prandtl number, which characterize the flow, are introduced. Successively, defining the concept of spatial filtering technique, the
compressible N-S equations are presented in the framework of large eddy
simulation (LES). To close the LES system of equations, the wall-adapting
local eddy-viscosity (WALE) model is introduced. For both N-S and LES
equations the formulation in tridimensional Cartesian space is given.
Since the N-S equations describes an initial-boundary value problem (IBVP),
it must be equipped with appropriate initial conditions and boundary conditions. The boundary conditions used in this work are introduced after
the presentation of the governing equations, whereas the initial conditions
are specified only in the next chapters, where flow simulations are presented.
13

CHAPTER 3. GOVERNING EQUATIONS

Dimensionless quantities like pressure coefficient, lift coefficient and the

drag coefficient, which can be used to assess the accuracy of a numerical
method to solve the fluid flow problems, are also discussed.
To conclude the chapter, the conservation law for the linear advection equation used in Chapter 6 to perform the Von Neumann analysis of the lowerupper symmetric Gauss-Seidel algorithm (LU-SGS) is defined.

3.1 Compressible Navier-Stokes equations

The compressible Navier-Stokes (N-S) equations are derived from the basic principles of conservation of mass, momentum, and energy with the assumption that the fluid, at the scale of interest, is a continuum (continuum
hypothesis), i.e. the macroscopic scale of the fluid motion is large compared
to the distance between the molecules of the fluid. It is important to notice
that, once the continuum hypothesis is invoked, continuous fields are obtained, and all the notions of the discrete molecular nature of the fluid can
be neglected. Consequently, the molecular scales cease to be relevant and
the fluid properties can be described by continuous functions of the position vector1 ~r Rdim , which belongs to a dim dimensions physical space,
and the time t R [120, 132]. Therefore, introducing the following notation: (~r, t) : Rdim R R for the mass density, ~u (~r, t) : Rdim R Rdim
for the velocity vector and et (~r, t) : Rdim R R for the specific total
energy, the vector of the conserved variables can be defined as

~ ,
w (~r, t) = ~u = m
et
et

(3.1)

where w (~r, t) : Rdim R Rvars , with the superscript vars denoting the
number of conservative variables, i.e. the number of scalar equations obtained from the the basic principles of conservation. Now, consider a fluid
in a domain with boundary surface denoted by , without radiation and
external volume forces. In this situation, the system of the compressible
N-S equations are mathematically described by the following convection1 Vectors

are assumed to be column vectors.

14

3.1. COMPRESSIBLE NAVIER-STOKES EQUATIONS

diffusion equations, written in conservative form
~ ~
+ f (w) = 0,
with f~ = m,
~
t
m
~
m
~ m
~
~~
~ f~m
+
with f~m
G
(w) ,
~ (w) = 0,
~ =
t



m
~
et ~ ~
~~
et G
(w) + ~q (w) ,
+ fet (w) = 0, with f~et =
t

(3.2a)
(3.2b)
(3.2c)

2
~
where f~ , f~m
~ and fet are the mass flux vector, the momentum flux vector
and the total energy flux vector, respectively. Moreover, in this expression,
~~
~ G
,
(w) : Rvars Rdim Rdim and q~ (w) : Rvars Rdim denote the
nabla/del operator, the stress tensor and the heat flux vector, which are
functions of the conserved variables. System (3.2) can also be rewritten in
a compact form as
w ~ ~
+ f (w) = 0,
(3.3)
t

where symbol ~f (w) : Rvars Rvars Rdim denotes the vector flux function

T
~f = f~ (w) , f~m
~
,
~ (w) , fet (w)

(3.4)

which is only function of the conserved variables. This implies that, both
stress tensor and heat flux vector are functions of the conserved variables.
Therefore, Equation (3.3) represents the conservative form of the compress~~
ible Navier-Stokes equations. The explicit dependence of G
and ~q on w will
be shown in the next sections.
Notice that, since the system (3.2) describes an initial-boundary value
problems it must be equipped with appropriate initial and boundary conditions, i.e.
w (~r, 0) = w0 (~r)
w| (t) = wb (t)

~r ,

t [t0 , tend ],

(3.5a)
(3.5b)

where t0 and tend denote the lower and the upper limits of the time variable
t.
2 The

symbol denotes the tensor product operator.

15

CHAPTER 3. GOVERNING EQUATIONS

3.1.1

Newtonian fluid

Generally speaking, a symmetric tensor can be decomposed in an isotropic

part and in a symmetric one. The latter represents the deviation from the
isotropic part which is called the deviatoric stress tensor. Therefore, ap~~
plying this decomposition to the stress tensor G,
introduced in Equations
(3.2b) and (3.2c), one obtains
~
~ = I~
~+~
G
~,

(3.6)

~~
where I,
and ~~ represent the isotropic stress tensor, the scalar number
~
which characterizes I~ and the deviatoric stress tensor, respectively. This
decomposition is not unique because is a free parameter. However, if
~
is set equal to the thermodynamic pressure P 3 , then tensor I~ becomes
the identity tensor and the deviatoric stress tensor ~~ becomes the viscous
stress tensor [36]. The latter is related to the symmetric velocity gradient
~~
tensor S
defined by
1
~~
S
(w) =
2

(

m
~



~
+

m
~

T )

(3.7)

Now, in the present thesis the working fluid is air. Air belongs to the fundamental class of Newtonian fluid, for which a linear relation between ~~
~~
and S
exists. Therefore, the viscous stress tensor ~~ can be expressed as

 
~
~
~~
~
~ m
~ (w) + d
~ (w) = 2 S
I,

(3.8)

where and d are the dynamic viscosity and the dilatation viscosity coefficients. The latter is related to a viscous stress caused by a volume
change. Both coefficients are in general functions of the temperature T
and the pressure P of the fluid, i.e. = (T, P ) and d = d (T, P ).
Moreover, they must satisfy the following conditions: > 0, (T, P ) and
d + 32 0, (T, P ) [134].
Now, assuming the Stokes hypothesis to be valid4 [120, 134], i.e. d = 23 ,
3 The mathematical/thermodynamic definition of both temperature and pressure will be
introduced in Section 3.1.3.
4 The Stokess hypothesis is valid for mono-atomic gases and for more complex gases (such
as air) in a wide range of thermodynamic conditions.

16

3.1. COMPRESSIBLE NAVIER-STOKES EQUATIONS

Equation (3.8) becomes
~
~ (w) = 2


  

m
~
1 ~
~
~
~

I~ .
S (w)
3

Therefore, inserting (3.9) in (3.6), with = P , one obtains


  

~
~
~
~
~ m
~ = P I~
~+2 S
~ (w) 1
I~ .
G
3

(3.9)

(3.10)

Notice that, the pressure, which appears explicitly in Equation (3.10), and
the temperature, whose contribution is hidden inside the coefficient ,
must be expressed as functions of the conserved variables, in order to as~
~ is only function of the conserved variables, i.e.
sert that the stress tensor G
~~
~~
G = G (w).

3.1.2 Thermal conductivity

Assume the Fouriers law of conduction to be valid [108]. Therefore, the
relation between the heat flux vector ~
q and the gradient of the temperature
~ is linear and can be expressed as
T
~
q = T,
~

(3.11)

where is is the fluids conductivity coefficient which is also function of

the thermodynamic conditions of the fluid, i.e. = (T, P ). The coefficient
must satisfy the following condition: > 0, (T, P ). Notice once more
that, in order to assert that ~
q = ~q (w), both temperature and pressure must
be expressed as functions of the conserved variable w.

3.1.3 Thermodynamic properties: ideal gas model

The compressible N-S equations defined by system (3.2) consist of three
equations (two scalar equations and one vectorial equation) with seven
unknowns: (~r, t), ~u (~r, t), et (~r, t), P (~r, t) and T (~r, t). Therefore, additional
equations (together with suitable initial and boundary conditions) must be
prescribed to complete (3.2). The missing equations are derived from the
fundamental equation of the thermodynamics which, for single-component
fluid at the chemical equilibrium, is the entropy function S. The entropy is
a function of the extensive parameters of the thermodynamic system such
as the internal energy E, the volume V and number of moles Nmol 5 , i.e.
5 In

the present thesis only single-component fluid is considered.

17

CHAPTER 3. GOVERNING EQUATIONS

S = S (E, V, Nmol ) [29]. The entropy function must satisfy three essential
properties [29]:
Homogeneous first-order function of the extensive parameters.
Continuous, differentiable and monotonically increasing function of
the internal energy. This implies that S can be inverted with respect
to the internal energy and that the energy is a single-valued, continuous, and differentiable function of S, V and Nmol , i.e. E = (S, V, Nmol ).
Super-additive function in the sense specified in [51].

The various partial derivatives of the fundamental equation in the form

E = (S, V, Nmol ) are the intensive parameters and they are defined as

E 
T (S, V, Nmol ) , the temperature,
(3.12a)
S V,Nmol

E 
P (S, V, Nmol ) , the pressure,
(3.12b)
V S,Nmol

E 
(S, V, Nmol ) ,
the electro-chemical potential.
(3.12c)
N S,V

Equations (3.12) are called equations of state. They are not independent
because of the Schwarz theorem which states that the partial derivatives
commute [2]. Therefore, the knowledge of two equations of state is the necessary and sufficient condition to define the thermodynamic state of the
fluid. In fact, the Gibbs-Duhem relations [29] combined with two equations of state allows the reconstruction of the fundamental equation of the
thermodynamics.
Now, consider the ideal gas model, which approximates well with the thermodynamic behavior of air in a wide range of thermodynamic conditions.
The fundamental equation of thermodynamic in specific variables, written
both in entropic and energetic representations, is then
#
"  1
e 1 v
,
(3.13a)
s = s(e, v) = s0 + R ln
e0
v0
e = e(s, v) = e0

e(1)(ss0 )/R
(v/v0 )1

(3.13b)

where e0 , v0 and s0 are the extensive specific value at an initial reference

thermodynamic state, with s0 defined by

R
 .
(3.14)
s0 =
1
T 0
18

3.1. COMPRESSIBLE NAVIER-STOKES EQUATIONS

The parameter R is the specific gas constant, which is about 287.06 J kg 1 K 1
for air, and is a second characteristic gas constant which will be defined
later on. Therefore, temperature T and pressure P obey to the following
equations:
1
e
=
e,
s
R
e
e
= ( 1) = ( 1) e.
P =
v
v
T =

(3.15a)
(3.15b)

Equation (3.15b) shows that e = e (T ) = ( 1) /R T . Consequently, recalling that the total energy per unit volume is defined as
 2 !
m
1
~
et = e +   ,
(3.16)
2

the specific internal energy can be expressed as a function of the conserved

variables, i.e.
 2
1 m
~
et
  = e (w) .
(3.17)
e=

2
This equation allows to express both pressure and temperature as a function of the conservative variables:
 2 !
1  m
~
1 et
  ,
(3.18a)
T (w) =
R

2
 2 !
~
1  m
t
(3.18b)
P (w) = ( 1) e   .
2

Therefore, system (3.2) together with Equations (3.18a) and (3.18b) results
in a closed system of five non-linear partial differential equations (PDEs),
which defines the complete system of compressible N-S equations for a
Newtonian fluid modelled by the thermodynamic ideal gas model.

To conclude this section, the definition of specific heat capacities and speed
of sound are introduced, for the ideal gas model. Using the law of conservation of energy [134], the specific heat capacities at constant volume and
constant pressure result in

R
e 
,
(3.19a)
=
cv =

T v=const
1


e 
v 
R
cP =
+P
=
+ R.
(3.19b)


T P =const
T P =const 1
19

CHAPTER 3. GOVERNING EQUATIONS

Equation (3.19a) shows that the heat capacity at constant volume cv is constant. A gas with constant cv is named polytropic gas. Combining Equation
(3.19a) with Equation (3.19b), the Meyers relation is obtained:
cP = cv + R.

(3.20)

This expression demonstrates that, also the heat capacity at constant pressure cP is constant. Now, inverting Equation (3.19a) with respect to the
parameters and using Meyers relation, one obtains
=

cP
,
cv

(3.21)

which shows that the second characteristic gas constant introduced in

(3.14) represents the ratio of the specific heat capacities.
Another thermodynamic quantity, which will be useful to define a dimensionless parameter, is the speed of sound c. For a polytropic ideal gas, c is
defined by [29, 134, 162]
s
p
P
2 P (s, v)
c = v
(3.22)
=
= R T,
v

which shows that, for the polytropic ideal gas model, the speed of sound
depends only on the temperature T .

3.1.4

Formulation in Cartesian space

In order to show the contributions of the convective and the diffusive fluxes,
separately, Equations (3.2) can be rewritten as


w ~ ~
~ ~fD w, w
~
+ fC (w) =
,
(3.23)
t


~
where ~fC (w(~r, t)) : Rvars Rvars Rdim and ~fD w(~r, t), w(~
r , t) : Rvars

Rvars Rdim are the convective and the diffusive flux vectors defined as


T
m
~ m
~
~
~
~
~fC = m,
~ m
~
et + P I~
+ P I,
,
(3.24)

~fD =

T

m
~ ~
~ (w) ~
q (w)
.
0, ~
~ (w) ,

20

(3.25)

3.1. COMPRESSIBLE NAVIER-STOKES EQUATIONS

~ ~fC (w) and the divergence
In (3.23), the divergence ofthe convective
flux

~
~ ~fD w, w
describe the transport (convection proof the diffusive flux
cess) of the conserved quantities and the mechanisms that dissipate the
conserved quantities (diffusion process), respectively.
Now, consider a general tridimensional (3D dim = 3) Cartesian physical
T
space with position vector
~ 6 defined by
~ = [1 , 2 , 3 ] . The convective
T
~
and the diffusive
 flux vectors are then defined as fC (w) = [fC , gC , hC ]
T
~
= [fD , g , hD ] , with
and ~fD w, w
D

fC =

gC =

hC =

u1
u21 + P
u1 u2
u1 u3
u1 (et + P )
u2
u1 u2
u22 + P
u2 u3
u2 (et + P )
u3
u1 u3
u2 u3
u23 + P
u3 (et + P )

6 In

(3.26a)

(3.26b)

(3.26c)

Section 3.1, where the Navier-Stokes equations were introduced using a general vectorial notation, the position vector was denoted by symbol ~
r.

21

CHAPTER 3. GOVERNING EQUATIONS

and

fD =

gD

hD

0
11
21
31
u1 11 + u2 21 + u3 31 q1

0
12
22
32
u1 12 + u2 22 + u3 32 q2

0
13
23
33
u1 13 + u2 23 + u3 33 q3

(3.27a)

(3.27b)

(3.27c)

In Equations (3.27), ij represents the ijcomponent of the viscous stress

tensor which, assuming both Newtonian fluid and the Stokes hypothesis
to be valid (see Equation (3.9)), can be rewritten in tensorial notation as


ij
Skk
ij = 2 Sij
3

i, j = 1, . . . , dim.

(3.28)

In this expression ij is the Kronecker delta function7 and Sij represents

~~
the ij-component of the symmetric velocity gradient tensor S,
which, according to Equation (3.7), can also be rewritten in tensorial notation as
1
Sij =
2

ui
uj
+
j
i

i, j = 1, . . . , dim.

(3.29)

Moreover, in Equations (3.27), qi represents the i component of the heat

flux vector, which, assuming the Fouriers law of conduction to be valid, can
be expressed as
T
i = 1, . . . , dim.
(3.30)
qi =
i
7 In

Section 3.1.1, the parameter was set equal to the thermodynamic pressure P and
~
consequently, tensor I~ became the identity tensor.

22

3.1. COMPRESSIBLE NAVIER-STOKES EQUATIONS

3.1.5 Dimensionless numbers

Consider a flow modelled by the compressible N-S equations, valid in a domain with boundary . Dimensional analysis shows that this flow is
completely characterized by three dimensionless numbers: Mach number,
Reynolds number and Prandtl number. The Mach Number is a value useful for analyzing fluid flow dynamics problems where compressibility is a
significant factor. It is defined by
|~uc |
,
(3.31)
c
where the quantity at the denominator c is the speed of sound, defined
by Equation (3.22) for the ideal gas model, and ~uc is a characteristic flow
velocity. Notice that, the subscript c denotes a characteristic value and
should not be confused with the speed of sound. If the Mach number is
less than one, then the local flow is subsonic, and if it is greater than one,
then the local flow is supersonic. Typically, a flow can be treated as incompressible if M . 0.3. For higher Mach numbers, compressibility must be
taken into account.
M=

The Reynolds number is defined by

Re =

c |~uc | Lc
|~uc | Lc
=
,
c
c

where is the kinematic viscosity coefficient given by

= .

(3.32)

(3.33)

The quantity Lc represents a characteristic length scale of the flow problem. The Reynolds number can be interpreted as a ratio of typical inertial
stresses and typical viscous stresses. In fact, its definition can be rewritten
as
2
c |~uc |
inertial forces
Re =
.
(3.34)

|~
uc |
viscous forces
c
Lc

If the Reynolds number is low, then the flow is dominated by the viscous
stresses, which results in a laminar flow. If the Reynolds number is high,
then the inertial stresses dominate and the flow is turbulent, characterized by random vortices and stochastic processes [132].
The Prandtl number P r is defined by
c cP,c
.
Pr =
c
23

(3.35)

CHAPTER 3. GOVERNING EQUATIONS

It is a measure of the ratio between momentum diffusivity and thermal
diffusivity, or the ratio of the rate by which momentum is transferred by
viscosity and the rate by which heat is transferred by conduction. The
latter interpretation can be understood from the following relation:
Pr =

c
c
c cP,c

momentum diffusivity
.
thermal diffusivity

(3.36)

According to Equation (3.35), the Prandtl number depends only on the

fluids properties, which for air, in a wide range of thermodynamic conditions around the standard values, can be assumed constant and equal to
0.72 [162].

3.2 Large eddy simulation

The four main numerical procedures for solving the N-S equations are
Direct Numerical Simulation (DNS), Large Eddy Simulation (LES), Detached Eddy Simulation (DES) and Reynolds Averaged N-S (RANS) approach. An elaborate overview of these techniques can be found in several
books [105, 132, 193]. The most accurate approach is DNS. DNS consists
in solving the N-S equations (system (3.2) together with Equations (3.18a)
and (3.18b)), resolving all the scales of motion, with initial and boundary
conditions appropriate to the flow considered. Conceptually it is the simplest approach and, when it can be applied, it is unrivalled in accuracy and
in the level of description provided. However, it is important to appreciate
that the cost is extremely high. In fact, since all the spatial scales, from
the smallest dissipative Kolmogorov scale (l ) up to the energy containing
integral scale (l), are needed to be resolved by the computational mesh,
the number of modes (points) required for a resolved DNS in 3D can be
estimated as [132],
N

mod,3D

l
l

3

u c lc
c

3

= Re9/4 .

(3.37)

The number of modes, for fully resolved DNS is enormous large, especially
for high Reynolds number flows encountered in most industrial applications, and consequently DNS is restricted to relatively low Reynolds number flows8 . To overcome this limitation and solve high-Reynolds number
flows, other numerical techniques like for instance LES, DES and RANS
8 DNS

is generally used as a research tool for analyzing the mechanics of turbulence, such
as turbulence production, energy cascade, energy dissipation, drag reduction etc.

24

3.2. LARGE EDDY SIMULATION

have been developed. In this thesis, the LES approach is used as a cheaper
alternative to DNS.
In LES, the larger 3D unsteady turbulent motions are directly represented,
whereas the effect of the smaller-scale motions are modelled. In computational expense, LES lies between RANS and DNS, and it is motivated by
the limitations of each of these approaches [132]. Because the large-scale
unsteady motions are represented explicitly, LES can be expected to be
more accurate and reliable than RANS for flows in which large-scale unsteadiness is significant - such as the flow over bluff bodies, which involves
unsteady separation and vortex shedding. As already discussed, the computational cost of DNS is high, and it increases as the cube of the Reynolds
number. Nearly all of the computational effort in DNS is expended on
the smallest, dissipative motions, whereas the energy and anisotropy are
contained predominantly in the larger scales of motion. In LES, the dynamic of the larger-scale motions (which are affected by the flow geometry
and are not universal) are computed explicitly, while the influence of the
smaller scales (which have a universal character) are being represented by
simple models. Thus, compared with DNS, the large computational cost of
explicitly representing the small-scale motions is avoided.
There are four conceptual steps in LES.
Decompose the velocity ~u (~
, t) into the sum of a filtered (or resolved)
component ~u (~
, t) and a residual (or subgrid-scale, sgs) component
represents the motion
~usgs (~
, t) ~u (~
, t). The filtered velocity field ~u
of the large eddies.
The equations for the evolution of the filtered fields are derived from
the compressible N-S equations (system (3.2) together with Equations (3.18a) and (3.18b)). Thus, the momentum equation contains
the residual stress tensor (or sgs stress tensor) ~~ sgs and the energy
equation contains both ~
~ sgs and the residual heat flux vector (or sgs
sgs
heat flux vector) ~
q .
Closure is obtained by modelling ~
~ sgs and ~qsgs , most simply by eddyviscosity [132] and eddy-diffusivity [50] models, respectively.
The resulting filtered equations, i.e. the filtered compressible NavierStokes equations, are solved numerically, providing one approximated
realization of the turbulent flow.
25

CHAPTER 3. GOVERNING EQUATIONS

3.2.1

Formulation in Cartesian space

Proceeding with the first two steps described in Section 3.2, one can define
the vector of the conserved variables for LES as

wLES

,
= ~u = ~u

et
et

(3.38)

where the symbols () and () represent the spatially filtered and the Favre
filtered fields, respectively. The Favre filtered is defined as g = g/, where
g is any field [97]. Consequently, the components of the convective and the
diffusive flux vector of the filtered compressible N-S equations are

fLES
=
C

gLES
=
C

hLES
C

u1

u21 + P

u1 u2

u1 u3

et + P
u
1

u2

u1 u2

u22 + P

u2 u3

u
2
et + P

u3

u1 u3

u2 u3

u23 + P

et + P
u
3
26

(3.39a)

(3.39b)

(3.39c)

3.2. LARGE EDDY SIMULATION

and

fLES
D

gLES
=
D

hLES
=
D

0
sgs

11 11

sgs
,

21 21

sgs

31 31
sgs
sgs
sgs
u
1 (
11 11
)+u
2 (
21 21
)+u
3 (
31 31
) q1 q1sgs
(3.40a)

0
sgs

12 12

sgs
,

22 22

sgs

32 32
sgs
sgs
sgs
u
1 (
12 12
)+u
2 (
22 22
)+u
3 (
32 32
) q2 q2sgs
(3.40b)

0
sgs

13 13

sgs
,

23 23

sgs

33 33
sgs
sgs
sgs
u
1 (
13 13
)+u
2 (
23 23
)+u
3 (
33 33
) q3 q3sgs
(3.40c)

where, according to Equations (3.28), (3.29) and (3.30),



ij

ij = 2 Sij
i, j = 1, . . . , dim,
Skk
3
1
Sij =
2

u
i
u
j
+
j
i

qi =

T
i

i, j = 1, . . . , dim,

i = 1, . . . , dim.

(3.41)

(3.42)

(3.43)

Notice that, according to the LES approach, the total energy et should be
calculated as follows
 2
m

1
~ 
et = e +   + k sgs ,
(3.44)
2
where k sgs is the subgrid-scale kinetic energy. For air, k sgs is very small
and can be neglected as shown by Lesieur and Comte [97].

27

CHAPTER 3. GOVERNING EQUATIONS

From the definitions of the fluxes components it is seen that both momentum and energy equations differ from the classical fluid dynamic equations
only for two terms which take into account the contributions from the unresolved scales. These contributions, represented by the specific subgridscale stress tensor ijsgs and by the subgrid heat-flux vector defined qisgs ,
appear when the spatial filter is applied to the convective terms and they
are defined as follows [56, 97]
ijsgs = (ug
i u
j )
i uj u



qisgs = cP Tg
ui Tu
i

i, j = 1, . . . , dim,

(3.45)

i = 1, . . . , dim.

(3.46)

The interactions of ~
~ sgs and ~
q sgs with the resolved scales have to be modeled through a subgrid-scale model because they cannot be determined
using only the resolved flow field wLES .
i ijsgs in Equations
Notice that, according to Vreman et al. [182], terms u
(3.40) improve the results only at moderate or high Mach number. Consequently, if the Mach number is low ui ijsgs can be neglected.

3.2.2

The wall-adapted local eddy-viscosity model

In the previous section, it has been shown that the smaller scales and
their interaction with the resolved scales have to be modeled through the
subgrid-scale stress tensor ~
~ sgs . Tensor ~
~ sgs can be modeled at different
levels of complexity. The most common approach is based on the eddyviscosity concept in which one assumes that the residual stress is proportional to the filtered rate of strain [132]:


ij
sgs
sgs
D

ij kk ij = 2 t Sij
,
(3.47)
Skk = 2 t Sij
3
where t is the turbulent kinematic viscosity or eddy-viscosity coefficient.
In the wall-adapted local eddy-viscosity (WALE) proposed by Nicoud and
Ducros [116], it is assumed that the eddy-viscosity t is proportional to the
square of the length scale of the filter and the filtered local rate of strain.
Although the model was originally developed for incompressible flows, it
can also be used for variable mass density flows by giving the formulation
as follows [56, 126]
 
~
2  ~
t = (C) S  .
(3.48)
28

3.2. LARGE EDDY SIMULATION

 
 ~
~  is defined as
Here S


d
where Sij
is given by

with

i3/2
h
 
d d
Sij
Sij
 ~
~
S
i5/2 h
i5/4 ,
 = h
d
d

+ Sij Sij
Sij Sij

ij 2
1 2
d
2
Sij
=

gij + gji
g ,
2
3 kk
2
gij
=

u
i u
k
.
xk xj

(3.49)

(3.50)

(3.51)

Note that in Equation (3.48) , i.e. the grid filter width, is an unknown
function and is often taken to be proportional to the smallest resolvable
length scale of the spatial discretization. In this thesis, the definition of
the grid filter function will be discussed in Section 4.2.6, where the highorder spectral difference method is presented.
The WALE model is specifically designed to return the correct wall-asymptotic
y +3 variation of the subgrid-scale viscosity t [116] and the constant
model coefficient C can be adjusted so that the correct amount of subgrid
dissipation is obtained. This model has the following properties:
It is invariant to any coordinate translation or rotation.
It is easily computed on any kind of computational grid.
It is a function of both the strain and the rotation rates.
It naturally goes to zero at the wall: neither damping function nor
dynamic procedure is needed to reproduce the effect of the no-slip
condition.
For the subgrid heat-flux vector qisgs , if an eddy diffusivity model [50] is
used, the following expression is obtained


t cP T
qisgs = cP Tg
ui Tui =
,
P rt xi

(3.52)

where the value of the turbulent Prandtl number P rt for air is usually set
to 0.72 [198] and the eddy-viscosity is computed by Equation (3.48).
29

CHAPTER 3. GOVERNING EQUATIONS

3.3 Boundary conditions

Most partial differential equations encountered in physical applications
are of three basic types: elliptic, parabolic and hyperbolic systems. Gustafsson and Sundstrom [59] classified the unsteady compressible N-S equations as being an incompletely parabolic system. They demonstrated that
the number of boundary conditions shown in Table 3.3 is necessary and
sufficient to define a well-posed problem. Notice that in Table 3.3, the
Inlet

Comp. Euler Eqs.

Comp. N-S Eqs.

Subsonic
4
5

Supersonic
5
5

Outlet

Subsonic
1
4

Supersonic
0
4

Table 3.1: Number of boundary conditions to be specified for both compressible

Euler and Navier-Stokes equations.

number of boundary conditions for the compressible Euler equations is also

given. These equations are used to model the flow of a compressible inviscid fluid, or the flow of a compressible viscous fluid in flow regions where
the effect of viscosity and heat conduction are negligible. The compressible
Euler equations may be derived from the compressible N-S equations
ne
~
~
glecting the contribution of the diffusive fluxes, i.e. fD w, w = 0.

In this work, the boundary conditions are specified by a weak imposition

approach, i.e imposing a boundary value through the boundary face flux.
This is achieve by introducing ghost solutions, which are defined such that
the average of the ghost solution, indicated by the subscript gho, and the
internal numerical solution, indicated by the subscript int, is equal to the
value prescribed by the boundary condition.
In this work the following boundary conditions are used: far field, mass
density-velocity inlet, pressure outlet and solid surface. In the next sections, these boundary conditions are introduced for the compressible N-S
equations completed with the ideal gas model. However, for obvious reasons, they are also valid for the LES equations presented in Section 3.2.

3.3.1

Far field

In real life, with external flow problems, the domain extends to infinity or at least far enough from the body to justify this assumption. However,
30

3.3. BOUNDARY CONDITIONS

when the governing equations of fluids motion are solved numerically, it
is not possible to represent an infinite domain and an artificial far field
boundary, as far away as possible from the region of interest, must be introduced. In addition, at this boundary, a suitable far field boundary condition, which minimizes the reflections of out-going waves and is accurate,
must be prescribed. Currently, there is no method available that is capable
of completely avoiding such spurious reflections.
The simplest choice to impose a far field boundary condition is to assume
that the flow far from the region of interest is undisturbed. That means
treating the boundary with a Dirichlet boundary condition. The ghost solution is then computed as
wgho = 2 w wint ,

(3.53)

where w is the undisturbed flow solution. The ghost values of the gradient is set to the internal value,




~
~
w
= w
,
(3.54)
gho

int

such that the average of the ghost gradient and the internal gradient is
equal to the internal gradient. This approach does not take into account
the flow physics, but it generally works well if the boundary is sufficiently
far away from the body. It can be combined with a buffer layer to avoid
strong spurious wave reflections. Such a buffer layer is a zone between the
domain of interest and the actual far field boundary, where outgoing and
spuriously reflected waves are damped. The damping can be accomplished
for instance by progressively increasing the size of the cells towards the
outflow, which increases the numerical damping introduced by the solution
method, or by introducing additional damping terms in the buffer layer
(see for instance Zhou and Wang [200]). The latter approach belongs to a
class of techniques called non-reflecting/absorbing boundary conditions.
Among the non-reflecting boundary condition approaches, the most used
technique is based on the characteristics of the compressible Euler equations [57, 74, 75, 98, 131, 170]. This approach was developed as one of
the first attempts to minimize the reflection of outgoing waves. It is based
on a one-dimensional (1D) local approximation of the flow for the normal
direction to the boundary. A number of physical variables equal to the
number of outgoing Riemann invariants is extrapolated from the internal
domain. The remaining physical variables at the boundary are computed
using the expressions that state that the incoming Riemann boundary invariants corresponding to the 1D approximation are zero. For instance, in
31

CHAPTER 3. GOVERNING EQUATIONS

a general tridimensional case, the number of conserved variables, i.e. the
number of the unknown in the compressible N-S equations, is five. According to the 1D characteristic boundary condition approach, at subsonic
inflow, four variables must be specified with free stream values and one
is computed from the interior solution. At subsonic outflow, one variable
must be specified and four others are computed from inside the computational domain. If the flow is supersonic, all variables must be specified at
an inflow boundary and all must be computed from the interior solution at
a supersonic outflow boundary. This approach causes less reflections than
a simple Dirichlet boundary condition and it works well for steady problems. However, the boundary should still be placed as far away as possible
from the region of interest. For multi-dimensional problems, the performance of the 1D characteristic boundary condition approach degrades if
the wave propagation direction is not aligned with the boundary face normal direction. It can also be combined with a buffer layer to better damp
outgoing and spuriously reflected waves. A family of boundary conditions
which simulate outgoing radiation are derived by Bayliss and Turkel [19].
Those boundary conditions were applied to the computation of steady state
flows and are shown to significantly accelerate the convergence to steady
state.

3.3.2

Inlet mass density and velocity

According to the 1D characteristic boundary condition approach, for a subsonic inflow, four variables must be specified and one must be computed
from the interior solution. This idea is used to impose inlet boundary conditions for flow problems where, for obvious reasons, the boundary or
a part of it is always associated to a subsonic inflow9 . In this case, both
mass density inlet and velocity profiles ~uinlet are prescribed and the ghost
solution is computed as
gho = 2 inlet int ,

(3.55a)

m
~ gho = gho ~ugho ,
et

gho

with ~ugho = 2 ~uinlet ~uint ,

2

m
Pint
1
~ gho 
=
,
+ gho 
1 2
gho 

(3.55b)
(3.55c)

where Pint denotes the internal numerical pressure at the boundary. In the
latter expression, the internal energy per unit volume has been expressed
using Equation (3.15b), which is valid for the ideal gas model.
9 For

instance, the inlet section of a pipe.

32

3.3. BOUNDARY CONDITIONS

3.3.3 Pressure outlet

This boundary condition is used for outlet boundary where, according to
the 1D characteristic boundary condition approach, one variables must be
specified. It is assumed that the pressure distribution P is known and
equal to Poutlet . Consequently, the ghost solution is computed as
gho = int ,

(3.56a)

m
~ gho = gho ~ugho ,

(3.56b)

et

gho

with ~ugho = ~uint ,

2

m
Pgho
1
~ gho 

=
+ gho 
, with Pgho = 2 Poutlet Pint .
1 2
gho 

(3.56c)

Notice that, also in (3.56c) the internal energy per unit volume has been
expressed using Equation (3.15b). This type of outlet boundary condition
works well if the outlet boundary is sufficiently far away from the region
of interest. To avoid strong spurious wave reflections, it can be combined
with a buffer layer.

3.3.4 Solid wall

At solid walls, the only physical boundary condition required by a viscous
fluid is the no-slip boundary condition, which states that the fluid velocity relative to the wall has to vanish. This translates into the following
mathematical expression:
~u|wall ~uwall = ~0.

(3.57)

Recalling the significance of the Reynolds number as the ratio of convective

to viscous effects, a high Reynolds number means that the flow system will
be dominated by its inviscid properties. This can be clearly seen when analyzing viscous flows, with the important exception of the near-wall regions,
where the viscous effects, leading to a boundary layer or frictional layer
configuration (see Figure 3.1), dominate the flow behavior. The higher the
Reynolds number, the thinner the boundary layer becomes.
The concept of boundary layer, therefore, implies that flows at high Reynolds
numbers can be divided up into two unequally large regions. In the bulk of
the flow region the viscosity can be neglected, and the flow corresponds to
the inviscid limiting solution. This is called the inviscid outer flow. The second region is the very thin boundary layer at the wall where the viscosity
plays a very important role. Within the boundary layer both laminar and
turbulent flow can occur. One then speaks of laminar boundary layer flows,
33

CHAPTER 3. GOVERNING EQUATIONS

2
Uout (1)

Inviscid region

u1 (1, 2) = 0.99 Uout (1)

Boundary layer
thickness

u1 (1, 2)

Viscous region
1

Figure 3.1: Representative velocity profile in a laminar boundary layer over a twodimensional flat plate.

and equivalently of turbulent boundary layer flows. High Reynolds number flow simulations exhibit strong gradients normal to walls and across
shear layers requiring much finer resolution of the solution in some direction compared to others. To keep mesh sizes manageable for such problems, meshes with highly anisotropic elements are necessary. This introduces a geometrical stiffness which the solver must also be able to deal
with.
It must be made absolutely clear that the concept of boundary layer thickness has been artificially introduced. The transition from boundary layer
flow to outer flow, at least in the case of laminar flows, takes place continuously, so that a precise boundary layer thickness cannot, in principle,
be given. Since this concept is so vivid, it is very often used in practice
though. Frequently the boundary is arbitrarily given as being at the point
where the velocity reaches a certain percentages of the outer velocity, e.g.
99% as shown in Figure 3.1.
A boundary condition for the temperature status of the solid wall should
also be specified. Constant temperature wall leads to a Dirichlet type
boundary condition

T |wall Twall = 0.
34

(3.58)

3.3. BOUNDARY CONDITIONS

The ghost solution is then computed as
gho = int ,

(3.59a)

m
~ gho = gho ~ugho ,
et

gho

with ~ugho = ~uint ,

(3.59b)
2

m
~ gho 
1
, with Tgho = 2 Twall Tint ,
= gho cv Tgho + gho 
2
gho 
(3.59c)

where Equation (3.19a) has been used to express the internal energy e as
a function of the temperature T .
In case of imposed heat flux, the solid wall is the source of a fixed heat
flux normal to the wall to or from the fluid flow, for instance when the solid
surface is part of a heat exchanger system. The flux will be positive for a
heated wall or negative for a cooled wall. This boundary condition, assuming the Fouriers law of conduction to be valid, is described by the following
mathematical expression:


~ 
~1n T

wall

qn,wall = 0,

(3.60)

where ~1n represent the normal unit vector to the wall and qn,wall is heat
flux normal to the wall defined as qn,wall = ~qwall ~1n . In this case, within
a thermal boundary layer the temperature rapidly changes from the value
imposed by the wall to the one of the flow away from the wall.
The Prandtl number is a measure of the relative thickness of the momentum10 to the thermal boundary layers. In heat transfer problems, the
Prandtl number controls the relative thickness of the momentum and thermal boundary layers. When P r < 1, it means that the heat diffuses very
quickly compared to the velocity (momentum).
In case of imposed heat flux, the ghost value of the temperature gradient
is specified as


~
T

gho



~
= T

int

hq

n,wall

10 The



~
+ ~1n T

int

~1n .

(3.61)

boundary-layer thickness is used for a thickness beyond which the velocity is essentially the free-stream velocity (see Fig. 3.1).

35

CHAPTER 3. GOVERNING EQUATIONS

3.4 Aerodynamic coefficients

The exact solution of the N-S equations is generally not available. Therefore, numerical solutions are mainly validated by comparison with experimental results and/or accurate and usually very expensive simulations.
Dimensionless quantities like pressure, lift and drag coefficients can also
be used to assess the accuracy of a numerical method to solve fluid flow
problems. These coefficients contain complex dependencies of the geometry of the flow problem. To correctly use them, the viscosity and compressibility effects of predicted case must be equal to those of the measured
and/or computed case. Otherwise, the prediction will be inaccurate.
2

Let the quantity dynamic pressure be defined as 12 c |~uc | . The pressure coefficient CP is then defined as the difference of the instantaneous pressure
in one point and a reference pressure, divided by the dynamic pressure, i.e.
CP =

P Pc
2.
1
uc |
2 c |~

(3.62)

The pressure coefficient is a dimensionless number which describes the

relative pressures throughout a flow field in fluid dynamics. Every point
in a fluid flow field has its own unique pressure coefficient.
The lift coefficient CL expresses the ratio between the component FL of
the global aerodynamic force in the direction perpendicular to the reference velocity ~uc and the force produced by the dynamic pressure times a
reference area of the object Ac , i.e.
CL =

FL
1
2 c

|~uc | Ac

(3.63)

The drag coefficient CD expresses the ratio between the component FD of

the global aerodynamic force in the direction parallel to the reference velocity ~uc and the force produced by the dynamic pressure times a reference
area of the object Ac . Its definition is
CD =

FD
1
2 c

|~uc | Ac

36

(3.64)

3.5. LINEAR CONVECTION EQUATION

3.5 Linear convection equation

The linear convection equation (also known as linear advection equation)
describes mathematically the conservation law of a scalar conserved variable w convected by a constant vectorial velocity field ~a. Using the notation
introduced for the N-S equations in Section 3.2.1, the linear convection
equation is defined by
w ~ ~
+ fC (w) = 0,
(3.65)
t
where ~fC (w) = ~aw. Although this model is very simple, it describes important physical phenomena, like the propagation of waves or propagation of
a chemical substance where the diffusion in the flow is negligible. Moreover, since this simple model has a general analytical solution [98, 100],
it is often used to assess the stability properties of a spatial discretization
scheme and, as it will be shown in Chapter 6, to study the stability and the
damping properties of time solvers.
For an initial condition w0 (~r), Equation (3.65) has the general solution
w (~r, t) = w0 (~r ~at) .

37

(3.66)

CHAPTER 3. GOVERNING EQUATIONS

38

Chapter 4

Spatial discretization
In the previous chapter, the Navier-Stokes and the filtered Navier-Stokes
equations, which govern all the physical problem considered in this thesis,
have been discussed. They form two systems of partial differential equations (PDEs) which describe initial-boundary value problems. In a very
limited number of test cases, the exact solution of the Navier-Stokes equations is known. Therefore, in the majority of the flow problems, the solution
must be computed numerically and the PDEs must be discretized both in
space and time. In order to achieve this, the method of lines (MOL) is used
in this work [75, 157]. The basic idea of the MOL is to replace the spatial
derivatives in the PDEs with algebraic approximations. Once this is done,
the spatial derivatives are no longer stated explicitly in terms of the spatial independent variables. Thus, only the initial value variable, typically
time in a physical problem, remains. With only one remaining independent
variable, we have then a system of ordinary differential equations (ODEs)
that approximate the original PDEs. Therefore, the space discretization is
separate from the time integration. A significant advantage of the method
is that it allows the solution to take advantage of the sophisticated general
purpose methods and software that have been developed for numerically
integrating ODEs. For the PDEs to which the method of lines is applicable,
the method typically proves to be quite efficient.
In this chapter, two compact methods that are specifically designed for
high-order accuracy on unstructured grids, namely the spectral volume
(SV) and the spectral difference (SD) methods, are presented for the spatial discretization of system of PDEs. The mathematical formulations of
the SV and SD methods are described in the following sections.
39

CHAPTER 4. SPATIAL DISCRETIZATION

4.1 Spectral volume method

The SV method, with applications to 1D scalar conservation laws, was
proposed in 2002 by Wang [183] as an alternative for the discontinuous
Galerkin (DG) method [37, 3941, 43, 140, 180]. Further development of
the SV method for 2D and for non-linear hyperbolic systems, such as the
Euler equations, was then reported in subsequent papers by Wang et al.
[186, 187, 191]. The extension to 3D for linear systems was described in
Liu et al. [104]. The method was applied to the 2D N-S equations in Sun et
al. [164] and was extended to curved boundary representation by Wang et
al. [188]. The first application of the SV method to the 3D N-S equations
was reported by Haga et al. [62]. A positive step towards addressing the
issue of stability was given by Van den Abeele et al. [174176], for 1D, 2D
and 3D SV schemes.

4.1.1

Discretization of convective term

The spectral volume method can be interpreted as a finite volume (FV)

method, where within each grid cells a miniature structured mesh of control volumes (CVs) is formed. In this way a unique stencil for the flux is
defined for each face, eliminating the need for searching operations, which
are needed in traditional high-order (> 2) FV methods on unstructured
grids. Volume-averaged conserved variables on the CVs, which are the
solution variables for the SV method, are then used to reconstruct a highorder polynomial inside the cell. To represent the solution as a polynomial
of degree p, i.e. a (p + 1)-th-order accurate scheme, we need N s pieces of independent information, or degrees of freedom (DOFs). Simplex cells (lines,
triangles and tetrahedrons) are then subdivided into
Ns =

(p + dim)!
p!dim!

(4.1)

CVs. Tensor product cells (quadrilaterals and hexahedrons) are subdivided

into
dim

N s = (p + 1)

(4.2)

CVs. For example, 1D, 2D and 3D spectral volumes supporting a quadratic

data reconstruction (p = 1) are shown in Figures 4.1(a), 4.1(b) and 4.1(c)
respectively.
Consider the following form of an hyperbolic system of conservation laws
40

4.1. SPECTRAL VOLUME METHOD

(a) 1D SV.

0.866

0.433

0
0

0.5

(b) 2D SV triangular cell.

(c) 3D SV tetrahedral cell.

Figure 4.1: Partitions for second-order (p = 1) SV method for 1D cell, 2D triangular

cell and 3D tetrahedral cell.

41

CHAPTER 4. SPATIAL DISCRETIZATION

with only convective fluxes:
w ~ ~
+ fC (w) = 0,
t

(4.3)

valid on a domain with boundary . Define the CV-averaged conserved

variable for the CV j which belongs to cell i as
Z
i,j = 1
w d,
(4.4)
W
i,j i,j
where i,j is the volume of the CV. Now, integrating the system (4.3) over
the CV j and applying Gausss theorem, one obtains
I
i,j
dW
1
~fC ~1n dS = Ri,j ,
=
(4.5)
dt
i,j i,j
where Ri,j is the residuals which governs the evolution of the solution variables in time. Given the N s CV-averaged conserved variables for all CVs
in cell with index i, a polynomial of degree p can be constructed such that
it is a (p + 1)-th-order accurate approximation of w inside the cell i.
The solution polynomial in each cell allows to evaluate the surface integrals over the internal faces that enclose the CVs with (p + 1)-th-order of
accuracy (i.e. the integral in Equation 4.5). Here, a quadrature-free approach, which does not need the Gaussian quadrature rules [67], is used
to perform the integration. At the external faces, which lie on a grid face
between two cells, two approximate values for the conserved variables are
available, from the solution polynomials corresponding to the two neighboring cells. In order to ensure numerical conservation, the contributions
of a face to its two neighboring cells should be equal in magnitude and
opposite in sign. Thus, a unique flux ~fC ~1n should be computed from the
two available solutions. In Godunovs original method, this was achieved
by computing an exact solution to the resulting Riemann problem [100].
Nowadays, approximate Riemann solvers replace the exact Riemann solver
and all of them are of the following form:
~
~
~ AR (WL , WR ) ~1n = fC (WL ) + fC (WR ) ~1n |A| WR WL ,
F
2
2

(4.6)

where WL and WR are the reconstructed conserved variables from the left
and the right side of a face and, by convention, the unit normal ~1n is oriented from the left to the right side. The first term in the right-hand-side
42

4.1. SPECTRAL VOLUME METHOD

of expression (4.6) is just the average of the convective fluxes on the left
and the right side of the face. The second term ensures that the data
used for the computation of the flux comes from the side from which the
physical quantities are propagating. This is called upwinding and it introduces a necessary amount of damping to stabilize the computations. The
matrix |A| should be defined such that a correct upwinding is ensured.
The different approximate Riemann solvers are distinguished by the corresponding definitions of this matrix; see for instance Rusanov [150] and
Roe [143]. The flux values obtained from the approximate Riemann solver
at each point are then used to evaluated the surface integrals over the external faces with (p + 1)-th-order of accuracy by means of an appropriate
quadrature-free approach.
The final expression of the spatial discretization with the SV method is
then
i,j
dW
dt

1
=
i,j

1
i,j

X Z

Niintf ac
m=1

i,j Sm

X Z

Niextf ac
m=1

i,j Sm

~fC ~1n dS
~ AR ~1n dS = Ri,j ,
F

(4.7)

where Sm is the surface of the m-th face. Expression (4.7) is a system of

ODEs in time that approximates the original set of PDEs. It can be solved
with any time marching scheme.

4.1.2 SV basis polynomials

In order to define the basis functions for the construction of the solution
polynomial, a mapped coordinate system ~ = [1 , 2 , 3 ]T is introduced on
each cell. The transformation from the standard to the physical element
in the global Cartesian coordinates for the cell i is given by

1,i
1,i (1 , 2 , 3 )
 

~ i = 2,i = 2,i (1 , 2 , 3 ) =
~ i ~ ,
3,i
3,i (1 , 2 , 3 )

(4.8)

where the corresponding Jacobian matrix and the Jacobian determinant

 
~
are denoted respectively by ~J i and Ji . The solution polynomial Wi ~ in
43

CHAPTER 4. SPATIAL DISCRETIZATION

cell i has the form

Ns
 
  X
i,j L
i,j ~ ,
W
Wi ~ =

(4.9)

j=1

 
i,j ~ are a complete polynomial basis (SV basis polynomials)
where the L
for the polynomials space of degree p. In order to solve the reconstruction
problem, one should recall that the solution polynomial over a CV must be
equal to the corresponding CV-averaged solution, i.e.
Z
 
1
~
i,j ,
~ Ji d = W
j = 1, ..., N s .
(4.10)
W
i
i,j j~
Therefore, inserting expression (4.9) into identity (4.10) the following set of
linear algebraic systems that define the SV basis polynomials are obtained:
Z
 
1
i,m ~ Ji d~ = jm ,
j, m = 1, ..., N s ,
(4.11)
L
~
i,j j
where jm is the Kronecker delta function. This shows that in general, the
SV basis polynomials depend on the local geometry of the corresponding
cell, through the Jacobian determinant Ji [46, 183, 186]. Consequently,
the coefficients of these polynomials should either be stored for each cell,
or recomputed on the fly when needed. The first approach requires a very
large amount of memory, while the second uses too much computational
time. However, if grids with simplex cells are considered, the transformation (4.8) is linear for all the cells, except for those with curved boundary
representation, and the Jacobian determinant is a constant proportional
to the volume of the cell [46], i.e.
Ji = d!i .

(4.12)

In addition, for simplex cells with a linear mapping, one has

~

j
i,j
=
,
i
~

with

Therefore, identity (4.11) reduces to

Z
 
1
i,m ~ d~ = jm ,
L
~
~
j j
44

1
.
d!

j, m = 1, ..., N s .

(4.13)

(4.14)

4.1. SPECTRAL VOLUME METHOD

Expression (4.14) shows that for simplex cells with a linear mapping, the
SV basis polynomials are the same for each cell. Consequently, only one
set of polynomial coefficients must be precomputed and stored before the
simulation. Only for boundary cells, where a higher-degree mapping is
sometimes needed to discretize properly curved geometries, might have a
non-linear mapping. These cells usually form only a small fraction of the
total number of cells, and consequently, the basis polynomial coefficients
can be precomputed and stored for each individual element before the actual simulation.
For grids with quadrilateral or hexahedral cells, the transformation to the
mapped coordinates is linear only for parallelograms and parallelepipeds.
However, these type of cells are not flexible enough to mesh a general geometry, and consequently, almost every known implementation of the SV
method is limited to simplex cells.

4.1.3 Discretization of diffusive terms

The proper discretization of viscous fluxes has been studied extensively in
the DG methods and shown to be very important for accuracy and stability.
In classical second-order finite volume methods, the solution gradients at
an interface are usually computed by averaging the gradients of the neighbour cells sharing the face in consideration. However, for higher-order elements, special care has to be taken in computing the solution gradients.
In the late 1970s and early 1980s, Arnold [5] introduced the discontinuous finite element methods known as penalty methods for elliptic operators. More recently, many researchers [15, 16, 27, 42, 44] have applied DG
methods to diffusive operators. One procedure was the local discontinuous
Galerkin (LDG) method, developed by Cockburn and Shu [42, 44]. Its simplicity and effectiveness have made it the main choice for discretizing the
viscous fluxes. In 1997, Bassi and Rebay [15] came up with a symmetrical
scheme to discretize the viscous fluxes. In 2000, again Bassi and Rebay
came up with a second concept (also referred to as BR2 [17]). A nice unifying framework of these approaches for the DG method can be found in
Arnold et al. [6], where more details on the aforementioned methods can
be found. In this thesis, the local SV (LSV) approach is considered. It is
based on the local DG (LDG) approach which was proposed by Cockburn
and Shu [42, 44].


~
. To evaluate it, an approxConsider the diffusive flux vector ~fD w, w
imation of the gradients of the conserved variables is needed. An approx45

CHAPTER 4. SPATIAL DISCRETIZATION

 
~ i ~ can be obtained by computing CV-averaged values as folimation
lows:
I

1
~ 
w~1n dS
w
=
i,j i,j
i,j
1
i,j

1
+
i,j
~ i,j .

X Z

Niintf ac
m=1

Wi ~1n dS

i,j Sm

X Z

Niextf ac
m=1

~1n dS
W

i,j Sm

(4.15)

=W
(WL , WR ) is an averaged value given by
where W
= WL + WR (WR WL ) .
W
2

(4.16)

In the latter expression defines the bias in the averaging operator. The
gradients in cell i can thus be approximated by the polynomial
Ns

 
  X

~
~
i,j ~ .
~ i,j L
~

w i =
i

(4.17)

j=1



~ i at an interThe diffusive flux vector is then approximated as ~fD Wi ,
~ L and
~ R of
~ are available, and again
nal face. At a cell-face, twovalues


~ =
~ WL , WR ,
~ L,
~ R must be used. With the LSV
an averaged value
~ is defined by
approach


~
~
~R
~ L + (WR WL ) ~1n ,
~ = L + R +

2
hf

(4.18)

where hf is a local length scale associated to a face, defined as

(JL + JR )
(JL + JR )
 = C 
,
hf = C 
T
T



 ~

~~ 1
~ 
~ 
1
~
~
~
1n,L 
1n,R 
 JJ
 JJ




L
R
46

(4.19)

4.1. SPECTRAL VOLUME METHOD

~

with ~1n the unit normal to the face in the mapped coordinate system associated to a cell. In this work, the constant C is chosen so that for simplex
cells with a linear transformation to the mapped coordinate system, hf
reduces to
dim L + R
,
(4.20)
hf =
dim + 1
S
which is the sum of the perpendicular distances between the neighboring
cell centers and the face. For tensor product cells, it should be chosen
analogously such that
L + R
hf =
.
(4.21)
2S
The parameter sets the amount of damping added to the gradient, while
defines again the bias in the averaging operator. In Arnold et al. [6], it
was shown that > 0 results in stable LDG schemes, and no stability problems have been observed for the LSV approach either [179]. Notice that the
LSV approach is not compact, since the neighbors of the neighboring cells
are needed for the contribution of the external faces to the residuals in a
cell. In this thesis, the penalty method described in Kannan et al. [91] for
which = 1 and = 0 is considered.

4.1.4 Spectral volume partition

In order to define the SV method, the partition of the cell into CVs has to
be chosen. Depending on the order of accuracy, the CV partition has a certain number of identifying parameters which strongly affect the stability
and accuracy properties of the method. Consequently, a suitable choice for
them is of paramount importance. The proper definition of CV partitions
for the SV method has been subject of several researches [33, 46, 65, 66,
174, 176, 187].

Second-order accurate scheme

The partitions into CVs for second-order (p = 1) accurate SV schemes are
uniquely defined by a set of rules of thumb [46, 186, 187]. They can be seen
for linear, triangular and tetrahedral cells in Figure 4.1. Their stability
was confirmed by the wave propagation analysis reported in a series of
paper by Van den Abeele et al. [174176]. The higher-order partitions are
defined as follows.
47

CHAPTER 4. SPATIAL DISCRETIZATION

Partitions for 1D cells
A general third-order (p = 2) 1D partition is shown in Figure 4.2(a). It
has one DOF, 3 ]0, 1[, which can be easily understood from the figure.
Figure 4.2(b) shows a general fourth-order (p = 3) 1D partition. There is
again one DOF, labeled 4 ]0, 1[.
0

0.58
3

0.58
+3

(a) p = 2.

0.78
4

0.78
+4

(b) p = 3.

Figure 4.2: General SV partitions for third- (p = 2) and fourth-order (p = 3) 1D

schemes.

The family of third- and fourth-order SV schemes

based on the Gauss

Lobatto points, i.e. 3 = 0.50 and 4 = cos 4 , was found to be weakly
unstable [174]. Van den Abeele and his co-workers carried out a detailed
analysis of the 1D SV partition [46, 174] for third-, fourth-, fifth- and sixthorder schemes. In the above works, stable third- (3 = 0.58 and 3 = 0.65)
and fourth-order (4 = 0.78 and 4 = 0.86) schemes that are optimal for
wave propagation were designed and tested. In the same study, the stability of a recently proposed family of schemes based on the Gauss-Legendre
quadrature points [79] was confirmed too.
Partitions for triangles
A general third-order partition of a triangular cell is plotted in Figure
4.3(a). It has two DOFs, which are defined as




|AE|
1
2
|AC|
and
3 =
,
(4.22)
0,
0,
3 =
|AB|
2
|AD|
3
48

4.1. SPECTRAL VOLUME METHOD

where points A, B, C, D and E are shown in the figure. Figure 4.3(b) shows
a general fourth-order partition of a triangular cell. In this case, there are
four DOFs:




|AE|
1
2
|AC|
,
4 =
,
0,
0,
4 =
|AB|
2
|AD|
3




|GD|
|AF |
1
2
4 =
and 4 =
.
(4.23)
0,
4 ,
|AD|
3
|AD|
3
Weak instabilities in three third-order SV schemes found in the literature
[104, 186, 188] were identified analytically and verified numerically in Van
den Abeele et al. [176]. They correspond to 3 = 14 and 3 = 23 [186], 3 = 14
and 3 = 41 [104] and 3 = 41 and 3 = 31 [188] respectively.
0.866

0.866

0.433

0.433

F
E

0
A

0.5

(a) p = 2.

0.5

(b) p = 3.

Figure 4.3: General SV partitions for third- (p = 2) and fourth-order (p = 3)

schemes of a triangular cell.

A stable and accurate third-order scheme was proposed by Van den Abeele
and his co-workers [176]. It is defined by the following choice of parameters: 3 = 0.1093621117 and 3 = 0.1730022492 (see Figure 4.4(a)). It was
found to be more accurate than the SV schemes proposed in Chen [33] and
for this reason is used in the present work.
Weak instabilities in two previously proposed fourth-order SV schemes
were also identified in Van den Abeele et al. [176]. Those partitions were
first presented in Wang and Liu [186] and Harris and Wang [65] and are
2
1
2
1
, 4 = 15
, 4 = 15
, 4 = 15
and 4 = 0.12061033,
defined by 4 = 15
49

CHAPTER 4. SPATIAL DISCRETIZATION

0.866

0.866

0.433

0.433

0
0

0.5

(a) p = 2.

0.5

(b) p = 3.

Figure 4.4: Stable third- (p = 2) and fourth-order (p = 3) SV partitions of a triangular cell [176] used in the present work.

4 = 0.12129456, 4 = 0.066666667, 4 = 0.312260947 respectively. Again

in Van den Abeele et al. [176], a stable and accurate fourth-order scheme
was proposed. This scheme shows better wave propagation properties than
the previously proposed SV schemes [33]. It is defined by 4 = 0.078,
4 = 0.104, 4 = 0.052 and 4 = 0.351 (see Figure 4.4(b)). This SV partition is used in the present work.
Partitions for tetrahedra
In the classical SV framework, the partition for a tetrahedral cell is fully
defined by the following three parameters:






|AE|
|AG|
1
2
3
|AC|
, 3 =
and 3 =
. (4.24)
0,
0,
0,
3 =
|AB|
2
|AD|
3
|AF |
4
Depending on the criteria for the selection of the CVs, third-order SV
scheme for tetrahedral cells allows four possible families of partitions which
are included in Figures 4.5(a), 4.5(b), 4.5(c) and 4.5(d). The first family of
partitions, Figure 4.5(a), was considered in Liu et al. [104] and Chen [32].
It is a simplified case, where it is imposed that the internal faces of the
corner CVs are planar. This way, two free parameters remain, namely 3
and 3 , with the third parameter 3 defined by
3 =

33 3
.
43 3
50

(4.25)

4.1. SPECTRAL VOLUME METHOD

The other two parameters, namely 3 and 3 , should satisfy the following
3
to satconstrains: 3 < 43 for 3 to be greater than zero, and 3 44
3 +1
3
isfy 3 4 . For the second and the third family of partitions, shown in
Figures 4.5(b) and 4.5(c) respectively, all parameters can be chosen freely.
The internal faces of the corner CVs are subdivided into two triangles. The
two families differ in the way these faces are subdivided. For the last family of partitions, shown in Figure 4.5(d), the internal faces of the corner
CVs are treated as a single bilinear quadrilateral face, and no additional
edge is introduced.

(a) First family.

(b) Second family.

(c) Third family.

(d) Fourth family.

Figure 4.5: SV partition families of a tetrahedral cell for third-order (p = 2) scheme

[175].

In Van den Abeele et al. [175], an extensive study of the variation of

51

CHAPTER 4. SPATIAL DISCRETIZATION

the stability properties in the three-dimensional parameter space corresponding to these schemes was performed. The above study indicates that
probably there is no partition that yields a stable third-order SV scheme
for tetrahedral cells. In 2010, Harris and Wang [66] have presented a
constrained minimization approach in the design of 3D third-order SV
schemes. Several new partitions have been proposed which have a reduced maximum real part of the Fourier footprint by up to 20% over the
original un-optimized partition proposed by Chen [32]. Numerical simulations have shown that the strength of the instability has been weakened by
about an order of magnitude for some cases by employing the constrained
minimization approach. However, also in this case, no fully stable thirdorder SV scheme for tetrahedral cells has been found to date.

4.2 Spectral difference method

An alternative to the SV method, is the spectral difference (SD) method. It
is also designed for high-order accuracy on unstructured grids. The SD
method originated in the staggered grid multi-domain spectral method
[93, 94]. It was generalized to simplex elements by Liu et al. [103]. Extension of the SD method to the Euler equations was described in Wang et
al. [189], and to the N-S equations in May and Jameson [111]. A 3D N-S
implementation of the SD method for hexahedral grids was presented in
Sun et al. [165]. In 2010, Parsani et al. [124, 126] coupled a high-order
SD scheme on unstructured hexahedral grids with the Wall-Adapted Local
Eddy-viscosity (WALE) model to perform LES.
The SD method has an important advantage over the DG and SV methods, in that no integrals have to be evaluated to compute the residuals,
thus avoiding the need for costly high-order integral evaluations.

4.2.1

Discretization of convective term

Consider a problem governed by the system of conservation laws given by

Equation (4.3) valid on a domain with boundary defined by . For each
~ with the transformation to the
cell i, take a mapped coordinate system ,
global Cartesian coordinate system given by (4.8), with Jacobian matrix
~
~J i and Jacobian determinant Ji . The convective fluxes projected in the

mapped coordinate system (~fC,i ) are thus related to the flux components in
52

4.2. SPECTRAL DIFFERENCE METHOD

the global coordinate system by
~

f
fC,i

C,i
~
~
~
1
~f = g = Ji ~J g
~
= Ji ~~J 1
C,i
C,i
i fC,i .
i
C,i
~
hC,i
hC,i

(4.26)

Therefore, system (4.3) can be written in the mapped coordinate system as

~

gC,i
hC,i
fC,i
~
(Ji w)
wi
~ ~ ~fC,i ,

=
t
t
1
2
3

(4.27)

with wi Ji w the conservative variables in the mapped coordinate system.

For a (p + 1)-th-order accurate dim-dimensional scheme, N s solution points
with index j are introduced at positions ~js in each cell i, with N s given by
(4.1) for simplex cells and by (4.2) for tensor product cells. Given the values
at these points, a polynomial of degree p that approximates the solution in
cell i, can be constructed. This polynomial is called solution polynomial,
and the conserved variables Wi,j at the solution points are the solution
variables of the SD method. The evolution of these variables is governed
by (4.27), evaluated at the solution points.
~

~ ~ ~f
To estimate the divergence of the mapped fluxes
C,i at the solution
points, a set of N f flux points with index l and at positions ~lf , supporting
a polynomial of degree p+1, is introduced. The evolution of the mapped flux
~
~

~ , which
vector ~fC in cell i is then approximated by a flux polynomial F
C,i
is obtained by reconstructing the solution variables at the flux points and
~
~ at these points. To ensure a sufficient coupling
evaluating the fluxes F
C,i
between the cells, a number of flux points need to lie on the faces or the
corners of the cell [103, 175, 189]. Examples of solution and flux point distributions are shown in Figure 4.6 for 1D, 2D triangular and 2D quadrilateral cells and for linear (p = 1) reconstruction.

In order to maintain conservation at a cell level, the flux component normal to a face must be continuous between two neighboring cells. Since
the solution at a face is in general not continuous, this requires the introduction of approximate Riemann solvers at those points. Two different approaches were discussed in Wang et al. [189]. The first approach
53

CHAPTER 4. SPATIAL DISCRETIZATION

(a) 1D SD.

0.866

0.433

0
0

0.5

(b) 2D SD triangular cell.

1
1

(c) 2D SD quadrilateral cell.

Figure 4.6: General solution () and flux points (N) distribution for second-order
(p = 1) 1D and 2D SD schemes.

54

4.2. SPECTRAL DIFFERENCE METHOD

involves the definition of multi-dimensional Riemann solvers, while the
second one uses multiple 1D Riemann solvers. One can find more information in the reference mentioned above. Here, the the first approach is used
~
~ in the
~ ~ F
[126, 129, 178]. Taking the divergence of the flux polynomial
C,i
solution points results in the following modified form of (4.27), describing
the evolution of the conservative variables in the solution points:


~ 
dWi,j
~  = Ri,j ,
~ C,i  = 1
~ ~ F
~ F
=
(4.28)
C,i 
dt
Ji,j
j
j
where Ri,j is the SD residual associated to Wi,j . This is a system of ordinary differential equations, in time, for the unknowns Wi,j , which can be
solved numerically using any method for such a system.

Notice that for p = 0, the SD method, like the DG and SV methods, reduces to the classical first-order accurate FV method. Furthermore, no
special treatment is required for cells with a non-linear transformation to
~
the mapped coordinate system, since the modified values of ~J i and Ji can
just inserted into expression (4.28).

4.2.2 SD basis polynomials

s
For a dim-dimensional (p + 1)-th order scheme,
 the N solution points support a set of solution basis polynomials Lsj ~ of degree p (SD solution
basis polynomials). Using these polynomials, the solution in cell i is approximated with a p-th order polynomial as follows
Ns
 
  X
Wi,j Lsj ~ .
Wi ~ =

(4.29)

j=1

In order to solve the reconstruction problem, one should recall that the solution polynomial at solution points must be equal to the solution variable,
i.e.
 
j = 1, ..., N s .
(4.30)
Wi ~js = Wi,j ,
Inserting expression (4.29) into identity (4.30) results in a set of linear
algebraic systems that define the SD basis polynomials:
 
s
= jm ,
j, m = 1, ..., N s .
(4.31)
Lsj ~m
55

CHAPTER 4. SPATIAL DISCRETIZATION

 
The N f flux points support a set of flux basis polynomials Lfl ~ of degree p + 1 (SD flux basis polynomials). The degree p + 1 flux polynomial in
cell i is then given by
Nf
 
X
~
~  
~ Lf ~ .
~
~
F
FC,i =
C,i,l l

(4.32)

l=1

In this case, the reconstruction problem is solved by imposing that the flux
polynomial at flux points must be equal to the flux at the flux points. Consequently, the flux basis polynomials are defined by a set of linear algebraic
systems of the form
 
f
= lm ,
l, m = 1, ..., N f .
(4.33)
Lfl ~m
Unlike the SV basis polynomials, the SD solution and flux basis polynomials are always independent of local geometry of the cell, and thus of the
cell index i. Consequently, independently of the type of cell, both SD solution and flux basis polynomials are the same for all cells and only two
sets of polynomial coefficients have to be precomputed and stored before
the actual simulation.

4.2.3

Discretization of diffusive terms

The treatment of the diffusive terms with the SD method is also derived
from approaches that were developed for the DG method. In this thesis,
the second approach of Bassi and Rebay (BR2) [17] is considered. This approach was reported for the first time in combination with the SD method
by Van den Abeele et al. [178].


~
. Defining the vectors J~i1 ,
Consider the diffusive flux vector ~fD w, w
J~i2 and J~i3 as
T

~
,
(4.34)
Ji ~J 1
= J~i1 J~i2 J~i3
i
~ at the solution points is
a polynomial approximation of the gradients w
obtained by
"
#

i J~ 1
i J~ 2
i J~ 3
1

W

i
i
i
~ 
~ i,j ,
w
=
(4.35)
+
+
Ji,j
1
2
3
i,j
j

56

4.2. SPECTRAL DIFFERENCE METHOD

i is a polynomial of degree p + 1, defined by its values at the flux

where W
points. In an internal flux point, this is just the value of the polynomial
of the two available
Wi . In a face flux point, it is equal to an average W,
values WL and WR :
= WL + WR .
(4.36)
W
2
 
~ i ~ , given by
The gradients in cell i are then approximated by
Ns

 
  X

~
~ i,j Lsj ~ .
~
~

w i =
i

(4.37)

j=1



~ i in an interThe diffusive flux vector is thus approximated as ~fD Wi ,
~ L and
~ R , of
~ are available. An
nal flux point. At a face, two values,

~
averaged value defined as
~
~
~
~
~ = WL + WR + L + R

2
2

(4.38)

~ L and
~ R associated to a cell-face
is used. In (4.38) the lifting operators
are again polynomials in the neighbouring cells, defined by their values at
the solution points:
~ L(R),j =

1
JL(R),j

~ ~ WL(R)

(4.39)

In this expression, WL(R) is a polynomial of degree p + 1, defined by its

values at the flux points:


T 



~
(W W )  J ~J 1
~1n~
~1n l curr. face
R,l
L,l


WL(R),l =
L(R),l

0
elsewise
(4.40)
The BR2 approach is fully compact, as only the immediate neighbors are
required for the computation of the residuals in a cell. Note that for p = 0,
~ L = W
~ R=
the gradients of the solution polynomials are always zero (W
0), and the gradients at a face are approximated by the lifting operators
57

CHAPTER 4. SPATIAL DISCRETIZATION

alone, which means that must be equal to one for consistency [90].
The normal component
of the diffusive flux vector is thus evaluated as

~fD WL +WR ,
~ ~1n in a face flux point. The tangential component is com2
~ i.
puted using the internal or the averaged value of
Notice that, if one is interested in the local SD approach (LSD), the def and
~ are identical to those discussed for the LSV approach,
initions of W
namely Equation (4.16) and Equation (4.18).

4.2.4

Component-wise flux point distribution

The flux point distribution for quadrilateral cells shown in Figure 4.6(c) is
not the one that is used in practical implementations of the SD method.
Instead, cells with component-wise flux point distributions, like the ones
shown in Figure 4.7, is used, [93, 94, 133, 164, 178].

1
1

Figure 4.7: Quadrilateral SD cells with general component-wise flux point distribution for second-order (p = 1) scheme. Solution (), 1 - (H) and 2 -flux points (N).

Different set of flux points are used for different components of the mapped
~
flux vector. For the general fi -component, a set of flux points that supports
~

a polynomial Fi of degree (p + 1) in 1 and of degree p in 2 (and in 3 , in

the case of hexahedral cells) is defined. These flux points are labeled 1 58

4.2. SPECTRAL DIFFERENCE METHOD

~

flux points and are used for the approximation of the derivative of fi to 1 .
~

For the gi - and hi -components, 2 - and 3 -flux points, supporting poly~

nomials Gi and Hi , are introduced analogously for the computation of the

derivatives to 2 and 3 respectively.
The main advantage of using component-wise flux point distributions is
that the reconstruction of the solution at the flux points and the computation of the flux derivatives at the solution points become one-dimensional
operations if the solution and flux points are aligned correctly, as shown
in Figure 4.7. This implies significant savings in computational time with
respect to the multi-dimensional operations that are required for general
quadrilateral cell SD schemes of the type shown in 4.6(c). Also notice that,
since each set of flux points is responsible for a single component of the
mapped flux, only the flux component normal to the face has to be computed at face flux points. Thus, a single one-dimensional Riemann flux
for the convective terms and the normal flux component for the diffusive
terms are used at face flux points. In this case, there are no corner flux
points.
Obviously, the use of component-wise flux point distributions is only possible for quadrilateral and hexahedral cells.

4.2.5 Solution and flux points distribution

In 2008, in Van den Abeele et al. [177] showed an interesting property of
the SD method, namely that it is independent of the positions of its solution
points in most general circumstances, for both simplex and tensor-product
cells. In the above work it has been shown that the distribution of the solution points has very little influence on the properties of the SD schemes,
and in fact, for linear problems, different distributions lead to identical results. This property greatly simplifies the design of SD schemes, since only
the flux point distributions have to be specified. It also implies an important improvement in efficiency, since the solution points can be placed at
flux points and thus a significant number of solution reconstructions can
be avoided.
1D cells
Respecting the symmetries of the cells and the minimum number of flux
points at a face, the flux point distributions of second-order SD scheme is
uniquely defined. Moreover, recalling that the SD method is independent
59

CHAPTER 4. SPATIAL DISCRETIZATION

of the positions of its solution points, the practical implementation of the
solution and flux points distribution is shown in Figure 4.8. Its stability
was confirmed by the wave propagation analysis shown in Van den Abeele
et al. [177].

Figure 4.8: Solution point distribution for second-order (p = 1) 1D SD cells, with

all solution points at flux points. Solution () and flux points (N).

Recently, Huynh [79] has proven that using the Legendre-Gauss quadrature points and the two end points of the cell as flux points results in stable
1D SD schemes of arbitrary orders of accuracy. Practical implementation
of higher-order 1D SD schemes is carried out respecting the symmetries of
the cell and placing most the solution points at flux points. For example,
the distribution of the points for third-order SD scheme is shown in Figure 4.9. In this figure, the free parameter 3 ]0, +1[, which was already
introduced for the 1D SV scheme, is also indicated. In order to obtain the
Legendre-Gauss quadrature points, 3 is set to 13 0.577.

0.58
3

0.58
+3

Figure 4.9: Solution point distribution for third-order (p = 2) 1D SD cells, with

symmetric and most solution points at flux points. Solution () and flux points (N).

Stability and accuracy properties of the 1D SD schemes proposed by Huynh

[79] was verified by Van den Abeele [46] with orders of accuracy up to
thirty.
2D triangular cells
Third-order SD schemes for triangular cells with different Riemann flux
approaches were examined using the wave propagation analysis in Van
den Abeele et al. [177]. The analysis indicates that no stable flux point
60

4.2. SPECTRAL DIFFERENCE METHOD

distribution for such schemes exists, with neither the semi-upwind nor the
full-upwind Riemann flux approach. The correctness of the analysis was
verified with numerical tests.
2D quadrilateral and 3D hexahedral cells
In Section 4.2.4 the component-wise flux point distributions has been discussed. The main advantage of that approach is that the reconstruction
of the solution at the flux points and the computation of the flux derivatives at the solution points become one-dimensional operations if the solution are aligned with the flux points. Following this idea, SD schemes
for quadrilateral and hexahedral cells are entirely defined by a 1D SD
cell, which is commonly labelled 1D source cell. Moreover, in Huynh [79]
it was proven that for quadrilateral and hexahedral cells, tensor product
flux point distributions based on a 1D flux point distribution consisting of
the end points and the Legendre-Gauss quadrature points, lead to stable
schemes for arbitrary values of p.
In Figures 4.10 and 4.11, symmetric second- and third-order quadrilateral locally 1D SD cells, with component-wise flux point distributions, are
illustrated.

1
1

Figure 4.10: Symmetric second-order (p = 1) quadrilateral locally 1D SD cells,

with component-wise flux point distributions. All solution points at flux points.
Solution (), - (H) and -flux points (N).

61

CHAPTER 4. SPATIAL DISCRETIZATION

1
1

0.58

Figure 4.11: Symmetric third-order (p = 2) quadrilateral locally 1D SD cells, with

component-wise flux point distributions. Most solution points at flux points. Solution (), - (H) and -flux points (N).

Always stable higher-order SD schemes for quadrilateral and hexahedral

cells are derived from their stable 1D counterparts [79].

4.2.6

Grid filter width for the subgrid-scale model

In Section 3.2.2 it has been shown that in the WALE model the grid filter
width is used to compute the turbulent eddy-viscosity, i.e. Equation
(3.48). In general is an unknown function and it has to be defined to
have a closed subgrid-scale model. Often, the grid filter width is taken to
be proportional to the smallest resolvable length scale of the discretization
and for a general cell with index i is usually approximated by
i =

dim
Y

hk

k=1

!1/dim

(4.41)

where hk is the size of the cell in the kdirection. However, at the face flux
points, two values of are available, labeled L and R . Consequently,
an averaged value for the filter width is generally used,
i = L + R .

2
62

(4.42)

4.2. SPECTRAL DIFFERENCE METHOD

For classical FV methods, Equation (4.42) uniquely defines the grid filter
width because for these schemes the flux points always lie at the cell face.
However, the same reasoning applied to a general SD method implies a
natural formulation of the grid filter width based on the Jacobian determinant of the transformation defined by Equation (4.8). In addition, since
in the SD scheme each cell has interior solution points and a high-order
polynomial approximation occurs in the cell, it is natural to choose the filter width depending on the order of the polynomial. For the SD method,
the order of the polynomial reconstruction is taken into account through
the division of the Jacobian determinant by the number of solution points,
i.e. N s (p, dim). Consequently, for each cell with index i and each flux
points with index l and positions fl , the following definition of filter width
is given:
"
 !#1/dim 
1/dim

Ji,l
1
~

~
=
det J i 
.
(4.43)
i,l =
Ns
Ns
f
l

Equation (4.43) uniquely defines the filter width for the internal flux points
but for the face flux points two values of the Jacobians determinant are
available, labeled JL,l and JR,l . Consequently, an averaged value is again
used, i.e.

i,l

det

 !

~
~J L 
+ det

fl

2 Ns

 ! 1/dim

~
~J R 

1/dim
 f

JL,l + JR,l
l

=
.

2 Ns

(4.44)

Notice that with this approach, the cell filter width is not constant in one
cell, but it varies because the Jacobian matrix is a function of the positions of the flux points. Moreover, for a given mesh, the number of solution
points depends on the order of the SD scheme, so that the grid filter width
varies by varying the order of the scheme. The proposed approach is valid
for high-order approaches and it is consistent because the filter width is a
function of the the polynomial order through the number of solution point.
In fact, the grid filter width decreases by increasing the polynomial order
of the approximation.
This new definition for the grid filter width with high-order SD scheme
was proposed in Parsani et al. [126], for 2D large eddy simulations with
63

CHAPTER 4. SPATIAL DISCRETIZATION

SD method (SD-LES). Recently, Parsani et al. [122, 123] applied the SDLES approach to 3D LES in a muffler.

4.3 Concluding remarks

In the previous sections, the mathematical formulations of the SV and SD
methods for a general system of conservation laws have been discussed.
With the SV method it has been shown that the basis polynomials depend on the local geometry of the corresponding cell, i.e. the CVs partition. Therefore, the polynomials coefficients should either be stored for
each cell, or recomputed on the fly when needed. The first approach requires a prohibitively large amount of memory, while the second uses too
much computational time. However, for grids with simplex cells, the transformation is linear for all the cells with straight edges and the SV basis
polynomials are the same for each element. Thus, only one set of polynomial coefficients can be precomputed and stored before the simulation.
For grids with quadrilateral or hexahedral cells, the transformation to the
mapped coordinates is linear only for respectively parallelograms or parallelepipeds. However, these type of cells are not flexible enough to mesh a
general geometry, and consequently, the implementation of the SV method
is limited to simplex cells. It has been pointed out that the SV method
is not uniquely defined for orders of accuracy higher than two. In fact,
like the SD method, the SV method has a certain number of parameters
that must be specified. These parameters define the partition of cells into
CVs and they affect strongly the stability and accuracy properties of the
method. For 1D and 2D SV schemes several stable and accurate partitions were proposed in literature. However, for the 3D case, no fully stable
third-order SV scheme for tetrahedral cells has been yet found and weak
instabilities have been discovered in all partitions for tetrahedra reported
in literature. Consequently, the application of 3D high-order SV methods to general flow problems is still an issue. These schemes are not yet
robust. Therefore, they can not be routinely employed in practical calculations, such as LES.
Unlike the SV basis polynomials, the SD solution and flux basis polynomials are always independent of the transformation to the mapped coordinate system, and consequently of the cell index. Therefore, independently
of the type of cell, both SD solution and flux basis polynomials are the same
for all cells and only two sets of polynomial coefficients have to be precomputed and stored before the simulation. The wave propagation analysis of
64

4.3. CONCLUDING REMARKS

the third-order SD schemes for triangular cells reported in literature indicates that no stable flux point distribution for such schemes exists. On the
contrary, for quadrilateral and hexahedral cells, tensor product flux point
distributions based on a 1D flux point distribution consisting of the end
points and the Legendre-Gauss quadrature points, was shown to lead to
stable schemes for arbitrary values of the order of accuracy. Consequently,
always stable higher-order SD schemes for quadrilateral and hexahedral
cells can be derived from their stable 1D counterparts.
Based on the important results summarized above, the coupling between
the SV method and the non-linear LU-SGS algorithm, which will be presented in Chapter 5, has been done only for 2D SV scheme and steadystate computations. In fact, this PhD work has been accomplished in the
framework of the IWT Project SBO 050163 (Simulation and design tools
towards the reduction of aerodynamic noise in confined flows), and its final goal was the development of an efficient N-S/LES solver for high-order
accurate schemes. In order to achieve the ultimate objective, since no stable 3D SV partition seems to exist, the author has mainly focused on the
promising high-order SD method. In this context, the SD method has been
coupled for the first time with the LES approach (SD-LES), through a new
definition of a grid filter width, as already discussed in Section 4.2.6.

65

CHAPTER 4. SPATIAL DISCRETIZATION

66

Chapter 5

Time discretization
In the previous chapter, the spectral volume and spectral difference methods have been used for the spatial discretization of the equations of fluid
dynamics. It has been shown that the spatial derivatives are no longer
stated explicitly in terms of the spatial independent variables and a system of ODEs in time has been obtained. This system approximates the
original set of PDEs and can theoretically be solved with any time marching scheme.
High-order explicit strong stability preserving (SSP) or total variation diminishing (TVD) Runge-Kutta (RK) schemes, developed by Shu [159], Shu
and Osher [119] and Gottlieb and Shu[58], were introduced for DG space
discretizations by Cockburn et al. [37, 39, 41, 43]. The same explicit methods were used for time marching SV and SD discretizations, respectively
by Wang et al. [183, 186, 191] and Liu et al. [103]. May and Jameson [111]
coupled their SD code with the popular four-stage fourth-order accurate
Runge-Kutta scheme [87]. The motivations for using these matrix-free explicit methods are high-order of accuracy in time, simplicity in implementation, and low computational cost per time step. However, explicit methods are subjected to severe Courant-Friedrichs-Lewy (CFL) stability limitations. In fact, with spatially high-order methods, the systems of ODEs
are generally stiff, and the stiffness increases with the order of the method.
Atkins and Shu [10] demonstrated that the increase of the order of accuracy of the DG method or the degree of the polynomials that are used
for the expansion in the local finite element space rapidly lowers the maximum CFL number and causes increasingly more stringent time step lim67

CHAPTER 5. TIME DISCRETIZATION

itations. Analogous limitations were encountered with the SV method by
Kannan et al. [91], Haga et al. [63] and Parsani et al. [121, 128, 130] and
with the SD method by Sun et al. [167] and Premasuthan et al. [133].
The CFL stability for spatially high-order discretizations is significantly
lower than the CFL stability of equivalent-order finite volume or finite
difference high-order methods. As a result, the computational effort of explicit methods for solutions of problems that require unstructured meshes
with spatially high-order discretizations is so high that the overall numerical scheme often becomes inefficient. This limitation is more dramatic
in case of viscous problems, where the solver must also be able to deal
with the geometrical stiffness imposed by the Navier-Stokes grids where
high-aspect ratios occur near walls. This is called grid clustering and it
increases with the Reynolds number. In the case of compressible solvers
there is an additional stiffness when solving for low speed flows caused by
the disparate eigenvalues of the system.
Implicit time integration schemes can be used to deal with these problems. These schemes can advance the solution with significantly larger
time steps compared to explicit methods. Two implicit time marching
schemes, namely the backward Euler (BE) scheme and second-order backward difference formula (BDF2) with variable time step, are used here for
the time discretization. The BE scheme, which is first-order accurate in
time, is used to solve steady flow problems, whereas the BDF2 is used for
unsteady flow simulations. Both schemes are A-stable, so that if they are
used to solve the linear Cauchy problem
 dy
dt = y(t), t > 0
(5.1)
y(0) = y0 ,
where is a complex number with negative real part ( C, R < 0),
their numerical solutions at time level tn , i.e. Y n , satisfy the condition
|Y n | 0, for tn [135]. This is a very important property because
the analytical solution of the Cauchy problem (5.1) is an exponential solution y (t) = et that goes to zero for t . The stability region of A-stable
schemes is then the complete left part of the complex plane, namely


A = z = tn C .

The A-stability is an important property of the time marching scheme for

solving systems of stiff ODEs, such as the systems arising from the spatial
discretization with high-order accurate schemes. In fact, the Fourier foot68

print1 of a spatial scheme, which gives an indication of the stiffness of the

problem, grows fast with increasing order of the method. Given a stable
spatial discretization, the A-stability guarantees the stability for the combination of the spatial scheme and the time marching scheme for any order
of accuracy of the space method. Notice that both BE and BDF2 schemes
are also L-stable, i.e. they are A-stable and their amplification factors go
to zero for z [135]. For more details about the Fourier footprint of
a spatial scheme and its stability condition, the stability region of a temporal discretization and the global stability condition, the reader is referred
to Appendix A.
Implicit temporal schemes imply the solution of non-linear algebraic systems. If the algorithm for such systems is not efficient then the schemes
may be more expensive than explicit ones. Recently, there has been some
research on RK schemes and backward difference formulae (BDF) [20, 30,
169]. In Bijl et al. [20], implicit Runge-Kutta (I-RK) solvers were investigated in combination with a standard cell-centered finite volume scheme
with artificial dissipation added for stability. It was observed that significant potential improvements in the temporal efficiency of implicit schemes
could be achieved from algebraic solver developments. In Swanson et al.
[169] the convergence of an explicit Runge-Kutta (E-RK) scheme with hmultigrid was accelerated by preconditioning with a fully implicit operator and the resulting RK/Implicit Residual scheme was used as a smoother
for an h-multigrid algorithm. It was demonstrated that the implicit preconditioner reduced the computational time of a well-tuned E-RK scheme
by a factor between four and ten. Both studies [20] and [169], concluded
that solver improvements can be more dramatic than improvements in integration techniques. For spatially high-order accurate methods, several
algorithms have been proposed and used to solve efficiently the non-linear
algebraic system. Element Jacobi methods were used by e.g. Helenbrook
and Atkins [69] for the DG method. Newton-Raphson GMRES solvers
with preconditioners were used in combination with DG schemes by Bassi
and Rebay [13], and in combination with SV and SD schemes by Van den
Abeele et al. [178, 179]. A matrix-free Krylov method was applied to DG
schemes by Rasetarinera and Hussaini [139] and to SD schemes by May
et al. [110]. Line-implicit solvers were combined with a DG method by
Fidkowski et al. [48].

1 The Fourier footprint is the collection of all eigenvalues in the complex plane for all wave
numbers of the modified dispersion relation which is obtained by wave propagation analysis
of a spatial scheme. It grows faster than linear with p.

69

CHAPTER 5. TIME DISCRETIZATION

In this work, an efficient algebraic algorithm, namely the non-linear lowerupper symmetric Gauss-Seidel algorithm (LU-SGS), is used to solve the
non-linear algebraic systems arising from the implicit time discretization.
The LU scheme was started by Jameson and Turkel [88] and later reformulated to use symmetric Gauss-Seidel by Jameson and Yoon [89] in the
context of second-order central schemes. It was recently rediscovered by
Sun et al. [167, 168] and adapted for use with SD schemes for steady state
computations. The first application of the non-linear LU-SGS in combination with the SV scheme and a full p-multigrid algorithm was reported
by Parsani et al. [128]. Applications of the non-linear LU-SGS in combination with p-multigrid to the SD method were reported by Premasuthan
et al. [133] and May et al. [110]. In 2009, in Parsani et al. [129] the
non-linear LU-SGS algorithm in combination with the BDF2 was coupled
with the SD method to solve 2D unsteady Navier-Stokes equations for laminar/turbulent test cases. More recently, in Parsani et al. [122, 124, 126],
the algorithm was used in combination with the SD method and an LES
approach for the simulation of 2D and 3D turbulent flows.
The formulation of the non-linear LU-SGS algorithm coupled with the
BE scheme and the BDF2 is described in the following sections. Extensions to other implicit time marching schemes is possible. The interested
reader is referred to Appendix D, where the non-linear LU-SGS algorithm
is used in combination with explicit-first-stage, single-diagonal-coefficient,
diagonally-implicit Runge-Kutta (ESDIRK) schemes. The coupling of the
non-linear LU-SGS solver with ESDIRK schemes was presented by Parsani
et al. [128] for quasi-1D Euler equations.

5.1 Backward Euler scheme

Consider the semi-discrete form of a system of conservation laws associated to a general spatial discretization :
Wcc
= Rcc (Wcc (t), {Wnb (t)}) = Rcc (W(t)) ,
t

(5.2)

where subscripts cc and nb denote the current cell and the neighbouring
cells that contribute to the residual Rcc respectively. For example, this system can be the semi-discrete form (4.7) or (4.28) with the inclusion of the
discretized diffusive flux, as shown in Sections 4.1.3 and 4.2.3. Approxi70

5.1. BACKWARD EULER SCHEME

mate the temporal derivative with the BE scheme

i

Wn+1
Wncc h
cc
Rcc Wn+1 Rcc (Wn ) = Rcc (Wn ) ,
t

(5.3)

where superscripts n and n+1 denote the time levels tn and tn+1 , and t =
tn+1 tn is the time step. Let Wcc = Wn+1
Wncc and Rcc (Wncc ) = Rncc .
cc
Linearizing the residual using the Taylor expansion about Wncc gives
Rn+1
Rncc
cc

n
X Rcc n
Rcc 
 Wnb .
W
+
cc
Wcc 
Wnb 

(5.4)

nb

Therefore, the fully linearized equations for (5.2) can be written as



n


X Rcc n
Rcc 
I
 Wnb = Rncc ,

Wcc
+
Wcc 
t
Wnb 

(5.5)

nb

where I is the identity matrix. However, Equation (5.5) requires too much
memory to store the LHS implicit Jacobian matrices. Consequently, the
most recent solution for the nb cells is used,


n


X Rcc n
I
Rcc 
m+1
n
 Wnb ,
W
=
R
+
+
cc
cc
Wcc 
t
Wnb 

(5.6)

nb

with superscript denoting the most recent solution when doing forward
and backward sweeps. Superscript m + 1 refers to the actual symmetric
Gauss-Seidel (SGS) sweep, i.e. Wm+1
= Wn+1,m+1
Wncc .
cc
cc
In Swanson et al. [169], Equation (5.6) is solved with a first order upwind scheme with the Jacobians calculated each time step and not stored.
However, this may not be an efficient solution for high-order methods beRcc
is time-consuming.
cause the computation of the Jacobian W
nb
To avoid the computation and the storage of the off-diagonal block matrices, expression (5.6) is further manipulated as follows [128, 168]:
Rncc +

X Rcc n
 Wnb
Wnb 
nb

Rcc (Wcc , {Wnb })

= Rcc
71

n
Rcc 
Wcc
Wcc 

n
Rcc 
Wcc .
Wcc 

(5.7)

CHAPTER 5. TIME DISCRETIZATION

m+1

n+1,m+1
^
Wm+1
Wcc = Wcc
Wcc . Consequently, comLet W
cc
cc
bining (5.7) with (5.6) results in expression (5.8)
n


Wcc
Rcc 
I ^ m+1

=
R

+
W
.
(5.8)
cc
cc
Wcc 
t
t

This implicit solver is denoted as LU-SGS + BE in the remainder of the

work, where BE stands for backward Euler scheme. The linear system
(5.8) is solved with multiple cell-wise symmetric forward and backward
m+1
^
and/or a maxsweeps with a prescribed tolerance on the change W
cc
imum number of symmetric forward and backward sweeps. The inverse of
the small Jacobian matrices in the left hand side of this expression is computed using a LU decomposition with pivoting at the beginning of each
time iteration, which makes the solution of the small linear algebraic systems much more efficient during subsequent SGS sweeps. If the time step
t is not too large, it is possible to freeze the cell-wise Jacobian matrices
during a certain number of time iterations, which also leads to an increase
in efficiency. Note that solving Equation (5.8) to machine zero implies that
Equation (5.2) is satisfied, i.e. the non-linear system of equations is solved
exactly. For this reason the algorithm is called non-linear LU-SGS algorithm. If Equation (5.2) is satisfied and the CFL number is smaller or
equal to one, the implicit LU-SGS + BE is actually a first-order accurate
physical time-advancing scheme. Because of the Gauss-Seidel nature of
this LU-SGS method, where the latest available solution in the neighboring cells is used to update the solution in a cell, information travels much
faster through the domain than with a traditional explicit solver, where
only the solution at the previous time step or stage is used to update the
solution in a cell. Besides the cell-wise implicitness of the algorithm, the
Gauss-Seidel nature is one reason why the LU-SGS method is much more
efficient than an explicit solver.
The initial guess for Wn+1
is Wncc . For steady flow problems, the L2 norm of
cc
the solution variation or right-hand side (RHS) of Equation (5.8) are monitored for convergence in this work. Moreover, for steady flow simulations,
it is not necessary to drive the RHS of Equation (5.8) to machine zero, but
it is more efficient to limit the maximum number of SGS sweeps to damp
high-frequency error components and/or set a threshold for some norm of
m+1
^
. More details about the values of the parameters
the change W
cc
used to solve the system (5.8) are given in Chapters 7 and 8 where the results of the numerical simulations are presented.
72

5.1. BACKWARD EULER SCHEME

Since only the diagonal block Jacobians are required to solve the system of
ODEs, the total number of real variables needed to store these Jacobians
on a grid with N tetrahedral cells with solution polynomial degree p is
N

2
(p + 1) (p + 2) (p + 3)
# physical variables .
6

(5.9)

For a grid with N hexahedral cells with solution polynomial degree p, this
number is
h
i2
3
N (p + 1) # physical variables .
(5.10)

From expressions (5.9) and (5.10), it is clear that the non-linear LU-SGS
method requires significantly less memory than classical methods that use
the full Jacobian matrix, like for instance the Newton-GMRES algorithm2
[92, 151]. Nevertheless, the required amount still increases with p to the
power six. For instance, in expression (5.9), the coefficient, which multiplies the number of physical variables, is 16, 100, 400 respectively for
p = 1, 2, 3.
In this study, the Jacobian matrices are numerically obtained using the
following numerical differencing
Rcc (Wcc + Wcc , Wnb ) Rcc (Wcc , Wnb )
Rcc
=
,
Wcc
Wcc

(5.11)

where Wcc is a small value computed as

Wcc = kWcc Wref k .

(5.12)

Wref and are respectively the reference conservative variables and a

small constant set equal to something larger than the square root of machine precision mach [92]. In the present work, with 64-bit double precision, was set to 105 . Notice that this numerical approach results in
straightforward operations for arbitrarily complex residual operator.
The cells numbering potentially has an important influence on the convergence of the scheme due to the Gauss-Seidel nature of the implicit LUSGS, see for instance Sharov and Nakahashi [158]. However, in practice
it is observed that the symmetric forward and backward sweeps strongly
2 For

more details about the Newton-GMRES method the reader is referred to Appendix C.

73

CHAPTER 5. TIME DISCRETIZATION

reduce the effect of the cells numbering and consequently only a small influence on the convergence is observed. For this reason, in the present
work, the order of the cells in the cell index list is not reordered with
any criteria or algorithm and it corresponds to the numbering of the mesh
generator. For instance, in the Gmsh software [54], the element numbers
are assigned by looping over geometrical entities of increasing dimensions
(points, then curves, then surfaces, then volumes) and numbering the cells
as they come in each entity.

5.2 Second-order backward difference formula

Consider again the system of ODEs (5.2). Discretizing the temporal derivative with the BDF2 with variable time step results in
1 + 2 n+1
2
W
(1 + ) Wn +
Wn1 = tn Rn+1 ,
1+
1+

(5.13)

t
where tn = tn+1 tn , = t
n1 . Expression (5.13) is again a non-linear
algebraic system, which has to be solved at each time iteration to find
the solution at the next iteration tn+1 starting from the two previous time
iterations tn and tn1 . Linearizing the residual about time iteration tn , one
obtains
!
n

X Rcc n
Rcc 
Wcc
 Wnb =
Wcc +
c1
tn
Wcc 
Wnb 
nb


n1
+ c1 Rncc ,
(5.14)
c2 Wncc Wcc
2

1+

with W = Wn+1 Wn , c1 = 1+2

and c2 = tn (1+2
) . Applying a SGS
algorithm to solve the linear algebraic system (5.14) results in

!

X Rcc n

 Wnb
+
Wnb 
nb


n1
,
(5.15)
+c2 Wncc Wcc

n

Rcc 
I
c1
Wm+1
= c1
+
cc
Wcc 
tn

Rncc

Inserting (5.7) into (5.15) and manipulating the resulting equation, one
74

5.3. TIME STEP

obtains the following final expression:
n




m+1
I
Rcc 
n
n1

^
W

W
+
c
=
c
R
+
W
c1
1
2
cc
cc
cc
cc
Wcc 
tn
Wcc

,
(5.16)
tn
m+1

^
where W
Wm+1
Wcc = Wn+1,m+1
Wcc . This implicit solver
cc
cc
cc
is denoted as LU-SGS + BDF2 in the remainder of the work. Note again
that, the LU-SGS algorithm acts directly on the non-linear algebraic system to be solved, i.e. the system. (5.13), which is the right hand side of
expression (5.16). The initial guess for Wn+1
is Wncc . At the first time itcc
eration (n = 0), since no solution at time level n = 1 is available, the
first-order BE scheme, namely Equation (5.8), is solved. Analogous properties and remarks indicated for the LU-SGS + BE are also valid for the
LU-SGS + BDF2.

5.3 Time step

The objective of time-accurate CFD calculations is the modeling of unsteady flow phenomena, such as vortex pairing or bluff body wakes. In
unsteady calculations, a global physical time step must be used throughout the entire flow-field, regardless of the local grid distribution. Because
for explicit schemes the CFL number has a limit which approximates the
upper limit of numerical stability (see Appendix A), the global time step
must be obtained from the maximum CFL number and the smallest grid
spacing in the field. Thus, the stability of the calculations in regions of
dense grid packing determines the time step for the entire flow-field, in
most of the cases. For example, if a wave is crossing a discrete grid, then
the time step must be less than the time for the wave to travel adjacent
grid points. Therefore, when the grid point separation is reduced, the upper limit for the time step also decreases. In essence, the numerical domain
of dependence must include the analytical domain of dependence in order
to assure that the scheme can access the information required to form the
solution [75, 99]. For the 1D linear convection equation with constant advection velocity a and on uniform grids , the stability limit on the time
step prescribed by the CFL condition is given by
t

C,max .
a

75

(5.17)

CHAPTER 5. TIME DISCRETIZATION

For 1D problems with constant diffusive coefficient D on uniform grids, a
similar stability limit does exist and is defined by
t

2
D,max .
D

(5.18)

Expressions (5.17) and (5.18) can be generalized to a multi-dimensional

problems on general grids with variable advection velocities and diffusive
coefficients:
i
C,max ,
(5.19)

t / P f ac 
Ni


~
S
a

1

~
m
m
n,m
m=1
and

2
t / P f ac i
D,max ,
Ni
2

S
m
m=1 D,m

(5.20)





where ~am ~1n,m  and D,m are respectively the maximum wave propagation velocity in the direction perpendicular to the cell face m and the maximum diffusive coefficient on the same face. For the N-S equations, D,m
is the maximum of kinematic viscosity and thermal diffusivity cP . For
general convection-diffusion equations, the global physical time step must
then be smaller or equal to the minimum between the maximum inviscid
time step (Equation (5.19)) and the viscous time step (Equation (5.20)) of
each cell. However, notice that for accuracy, a smaller value for t may be
required. In fact, the time step must also be chosen based on the physics
of the problem. This criterion is generally used for implicit time marching schemes which do not suffer from a restrictive CFL condition. Consequently, for implicit methods the time step usually does not change when
the grid is changed.
In case of steady computations, as one is not interested in the transient
behavior of the solution, a local time step can be chosen, whereby each
cell i progresses at its maximum possible time step ti , without locally
violating the stability conditions (5.19) and (5.20). This provides significant convergence acceleration at the cost of the transient time consistency
since each cell has its own time step. This approach is known as local time
stepping technique and can be applied with any explicit or implicit time
marching scheme.

76

Chapter 6

Analysis of the non-linear

LU-SGS algorithm
In this chapter, the stability and the smoothing properties of the LU-SGS
+ BE solver in combination with the SV and SD methods is analyzed. The
main goal is to investigate the behavior of the LU-SGS + BE solver for
a mesh with high-aspect ratio cells. These grids typically occur near the
walls to resolve the viscous boundary layer with an economical distribution of grid points.
The properties of the LU-SGS + BE are evaluated with a Von Neumann
analysis for a model 2D linear advection. The analysis is performed for
second- to fourth-order SV and SD schemes using the stable SV partitions
and the stable SD distributions of flux points presented in Chapter 4. For
more details about the methodology of this analysis technique, the reader
is referred to Appendix A. The SV results that are discussed in this chapter
are published in Parsani et al. [127, 130].

6.1 Summary of the methodology

The methodology of the analysis can be summarized as follows. Consider
the linear advection equation, discussed in Section 3.5, and given by
w ~
+ (~aw) = 0.
(6.1)
t
Equation (6.1) is discretized in space by introducing the spatial derivatives
corresponding to a spatial discretization method. A uniform grid with peri77

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

odic boundary conditions is considered. The grid is defined by a generating
pattern, which is the smallest part from which the full grid can be reconstructed by periodically repeating the pattern in all directions. For the
2D case, the generating patterns for uniform triangular and quadrilateral
meshes are shown in Figures 6.1(a) and 6.1(b) respectively. The gener~ 1 and B
~ 2 and their
ating pattern is completely defined by the vectors B
~ 1,
non-dimensional form is obtained by scaling them with the length of B

~
~
~
~
denoted by B: B1 B B1 and B2 B B2 . If the dimensionless vector
~ 1 is chosen as [1 0]T , then the dimensionless mesh is completely defined
B
~ 2 . In 2D, the advection speed ~a in Equation
by the two components of B

(a) Triangular mesh.

(b) Quadrilateral mesh.

Figure 6.1: Generating patterns for 2D uniform meshes.

(6.1) is defined by its amplitude a and orientation angle :



cos
~a = a
.
sin

(6.2)

After application of the space method to (6.1) on a uniform quadrilateral

or triangular mesh, the following system of ODEs is obtained:
dWi,j
a  0,0
T Wi,j + T1,0 Wi1,j + T0,1 Wi,j1
+
dt
B


+T+1,0 Wi+1,j + T0,+1 Wi,j+1 = 0,

(6.3)

where the five dimensionless matrices T are defined by the spatial discretization method. They depend on the generating pattern, the advection
speed orientation angle and the Riemann flux. The column vector Wi,j
contains the N s,GP solution variables in the generating pattern with indices i and j (from within one quadrilateral cell or two triangular cells).
78

6.1. SUMMARY OF THE METHODOLOGY

Inserting the following Fourier wave

(t) eI~k(iB~ 1 +j B~ 2 )B
Wi,j (t) = W
(t) eIK~1k (iB~ 1 +j B~ 2 ) ,
= W

(6.4)

into Equation (6.3) results in

dW
a  0,0
~ ~
~ ~
T + T1,0 eIK 1k B1 + T0,1 eIK 1k B2
+
dt
B

~ ~
~ ~
= 0,
+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 W
where I
given by

(6.5)

1 is the imaginary unit number and ~k is the wave vector

~k = k

cos
sin

= k ~1k .

(6.6)

Approximating the time derivative with the BE scheme, Equation (6.7) is

obtained,

n+1 W
n + T0,0 + T1,0 eIK~1k B~ 1 + T0,1 eIK~1k B~ 2
W
 n+1
~ ~
~ ~

+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 W

= 0, (6.7)
where =

a t
B

is the CFL number. An expression for the amplification

n+1 = Gd W
n , can be obtained from system (6.7):
matrix Gd , defined by W
h

~ ~
~ ~
Gd = I + T0,0 + T1,0 eIK 1k B1 + T0,1 eIK 1k B2
i1
~ ~
~ ~
.
(6.8)
+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2

The matrix Gd represents the amplification matrix of the direct inversion

method which is marked as direct + BE in this work.
Applying a Gauss-Seidel algorithm to (6.7), the expression for the amplification matrix of the LU-SGS algorithm can be found. In this case, because
of the Gauss-Seidel nature of the non-linear LU-SGS algorithm, where the
latest available solution in the neighboring cells is used to update the solution in a cell, the procedure to obtain the amplification matrix for the
LU-SGS depends on the generating pattern structure, through the matrix
T0,0 . In fact, the column vector Wi,j in (6.3) contains the N s,GP solution
79

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

variables of the generating pattern which corresponds to one quadrilateral
cell or two triangular cells, as illustrated in Figures 6.1(b) and 6.1(a).
Consider a mesh made by triangular cells. After application of the GaussSeidel algorithm, the amplification matrix of the first forward sweep is
found:
h

i1
~
~
0,1 IK~1k B
0,0
0,0
1,0 IK~1k B
1 + T
2
e
GTf,1R = I + T0,0
+
T
+
T
+
T
e
1,1
2,1
2,2
i
h

~
~
0,0
0,+1 IK~1k B
+1,0 IK~1k B
2
1 + T
,
(6.9)
e
I T1,2 + T
e

0,0
where T1,1
and T0,0
2,2 represent respectively the contribution to the residual
of the first and second cell of the generating pattern to themselves, while
0,0
0,0
T1,2
and T2,1
represent the cross contributions of both cells of the generating pattern. Matrix I is the unit or identity matrix.

The amplification matrix of the first SGS sweep is given by

i1
h

~
~
R
0,0
0,0
0,+1 IK~1k B
+1,0 IK~1k B
2
1 + T
e
GTSGS,1
= I + T0,0
+
T
+
T
+
T
e
1,1
1,2
2,2
i
h

~
~
0,1 IK~1k B
0,0
1,0 IK~1k B
1 + T
2
.
(6.10)
I T2,1 + T
e
e

R
From the amplification matrix GTSGS,1
, the amplification matrix of the m-th
SGS sweep may be computed using the following two-step recursive procedure:

1: Compute the amplification matrix of the m-th forward sweep

i1
h

~
~
R
0,0
0,0
1,0 IK~1k B
0,1 IK~1k B
1 + T
2
GTf,m
= I + T0,0
e
e
1,1 + T2,1 + T2,2 + T
i

h

~
~
TR
0,+1 IK~1k B
+1,0 IK~1k B
2
1 + T
G
e
I T0,0
+
T
e
SGS,m1 .
1,2

(6.11)

2: Compute the amplification matrix of the m-th SGS sweep

i1
h

~
~
R
0,0
0,0
0,+1 IK~1k B
+1,0 IK~1k B
2
1 + T
e
GTSGS,m
= I + T0,0
e
1,1 + T1,2 + T2,2 + T

i
h

~
~
TR
1,0 IK~1k B
0,1 IK~1k B
1 + T
2
G
I T0,0
+
T
e
e
f,m .
2,1

(6.12)

80

6.1. SUMMARY OF THE METHODOLOGY

With quadrilateral cells, the amplification matrices for the first forward
sweep and the first SGS sweep are given by Equations (6.13) and (6.14)
respectively.
i1
h

~ ~
~ ~
0,0
GQD
+ T1,0 eIK 1k B1 + T0,1 eIK 1k B2
f,1 = I + T
i
h

~ ~
~ ~
(6.13)
I T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2
i1
h

~
~
0,0
+1,0 IK~1k B
0,+1 IK~1k B
1 + T
2
=
I
+

T
+
T
e
GQD
e
SGS,1
i
h

~ ~
~ ~
I T1,0 eIK 1k B1 + T0,1 eIK 1k B2

(6.14)

In this case, the matrix T0,0 contains only the contributions of the N s,GP
solution variables of the generating pattern which corresponds to one quadrilateral cell. From the amplification matrix GQD
SGS,1 , the amplification matrix of m-th SGS sweep may be computed using again the following twostep recursive procedure:
1: Compute the amplification matrix of the m-th forward sweep
h

i1
~
~
0,0
1,0 IK~1k B
0,1 IK~1k B
1 + T
2
GQD
e
e
f,m = I + T1,1 + T
i

h

~ ~
~ ~
I +T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 GQD
SGS,m1 .

2: Compute the amplification matrix of the m-th SGS sweep

i1
h

~ ~
~ ~
0,0
GQD
+ T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2
SGS,m = I + T
i

h

~ ~
~ ~
I T1,0 eIK 1k B1 + T0,1 eIK 1k B2 GQD
f,m .

(6.15)

(6.16)

Let m (m = 1, 2, . . . , N s,GP ) be an eigenvalue of the general amplification

matrix G = G (K, , ) and = (G (K, , )) represents the eigenvalue
spectrum of G. Then g = g (K, , ) max | (G)| is the amplification factor for a given (K, , ). In order for the global discretization to be stable,
g 1 should be satisfied i.e. (G) lies inside the unit circle of the complex
plane (stability boundary) for all K, and . The range of K is one period
of the Equation (6.7) and it is marked PK in this work. For a fixed shape
of the generating pattern, PK is a function of , i.e. PK = PK (). The ex~,
pression from which PK can be computed, is obtained by substituting B
1
~ 2 and in the exponential terms of Equation (6.7) and using the Eulers
B
81

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

formula, which gives a relation between trigonometric functions and complex numbers. Following this reasoning, a trigonometric function f , which
has exactly the same period as Equation (6.7) is obtained. In general, it is
not possible to find a closed formulation for PK = PK (). Therefore, the
period PK is computed taking the inverse fast Fourier transformation of
the function f .
If the scheme is used as a smoother for multigrid, then it must have good
damping of high-frequency error components, i.e. it should cluster the amplification matrix eigenvalues corresponding to the high-frequency modes
towards the origin. Therefore, for high-frequency modes g 1 for all
and should be satisfied. In addition, we desire that the CFL number be
sufficiently large to produce significant reduction (if not elimination) of the
convergence slow-down effects that are associated with high-aspect ratio
mesh cells. A large CFL number also facilitates the expulsion of error components. At the same time the capability for large CFL numbers must not
compromise the high-frequency damping property of the scheme.
In the next sections, the damping properties of the LU-SGS + BE method
in combination with the SV and SD schemes will be analyzed, for a high
CFL number and a mesh with high aspect ratio cells. The latter choice
accounts for the effects of the the geometrical stiffness imposed by the
Navier-Stokes grids where high-aspect ratios occur near walls. However,
since the performance of the direct inversion method is better than that
of any approximate method, the direct inversion method is examined as a
baseline for comparison with the LU-SGS algorithm. An upwind Riemann
flux is employed and the direction of the wave propagation velocity is
set to 6 (a1 > 0, a2 < 0). A negative 2 velocity component is chosen
to avoid having the sweep directions of the LU-SGS method aligned with
the flow direction. In fact, if the flow direction corresponds to the sweep
directions, the LU-SGS becomes a direct solver, e.g. in case of a purely
upwind scheme. The smoothing properties of the direct + BE and the LUSGS + BE methods are shown for three values of the solution orientation
(or Fourier wave angle) , i.e 6 , 0 and 3 .

6.2 SV method for triangular cells

In this section the LU-SGS + BE solver in combination with second- to
fourth-order SV schemes is analyzed. In order to explain the procedure
used to construct the amplification factor as a function of the wave number,
82

6.2. SV METHOD FOR TRIANGULAR CELLS

first a grid with isotropic cells, and a small CFL number are considered.
This allows to have better understanding of the necessary steps to carry
out the task.

6.2.1 Second-order SV method

Consider a mesh built from equilateral triangles. Such a mesh is obtained
with the following choice for the dimensionless vectors B1 and B2 ,
B1

1
0

B2

1
2
3
2

(6.17)
|B |

and leads to a mesh with isotropic cells, i.e. the aspect ratio, AR = |B1 | = 1.
2
Figure 6.2 shows the eigenvalue spectrum of the amplification matrix of
the LU-SGS + BE method and the direct + BE method on this mesh, for
CFL= 1, varying the direction of the wave propagation velocity ~a, the
wave number K, and the solution orientation defined by the angle . For
the LU-SGS + BE one, two and three SGS sweeps were used. It is seen
that the global discretization is stable for all K, and . Figure 6.2 shows
that with an increasing number of SGS sweeps, the eigenvalue spectrum
of the LU-SGS + BE method approaches that of the direct + BE method.
The aim is to plot the amplification factor g as a function of the wave
number K. Hence, the following paragraphs first present the analysis of
the smoothing properties done for the direct + BE solver and describe the
methodology used to plot correctly the amplification factor g. First, consider a solution with an orientation = = 6 . For this choice, Equation
(6.7) is periodic in K with a period equal to 43 . However, Equation (6.7)
has N s,GP eigenvalues for each K, and each eigenvalues corresponds to a
wave number K + l 43 , with l an integer number. Consequently, to get
the correct damping properties of the solver, each mode should be shifted
by a multiple of 43 along the wave number axis. The actual wave number K + l 43 to which an eigenvalue m (m = 1, 2, . . . , N s,GP ) corresponds
has to be determined by examining the spatial shape of the eigenmodes
m eIK~1k (iB~ 1 +j B~ 2 ) , where V
m are the eigenvectors of the amplification maV
trix. The shape of these eigenmodes are shown in Figure 6.3 for K = l 3
with l = 0, . . . , 5.
In Figure 6.4 the eigenvalue spectrum (Gd ) is plotted as a function of
the wave number K, for = = 6 and CFL= 1. For small wave numbers
83

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

(a) 1 SGS sweep.

(b) 2 SGS sweeps.

(c) 3 SGS sweeps.

(d) direct + BE

Figure 6.2: Effect of the SGS sweeps on the eigenvalue spectrum of the LU-SGS +
BE scheme. Second-order (p = 1) SV method, [0, 2], [0, 2], K [0, PK ()],
AR = 1, CFL=1.

K, the shape of the eigenmodes reveals the correct wave number to which
an eigenvalue corresponds. In fact, three curves can readily be selected
by examining the spatial shape of the eigenvector (see Figure 6.3), namely
the curves marked by the plus (+), the square () and the circle () signs.
The other curves, marked by the () sign, correspond to eigenvalues which
belong to a higher wave number range of K compared to the eigenvalues
of the selected curves. From Figure 6.3 it is clear that the amplitude of
the discontinuities in the numerical solution at the cell interfaces grow
with increasing wave number K which makes it difficult to analyze the
eigenvectors and to select the actual wave number K + l 43 to which an
eigenvalue corresponds. Therefore, for a high-wave number, the effort of
84

6.2. SV METHOD FOR TRIANGULAR CELLS

4
6

6
10

10
6

0
0

10

12

14

16

0
0

6
4

14

16

10

12

14

16

12

14

16

(b) K =

12

(a) K = 0.

10

.
3

10

10
6

0
0

10

12

14

0
0

.
3

(d) K = 3

(c) K = 2

16

.
3

6
10

10
6

0
0

10

12

14

16

0
0

(e) K = 4

.
3

10

(f) K = 5

.
3

Figure 6.3: Eigenmodes of the direct + BE scheme. Second-order (p = 1) SV

method, = = 6 , AR = 1, CFL= 1, K = l 3 , l = 0, . . . , 5.

associating each eigenvalue with the corresponding wave number K has

not been performed. However, the eigenvalues which correspond to a high85

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

0.5
0.4

0.8

0.3

0.7

0.2

0.6

Im[(G(K))]

Re[(G(K))]

1
0.9

0.5
0.4

0.1
0
0.1

0.3

0.2

0.2

0.3

0.1

0.4

0
0

0.5
0

(a) Real part (R vs. K).

(b) Imaginary part (I vs. K).

Figure 6.4: Real and imaginary parts of the eigenvalue spectrum of the amplification matrix for the direct + BE scheme. Second-order (p = 1) SV method,
= = 6 , AR = 1, CFL= 1.

wave number K, thus the eigenvalues marked by the sign (), are well
damped as observed in Figure 6.4.
Therefore, shifting these curves in the appropriate way, leads to plots as in
Figure 6.5. Because of the symmetry (Re[(G (K))] = Re[(G (K))] and
Im[(G (K))] = Im[(G (K))]), the curves are only shown for positive
wave numbers K [0, 23 PK ].
In this case, the direct solver is a physically accurate scheme because
CFL= 1. Consequently, its amplification factor should closely follow the
exact amplification factor relation given by Equation (6.18),
Gexact =

w(t + t)
= eIkat .
w(t)

(6.18)

This expression can be approximated by a Taylor series as

Gexact 1 IKat

K 2 a2 t2
+ o(K 3 t3 ),
2

(6.19)

for wave numbers K close to zero (Kt 1). In Figures 6.5(a) and 6.5(b),
the real and the imaginary parts of the amplification factor of the direct
solver for Kt close to zero are plotted along with their counterparts in
Equation (6.19). It is seen that the direct + BE method follows closely the
86

6.2. SV METHOD FOR TRIANGULAR CELLS

1

0.9

x 10

Analytic

0.1

0.8
0.7

0
Direct solver

Direct solver

Im[g(G(K))]

Re[g(G(K))]

Analytic

0.6
0.5

0.2

0.3

0.4
0.3

0.4

0.2
0.1
0

0.5
0

10

(a) Real part (R vs. K).

10

(b) Imaginary part (I vs. K).

1
0.9
0.8

g(G(K))

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

10

(c) Amplification factor (g vs. K).

Figure 6.5: Real and imaginary parts of the shifted eigenvalues of the direct + BE
scheme and corresponding amplification factor shape. Second-order (p = 1) SV
method, = = 6 , AR = 1, CFL= 1.

theoretical results for Kt 1. In Figure 6.5(c) the amplification factor

g of the direct + BE inversion method is plotted as a function of the wave
number K [0, 23 PK ], for = = 6 . It is seen that the direct + BE solver
has good damping properties for high-frequency solution components.
The procedure described above may be applied to the amplification matrix of the LU-SGS + BE scheme. In Figure 6.6 the eigenvalue spectrum
of the iterative solver is plotted as a function of the wave number K, for
= = 6 , CFL= 1 and one SGS sweep. In this figure, three curves
87

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

1

0.6

0.9

0.5

0.8

0.4
0.3

0.7

Im[(G(K))]

Re[(G(K))]

0.2

0.6
0.5
0.4

0.1
0
0.1
0.2

0.3

0.3

0.2

0.4

0.1
0
0

0.5

0.6
0

(a) Real part (R vs. K).

(b) Imaginary part (I vs. K).

Figure 6.6: Real and imaginary parts of the eigenvalue spectrum of the amplification matrix for the LU-SGS + BE scheme. Second-order (p = 1) SV method,
= = 6 , AR = 1, CFL= 1, one SGS sweep.

marked with the plus sign (+), the square () and the circle () are shown
(cf. Figure 6.4). Shifting these curves in the appropriate way the amplification factor curve of Figure 6.7 for K [0, 32 PK ] is found. In the same plot,
the amplification factor for two and three sweeps is also shown. It is seen
1

Direct solver
1 SGS
2 SGS
3 SGS

0.9
0.8

g(G(K))

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0

10

12

Figure 6.7: Effect of the SGS sweeps on the amplification factor of the LU-SGS +
BE scheme. Second-order (p = 1) SV method, = = 6 , K [0, 23 PK ], AR = 1,
CFL= 1.

88

6.2. SV METHOD FOR TRIANGULAR CELLS

that the discontinuity, obtained with one SGS sweep, disappears when two
SGS sweeps are employed. For three SGS sweeps the amplification factor
of the LU-SGS + BE method is indistinguishable from that of the direct +
BE inversion method. This explains why for the analysis of a mesh with
AR = 1 and CFL= 1, a maximum number of three SGS sweeps was used.
Consider the same equilateral triangular mesh but with a CFL number
of 106 . The eigenvalue spectrum of the amplification matrix of the direct +
BE and the LU-SGS + BE methods is plotted in Figure 6.8, for this case,
varying again the direction of the wave propagation velocity ~a, the wave
number K and the angle . For the LU-SGS + BE method, one, ten and

(a) 1 SGS sweep.

(b) 10 SGS sweeps.

(c) 100 SGS sweeps.

(d) direct + BE.

Figure 6.8: Effect of the SGS sweeps on the eigenvalue spectrum of the LU-SGS +
BE scheme. Second-order (p = 1) SV scheme, [0, 2], [0, 2], K [0, PK ()],
AR = 1, CFL= 106 .

89

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

one hundred SGS sweeps are employed. It is seen that for CFL = 106 , the
LU-SGS + BE method is still stable for all K, and . However, using too
few SGS sweeps (1-3 sweeps), the scheme exhibits poor damping behavior
for some values of the angle and a range of frequencies K.
In Figures 6.9(a), 6.9(b) and 6.9(c) the amplification factor for = 6 ,
= 0 and = 3 is shown for CFL= 106 . In these figures, the amplification factor of the direct solver is also plotted for comparison. Figure
6.9(b) shows that for = 0 the amplification factor exhibits a discontinu1

0.8
0.7

0.8
0.7

0.6
0.5

0.6
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

g(G(K))

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

0
0

10

10

12

(a) = = 6 .

(b) = 6 , = 0.
1

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8

g(G(K))

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0

10

12

(c) =
,=
6

.
3

Figure 6.9: Effect of the SGS sweeps on the amplification factor of the LU-SGS +
BE scheme, for different Fourier wave angles . Second-order (p = 1) SV method,
AR = 1, CFL= 106 .

90

6.2. SV METHOD FOR TRIANGULAR CELLS

ity for a wave number equal to half the period of Equation (6.7), i.e. for
K = 2. In fact, the amplification factor starts from one for K = 0, and
it decreases very fast to zero for K > 0. However, when K = 2, it jumps
to a value which is again close to (0.9 1) and then it rapidly decreases
again for high-wave numbers. Notice that the discontinuity occurs at a
low-frequency wave number, i.e. at a frequency which is 16 of the whole
wave number covered by the spatial scheme. By increasing the number of
SGS sweeps, the amplitude of the discontinuity decreases but for one hundred SGS sweeps the curve is still discontinuous and it differs from that of
the direct + BE method.
Figures 6.9(a) and 6.9(c) show that, for = 6 and = 3 , the highfrequency components are well damped and the amplification factor does
not exhibit jumps as for = 0. Moreover, for these two values of , the
amplification factor also shows a discontinuity which disappears when five
or more SGS sweeps are employed.
In order to study the damping properties of the LU-SGS + BE scheme
when the geometrical stiffness imposed by the Navier-Stokes grids occurs
near walls, consider a mesh obtained for the following choice of the dimensionless vectors B1 and B2 ,




1
0

.
(6.20)
B1 =
,
B2 =
1
0
100
This mesh has anisotropic cells with AR = 100. As for a mesh with AR = 1,
in Figures 6.10(a) , 6.10(b), 6.10(c) the amplification factor of both direct
and LU-SGS methods is plotted as a function of the wave number K, for
= 6 , = 0 and = 3 . In these cases, the period PK for the selected
2
2
values of the solution orientation is respectively, sin
, 2 and
sin .
6

It is seen that, when few SGS sweeps (3 5) are employed, the damping properties of the LU-SGS + BE solver for = 6 and = 3 are similar
to that of the direct + BE method, Figures 6.10(a) and 6.10(c). For = 0
(Figure 6.10(b)) the behavior is different. In fact, the low-frequency components are not damped at all and the amplification factor for a wave number
range [0, ) is close to one. However, for K > it jumps to a value close to
zero and the high-frequency error components are still well damped.

91

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

1

0.8
0.7

0.9
0.8

0.6
0.5
0.4

0.6
0.5
0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.7

g(G(K))

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

20

40

60

80

0
0

100

(a) = = 6 .

(b) = 6 , = 0.
1

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8

g(G(K))

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0

20

40

60

80

100

,=
(c) =
6

.
3

Figure 6.10: Effect of the SGS sweeps on the amplification factor of the LU-SGS +
BE scheme, for different Fourier wave angles . Second-order (p = 1) SV method,
AR = 100, CFL= 106 .

6.2.2

Third-order SV method

The analysis for third-order SV method was performed using the mesh
with AR = 100 and a CFL number of = 106 . For the LU-SGS + BE solver
one, three, five, ten and one hundred SGS sweeps were employed. In Figure 6.11 the amplification factor is plotted as a function of the wave number, respectively for = 6 , = 0, = 3 .
For the third-order SV scheme, the damping properties of the LU-SGS +
92

6.2. SV METHOD FOR TRIANGULAR CELLS

1

0.8
0.7

0.8
0.7

0.6
0.5

0.6
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

100

200

300

400

500

600

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

g(G(K))

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

0
0

700

(a) = =
.
6

(b) = 6 , = 0.
1
Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8

g(G(K))

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0

100

200

300

400

(c) = 6 , =

.
3

Figure 6.11: Effect of the SGS sweeps on the amplification factor of the LU-SGS
+ BE scheme, for different Fourier wave angles . Third-order (p = 2) SV method,
AR = 100, CFL= 106 .

BE solver for = 6 and = 3 are close to those of the direct + BE method

when few SGS sweeps are employed, Figures 6.11(a) and 6.11(c). However,
when = 0, the low-frequency error components are not damped and the
amplification factor for a wave number range [0, ) is close to one. Subsequently, for K > the amplification factor jumps immediately to a value
close to zero and the high-frequency error components are nicely damped,
see Figure 6.11(b). This behavior has also been observed for second-order
SV scheme, as discussed in the previous section. Notice that the amplification factor shows a discontinuity for = 6 and = 3 at high-frequency,
93

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

when one or three SGS sweeps are employed. This discontinuity disappears by increasing the number of SGS sweeps.

6.2.3

Fourth-order SV method

The analysis of the fourth-order SV method was also performed on a mesh

with AR = 100 and for a CFL number of 106 . In Figure 6.12 the amplification factor is plotted as a function of the wave number for = 6 ,
1

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8

0.8
0.7

0.6

g(G(K))

g(G(K))

0.7

0.5

0.6
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

100

200

300

400

500

600

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

0
0

700

(a) = = 6 .

, = 0.
(b) =
6
1

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8
0.7

g(G(K))

0.6
0.5
0.4
0.3
0.2
0.1
0
0

100

200

300

400

(c) =
,=
6

.
3

Figure 6.12: Effect of the SGS sweeps on the amplification factor of the LU-SGS +
BE scheme, for different Fourier wave angles . Fourth-order (p = 3) SV method,
AR = 100, CFL= 106 .

94

6.3. SD METHOD FOR QUADRILATERAL CELLS

= 0, = 3 . One, three, five, ten and hundred SGS sweeps were used
for the LU-SGS + BE solver. It is seen that also for the fourth-order SV
scheme, the amplification factor of the LU-SGS + BE scheme gets closer
to that of the direct + BE solver for = 6 and = 3 when more SGS
sweeps are used. However, for = 0, the same behavior observed as for
the third-order SV scheme is found. In fact, the amplification factor for
a wave number range [0, ) is close to one and for K > it jumps immediately to a value close to zero, showing that the high-frequencies error
components are nicely damped.

6.3 SD method for quadrilateral cells

In this section the main results obtained by investigating the smoothing properties of the LU-SGS + BE solver in combination with second- to
fourth-order SD schemes are shown. The aim is to show that the LU-SGS
+ BE solver is also a good smoother for the SD method. To construct the
amplification factor g as a function of the dimensionless wave number K,
the procedure explained for the SV schemes is used.
A mesh with quadrilateral cells and aspect ratio AR of 100 was considered. This mesh was obtained by choosing the dimensionless vectors B1
and B2 as shown in (6.20). For the sake of consistency, the CFL number
was set to 106 and one, three, five, ten and hundred SGS sweeps were used
for the LU-SGS + BE solver.

6.3.1 Second-order SD method

The amplification factor of the LU-SGS + BE for the second-order SD
scheme for quadrilateral cells is shown in Figure 6.13, for = 6 , = 0,
= 3 . In this figure, the amplification factor of the direct solver is also
plotted for comparison. It is seen that the iterative solver, in combination
with the SD scheme, is stable, i.e. g 1, and its amplification factor gets
closer to that of the direct + BE solver by increasing the number of SGS
sweeps. In addition, Figure 6.13(b) shows that for = 0, the low-frequency
error components are weakly damped for a wave number range [0, ). For
high-wave numbers the amplification factor is close to zero, showing that
the high-frequency error components are nicely damped.
Upon comparison of the amplification factor for the second-order SV scheme
and the amplification factor for the second-order SD scheme, it is seen that
the LU-SGS + BE solver gives similar results for both spatial operators.
95

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

1

0.8
0.7

0.8
0.7

0.6
0.5

0.6
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

20

40

60

80

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

g(G(K))

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

0
0

100

(a) = = 6 .

(b) = 6 , = 0.

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8
0.7

g(G(K))

0.6
0.5
0.4
0.3
0.2
0.1
0
0

20

40

60

(c) =
,=
6

80

100

.
3

Figure 6.13: Effect of the SGS sweeps on the amplification factor of the LU-SGS +
BE scheme, for different Fourier wave angles . Second-order (p = 1) SD method,
AR = 100, CFL= 106 .

Note that when hundred SGS sweeps are used, the amplification is almost
indistinguishable from that of the direct + BE inversion method.

6.3.2

Third-order SD method

Figure 6.14 shows the amplification factor of the LU-SGS + BE for thirdorder SD scheme. The plots in this figure show once more that the amplification factor of the iterative solver gets closer to that of the direct + BE
96

6.3. SD METHOD FOR QUADRILATERAL CELLS

solver for = 6 and = 3 by increasing the number of SGS sweeps. For
= 0, the same behavior as for the second-order SD scheme is observed.
As for the third-order SV case, the amplification factor shows a discontinuity for = 6 and = 3 at high-frequency, when one or three SGS sweeps
are employed. The discontinuity disappears with an increasing number of
SGS sweeps.
1

0.8
0.7

0.8
0.7

0.6
0.5

0.6
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

100

200

300

400

500

600

0
0

700

(a) = = 6 .

(b) = 6 , = 0.

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8
0.7

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

g(G(K))

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

0.6
0.5
0.4
0.3
0.2
0.1
0
0

100

200

(c) = 6 , =

300

400

.
3

Figure 6.14: Effect of the SGS sweeps on the amplification factor of the LU-SGS
+ BE scheme, for different Fourier wave angles . Third-order (p = 2) SD method,
AR = 100, CFL= 106 .

97

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

6.3.3

Fourth-order SD method

The amplification factor of the LU-SGS + BE for the fourth-order SD scheme

for quadrilateral cells is shown in Figure 6.15, for = 6 , = 0, = 3 .
It can be seen that few SGS sweeps are in general sufficient to damp highfrequency error components and get an amplification factor fairly close to
that of the direct solver, for = 6 and = 3 . Similar results as for the
1

0.8
0.7

0.8
0.7

0.6
0.5

0.6
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
0

100

200

300

400

500

600

0
0

700

(a) = = 6 .

(b) =
, = 0.
6

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9
0.8
0.7

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

g(G(K))

g(G(K))

Direct solver
1 SGS
3 SGS
5 SGS
10 SGS
100 SGS

0.9

0.6
0.5
0.4
0.3
0.2
0.1
0
0

100

200

(c) =
,=
6

300

400

.
3

Figure 6.15: Effect of the SGS sweeps on the amplification factor of the LU-SGS +
BE scheme, for different Fourier wave angles . Fourth-order (p = 3) SD method,
AR = 100, CFL= 106 .

foregoing analysis are found for = 0. Indeed, the low-frequency error

components are weakly damped for a wave number range [0, ) and the
98

6.4. REMARKS
high-frequency ones are completely removed.

6.4 Remarks
The analysis has demonstrated that the LU-SGS + BE scheme is always
stable for any choice of the convective velocity direction and the solution
orientation for the second- to fourth-order SV and SD schemes. Furthermore, the analysis has shown that the smoothing properties of the implicit
solver depend strongly on the orientation of the solution defined by the angle . It has been shown that by increasing the number of the SGS sweeps,
the damping behavior of the LU-SGS + BE scheme gets closer to the damping properties of a direct solver + BE scheme.
The analysis was performed for a CFL of 106 and on a mesh with an aspect
ratio of hundred. This choice has enabled to take into account the effects
of the the geometrical stiffness imposed by the Navier-Stokes grids where
high-aspect ratios occur near walls. It has been found that 5 6 sweeps
are in general sufficient to get a good damping of the high-frequency error components. Moreover, it has been shown that the amplification factor may have a discontinuity for a certain wave number. This depends
on the direction of the harmonic wave solution and the number of SGS
sweeps. Nevertheless, the high-frequency error components are always
well damped. Besides, for a specific direction of the harmonic wave solution ( = 0) of the 2D linear convection equation, the implicit iterative
solver has some difficulties to damp low-frequency error components for a
wave number range equal to half of the period of Equation (6.7). For wave
numbers higher than half of the period the amplification factor is close to
zero.

99

CHAPTER 6. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

100

Chapter 7

Application I: spectral
volume method
In this chapter, the main results achieved by applying the spectral volume method and the non-linear LU-SGS algorithm with backward Euler scheme to problems governed by the 2D steady compressible NavierStokes equations are presented. The coupling between the SV method and
the non-linear LU-SGS algorithm has been limited to this type of flows
because, as discussed in Section 4.1.4, no stable high-order 3D SV partition has been found until now. Therefore, since the goal of this PhD research has been the development of an efficient high-order N-S/LES solver,
the application of the implicit time marching scheme for 2D and 3D unsteady laminar and turbulent flow simulations has been done for the SD
method, which was proven to be stable for any order of accuracy and multidimensional problems by several researchers. Examples of flow simulations with the implicit SD solver will be shown in the next chapter.
For the SV method, the implementation of the LU-SGS + BE scheme was
done in the 2D-SV-TRI code, which is a C++ code developed at the Vrije
Universiteit Brussel, Department of Mechanical Engineering, Fluid Mechanics and Thermodynamics Research Group. In this code, the 2D SV
method can also be combined with explicit Runge-Kutta (E-RK) time marching schemes and for steady-state flow simulations, the (pseudo) time integration can be accelerated by a full p-multigrid strategy, as discussed in
Appendix B.
The grids for all the SV test cases were generated with Gmsh [54]. The cal101

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

culations were performed on a Linux workstation with an Intel T2500 Core
Duo (2.0 GHz) processorTM . Wherever possible, direct validation against
experimental data and/or reference solutions available in literature are
made. The results presented in the following sections are published in
Parsani et al. [127, 130].

7.1 Two-dimensional laminar steady flow simulations

The LU-SGS + BE scheme in combination with a full multigrid (FMG) Vcycle p-multigrid algorithm was used to compute the 2D steady laminar
flow past a circular cylinder and over a NACA0012 airfoil, and the steady
laminar flow in a channel with a backward-facing step. In order to assess
the convergence properties of the LU-SGS + BE scheme, the performance
of the latter scheme is compared with that of a family of E-RK smoothers
available in the 2D-SV-TRI code. The E-RK schemes compute the solution
n
at the new time tn+1 , i.e. Wn+1
i,j , starting from the solution at time t , i.e.
n
Wi,j , using an explicit NRK -stage procedure. In this thesis, E-RK schemes
of the following form are considered [138]:
W0i,j

Wni,j

Wm
i,j

m1
1
2
3 t
Cm
W0i,j + Cm
Wm1
+ Cm
i,j
i,j Ri,j

Wn+1
i,j

RK
WN
i,j ,

1 m NRK

1
2
3
where the coefficients Cm
, Cm
, Cm
are listed in Table 7.1. These E-RK
schemes were introduced in Ramboer et al. [138] in order to minimize the
total error arising from the spatial and temporal discretizations. Because
of this, they are also denoted optimized E-RK schemes in the reminder of
this work.

Table 7.1: Coefficients of different NRK -stage E-RK schemes available in the 2DSV-TRI code.

Scheme
Opt. E-RK2

1
Cm
1

2
Cm
0

C13
1
4

C23
1

C33
-

C43
-

C53
-

Opt. E-RK3

1
5

1
2

Opt. E-RK5

1 (if m = 1)
3
1 Cm
(if m 6= 1)

0 (if m = 1)
3
Cm
(if m 6= 1)

85
1300

1
10

9
50

1
4

132
300

102

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

The FMG V-cycle p-multigrid strategies adopted here are depicted in Figure 7.1, for both E-RK and LU-SGS + BE smoothers. The numbers of
smoothing sweeps were empirically determined in order to minimize the
computer time (CPU-time) needed to perform the simulations. Note that
these numbers are independent of the test case.

(a) Optimized E-RK schemes.

(b) LU-SGS + BE scheme.

Figure 7.1: FMG V-cycle with number of sweeps on each p-multigrid level.

For the LU-SGS + BE scheme, the non-linear system (5.8) was solved with
multiple cell-wise symmetric Gauss-Seidel (SGS) sweeps with a prescribed
tolerance of 106 on the change of the L2 norm of the solution variation
m+1
^
and/or a maximum number of 6 SGS sweeps. For all the calcuW
cc
lations, a high CFL number, varying between 104 and 106 , was used. The
former choices allows strong damping of high-frequency error components
and gives good properties in terms of CPU-time and convergence rate for
the present computations. Notice that, in Chapter 6, from the Von Neumann analysis of the 2D linear advection equation, it was found that a
maximum number of 5 6 SGS sweeps are in general sufficient to get good
damping of high-frequency error components for a CFL number of 106 .
All the test cases were solved with a local time stepping technique, as described in Section 5.3. A Roe FDS flux as a Riemann solver and the LSV
approach for the diffusive terms were used. For convergence, the residual
of the continuity equation was monitored. Here, the residuals are normalized by the corresponding residuals of the first iteration and they are
shown in base ten logarithmic scale.
103

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

7.1.1

Flow over a circular cylinder

The 2D steady compressible viscous flow past a circular cylinder is considered in this section. The computational domain is shown in Figure 7.2
where the incoming uniform flow is from left to right, i.e. in the 1 direction. The simulation is conducted at Reynolds number Re of 40 and
2

20D

10D

30D

Figure 7.2: Configuration of the 2D circular cylinder test case.

free-stream Mach number M of 0.15. The Reynolds number is based on

the free-stream velocity module |~u | and the diameter of the cylinder D.
The Prandtl number P r is set to 0.72, which is the standard value for air.
At this conditions, the flow is laminar and steady, with a recirculation zone
behind the cylinder.
At the cylinder wall, which is assumed to be adiabatic, the no-slip boundary condition discussed in Section 3.3.4 is imposed. Maximum secondorder (p = 1) SV results are given because of the curved boundaries. A
first-order interpolation is used for the boundary shape in the 2D-SV-TRI
code, but high-order schemes would require a more accurate interpolation,
especially on the relatively coarse grids that are being used in combination with high-order polynomial representations of the solution; see for instance Wang and Liu [188]. At the far field boundary, which is sufficiently
far from the cylinder, the flow is prescribed to be uniform. This corresponds
to a Dirichlet boundary condition, as discussed in Section 3.3.1.
First, a medium grid with 3, 744 triangular cells with a maximum aspectratio ARmax of 2 was used. The mesh was generated in such a way that
the cell aspect ratio close to the cylinder wall could be easily controlled, as
illustrated in Figure 7.3.
104

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

Figure 7.3: Detail of the grid for the steady laminar flow over a circular cylinder.
3, 744 triangular cells with ARmax = 2.

In the first part of Table 7.2, the CFL numbers for both E-RK and LUSGS + BE schemes are listed. For the E-RK scheme two CFL numbers are
indicated; the first one is the convective CFL number C while the second
one is the viscous CFL number D . According to the local time stepping
technique, two local time steps are then computed and the solution at the
next time level is calculated using the minimum between these two values.
In this table, the exponent n denotes the time iteration index. For the LUSGS + BE scheme both maximum convective and viscous CFL numbers
were set to 106 , for each p-multigrid level.
Table 7.2: CFL number for two-level V-cycles p-multigrid on two different grids, for
the steady laminar flow over a circular cylinder.

Scheme
Opt. E-RK
LU-SGS + BE

3, 744 cells, ARmax = 2

SV p = 0
SV p = 1
7, 0.5
4, 0.3
10n

10n

5, 440 cells, ARmax = 162

SV p = 0
SV p = 1
1.5, 0.01
0.35, 0.005
10n

10n

The Mach number contour and the stream traces obtained with the secondorder SV method and the LU-SGS + BE scheme are shown in Figure 7.4.
It can be seen that, the flow field around the cylinder is well resolved and
the recirculation zone behind the cylinder is clearly visible.
In Figure 7.5, the convergence histories of the E-RK and the LU-SGS +
BE schemes are compared. The computations were stopped when the L2
norm of the residuals was reduced 12 orders of magnitude. In Table 7.3, the
105

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

M
0.18
0.17
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01

Figure 7.4: Mach number contour and stream traces, for the steady laminar flow
over a circular cylinder, obtained with second-order (p = 1) SV method and the LUSGS + BE scheme. Grid with 3, 744 triangular cells and ARmax = 2. M = 0.01.

number of V-cycles for each multigrid level (VMG1 and VMG2), the total
number of V-cycles (VMG) and the total CPU-time are listed. For this test
case, once the VMG2 cycle is used and the second-order accurate solution
is computed, the smoothing sweeps at p = 0 takes 69% of the total CPU
time listed in Table 7.3. The super-script indicates that the residuals is
not yet converged. In fact, after 10, 000 V-cycles the E-RK scheme reduced
2

Opt. ERK
LUSGS + BE

Log(||Res|| )

2
4
6
8
10
12
0

200

400
600
Cycles

800

1000

Figure 7.5: Convergence histories of the optimized E-RK scheme and the LU-SGS
+ BE scheme for the steady laminar flow over a circular cylinder. Second-order
(p = 1) SV method, grid with 3, 744 triangular cells and ARmax = 2.

106

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

Table 7.3: Number of V-cycles and total CPU-time [s] of the optimized E-RK scheme
and the LU-SGS + BE scheme for the steady laminar flow over a circular cylinder.
Second-order (p = 1) SV method on two different grids.

Grid
3, 744 cells,
ARmax = 2
5, 440 cells,
ARmax = 162

Scheme
Opt. E-RK

VMG1
500

VMG2
9, 500

VMG
10, 000

CPU-time
145, 094

LU-SGS + BE

78

401

479

17, 576

Opt. E-RK

500

9, 500

10, 000

211, 859

LU-SGS + BE

77

418

495

24, 861

the L2 norm of the residuals by 9 orders of magnitude. From this table it

is seen that the LU-SGS + BE scheme is almost two orders of magnitude
faster than the E-RK scheme in terms of CPU-time.
In order to study the convergence properties of the LU-SGS + BE solver,
this test case was also computed on a grid with 5, 440 triangular cells with
ARmax = 162. Table 7.2 shows also the CFL number used for this new
mesh. Note that the power law of the CFL number and its maximum value
for LU-SGS + BE scheme were the same as for the ARmax = 2 case.
In Figure 7.6 the convergence histories of the E-RK and LU-SGS + BE

Opt. ERK
LUSGS + BE

Log(||Res||2)

2
4
6
8
10
12
0

200

400
600
Cycles

800

1000

Figure 7.6: Convergence histories of the optimized E-RK scheme and the LU-SGS
+ BE scheme for the steady laminar flow over a circular cylinder. Second-order
(p = 1) SV method, grid with 5, 440 triangular cells and ARmax = 162.

107

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

schemes for this mesh are shown. In Table 7.3, the number of V-cycles for
each multigrid level, the total number of V-cycles and the total CPU-time
are listed. It is observed that whereas the convergence of the E-RK scheme
is significantly slowed down because of the increased ARmax , this is not the
case for the LU-SGS + BE scheme. The required number of V-cycles to reduce the residual 12 orders of magnitude (495 V-cycles) is almost identical
as for the ARmax = 2 case (479 V-cycles).
In Table 7.4 the drag coefficients CD for both simulations are summarized.
Good agreement with the experimental data reported in Barlow et al. [11]
is found.
Table 7.4: Drag coefficient for the steady laminar flow over a circular cylinder,
obtained with second-order (p = 1) SV method and the LU-SGS + BE scheme on
two different grids. Comparison with experimental measurements [11].

CD

Exp. [11]

3, 744 cells, ARmax = 2

5, 440 cells, ARmax = 162

1.536

1.526

1.531

To conclude this study on the cylinder, the effect of the low Mach number on the convergence rate of the LU-SGS + BE solver is presented. In
Figure 7.7, the convergence histories of the LU-SGS + BE scheme for both
meshes (ARmax = 2 and ARmax = 162) are shown for three values of the
free-stream Mach number M : 0.15, 0.05, 0.005. The Reynolds and Prandtl
numbers were again fixed to 40 and 0.72, respectively. The power law of the
CFL number and its maximum value for LU-SGS + BE scheme was the
same as used for the mesh with ARmax = 2 and ARmax = 162. The computations were stopped when the L2 norm of the residuals was reduced 12
orders of magnitude.
In Table 7.5 the number of V-cycles for each multigrid level and the total number of V-cycles for the three values of the Mach number are listed.
In this table, two values for the total number of V-cycles are indicated.
The first one is the total number to reduce the L2 norm of the residuals 10
orders of magnitude (VMG (-10)), while the second one is the total number for a reduction of 12 orders of magnitude (VMG (-12)). It is seen that
with both meshes, the required number of V-cycles to transfer the solution
from a first-order polynomial approximation to a second-order polynomial
approximation (VMG1) decreases when the Mach number is reduced. The
switch to a finer level is made when the L2 norm of the coarse level resid108

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

2

M = 0.15

M = 0.15

M = 0.05

M = 0.05

M = 0.005
2

Log(||Res|| )

Log(||Res|| )

M = 0.005
2

6
8

6
8

10

10

12

12

14
0

200

400
600
Cycles

800

14
0

1000

(a) 3, 744 triangular cells, ARmax = 2.

200

400
600
Cycles

800

1000

(b) 5, 440 triangular cells, ARmax = 162.

Figure 7.7: Effect of the free-stream Mach number on the convergence histories
of the LU-SGS + BE scheme for the steady laminar flow over a circular cylinder.
Second-order (p = 1) SV method on two different grids.

uals is smaller than a factor switch times the L2 norm of the fine level
residual. In the present work switch is set to 0.001.
Figure 7.7 shows that the convergence rate of the LU-SGS + BE method
for a residual norm higher than 1011 is not slowed by a decrease of the
Mach number. On the contrary, for M = 0.005, the required number of V-

Table 7.5: Effect of the free-stream Mach number on the number of V-cycles of the
LU-SGS + BE scheme for the steady laminar flow over a circular cylinder. Secondorder (p = 1) SV method on two different grids.

Grid
3, 744 cells,
ARmax = 2
5, 440 cells,
ARmax = 162

M
0.15

VMG1
78

VMG2
401

VMG (-10)
400

VMG (-12)
479

0.05

58

424

323

482

0.005

26

695

232

721

0.15

77

418

425

495

0.05

48

537

325

585

0.005

26

913

266

939

109

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

cycles to reduce the residual norm 10 orders of magnitude is smaller than
the computation with M = 0.15. The low Mach number, combined with a
mesh with high-aspect ratio, slightly affected the convergence rate of the
LU-SGS + BE solver. For M = 0.005 and the mesh with ARmax = 162, the
LU-SGS + BE scheme took 34 V-cycles more than that for the mesh with
ARmax = 2. For a residual norm smaller than 1011 the convergence rate of
the LU-SGS + BE solver slows. This behavior is caused by round-off errors
which are more influential at low Mach number.

7.1.2

Flow over a NACA0012 airfoil

The compressible steady laminar flow simulation over a NACA0012 airfoil

is conducted at free-stream Mach number M of 0.5, Reynolds number Re,
based on the free-stream velocity module |~u | and the airfoil chord c, of
5 103 and Prandtl number P r of 0.72. In Figure 7.8 the configuration of
the test case is illustrated, where the incoming flow is from left to right.
The airfoil is placed on the 1 axis (2 = 0) of the computational domain.
2

12 c

1
R = 6c

12 c
Figure 7.8: Configuration of the 2D NACA0012 airfoil test case.

At the left-hand-side boundary (the inflow) the flow is prescribed to be uniform with zero angle of attack. The same boundary conditions are also applied to the upper and lower boundaries. At the right-hand-side boundary
(the outflow), far enough from the profile, only the pressure is prescribed.
At the airfoil wall, which is assumed to be adiabatic, the no-slip boundary
condition is imposed. Maximum second-order (p = 1) SV results are given
because of the curved boundaries, as discussed for the previous test case.
110

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

Table 7.6: CFL number for two-level V-cycles p-multigrid, for the steady laminar
flow over a NACA0012 airfoil at zero angle of attack. Grid with 6, 878 triangular
cells and ARmax = 2.5.

Scheme
Opt. E-RK

SV p = 0
7, 0.5

SV p = 1
4, 0.1

LU-SGS + BE

10n

10n

A grid with 6, 878 triangular cells with ARmax = 2.5 was used.
In Table 7.6 the CFL numbers for both E-RK and LU-SGS + BE smoothers
are listed. The maximum CFL number for the implicit scheme was set to
106 . In Figure 7.9, the convergence histories of the E-RK and LU-SGS +
BE schemes are compared. The computations were stopped when the L2
norm of residuals was reduced by 10 orders of magnitude.
In Table 7.7, the number of V-cycles for each multigrid level, the total
number of V-cycles and the total CPU-time are listed. This table shows
that the LU-SGS + BE scheme is approximately six times faster than the
E-RK scheme in terms of CPU-time. For this test case, once the finest grid

Opt. ERK
LUSGS + BE

Log(||Res||2)

2
4
6
8
10
12
0

100

200

300
400
Cycles

500

600

700

Figure 7.9: Convergence histories of the optimized E-RK scheme and the LU-SGS
+ BE scheme for the steady laminar flow over a NACA0012 airfoil at zero angle
of attack. Second-order (p = 1) SV method, grid with 6, 878 triangular cells and
ARmax = 2.5.

111

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

Table 7.7: Number of V-cycles and total CPU-time [s] of the optimized E-RK scheme
and the LU-SGS + BE scheme for the steady laminar flow over a NACA0012 airfoil
at zero angle of attack. Second-order (p = 1) SV method, grid with 6, 878 triangular
cells and ARmax = 2.5.

Scheme
Opt. E-RK

VMG1
35

VMG2
606

VMG
641

CPU-time
16, 589

LU-SGS + BE

13

49

62

2, 784

solution is reached, i.e. the VMG2 cycle is used and the second-order accurate solution is computed, the coarse grid solution takes 67% of the total
CPU-time listed in Table 7.7.
In Figure 7.10, the Mach number contour and the stream traces are shown.
This plot visually compares well with the results shown in Sun et al. [167].
The distribution of the skin friction coefficient Cf and the pressure coeffi-

M
0.58
0.54
0.5
0.46
0.42
0.38
0.34
0.3
0.26
0.22
0.18
0.14
0.1
0.06
0.02

Figure 7.10: Mach number contour and stream traces for the steady laminar flow
over a NACA0012 airfoil at zero angle of attack, obtained with second-order (p = 1)
SV method and the LU-SGS + BE scheme. Grid with 6, 878 triangular cells and
ARmax = 2.5. M = 0.04.

cient CP on the airfoil surface are plotted in Figure 7.11, where the horizontal axis is the normalized coordinate respect to the leading edge of the
airfoil (1 /c). The skin friction coefficient is defined by
Cf =

w
,
1
u |2
2 |~
112

(7.1)

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

where w is the local wall shear stress. Figure 7.11 shows good agreement
with the experimental measurements presented in the AGARD Report AR138 [1].
0.15

0.5
Experiment
p=1

Experiment
p=1
0

CP

Cf

0.1

0.05

0.05
0

0.5

0.2

0.4

1 /c

0.6

0.8

(a) Skin friction coefficient (Cf vs. 1 /c).

1.5
0

0.2

0.4

1 /c

0.6

0.8

(b) Pressure coefficient (CP vs. 1 /c).

Figure 7.11: Distribution of skin friction and pressure coefficients on the

NACA0012 airfoil surface at zero angle of attack, obtained with second-order
(p = 1) SV method and the LU-SGS + BE scheme on a grid with 6, 878 triangular cells and ARmax = 2.5. Comparison with experimental measurements [1].

To conclude this study on the NACA0012 airfoil, a comparison between the

convergence behavior of the LU-SGS + BE scheme and the RK3/Implicit
Residual scheme, proposed in Swanson et al. [169], is presented. The latter scheme is combined with a classical second-order FV scheme. In [169],
the convergence of a three-stage E-RK scheme with h-multigrid is accelerated by preconditioning with a fully implicit operator, whose inverse is
approximated with three point-wise SGS iterations. The aim is to compare
the convergence behavior of the schemes, i.e. only the number of multigrid cycles to reduce the L2 norm of the residuals 12 orders of magnitude.
It should be noted that a comparison of the efficiency is difficult as both
simulations have a different number of degrees of freedom (DOFs): the
second-order (p = 1) SV scheme for 2D problems on triangular grids has
six times the number of DOFs of a classical FV scheme on quadrilateral
grids. For this reason and because the simulations were done on different
computers, no CPU-time comparisons is shown. The same conditions as in
the previous computation were used except for the angle of attack which
was set to 2.5 .
The RK3/implicit scheme used a structured quadrilateral mesh with 16, 384
cells while the LU-SGS + BE scheme used a triangular mesh constructed
from the quadrilateral mesh, as illustrated in Figure 7.12. The number of
113

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

Figure 7.12: Detail of the SV grid for the comparison with the RK3/Implicit Residual scheme. Steady laminar flow over a NACA0012 airfoil at 2.5 angle of attack,
grid with 32, 768 triangular cells and ARmax = 136.

triangular cells for the LU-SGS + BE scheme is then 32, 768. The maximum
aspect ratio is ARmax = 136. In the calculation of the RK3/Implicit Residual the CFL number was 16 during the first 8 multigrid cycles and then, it
was increased to 103 . For the LU-SGS + BE scheme the CFL number was
set to 106 during the entire calculation. In Figure 7.13, the convergence
histories of the RK5 Standard scheme, the RK3/Implicit Residual scheme
and the LU-SGS + BE scheme are compared.
In Table 7.8 the total number of V-cycles are listed. The LU-SGS + BE requires 19 cycles more than the RK3/Implicit Residual code and nearly one
order of magnitude less than the RK5 Standard code. In terms of multigrid

Table 7.8: Number of V-cycles of the RK5 Standard scheme, the RK3/Implicit
Residual scheme and the LU-SGS + BE scheme for the flow over a NACA0012
airfoil at 2.5 angle of attack. Second-order FV/SV methods, structured quadrilateral/triangular grid, ARmax = 136.

Scheme
RK5 Standard

Total Number of Cycles

773

RK3/Implicit Residual, Swanson et al. [169]

58

LU-SGS + BE

77
114

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

0
2

RK5 Standard (Swanson, Turkel)

RK3/Implicit residual (Swanson, Turkel, Rossow)
LUSGS + BE

Log(||Res||2)

4
6
8
10
12
14
0

200

400
600
Cycles

800

1000

Figure 7.13: Convergence histories of the RK5 Standard scheme, the RK3/Implicit
Residual scheme and the LU-SGS + BE scheme for the flow over a NACA0012
airfoil at 2.5 angle of attack. Second-order FV/SV methods, structured quadrilateral/triangular grid, ARmax = 136.

cycles, the RK3/Implicit Residual scheme is more efficient than the LUSGS + BE scheme. However, the work per DOF of the RK3/Implicit Residual scheme is approximately twice that of the LU-SGS + BE scheme. The
RK3/Implicit Residual scheme employs 12 point-wise SGS (3 SGS sweeps
plus a residual calculation times three RK stages) per DOF, while the LUSGS + BE scheme requires only 6 SGS sweeps per DOF. However, the
LU-SGS + BE scheme requires the computation of the LHS of Equation
(5.8), which is time consuming since its size is quite large. In fact, the total number of diagonal block matrix elements that have to be stored for a
tetrahedral mesh with N cells with polynomial degree p increases with p4
in 2D and p6 in 3D.
Table 7.9: Drag coefficient for the flow over a NACA0012 airfoil at 2.5 angle of attack obtained with the RK5 Standard scheme, the RK3/Implicit Residual scheme
and the LU-SGS + BE scheme. Second-order FV/SV methods, structured quadrilateral/triangular grid, ARmax = 136.

Scheme
RK5 Standard

CD
0.0568425

CL
0.0331927

RK3/Implicit Residual, Swanson et al. [169]

0.0568523

0.0333319

LU-SGS + BE

0.0562196

0.0331650

115

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

Finally, in Table 7.9 the drag coefficient CD and the lift coefficient CL for
the three simulations are summarized. This table shows good agreement
between the results of the three codes.

7.1.3

Flow in a channel with a backward-facing step

This test case is fundamental in design and geometry and consequently it

is found in a variety of engineering applications. The flow separation process caused by the sudden change in geometry has been used extensively
in applications, usually in order to create a recirculation region or a sudden change in pressure.
In Figure 7.14 the characteristic lengths of the channel and a sketch of the
main flow features are illustrated. The incoming flow is from left to right.
A sudden expansion of H/h = 2 is used, see Barton [12]. The position of
the inlet section (Lin ) is fixed to 10 h. An inlet Mach number Minlet of 0.2,
a Reynolds number Re of 800, based on twice the inlet channel height and
the inlet bulk velocity, and a Prandtl number P r of 0.72 are imposed. Uniform mass density and velocity profiles are prescribed at the inlet section
so that the Reynolds number is equal to the prescribed value. At the channel walls, which are assumed to be adiabatic, no-slip boundary conditions
are imposed. At the right-hand-side boundary (the outflow), a pressure
outlet boundary condition is used.
2
l3

H h

l2

1
l1

Lin

16H

Figure 7.14: Configuration of the 2D backward facing step test case and locations
of the recirculating regions.

116

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

Table 7.10: CFL number for three-level V-cycles p-multigrid, for the backward facing step flow.

Scheme
Opt. E-RK

SV p = 0
3.5, 0.7

SV p = 1
0.5, 0.1

SV p = 2
0.05, 0.001

SV p = 3
0.01, 0.0001

LU-SGS + BE

10n

10n

10n

10n

Firstly, an unstructured grid with 8, 503 triangular cells with a maximum

aspect ratio of 42 was used. In Table 7.10, the CFL numbers are listed for
both E-RK and LU-SGS + BE schemes. The solution was computed with
third- (p = 2) and fourth-order (p = 3) SV methods. For the LU-SGS +
BE smoother the CFL number started from one and it reached its maximum value using the CFL-law indicated in Table 7.10. The maximum CFL
number was set to 106 for first- and second-order SV schemes and to 104 for
third- and fourth-order SV schemes. This choice allow to have stable computations and a better convergence in terms of CPU-time. In Figure 7.15,
the convergence histories of the optimized E-RK scheme and the LU-SGS
scheme are compared, for third- and fourth-order SV methods.
Tables 7.11 and 7.12 list the number of V-cycles for each multigrid level,

Opt. ERK
LUSGS + BE

Opt. ERK
LUSGS + BE

Log(||Res|| )

Log(||Res|| )

4
6

4
6

10

10

12
0

200

400
600
Cycles

800

12
0

1000

(a) p = 2.

200

400

600 800
Cycles

1000 1200 1400

(b) p = 3

Figure 7.15: Convergence histories of the optimized E-RK scheme and the LU-SGS
+ BE scheme for the backward-facing step flow. Third- (p = 2) and fourth-order
(p = 3) SV methods, grid with 8, 503 triangular cells and ARmax = 42.

117

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

the total number of V-cycles and the total CPU-time to obtain a reduction
of 10 orders of magnitude of the L2 residual norms, resp. for the thirdand fourth-order SV schemes. The super-script indicates that the residuals are not yet converged. From this table it is seen that the LU-SGS +
BE scheme is more than two orders of magnitude faster than the E-RK
scheme for both third- and fourth-order SV methods.
Table 7.11: Number of V-cycles and total CPU-time [s] of the optimized E-RK
scheme and the LU-SGS + BE scheme for the backward-facing step flow. Thirdorder (p = 2) SV method, grid with 8, 503 triangular cells and ARmax = 42.

Scheme
Opt. E-RK

VMG1
250

VMG2
421

VMG3
9, 329

VMG
10, 000

CPU-time
50, 264

LU-SGS + BE

32

20

244

296

12, 032

Table 7.12: Number of V-cycles and total CPU-time [s] of the optimized E-RK
scheme and the LU-SGS + BE scheme for the backward-facing step flow. Fourthorder (p = 3) SV method, grid with 8, 503 triangular cells and ARmax = 42.

Scheme
Opt. E-RK

VMG1
250

VMG2
421

VMG3
491

VMG4
8, 838

VMG
10, 000

CPU-time
85, 426

LU-SGS + BE

32

20

79

315

446

29, 651

In order to study the convergence properties of the LU-SGS + BE solver

this test case was also computed using a finer grid, obtained by doubling
the number of cells of the previous mesh in each direction. Therefore, the
new mesh has 34, 015 triangular cells with still a maximum aspect ratio of
42. The power law of the CFL number and its maximum value for LU-SGS
+ BE scheme are the same as for the previous grid.
Table 7.13: Number of V-cycles and total CPU-time [s] of the LU-SGS + BE scheme
for the backward-facing step flow. Third- (p = 2) and fourth-order (p = 3) SV
methods, grid with 34, 015 triangular cells and ARmax = 42.

SV method
p=2
p=3

VMG1
40
40

VMG2
23
23

VMG3
249
83

118

VMG4
320

VMG(-10)
312
465

CPU-time
62, 566
160, 115

7.1. TWO-DIMENSIONAL LAMINAR STEADY FLOW SIMULATIONS

In Figure 7.16, the convergence histories of the LU-SGS + BE scheme are
compared for both grids. It can be seen that, according to the multigrid
theory, the asymptotic convergence rate is unaffected by the mesh size.
Notice that, as shown in Table 7.13, there is a small grid effect on the
FMG procedure, where slightly more cycles are needed on the fine grid to
reach the maximum order of accuracy. This feature was also reported in
Van den Abeele et al. [174] and Bassi et al. [14], respectively for explicit
and implicit smoothers.
2

Grid with 8,503 cells

Grid with 34,015 cells

Grid with 8,503 cells

Grid with 34,015 cells

Log(||Res|| )

Log(||Res|| )

4
6

4
6

10

10

12
0

100

200
Cycles

300

12
0

400

(a) p = 2.

100

200
300
Cycles

400

500

(b) p = 3.

Figure 7.16: Convergence histories of the LU-SGS + BE scheme for the backwardfacing step flow on two different grids. Third- (p = 2) and fourth-order (p = 3) SV
methods.

In Figure 7.17, the stream traces obtained with fourth-order SV method

and the LU-SGS + BE scheme are shown.

Figure 7.17: Stream traces for the backward facing step flow obtained with fourthorder (p = 3) SV method and the LU-SGS + BE scheme. Grid with 8, 503 triangular
cells and ARmax = 42.

119

CHAPTER 7. APPLICATION I: SPECTRAL VOLUME METHOD

Table 7.14 compares the re-attachment and separation positions for both
grids with those reported in Barton [12]. For both grids, the results are in
good agreement with the reference data.
Table 7.14: Re-attachment and separation positions for the backward-facing step
flow obtained with third- (p = 2) and fourth-order (p = 3) SV methods and the LUSGS + BE scheme on two different grids. Comparison with reference numerical
solution [12].

Barton with inlet channel [12]

SV p = 2, 8, 503 cells
SV p = 3, 8, 503 cells
SV p = 2, 34, 015 cells
SV p = 3, 34, 015 cells

120

l1 h
11.51
11.76
11.72
11.73
11.69

l2 h
9.14
9.28
9.24
9.27
9.24

l3 h
20.66
20.74
20.71
20.70
20.69

Chapter 8

Application II: spectral

difference method
In this chapter, the main results achieved by applying the spectral difference method and the non-linear LU-SGS algorithm to problems governed
by 2D and 3D compressible Navier-Stokes and filtered Navier-Stokes equations (LES approach) are presented.
The non-linear LU-SGS algebraic solver and the LES approach for the
SD method have been implemented in the COOLFluiD code which was
developed at the von Karman Institute for Fluid Dynamics. COOLFluiD
is a collaborative simulation environment written in C++. It focuses on
complex computational fluid dynamics (CFD) involving multi-physics phenomena. Every physical model or numerical method is a separate plugin component which is loaded on demand. This creates high-performant
solvers, each dedicated to a specific application, while reusing the same
components. The COOLFluiD platform is also a collaborative platform,
where diverse European research centers join together their developments.
More information about COOLFluiD can be found in the PhD theses of T.
Quintino [136], A. Lani [95] and T. Wuilbaut [195], and on the COOLFluiD
project web site [137].
A Roe FDS flux as a Riemann solver and the BR2 approach for the diffusive
terms were used. The grids for all the test cases were generated with Gmsh
[54] and Fluent/GambitTM . The former code allows second-order polynomial approximation of the boundary elements. In fact, high-order schemes
require an accurate representation of curved boundaries, especially on the
121

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

relatively coarse grids that are being used in combination with high-order
polynomial representations of the solution. The calculations were done on
a server with Dual Core AMD OpteronTMprocessors with a clock-speed of
2,412MHz. Wherever possible, direct validation against experimental data
and/or reference solutions available in literature are made.

8.1 Steady laminar flow simulations

In this section, the 2D steady laminar flow around a NACA0012 airfoil at
zero angle of attack and the 3D steady laminar flow through a 90 bending
square duct are considered. In order to assess the convergence properties
of the LU-SGS + BE scheme, a Newton-Raphson GMRES algebraic solver
with BE scheme, as discussed in Appendix C, is used as reference efficient
time marching scheme. In COOLFluiD, the PETSc suite1 is used for the
GMRES algorithm [3].
For convergence, the residual of the continuity equation was considered.
In this work, the residuals are normalized by the corresponding residuals
of the first iteration and they are shown in base ten logarithmic scale.

8.1.1

Flow over a NACA0012 airfoil

The simulation is conducted at Reynolds number Re of 5, 000 and freestream Mach number M of 0.5. The Reynolds number is based on the
module of the free-stream velocity vector |~u | and the airfoil chord c. The
Prandtl number P r is set to 0.72, which is the standard value for air. In
Figure 8.1 the configuration of the test case is illustrated, where the incoming flow is from left to right. The airfoil is placed on the 1 axis (2 = 0) of
the computational domain. At the left-hand-side boundary (the inflow) the
flow is prescribed to be uniform with zero angle of attack, as described in
Section 3.3.1. The same boundary conditions are also applied to the upper
and lower boundaries. At the right-hand-side boundary (the outflow), sufficiently far from the airfoil, only the pressure is prescribed, as discussed
in Section 3.3.3. At the airfoil wall, which is assumed to be adiabatic, the
no-slip boundary condition described in Section 3.3.4 is imposed.
A grid with 2, 925 quadrilateral cells with ARmax = 9 near the airfoil wall
was used. In Figure 8.2, two details of this grid are shown. The compu1 PETSCc

suite is a data structures and routines for the scalable (parallel) solution of
scientific applications modeled by partial differential equations.

122

8.1. STEADY LAMINAR FLOW SIMULATIONS

2

12 c

1
R = 6c

12 c

Figure 8.1: Configuration of the 2D NACA0012 airfoil test case.

tations were performed using second- (p = 1) to fourth-order (p = 3) SD

methods, with the airfoil wall represented by a quadratic mapping. Freestream flow conditions were used as initial solution.
Figure 8.3 compares the experimental distribution of the pressure coefficient CP reported in [1] with those obtained using second- to fourth-order
SD methods and the LU-SGS + BE scheme. In this figure, the horizontal
axis is the normalized coordinate respect to the leading edge of the airfoil
(1 /c). It can be seen that, the SD results get closer to the experimental
data by increasing the order of the polynomial reconstruction. Note that,
third- and fourth-order accurate solutions are nearly indistinguishable, indicating the order-independent flow convergence.

(a) Airfoil and wake.

(b) Airfoil.

Figure 8.2: Two details of the grid for the steady laminar flow over a NACA0012
airfoil at zero angle of attack. 2, 925 quadrilateral cells with ARmax = 9.

123

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

0.5

CP

0
Experiment
p=1
p=2
p=3

0.5

1.5
0

0.1

0.2

0.3

0.4

0.5
1 /c

0.6

0.7

0.8

0.9

Figure 8.3: Effect of the p-refinement on the distribution of the pressure coefficient on the airfoil surface, for the steady laminar flow over a NACA0012 airfoil
at zero angle of attack, obtained with SD methods and the LU-SGS + BE scheme.
Comparison with experimental measurements [1].

The Reynolds number is near to the upper limit for a steady laminar flow.
A feature of this test case is the separation region of the flow occurring
near the trailing edge, which causes the formation of a small recirculation
bubble that extends in the near-wake region of the airfoil. The comparison
between the separation point locations and the drag coefficients CD computed using second- to fourth-order SD methods are summarized in Table
8.1. It is seen that the order-independent convergence is also verified for
these quantities.

Table 8.1: Effect of the p-refinement on the drag coefficient and the location of the
separation point, for the steady laminar flow over a NACA0012 airfoil at zero angle
of attack, obtained with SD methods and the LU-SGS + BE scheme.

SD method
p = 1 (11, 700 DOFs)

CD
0.05448

Separation point
86.81%

p = 2 (26, 325 DOFs)

0.05476

81.43%

p = 3 (46, 800 DOFs)

0.05476

81.42%

124

8.1. STEADY LAMINAR FLOW SIMULATIONS

Algebraic solver performance
The non-linear LU-SGS and the Newton-Raphson GMRES solvers, combined with the BE scheme, were both used to solve the non-linear algebraic
system arising from the spatial and temporal discretizations. A local time
stepping technique, as described in Section 5.3, was also used. The CFL
number was computed using the following power law


1
n1
= min
(8.1)
2
, max ,
4
where the superscript n indicates the time iteration index. The upper limit
max depends on the specific test case and will be defined further on. For
the non-linear LU-SGS algorithm, a maximum of thirty SGS sweeps per
time iteration and/or a prescribed tolerance of 105 on the change of the L2
m+1
^
were imposed (see system (5.8) in
norm of the solution variation W
cc
Section 5.1). The GMRES algorithm was preconditioned with the additive
Schwarz method [28] and its convergence criterion was a linear system
residual drop of five orders of magnitude. An upper limit of two thousand
of Krylov sub-spaces was used. The computations were done on two processors.
The convergence history versus the number of iterations and the wall time
to drop the L2 norm of the residuals 10 orders of magnitude are plotted
in Figures 8.4 and 8.5, for second- to fourth-order SD methods. The upper
limit of the CFL number max for the non-linear LU-SGS + BE scheme was
set to 80, 40 and 20, respectively for second- third- and fourth-order accurate solutions. For the Newton-Raphson GMRES solver these limits were
fixed to 150, 100 and 80. This was necessary to obtain a fast convergence
and to maintain stability.
Figure 8.4 shows that, for second- and third-order SD methods, the NewtonRaphson GMRES solver needs fewer iterations as compared to the nonlinear LU-SGS solver. On the contrary, for fourth-order SD scheme, the
non-linear LU-SGS solver converged in fewer iterations as compared to
the Newton-Raphson GMRES solver. In term of wall time, the NewtonRaphson GMRES solver is three times faster than the non-linear LU-SGS
solver, for second-order SD method. However, for third-order SD method,
the former algorithm needs almost the same amount of time as compared
to the non-linear LU-SGS solver. In fact, the GMRES method had difficulties to converge and used about one thousand Krylov subspace at almost
each Newton iteration. This explains the large computational time. For
125

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

fourth-order accurate calculation, the non-linear LU-SGS solver performed
better, although it used almost thirty SGS sweeps at each time iteration.
2

p=1
p=2
p=3

2
Log(||Res||2)

Log(||Res||2)

10
0

100

200

Iter

300

400

10
0

500

p=1
p=2
p=3

(a) Non-linear LU-SGS.

100

200

300

Iter

400

500

600

700

(b) Newton-Raphson GMRES.

Figure 8.4: Non-linear LU-SGS and Newton-Raphson GMRES convergence histories for the steady laminar flow over a NACA0012 airfoil at zero angle of attack.
SD method with p-refinement.

p=1
p=2
p=3

2
Log(||Res||2)

Log(||Res||2)

10
0

5000

10000
15000
Wall Time [s]

20000

10
0

25000

(a) Non-linear LU-SGS.

p=1
p=2
p=3

10000

20000 30000 40000

Wall Time [s]

50000

60000

(b) Newton-Raphson GMRES.

Figure 8.5: Non-linear LU-SGS and Newton-Raphson GMRES convergence histories for the steady laminar flow over a NACA0012 airfoil at zero angle of attack.
SD method with p-refinement.

126

8.1. STEADY LAMINAR FLOW SIMULATIONS

The total memory requirements for these computations are summarized
in Table 8.2. It can be seen that the Newton-Raphson GMRES method
requires far more memory than the non-linear LU-SGS method.
Table 8.2: Non-linear LU-SGS and Newton-Raphson GMRES memory requirements [MB] for the steady laminar NACA0012 airfoil computations. SD method
with p-refinement.

SD method
p = 1 (11, 700 DOFs)

Non-linear LU-SGS
30

Newton-Raphson GMRES
174

p = 2 (26, 325 DOFs)

65

771

p = 3 (46, 800 DOFs)

155

2,325

8.1.2 Flow through a 90 bending square duct

The experiment of Humphrey et al. [78], which measured the flow through
a strongly curved 90 bend square duct, is used as a 3D laminar steady
state test case in the present section. In Figure 8.6 the configuration of
the test case is illustrated. The side length of the square cross section is

ri = 1.8H

5H

F low

10H

Figure 8.6: Configuration of the 3D 90 bending square duct test case.

127

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

denoted by H. A straight duct of 10H length is located ahead of the bend
in order to produce inlet laminar flow that is fully developed. The bend is
located in the vertical plane with a 5H length of straight duct attached to
its downstream. The bend has an inner radius of ri = 1.8H and an outer
radius of ro = 2.8H.
A Reynolds number Re of 790 and a Prandtl number P r of 0.72 are imposed. Uniform mass density and velocity profiles are prescribed at the
inlet section so that the Reynolds number based on the side of the square
cross section H is equal to the prescribed value. In order to obtain a quasiincompressible flow field, the inlet Mach number Minlet is set to 0.1. At the
outflow the pressure is prescribed to be uniform. At the walls, which are
assumed to be adiabatic, no-slip boundary conditions are imposed.

(a) Cross section.

(b) Section in the symmetry plane.

Figure 8.7: Grid for the steady laminar flow in square pipe with 90 bend. 4, 505
hexahedral cells with ARmax = 13.

A coarse grid with 4, 505 hexahedral cells and ARmax = 13 was used (see
Figure 8.7). This test case was computed with second- (p = 1) and thirdorder (p = 2) SD methods. In Figure 8.8, the Mach number contours obtained with these schemes are shown. It can be seen that the solution gets
smoother by increasing the order of the polynomial reconstruction, indicating the solution is more accurate. Note that, the third-order solution
was computed using a second-order polynomial approximation (quadrating mapping) of the boundary elements.

128

8.1. STEADY LAMINAR FLOW SIMULATIONS

M
0.15
0.13
0.11
0.09
0.07
0.05
0.03
0.01

(a) p = 1.

M
0.15
0.13
0.11
0.09
0.07
0.05
0.03
0.01

(b) p = 2.

Figure 8.8: Mach number contours in the symmetry plane under p-refinement, for
the steady laminar flow in square pipe with 90 bend, obtained with SD methods
and the LU-SGS + BE scheme. M = 0.02.

In order to study the order-independent flow convergence, the stream-wise

(axial) velocity profile at four locations, namely at 0 , 30 , 60 and 90 along
the bend, was analyzed. In Figure 8.9, the numerically obtained streamwise velocity profiles in the symmetry plane are shown. In this figure the
experimental data of Humphrey et al. [78] are also plotted for comparison.
The horizontal axis is the normalized radial distance and the vertical axis
is the normalized velocity in the stream-wise direction. The second-order
computation, with only 36, 040 DOFs, is under-resolved. For the thirdorder computation, which had 121, 635 DOFs, the match with the experiment is very good at the first two stations. However, when the flow begins
to form swirls at = 60 , a slight discrepancy between the computed and
experimental results appears. This deviation can also be found in other
numerical calculations [145, 196] and is thought to be caused by any small
129

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

change in the Reynolds number. Nevertheless, the peak of stream-wise
velocity near the outside wall and the second peak near the inside wall at
= 60 and = 90 are well captured by third-order SD scheme.

Experiment
p=1
p=2

Axial velocity/Bulk velocity

Axial velocity/Bulk velocity

1.5

0.5

0
0

Experiment
p=1
p=2

1.5

0.5

0.2

0.4
0.6
(r ri ) / (r0 ri )

0.8

0
0

0.2

(a) = 0 .
Experiment
p=1
p=2

1.5

0.5

0
0

0.8

0.8

(b) = 30 .

Axial velocity/Bulk velocity

Axial velocity/Bulk velocity

0.4
0.6
(r ri ) / (r0 ri )

Experiment
p=1
p=2

1.5

0.5

0.2

0.4
0.6
(r ri ) / (r0 ri )

0.8

0
0

(c) = 60 .

0.2

0.4
0.6
(r ri ) / (r0 ri )

(d) = 90 .

Figure 8.9: Axial velocity profiles at four locations along the bend under prefinement, for the steady laminar flow in square pipe with 90 bend, obtained
with SD methods and the LU-SGS + BE scheme. Comparison with experimental
measurements [78].

130

8.1. STEADY LAMINAR FLOW SIMULATIONS

Algebraic solver performance
The non-linear LU-SGS and the Newton-Raphson GMRES algorithms, combined with the BE scheme, were both used to perform the pseudo time
integration to get steady-state solutions. A local time stepping technique
was used. The CFL-law was the same as for the steady laminar flow past
a NACA0012 airfoil presented in the previous section. For the non-linear
LU-SGS algorithm, a maximum of hundred SGS sweeps per time iteration and/or a prescribed tolerance of 105 on the change of the L2 norm of
m+1
^
were imposed. The GMRES linear system
the solution variation W
cc
solver was used in combination with the additive Schwarz preconditioning.
Its convergence criterion was a linear system residual drop of five orders
of magnitude, with a maximum of two thousand Krylov sub-spaces. The
computations were done on nine processors. For both algebraic solvers, the
upper limit of the CFL number max was set to 800 and 20 for second- and
third-order SD methods respectively.
The convergence history versus the number of iterations and the wall time
to drop the L2 norm of the residuals 10 orders of magnitude are plotted in
Figures 8.10 and 8.11 respectively. It can be observed that the non-linear
LU-SGS solver converged with both second- and third-order SD methods,
whereas the Newton-Raphson GMRES solver did converge only for secondorder SD method. Moreover, for second-order accurate solution the nonlinear LU-SGS solver performed better than the Newton-Raphson GMRES
solver. The GMRES solver had some difficulties to converge and used almost thousand five hundred Krylov sub-spaces at each Newton iteration.
Note that, for the non-linear LU-SGS solver, third-order accurate solution
required twice the number of iterations needed by the second-order accurate computation. However, since the total computational load required
by both SD method and non-linear LU-SGS solver does not scale linearly
with the order of accuracy, the wall time required by the third-order accurate solution is not twice of that of the second-order accurate simulation.
The total memory requirements for these computations are summarized
in Table 8.3. It can be seen that, the non-linear LU-SGS solver needs
much less memory than the Newton-RGMRES solver and the difference
between them is more pronounced than for the previous test case, which
was two-dimensional. Nevertheless, the amount of memory required by
the non-linear LU-SGS method is still quite large and it increases rapidly
with increasing polynomial orders p, as discussed in Section 5.1.
131

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

p=1
p=2

2
Log(||Res||2)

Log(||Res||2)

10
0

20

40

Iter

60

80

10
0

100

p=1
p=2

(a) Non-linear LU-SGS.

200

400

Iter

600

800

1000

(b) Newton-Raphson GMRES.

Figure 8.10: Non-linear LU-SGS and Newton-Raphson GMRES convergence histories for the steady laminar flow in square pipe with 90 bend. SD method with
p-refinement.

p=1
p=2

2
Log(||Res||2)

Log(||Res||2)

10
0

5000

10000 15000 20000

Wall Time [s]

25000

10
0

30000

(a) Non-linear LU-SGS.

p=1
p=2

50000
100000
Wall Time [s]

150000

(b) Newton-Raphson GMRES.

Figure 8.11: Non-linear LU-SGS and Newton-Raphson GMRES convergence histories for the steady laminar flow in square pipe with 90 bend. SD method with
p-refinement.

132

8.2. UNSTEADY LAMINAR FLOW SIMULATIONS

Table 8.3: Non-linear LU-SGS and Newton-Raphson GMRES memory requirements [MB] for the steady laminar square pipe with 90 bend computations. SD
method with p-refinement.

SD method
p = 1 (36, 040 DOFs)

Non-linear LU-SGS
250

Newton-Raphson GMRES
2,648

p = 2 (121, 635 DOFs)

1,919

11,452

8.2 Unsteady laminar flow simulations

In this section, the 2D unsteady laminar flow over an open cavity, a square
cylinder and a circular cylinder are considered. Time marching scheme
was done using the LU-SGS + BDF2 scheme discussed in Section 5.2. The
non-linear system (5.16) was solved with a prescribed tolerance of 106
m+1
^
and/or a
on the change of the L2 norm of the solution variation W
cc
maximum number of hundred SGS sweeps. During the calculations, the
maximum number of SGS sweeps was never required. However, during
the initial time steps (depending on the test case and the initial solution)
the number of inner LU-SGS sweeps was between thirty and forty. Afterwards, the number decreased and reached a values which was between
seven and ten (depending again on the test case). The computations were
done with eight processors.
The flow solution of the first two test cases was used by the present author
to provide the acoustic sources for aerodynamic sound field simulation in
the time domain with a Ffowcs-Williams Hawkings (FW-H) approach, see
Parsani et al. [123].

8.2.1 Flow over an open cavity

The flow in an open cavity is dominated by vorticity production and transport and is very complex. So far, several numerical and experimental investigations have been performed by researchers [4, 8, 55, 96, 141, 148].
The case studied in the present section has a length to depth ratio L/D = 4,
a inlet Mach number of Minlet = 0.15, and a Reynolds number Re =
1.5 103 . The Reynolds number is based on the module of the far-field
velocity vector |~u | and the depth of the cavity D. At this Reynolds number, the flow is laminar. Because no experimental data exist for this case,
the 2D DNS data presented in Larsson et al. [96] are used as reference
solution.
133

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

16D

1
4.3D

20D

4D

Figure 8.12: Configuration of the 2D cavity test case.

The computational domain is shown in Figure 8.12, where the flow is from
left to right. The cavity is discretized in the stream-wise and normal directions with 50 and 30 points respectively. The maximum aspect ratio ARmax
of the first layer of the cells close to the walls is 30. At the right boundary,
a damping zone (buffer layer) is introduced by progressively increasing the
size of the cells towards the outflow. This type of buffer layer increases the
numerical damping introduced by the solution method, avoiding strong
spurious wave reflections which might contaminate the flow field. The
buffer layer contains 5 cells and its length is 4D. The total number of
quadrilateral cells is 11, 586. At the inflow boundary, the incoming boundary layer thickness and the mass density profiles are set equal to those
of the reference DNS solution [96]. At the outlet, which is sufficiently far
from the cavity, only the pressure is prescribed. At the top boundary the
1D characteristic boundary condition is used. The walls are assumed to be
adiabatic and the no-slip boundary condition is used there.
The test case was solved with third- (p = 2) and fourth-order (p = 3) SD
methods. Therefore, the total number of DOFs is 104, 274 and 185, 376, respectively for third- and fourth-order computations. Notice that, the number of DOFs for the highest-order SD scheme is much lower than the number of DOFs used in the DNS of Larsson et al. [96] and Ask et al. [9].
In the former reference, the number of DOFs is 1, 120, 080, whereas in the
latter one, it is of the order of 500, 000. The time step t used for the computations started from 0.00001 and increased linearly till 0.005. This time
step allowed about 3, 280 time samples per flow period.
134

8.2. UNSTEADY LAMINAR FLOW SIMULATIONS

M
0.22
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02

Figure 8.13: Mach number contour and stream traces, for the unsteady laminar
flow over an open cavity, obtained with fourth-order (p = 3) SD method. M =
0.02.

In Figure 8.13, the Mach number contour and the stream traces obtained
with fourth-order SD method are plotted. This figure shows that, despite
the geometric simplicity, many complicated flow patterns characterize the
flow.
In Figure 8.14, the cavity drag coefficient per unit width CD at statistically stationary state computed with fourth-order SD method is shown.
The Strouhal number associated to the frequency of the fluctuating drag
coefficient fCD and defined by
St =

fCD L
,
|~u |

the time averaged drag coefficient hCD i and the mean drag pressure coP
efficient2 hCD
i are listed in Table 8.4. In this table, reference coefficients
are also presented for comparison [96]. It can be seen that the mean drag
coefficients and the Strouhal number are in good agreement with the DNS
solution. Moreover, the accuracy of the solution improves by increasing
the order of the SD scheme. In fact, the coefficients obtained with fourthorder accurate solution are very close to those reported in Larsson et al.
[96]. This is a excellent achievement since the number of spectral difference DOFs is significantly lower than that of the DNS computations [9, 96].
2 The

pressure drag coefficient is evaluated through integration of the components in the

free-stream direction of the pressure forces acting on the after-body.

135

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

CD

0.5

0.5

20

10

0
t |~u | /D

10

20

Figure 8.14: Cavity drag coefficient obtained with fourth-order (p = 3) SD method.

Table 8.4: Effect of the p-refinement on the Strouhal number and the mean drag
coefficients for the unsteady laminar flow over an open cavity, obtained with SD
methods. Comparison with DNS reference solution [96].

Solution
DNS [96]

St
0.243

hCD i
0.377

P
hCD
i
0.402

SD p = 2 (104, 274 DOFs)

0.246

0.385

0.409

SD p = 3 (185, 376 DOFs)

0.243

0.379

0.403

The total memory requirements for the present computations are summarized in Table 8.5. It should be noted that the fourth-order accurate
simulation required about three times more memory as compared to the
third-order one.
Table 8.5: Non-linear LU-SGS memory requirements [MB] for the unsteady laminar flow simulations over an open cavity. SD method with p-refinement.

SD method
p = 2 (104, 274 DOFs)

Non-linear LU-SGS
257

p = 3 (185, 376 DOFs)

612

136

8.2. UNSTEADY LAMINAR FLOW SIMULATIONS

8.2.2 Flow past a square cylinder

The flow past a square cylinder represents a benchmark test case for external flow past bluff bodies. This flow has been the subject of intense
experimental and numerical research in the past [117, 118, 161] and it is
a relevant application in the field of aeroacoustics. A practical example
can be found in automotive applications such as the noise generated by a
luggage carrier system.
In Figure 8.15 the configuration of the test case is illustrated. At the
left boundary (the inflow), the flow is prescribed to be uniform. The same
boundary condition is applied to the upper and lower boundaries of the
domain. At the right boundary, which is sufficiently far from the cylinder,
uniform pressure profile is prescribed. At the cylinder wall, which is assumed to be adiabatic, the no-slip boundary condition is imposed.
2

11H

10H

25H

Figure 8.15: Configuration of the 2D square cylinder test case.

The flow has a free-stream Mach number of M = 0.5, and a Reynolds

number of Re = 200. The Reynolds number is based on the module of
the free-stream velocity vector |~u | and the height of the cylinder H. At
this Reynolds number, the regime is laminar and it usually persists up to
a Reynolds number of about 400. Moreover, the vortex shedding is characterized by one very well-defined frequency [117] and the flow field can
be described well with a 2D model. In fact, in Murakami et al. [114],
it has been shown that the 2D approach gives an accurate description of
the physical behavior of the problem, when the vortex shedding is concentrated around one frequency, which is the case for Re . 400. Because no
137

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

experimental data exist for this case, the 2D DNS results presented in De
Roeck et al. [144] are used as reference solution.
The test case was solved using second- (p = 1) to fifth-order (p = 4) SD
methods on a very coarse mesh with 2, 205 quadrilateral cells, as shown in
Figure 8.16. Notice that, the number of DOFs for the fifth-order accurate
solution is 55, 125 which is about 22% of the number of DOFs used for the
reference DNS solution [144]. The maximum cell aspect ratio ARmax close
to the cylinder was small. The time step t used for the computations
started from 0.00001 and increased linearly till 0.08. This allowed about
180 time samples per period of the vortex shedding.

Figure 8.16: Grid for the unsteady laminar flow past a square cylinder. 2, 205
quadrilateral cells and ARmax = 3.

The instantaneous Mach number contours at the end of the simulations,

obtained with second- to fifth-order accurate SD schemes, are shown in
Figure 8.17. The initial phase of the computations, where the solution
changes from a uniform flow field, to a fully developed unsteady flow field
with a von Karman vortex street, is difficult to resolve and changes significantly for different order of accuracy. Because of this, there is always a
phase shift between the solutions at the end of each calculation. The improvement in the quality of the solution is however clearly visible.
The evolution of the lift CL and the drag CD coefficients obtained with
fourth-order SD scheme, are shown in Figure 8.18. The Strouhal number
associated to the frequency of the fluctuating lift coefficient fCL , the time
averaged drag coefficient hCD i, the root mean square (RMS) value of the
138

8.2. UNSTEADY LAMINAR FLOW SIMULATIONS

0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05

0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05

(a) p = 1.

(b) p = 2.
M

0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05

0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05

(c) p = 3.

(d) p = 4.

Figure 8.17: Instantaneous Mach number contours under p-refinement, for the
unsteady laminar flow past a square cylinder, obtained with SD methods. M =
0.05.
RMS
drag coefficient CD
, and the RMS value of the lift coefficient CLRMS are
listed in Table 8.6. The Strouhal number associated to the frequency of the
fluctuating lift coefficient fCL is given by

St =

fCL H
.
|~u |

In Table 8.6, the aerodynamic results presented in De Roeck et al. [144]

are also listed for comparison. From this table, it can be seen that the
accuracy of the results improve by increasing the order of the SD scheme.
Moreover, although the grid is very coarse, with fifth-order SD scheme,
the accuracy of the aerodynamic coefficients is comparable with that of the
LES solutions.

139

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

0.6

1.8
1.6

0.4

1.4
0.2

CD

CL

1.2
0

1
0.2
0.8
0.4

0.6
0

0.6

100

200

Time

300

400

0.4
0

500

100

(a) CL .

200

Time

300

400

500

(b) CD .

Figure 8.18: Evolution of the lift CL and the drag CD coefficient in time, for the
unsteady laminar flow past a square cylinder, obtained with fifth-order (p = 4) SD
method.

Table 8.6: Effect of the p-refinement on the Strouhal number and the aerodynamic
coefficients for the unsteady laminar flow past a square cylinder, obtained with SD
methods. Comparison with DNS and LES reference solutions [144].

Solution
DNS [144]

St
0.146

hCD i
1.55

RMS
CD
0.019

CLRMS
0.232

Fine LES: non refl. BC [144]

0.158

1.51

0.009

0.135

Coarse LES: non refl. BC [144]

0.178

1.46

0.004

0.103

SD p = 1 (8, 820 DOFs)

0.136

1.06

0.000

0.060

SD p = 2 (19, 845 DOFs)

0.143

1.35

0.001

0.103

SD p = 3 (35, 280 DOFs)

0.142

1.49

0.008

0.194

SD p = 4 (55, 125 DOFs)

0.143

1.50

0.012

0.201

140

8.2. UNSTEADY LAMINAR FLOW SIMULATIONS

The total memory requirements for the present computations are summarized in Table 8.7. It should be noted that the required memory increases
rapidly with the polynomial order p.
Table 8.7: Non-linear LU-SGS memory requirements [MB] for the unsteady laminar flow simulations past a square cylinder. SD method with p-refinement.

SD method
p = 1 (8, 820 DOFs)

Non-linear LU-SGS
23

p = 2 (19, 845 DOFs)

49

p = 3 (35, 280 DOFs)

117

p = 4 (55, 125 DOFs)

319

8.2.3 Flow past a circular cylinder

The experimental work of Roshko [147] locates the beginning of the laminarto-turbulent transition at Reynolds numbers 200 300 based on the diameter of the cylinder D and the module of the free-stream velocity vector |~u |.
Beyond this Reynolds number but less than 3105 the wake of the cylinder
is completely turbulent and the boundary layer separation is laminar. It
is known that the flow around the cylinder is 2D only when Re < 200. For
larger Reynolds number, the vortex shedding is completely 3D. However,
in this section the results obtained for the flow over a circular cylinder
for Reynolds numbers 75, 150, 300, 800, 103 are investigated using a 2D
formulation with fourth-order (p = 3) SD scheme and quadratic boundary elements representation. The free-stream Mach number M and the
Prandtl P r are set to 0.05 and 0.72 respectively. The aim is to compare
the present results with the experimental measurements of Wieselberger
[192] and the 2D numerical solution of Henderson [72].
In Figure 8.19 the test problem is illustrated. At the left-hand-side boundary (the inflow) the flow is prescribed to be uniform and the same boundary
conditions are applied to the upper and lower boundaries. At the righthand-side boundary (the outflow) only the pressure is prescribed. At the
cylinder wall, which is assumed to be adiabatic, the no-slip boundary condition is imposed. A mesh with 4, 298 quadrilateral cells with a maximum
aspect ratio ARmax of 9 close to the circular cylinder was used for all computations. A detailed view of the grid is shown in Figure 8.20.

141

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

2

20D

10D

30D

Figure 8.19: Configuration of the 2D circular cylinder test case.

The time step t started from 0.00025 and increased linearly till 0.25. However, when the flows reached a statistically steady state solutions (after the
transient), the time step could be increased up to 0.65 without any stability problems. Nevertheless, for these calculations a time step of 0.25 was
employed to guarantee about 150 samples for each period of the vortex
shedding.

(a) Entire cylinder.

(b) North-east quarter of the cylinder.

Figure 8.20: Two details of the grid for the unsteady flow over a circular cylinder.
4, 298 quadrilateral cells with ARmax = 9.

142

8.2. UNSTEADY LAMINAR FLOW SIMULATIONS

s
0.0044
0.004
0.0036
0.0032
0.0028
0.0024
0.002
0.0016
0.0012
0.0008
0.0004
0

(a) Re = 300.

s
0.0044
0.004
0.0036
0.0032
0.0028
0.0024
0.002
0.0016
0.0012
0.0008
0.0004
0

(b) Re = 800.

s
0.0044
0.004
0.0036
0.0032
0.0028
0.0024
0.002
0.0016
0.0012
0.0008
0.0004
0

(c) Re = 103 .

Figure 8.21: Instantaneous entropy contours, for the unsteady flow past a circular
cylinder at different Reynolds numbers, obtained with fourth-order (p = 3) SD
method.

143

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

In Figure 8.21 the instantaneous entropy contours are shown to give an

impression of the flow field at three different Reynolds numbers, namely
300, 800 and 103 . In Table 8.8 the Strouhal numbers associated to the
fluctuating lift coefficient are compared with the experimental results of
Wieselberger [192]. From this table, it can be seen that the agreement between the present 2D simulations and the experimental values is good up
to Re = 300. However, for higher Reynolds number the frequency of the
vortex shedding is overestimated.
Table 8.8: Strouhal numbers for the unsteady flow past a circular cylinder at different Reynolds numbers, obtained with fourth-order (p = 3) SD method. Comparison
with experimental measurements [192].

Exp. [192]

Re = 75
0.148

Re = 150
0.180

St
Re = 300
0.200

Re = 800
0.204

Re = 1000
0.209

SD p = 3

0.148

0.181

0.201

0.233

0.240

In Figure 8.22 the time averaged drag coefficient hCD i of the circular cylinder is plotted in function of the Reynolds number. This figure shows that
3
Experiment
2D Henderson
2D SD with p = 3

2.5

hCD i

1.5

0.5

0 0
10

10

10

10

Re

10

10

10

10

Figure 8.22: Variation of the time averaged drag coefficient with Reynolds number
for the flow past a circular cylinder, obtained with fourth-order (p = 3) SD method.
Comparison with experimental measurements [147, 192] and the 2D numerical
simulations [72].

144

8.3. LARGE EDDY SIMULATIONS

for Re = 75, the mean drag coefficient of the present computation matches
very well with the experimental results of Wieselberger [192] and Roshko
[147]. For higher Reynolds numbers, the mean drag coefficients of the 2D
simulations overestimate the experimental ones but they agree very well
with the 2D numerical results of Henderson [72], in which high-resolution
computer simulations is used to quantify the change in drag around a 2D
circular cylinder, for different Reynolds numbers. The reason is that for
low Reynolds number (Re - 100), the 3D flow features are still weak and
the vortex axes are nearly aligned with the cylinder axis in the near wake.
For higher Re the boundary layer is sharper, the velocity gradients are
larger, resulting in the release of stronger vortices which are significantly
distorted and possess components besides that in the span-wise direction
and which cannot be taken into account by a 2D simulation.
The total memory requirement for the present computations is summarized in Table 8.9.
Table 8.9: Non-linear LU-SGS memory requirements [MB] for the unsteady flow
simulations past a circular cylinder.

SD method
p = 3 (68, 768 DOFs)

Non-linear LU-SGS
228

8.3 Large eddy simulations

In this section, two 2D turbulent flows, namely the flow past a square
cylinder at two different Reynolds numbers, and the 3D turbulent flow in
a muffler are considered. These test cases are computed by solving the filtered compressible Navier-Stokes equations in combination with the walladapted local eddy-viscosity model (WALE), as discussed in Sections 3.2
and 3.2.2. However, since one property of a good LES model is that its use
in a laminar or low Reynolds number flow results in a solution which is
very close to the solution obtained by solving the Navier-Stokes equations
[82], the sensitivity of the SD-LES method is first investigated by solving
the 2D laminar flow around a NACA0012 airfoil.
Time marching scheme was done using the LU-SGS + BDF2 scheme. The
non-linear system (5.16) was solved with a prescribed tolerance of 106
m+1
^
and/or a
on the change of the L2 norm of the solution variation W
cc
145

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

maximum number of hundred SGS sweeps. During the calculations, the
maximum number of symmetric Gauss-Seidel sweeps was never required.
However, during the initial time steps (depending on the test case and the
initial solution) the number of inner LU-SGS sweeps was between forty
and forty-five. Afterwards, the number decreased and it reached a values
which was between eight and twelve (depending again on the test case).
The 2D test cases were done with eight processors, while the 3D one was
carried out using forty processors.

8.3.1

Flow around a NACA0012 airfoil

The compressible laminar flow simulation over a symmetric NACA0012

airfoil is conducted at zero angle of attack and free-stream Mach number
of M = 0.4. The characteristic Reynolds number based on the chord c
and the module of the free-stream velocity |~u | is specified as Re = 5 104 .
The configuration of the test case is the same of that used in the previous
NACA0012 test cases, i.e. Figure 8.1. Therefore, the airfoil is placed on
the 1 axis (2 = 0) of the computational domain. At the left-hand-side
boundary (the inflow) the flow is prescribed to be uniform and the same
boundary conditions are applied to the upper and lower boundaries. At
the right-hand-side boundary (the outflow), sufficiently far from the airfoil, only the pressure is prescribed. At the airfoil wall, which is assumed
to be adiabatic, the no-slip boundary condition is imposed.
The test case was solved using fourth-order (p = 3) SD scheme and a
mesh with 19, 874 quadrilateral cells and quadratic boundary elements.
The maximum aspect ratio ARmax of the first layer of the cells close to the
airfoil is 30. The time step used for the computation started from 0.00001
and increased linearly till 0.00125. This time step allowed about 300 time
samples per period of the vortex shedding.
The instantaneous entropy contour obtained using the WALE model is
shown in Figure 8.23 to give an impression of the flow field around the
airfoil. The flow around the body is almost symmetric on the top and the
bottom sides of the airfoil and an unsteady wake develops downstream of
the trailing edge.
The non-dimensional value of the frequency associated to the oscillating
146

8.3. LARGE EDDY SIMULATIONS

s
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0

Figure 8.23: Instantaneous entropy contour for the unsteady laminar flow over a
NACA0012 airfoil at zero angle of attack, obtained with fourth-order (p = 3) SDLES method.

lift coefficient fCL , i.e. the reduced frequency

fCL ,r =

fCL 2c
,
2 )
(1 M

(8.2)

where c is the free-stream speed of sound, is listed in Table 8.10. In this

table the 2D DNS result of Sandberg et al. [153], where a fourth-order
numerical scheme without upwinding, artificial dissipation or explicit filtering [154] is used, is also indicated. For DNS, a mesh with 2, 243 692
non-equidistantly spaced points in the tangential and in the normal direction with respect to the airfoil surface was employed. From Table 8.10 it
can be seen that the reduced frequency is in good agreement with the DNS
value for both computations. In fact, the error with and without subgridscale model is respectively 1.04% and 1.06%. This is a good achievement
since the grid employed is very coarse compared to the grid used for a DNS
computation. Moreover, as was expected, the subgrid-scale model does not
affect the frequency of the vortex shedding when the laminar flow is well
resolved.
To conclude this study, in Figure 8.24 the non-dimensional time averaged
center-line (2 = 0) velocity component h
u1 i/ |~u | is compared with the 2D
DNS solution of Sandberg et al. [153]. It can be seen that, after an initial
peak, h
u1 i becomes negative before increasing to positive values further
downstream, i.e. a reverse flow region is present. This reverse flow region is caused by flow oscillating around the trailing-edge corner at the
wake frequency. Figure 8.24 shows also that the solutions obtained with
147

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

Table 8.10: Reduced frequencies for the unsteady laminar flow over a NACA0012
airfoil at zero angle of attack, obtained with fourth-order (p = 3) SD and SD-LES
methods. Comparison with DNS reference solution [153].

Solution
DNS [153]

fCL ,r
6.580

SD no model

6.650

SD-LES

6.649

and without subgrid-scale model are almost indistinguishable. Therefore,

the new coupling between the SD method and LES replicates the laminar
flows obtained by solving the pure Navier-Stokes equations. Moreover, the
comparisons show a good agreement between the present numerical results and the 2D DNS solution of Sandberg et al. [153].

0.3
DNS
SD no model
SD-LES

0.25
0.2

h
u1 i/ |~u |

0.15
0.1
0.05
0
0.05
0.1
0.15
0.05

0.05

1 /c

0.1

0.15

Figure 8.24: Time averaged center-line (2 = 0) velocity profiles h

u1 i/ |~
u |, for the
unsteady laminar flow over a NACA0012 airfoil at zero angle of attack, obtained
with fourth-order (p = 3) SD and SD-LES methods. Comparison with DNS reference solution [153].

The total memory requirement for the present computations is summarized in Table 8.11. The SD-LES calculation required slightly more memory as compared to the calculation without subgrid-scale model because
few more variable
148

8.3. LARGE EDDY SIMULATIONS

Table 8.11: Non-linear LU-SGS memory requirement [MB] for the unsteady flow
simulations over a NACA0012 airfoil at zero angle of attack.

SD method
p = 3 (317, 984 DOFs)

Non-linear LU-SGS
1,060

8.3.2 Flow around a square cylinder at Re = 104

The purpose of this test case is to compare the quality of the SD-LES approach with the 2D DNS solution presented by Wissink in [194]. In Figure
8.25, the configuration of the test case is illustrated. The cylinder is placed
2

12H

5.5H

15.5H

Figure 8.25: Configuration of the 2D square cylinder test case.

on the length axis (2 = 0) of the computational domain. At the left-handside boundary (the inflow) the flow is prescribed to be uniform. The same
conditions are applied to the upper boundary and to the lower boundary.
At the right-hand-side boundary, far enough from the cylinder, only the
pressure is prescribed. For the cylinder wall, which is assumed to be adiabatic, the no-slip boundary condition is used.
The free-stream Mach number M is set to 0.05, so that the flow is almost incompressible. The Prandtl number is set to the standard value
for air, i.e. 0.72. The Reynolds number Re, based on the module of freestream velocity |~u | and the height of the square cylinder H, is 104 . The
DNS of Wissink [194] was performed on a structured mesh with 400 400
grid points using a spatial discretization that consists of a combination
of a seventh-order upwind-biased method for the convective terms and an
149

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

Figure 8.26: Streamlines for the turbulent flow past a square cylinder at Re = 104 ,
obtained with third-order (p = 2) SD-LES method.

eighth-order central method for the diffusive terms. In the present work,
the test case was solved using third-order (p = 2) SD scheme with and
without the subgrid-scale model. A mesh with 12, 622 quadrilateral cells
with a maximum aspect ratio ARmax of 33 close to the square cylinder was
used. Notice that the total number of DOFs is 113, 598 which is almost 71%
of the DOFs used in [194]. The time step used for the computation started
from 0.00001 and increased linearly up to 0.0025.
At the beginning of the simulation (starting with uniform flow field), rather
small vortices of opposite sign are shed in parallel. However, this symmetric flow is not a stable solution at Re = 104 . After breaking the symmetry,
the flow behind the body behaves very chaotically and is fully turbulent. A
lot of very small vortices and filaments appear just behind the cylinder as
shown in Figure 8.26. Further away from the cylinder these small vortices
merge to form larger vortical structures.
In Figure 8.27, the number of inner LU-SGS sweeps as a function of the
time iteration is shown, for the calculation with subgrid-scale model. It can
be seen that, the maximum number of SGS sweeps was never used. However, during the initial time steps the number of inner LU-SGS sweeps
increased up to forty-six. Afterwards, it decreased to an averaged value of
ten. For this reason, the evolution of the LU-SGS sweeps in Figure 8.27 is
shown up to hundred and fifty time iterations. Note that, the transitional
behavior in Figure 8.27 depends on the initial solution. In the present
work, an uniform velocity field equals to the inlet boundary condition was
imposed at the beginning of the computation.

150

8.3. LARGE EDDY SIMULATIONS

50
45
40

LUSGS sweeps

35
30
25
20
15
10
5
0
0

50
100
Time Iterations

150

Figure 8.27: Number of inner LU-SGS sweeps as a function of the time iteration,
for the turbulent flow past a square cylinder at Re = 104 . Third-order (p = 2)
SD-LES method.

After breaking the symmetry, the resolved turbulent properties were obtained by statistically averaging, for approximately 28 shedding cycles. In
Figure 8.28, the non-dimensional time averaged velocity profiles h
u1 i/ |~u |
at the upper side of the cylinder are plotted. The 2D DNS solution presented in Wissink [194] is also shown for comparison. It can be observed
that, the SD method in combination with LES works very well and improves the accuracy of the results obtained without subgrid-scale model.
The improvement is clearly visible at trailing edge of the cylinder (Figures
8.28(c) and 8.28(d)), where the LES solution captures quite well the variation of the velocity profiles close to the wall, whereas the solution obtained
without subgrid-scale model differs from the DNS solution. Nevertheless,
it is very important to note that the SD scheme is already able to predict
all the features of the flows, showing the potential of high-order methods.
The h
u1 i/ |~u | profiles plotted in Figure 8.29 were gathered at various
stations 1 = constant behind the cylinder. This figure shows once more
that the SD-LES simulation works well and improves the accuracy of the
results obtained solving the pure Navier-Stokes equations. In fact, the
velocity profiles obtained with subgrid-scale model agree well with those
computed with the DNS.

151

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

1
0.95

DNS
SD no model
SDLES

0.95

0.9

0.9

0.85

0.85
0.8
2 /H

2 /H

0.8
0.75
0.7

0.7
0.65

0.6

0.6

0.55

0.55
0

(a)
1
0.95

0.5
h
u1 i/ | ~u |

1
H

0.5
0.5

1.5

DNS
SD no model
SDLES

0.95

0.9

0.9

0.85

0.85

2 /H

2 /H

0.7

0.7
0.65

0.6

0.6

0.55

0.55
0

(c)
1

0.5
h
u1 i/ | ~u |

1
H

0.5
0.5

1.5

1.5

= 0.36.

DNS
SD no model
SDLES

(d)

= 0.24.
1

DNS
SD no model
SDLES

0.95

0.9

0.9

0.85

0.85

0.5
h
u1 i/ | ~u |

1
H

1.5

= 0.12.

DNS
SD no model
SDLES

0.8
2 /H

0.8
0.75

0.75

0.7

0.7

0.65

0.65

0.6

0.6

0.55

0.55

0.5
0.5

1
H

0.75

0.65

0.95

0.5
h
u1 i/ | ~u |

0.8

0.75

0.5
0.5

(b)

= 0.48.

0.8

2 /H

0.75

0.65

0.5
0.5

DNS
SD no model
SDLES

(e)

0.5
h
u1 i/ | ~u |

1
H

0.5
0.5

1.5

(f)

= 0.0.

0.5
h
u1 i/ | ~u |

1
H

1.5

= 0.10.

Figure 8.28: Time averaged velocity profiles h

u1 i/ |~
u | at various stations along
the upper side of the square cylinder at Re = 104 , obtained with third-order (p = 2)
SD and SD-LES methods. Comparison with DNS reference solution [194].

152

8.3. LARGE EDDY SIMULATIONS

1

DNS
SD no model
SDLES

0.95

0.95

0.9

0.9

0.85

0.85
0.8
2 /H

2 /H

0.8

DNS
SD no model
SDLES

0.75

0.75

0.7

0.7

0.65

0.65

0.6

0.6

0.55

0.55

0.5
0.5

0.5
h
u1 i/ | ~u |

(a)
1

0.9

1
H

0.5
0.5

1.5

(b)

= 0.21.

0.5
h
u1 i/ | ~u |

1
H

1.5

= 0.32.

0.54

DNS
SD no model
SDLES

DNS
SD no model
SDLES

0.535
0.53
0.525

2 /H

2 /H

0.8

0.7

0.52
0.515

ZOOM

0.51

0.6

0.505
0.5
0.5

0.5
1
h
u1 i/ | ~u |

(c)
1

0.9

0.5
0.1 0.08 0.06 0.04 0.02
h
u1 i/ | ~u |

1.5

1
H

= 0.42.
0.54

DNS
SD no model
SDLES

DNS
SD no model
SDLES

0.535
0.53
0.525
2 /H

2 /H

0.8

0.7

0.52
0.515
0.51

ZOOM

0.6

0.505
0.5
0.5

0.5
1
h
u1 i/ | ~u |

0.5

1.5

(d)

1
H

0.05
0.1
h
u1 i/ | ~u |

0.15

= 0.49.

Figure 8.28: Time averaged velocity profiles h

u1 i/ |~
u | at various stations along
the upper side of the square cylinder at Re = 104 , obtained with third-order (p = 2)
SD and SD-LES methods. Comparison with DNS reference solution [194].

153

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

6

DNS
SD no model
SDLES

2
2 /H

2 /H

0.2

0.4

(a)
6

0.6
0.8
h
u1 i/ | ~u |

1
h

6
0

1.2

0.4

(b)
6

DNS
SD no model
SDLES

0.6
0.8
h
u1 i/ | ~u |

1
h

1.2

1.2

= 3.0.

DNS
SD no model
SDLES

2 /H

6
0

0.2

= 2.0.

2
2 /H

6
0

DNS
SD no model
SDLES

0.2

0.4

(c)

0.6
0.8
h
u1 i/ | ~u |

1
h

6
0

1.2

0.2

0.4

(d)

= 4.0.
6

0.6
0.8
h
u1 i/ | ~u |

1
h

= 5.0.

DNS
SD no model
SDLES

2 /H

6
0

0.2

0.4

(e)

0.6
0.8
h
u1 i/ | ~u |

1
h

1.2

= 6.0.

Figure 8.29: Time averaged velocity profiles h

u1 i/ |~
u | at various stations through
the turbulent wake of the square cylinder at Re = 104 , obtained with third-order
(p = 2) SD and SD-LES methods. Comparison with DNS reference solution [194].

154

8.3. LARGE EDDY SIMULATIONS

In order to study the effect of the polynomial order on the LES solution,
a simulation with second-order spatial discretization (p = 1) on the same
grid was also performed. Thus, the number of DOFs was reduced by a factor of 2.25. In Figures 8.30 and 8.31, the results obtained at three locations
on the side of the square cylinder and at two locations downstream in the
wake are shown. One can compare the solutions and see that the solution
obtained with the second-order SD-LES scheme is much less accurate than
the one obtained with the third-order SD-LES scheme. The second-order
accurate solution captures the main flow features but the velocity profiles
show a deviation from the DNS solution.
1
0.95

DNS
SD-LES p = 1
SD-LES p = 2

0.95

0.9

0.9

0.85

0.85
0.8
2 /H

2 /H

0.8
0.75

0.75

0.7

0.7

0.65

0.65

0.6

0.6

0.55

0.55

0.5

0.5

(a)
1

1
h

0.5
h
u1 i/ | ~u |

0.5
0.5

1.5

(b)

= 0.12.

0.5
h
u1 i/ | ~u |

1
h

1.5

= 0.10.

0.54

DNS
SD-LES p = 1
SD-LES p = 2

0.535

0.9

DNS
SD-LES p = 1
SD-LES p = 2

0.53
0.525

0.8
2 /H

2 /H

DNS
SD-LES p = 1
SD-LES p = 2

0.7

0.52
0.515
0.51

ZOOM

0.6

0.505
0.5
0.5

0.5
1
h
u1 i/ | ~u |

0.5
0.1 0.08 0.06 0.04 0.02
h
u1 i/ | ~u |

1.5

(c)

1
h

= 0.42.

Figure 8.30: Time averaged velocity profiles h

u1 i/ |~
u | at two selected locations
along the upper side of the square cylinder at Re = 104 , obtained with second(p = 1) and third-order (p = 2) SD-LES method. Comparison with DNS reference
solution [194].

155

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

6

DNS
SD-LES p = 1
SD-LES p = 2

2
2 /H

2 /H

6
0

0.2

0.4

(a)

0.6
0.8
h
u1 i/ | ~u |

1
h

DNS
SD-LES p = 1
SD-LES p = 2

6
0

1.2

0.2

0.4

(b)

= 3.0.

0.6
0.8
h
u1 i/ | ~u |

1
h

1.2

= 4.0.

Figure 8.31: Time averaged velocity profiles h

u1 i/ |~
u | at two selected locations
through the turbulent wake of the square cylinder at Re = 104 , obtained with
second- (p = 1) and third-order (p = 2) SD-LES method. Comparison with DNS
reference solution [194].

The total memory requirement for the present computations is summarized in Table 8.12.
Table 8.12: Non-linear LU-SGS memory requirement [MB] for the turbulent flow
simulations over a square cylinder at Re = 104 .

SD method
p = 2 (113, 598 DOFs)

Non-linear LU-SGS
284

Estimation of the SD-LES computational cost

To conclude the study of this test case, a simple estimation of the Reynolds
number scaling of the SD-LES computational cost is given. In Pope [132]
it is shown that the computational cost for a DNS, i.e. the total number
of modes to resolve the entire energy spectrum in 3D, is proportional to a
power of the Reynolds number
9/4

N mod,3D ReL ,
where ReL is the Reynolds number based on the length scale L characterizing the large eddies. For the present test cases L can be chosen equal
to the height of the square cylinder, i.e. H. For a classical FV scheme,
156

8.3. LARGE EDDY SIMULATIONS

N mod,3D corresponds to the number of DOFs. Using the same relation presented above, the number of DOFs needed to perform a DNS with the SD
scheme can be roughly estimated as
6/4

9/4

cells,3D
NSDDN
S

ReL
s
N (p, dim)

dim = 3,

cells,2D
NSDDN
S

ReL
s
N (p, dim)

dim = 2.

Consequently, for the flow past a square cylinder at Re = 104 , in order

to perform a 2D DNS with a third-order SD scheme, 3.7 104 cells are
needed. Hence, the averaged CPU-time per iteration, to perform the 2D
DNS simulation can be computed as
iter
CP USDDN
S

cells,2D
iter
NSDDN
CP USDno
model
= cells,2D S
,
iter
CP
U
NSDLES
SDLES

iter
iter
where CP USDno
model and CP USDLES are the averaged CPU-times per
time iteration required by the calculations without and with subgrid-scale
iter
model. For the present test case, CP USDDN
S is about 2.5. Therefore, the
SD-LES simulation is almost 2.5 times cheaper than the SD-DNS.

As mentioned in Section 3.2, in this work the LES approach is used as

a cheaper alternative to DNS. However, for sake of consistency, one should
compare the LES cost with that of a RANS calculation, which represents
the cheapest approach to numerically simulate the behavior of unsteady
turbulent flows. With the RANS approach, the Navier-Stokes equations
are ensemble-averaged, converting turbulent fluctuations into Reynolds
stresses, while leaving the large scale, rotational motions to be resolved
as unsteady phenomena. The LES approach, on the other hand, employs
a spatial averaging over a scale sufficient to remove scales not resolved
by the particular grid being used. The subgrid scale turbulence is then
modeled. A practical difference is in the degree of mesh resolution required: LES resolves the larger eddies of the turbulence itself, whereas
the unsteady RANS approach models the turbulence and resolves only unsteady, mean flow structures - primarily larger than the turbulent eddies.
Consequently, LES typically requires much higher grid resolution, at least
locally, and is therefore more costly. In addition, LES resolves the complete range of scales of random motion, up to the cut-off frequency, while
unsteady RANS aims to capture a single frequency (e.g. corresponding
to coherent shedding) and to model the random motions using standard
turbulence closures. Therefore, LES requires very long integration time
with a small time step to build a statistically-averaged solution; on the
157

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

other hand, a few shedding periods are usually enough to obtain accurate phase-averaged solution with RANS, thus limiting its overall computational cost. Because of the grid resolution and the long integration time,
an LES simulation may be two or more orders of magnitude slower as compared to a RANS simulation. However, one should take into account that
the classical RANS approaches are usually not able to distinguish between
quasi-periodic large scale and turbulent chaotic small scale features of the
flow field. This leads to huge problems when the flow field is governed by
both phenomena. A typical representative is a bluff-body flow. Generally,
the RANS approach is not able to reproduce the unsteady characteristics
of the flow field reasonably, resulting in an inadequate description of unsteady phenomena, such as vortex formation and shedding behind bluff or
inclined bodies [81].

8.3.3

Flow around a square cylinder at Re = 2.2 104

The purpose of this test case is to evaluate the quality of the SD-LES approach by comparing the mean flow field and the turbulent kinetic energy
results with some reference solutions. In literature 2D and 3D large eddy
simulations of the flow past a square cylinder have been performed, for
instance see Murakami et al. [114] and Rodi [142]. Moreover, Breuer and
Pourquie [26] and Bouris et al. [23] also performed 2D and 3D LES computations and compared the results with the experimental measurements
et al. [47] and Lyn et al. [106]. It was demonstrated that the
of Durao
characteristic of this type of flow is its quasi-two-dimensional character
and the presence of periodic vortex shedding from the front corners of the
square rod which introduces a low-frequency variation of the velocity field
behind the rod in addition to the high-frequency turbulence fluctuations.
In Breuer and Pourquie [26] and Bouris et al. [23], it has been stated that
2D LES calculations are clearly inferior to 3D ones since certain important features of 3D turbulence are not resolved. The three dimensionality of turbulence cannot be questioned, however, in the present work, we
want to show that the importance of detailed simulation of the quasi-twodimensional mechanisms can be achieve performing a 2D LES with the
new combination of SD and subgrid-scale model.
The same domain, boundary conditions and grid used for the previous test
case were employed. The free-stream Mach number M is set to 0.05 so
that the flow is again almost incompressible. The Reynolds number Re,
based on the module of the free-stream velocity | ~u | and height of the
square cylinder H, is 2.2 104 . The flow was computed using third-order
158

8.3. LARGE EDDY SIMULATIONS

(p = 2) SD scheme with and without the subgrid-scale model. The time
step used for the computation started from 0.00001 and increased linearly
up to 0.002. Figure 8.32 shows the number of inner LU-SGS sweeps in
function of the time iteration. One can see that the maximum number
of symmetric Gauss-Seidel sweeps was never reached. However, during
the initial steps the number of inner LU-SGS sweeps increased up to forty
five. Afterwards, the number of sweeps decreased to nine and it oscillated
around this value. For this reason, the evolution of the LU-SGS sweeps in
Figure 8.32 is shown up to hundred and fifty time iterations. Note once
more, that the transitional behavior in Figure 8.32 depends on the initial
solution. A uniform velocity field equals to the inlet boundary condition
was imposed at the beginning of the computation.
45
40

LUSGS sweeps

35
30
25
20
15
10
5
0
0

50

100

150

Time Iterations

Figure 8.32: Number of inner LU-SGS sweeps as a function of the time iteration,
for the turbulent flow past a square cylinder at Re = 2.2 104 . Third-order (p = 2)
SD-LES method.

In the present test case, the 2D LES calculation with a second-order finite volume scheme and the Smagorisky-Lilly model of Bouris et al. [23]
is used as reference 2D LES solution. In the latter work, a mesh with
350 300 points is employed. Therefore, the total number of DOFs of the
present SD calculation is practically the same as compared to the number
of DOFs used in Bouris et al. [23].
As for the previous calculation, inlet conditions are based on a uniform
flow with non-fluctuating velocity profiles. Table 8.13 shows the present
predictions of the dominant vortex shedding frequency in non-dimensional
form St and the time averaged drag coefficient hCD i, as well as the results
159

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

reported in Bouris et al. [23]. In this table, the abbreviation (k ) , RSE1
and RSE2 stand, respectively, for k model with 2 layers, Reynolds stress
equation with wall function and Reynolds stress equation with two layers
[49, 142]. The results of the SD-LES are in good agreement with the experimental measurements while the solution without the subgrid-scale model
underestimates both the Strouhal number St and the time averaged drag
coefficient hCD i.
Table 8.13: Strouhal numbers and mean drag coefficients, for the turbulent flow
past a square cylinder at Re = 2.2 104 , obtained with third-order (p = 2) SD and
SD-LES methods. Comparison with other numerical simulations [23] and experimental measurements [47, 106].

St
hCD i

(k )
0.124

RSE1
0.136

RSE2
0.159

2D LES [23]
0.134

SD
0.121

SD-LES
0.133

Exps.
0.132
0.139

1.179

2.15

2.43

2.18

1.98

2.21

2.05
2.23

Time averaged results were obtained integrating the data over approximately 20 shedding cycles and the mean center-line velocity is presented
in Figure 8.33. Note that, the horizontal normalized coordinate is denoted by 1 and corresponds to the coordinate 1 shifted by H/2 so that
1 = 1 + H/2. Although the SD scheme without subgrid-scale model overpredicts the value of the reverse velocity, further downstream it gives a
solution which is close to the 2D LES solution of Bouris et al. [23]. Moreover, it can be clearly seen that the modeling of the subgrid-scale stress
tensor improves the accuracy of the results. In fact, good agreement be et al. [47]
tween the predicted results and experimental data of Durao
throughout the comparison domain shows that the quality of a high-order
SD method increases when it is coupled with large eddy simulation. In the
other parts of the comparison domain the method without subgrid-scale
model is in a good agreement with LES of Bouris et al. [23]. Note that,
the SD-LES method captures the peak of the mean stream-wise velocity
considerably better than the others.
Figure 8.34 shows the 2D resolved total turbulent kinetic energy defined
by


1
ktot =
hu12 i + hu22 i ,
(8.3)
2
160

8.3. LARGE EDDY SIMULATIONS

SD no model
SD-LES
Exp. Dur
ao et al.
Exp. Lyn et al.
2D LES Bouris et al.
3D LES Murakami et al.
3D LES Breuer et al.

0.8

h
u1 i/ | ~u |

0.6
0.4
0.2
0
0.2
0.4
2

1 /H

Figure 8.33: Time averaged center-line (2 = 0) velocity profiles h

u1 i/ |~
u | behind
the square cylinder at Re = 2.2 104 , obtained with third-order (p = 3) SD and
SD-LES methods. Comparison with other numerical simulations [23] and experimental measurements [47, 106].

SD no model
SD-LES
Exp. Dur
ao et al.
Exp. Lyn et al.
2D LES Bouris et al.
3D LES Murakami et al.

0.8

ktot / | ~u |2

0.6
0.4
0.2
0
0.2
0.4
2

1 /H

Figure 8.34: Total turbulent kinetic energy ktot / | ~

u |2 along the center-line (2 =
0) behind the square cylinder at Re = 2.2 104 , obtained with third-order (p = 3)
SD and SD-LES methods. Comparison with other numerical simulations [23] and
experimental measurements [47, 106].

161

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

where hu1 i and hu2 i are the time averaged fluctuations in 1 and 2 directions respectively. One can compare the results with the experimental
data and see that the SD-LES method undershoots the peak experimental
value somewhat less than the 2D LES of Bouris et al. [23] and predicts the
results well elsewhere. The approach without subgrid-scale model overshoots the peak experimental value less than does the 2D LES of Bouris et
al. [23], whereas throughout most of the domain both approaches are on
the top of each other.
Overall, the results indicate that although the SD method without subgridscale model is already able to predict all the features of the flow, the SDLES approach gives better results, for the stream-wise mean velocity and
the total kinetic energy.

8.3.4

Flow in a muffler

Mufflers are commonly used in a wide variety of applications. Industrial

flow ducts as well as internal combustion engines frequently make use of
silencing elements to attenuate the noise levels carried by the fluids and
radiated to the outside atmosphere by the exhausts. Restrictive environmental legislation requires that silencer designers use high performance
and reliable techniques. Various techniques are currently available for the
modeling and testing of duct mufflers. Empirical, analytical and numerical
techniques have been used and proven reliable under controlled conditions.
Design of a complete muffler system is, usually, a very complex task. Each
element is selected by considering its particular performance, cost and its
interaction effects on the overall system performance and reliability.
The main purpose of this section is to evaluate the accuracy and the reliability of the implicit SD-LES solver by performing the simulation of a
confined flow in an industrial geometry. The 3D turbulent flow in a muffler is considered as a test case. The results are compared with the particle
image velocimetry (PIV) measurement performed by Dr. Michael Bilka at
the Department of Environmental and Applied Fluid Dynamics of the von
Karman Institute for Fluid Dynamics [22]. In Figure 8.35, the geometry
of the muffler and its characteristic dimensions are illustrated, where the
flow is from left to right. This configuration was selected because is representative of the physical mechanisms involved in a generic muffler and it
is a benchmark test case of the IWT Project SBO 050163 (Simulation and
design tools towards the reduction of aerodynamic noise in confined flows).

162

8.3. LARGE EDDY SIMULATIONS

0.625 d

0.625 d

5d

7.5 d

5d

4.75 d

Figure 8.35: Configuration of the 3D muffler test case.

At the inlet, mass density and velocity profiles are imposed. The inlet
velocity profile in the 3 direction is given by


 
1 1
d/2
r
u3 = umax
.
tanh 2.2

2 2
d/2
r
At the the outlet only the pressure is prescribed. In accordance to the experiments, the inlet Mach number Minlet and the Reynolds number, based
on maximum velocity at the inlet umax and the diameter of the inlet/outlet
d, are set respectively to 4.64 104 and 0.05.
The flow is computed using second-order (p = 1) SD scheme on two grids
with 195, 072 and 337, 505 hexahedral elements respectively. Therefore, the
total number of DOFs is approximately 1.6 and 2.7 millions. The maximum aspect ratio is 22 for the coarse mesh and 15 for the other one. The
fine grids was obtained by refining the coarse one mainly in the expansion
chamber, where the flow gets turbulent. In Figure 8.36 the coarsest mesh
is shown. The time-step used for the computations started from 0.00001
and increased linearly up to 0.001. Notice that, although the Mach number
is rather small, no specific difficulties of convergence and/or accuracy were
observed.
The computation is validated on the center plane of the expansion coinciding with the center planes of the inlet and outlet pipes. The laser sheet
was generated by a 200 mJ Nd-YAG double-pulsed laser passing through a
spherical and cylindrical lens. A smoke generator was used to see the flow
with fine oil droplets (1 m). A digital CCD camera was used (PCO Sensicam, resolution - 1280 1024) to record the images with a region of interest
limited to 1280 832. The pulse separation of the between the images of
163

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

(a) 1 -2 plane

(b) 2 -3 plane.

Figure 8.36: Plane views of the 195, 072 cells grid used for the muffler simulation
[54].

the PIV couple correspond to a maximum displacement of 6 pixels. In order

to keep a high spatial resolution over the measurement domain, the measurements are divided into 3 zones and with a field of view of 100 55 mm2
with a resolution of 12 pixels per millimeter. All of these measurements are
taken on the symmetrical center plane of the muffler. It should be noted
that the circular nature of the geometry acts as a lens causing a change
in magnification which prevents from capturing images close to the wall.
It is found that outside 1cm from the wall the magnification effect is negligible and as the mean flow direction is in the direction of constant magnification (i.e along the height of the cylinder) no corrections are deemed
necessary. For each region, 1,000 image pairs have been acquired and preprocessed with the home-made cross-correlation algorithm WIDIM (Window Displacement Iterative Multigrid). This program is based on an iterative multigrid predictor-corrector method, handling the window distortion, for better resolution of shear flows, and the sub-pixel window displacement, to limit pixel-locking. The predictor-corrector method is then
validated for each grid size if the signal-to-noise (SN) ratio is above 1.5
164

8.3. LARGE EDDY SIMULATIONS

X

Velocity Magnitude: 2 4 6 8 10 12 14 16 18 20 22 24 26 28

Figure 8.37: Contour of the time averaged velocity vector magnitude for the turbulent flow in a muffler, obtained with second-order (p = 1) SD-LES method. Grid
with 337, 505 hexahedral cells and ARmax = 12.

[156]. Mean quantities are then computed from the PIV data and profiles
are extracted to compare with the LES.
Figure 8.37 shows the contour of the time averaged velocity vector magnitude in the symmetry plane obtained with second-order (p = 1) SD-LES
method on the grid with 337, 505 hexahedral cells. In Figure 8.38, the
non-dimensional mean velocity profile in the axial direction h
u3 i/umax is
shown for six different cross sections in the expansion chamber, where
the PIV measurements were done. In this figure, the PIV data are also
plotted for comparison. Figure 8.39 shows the non-dimensional Reynolds
stress hu2 u3 i/u2max at the same cross sections. It can be seen that the mean
velocity profiles computed with the SD-LES approach, are in good agreement with the experiments throughout the comparison domain for both
grids. The effect of the mesh refinement is clearly visible for the Reynolds
stresses. In fact, although the computation with the coarse mesh captures
reasonably well the two peaks in the shear layer, it underestimates widely
the Reynolds stress away from the jet region. On the contrary, the calculation with the fine grid gives good results throughout the domain of
interest. This is a good achievement since the number of DOFs for the fine
mesh is only 2.7 millions.

165

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

1.2

0.8

0.8

0.6

0.6
hu3 i/umax

hu3 i/umax

1.2
SDLES coarse grid
SDLES fine grid
PIV

0.4
0.2

0.4
0.2

0.2

0.2
0.4

0.4
3

2 /d

(a) 1d downstream.
1.2

SDLES coarse grid

SDLES fine grid
PIV

0.8

0.8

0.6

0.6

0.4
0.2

0
0.2

0.4
1

2 /d

SDLES coarse grid

SDLES fine grid
PIV

1
0.8

0.6

0.6
hu3 i/umax

hu3 i/umax

SDLES coarse grid

SDLES fine grid
PIV

2 /d

(d) 4d downstream.
1.2

0.8

0.4
0.2

0.2

0
0.2

0.4

SDLES coarse grid

SDLES fine grid
PIV

0.4

0.2

0.4
2

(c) 3d downstream.

2 /d

0.2

0.2

1.2

0.4

(b) 2d downstream.
1.2

hu3 i/umax

hu3 i/umax

SDLES coarse grid

SDLES fine grid
PIV

0.4
2

2 /d

(e) 5d downstream.

2 /d

(f) 6d downstream.

Figure 8.38: Time-averaged velocity profile in the axial direction at six cross sections in the expansion chamber (h
u3 i/umax vs. 2 /d), for the turbulent flow in a
muffler, obtained with second-order (p = 1) SD-LES method on coarse and fine
grids. Comparison with experimental measurements (PIV) [22].

166

8.3. LARGE EDDY SIMULATIONS

0.02

0.02

SDLES coarse grid

SDLES fine grid
PIV

0.015

0.01

0.01

0.005

0.005

hu2 u3 i/u2max

hu2 u3 i/u2max

0.015

0
0.005

0
0.005

0.01

0.01

0.015

0.015

0.02
3

2 /d

0.02
3

(a) 1d downstream.
0.02

SDLES coarse grid

SDLES fine grid
PIV

0.015

0.01

0.01

0.005

0.005

0
0.005

0.015

0.015

2 /d

0.02
3

(c) 3d downstream.
0.02

0.02

SDLES coarse grid

SDLES fine grid
PIV

0.015

0.01

0.01

0.005

0.005

0
0.005

0.01
0.015

2 /d

SDLES coarse grid

SDLES fine grid
PIV

2 /d

0.02
3

(e) 5d downstream.

SDLES coarse grid

SDLES fine grid
PIV

0.005

0.015

0.01

0.02
3

(d) 4d downstream.

hu2 u3 i/u2max

hu2 u3 i/u2max

0.015

2 /d

0.01

0.005

0.01

0.02
3

(b) 2d downstream.
0.02

hu2 u3 i/u2max

hu2 u3 i/u2max

0.015

SDLES coarse grid

SDLES fine grid
PIV

2 /d

(f) 6d downstream.

Figure 8.39: Reynolds stress at six cross sections in the expansion chamber
(hu2 u3 i/u2max vs. 2 /d), for the turbulent flow in a muffler, obtained with secondorder (p = 1) SD-LES method on coarse coarse and fine grids. Comparison with
experimental measurements (PIV) [22].

167

CHAPTER 8. APPLICATION II: SPECTRAL DIFFERENCE METHOD

The total memory requirement for the present computations are summarized in Table 8.14. This table clearly illustrates that, although the simulations were done with a second-order SD-LES scheme, a large amount of
memory is required.
Table 8.14: Non-linear LU-SGS memory requirements [MB] for the turbulent flow
simulations in a muffler. Second-order (p = 1) SD method with grid refinement.

Grid
195, 072 hexahedral cells (1, 560, 576 DOFs)
337, 505 hexahedral cells (2, 700, 040 DOFs)

168

Non-linear LU-SGS
10,831
18,740

Chapter 9

Conclusions and future

directions
The present thesis is focused on the development of an efficient NavierStokes (N-S) / large eddy simulation solver on unstructured grids for highorder accurate schemes. Two spatially high-order accurate methods for
unstructured grids have been considered, namely the spectral volume (SV)
and the spectral difference (SD) methods. To fully exploit the potential benefits of such high-order spatial accurate operators, an efficient non-linear
algebraic solver, denoted as non-linear LU-SGS algorithm, has been investigated. This solver has been used to invert the non-linear algebraic systems arising from implicit time discretizations with the well-known backward Euler (BE) scheme and the second-order backward difference formula
(BDF2). The evaluation of the solver performances, both with analysis and
computation, have been two main cores of this thesis.
The need for paying particular attention to subgrid models for the simulation of general turbulent has been recognized. In this framework, a highorder SD scheme has been coupled with the local eddy-viscosity (WALE)
model to perform large eddy simulations and a new procedure to calculate
the grid filter width has been proposed. The accuracy and the reliability
of the implicit SD solver have been tested by solving two-dimensional (2D)
and three-dimensional (3D) turbulent test cases.
169

CHAPTER 9. CONCLUSIONS AND FUTURE DIRECTIONS

9.1 Achievements
The damping properties of the non-linear LU-SGS algorithm combined
with the BE scheme (LU-SGS + BE) have been analyzed by means of a Von
Neumann analysis for a model 2D linear advection. The analysis has been
performed with a CFL number of one million, on a mesh with cell aspect
ratio of hundred. The latter choice has enabled to account for the effects of
the geometrical stiffness imposed by the Navier-Stokes grids where highaspect ratios occur near walls. The most important results of this analysis
can be summarized as follows:
The LU-SGS + BE scheme is always stable for any choice of the convective velocity direction and the solution orientation for second- to
fourth-order SV and SD methods.
Five/six symmetric Gauss-Seidel (SGS) sweeps are in general sufficient to get a good damping of high-frequency error components.

These results have been used to simulate some 2D steady compressible

flows with the 2D-SV-TRI in-house code. 2D-SV-TRI is a 2D SV-based code
which was developed at the Vrije Universiteit Brussel, Department of Mechanical Engineering. In combination with a full multigrid (FMG) V-cycle
p-multigrid algorithm, the LU-SGS + BE scheme was used to compute the
2D steady laminar flow past a circular cylinder at Re = 40 and different
Mach numbers and the flow over a NACA0012 airfoil at Re = 5 103 and
M = 0.5. A steady laminar flow in a channel with a backward-facing
step at Re = 800 and Minlet = 0.2 was also considered. In order to assess
the convergence properties of the LU-SGS + BE scheme, the performance
of the latter scheme has been compared with that of a family of explicit
Runge-Kutta (E-RK) smoothers available in the 2D-SV-TRI code. The most
important results of this study can be summarized as follows:
The convergence rate of the LU-SGS + BE scheme is not greatly affected by the cell aspect ratio of the mesh.
The convergence rate of the LU-SGS + BE scheme is slightly affected
by the low Mach number when meshes with isotropic cells are used.
Effects of the low Mach number on the convergence rate of the LUSGS + BE scheme are visible when meshes with anisotropic cells and
high-aspect ratio are used.
170

9.1. ACHIEVEMENTS
For second-order SV method, the LU-SGS + BE scheme reduces the
computational time by a factor of up to 5 100 (depending on the
aspect ratio of the mesh) as compared to a well tuned E-RK scheme.
For third- and fourth-order SV methods, the LU-SGS + BE scheme reduces the computational time by 2 orders of magnitude as compared
to a well tuned E-RK scheme.
The convergence rate of the LU-SGS + BE for the NACA0012 airfoil simulation is comparable with that of a very efficient reference
implicit scheme proposed by Swanson, Turkel and Rossow [169].

The coupling between the SV method and the non-linear LU-SGS algorithm has been limited to 2D steady flow problems because no stable highorder 3D SV partition has been found to date.
The non-linear LU-SGS solver has also been implemented in a C++ code,
namely the COOLFluiD collaborative simulation environment developed
at the Von Karmann Institute for Fluid Dynamics. The algorithm has been
combined with the BE scheme and the BDF2, for a SD method. The LUSGS + BE scheme has been tested by solving 2D steady laminar flow over a
NACA0012 airfoil at Re = 5103 and M = 0.5 and the 3D steady laminar
flow through a 90 bending square duct. In order to assess the convergence
properties of the LU-SGS + BE scheme, a Newton-Raphson GMRES algebraic solver, has been used as reference efficient algebraic solver. A study
of the methods performances has shown that:
In term of wall time LU-SGS + BE scheme for high-order SD discretizations performs better than the Newton-Raphson GMRES solver
with the BE scheme.
The non-linear LU-SGS algorithm needs far less memory than the
Newton-Raphson GMRES method, however its application to 3D flow
problems still requires a large amount of memory, when high-order
polynomial reconstructions are used.

To demonstrate the capabilities of the SD method in combination with the

non-linear LU-SGS solver for the solution of unsteady flow problems, the
following 2D test cases have been investigated: flow over an open cavity at
Re = 1.5103 and Minlet = 0.15, flow past a square cylinder at Re = 200 and
M = 0.5, and flow past a circular cylinder at Re = 75, 150, 300, 800, 103 .
171

CHAPTER 9. CONCLUSIONS AND FUTURE DIRECTIONS

Time marching was done with the BDF2. The main conclusions of this
study can be summarized as follows:
For a given coarse grid the accuracy of the flow solution improves by
increasing the order of the SD scheme.
For each test case, the solution computed with highest order polynomial interpolation agrees well with experimental data and/or reference numerical solutions available in literature.
Very good results can be achieved using a number of degrees of freedom which is significantly lower than that of numerical solutions
computed with classical spatial discretizations, e.g. finite volume
(FV) or finite difference (FD).

An important achievement of the present work is the coupling of a highorder SD scheme with the WALE model (SD-LES) to perform large eddy
simulations. The accuracy and the reliability of the SD-LES approach have
been tested by solving the 2D flow over a NACA0012 airfoil at Re = 5 105,
M = 0.4 and zero degree angle of attack, the 2D turbulent flow past a
square cylinder at Re = 104 and Re = 2.2 104 and M = 0.05, and the
3D turbulent flow in a muffler at Re = 4.64 104 and Minlet = 0.05. Simulations without subgrid-scale model were also performed for the 2D test
cases. Time marching was done with BDF2 in combination with the nonlinear LU-SGS solver. The main results of this analysis can be summarized
as follows:
The subgrid-scale model does not affect the numerical solution when
the laminar flow is well resolved, i.e. the SD-LES computation replicates the laminar flow obtained by solving the pure Navier-Stokes
equations.
No specific difficulties of convergence have been observed for low
Mach number flows.
During the initial time steps the number of inner LU-SGS sweeps
oscillates between forty and fifty; afterwards, it decreases to an averaged value of about ten.
Although the SD method without subgrid-scale model is already able
to predict all the features of the flows, the SD-LES approach is more
accurate.
172

9.2. FUTURE WORK

An estimation of the SD-LES computational cost has shown that the
new approach is about 2.5 times faster than a DNS performed with
the same SD scheme.
The SD-LES results agree well with numerical or experimental data
reported in literature.

9.2 Future work

With this work, the initial development of an efficient Navier-Stokes/LES
solver on unstructured grids for high-order accurate schemes was established. The capabilities and the advantages of the current available tool
have been clearly illustrated by its application to different types of fluid
flow problems. However, more research and development are needed in order to make this solver - and in general high-order solvers - a valuable tool
for the industry. A personal vision and an overview of the main tasks and
issues that need to be solve for the application of any high-order methods
for unstructured grids to practical applications is given below.

9.2.1 Compact high-order accurate spatial methods

The stability problems that occur with high-order SV schemes for tetrahedral cells, is the most critical issue that remains to be resolved for the usability of the SV method. Several new SV partitions were proposed which
have a reduced maximum real part of the Fourier footprint by up to 20%
over the original un-optimized partition reported in literature. Numerical
simulations showed that the strength of the instability was weakened by
about an order of magnitude for some cases by employing a constrained
minimization approach. However no fully stable third-order SV scheme
for tetrahedral cells has been found to date. Consequently, high-order SV
schemes for tetrahedral cells are not yet robust. Therefore, they can not be
routinely employed in practical calculations with complicated geometries
and complex physics.
Although the SD schemes for quadrilateral and hexahedral cells are very
promising, a crucial issue for the application of the SD method to real complex geometries and industrial flow problems is the solution of the stability
problems with high-order SD schemes for triangular and tetrahedral cells.
In fact, only if this issue will be solved, high-order SD schemes could be
used in modern applications where unstructured hybrid grids are required
173

CHAPTER 9. CONCLUSIONS AND FUTURE DIRECTIONS

for efficient automated grid generations of complex geometries. Nevertheless, the flexibility of the SD method could already be improved by extending the schemes formulation to more general quadrilateral and hexahedral
cells, where hanging nodes are allowed. Although several algorithms have
been developed for the generation of pure hexahedral meshes for complex
geometries, see for instance Yasushi et al. [197], hanging nodes are in fact
absolutely necessary for an efficient automated grid generation for general
geometries, when hybrid grids can not be used.
Recently, Huynh [79, 80] developed a new high-order formulation for 1D
conservation laws based on the idea of flux reconstruction. The formulation was capable of unifying several high-order methods including the DG
method and the SV and SD methods into a single family. The idea of flux
reconstruction was generalized into a lifting collocation penalty (LCP) approach on hybrid grids by Wang and Gao [52, 184] and Haga et al. [61], respectively for 2D Euler and 3D Navier-Stokes equations. The LCP method
has a nodal formulation, with a cell-wise discontinuous polynomial solution space. The solution is interpolated from the solution at a set of solution points. In addition, a new flux polynomial is reconstructed. The LCP
method does not require the evaluation of any integrals. With a careful
selection of the solution and flux points, the approach can be made simple
and efficient to implement for hybrid grids. For these reasons the LCP
approach is a very promising method, which could be used as a valid alternative to the DG method and the SV and SD methods.
Although LES is by now a well researched approach for spatially low-order
accurate methods, specific studies on the capabilities of their high-order
counterparts to resolve accurately the small scales with an unresolved
DNS type approach or a LES approach are necessary. In this context, it
would be very interesting to investigate in more detail the computational
cost of SD schemes to perform DNS, unresolved DNS and LES. In order
to achieve that, an analysis of the number of Fourier modes in each direction for adequate resolution of isotropic turbulence could be performed,
as shown by Pope [132]. These studies should be carried out considering
the dispersion and the dissipative properties of high-order schemes. These
results are expected to give a better understanding of the capabilities of
spatially high-order schemes for the simulation of turbulent flows. In addition, the outcome of this analysis will also be useful in mesh generation
with high-order schemes. Further investigation of the subgrid scale models is also necessary.

174

9.2. FUTURE WORK

Recently, Zhou and Wang [201] applied a SD method in combination with a
perfectly matched layer (PML) to the simulation of CAA benchmark problems, while the author of this thesis has exploited the strength of a highorder spectral difference method to provide the acoustic sources for aerodynamic sound field simulations with a Ffowcs-Williams Hawkings (FW-H)
approach. However, with the aim of applying high-order methods to computational aeroacoustics (CAA), further investigation of the capabilities of
these schemes to simulate CAA problems or to provide acoustic sources for
aerodynamic sound field simulation is also very important.

9.2.2 Time integration/solution iterative approaches

The spatial discretization of nonlinear hyperbolic PDE to high order of
accuracy on unstructured meshes can be accomplished by local discretization methods, such as the DG method, the SV and the SD methods or
the new promising LCP approach. While a high-order spatial discretization enables one to achieve better resolution with less degrees of freedom
as compared to current state-of-the-art second-order methods, the overall
efficiency and robustness of a numerical scheme for large-scale applications depends on the solution methodology for the (non-linear) system of
equations arising from the discretization. High-order spatial discretization operators are usually much stiffer than their low-order counterparts.
Classical explicit time marching algorithms, such as explicit Runge-Kutta
schemes, have an upper limit for the time step that can be taken for stability reasons. Such classical algorithms can be very inefficient in combination with high-order spatial schemes, where the maximum time step tends
to be very small. Consequently, high-order spatial discretizations necessitate the use of an implicit method for robust simulations at a reasonable
cost. Several implicit time integration/solution iterative approaches have
been used in literature, e.g. the preconditioned Newton-Raphson GMRES
solver, the non-linear LU-SGS algorithm, the line-implicit algorithm, to
list a few.
In this thesis, the non-linear LU-SGS algorithm, in combination with two
implicit time discretization schemes, has been investigated extensively.
This algebraic solver proved to be very efficient and robust. However, additional studies can be performed, for instance applying the Fourier analysis
to the discretized form of the linearized time-dependent Euler equations
when solved with the LU-SGS + BE scheme. The outcome of this analysis is expected to give additional information on the properties of the time
marching/iterative technique. Moreover, specific studies on the capabil175

CHAPTER 9. CONCLUSIONS AND FUTURE DIRECTIONS

ities of the implicit solver to deal with high Reynolds number and very
large aspect ratio are necessary.
Although non-linear LU-SGS algorithm requires significantly less memory
than the classical methods that use the full Jacobian matrix (for instance
the Newton-Raphson GMRES algorithm), the required amount still increases with order of accuracy to the power six, for 3D computations. This
large memory requirement can be a serious limitation for the applicability
of high-order methods to very large 3D problems. The key issue is then to
develop an algebraic solver that requires a small amount of memory and
is able to solve efficiently the algebraic non-linear systems. Good candidates may be matrix-free Newton-Krylov methods and the more recently
developed RK3/Implicit Residual scheme which was proposed by Swanson, Turkel and Rossow [169]. With the latter approach, the preconditioner would be based on a first-order upwind scheme instead of high-order
spatial discretization operators. This way the complete Jacobian with no
simplifications could be easily calculated on the fly and requires no storage. In addition, these small calculations could be accelerated by using a
graphics processing unit (GPU). For these reasons, the RK3/Implicit Residual scheme is a very promising method, which is certainly worth further
investigation in combination with spatially high-order compact schemes.
Geometric and p-multigrid methods are also very promising acceleration
approaches and can be used to exploit the strength of the algebraic solvers.
These two techniques are essential to perform both steady and unsteady
simulations at a reasonable cost for industrial flow applications.
A last topic of future interesting research coming to mind is the aspect
of implicit high-order accurate temporal methods. Since time accuracy is
also an important aspect in unsteady numerical simulations, it would be
very interesting to investigate the performances of these implicit schemes
in combination with compact high-order accurate spatial methods. Successful application of them to the simulation of compressible turbulence,
vortex dominated flows and aeroacoustics is expected to have an important
impact on the usability of high-order solvers for unstructured grids. The
fourth-order explicit-first-stage singly diagonally implicit Runge-Kutta (ESDIRK4) scheme or the fourth-order modified extended backward differentiation formula (MEBDF4) may be good starting candidates. It is very
important to point out that although research in this direction is needed
to achieve the full potential of high-order schemes, it already appears that
algebraic solver technology is again the area that needs significant attention.
176

Appendix A

Time integration methods

for space-discretized
equations
In this appendix, a method to analyze the linear stability for the combination of a space and a time discretization will be discussed. It is based
on the method of lines (MOL) introduced in the early 1960s by Sarmin and
Chudov [155]. The MOL most often refers to the construction or analysis of
numerical methods for partial differential equations that proceeds by first
discretizing the spatial derivatives only and leaving the time variable continuous. This leads to a system of ordinary differential equations to which
a numerical method for initial value ordinary equations can be applied.
This way, the analysis of the space discretization and time integration can
be performed separately with a Von Neumann type of approach. This procedure allows to answer to the fundamental question on the criteria to be
satisfied by a time integration method, applied to a given space discretization, so that it leads to a stable fully discretization. It can be applied to any
system of linear PDEs. In the present appendix, this procedure is applied
to 2D linear advection equation.
The methodology to analyze the stability properties and the amplification
matrix of the non-linear LU-SGS algorithm in combination with the backward Euler difference is also discussed. The method corresponds to an
application of the Von Neumann analysis and uses the data obtained from
the stability analysis of the spatial discretization.
177

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

More information on the numerical solution of partial differential equations, including other methods for the stability analysis of spatial and temporal schemes can be found in Isaacson and Keller [83], Smith [160] and
Hirsch [75].

A.1 Stability of spatial discretizations

Consider the linear advection equation, discussed in Section 3.5, and given
by
w ~
+ (~aw) = 0.
(A.1)
t
The initial solution (t = 0) is a spatial Fourier wave defined by
W (~
, 0) =

W0 (~
) = A0 eI k~

(A.2)

where
~ is the position vector, I 1 is the imaginary unit number and
~k is the wave vector given by


~k = k cos = k ~1k .
(A.3)
sin
Equation (A.1) is discretized in space by introducing the spatial derivatives
corresponding to a spatial discretization method. A uniform grid with periodic boundary conditions is considered. The grid is defined by a generating
pattern, which is the smallest part from which the full grid can be reconstructed by periodically repeating the pattern in all directions. For the
2D case, the generating patterns for uniform triangular and quadrilateral
meshes are shown in Figures A.1(a) and A.1(b). The generating pattern
~ 1 and B
~ 2 and their non-dimensional
is completely defined by the vectors B
~ 1 , denoted by B:
form is obtained by scaling them with the length of B

~
~
~
~
~ 1 is chosen as
B1 B B1 and B2 B B2 . If the dimensionless vector B
T
[1 0] , then the dimensionless mesh is completely defined by the two com~.
ponents of B
2
In 2D, the advection speed ~a in Equation (A.1) is defined by its amplitude
a and orientation angle :


cos
~a = a
.
(A.4)
sin
178

A.1. STABILITY OF SPATIAL DISCRETIZATIONS

(a) Triangular mesh.

(b) Quadrilateral mesh.

Figure A.1: Generating patterns for 2D uniform meshes.

At the cell faces, which lie on a grid face between two cells, two values
for the convected variables are available, from the solution corresponding
to the two neighboring cells. In order to ensure numerical conservation,
the contributions of a face to its two neighboring cells should be equal in
magnitude and opposite in sign. Thus, a unique flux should be computed
from the two available solutions. In the present analysis, the following
approximate Riemann flux is used:

 W W
WL + WR

 R
L
~a ~1n 
,
F~ AR (WL , WR ) ~1n = ~a ~1n
2
2

(A.5)

where ~1n is the unit normal oriented from the left to the right side and
indices L and R indicate respectively the left and right neighboring cell to
a face. is an upwinding parameter, with = 1 resulting in an upwind
flux and = 0 in a central flux.
After application of the space method to (A.1) on a uniform quadrilateral
or triangular mesh, the following system of ODEs is obtained:
dWi,j
a  0,0
+
T Wi,j + T1,0 Wi1,j + T0,1 Wi,j1
dt
B


+T+1,0 Wi+1,j + T0,+1 Wi,j+1 = 0,

(A.6)

where the five dimensionless matrices T are defined by the spatial discretization method. They depend on the generating pattern, the advection
speed orientation angle and the upwinding parameter . The column
vector Wi,j contains the N s,GP solution variables in the generating pattern
with indices i and j (from within one quadrilateral cell or two triangular
cells).
179

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

We seek for a spatial Fourier wave solution of the form

Wi,j (t)

(t) eI~k(iB~ 1 +j B~ 2 )B
= W
(t) eIK~1k (iB~ 1 +j B~ 2 ) ,
= W

(A.7)

where K is the dimensionless wave number, defined as kB. Inserting

expression (A.7) in Equation (A.6) results in

dW
a
= T W,

=
TW
dt
B

T =

a
T
B

with the matrix T defined by


~ ~
~ ~
T = T0,0 + T1,0 eIK 1k B1 + T0,1 eIK 1k B2

~ ~
~ ~
+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 .

(A.8)

(A.9)

The stability analysis of the space discretization is based on the eigenvalues structure of the matrix T , since the exact solution of the system
(A.8) is directly determined by the eigenvalues and eigenvectors of T . Let
m = m,R + Im,I , m = 1, . . . N s,GP be the complex eigenvalues of the
matrix T solution of the eigenvalue equation


det T I = 0
(A.10)

and Vm the associated eigenvectors. Expression (A.10) defines the numerical dispersion relation of the spatial scheme, i.e. the relation between
eigenvalues , the dimensionless wave number K, the wave orientation
angle and the the convection speed orientation angle . This relation
should be compared with the exact dispersion relation ex,R = 0, ex,I =
K cos ( ) to assess the capability of the spatial scheme to model wave
propagation [75, 76, 176].
Since, the eigenvectors Vm form a complete set of basis vectors in the considered space of solution variables, the exact solution of (A.8) can always
be written as a linear combination of these eigenvectors:
(t) =
W

s,GP
NX

m
W
(t) Vm ,

m=1

180

(A.11)

A.1. STABILITY OF SPATIAL DISCRETIZATIONS

m
where the W
coefficients depend only on time. These coefficients are obtained from the differential system (A.8), by inserting (A.11), leading to
N s,GP homogeneous modal equations:

dW
m
,
= a m W
= m W
m
m
dt
B

m = 1, ..., N s,GP ,

(A.12)

s,GP
are the eigenvalues of the matrix T
where m , = m B
a , m = 1, . . . N
defined by Equation (A.9). The solution of these ODEs is of the form

(t) = em t = e(a/B)m t ,
W
m

m = 1, ..., N s,GP .

(A.13)

Therefore, expression (A.11) can be written as linear combination of the

N s,GP eigenmode solutions em t Vm :

(t) =
W

s,GP
NX

0 em t V ,
W
m
m

(A.14)

m=1

0
m
where the coefficients W
ensure that the initial condition (A.2) is satisfied:
s,GP
NX
0
0
m
W
Vm .
(A.15)
W (~
) =
m=1

It can be easily seen from expression (A.14) that, the ODE system is wellposed, or stable, if the exact solution remains bounded. This implies that
all the modal components are also bounded, since if any of them would
grow exponentially with time, the full solution would also have this unwanted behavior. Hence, for stability, the real part of the eigenvalues m
must be negative or zero. In addition, if an eigenvalue is zero, it has to be a
simple eigenvalue. Therefore, the left-half of the complex plane, including
the imaginary axis, is the region of stability for the exact solution of the
semi-discretized system of equations.
The collection of all eigenvalues m for all wave numbers K is often called
Fourier footprint. The Fourier footprint of a spatial scheme gives an indication of the stiffness of the problem1 and grows very fast with increasing
the order of the method. For instance, for the linear advection equation the
growing is faster than proportionally with the order of accuracy of the spatial method, as illustrated in Figures A.2 and A.3 for second-, third- and
1 An ordinary differential equation problem is stiff if the solution being sought is varying
slowly, but there are nearby solutions that vary rapidly, so the numerical method must take
small steps to obtain satisfactory results. Stiffness is an efficiency issue. If we werent concerned with how much time a computation takes, we wouldnt be concerned about stiffness.

181

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

15

15

10

10

10

10

15
35

Im(Eigenvalue)

15

Im(Eigenvalue)

Im(Eigenvalue)

fourth-order accurate spectral volume and spectral difference schemes for

triangular and quadrilateral cells respectively.

10

30

25

20
15
Re(Eigenvalue)

10

10

15
35

(a) Second-order (p = 1).

30

25

20
15
Re(Eigenvalue)

10

15
35

(b) Third-order (p = 2).

30

25

20
15
Re(Eigenvalue)

10

(c) Fourth-order (p = 3).

15

15

10

10

10

0
5
10
15
20

Im(Eigenvalue)

15

Im(Eigenvalue)

Im(Eigenvalue)

Figure A.2: Fourier footprints of second- to fourth-order SV schemes for triangular

cells for the linear advection equation.

0
5
10

18

16

14

12
10
8
Re(Eigenvalue)

10

15
20

(a) Second-order (p = 1).

0
5

18

16

14

12
10
8
Re(Eigenvalue)

(b) Third-order (p = 2).

15
20

18

16

14

12
10
8
Re(Eigenvalue)

(c) Fourth-order (p = 3).

Figure A.3: Fourier footprints of second- to fourth-order SD schemes for quadrilateral cells for the linear advection equation.

A.2 Stability of time discretizations

In the previous section it has been shown that, the real part of the eigenvalues m of the Fourier footprint must be negative or zero, for stability.
Only if this condition is satisfied the exact solution of the semi-discretized
system of equations remains bounded. However, it is quite possible for the
numerical solution of an ODE to grow unbounded even though its exact
solution is well behaved. Of course, there are cases for which the exact
solution is well grows unbounded, but for the discussion of the stability,
only cases in which the exact solution is bounded are considered. Consider
a differential equation of the form


dy
dt

= f (y, t) , t > 0
y(0) = y0 .
182

A.2. STABILITY OF TIME DISCRETIZATIONS

The aim of all numerical methods for solution of this differential equation
is to obtain the solution at time tn+1 = tn + t, given the
solution for
0 t tn . This process continues; i.e., once Y n+1 = Y tn+1 is obtained,
then Y n+2 is calculated and so on until the final time, tend . The stability
analysis of the numerical method, is performed in terms of the parameters
of the numerical method (mainly the step size t) for which the numerical solution remains bounded. In this context we have three classes of
numerical methods:
Unconditionally stable: numerical solution does not grow unbounded
(blow up) with any choice of parameters such as the time step. The
cost of such robustness must be investigated.
Unstable scheme: numerical solution grows unbounded with any choice
of parameters. Such numerical schemes are not useful.
Conditionally stable scheme: with certain choices of the parameters
the numerical solution remains bounded. Hopefully, the cost of the
calculation does not become prohibitively large.
For convenience and feasibility of analytical treatment, stability analysis
of time discretization schemes is usually performed on the model problem,
consisting of
 dy
dt = y(t), t > 0
(A.16)
y(0) = y0 .

The model equation has an analytical solution y (t) = et . Here, is a

complex constant, i.e = R + II , with the real part R < 0 to ensure that the solution does not grow with t. This generalization allows to
readily apply the results of the next analysis to system of ODEs. In fact, in
Section A.1, it has been shown that systems of ODEs, obtained for instance
from spatial discretizations of PDEs, can reduce to uncoupled ODEs of the
form (A.16) with complex coefficients. These coefficients correspond to the
eigenvalues m of the matrix T, which is defined by Equation (A.9).

A.2.1 Forward Euler scheme

The forward Euler scheme (FE) uses the first-order forward difference formula to approximate the time derivatives:
Y n+1 = Y n + t f (Y n , tn ) .

(A.17)

Applying this scheme to the model problem (A.16) leads to

Y n+1 = Y n + t Y n = (1 + t) Y n .
183

(A.18)

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

For complex , we have
Y n+1 = (1 + R t + II t) Y n = G Y n ,

(A.19)

where G = (1 + R t + II t) is called amplification factor. The numerical solution is stable (i.e., remains bounded as time becomes large) if
|G| 1.

(A.20)

For the system (A.8), the stability condition is then given by

|G| = max |1 + m | 1,

(A.21)

where m are the complex eigenvalues of the matrix T and =

Courant-Friedrichs-Lewy (CFL) number.

at
B

is

The general form of the amplification factor of the FE scheme can be written as
G (z ) = 1 + z .
(A.22)
The region in the complex plane defined by |G (z )| 1 is called stability
5
4
3
2

1.8
1.4
1.2
0.8
1
0.6

1.6
8
1.

2
1.
6

0.6
0.8

1.6

1.2
1.4

1.
6
1.8

0.4

2
0.

1.2
1.4

Im(z)

1.8

3
4
5
5

1
Re(z)

Figure A.4: Amplification factor module for the forward Euler scheme.

region. For the FE scheme, it is a circle with a radius of one in the complex
184

A.2. STABILITY OF TIME DISCRETIZATIONS

plane, centered around 1, as shown in Figure A.4. The FE scheme is
then a conditionally stable scheme. The stability condition for the complete
discretization is that the Fourier footprint of the spatial scheme scaled
with the CFL number, i.e. m for m = 1, ..., N s,GP , lies entirely inside the
stability region of the FE scheme. Comparing the Fourier footprints shown
in Figures A.2 and A.3 with the stability region of the FE scheme, one can
conclude that, combinations of high-order accurate spatial discretizations
with FE method lead to stable complete discretizations only when a very
small CFL number is used.

A.2.2 Backward Euler scheme

The backward Euler (BE) scheme is given by the following formula:


(A.23)
Y n+1 = Y n + t f Y n+1 , tn+1 .

Note that in contrast with the FE scheme, the BE scheme does not allow
to easily obtain the solution at the next time step. If f is non-linear, then
a non-linear algebraic equation must be solved at each time step to obtain
Y n+1 . Therefore, the computational cost per time step for this scheme is,
apparently, much higher than that for the FE scheme. An efficient algebraic solver is then needed to fully fulfill the potential of the Euler method,
which, as it will be shown further on, has better stability properties.
Applying the BE scheme to the model problem (A.16), one obtains
Y n+1 = Y n + t Y n+1 .

(A.24)

Solving for Y n+1 produces

Y n+1 =

1
Yn
1 t

(A.25)

or
Y n+1 = G Y n ,

(A.26)

1
.
1 t

(A.27)

where
G=

Considering complex , the following expression for the amplification factor G is obtained
1
.
(A.28)
G=
(1 R t) II t
185

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

The denominator is a complex number and can be written as
G=

1
,
AeI

(A.29)

where
A=

q
2
(1 R ) + 2I t2 ,

= tan1

I t
.
1 R t

(A.30)

For stability, the modulus of G must be less than or equal to 1; i.e.,

|G| =

eI
1
=
1 R 0,
A
A

where
= tan1

GI
.
GR

(A.31)

(A.32)

The amplification factor of the BE scheme can be written in the following

general form:
1
G (z ) =
.
(A.33)
1z
The stability region of the BE scheme is then the entire complex plane
minus the region enclosed by a circle with a radius of one and centered
around +1, as shown in Figure A.5.
Combining the spatially stability condition derived in Section A.1 (i.e. m,R
0 for m = 1, ..., N s,GP ) for the 2D linear advection equation with the stability condition of the BE scheme, one can conclude that there is no restriction on the CFL number for the stability of the complete space and
time discretization. In addition, the numerical solution Y n obtained with
the BE scheme satisfies the condition |Y n | 0, for tn , i.e. the BE
scheme is A-stable [135]. The A-stability property is important to solve the
systems of stiff ODEs arising from the discretization of the fluid dynamic
equations with a spatially high-order numerical scheme. In fact, A-stable
methods do not exhibit instability problems. Notice that |G (z )| tends also
to zero for large values of z . Thus the BE scheme is also L-stable [135].

A.2.3 Second-order backward difference formula

The backward difference formulae (BDF) represent a class of linear multistep methods which was discovered in 1952, together with the phenomenon
186

A.2. STABILITY OF TIME DISCRETIZATIONS

5

0.2
0.2

0.2

3
0.4

2
0.6

0.8

0.2

0.
6

0.6

0.4

0.8

0.4

Im(z)

1.5

0.4

2
0

0.4

0.
2

4
5
5

0.2

0.2

0
1
Re(z)

Figure A.5: Amplification factor module for the backward Euler scheme.

of stiffness, by Curtiss & Hirschfelder [45] in calculations for chemistry,

and their extreme importance for stiff problems has been recognized since
the work of Gear in 1971 [53]. With the BDF, the time derivative dy
dt at
time tn+1 is approximated with the derivative of the k + 1 degree polynomial, which is constructed by interpolating the solution at k + 2 time
points tn , tn1 , . . . tnk . This procedure leads to schemes of the following
form [135]:
Y n+1 =

k
X
j=0

aj Y nj + t b1 f Y n+1 , tn+1 ,

(A.34)

with b1 6= 0. For k = 0 (first-order interpolation polynomial), assuming

a0 = 1 and b1 = 1, one obtains the BE scheme (A.23). Thus, the BE
scheme can be considered as a first-order backward difference formula.
The second-order (k = 1) backward difference scheme (BDF2) with constant time step is defined by a0 = 34 , a1 = 31 and b1 = 32 , i.e.
Y n+1 =

4 n 1 n1 2
Y Y
+ t f Y n+1 , tn+1 .
3
3
3
187

(A.35)

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

Applying this scheme to the model problem (A.16), one obtains
1
2
4
(A.36)
Y n+1 = Y n Y n1 + t Y n+1 .
3
3
3
The amplification factor of the BDF2 is then the solution of the following
equation
(3 2z ) G2 4G + 1 = 0,
(A.37)

which has been obtained by substituting the general expression G = Y n+1 /Y n

in (A.36). Equation (A.37) has two roots:

2 1 + 2z
.
(A.38)
G =
3 2z
The amplitude of G is always smaller than one, while |G+ (z )| is smaller
than one if zR 0. The module of G+ (z ) is shown in Figure A.6. The BDF2
is both A-stable and L-stable.
5
0.6
0.4

0.
6

3
0.8

0.8

2
1

0.6

1
0.8

3
0.4

4
5
5

0.8

0.8

4
0.

1.5

1.5

Im(z)

5
1.

0.6

0.6

0
1
Re(z)

Figure A.6: Amplification factor module for the second-order backward difference
formula.

A.2.4 Higher-order backward difference formulae

Following the procedure described at the beginning of Section A.2.3, higherorder BDF (order of accuracy higher than two) can be constructed. However, implicit multi-step methods can only be A-stable if their order is at
188

A.2. STABILITY OF TIME DISCRETIZATIONS

most two. This result is known as the second Dahlquist barrier [135].
In Table A.1 the coefficients of all high-order zero-stable BDF are listed.
A scheme for which perturbations remain bounded in the limit t 0
is said to be zero-stable [135]. The high-order BDF considered here are
also L-stable. The limit of the stability zone (|G (z)| = 1) of the BDF for
Table A.1: Coefficients of the zero-stable BDF.

k
2

a0

48
11

300
137

360
147

18
11

a1
9
11

36
25

300
137
450
147

a2
2
11
16
25
200
137
400
147

a3
0

a4
0

a5
0

b1

3
25

12
25

12
137

60
137

72
147

10
147

60
147

75
137

225
147

6
11

k = 2, 3, 4, 5 are shown in Figure A.7. It can be easily seen that these

schemes are not A-stable (second Dahlquist barrier). In fact, the curves
for which |G (z)| = 1 cross the imaginary axis. In addition, the region of
instability in the left-half complex plane increases by increasing the order
of accuracy of the BDF, as illustrated in Figure A.7(b).
25

BDF3
BDF4
BDF5
BDF6

20
15

BDF3
BDF4
BDF5
BDF6

10
1
Im(z)

Im(z)

5
0

5
1

10
15

20
25

10

10
Re(z)

20

30

(a) Limit of the stability zones.

0.5

0
Re(z)

0.5

(b) Zoom near the origin of the axes.

Figure A.7: Limit of the stability regions of high-order zero-stable backward difference formulae.

As shown in Figures A.2 and A.3, both high-order SV and SD schemes

have a wide range of eigenvalues in the vertical direction and very close to
189

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

the imaginary axis. This might lead to an unwanted reduction of the maximum CFL value when associated with high-order accurate BDF2 . In fact,
for the complete discretization with a general spatial scheme and a general time marching scheme, the stability condition is then that the Fourier
footprint of the spatial scheme, scaled with the CFL-number, lies entirely
inside the stability zone of the time marching scheme.
Notice that the amplification factor G (z) can be defined for a general time
marching scheme and the stability zone, defined by |G (z)| 1, can always be determined. Therefore, it is always possible to find a value of
the time step t and/or the CFL number which leads to a stable complete
discretization. However, for spatially high-order accurate schemes, this
value might be very small because such spatial discretizations generate
very often systems of stiff ODEs. This is very common in computational
fluid dynamics where the time integration scheme must also be able to
deal with the geometrical stiffness imposed by the Navier-Stokes grids. In
the case of compressible solvers there is an additional stiffness when solving for low speed flows caused by the disparate eigenvalues of the system.
The A-stability property is then very important to solve the systems of stiff
ODEs.

A.3 Method for the analysis of the non-linear

LU-SGS algorithm
The aim of this section is to present a methodology to analyze the smoothing properties of the non-linear LU-SGS algorithm when combined with
the backward Euler scheme. This scheme, denoted as LU-SGS + BE, was
presented in Section 5.1 and is used in Chapters 7 and 8 as a smoother to
solve steady flow problems. Its analysis is then fundamental because essential for efficiency is that the solver is a good smoother of high-frequency
error components.
The damping properties of the LU-SGS + BE are evaluated with a Von
Neumann stability analysis for a model 2D linear advection. This model
problem was already discussed in Section 3.5. The starting point of the
analysis is the 2D semi-discretized linear advection equation (A.8). Since
the direct inversion method performs better than any approximate method,
the procedure to derive its amplification matrix is also illustrated.
2 For

accuracy, a smaller value for t may be required.

190

A.3. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

A.3.1 Direct inversion method

Consider the 2D semi-discretized linear advection equation (A.8) obtained
by substituting the spatial Fourier wave (A.7) in Equation (A.6). For convenience, this equation is repeated here:

dW
a  0,0
~ ~
~ ~
T + T1,0 eIK 1k B1 + T0,1 eIK 1k B2
+
dt
B

~ ~
~ ~
= 0,
(A.39)
+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 W

where a and B are the module of the advection speed vector and the reference length scale respectively. In (A.39), the five dimensionless matrices
are defined by the spatial discretization method and correspond to the discretized spatial derivatives. They depend on the generating pattern, the
advection speed orientation angle and the upwinding parameter of the
approximate Riemann flux (A.5).
Approximating the time derivative with the BE scheme, discussed in Section A.2.2 and defined by (A.23), Equation (A.40) is obtained,

n+1 W
n + T0,0 + T1,0 eIK~1k B~ 1 + T0,1 eIK~1k B~ 2
W
 n+1
~ ~
~ ~

= 0, (A.40)
+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 W
t
is the CFL number. The linear system (A.40) may also be
where = aB
written in a more compact form, i.e.

n+1 = W
n,
EW

(A.41)

where the matrix E is obtained by assembling the matrices T0,0 , T1,0 ,

T0,1 , T+1,0 and T0,+1 . An expression for the amplification matrix Gd ,
n+1 = Gd W
n , can be obtained from system (A.40):
defined by W
h

~ ~
~ ~
Gd = I + T0,0 + T1,0 eIK 1k B1 + T0,1 eIK 1k B2
i1
~ ~
~ ~
.
(A.42)
+T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2

The matrix Gd represents the amplification matrix of the direct inversion

method which is marked as direct + BE in this work.

A.3.2 Non-linear LU-SGS algorithm

The analysis described for the direct inversion method may be applied to
the LU-SGS + BE scheme to get its amplification matrix. The LU-SGS +
191

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

BE method is used to solve a non-linear system of equations. However,
when it is employed to solve a linear problem, it is identical to the classical
LU-SGS algorithm proposed by Jameson and Turkel [88]. The main idea
of the linear LU-SGS method is to split the matrix E in Equation (A.41)
into a diagonal, a strictly lower and a strictly upper matrix
E=D+L+U

(A.43)

and solve the linear system with forward and backward Gauss-Seidel sweeps:
m+1/2 + U W
m =W
n,
(D + L) W
m+1

(D + U) W

m+1/2

+ LW

(A.44a)
n

,
=W

(A.44b)

where m = 0, 1, 2, 3, . . . is the actual SGS sweep index. The first equa m+1/2 , corresponds to the forward sweep and the section, with solution W
m+1 represents the backward sweep. Note that
ond one, with solution W
m=0
n

.
W
=W
Because of the Gauss-Seidel nature of the non-linear LU-SGS algorithm,
where the latest available solution in the neighboring cells is used to update the solution in a cell, the procedure to obtain the amplification matrix
for the LU-SGS depends on the generating pattern structure, through the
matrix T0,0 . In fact, the column vector Wi,j in Equation (A.6) contains
the N s,GP solution variables of the generating pattern which corresponds
to one quadrilateral cell or two triangular cells, as illustrated in Figures
A.1(b) and A.1(a). In the next sections, the procedure for both grids is
discussed.
Uniform triangular mesh
In case of uniform triangular grid, the solution variables of the generating
pattern are the solution variables of two triangular cells. Consequently,
the matrix T0,0 has to be split in the following form:
0,0
0,0
0,0
T0,0 = T0,0
1,1 + T1,2 + T2,1 + T2,2 ,

(A.45)

0,0
where T0,0
1,1 and T2,2 represent respectively the contribution to the residual
of the first and second cell of the generating pattern to themselves, while
0,0
T0,0
1,2 and T2,1 represent the cross contributions of both cells of the generating pattern. After substituting Equation (A.45) into Equation (A.40), the

192

A.3. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

equation of the amplification matrix for the first forward sweep becomes

1/2 W
n + T0,0 W
1/2 + T0,0 W
n + T0,0 W
1/2 + T0,0 W
1/2
W
1,1
1,2
2,1
2,2
~

1/2

1/2

+ T0,1 eIK 1k B2 W


n
n
~ ~
~ ~
,
+ T0,+1 eIK 1k B2 W
+T+1,0 eIK 1k B1 W

+T1,0 eIK 1k B1 W

(A.46)

from which the amplification matrix of the first forward sweep (A.47) is
found.
h

i1
~
~
0,0
0,0
0,1 IK~1k B
1,0 IK~1k B
1 + T
2
GTf,1R = I + T0,0
e
e
1,1 + T2,1 + T2,2 + T
i
h

~
~
0,+1 IK~1k B
+1,0 IK~1k B
2
1 + T
(A.47)
e
I T0,0
+
T
e
1,2
For the first backward sweep, in accordance with Equation (A.44b), one
obtains

1W
n + T0,0 W
1 + T0,0 W
1 + T0,0 W
1/2 + T0,0 W
1
W
1,1
1,2
2,1
2,2
~ ~ 1/2
~ ~ 1/2
+ T0,1 eIK 1k B2 W
+T1,0 eIK 1k B1 W

~ ~ 1
~ ~ 1
+ T0,+1 eIK 1k B2 W
+T+1,0 eIK 1k B1 W
.

(A.48)

Hence, the amplification matrix of the first SGS sweep is given by

h

i1
~
~
R
0,0
0,0
+1,0 IK~1k B
0,+1 IK~1k B
1 + T
2
GTSGS,1
= I + T0,0
e
e
1,1 + T1,2 + T2,2 + T
i
h

~
~
0,1 IK~1k B
1,0 IK~1k B
2
1 + T
.
(A.49)
e
I T0,0
e
2,1 + T
R
From the amplification matrix GTSGS,1
, the amplification matrix of m-th
SGS sweep may be computed using the following two-step recursive procedure:

1: Compute the amplification matrix of the m-th forward sweep

i1
h

~
~
R
0,0
0,0
1,0 IK~1k B
0,1 IK~1k B
1 + T
2
e
GTf,m
= I + T0,0
+
T
+
T
+
T
e
1,1
2,1
2,2

i
h

~
~
R
+1,0 IK~1k B
0,+1 IK~1k B
1 + T
2
GTSGS,m1
.
I T0,0
e
e
1,2 + T

(A.50)

193

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

2: Compute the amplification matrix of the m-th SGS sweep
i1
h

~
~
R
0,0
0,0
+1,0 IK~1k B
0,+1 IK~1k B
1 + T
2
GTSGS,m
= I + T0,0
e
e
1,1 + T1,2 + T2,2 + T

i
h

~
~
TR
1,0 IK~1k B
0,1 IK~1k B
1 + T
2
G
I T0,0
+
T
e
e
f,m .
2,1

(A.51)

Uniform quadrilateral mesh

In contrast with triangular meshes, the analysis of the amplification matrix for quadrilateral grids does not require the splitting of matrix T0,0 .
In fact, the generating pattern corresponds to only one quadrilateral cell.
Therefore, the equation for the first forward sweep becomes

1/2
1/2 + T0,1 eIK~1k B~ 2 W
1/2 W
n + T0,0 W
1/2 + T1,0 eIK~1k B~ 1 W
W

~ ~ n
~ ~ n
,
(A.52)
+ T0,+1 eIK 1k B2 W
+T+1,0 eIK 1k B1 W

from which the amplification matrix for the first forward sweep (A.53) is
found.
i1
h

~
~
0,1 IK~1k B
0,0
1,0 IK~1k B
2
1 + T
e
GQD
=
I
+

T
+
T
e
f,1
i
h

~ ~
~ ~
(A.53)
I T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2

For the first backward sweep, in accordance with Equation (A.44b), one
obtains

1/2
1/2 + T0,1 eIK~1k B~ 2 W
1W
n + T0,0 W
1 + T1,0 eIK~1k B~ 1 W
W

~ ~ 1
~ ~ 1
.
(A.54)
+ T0,+1 eIK 1k B2 W
+T+1,0 eIK 1k B1 W
Therefore, the amplification matrix of the first SGS sweep is given by
i1
h

~
~
0,0
+1,0 IK~1k B
0,+1 IK~1k B
1 + T
2
GQD
=
I
+

T
+
T
e
e
SGS,1
i
h

~ ~
~ ~
I T1,0 eIK 1k B1 + T0,1 eIK 1k B2 .

(A.55)

From the amplification matrix GQD

SGS,1 , the amplification matrix of m-th
SGS sweep may be computed using again the following two-step recursive
procedure:
194

A.3. ANALYSIS OF THE NON-LINEAR LU-SGS ALGORITHM

1: Compute the amplification matrix of the m-th forward sweep

i1
h

~
~
0,1 IK~1k B
1,0 IK~1k B
0,0
2
1 + T
e
e
GQD
f,m = I + T1,1 + T
i

h

~ ~
~ ~
I +T+1,0 eIK 1k B1 + T0,+1 eIK 1k B2 GQD
SGS,m1 .

(A.56)

2: Compute the amplification matrix of the m-th SGS sweep

i1
h

~
~
0,+1 IK~1k B
0,0
+1,0 IK~1k B
2
1 + T
e
GQD
=
I
+

T
+
T
e
SGS,m
i

h

~ ~
~ ~
I T1,0 eIK 1k B1 + T0,1 eIK 1k B2 GQD
f,m .

(A.57)

A.3.3 Eigenvalue spectrum of the amplification matrix

It is well known in many matrix theory that the properties of a matrix are
fully contained in its eigenvalue and eigenvector spectrum, as shown in
Section A.1 for the system of semi-discretized PDEs. The analysis of a matrix, or an operator, through the eigenvalues and eigenvectors, represents
therefore a most profound investigation of their properties. Application of
this analysis to the amplification matrix can can be used to asses the stability and smoothing properties of the solver.
Let m (m = 1, 2, . . . , N s,GP ) be an eigenvalue of the general amplification
matrix G = G (K, , ) and = (G (K, , )) represent the eigenvalue
spectrum of G. Then g = g (K, , ) max | (G)| is the amplification factor for a given (K, , ). In order for the global discretization to be stable,
g 1 should be satisfied i.e. (G) lies inside the unit circle of the complex
plane (stability boundary) for all K, and . The range of K is one period
of the Equation (A.40) and it is marked PK in this work. For a fixed shape
of the generating pattern, PK is a function of , i.e. PK = PK (). The ex~,
pression from which PK can be computed, is obtained by substituting B
1
~ 2 and in the exponential terms of Equation (A.40) and using the Eulers
B
formula, which gives a relation between trigonometric functions and complex numbers. Following this reasoning, a trigonometric function f , which
has exactly the same period as Equation (A.40) is obtained. In general, it
is not possible to find a closed formulation for PK = PK (). Therefore, the
period PK is computed taking the inverse fast Fourier transformation of
the function f .
Equation (A.40) is periodic in K with a period PK . However, for each K,
195

APPENDIX A. TIME INTEGRATION METHODS FOR ODES

there are N s,GP /2 and N s,GP eigenmodes for triangular and the quadrilateral meshes respectively. Consequently, each eigenmode can be interpreted
as corresponding to a certain wave number K +l PK , with l an integer number. The actual wave number K + l PK to which an eigenvalue m corresponds should be determined by examining the accompanying eigenmode
m eIK~1k (iB~ 1 +j B~ 2 ) , where V
m are the eigenvectors of the amplificashape V
tion matrix.
If the scheme is used as a smoother for multigrid, then it must have good
damping of high-frequency error components, i.e. it should cluster the amplification matrix eigenvalues corresponding to the high-frequency modes
towards the origin. Therefore, for high-frequency modes g 1 for all
and should be satisfied. In addition, we desire that the CFL number be
sufficiently large to produce significant reduction (if not elimination) of the
convergence slow-down effects that are associated with high-aspect ratio
cells. A large CFL number also facilitates the expulsion of error components. At the same time the capability for large CFL numbers must not
compromise the high-frequency damping property of the scheme.

196

Appendix B

p-Multigrid
In this appendix a brief introduction to the p-multigrid algorithm is given.
The main idea of multigrid is based on the observation that error-smoothing
operators are generally efficient in eliminating high-frequency errors, but
less adequate for the low-frequency errors. The multigrid strategy is to
switch to a coarser representation of the solution, where the low-frequency
errors of the fine representation occur as high-frequency modes, which can
thus be efficiently damped out. In the traditional h-multigrid approach,
this is done by switching to a coarser spatial grid. With a p-multigrid algorithm, a high-order solution representation is transferred to a lower-order
one.
The p-multigrid is an iterative algorithm in which systems of equations
arising from compact, high-order space discretizations, such as spectral
volume and spectral difference formulations, are solved by recursively iterating on solution approximations of different polynomial order. For example, to solve equations derived using a polynomial approximation order
of p = 4, the solution can be iterated on at an approximation order of
p = 4, 3, 2, 1. The p component of this algorithm was proposed by Rnquist
and Patera [146] and analyzed by Maday and Munoz [107] for 1D, Galerkin
spectral element discretization of Laplace equation. Helenbrook [68] combined p-multigrid with standard low-order multigrid and applied it to an
unstructured stream-wise-upwind-Petrov-Galerkin (SUPG) discretization
of the incompressible Navier-Stokes equations. In the context of DG approximations, Hemker et al. [71] analyzed block Jacobi smoothing strategies with h-multigrid for 1D diffusion problems. In addition, p-multigrid,
or multi-order, solution strategies have been studied for high-order DG by
197

APPENDIX B. P -MULTIGRID
Helenbrook et al. [70], Bassi and Rebay [18] and Fidkowski et al. [48],
showing several advantages such as ease of implementation and orderindependent convergence rates. It was also applied to the SV method by
Van den Abeele et al. [174], Parsani et al. [128, 130] and Kannan et al.
[90], and to the SD method by Premasuthan et al. [133] and May et al.
[110].

B.1

Full approximation scheme

A two-level full approximation scheme algorithm (FAS) at p and p1 levels

can be summarized in the following way [24]. To solve a fine level problem
Rp (Wp ) = 0 perform the following operations:
Perform 1 smoothing sweeps on the fine level p to improve the solution Wp : Wp Gp Wp .
Transfer the state and the residual to the coarse level p 1:
p1
Wp1
Ip Wp ,
0




Ip1
Rp (Wp ).
Rp1 (Wp ) = Rp1 Wp1
f p1 Rp1 Wp1
p
0
0



Solve the coarse level problem: Rp1 Wp1 = f p1 .

Prolongate the coarse

level error
 and correct the fine level state:

.
Wp Wp + Ipp1 Wp1 Wp1
0

Perform 2 smoothing sweeps on the fine level to improve the solution

Wp : Wp Gp Wp .

In this algorithm, Gp represents an arbitrary smoothing operator on the

fine level and f p1 is the so-called forcing function. When the fine level
residual is zero, the coarse level correction is zero since Wp1
= Wp1 .
0
The coarse level problem can again be solved using a FAS algorithm, and
so on. In this way, one arrives at a V-cycle. A further increase in efficiency
can be achieved by initializing the solution on coarser levels, i.e. using
the full multigrid approach (FMG). In this way, a better initial solution is
provided for the fine levels, which will also improve the robustness of the
method. To determine when to switch approximation orders in the FMG
process, a residual-based approach can be used.
198

B.2. TRANSFER OPERATORS

B.2 Transfer operators

The prolongation and restriction operators between orders are local to the
elements. Their definitions are given for the SV method, with omission of
the cell index i. The definition for the SD method is analogous.
Prolongation state operator Ipp1 . For the SV method, the solution
PN s
within a cell is represented by a polynomial j=1 W
j Lj . The coarse
p1 can be written as a function of the fine level
level polynomials L
j
pm :
polynomials L
Ns

p1 =
L
j

p
X

p1 p
jm
Lm

s
j = 1, . . . , Np1
,

(B.1)

m=1
s
where Np1
and Nps are the number of CVs within a cell on the coarse
and fine level. By equating the fine level solution to the
coarse level
p1
.
solution, the following expression for Ipp1 is found: Ipp1 mj jm

p
State restriction operator Ip1 . This operator projects the fine level
solution
level polynomial basis. It is defined as
 p1  onto the coarse

1
I
= P Q
, with the matrices P and Q defined by
p
jm

Pjm =

jm

Qjm =

p1 L
p1
L
m d
j

p1 L
pm d
L
j

s
j, m = 1, . . . , Np1
,

(B.2)

s
j = 1, . . . , Np1
m = 1, . . . , Nps . (B.3)

Residual restriction operator Ip1

. The SV residuals are CV-averaged
p
quantities, like the SV solution variables. It is defined as


Ip1
p

jm

 p1 
I
p1
p
j

199

jm

1
.
pm

(B.4)

APPENDIX B. P -MULTIGRID

200

Appendix C

Newton-Raphson GMRES
solver

The need to solve non-linear systems of algebraic equations is ubiquitous throughout computational physics. Such systems typically arise from
the discretization of partial differential equations (PDEs), whether scalar
(such as heat conduction) or a system of coupled equations (such as the
Navier-Stokes equations). One may be interested in the steady-state solution of these equations (a boundary value problem) or in their dynamical
evolution (an initial value problem). For boundary value problems, nonlinear iterative methods are desirable. The same is true for multiple timescale initial value problems, when discretized implicitly at each time step.
In this appendix, the the Newton-Raphsons method coupled with a generalized minimum residual (GMRES) method, developed by Saad and Schultz
[151], is presented. This solver is available in the COOLFluiD collaborative simulation environment, developed at the von Karman Institute
for Fluid Dynamics [95, 136, 137]. In Chapter 8, the Newton-Raphson
GMRES algorithm is used as reference efficient algebraic solver to assess
the convergence properties of the non-linear LU-SGS algorithm for steadystate flow simulations.
201

APPENDIX C. NEWTON-RAPHSON GMRES SOLVER

C.1

Newton-Raphson algorithm

The overall goal of a non-linear algebraic solver is obtaining a solution to

the following problem:
Given F : RN RN ,

find x such that F (x ) = 0,

where N is the length of the unknown column vector x . For example,
considering the non-linear algebraic system associated to a general spatial
discretization and the backward Euler time marching scheme, this translates in: find the solution vector at the new time level Wn+1 , starting from
the solution vector at the previous time level Wn , of the following nonlinear algebraic system


Wn+1 Wn
(C.1)
= R Wn+1 = Rn+1 .
t
In the present work, the column vector W contains all the solution variables on the grids. The non-linear system (C.1) can be solved with any
method for such a system, e.g. the Newton-Raphson algorithm. Writing
the m-th approximation of Wn+1 as Wn+1,m , and the update to the (m +
1)-th approximation as Wn+1,m+1 = Wn+1,m+1 Wn+1,m , the NewtonRaphson algorithm is given by
m


I
Wm Wn
R 
m+1
m
+
W
=
R

,
(C.2)

W 
t
t
where I is the unity matrix and superscripts n and n + 1 denote the time
levels tn and tn+1 . Time step is given by t = tn+1 tn . This expression
is a linear algebraic system, which should be inverted at each inner iteration m of the Newton-Raphson algorithm. This method is an extremely
powerful technique- in general the convergence is quadratic: the error is
essentially squared at each step.

Solving

system (C.2) may be computationally expensive, on the order of
O N 3 arithmetic operations for direct linear algebra methods, where N
is the number of unknown of the system. In large-scale problems, where N
is 104 , 105 or greater, direct solutions are often infeasible. An alternative
is to use iterative linear algebra methods to form an approximation of the
solution. Using an iterative linear solver to obtain approximate Newton
steps results in a Newton iterative method, or a truncated Newton method.
Candidates for this are for instance the Jacobi method or the (symmetric)
Gauss-Seidel method, with or without relaxation [135]. Here, the GMRES
algorithm is discussed for the inversion of the linear systems.
202

C.2. GMRES

C.2 GMRES
GMRES is a particular Krylov subspace method and a key to derive it
is the Arnoldi process [7]. This method constitutes a class of algorithms
designed to solve a linear algebraic problem:
Given A RN N and b RN ,
find x such that Ax = b.

A Krylov subspace method begins with an initial x(0) and at the l-th step,
determines an iterate x(l) through a correction in the l-th Krylov subspace,
which is defined by the following set of basis vectors
r(0) , Ar(0) , A2 r(0) , ..., Al1 r(0) ,

(C.3)

where r(0) = b Ax(0) is the initial residual1 . In GMRES, each iterate x(l)
is chosen to minimize the residual norm



 (l)
(C.4)
Ax b .

This is a linear least squares problem, which should be solved at each iteration l. The convergence of the GMRES algorithm is monotonous, since
the Krylov subspace m contains the entire subspace l 1. It is guaranteed
to converge when l is equal to the size of the matrix A. However, it often
performs much better and a good approximation of the exact solution x is
obtained after a few iterations, for certain classes of matrices [171]. Preconditioning methods, like ILU-preconditioners or the additive Schwarz
method, are often used to accelerate the convergence of the GMRES algorithm. Like the full orthogonalization method or Arnoldis method for
linear system, the GMRES requires a big computational cost. For this reason, two variants of the classical GMRES have been developed, namely the
l-step restart GMRES and the Quasi-GMRES. More details on these two
variants can be found in Saad and Schultz [151] and Saad and Wu [152].
With methods that are based on the solution of a sparse linear system
like (C.2), the amount of memory required to store the sparse matrix may
be a serious problem. For instance, on a grid with N tetrahedral cells
with solution polynomial degree p, and with classical diffusive treatment
approach, i.e. the LSV/LSD approach (see for instance Section 4.1.3 and
4.2.3), the residuals in one cell depend on the solution in the current cell,
1 Here,

the iteration index is enclosed between brackets to avoid confusion with the exponent of the matrix A in Equation (C.3).

203

APPENDIX C. NEWTON-RAPHSON GMRES SOLVER

the neighbor cells, and the neighbors of the neighbor cells. Therefore, the
total number of non-zero entries in the sparse matrix is then about
17N

2
(p + 1) (p + 2) (p + 3)
# physical variables .
6

(C.5)

With fully compact diffusive approaches, like the BR2 approach described
in Section 4.2.3, only the immediate neighbors are required for the computation of the residuals in a cell. In this case, the number of non-zero entries
is smaller in comparison with the LSV/LSD approach and it is about


(p + 1) (p + 2) (p + 3)
# physical variables
5N
6

2

(C.6)

These numbers rapidly increase with p, to the sixth power. Upon comparison with the number of non-zero elements in the Jacobian matrix of
the non-linear LU-SG method, (5.9), it is clear that the GMRES method
requires significantly more memory than the non-linear LU-SG algorithm.

204

Appendix D

ESDIRK schemes
Implicit multi-step backward difference formulae (BDF) compute each solution vector update to design order of accuracy using one non-linear equation solve per time step. Unfortunately, they are not A-stable above secondorder (BDF2) and self-starting. Practical experience indicates that large
scale engineering computations are seldom stable if run with the fourthorder backward difference formula (BDF4) [112]. The third-order backward difference formula (BDF3), with its smaller region of instability, is
often stable but diverges for certain problems and some spatial operators.
The focus of the studies of Bijl et al. [20, 21], Carpenter et al. [31] and
Isono and Zingg [84] is on high-order implicit Runge-Kutta (I-RK) methods, in particular explicit-first-stage, single-diagonal-coefficient, diagonallyimplicit Runge-Kutta (ESDIRK) methods of various orders. Implicit multistage schemes do not face the A-stability restriction of the BDF, and ESDIRK schemes can be of arbitrarily high-order while retaining A-stability
[21]. In general, higher-order schemes in both space and time become more
attractive as the need for accuracy increases, i.e. error tolerances decrease.
Consequently, ESDIRK family of schemes, especially the fourth-order variant, could be an efficient alternative to second-order methods (of which the
BDF2 which is A-stable is the most popular) for problems requiring high
accuracy, such as turbulent numerical simulations and long-range wave
propagation in computational aeroacoustics.
A general ESDIRK scheme with s stages is given by the following [20]:
205

APPENDIX D. ESDIRK SCHEMES

 
Wk Wn X
akj R Wj = 0,

t
j=1

(D.1a)

Wn+1 = Wn + t

(D.1b)

s
X
j=1

 
bj R W j ,

where superscripts n and n + 1 denote the time levels tn and tn+1 , and
aij and bj are the stage and the main scheme weights. Table D.1 shows a
Butcher table of the coefficients for a general six-stage ESDIRK scheme,
where ci are the abscissae that denote the point in the time, t+ ci t, where
the stage is evaluated. ESDIRK schemes differ from traditional SDIRK
(see Hairer and Wanner [64]) methods by the choice a11 = 0. This means
that the first stage is explicit, i.e Wk=1 = Wn .
Table D.1: Sample Butcher table for a general six-stage ESDIRK scheme.

0
c2
c3
c4
c5
c6

0
a21
a31
a41
a51
a61
b1

0
a22
a32
a42
a52
a62
b2

0
0
a33
a43
a53
a63
b3

0
0
0
a44
a54
a64
b4

0
0
0
0
a55
a65
b5

0
0
0
0
0
a66
b6

Expression (D.1a) is a non-linear system of algebraic equations which must

be solved at each stage. Applying the non-linear LU-SGS algorithm to
(D.1a), results in the following general equation for the stage k:


akk

n

Rcc 
I ^ k,m+1
= akk Rcc (W ) +
+
Wcc
Wcc 
t

k
X
j=1

 
akj Rcc Wj

Wk,m
cc
,
t

(D.2)

where the subscripts cc and denote the current cell1 and the most recent solution when doing forward and backward sweeps respectively, and
k,m+1
^
W
= Wk,m+1 Wk,m , Wk,m = Wk,m Wn .
cc
cc

cc

cc

1 The

cc

cc

non-linear LU-SGS algorithm solves the non-linear system of equations with multiple cell-wise symmetric forward and backward sweeps.

206

The coupling between ESDIRK schemes and the non-linear LU-SGS algorithm was presented for the first time in 2007 by Parsani et al. [128].

207

APPENDIX D. ESDIRK SCHEMES

208

List of publications
Journal articles
Accepted
1. M. Parsani, G. Ghorbaniasl, C. Lacor, and E. Turkel. An implicit
high-order spectral difference approach for large eddy simulation. J.
Comput. Phys., 229(14):5373-5393, 2010.
2. M. Parsani, K. Van den Abeele, C. Lacor, and E. Turkel. Implicit LUSGS algorithm for high-order methods on unstructured grids with pmultigrid strategy for solving the steady Navier-Stokes equations. J.
Comput. Phys., 229(3):828-850, 2010.
3. K. Van den Abeele, G. Ghorbaniasl, M. Parsani, and C. Lacor. A stability analysis for the spectral volume method on tetrahedral grids.
J. Comput. Phys., 228(2):257-265, 2009.

Submitted
1. M. Parsani, G. Ghorbaniasl, and C. Lacor. Validation and application of an LES-high-order spectral difference method for flow induced
noise simulation. J. Comput. Acoust., Submitted, 2010.
2. M. Parsani, G. Ghorbaniasl, and C. Lacor. Large eddy simulation of
a muffler using an implicit spectral difference approach. J. Comput.
Phys., To be submitted, 2010.
3. M. Parsani, G. Ghorbaniasl, and C. Lacor. Stability analysis and
application of 3rd- and 4th-order spectral volume scheme with pmultigrid for solving the steady Navier-Stokes equations. J. Comput.
Phys., First revision, November 2010.
209

LIST OF PUBLICATIONS

Lecture notes
1. K. Van den Abeele, M. Parsani, and C. Lacor. Spectral volume and
spectral difference methods: wave propagation analysis and efficient
solvers. In Lecture notes of Von Karman Institute Lecture Series: 35th
CFD VKI/ADIGMA Course on Very High-Order Discretization Methods, Sint-Genesius-Rode, Belgium, October 2008. VKI LS 2008-08,
ISBN 978-2-930389-88-5.
2. K. Van den Abeele, M. Parsani, and C. Lacor. The high-order spectral difference method for unstructured grids. Cours-confrence Technologie et societe, Coll`ege Belgique, Bruxelles, Belgique, September
2009.

Conference proceedings
1. M. Parsani, G. Ghorbaniasl, and C. Lacor. Development of LES
high-order spectral difference method for flow induced noise simulation. In Proceedings of 16th AIAA/CEAS Aeroacoustics Conference,
Stockholm, Sweden, June 2010. AIAA Paper 2010-3816.
2. M. Parsani, G. Ghorbaniasl, M. Bilka, C. Lacor, and E. Turkel. On
the coupling between a high-order spectral difference method and
large eddy simulation. In Proceedings of the V European Conference
on Computational Fluid Dynamics (ECCOMAS CFD 2010). Mynisimposium: Towards industrial application of higher order methods, Lisbon, Portugal, June 2010.
3. M. Parsani, K. Van den Abeele, D. Vucinic, and C. Lacor. An implicit LU-SGS algorithm for high-order spectral difference method
on unstructured grids for turbulent compressible flows. In International Symposium on Coupled Methods in Numerical Dynamics,
Split, Croatia, September 2009.
4. M. Parsani, K. Van den Abeele, C. Lacor, and E. Turkel. Simulation
of compressible turbulent flows with an implicit LU-SGS algorithm
for high-order spectral difference method on unstructured grids. In
Proceedings of 39th AIAA Fluid Dynamics Conference and Exhibit,
San Antonio, Texas, June 2009. AIAA Paper 2009-4143.
5. K. Van den Abeele, M. Parsani, and C. Lacor. An implicit spectral
difference Navier-Stokes solver for unstructured hexahedral grids. In
210

LIST OF PUBLICATIONS
Proceedings of 47th AIAA Aerospace Sciences Meeting and Exhibit,
Orlando, Florida, January 2009. AIAA Paper 2009-0181.
6. M. Parsani, K. Van den Abeele, C. Lacor, and E. Turkel. Implicit
LU-SGS algorithm for high-order methods on unstructured grids. In
Proceedings of Annual Seminar of the ERCOFTAC Belgian Pilot Centre, Leuven, Belgium, December 2008.
7. M. Parsani, K. Van den Abeele, and C. Lacor. Implicit time integration algorithms for high-order methods on unstructured tetrahedral
grids With p-multigrid strategy. In Proceedings of 8th World Congress
on Computational Mechanics / 5th European Congress on Computational Methods in Applied Sciences and Engineering, Venice, Italy,
July 2008.
8. K. Van den Abeele, M. Parsani, C. Lacor, and T. Quintino. A spectral
volume Navier-Stokes solver on unstructured tetrahedral grids. In
Proceedings of 8th World Congress on Computational Mechanics / 5th
European Congress on Computational Methods in Applied Sciences
and Engineering, Venice, Italy, July 2008.
9. M. Parsani, K. Van den Abeele, and C. Lacor. An efficient LU-SGS
solver for high-order spectral volume schemes on unstructured grids.
In Proceedings Modern Techniques for Solving Partial Differential
Equations, Brussels, Belgium, June 2008.
10. M. Parsani, K. Van den Abeele, and C. Lacor. Implicit LU-SGS time
integration algorithm for high-order spectral volume method with
p-multigrid strategy. In Proceedings of West-East High Speed Flow
Field Conference, Moscow, Russia, November 2007.

211

LIST OF PUBLICATIONS

212

Bibliography
[1] Experimental data base for computer program assessment. AGARD
Report no. AR-138, May 1979.
[2] Partial derivative, Encyclopaedia of Mathematics. M. Hazewinkel,
Kluwer Academic, 2001.
[3] PETSc project web site, 2010. http://www.mcs.anl.gov/petsc/petscas/.
[4] K. K. Ahuja and J. Mendoza. Effects of the cavity dimensions, boundary layer, and temperature on the cavity noise with emphasis on
benchmark data to validate computational aeroacoustic codes. CR
4653, NASA, 1995.
[5] D. N. Arnold. An interior penalty finite element method with discontinuous elements. SIAM J. Numer. Anal., 19(4):742760, 1982.
[6] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini. Unified analysis of discontinuous Galerkin methods for elliptic problems. SIAM
J. Numer. Anal., 39(5):17491779, 2002.
[7] W. E. Arnoldi. The principle of minimized iterations in the solution of
the matrix eigenvalue problem. Quart. Appl. Math., 9:1729, 1951.
[8] G. Ashcroft and X. Zhang. A computational investigation of the noise
radiated by flow-induced cavity oscillations. AIAA paper, 2001-1914,
2001.
[9] J. Ask and L. Davidson. An acoustic analogy applied to the laminar
upstream flow over an open 2D cavity. C. R. Mecanique, 333(9):660
665, 2005.
213

BIBLIOGRAPHY
[10] H. L. Atkins and C.-W. Shu. Quadrature-free implementation of discontinuous Galerkin method for hyperbolic equations. AIAA Journal, 36:775782, 1998.
[11] J. B. Barlow, W. H. Rae, and A. Pope. Low-Speed Wind Tunnel Testing. John Wiley & Sons Inc., Canada, first edition, 1999.
[12] I. E. Barton. The entrance effect of laminar flow over a backwardfacing step geometry. Internat. J. Numer. Methods Fluids, 25:633
644, 1997.
[13] F. Bassi, A. Crivellini, S. Rebay, and M. Savini. Discontinuous
Galerkin solution of the Reynolds-averaged Navier-Stokes and k
turbulence model equations. Comput. Fluids, 34(4-5):507540, 2005.
[14] F. Bassi, A. Ghidoni, S. Rebay, and P. Tesini. High-order accurate
p-multigrid discontinuous Galerkin solution of the Euler equations.
Internat. J. Numer. Methods Fluids, 60(8):847865, 2008.
[15] F. Bassi and S. Rebay. A high-order accurate discontinuous finite element method for the numerical solution of the compressible NavierStokes equations. J. Comput. Phys., 131(2):267279, 1997.
[16] F. Bassi and S. Rebay. High-order accurate discontinuous finite
element solution of the 2D Euler equations. J. Comput. Phys.,
138(2):251285, 1997.
[17] F. Bassi and S. Rebay. Discontinuous Galerkin Methods: Theory, Computation and Applications, chapter GMRES discontinuous Galerkin solution of the compressible Navier-Stokes equations,
pages 197208. Springer Berlin, 2000.
[18] F. Bassi and S. Rebay. Numerical solution of the Euler equations
with a multiorder discontinuous finite element method. In Proceedings of Second International Conference on Computational Fluid Dynamics, Sydney, Australia, 2002.
[19] A. Bayliss and E. Turkel. Far field boundary conditions for compressible flows. J. Comput. Phys., 48(2):182199, 1982.
[20] H. Bijl, M. H. Carpenter, and V. N. Vatsa. Time integration schemes
for the unsteady Navier-Stokes equations. AIAA paper, 2001-2612,
2001.
214

BIBLIOGRAPHY
[21] H. Bijl, M. H. Carpenter, V. N. Vatsa, and C. A. Kennedy. Implicit
time integration schemes for the unsteady compressible NavierStokes equations. J. Comput. Phys., 179(1):313329, 2002.
[22] M. Bilka and J. Anthoine. Experimental investigation of flow noise
in a circular expansion using PIV and acoustic measurements. AIAA
paper, 2008-2952, 2008.
[23] D. Bouris and G. Bergeles. 2D LES of vortex shedding from a square
cylinder. J. Wind Eng. Ind. Aerodyn., 80:3146, 1999.
[24] A. Brandt. Guide to Multigrid Development. Springer, Berlin, 1982.
[25] A. Brandt. Multigrid techniques: 1984 guide with applications to
fluid dynamics. GMD-FIT GMD-Studie 85, 1985.
[26] M. Breuer and M. Poriquie. First experiences with LES of flows past
bluff bodies. In Proceedings of Third International Symposium on
Engineering Turbulence Modelling and Measurements, Amsterdam,
Nederlands, 1996.
[27] F. Brezzi, G. Manzini, D. Marini, P. Pietra, and A. Russo. Discontinuous Galerkin approximations for elliptic problems. Numer. Methods
Partial Differ. Equ., 16:365378, 2000.
[28] X.-C. Cai and M. Sarkis. A Restricted additive Schwarz preconditioner for general sparse linear systems. SIAM J. Sci. Comput.,
21(2):792797, 1999.
[29] H. B. Callen. Thermodynamics and an introduction to thermostatics.
John Wiley & Sons Inc., Chichester - New York - Brisbane - Toronto
- Singapore, second edition, 1985.
[30] M. H. Carpenter, S. A. . Viken, and E. J. Nielsen. The efficiency of
high order temporal schemes. AIAA paper, 2003-0086, 2003.
[31] M. H. Carpenter, S. A. Viken, and E. J. Nielsen. The efficiency of
high-order temporal schemes. AIAA paper, 2003-0086, 2003.
[32] Q.-Y. Chen. Partitions for spectral (finite) volume reconstruction in
the tetrahedron. J. Sci. Comput., 29(3):299319, 2006.
[33] Q.-Y. Chen. Partitions of a simplex leading to accurate spectral
(finite) volume reconstruction. SIAM J. Sci. Comput., 27(4):1458
1470, 2006.
215

BIBLIOGRAPHY
[34] R. F. Chen and Z. J. Wang. Fast block lower-upper symmetric Gauss
Seidel scheme for arbitrary grids. AIAA Journal, 38(12):22382245,
2000.
[35] Y.-H. Choi and C. L. Merkle. The application of preconditioning in
viscous flows. J. Comput. Phys., 105(2):207223, 1993.
[36] T. J. Chung. General Continuum Mechanics. Cambridge University
Press, Cambridge - New York - Melbourne - Madrid - Cape Town Paulo, second edition, 2007.
Sao
[37] B. Cockburn, S. Hou, and C.-W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation
laws IV: the multidimensional case. Math. Comput., 54(190):545
581, 1990.
[38] B. Cockburn, G. Karniadakis, and C.-W. Shu.
Discontinuous
Galerkin Methods. Theory, Computation and Applications. Lecture Notes in Computational Science and Engineering, 11, chapter
The development of discontinuous Galerkin methods, pages 350.
Springer Verlag, February 2000.
[39] B. Cockburn, S.-Y. Lin, and C.-W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws III: one-dimensional systems. J. Comput. Phys., 84(1):90
113, 1989.
[40] B. Cockburn and C.-W. Shu. The Runge-Kutta local projection P 1
Discontinuous Galerkin finite element method for scalar conservation laws. In Proceedings of First Fluid Dynamics Congress. University of Minnesota, Institute for Mathematics and its Applications
Preprint Series # 388, July 1988.
[41] B. Cockburn and C.-W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws II:
general framework. Math. Comput., 52(186):411435, 1989.
[42] B. Cockburn and C.-W. Shu. The local discontinuous Galerkin
method for time-dependent convection-diffusion systems. SIAM J.
Numer. Anal., 35(6):24402463, 1998.
[43] B. Cockburn and C.-W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws V:
multidimensional systems. J. Comput. Phys., 141(2):199224, 1998.
216

BIBLIOGRAPHY
[44] B. Cockburn and C.-W. Shu. Runge-kutta discontinuous galerkin
methods for convection-dominated problems. J. Sci. Comput.,
16(3):173261, 2002.
[45] C. F. Curtiss and J. O. Hirschfelder. Integration of stiff equations.
Proc. Nat. Acad. Sci., 38:235243, 1952.
[46] K. V. den Abeele. Development of high-order accurate methods on
unstructured grids. PhD thesis, Vrije Universiteit Brussel, Brussel,
May 2009.
M. Heitor, and J. Pereira. Measurements of turbulent and
[47] D. Durao,
periodic flows around a square cross-section cylinder. Exp. Fluids,
6:298304, 1995.
[48] K. J. Fidkowski, T. A. Oliver, J. Lu, and D. L. Darmofal. p-Multigrid
solution of high-order discontinuous Galerkin discretizations of the
compressible Navier-Stokes equations. J. Comput. Phys., 207:92
113, 2005.
[49] R. Franke and W. Rodi. Calculation of vortex shedding past a square
cylinder with various turbulence models. In Proceedings of Eight
Symposium on Turbulent Shear Flows, Munich, Germany, 1991.
[50] C. Fureby and G. Tabor. Mathematical and physical constraints on
large-eddy simulations. Theor. Comp. Fluid. Dyn., 9(2):85102, 1997.
[51] L. Galgani and A. Scotti. On subadditivity and convexity properties
of thermodynamic functions. Pure and Applied Chemistry, 22:229
235, 1970.
[52] H. Gao and Z. J. Wang. A high-order lifting collocation penalty formulation for the Navier-Stokes equations on 2D mixed grids. AIAA
paper, 2009-3784, 2009.
[53] C. W. Gear. Numerical initial value problems in ordinary differential
equations. Prentice Hall, first edition, 1971.
[54] C. Geuzaine and J.-F. Remacle. Gmsh: a three-dimensional finite
element mesh generator with built-in pre- and post-processing facilities. Int. J. Numer. Methods Eng., 79(11):13091331, 2009.
[55] M. Gharib and A. Roshko. The effect of flow oscillations on cavity
drag. J. Fluid Mech., 177:501530, 1987.
217

BIBLIOGRAPHY
[56] G. Ghorbaniasl. Computational aeroacoustic-noise prediction using
hybrid methodologies. PhD thesis, Vrije Universiteit Brussel, Brussel, September 2009.
[57] E. Godlewsky and P. A. Raviart. Numerical approximation of hyperbolic system of conservation laws, volume 118. Springer-Verlag New
York Inc., New York - Berlin - Heidelberg - Barcelona - Hong Kong London - Milan - Paris - Santa Clara - Singapore - Tokyo, 1996.
[58] S. Gottlieb and C.-W. Shu. Total variation diminishing Runge-Kutta
schemes. Math. Comput., 67(221):7385, 1998.
[59] B. Gustafsson and A. Sundstrom. Incompletely parabolic problems
in fluid dynamics. J. SIAM Appl. Math., 35(2):343357, 1978.
[60] W. Hackbush. Multi-grid Methods and Applications.
Verlag, Berlin - New York, 1985.

Springer-

[61] T. Haga, H. Gao, and Z. J. Wang. A high-order unifying discontinuous

formulation for 3d mixed grids. AIAA paper, 2010-540, 2010.
[62] T. Haga, N. Ohnishi, K. Sawada, and A. Masungana. Spectral volume computation of flow field in aerospace application using Earth
simulator. AIAA paper, 2006-2823, 2006.
[63] T. Haga, K. Sawada, and Z. J. Wang. An implicit LU-SGS scheme for
spectral volume method on unstructured tetrahedral grids. Commun. Comput. Phys., 6(5):978996, 2009.
[64] E. Hairer and G. Wanner. Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic Problems. Springer-Verlag,
Berlin, second edition, 1996.
[65] R. Harris and Z. J. Wang. Partition design and optimization for highorder spectral volume schemes. AIAA paper, 2009-1333, 2009.
[66] R. Harris and Z. J. Wang. Partition design and optimization for highorder spectral volume schemes on tetrahedral grids. AIAA paper,
2010-720, 2010.
[67] R. Harris, Z. J. Wang, and Y. Liu. Efficient quadrature-free highorder spectral volume method on unstructured grids: theory and 2D
implementation. J. Comput. Phys., 227(3):16201642, 2008.
[68] B. T. Helenbrook. A two-fluid spectral element method. Comp. Meth.
Appl. Mech. Eng., 191(3-5):273294, 2001.
218

BIBLIOGRAPHY
[69] B. T. Helenbrook and H. L. Atkins. Application of p-multigrid to
discontinuous Galerkin formulations of the Poisson equation. AIAA
Journal, 44(3):566575, 2005.
[70] B. T. Helenbrook, D. J. Mavriplis, and H. L. Atkins. Analysis of pmultigrid for continuous and discontinuous finite element discretizations. AIAA paper, 2003-3989, 2003.
[71] P. Hemker, W. Hoffmann, and M. van Raalte. Two-level Fourier analysis of a multigrid approach for discontinuous Galerkin discretization. SIAM J. Sci. Comput., 25(3):10181041, 2004.
[72] R. D. Henderson. Details of the drag curve near the onset of vortex
shedding. Phys. Fluids, 7:21022104, 1995.
[73] J. Hesthaven and D. Gottlieb. Stable spectral methods for conservation laws on triangles with unstructured grids. Comput. Meth. Appl.
Mech. Eng., 175:361381, 1999.
[74] C. Hirsch. Numerical computations of internal and external flows.
Volume 1. John Wiley & Sons Inc., New York - Chichester - Brisbane
- Toronto - Singapore, first edition, 1988.
[75] C. Hirsch. Numerical computations of internal and external flows.
Volume 1. John Wiley & Sons Inc., New York - Chichester - Brisbane
- Toronto - Singapore, second edition, 2007.
[76] F. Q. Hu, M. Y. Hussaini, and P. Rasetarinera. An analysis of the
discontinuous Galerkin method for wave propagation problems. J.
Comput. Phys., 151:921946, 1999.
[77] P. G. Huang, Z. J. Wang, and Y. Liu. An implicit space-time spectral
difference method for discontinuity capturing using adaptive polynomials. AIAA paper, 2005-5255, 2005.
[78] J. Humphrey, A. Taylor, and J. Whitelaw. Laminar flow in a square
duct of strong curvature. J. Fluid Mech, 83(3):509527, 1977.
[79] H. T. Huynh. A flux reconstruction approach to high-order schemes
including discontinuous Galerkin methods. AIAA paper, 2007-4079,
2007.
[80] H. T. Huynh. A reconstruction approach to high-order schemes including discontinuous Galerkin for diffusion. AIAA paper, 20090403, 2009.
219

BIBLIOGRAPHY
[81] G. Iaccarino and P. Durbin. Unsteady 3D RANS simulations using
the v 2 -f model. Center for Turbulence Research, Annual Research
Briefs, pages 263269, 2000.
[82] T. Iliescu, V. Jhon, W. J. Layton, G. Matthies, and L. Tobiska. A
numerical study of a class of LES models. Int. J. Comput. Fluid
Dyn., 17(1):7585, 2003.
[83] E. Isaacson and H. B. Keller. Analysis of numerical methods. John
Wiley & Sons Inc., New York, 1994.
[84] S. Isono and D. W. Zingg. A Runge-Kutta-Newton-Krylov algorithm
for fourth-order implicit time marching applied to unsteady flows.
AIAA paper, 2004-0433, 2004.
[85] A. Jameson. A proof of the stability of the spectral difference method
for all orders of accuracy. J. Sci. Comput., 45(1-3):348358, 2010.
[86] A. Jameson and D. A. Caughey. How many steps are required to
solve the Euler equations of steady, compressible flow: in search of a
fast solution algorithm. AIAA paper, 2001-2673, 2001.
[87] A. Jameson, W. Schmidt, and E. Turkel. Numerical solution of the
Euler equations by finite volume methods using Runge-Kutta time
stepping schemes. AIAA paper, 81-1259, 1981.
[88] A. Jameson and E. Turkel. Implicit schemes and LU decompositions.
Math. Comput., 37(156):385397, 1981.
[89] A. Jameson and S. Yoon. Lower-upper implicit schemes with multiple grids for the Euler equations. AIAA Journal, 25(7):929935,
1987.
[90] R. Kannan, Y. Sun, and Z. J. Wang. A study of viscous flux formulations for an implicit p-multigrid spectral volume Navier-Stokes
solver. AIAA paper, 2008-783, 2008.
[91] R. Kannan and Z. J. Wang. A study of viscous flux formulations for
a p-multigrid spectral volume Navier-Stokes solver. J. Sci. Comput.,
41(2):165199, 2009.
[92] D. A. Knoll and D. A. Keyes. Jacobian-free Newton-Krylov methods: a survey of approaches and applications. J. Comput. Phys.,
193(2):357397, 2004.
220

BIBLIOGRAPHY
[93] D. A. Kopriva. A conservative staggered-grid Chebyshev multidomain method for compressible flows. II semi-structured method. J.
Comput. Phys., 128(2):475488, 1996.
[94] D. A. Kopriva and J. H. Kolias. A conservative staggered-grid Chebyshev multidomain method for compressible flows. J. Comput. Phys.,
125(1):244261, 1996.
[95] A. Lani. An object-oriented and high-performance platform for
aerothermodynamics simulation. PhD thesis, Universite Libre de
Bruxelles, von Karman Institute for Fluid Dynamics, December
2008.
[96] J. Larsson, L. Davidson, M. Olsson, and L.-E. Eriksson. Aeroacoustic
investigation of an open cavity at low Mach number. AIAA Journal,
42(12):24622473, 2004.
[97] M. Lesieur and P. Comte. Favre filtering and macro-temperature in
large eddy simulation of compressible turbulence, volume 329, pages
363368. C. R. Acad. Sci. Paris, 2001.
[98] R. J. LeVeque. Simulation and modeling of turbulent flows. Lecture
in Mathematics. Birkhauser-Verlag, Basel - Boston - Berlin, 1994.
[99] R. J. LeVeque. Numerical methods for conservation laws. Lecture
in Mathematics. Birkhauser-Verlag, Basel - Boston - Berlin, second
edition, 1999.
[100] R. J. LeVeque. Finite volume methods for hyperbolic problems.
Cambridge University Press, Cambridge - New York - Melbourne
- Madrid - Cape Town, 2002.
[101] C. Liang, R. Kannan, and Z. J. Wang. A p-multigrid spectral difference method with explicit and implicit smoothers on unstructured
triangular grids. Comput. Fluids, 38(2):254265, 2009.
[102] C. Liang, S. Premasuthan, A. Jameson, and Z. J. Wang. Large eddy
simulation of compressible turbulent channel flow with spectral difference method. AIAA paper, 2009-402, 2009.
[103] Y. Liu, M. Vinokur, and Z. J. Wang. Spectral difference method
for unstructured grids I: basic formulation. J. Comput. Phys.,
216(2):780801, 2006.
221

BIBLIOGRAPHY
[104] Y. Liu, M. Vinokur, and Z. J. Wang. Spectral (finite) volume method
for conservation laws on unstructured grids V: extension to threedimensional systems. J. Comput. Phys., 212(2):454472, 2006.
[105] J. L. Lumley.
Simulation and modeling of turbulent flows.
ICASE/LaRC in Computational Science and Engineering. Oxford
University Press Inc., New York - Oxford, 1996.
[106] D. Lyn, S. Einav, W. Rodi, and J. Park. A laser doppler velocimetry study of ensemble-averaged characteristics of the turbulent near
wake of a square cylinder. J. Fluid Mech, 304:285319, 1995.
[107] Y. Maday and R. Munoz. Spectral element multigrid. Part 2. Theoretical justification. Technical Report ICASE 88-73, 1988.
[108] P. Marinov and G. Lebon. A generalized Fouriers law in thermodynamically simple materials of the differential type. Acta Mech.,
22(3-4):257265, 1975.
[109] D. J. Mavriplis. Multigrid strategies for viscous flow solvers on
anisotropic unstructured meshes. J. Comput. Phys., 145(1):141165,
1998.
[110] G. May, F. Iacono, and A. Jameson. Efficient algorithms for highorder discretizations of the Euler and Navier-Stokes equations.
AIAA paper, 2009-0182, 2009.
[111] G. May and A. Jameson. A spectral difference method for the Euler
and Navier-Stokes equations. AIAA paper, 2006-304, 2006.
[112] N. D. Melson, M. D. Sanetrik, and H. L. Atkins. Time accurate
Navier-Stokes calculations with multigrid acceleration. In Proceedings of the Sixth Copper Mountain conference on multigrid methods,
pages 423437, Copper Mountain, Colorado, April 1993.
[113] W. A. Mulder. A high-resolution Euler solver based on multigrid,
semi-coarsening, and defect correction. J. Comput. Phys., 100(1):91
104, 1992.
[114] S. Murakami and A. Mochida. On turbulent vortex shedding flow
past a square cylinder predicted by CFD. J. Wind Eng. Ind. Aerodyn.,
54-55:191211, 1995.
[115] C. R. Nastase and D. J. Mavriplis. High-order discontinuous
Galerkin methods using hp-multigrid approach. J. Comput. Phys.,
213(1):330357, 2006.
222

BIBLIOGRAPHY
[116] F. Nicoud and F. Ducros. Subgrid-scale stress modelling based on
the square of the velocity gradient tensor. Flow. Turbul. Combust.,
62(3):183200, 1999.
[117] C. Norberg. Flow around rectangular cylinders: pressure forces
and wake frequencies. J. Wind Eng. Ind. Aerodyn., 49(1-3):187196,
1993.
[118] A. Okajima, T. Nagahisa, and A. Rokugoh. A numerical analysis of
flow around rectangular cylinders. JSME Int J., Ser. II, 33:702711,
1990.
[119] C.-W. S. S. Osher. Efficient implementation of essentially nonoscillatory shock-capturing schemes. J. Comput. Phys., 77:439471,
1988.
[120] R. L. Panton. Incompressible Flow. John Wiley & Sons Inc., New
York - Chichester - Brisbane - Toronto - Singapore, second edition,
1996.
[121] M. Parsani, K. V. den Abeele, and C. Lacor. Implicit time integration algorithms for high-order methods on unstructured tetrahedral grids with p-multigrid strategy. In Proceedings of 8th
World Congress on Computational Mechanics (WCCM8)/5th European Congress on Computational Methods in Applied Sciences and
Engineering (ECCOMAS 2008), Venice, Italy, July 2008.
[122] M. Parsani, G. Ghorbaniasl, M. Bilka, C. Lacor, and E. Turkel. On
the coupling between a high-order spectral difference method and
large eddy simulation. In Proceedings of V European Conference on
Computational Fluid Dynamics (ECCOMAS CFD 2010), Mynisimposium: Towards industrial application of higher order methods,
Lisbon, Portugal, June 2010.
[123] M. Parsani, G. Ghorbaniasl, and C. Lacor. Development of LES highorder spectral difference method for flow induced noise simulation.
AIAA paper, 2010-3816, 2010.
[124] M. Parsani, G. Ghorbaniasl, and C. Lacor. Large eddy simulation of
a muffler using an implicit spectral difference approach. J. Comput.
Phys., Submitted, 2010.
[125] M. Parsani, G. Ghorbaniasl, and C. Lacor. Validation and application
of an les-high-order spectral difference method for flow induced noise
simulation. J. Comput. Acoust., Submitted, 2010.
223

BIBLIOGRAPHY
[126] M. Parsani, G. Ghorbaniasl, C. Lacor, and E. Turkel. An implicit
high-order spectral difference approach for large eddy simulation.
J. Comput. Phys., 229(14):53735393, 2010.
[127] M. Parsani, G. Ghorbaniasl, K. Van den Abeele, and C. Lacor. Stability analysis and application of 3rd- and 4th-order spectral volume
scheme with p-multigrid for solving the steady Navier-Stokes equations. J. Comput. Phys., First revision, November 2010.
[128] M. Parsani, K. Van den Abeele, and C. Lacor. Implicit LU-SGS time
integration algorithm for high-order spectral volume method with
p-multigrid strategy. In Proceedings of West-East High Speed Flow
Field Conference, Moscow, Russia, November 2007.
[129] M. Parsani, K. Van den Abeele, C. Lacor, and E. Turkel. Simulation of compressible turbulent flows with an implicit LU-SGS algorithm for high-order spectral difference method on unstructured
grids. AIAA paper, 2009-4143, 2009.
[130] M. Parsani, K. Van den Abeele, C. Lacor, and E. Turkel. Implicit LUSGS algorithm for high-order methods on unstructured grid with pmultigrid strategy for solving the steady Navier-Stokes equations.
J. Comput. Phys., 229(3):828850, 2010.
[131] T. J. Poinsot and S. Lele. Boundary conditions for direct simulation of
compressible viscous flows. J. Comput. Phys., 101(1):104129, 1992.
[132] S. B. Pope. Turbulent Flows. Cambridge University Press, Cambridge - New York - Melbourne - Madrid - Cape Town, 2003.
[133] S. Premasuthan, C. Liang, A. Jameson, and Z. J. Wang. A p-multigrid
spectral difference method for viscous compressible flow using 2D
quadrilateral meshes. AIAA paper, 2009-0950, 2009.
[134] L. Quartapelle, F. Auteri, and A. Guardone. Fluid dynamics course
notes I. Technical report, Politecnico di Milano, Department of
aerospace engineering, 2007.
[135] A. Quarteroni, R. Sacco, and F. Saleri. Numerical mathematics.
Springer-Verlag, New York - Berlin - Heidelberg, 2002.
[136] T. Quintino. A component environment for high-performance scientific computing: design and implementation. PhD thesis, Katholieke
Universiteit Leuven, von Karman Institute for Fluid Dynamics, December 2008.
224

BIBLIOGRAPHY
[137] T. Quintino, D. Kimpe, T. Wuilbaut, and A. Lani. COOLFluiD project
web site, 2010. http://coolfluidsrv.vki.ac.be/coolfluid.
[138] J. Ramboer, T. Broeckhoven, S. Smirnov, and C. Lacor. Optimization
of time integration schemes coupled to spatial discretization for use
in CAA applications. J. Comput. Phys., 213:777802, 2006.
[139] P. Rasetarinera and M. Y. Hussaini. An efficient implicit discontinuous Galerkin method. J. Comput. Phys., 172(2):718738, 2001.
[140] W. Reed and T. Hill. Triangular mesh methods for the neutron transport equation. Technical Report LA-UR 73-479, Los Alamos National
Laboratory, 1973.
[141] D. Rockwell and E. Naudasher. Review - self-sustaining oscillations
of flow past a cavity. J. Fluid Eng.-T. ASME, 100(2):152165, 1978.
[142] W. Rodi. On the simulation of turbulent flow past bluff bodies. J.
Wind Eng. Ind. Aerodyn., 46-47:319, 1993.
[143] P. L. Roe. Approximate Riemann solvers, parameter vectors and
difference schemes. J. Comput. Phys., 43(2):357372, 1981.
[144] W. D. Roeck, G. Rubio, M. Baelmans, P. Sas, and W. Desmet. The
influence of flow domain modelling on the accuracy of direct and hybrid aeroacoustic noise calculations. AIAA paper, 2006-2419, 2006.
[145] S. Rogers, D. Kwak, and C. Kiris. Steady and unsteady solutions of the incompressible Navier-Stokes equations. AIAA Journal,
29(3):603610, 1991.
[146] E. M. Rnquist and A. T. Patera. Spectral element multigrid I.
Formulation and numerical results. J. Sci. Comput., 2(4):389406,
1987.
[147] A. Roshko. Perspectives on bluff body aerodynamics. J. Wind Eng.
Ind. Aerodyn., 49:79100, 1993.
[148] J. E. Rossiter. Wind tunnel experiments on the flow over rectangular
cavities at subsonic and transonic speeds. TR 3438, Aeronautical
Research Council, October 1964.
[149] C.-C. Rossow. Efficient computation of compressible and incompressible flows. J. Comput. Phys., 220(2):879899, 2007.
225

BIBLIOGRAPHY
[150] V. V. Rusanov. Calculation of interaction of non-steady shock waves
with obstacles. J. Comput. Math. Phys. USSR, 1:267279, 1961.
[151] Y. Saad and M. H. Schultz. GMRES: A generalized minimal residual
algorithm for solving nonsymmetric linear systems. SIAM J. Sci.
Stat. Comput., 7(3):856869, 1986.
[152] Y. Saad and K. Wu. DQGMRES: a direct quasi-minimal residual algorithm based on incomplete orthogonalization. Numer. Linear Algebra Appl, 3(4):329343, 1996.
[153] R. Sandberg, L. Jones, N. Sandham, and P. Joseph. Direct numerical
simulations of tonal noise generated by laminar flow past airfoils. J.
Sound Vibrat., 320(4-5):838858, 2008.
[154] N. D. Sandham, Q. Li, and H. C. Yee. Entropy splitting for high-order
numerical simulation of compressible turbulence. J. Comput. Phys.,
178(2):307322, 2002.
[155] E. N. Sarmin and L. A. Chudov. On the stability of the numerical
integration of systems of ordinary differential equations arising in
the use of the straight line method. U.S.S.R. Comput. Math. and
Math. Phys., 3(6):15371543, 1963.
[156] F. Scarano. Particle image velocimetry: development and application.
PhD thesis, Universita` Degli Studi di Napoli Federico II, Naples,
2000.
[157] W. E. Schiesser. The Numerical Method of Lines. Academic Press,
Inc., San Diego - New York - Boston - London - Sydney - Tokyo Toronto, first edition, 1991.
[158] D. Sharov and K. Nakahashi. Reordering of hybrid unstructured
grids for lower-upper symmetric Gauss-Seidel computations. AIAA
Journal, 36(3):484486, 1998.
[159] C.-W. Shu. Total variation diminishing Runge-Kutta schemes. SIAM
J. Sci. Stat Comp, 9:10791084, 1988.
[160] G. D. Smith. Numerical solution of partial differential equations.
Oxford Applied Mathematics and Computing Science. Oxford University Press Inc., Oxford - New York, third edition, 2004.
[161] A. Sohankar, C. Norberg, and L. Davidson. Numerical simulation of
unsteady low-reynolds number flow around rectangular cylinders at
incidence. J. Wind Eng. Ind. Aerodyn., 69-71:189201, 1997.
226

BIBLIOGRAPHY
[162] H. Sohr. The Navier-Stokes equations. An elementary functional analytic approach. Birkhauser Verlag, Basel - Boston - Berlin, 2001.
[163] Y. Sun and Z. J. Wang. Evaluation of discontinuous Galerkin and
spectral volume methods for scalar and system conservation laws on
unstructured grids. Internat. J. Numer. Methods Fluids, 45:819838,
2004.
[164] Y. Sun, Z. J. Wang, and Y. Liu. Spectral (finite) volume method
for conservation laws on unstructured grids VI: extension to viscous
flow. J. Comput. Phys., 215(1):4158, 2006.
[165] Y. Sun, Z. J. Wang, and Y. Liu. High-order multidomain spectral
difference method for the Navier-Stokes equations on unstructured
hexahedral grids. Commun. Comput. Phys., 2(2):310333, 2007.
[166] Y. Sun, Z. J. Wang, and Y. Liu. Efficient implicit non-linear LU-SGS
approach for compressible flow computation using high-order spectral difference method. Commun. Comput. Phys., 5:760778, 2009.
[167] Y. Sun, Z. J. Wang, and Y. Liu. Efficient implicit non-linear LUSGS approach for compressible flow computation using high-order
spectral difference method. Commun. Comput. Phys., 5(2-4):760
768, 2009.
[168] Y. Sun, Z. J. Wang, Y. Liu, and C. L. Chen. Efficient implicit LU-SGS
algorithm for high-order spectral difference method on unstructured
hexahedral grids. AIAA paper, 2007-313(7), 2007.
[169] R. C. Swanson, E. Turkel, and C.-C. Rossow. Convergence acceleration of Runge-Kutta schemes for solving the Navier-Stokes equations. J. Comput. Phys., 224(1):365388, 2007.
[170] K. Thompson. Time-dependent boundary conditions for hyperbolic
systems. J. Comput. Phys., 68(1):124, 1987.
[171] L. N. Trefethen and D. Bau. Numerical Linear Algebra. Society for
Industrial and Applied Mathematics, 1997.
[172] U. Trottenberg, C. W. Oosterlee, and A. Schuller. Multigrid. Academic Press, London - San Diego, 2001.
[173] E. Turkel. Preconditioned methods for solving the incompressible
and low speed compressible equations. J. Comput. Phys., 72(2):277
298, 1987.
227

BIBLIOGRAPHY
[174] K. Van den Abeele, T. Broeckhoven, and C. Lacor. Dispersion and
dissipation properties of the 1D spectral volume method and application to a p-multigrid algorithm. J. Comput. Phys., 224(2):616636,
2007.
[175] K. Van den Abeele, G. Ghorbaniasl, M. Parsani, and C. Lacor. A stability analysis for the spectral volume method on tetrahedral grids.
J. Comput. Phys., 228(2):257265, 2009.
[176] K. Van den Abeele and C. Lacor. An accuracy and stability study
of the 2D spectral volume method. J. Comput. Phys., 226(1):1007
1026, 2007.
[177] K. Van den Abeele, C. Lacor, and Z. J. Wang. On the stability and accuracy of the spectral difference method. J. Sci. Comput., 37(2):162
188, 2008.
[178] K. Van den Abeele, M. Parsani, and C. Lacor. An implicit spectral
difference Navier-Stokes solver for unstructured hexahedral grids.
AIAA paper, 2009-0181, 2009.
[179] K. Van den Abeele, M. Parsani, C. Lacor, and T. Quintino. A spectral volume Navier-Stokes solver on unstructured tetrahedral grids.
In Proceedings of 8th World Congress on Computational Mechanics (WCCM8)/5th European Congress on Computational Methods in
Applied Sciences and Engineering (ECCOMAS 2008), Venice, Italy,
July 2008.
[180] B. Van Leer. Towards the ultimate conservative difference scheme
IV. A new approach to numerical convection. J. Comput. Phys.,
23:276299, 1977.
[181] V. Venkatakrishnan and D. J. Mavriplis. Implicit solvers for unstructured meshes. J. Comput. Phys., 105(1):8391, 1993.
[182] B. Vreman, B. Geurts, and H. Kuerten. Subgrid modeling in les of
compressible flow. Appl. Sci. Res., 54(3):191203, 1995.
[183] Z. J. Wang. Spectral (finite) volume method for conservation
laws on unstructured grids: basic formulation. J. Comput. Phys.,
178(1):210251, 2002.
[184] Z. J. Wang and H. Gao. A unifying lifting collocation penalty formulation for the Euler equations on mixed grids. AIAA paper, 20090401, 2009.
228

BIBLIOGRAPHY
[185] Z. J. Wang and H. Gao. A unifying lifting collocation penalty
formulation including the discontinuous Galerkin, spectral volume/difference methods for conservation laws on mixed grids. J.
Comput. Phys., 228(21):81618186, 2009.
[186] Z. J. Wang and Y. Liu. Spectral (finite) volume method for conservation laws on unstructured grids II: extension to two-dimensional
scalar equation. J. Comput. Phys., 179(2):665697, 2002.
[187] Z. J. Wang and Y. Liu. Spectral (finite) volume method for conservation laws on unstructured grids III: one dimensional systems and
partition optimization. J. Sci. Comput., 20(1):137157, 2004.
[188] Z. J. Wang and Y. Liu. Extension of the spectral volume method to
high-order boundary representation. J. Comput. Phys., 211(1):154
178, 2006.
[189] Z. J. Wang, Y. Liu, G. May, and A. Jameson. Spectral difference
method for unstructured grids II: extension to the Euler equations.
J. Sci. Comput., 32(1):4571, 2007.
[190] Z. J. Wang, Y. Sun, C. Liang, and Y. Liu. Extension of the SD method
to viscous flow on unstructured grids. In Proceedings of the 4th international conference on computational fluid dynamics, Ghent, Belgium, July 2006.
[191] Z. J. Wang, L. Zhang, and Y. Liu. Spectral (finite) volume method
for conservation laws on unstructured grids IV: extension to twodimensional Euler equations. J. Comput. Phys., 194(2):716741,
2004.

[192] C. Wieselberger. Neuere Feststellungen uber

die Gesetze des

Flussigkeitsund luftwiderstands. Phys. Z., 22:321328, 1921.

[193] D. C. Wilcox. Turbulence modeling for CFD. DWC Industries, Inc.,
California, third edition, 2006.
[194] J. G. Wissink. DNS of 2D turbulent flow around a square cylinder.
Internat. J. Numer. Methods Fluids, 25(1):5162, 1998.
[195] T. Wuilbaut. Algorithmic developments for a multiphysics framework. PhD thesis, Universite Libre de Bruxelles, von Karman Institute for Fluid Dynamics, December 2008.
229

BIBLIOGRAPHY
[196] J. Y. Yang, S. C. Yang, Y. N. Chen, and C. A. Hsu. Implicit weighted
ENO schemes for three-dimensional incompressible Navier-Stokes
equations. J. Comput. Phys., 146(1):464487, 1998.
[197] I. Yasushi, A. M. Shih, and B. K. Soni. Octree-based reasonablequality hexahedral mesh generation using a new set of refinement
templates. Int. J. Numer. Meth. Engng., 77(13):18091833, 2009.
[198] T. A. Zang, R. B. Dahlburg, and J. P. Dahlburg. Direct and large
eddy simulation of three dimensional compressible navierstokes turbulence. Phys. Fluids, A4:127140, 1992.
[199] M. Zhang and C.-W. Shu. An analysis of and a comparison between
the discontinuous Galerkin and the spectral finite volume methods.
Comput. Fluids, 34:581592, 2005.
[200] Y. Zhou and Z. J. Wang. Absorbing boundary conditions for the Euler
and Navier-Stokes equations with the spectral difference method. J.
Comput. Phys., 229(23):87338749, 2010.
[201] Y. Zhou and Z. J. Wang. Simulation of CAA benchmark problems
using high-order spectral difference method and perfectly matched
layers. AIAA paper, 2010-838, 2010.

230