PHD Aldo Hennink

LowMach Number Flow and the
Discontinuous Galerkin Method

LowMach Number Flow and the
Discontinuous Galerkin Method
Dissertation
for the purpose of obtaining the degree of doctor

at Delft University of Technology
by the authority of the Rector Magnificus, Prof. dr. ir. T. H. J. J. van der Hagen,
chair of the Board for Doctorates
to be defended publicly on
Tuesday 22 Februari 2022 at 10:00 o’clock
by
Aldo HENNINK
Master of Science in Applied Physics, Delft University of Technology, The Netherlands

born in Amsterdam, the Netherlands.
This dissertation has been approved by the promotors.
Composition of the doctoral committee:

Rector Magnificus, chairperson
Prof. dr. ir. J. L. Kloosterman Delft University of Technology, promotor
Dr. ir. D. Lathouwers, Delft University of Technology, promotor
Independent members:
Prof. dr.ing. E. Laurien, University of Stuttgart, Germany
Prof. dr. ir. B. Koren, Eindhoven University of Technology, the Netherlands
Prof. dr. S. Hickel, Delft University of Technology
Dr. R. Pecnik, Delft University of Technology
Prof. dr. P. Dorenbos, Delft University of Technology, reserve member
Part of this work was funded by the sCO2 HeRo project that has received funding
from the European research and training program 2014–2018 (grant agreement
ID 662116).
Part of this work was supported by the ENEN+ project that has received funding
from the Euratom research and training Work Programme 2016–2017 — 1 #755576.
Keywords: LowMach Number Flow, Pressure Correction, Discontinuous

Galerkin
Copyright © 2022 by A. Hennink

ISBN XXXXXXXXXXXXX
An electronic version of this dissertation is available at
http://repository.tudelft.nl/.
Contents
Nomenclature vii
Summary ix
Samenvatting xi
1 Introduction 1
1.1 Mathematical Problem Setting . . . . . . . . . . . . . . . . . . . 1
1.2 Overview of the Numerical Method . . . . . . . . . . . . . . . . 3
1.2.1 Why Use a Discontinuous Galerkin Method? . . . . . . 4
1.3 Thesis Overview. . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Spatial Discretization with a Discontinuous Galerkin Method 9
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Overview of DG discretizations . . . . . . . . . . . . . . . 9
2.1.2 Construction of the Solution Space . . . . . . . . . . . . 10
2.2 Discrete Continuity Equation . . . . . . . . . . . . . . . . . . . 12
2.3 Discrete Momentum Equation . . . . . . . . . . . . . . . . . . . 14
2.3.1 Discretization of the Viscous Stress . . . . . . . . . . . . 14
2.4 Discrete Enthalpy Equation . . . . . . . . . . . . . . . . . . . . 16
2.5 Discretization of the Convection . . . . . . . . . . . . . . . . . . 16
2.5.1 Choice of the Numerical Flux. . . . . . . . . . . . . . . . 17
2.5.2 Solution Spaces for the Enthalpy and the Pressure . . 19
2.5.3 Proper Treatment of Dirichlet Boundary Conditions . . 20
2.6 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.7 Test Case: A Heated Backwardfacing Step . . . . . . . . . . . 23
2.8 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 26
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Pressure Correction 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.1 Fully Discrete Linear System . . . . . . . . . . . . . . . . 33
3.2 Pressure Correction Method . . . . . . . . . . . . . . . . . . . . 35
3.3 Verification with Manufactured Solutions . . . . . . . . . . . . 38
3.3.1 TaylorGreen Vortex . . . . . . . . . . . . . . . . . . . . . 38
3.3.2 Variableproperty Manufactured Solution . . . . . . . . 39
3.4 Validation with Flow Past a Circular Obstacle. . . . . . . . . . 42
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.1 Pressure Correction with Equalorder Discretizations . 48
v
vi Contents
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4 Handling the Enthalpy Equation for LowMach Number Flow 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1.1 In Between Compressible and Incompressible. . . . . . 53
4.1.2 Which Enthalpy Equation Should be Solved? . . . . . . 54
4.2 The Temporal Density Gradient . . . . . . . . . . . . . . . . . . 57
4.3 Linearizing the Temporal Derivative of the Enthalpy . . . . . . 58
4.3.1 Error Estimates and Stability . . . . . . . . . . . . . . . 59
4.3.2 Proper Scaling of the Enthalpy Equation . . . . . . . . . 61
4.3.3 Special Case of an Ideal Gas . . . . . . . . . . . . . . . . 62
4.4 Test Case for the Spaceindependent Enthalpy Equation . . . 63
4.5 Test Cases with LowMach Number Flow. . . . . . . . . . . . . 65
4.5.1 Variabledensity Manufactured Solution . . . . . . . . . 65
4.5.2 Validation with Flow Past a Heated Circular Obstacle . 72
4.6 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 73
Appendices 76
A Derivations of the Results in Section 4.3.1. . . . . . . . . . . . 76
A.1 Derivations for Method #1 . . . . . . . . . . . . . . . . . 76
A.2 Derivations for Method #2 . . . . . . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5 Channel Flow and Large Eddy Simulation 83
5.1 Introduction and Governing Equations . . . . . . . . . . . . . . 83
5.2 Subfilter Scale models . . . . . . . . . . . . . . . . . . . . . . . 86
5.3 Numerical Simulation . . . . . . . . . . . . . . . . . . . . . . . . 89
5.3.1 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.3.2 Including a Variable Density . . . . . . . . . . . . . . . . 91
5.4 Test Case: Infinite Plane Channel Flow. . . . . . . . . . . . . . 92
5.4.1 Dimensionless Analysis . . . . . . . . . . . . . . . . . . . 93
5.4.2 Initial Condition . . . . . . . . . . . . . . . . . . . . . . . 94
5.4.3 Domain Size and Mesh . . . . . . . . . . . . . . . . . . . 94
5.4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6 Conclusion 107
Acknowledgements 109
Curriculum Vitæ 111
List of Publications 113
Nomenclature
General
∇𝑖 gradient in the 𝑖 th Cartesian coordinate (e.g., ∇2 ≔ 𝜕/𝜕𝑦)
𝜌𝑇 , ℎ𝜌 , 𝐻ℎ , … partial derivatives with respect to fluid properties (e.g., 𝜌𝑇 ≔
𝜕𝜌/𝜕𝑇)
𝐮, 𝐅, … vectorvalues quantities (e.g., velocity, force,)
𝒎, 𝒉, 𝒑, … vectors in the discrete global linear system
𝑴, 𝑫, 𝑪, … matrices in the discrete global linear system
List of symbols
Independent variables and domain:
𝐫 = [𝑥, 𝑦, 𝑧] Cartesian spatial coordinate
𝑡 time
Ω spatial domain
𝐧 outward normal vector
𝜕Ω, 𝜕𝑇, … boundaries of Ω, 𝑇, …
𝜕ΩD , 𝜕ΩN parts of 𝜕Ω with a Dirichlet (D) or Neumann (N) boundary con
dition
Related to the spatial discretization:
𝒯 the set of elements (i.e., the computational mesh)
ℱ the set of all faces in the mesh
ℱi , ℱ D , ℱN sets of all internal (i), Dirichlet (D), or Neumann (N) faces
𝒫𝑋 polynomial order of a numerical solution space of quantity 𝑋
Variables in the transport equations:
𝑝th thermodynamic pressure
𝑝 ‘incompressible’ (or ‘hydrodynamic’, or ‘mechanical’) pressure
𝐦 mass flux
𝐮 ≔ (1/𝜌)𝐦 velocity
𝜏 viscous stress tensor
𝐅 volumetric external force
𝑄 volumetric heat source
𝜙D Dirichlet value of a quantity 𝜙 at a boundary (i.e., the inhomo
geneous part of a Dirichlet boundary condition)
𝐟N stress at the outflow boundary (i.e., the inhomogeneous part of
a Neumann boundary condition for the momentum)
vii
viii Nomenclature
𝑞N heat flux out of the domain at the boundary (i.e., the inhomoge
neous part of a Neumann boundary condition for the enthalpy)
Fluid properties:
𝑇 temperature
𝜌 density
ℎ specific enthalpy
𝐻 = 𝜌ℎ volumetric enthalpy
ℎ0 enthalpy offset
𝑐𝑝 ≔ ℎ𝑇 specific heat capacity at constant thermodynamic pressure
𝑘 thermal conductivity
𝛼 ≔ 𝑘/ (𝜌𝑐𝑝 ) thermal diffusivity
𝜇 dynamic viscosity
𝜈 ≔ 𝜇/𝜌 kinematic viscosity
Dimensionless numbers:
Pr ≔ 𝜈/𝛼 Prandtl number
Re Reynolds number
Nu Nusselt number
𝑓D Darcy friction factor
St Strouhal number
Summary
This thesis describes a numerical method for computational fluid dynamics. Special
attention is paid to lowMach number flows.
The spatial discretization is a discontinuous Galerkin method, based on modal
basis functions. The convection is discretized with the local LaxFriedrichs flux. The
diffusion in the enthalpy equation is discretized with the symmetric interior penalty
method, which is generalized in a straightforward manner for the viscous stress in
the momentum equation. The numerical method does not deviate fundamentally
from previous literature.
The temporal derivatives in the enthalpy and momentum equations are dis
cretized with a secondorder backward finite difference method. An algorithmic
pressure correction scheme is used decouple the momentum and the continuity
equations, giving rise to explicit artificial boundary conditions. If the pressure and
the momentum are discretized with an equalorder polynomial space, then the pres
sure equation is stabilized with an extra penalty term to suppress the discontinuities
in the solution, as explained in chapter 2.
Using a timesplitting method is far more difficult when the flow is compressible,
due the variable density. LowMach number flows also do not lend themselves well
to solving the coupled transport equations, because the density is a function of the
enthalpy, not the pressure. This differs from highMach number flows, where one
can solve a transport equation for the density. Chapter 4 describes in great detail
how a nonconstant density can be incorporated into a timesplitting scheme for
lowMach number flows.
Chapter 4 also discusses the best form of the enthalpy transport equation to
solve (primitive or conservative), and for which variable (primitive or conserved).
This question arises in lowMach number flows, because the density is a function of
the temperature. Here the conservative transport equation is solved for the specific
enthalpy.
The main difficulty with this approach is that the temporal enthalpy derivative
is nonlinear due to the variable density. This can be addressed with an easily
implemented adjustment of the finite difference scheme (‘method #2’ in sections
4.3–4.4). The resulting discretization displays secondorder temporal accuracy (as
measured in the spatial 𝐿2 norm) for the enthalpy and the mass flux, without having
to iterate within a time step.
Furthermore, the enthalpy transport equation needs to be stabilized with a sim
ple change of variables, in which the specific enthalpy is ‘offset’ by a constant.
Though it may be counterintuitive, the enthalpy offset is critical to the stability and
the accuracy of the temporal discretization. This would also be true if one were to
solve for the volumetric enthalpy, because the enthalpy offset determines whether
there is a onetoone mapping between the volumetric enthalpy and the density.
ix
x Summary
The spatial and temporal discretizations and their implementations are exten
sively verified and validated with the test cases at the end of the chapters. In
particular, sections 3.3.1, 3.3.2, and 4.5.1 feature exhaustive tests with manufac
tured solutions with nontrivial fluid properties. Sections 2.7, 3.4, and 4.5.2 contain
validations for laminar flows. Chapter 5 also shows simulations of turbulent flows.
Samenvatting
Deze scriptie beschrijft een numerieke methode voor vloeistofdynamica. Er gaat
extra aandacht uit naar stromingen met een laag Machgetal.
De spatiële discretisatie is een discontinue Galerkinmethode, gebaseerd op
modale basisfuncties. Voor de discretisatie van de convectie wordt gebruikt ge
maakt van de LaxFriedrichsflux. Voor de discretisatie van de diffusieve term in de
enthalpievergelijking wordt gebruikt gemaakt van de ‘symmetric interior pentaly’
methode, die in een algemenere vorm ook toepasbaar is op de viskeuze term in
the impulsvergelijking. De numerieke method wijkt niet fundamenteel af van voor
gaande literatuur.
The tijdsafgeleide in de enthalpie en impulsvergelijkingen worden discreet ge
maakt met een impliciete eindigverschilmethode van tweede orde. Een algoritmisch
drukcorrectieschema ontkoppelt de drukvergelijking en de impulsvergelijking, wat
resulteert in expliciete artificiële randvoorwaarden. Als voor de druk en de im
puls een polynome oplossingsruimte van gelijke order wordt gebruikt, dan moet de
drukvergelijking gestabiliseerd worden met een extra ‘penalty’term om de discon
tinuïteiten in de oplossing in toom te houden, zoals wordt uitgelegd in hoofdstuk
2.
In een compressibele stroming wordt de tijdsplitsingsmethode aanzienlijk be
moeilijkt door de variabele dichtheid. Voor stromingen met een laag Machgetal
is het ook niet makkelijk om de gekoppelde transportvergelijkingen op te lossen,
omdat de dichtheid een functie is van de enthalpie, niet van de druk. Hoofdsuk 4
detailleert hoe een nietconstante dichtheid opgenomen kan worden in een tijds
plitsingsschema voor stromingen met een laag Machgetal.
Hoofdstuk 5 behandelt ook in welke vorm de enthalpievergelijking het best kan
worden opgelost (primitief of conservatief) en voor welke variabele (de specifieke
of de volumetrische enthalpie). Deze vraagstukken spelen een rol bij stromingen
met een laag Machgetal, omdat de dichtheid een functie is van de temperatuur. In
dit werk wordt de conservatieve enthalpievergelijking opgelost voor de specifieke
enthalpie.
De grootste uitdaging bij deze aanpak is de nietlineaire tijdsafgeleide van de
enthalpie, als gevolg van de variabele dichtheid. Dit probleem kan worden ge
adresseerd met een eenvoudig te implementeren aanpassing van de eindigver
schilmethode (‘methode #2’ in sections 4.3–4.4). Dit resulteert in tweedeorder
nauwkeurigheid in de tijd (gemeten in de spatiële 𝐿2 norm) voor de enthalpie en
de massaflux, zonder te hoeven itereren binnen een tijdstap.
Daarnaast moet de enthalpievergelijking gestabiliseerd worden met een eenvou
dige substitutie van variabelen, waarbij een constante bij de specifieke enthalpie
wordt opgeteld. Deze ‘enthalpieverspringing’ (‘enthalpy offset’) is essentieel voor
de stabiliteit en de nauwkeurigheid van de discretisatie. Dat zou ook het geval
xi
xii Samenvatting
geweest zijn als we de volumetrische enthalpie zouden hebben gekozen als onbe
kende variabele, omdat de enthalpieverspringing bepaalt of er éénopéénverband
bestaat tussen de volumetrische enthalpie en de dichtheid.
De discretisaties in ruimte en tijd worden uitvoerig geverifieerd en gevalideerd
met de tests aan het einde van de hoofdstukken. Hierbij zijn vooral de tests met
artificiële oplossingen en niettriviale vloeistofeigenschappen noemenswaardig, zie
secties 3.3.1, 3.3.2 and 4.5.1. Secties 2.7, 3.4 and 4.5.2 behandelen validaties met
laminaire stromingen. Hoofdstuk 5 toont ook simulaties met turbulente stromingen.
1
Introduction
This thesis is on computational fluid dynamics for incompressible flows and flows
in the lowMach number limit. Various numerical issues are explored, related to the
spatial discretization, the timesplitting method, and turbulent flows. Special atten
tion is paid to how the discretizations can be adjusted when the density depends
on the temperature.
These are fairly unrelated topics, and it would be infeasible to cover all the basics
in a single chapter. The following chapters therefore contain their own introduc
tions, including extensive literature reviews. Here we merely state the mathematical
problem (i.e., the governing equations with boundary conditions).
It is assumed throughout the thesis that the reader is familiar with common
concepts in computational fluid dynamics.
This text is also not meant as an introduction to the discontinuous Galerkin (DG)
method, though section 1.2 discusses some reasons for its recent gain in popularity,
and can be read without prior knowledge. Chapter 2 (‘Spatial Discretization with a
Discontinuous Galerkin Method’ ) is technically selfcontained, but focuses on what
differs from other literature, while skipping over some important basics. The reader
can instead consult one of several monographs, of which the one by Hesthaven and
Warburton [1] likely provides the gentlest introduction.
1.1. Mathematical Problem Setting

The lowmach number limit of the compressible flow equations can be obtained in a
conceptually straightforward manner by expanding each variable 𝜙 into a Maclaurin
2
series in the Mach number Ma, that is, 𝜙 = 𝜙(0) + 𝜙(1) Ma + 𝜙(2) Ma + …. Since the
equations must hold for all Ma, this leads to a series of equations, corresponding to
0 1 2 3
the coefficients of Ma , Ma , Ma , …. The terms of order 𝒪 (Ma ) are neglected.
A detailed derivation can be found in several places, including the seminal paper on
this approach by Paolucci [2].
1
2 1. Introduction
In the resulting transport equations, the pressure is split into two parts: the
1 0
thermodynamic pressure 𝑝th (of order Ma ), which is used to evaluate the equation
2
of state, and a pressure 𝑝 (of order Ma ) in the momentum equation that acts
as a Lagrange multiplier to enforce continuity, just like in the incompressible flow
equations [2]. We simply refer to 𝑝 as the pressure; it has also been called the
‘hydrodynamic pressure’1 (e.g, [3, 4]), the ‘mechanical pressure’ (e.g., [5]), and
the ‘incompressible pressure’ (e.g., [6]).
Assuming a constant thermodynamic pressure, the transport equations in the
lowMach number limit are
𝜕𝜌
+∇⋅𝐦=0 , (1.1a)
𝜕𝑡
𝜕𝐦
+ ∇ ⋅ (𝐮 𝐦) = ∇ ⋅ 𝜏 − ∇𝑝 + 𝐅 , (1.1b)
𝜕𝑡
𝜕𝜌ℎ
+ ∇ ⋅ (𝐦ℎ) = −∇ ⋅ 𝐪 + 𝑄 (1.1c)
𝜕𝑡
on a domain Ω. Here 𝑡 is the time, 𝜌 is the density, 𝐮 is the velocity, 𝐦 ≔ 𝜌𝐮 is the
mass flux, ℎ is the specific enthalpy, and 𝐅 and 𝑄 are known external sources. The
pressure 𝑝 has no effect on any of the fluid properties.
The general form of the lowMach number equations has a nonconstant ther
modynamic pressure 𝑝th = 𝑝th (𝑡) and a density that is a function of both 𝑝th and
ℎ. The enthalpy equation then gets an extra term d𝑝th /d𝑡, which can be estimated
by integrating the enthalpy equation. Since all variabledensity flows in this work
have an outlet, where 𝑝th is fixed, we simply have d𝑝th /d𝑡 = 0.
Assuming a Newtonian fluid, the stress tensor is
2
𝜏𝑖𝑗 = 𝜇 (∇𝑖 𝑢𝑗 + ∇𝑗 𝑢𝑖 − (∇ ⋅ 𝐮) 𝛿𝑖𝑗 ) . (1.2)
3
The heat flux is

𝑘
𝐪 = −𝑘∇𝑇 = − ∇ℎ , (1.3)
𝑐𝑝
where 𝑇 is the temperature. 𝑘 is the thermal conductivity, and 𝑐𝑝 ≔ 𝜕ℎ/𝜕𝑇 is the
specific heat capacity. The last equality is technically an approximation because
it neglects the dependence of the temperature on the pressure, but this is highly
accurate for lowMach number flows, even for strongly variable fluid properties in
supercritical fluids [3]. The transport properties (𝜇, 𝑘) and the specific heat capacity
(𝑐𝑝 ) are a function of 𝑇, but do not depend significantly on the pressure.
We consider two types of boundaries:
• Dirichlet boundaries, denoted by 𝜕ΩD , on which the mass flux and the tem
perature (and, therefore, the enthalpy and the fluid properties) are given, that
is, 𝐦 = 𝐦D and 𝑇 = 𝑇D ;
1 This
is perhaps slightly confusing because of the double meaning of the ‘dynamic pressure’ in the
Bernoulli equations.
1.2. Overview of the Numerical Method 3
• outflow boundaries, denoted by 𝜕ΩN , where

1
(𝜏 − 𝑝𝐼) ⋅ 𝐧 = 𝐟N and 𝑘∇𝑇 ⋅ 𝐧 = 𝑞 N (1.4)
are prescribed.
Periodic boundary conditions do not require special attention, because correspond
ing mesh elements at opposite periodic boundaries can simply be treated as internal
elements. Of course the mass flux and the temperature must also be equipped with
appropriate initial conditions, but these will not play a significant role.
1.2. Overview of the Numerical Method

The goal of a numerical method is to find an approximate solution to the system
1.1a–1.1c. Since computers work efficiently with blocks of data, it is natural to
represent the state of a fluid by a list of numbers, which form a vector 𝝓 ∈ ℝ𝑁 .
This is known as a discretization, because it transforms the continuous problem into
a problem with a countable number (𝑁) of degrees of freedom. In this thesis the
terms ‘numerical method’ and ‘discretization’ are used interchangeably.
The type of discretization is determined by the way in which 𝝓 represents an
unknown continuous state. For the spatial discretization we use a finite element
𝑁
method, which is based on 𝑁 predefined basis functions {𝜉𝑖 }𝑖=1 . A generic exact
quantity 𝜙̃ is approximated by the linear combination 𝜙̃ ≈ 𝜙 = 𝝓 𝜉𝑘 . In other
𝑘
words, the coefficients of the solution vector 𝝓 are the weights of the basis func
tions.
An effective way to use the computing power is to solve a linear system
𝑨𝝓 = 𝒃 (1.5)
for 𝝓, which is therefore known as the solution vector. In practice such a large
linear system can only be solved approximately, and there is an enormous body of
research on how this can be done. This typically has a large impact on the overall
efficacy of the numerical method, but that is not the topic of this thesis. Section
2.6 outlines how we solve the linear equations, and we make occasional reference
to the linear solvers when they are relevant to the rest of the discretization (such
as in section 3.5.1), but the reader can assume that the above equation is solved
with a negligible error.
That leaves the question of how to construct the matrix 𝑨 and righthand side
𝒃. In a finite element method this is done with discrete bilinear and linear operators
𝑎(⋅, ⋅) and 𝑏(⋅) that generate the entries in 𝑨 and 𝒃 with the basis functions, that is,
𝑨𝑖𝑗 = 𝑎(𝜉𝑖 , 𝜉𝑗 ) and 𝒃𝑖 = 𝑏(𝜉𝑖 ) . (1.6)

The linear system is equivalent to the problem
Find 𝝓 ∈ ℝ𝑁 , such that, for all 𝒗 ∈ ℝ𝑁 ,
(1.7)
𝒗⊺ 𝑨𝝓 = 𝒗⊺ 𝒃 .
4 1. Introduction
The test vector 𝒗 has a corresponding test function 𝑣 = 𝒗𝑘 𝜉𝑘 . Using the linearity
1
of 𝑎 and 𝑏, we have 𝒗⊺ 𝑨𝝓 = 𝒗𝑘 𝑎 (𝜉𝑘 , 𝜉𝑞 ) 𝝓 = 𝑎 (𝒗𝑘 𝜉𝑘 , 𝝓 𝜉𝑞 ) = 𝑎 (𝑣, 𝜙) and
𝑞 𝑞
𝒗⊺ 𝒃 = 𝒗𝑘 𝑏 (𝜉𝑘 ) = 𝑏 (𝒗𝑘 𝜉𝑘 ) = 𝑏 (𝑣), so the linear system can also be written as
Find 𝜙 ∈ 𝑉𝜙 , such that, for all 𝑣 ∈ 𝑉𝜙 ,

(1.8)
𝑎(𝜙, 𝑣) = 𝑏(𝑣) ,
𝑁
where 𝑉𝜙 = span ({𝜉𝑖 }𝑖=1 ) is the numerical solution space (i.e., the span of all basis
functions). Instead of deriving the linear system directly, we work with statements
like Eq. 1.8, which is known as the discrete weak form.
The precise type of finite element method we use is a discontinuous Galerkin
(DG) method, characterized by the fact that the basis functions are not continuous,
and neither is the numerical approximation. The term ‘Galerkin’ refers to the fact
that the test function in Eq. 1.8 lies in the same solution space as the numerical
solution. As a result of the discontinuities, the DG method shares many characteris
tics with the finite volume method, especially when applied to hyperbolic problems,
as will become clear in Chapter 2.
In practice we have to decouple the spatial and temporal discretizations to make
the computations feasible. For the temporal discretization we use a finite difference
𝑛 𝑛+1
method, resulting in solution vectors {𝝓 , 𝝓 , …} to approximate the solution at
discrete times {𝑡𝑛 , 𝑡𝑛+1 , …}. We also split the full coupled state of the fluid into three
separate solution vectors for the mass flux, pressure, and enthalpy (𝒎, 𝒑, and 𝒉),
as is common for incompressible flows.
1.2.1. Why Use a Discontinuous Galerkin Method?

The discontinuous Galerkin method initially received little attention outside of the
field of particle transport, where it was introduced in 1973 [7]. This has changed
over the last one or two decades, and the DG method is now a very active area of
research with applications in many fields, including computational fluid dynamics.
Each author seems to prefer the DG method for his or her own reasons; the liter
ature does not agree on the most important advantage over other discretizations.
We list a few possible advantages here.
Early work by Collis [8] focused on the weak imposition of Dirichlet boundary
conditions in an attempt to explain their remarkably accurate results with few de
grees of freedom. Ern and Guermond [9] provide a more theoretical analysis of
weak Dirichlet boundary conditions for Friedrich’s system. Weak boundary condi
tions have long been known to be superior to strongly imposed boundary condi
tions, where the numerical solution satisfies the boundary condition at every point
on the boundary, as had been common in classical finite element methods. A weak
boundary condition can be shown to act as an implicit filter in badly resolved flows
[10].
Weak Dirichlet boundary conditions with a userdefined penalty parameter also
provided the inspiration for the first penalty methods to discretize elliptic problems
1.2. Overview of the Numerical Method 5
with a discontinuous solution space. The unified analysis of DG methods for elliptic
problems by Arnold et al. [11] was foundational for extending the DG method to 1
viscous flow. The required penalty parameter has been estimated by many authors,
including Shahbazi [12]. They later also established the DG method for convection
diffusion problems for incompressible flow with a timesplitting method [13].
An interesting property of the DG solution space is that it supports a lowpass
filter in a natural way, by separating the highorder and loworder polynomials in the
solution space, giving it great potential for a variational multiscale method. Collis
[14] has discussed this as early as 2002. See Hughes et al. [15] for a comparison of
continuous and discontinuous solution spaces in this context. Variational multiscale
with DG is still an ongoing area of research (e.g., [16, 17]).
There are other applications for the natural support for scale separation. For ex
ample, Atkins and Helenbrook [18] and Helenbrook and Atkins [19] have introduced
polynomialbased multigrid methods to DG discretizations in 2005. Separating the
polynomials can also be used for the coarsegrid projection in a dynamic Large Eddy
simulation [20].
Most authors recognize the arbitrarily high order of accuracy on unstructured
meshes as a major benefit of DG discretizations. Several papers have applied the
DG method to nontrivial geometries, such as airfoils. This geometric flexibility could
be valuable for some industrial applications (as discussed in, e.g., [21]). It also
provides support for adaptivity (e.g., [22]). Furthermore, the unstructured nature of
DG methods has been used for arbitrary spacetime domains and moving boundary
problems (e.g., [23–26]).
The most important numerical characteristic of any discretization is perhaps how
well it handles unresolved flows. Turbulent flow simulations are rarely fully resolved,
and that is particularly true for large eddy simulations (LES), where it is standard
practice to rely on a moderate amount of numerical dissipation. This is likely to
increase in importance, as more practical flows are starting to fall within the range
of applications of LES.
There are indications that the DG method handles unresolved flows well. One
reason is that the weak form in a DG method can be written as a local conservation
law for each element in the mesh, much like in finite volume methods. This offers
stability for badly resolved flows, despite the high order of accuracy. Another reason
is the frequency spectrum of the numerical dissipation, which many authors have
studied recently (e.g., [27]). Highorder DG discretizations acts as a lowpass filter.
Moura et al. [28] have argued that this makes them particularly suitable for implicit
large eddy simulations.
More recently, there has been much attention to how the computational struc
ture of DG discretizations can be implemented efficiently on modern hardware archi
tectures. The DG method results in a block matrix, wherein the blocks are coupled
with a compact stencil. Chapelier et al. [17] have claimed that this is beneficial for
MPI parallelization. The block structure also lends itself to acceleration with GPUs
(e.g., [29]). As mentioned by Kronbichler et al. [30], ‘the most common obstacle to
highorder methods is a lack of competitive implementations in generic numerical
software packages, which has limited their application mostly to specialized codes
6 References
targeting’. This makes it difficult to compare the efficacy of DG methods to the

1 more traditional finite volume methods for industrial flow problems.
1.3. Thesis Overview

The rest of the thesis is structured as follows.
• Chapter 2 first defines the basis functions, and then derives discrete bilinear
and linear operators (𝑎 and 𝑏 in Eq. 1.8) from the governing equations and
the boundary conditions.
• Chapter 3 treats the temporal discretization, which is based on a pressure
correction method.
• Chapter 4 discusses several issues with the temporal discretization of variable
density flows.
• Chapter 5 contains simulations of turbulent plane channel flow, some of which
were performed with an LES model.
There is little relation between the chapters, and each can be read independently.
References
[1] J. S. Hesthaven and T. Warburton, Nodal Discontinuous Galerkin Methods
(Springer, New York, 2008).
[2] S. Paolucci, On the Filtering of Sound from the NavierStokes Equations, Tech.
Rep. SAND828257 (Sandia National Laboratories, Ananlytical Thermal/Fluid
Mechanics Division, Livermore, 1982).
[3] J. W. R. Peeters, Turbulence and turbulent heat transfer at supercritical pres
sure, doctoral thesis, Delft University of Technology (2016).
[4] F. C. Nicoud, Numerical study of a channel flow with variable properties, in
Annual Research Briefs (Center for Turbulence Research, 1998) pp. 289–310.
[5] M. Hassanaly, H. Koo, C. F. Lietz, S. T. Chong, and V. Raman, A minimally
dissipative lowMach number solver for complex reacting flows in OpenFOAM,
Computers & Fluids 162, 11 (2018).
[6] B. Müller, LowMachNumber Asymptotics of the NavierStokes Equations,
Journal of Engineering Mathematics 34, 97 (1998).
[7] W. H. Reed and T. R. Hill, Triangular mesh methods for the neutron trans
port equation, in Proceedings of the American Nuclear Society (University of
California, Los Alamos Scientific Laboratory, Los Alamos, New Mexico 87544,
1973).
[8] S. S. Collis, Discontinuous Galerkin methods for turbulence simulation, in Sum
mer Program 2002, Center for Turbulence Research (2002) pp. 155–167.
References 7
[9] A. Ern and J.L. Guermond, Discontinuous Galerkin Methods for Friedrichs’
Systems, in Numerical Mathematics and Advanced Applications, edited by A. B. 1
de Castro, D. Gómez, P. Quintela, and P. Salgado (Springer Berlin Heidelberg,
Berlin, Heidelberg, 2006) pp. 79–96.
[10] Y. Bazilevs and T. Hughes, Weak imposition of Dirichlet boundary conditions

in fluid mechanics, Computers & Fluids 36, 12 (2007).
[11] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini, Unified analysis of dis
continuous Galerkin methods for elliptic problems, SIAM Journal on Numerical
Analysis 39, 1749 (2002).
[12] K. Shahbazi, An explicit expression for the penalty parameter of the interior
penalty method, Journal of Computational Physics 205, 401 (2005).
[13] K. Shahbazi, P. F. Fischer, and C. R. Ethier, A highorder discontinuous Galerkin
method for the unsteady incompressible NavierStokes equations, Journal of
Computational Physics 222, 391 (2007).
[14] S. S. Collis, The DG/VMS Method for Unified Turbulence Simulation, in 32nd
AIAA Fluid Dynamics Conference and Exhibit (Houston, Texas, 2002).
[15] T. J. R. Hughes, G. Scovazzi, P. B. Bochev, and A. Buffa, A multiscale dis
continuous Galerkin method with the computational structure of a continuous
Galerkin method, Computer Methods in Applied Mechanics and Engineering
195, 2761 (2006).
[16] M. Drosson, Development of a highorder interior penalty discontinuous

Galerkin method for compressible turbulent flows, Phd thesis, University of
Liège (2013).
[17] J.B. Chapelier, M. De La Llave Plata, F. Renac, and E. Lamballais, DNS of

Canonical Turbulent Flows Using the Modal Discontinuous Galerkin Method, in
Direct and LargeEddy Simulation IX, edited by J. Fröhlich, H. Kuerten, B. J.
Geurts, and V. Armenio (Springer International Publishing, Dresden, 2015)
pp. 91–96.
[18] H. L. Atkins and B. T. Helenbrook, Numerical Evaluation of PMultigrid Method

for the Solution of Discontinuous Galerkin Discretizations of Diffusive Equa
tions, in 17th AIAA Computational Fluid Dynamics Conference (Toronto, On
tario, 2005).
[19] B. Helenbrook and H. Atkins, Application of PMultigrid to Discontinuous

Galerkin Formulations of the Poisson Equation, in 17th AIAA Computational
Fluid Dynamics Conference (Toronto, Ontario, 2005).
[20] A. Abbà, L. Bonaventura, M. Nini, and M. Restelli, Dynamic models for Large
Eddy Simulation of compressible flows with a high order DG method, Comput
ers & Fluids 122, 209 (2015), arXiv:1407.6591 .
8 References
[21] K. Hillewaert, Development of the discontinuous Galerkin method for high

1 resolution, large scale CFD and acoustics in industrial geometries, Phd thesis,
Université catholique de Louvain (2013).
[22] K. A. Cliffe, E. J. C. Hall, and P. Houston, Adaptive Discontinuous Galerkin

Methods for Eigenvalue Problems Arising in Incompressible Fluid Flows, SIAM
Journal on Scientific Computing 31, 4607 (2010).
[23] J. J. W. Van der Vegt and H. Van der Ven, SpaceTime Discontinuous Galerkin
Finite Element Method with Dynamic Grid Motion for Inviscid Compressible
Flows: I. General Formulation, Journal of Computational Physics 182, 546
(2002).
[24] V. Ambati and O. Bokhove, Spacetime discontinuous galerkin discretization
of rotating shallow water equations, Journal of Computational Physics 225,
1233 (2007).
[25] S. Rhebergen, B. Cockburn, and J. J. W. van der Vegt, A spacetime dis

continuous Galerkin method for the incompressible NavierStokes equations,
Journal of Computational Physics 233, 339 (2013).
[26] M. Tavelli and M. Dumbser, A pressurebased semiimplicit spacetime discon
tinuous Galerkin method on staggered unstructured meshes for the solution
of the compressible NavierStokes equations at all Mach numbers, Journal of
[27] M. Alhawwary and Z. J. Wang, Fourier analysis and evaluation of DG, FD and
compact difference methods for conservation laws, Journal of Computational
Physics 373, 835 (2018).
[28] R. Moura, S. Sherwin, and J. Peiró, Linear dispersiondiffusion analysis and

its application to underresolved turbulence simulations using discontinuous
Galerkin spectral/hp methods, Journal of Computational Physics 298, 695
(2015).
[29] Y. Xia, J. Lou, H. Luo, J. Edwards, and F. Mueller, OpenACC acceleration of
an unstructured CFD solver based on a reconstructed discontinuous Galerkin
method for compressible flows, International Journal for Numerical Methods
in Fluids 78, 123 (2015).
[30] M. Kronbichler, B. Krank, N. Fehn, S. Legat, and W. A. Wall, A New HighOrder
Discontinuous Galerkin Solver for DNS and LES of Turbulent Incompressible
Flow, in New Results in Numerical and Experimental Fluid Mechanics XI, edited
by A. Dillmann, G. Heller, E. Krämer, C. Wagner, S. Bansmer, R. Radespiel, and
R. Semaan (Springer International Publishing, Cham, 2018) pp. 467–477.
Spatial Discretization with a
2
Discontinuous Galerkin
Method
2.1. Introduction
This chapter details the spatial discretization, which is based on the discontinuous
Galerkin (DG) finite element method. To simplify the discussion, we consider the
stationary transport equations
∇⋅𝐦=0 , (2.1a)
∇ ⋅ (𝐮 𝐦) = ∇ ⋅ 𝜏 − ∇𝑝 + 𝐅 , (2.1b)
𝑘
∇ ⋅ (𝐦 ℎ) = ∇ ⋅ ( ∇ℎ) + 𝑄 , (2.1c)
𝑐𝑝
and postpone including the time variable to chapters 3 and 4. The temporal dis
cretization that is discussed there does not depend on the spatial discretization, and
understanding this chapter is not essential to understanding the rest of the thesis.
2.1.1. Overview of DG discretizations

The discontinuous Galerkin method can be thought of as a highorder finite volume
method, replacing the constant solution within an element with an arbitrarily rich
solution space that is defined everywhere on the interior of the element. The result
ing numerical solution is discontinuous at the element boundaries. Finite volume
and discontinuous Galerkin methods have in common that they discretize the weak
form, rather than the original partial differential equation. As a result, numerical
Parts of this chapter have been published in [1].
9
10 2. Spatial Discretization with a Discontinuous Galerkin Method
fluxes that were developed for onedimensional finite volume schemes carry over
naturally to the DG method, making it particularly suitable for hyperbolic systems
(see, e.g., the review [2], and a unified analysis of finite volume and DG discretiza
tions in [3]).
DG methods for the diffusion operator have matured more recently. The dif
2 fusion equation has an ‘instant smoothing’ property, meaning that it has a differ
entiable solution everywhere in the domain, even when there are discontinuities
in the initial condition, the boundary conditions, the forcing term, or the diffusion
parameter. Unsurprisingly, a discontinuous solution space is not the most natural
choice for these problems. Nevertheless, several DG approaches have emerged.
We mention two of them here. See the standard reference [4] for a clear and
thorough analysis.
The first approach is to split the Laplace operator into two equations with first
order derivatives, resulting in two discretizations, which can then be merged. This
socalled ‘local DG’ (LDG) method is straightforward, but it leads to a large stencil,
where each element is coupled to the neighbors of its neighbors.
We avoid this by using another approach, namely the symmetric interior penalty
(SIP) method. Interior penalty methods have a local stencil, where each element is
only coupled to it direct neighbors. One could take the following heuristic viewpoint
of this method. Think of each element as a separate domain in which a spectral
method is used to solve the temperature diffusion equation. The elements are
coupled by providing each other’s boundary conditions, of which we need two at
each interior face: one for each neighbor. One of these boundary conditions is
natural to a finite element framework: the heat flux must be the same at both
sides of the face. The other boundary condition is that the temperature must be
continuous, but enforcing this strongly would break the discontinuous nature of the
solution space. Therefore continuity is enforced weakly by penalizing the jumps of
the solution at the faces.
This gives rise to a userdefined penalty parameter that must be ‘large enough’
in order for the discrete bilinear form to be coercive, which is a sufficient condition
for stability. This has been seen as a disadvantage, but several authors (e.g., [5,
6]) have found useful estimates, and this has been extended to various types of
elements and unstructured meshes [7].
2.1.2. Construction of the Solution Space

To construct the finite element solution space, the domain Ω with outward normal 𝑛
is meshed into a set of nonoverlapping elements 𝒯. The boundaries of an element
are called faces. Internal faces each have two neighbors; boundary faces have only
one.
We adopt the following notation from di Pietro and Ern [8]. The set of internal
faces is denoted by ℱi . The set of faces that lie on the Dirichlet (resp. Neumann)
boundary of the domain are denoted by ℱD (resp. ℱN ). Sets of multiple types of
faces are sometimes abbreviated with the obvious notation ℱD,i ≔ ℱD ⋃ ℱi . The
set of faces of an element 𝑇 ∈ 𝒯 is ℱ𝑇 . Similarly, 𝒯𝐹 is the set of neighbors of face
𝐹. Each internal face 𝐹 has a normal vector 𝐧𝐹 that points in an arbitrary but fixed
2.1. Introduction 11
direction. On boundary faces, 𝐧𝐹 coincides with the outward normal of Ω.

The solution space is spanned by the basis functions, each of which has support
on one element, meaning that it is zero elsewhere. The numerical solution is piece
wise continuous: it is continuous within each element, but discontinuous across
the faces. These discontinuities are described by the jump and average operators,
which are defined on internal faces as 2
1 −
J𝑥K ≔ 𝑥 + − 𝑥 − and {𝑥} ≔ (𝑥 + 𝑥 + ) , (2.2)
2
where
𝑥 ± (𝐫 ∈ 𝐹) ≔ lim 𝑥(𝐫 ∓ 𝜖 𝐧𝐹 ) (2.3)
𝜖↓0
indicate the values at the two sides. On vectors they act elementwise. On boundary
elements, both the jump and the average are defined as the internal numerical value
(i.e., the ‘trace’). The reason for this convention at the boundary is that it allows
for a concise notation of the elementwise application of the divergence theorem:
for an arbitrary piecewise continuous vector 𝐯,
∑ ∫ ∇ ⋅ 𝐯 = ∑ ∫ 𝐯 ⋅ 𝐧𝑇 = ∑ ∫ J𝐯K ⋅ 𝐧𝐹 , (2.4)
𝑇 𝜕𝑇 𝐹
𝑇∈𝒯 𝑇∈𝒯 𝐹∈ℱD,N,i
where 𝐧𝑇 is the outward normal of element 𝑇.
Modal vs. Nodal Basis Functions

Though arbitrary types of basis functions are theoretically possible, discontinuous
Galerkin methods almost always use polynomials. An exception is wall function
enrichment, in which the solution space of wallbounded elements is enriched with
functions that contain some physically motivated, a priori information on the flow
near the wall flow (e.g., [9, 10]). The polynomials can be either modal or nodal.
Nodal functions are defined through nodes, which are coordinates such that the
𝑖 th basis function equals 𝛿𝑖𝑗 on the 𝑗th node, that is, the function is nonzero on
exactly one of the nodes. This has a computational advantage when evaluating the
integrals in the discrete weak forms: the nodes can be placed on the quadrature
points, making the numerical quadrature a sparse sum. Many recent DG implemen
tations use this approach to speed up the matrix assembly. The book by Hesthaven
and Warburton [11] is a good introduction to this topic, and to the DG method in
general.
Nodal functions are often defined such that they span a tensor product of one
dimensional polynomial spaces. For example, a secondorder polynomial approx
imation on a twodimensional element would contain all functions in {1, 𝑥, 𝑥 2 } ×
{1, 𝑦, 𝑦 2 } = {1, 𝑦, 𝑦 2 , 𝑥, 𝑥𝑦, 𝑥𝑦 2 , 𝑥 2 , 𝑥 2 𝑦, 𝑥 2 𝑦 2 }. In this case a polynomial order 𝒫
actually means that all polynomials of order less than or equal to 𝒫 lie in the so
𝑑
lution space, but it also contains some higherorder functions. There are (𝒫 + 1)
basis functions in each 𝑑dimensional element. Note that the basis depends on the
orientation of the coordinate axes.
We use a modal basis, in which the basis functions are constructed such that they
are hierarchical and orthogonal in the 𝐿2 norm. This is just an implementation issue;
all equations in this thesis are valid for nonorthogonal bases. The solution space
within each element is simply spanned by all polynomials up to an order 𝒫. There
are ( 𝒫 ) = (𝒫 + 𝑑)!/(𝑑! 𝒫!) linearly independent polynomials in a 𝑑dimensional
2 𝒫+𝑑
element. In this chapter the polynomial order is the same on all elements, though
this is not a requirement of the numerical method. The order of the polynomials
for the unknown 𝑋 is denoted by 𝒫𝑋 .
2.2. Discrete Continuity Equation

The DG bilinear form for the divergence operator can be found in many previous
works, including [12], [13, p. 92], and [8, pp. 250–252], and is given by
𝑎div (𝐯, 𝑞) = − ∑ ∫ 𝑞∇ ⋅ 𝐯 + ∑ ∫ {𝑞} J𝐯K ⋅ 𝐧𝐹 , (2.5)

𝑇 𝐹
𝑇∈𝒯 𝐹∈ℱD,i
so that a consistent discrete weak form of the continuity equation with 𝜕𝜌/𝜕𝑡 = 0
(Eq. 2.1a) is
Find 𝑝 ∈ 𝑉𝑝 , such that, for all 𝑞 ∈ 𝑉𝑝 ,
(2.6)
− 𝑎div (𝐦, 𝑞) = − ∫ 𝑞 𝐦D ⋅ 𝐧 ,
𝜕ΩD
where 𝑉𝑝 is the solution space of the pressure, and 𝐦 is the numerical (discontinu
ous) mass flux.
The weak form of the continuity equation can only be stable if the divergence
operator is surjective. This means that, for every 𝑞 in the pressure space, there
exists a 𝐯 in the velocity (or mass flux) space, such that 𝑞 is the divergence of 𝐯
(and satisfying a constraint on the norm of 𝐯; see [8, pp. 246–252] for a precise
analysis). Surjectivity can be shown to be equivalent to an infsup condition on the
bilinear form of the divergence operator.
The continuous divergence operator is surjective, but the discrete divergence
operator 𝑎div is not for equalorder discretizations of the pressure and the mass
flux (i.e., 𝒫𝑚 = 𝒫𝑝 ). This makes intuitive sense, since the divergence of the mass
flux would lie in a lowerorder polynomial space than the pressure. A solution is
to set 𝒫𝑚 > 𝒫𝑝 , in which case infsup stability has been proven [14]. The lack
of infsup stability is a general aspect of equalorder finite element methods for
incompressible flows. See, for example, John [15] for an extensive discussion in
the context of continuous finite elements.
This is unfortunate, because equalorder methods have often been found more
efficient than mixedorder methods, and therefore methods have been devised to
stabilize the pressure for equalorder DG discretizations. Cockburn et al. [16] as
sumed a homogeneous kinematic viscosity and added a pressure stabilization term,
2.2. Discrete Continuity Equation 13
so that Eq. 2.6 becomes
Find 𝑝 ∈ 𝑉𝑝 , such that, for all 𝑞 ∈ 𝑉𝑝 ,

(2.7)
− 𝑎div (𝐦, 𝑞) + 𝑎stab (𝑝, 𝑞) = − ∫ 𝑞 𝐦D ⋅ 𝐧 ,
𝜕ΩD
2
where
𝑎stab (𝑝, 𝑞) = ∑ ∫ 𝜁𝐹 J𝑝K J𝑞K (2.8)
𝐹
𝐹∈ℱ𝑖
for equalorder discretizations, and 𝑎stab (𝑝, 𝑞) = 0 for mixedorder discretizations.

The penalty parameter for the pressure discontinuities is 𝜁𝐹 = 𝛾0 ‖𝐹‖leb /𝜈, where
‖⋅‖leb is the Lebesgue measure (which is the length, area, or volume in 1, 2, or 3
dimensions). We adjust the above penalty term to a variable viscosity in the obvious
way: by taking the pointwise maximum value of 𝜁𝐹 on both sides of the face, that
is,
1
𝜁𝐹 = 𝛾0 ‖𝐹‖leb max [ ] . (2.9)
𝑇∈𝒯𝐹 𝜈
𝑇
We set 𝛾0 = 1 without investigating other values. We offer no proof for the validity
of this handling of the variable viscosity, but extensive tests in this thesis (especially
those with manufactured solutions in sections 3.3.2 and 4.5.1) will show that the
discretization is stable.
Another approach to stabilizing equalorder DG methods was taken by Botti
and Di Pietro [17], who used continuous finite elements for the pressure, simply
disposing of the pressure discontinuities altogether. This seems logical, since the
transport equations for an incompressible flow imply a Poisson equation for the pres
sure, and the continuous Galerkin (CG) method is the most effective discretization
for purely diffusive problems, whereas the DG method mostly thrives for hyperbolic
problems. (Though it has been shown that DG and CG methods are asymptotically
equally efficient in the limit of (very) high polynomial orders of approximation [18].)
A possible disadvantage would be the reduced mesh generality compared to a pure
DG discretization. We have not pursued this approach, partly because it would be
demanding to implement.
Krank et al. [19] have instead focused on the momentum equation to stabilize
equalorder DG discretizations. They suppressed the local violation of the continu
ity equation by modifying the discrete NavierStokes equation with elementwise
penalty terms for ∇ ⋅ 𝐮 within an element, and for the jump of 𝐧 ⋅ 𝐦 across the
faces. This assumes a divergencefree velocity field, which is generally not valid for
lowMach number flow. We will use this approach in chapter 5.
These ideas for pressure stabilization can be compared to the artificial compress
ibility method, in which the continuity equation is perturbed with a compressibility
term (1/𝑐 2 )𝜕𝑝/𝜕𝑡 for some parameter 𝑐 > 0. The numerical fluxes in the DG weak
form are then obtained by solving a Riemann problem for the discontinuities at a
face. This also gives rise to penalty terms for the jump of the pressure (as in Eq.
2.8), and the jump of 𝐧 ⋅ 𝐦 across a face (ref., e.g, [20]).
2.3. Discrete Momentum Equation

The discrete weak form of the stationary momentum equation (Eq. 2.1b) is
Find 𝐦 ∈ 𝑉𝑚 , such that, for all 𝐯 ∈ 𝑉𝑚 ,
2 𝑎conv (𝐮, 𝑚𝑘 , 𝑣𝑘 ) + 𝑎visc (𝐦, 𝐯) = 𝑙 conv (𝐮, 𝑚D
𝑘 , 𝑣𝑘 ) + 𝑙
visc
(𝐯) − 𝑎div (𝐯, 𝑝) + ∫ 𝐅 ⋅ 𝐯 ,
Ω
(2.10)
where 𝑉𝑚 is the solution space of the mass flux, and 𝑝 is the numerical (discontinu
ous) pressure. Note that the divergence operator 𝑎div doubles as a gradient oper
ator. Integrating Eq. 2.5 by parts, and using the fact that J𝑞𝐯K = J𝑞K {𝐯} + {𝑞} J𝐯K
on an interior face, gives
𝑎div (𝐯, 𝑞) = ∑ ∫ 𝐯 ⋅ ∇𝑞 − ∑ ∫ J𝑞K {𝐯} ⋅ 𝐧𝐹 , (2.11)

𝑇 𝐹
𝑇∈𝒯 𝐹∈ℱN,i
demonstrating consistency, since the last term is zero when the continuous pressure
is substituted for 𝑞.
The discretization of the convection 𝑎conv and 𝑙 conv will be given in section 2.5.
Solving for the mass flux 𝐦 instead of the velocity 𝐮 complicates the treatment of
the viscous term, which is linear in ∇𝐮, not ∇𝐦. Section 2.3.1 details how this can
be handled with a DG method.
2.3.1. Discretization of the Viscous Stress

To derive a discretization for the viscous term, rewrite the viscous stress in terms
of the mass flux as 𝜏 = 𝐿visc (𝑚), where
𝜇 2
𝐿visc
𝑖𝑗 (𝐦) = (𝐵 + 𝐵𝑗𝑖 − 𝐵𝑘𝑘 𝛿𝑖𝑗 ) , (2.12)
𝜌 𝑖𝑗 3
is a linear operator, with 𝐵𝑖𝑗 = 𝜌∇𝑖 𝑢𝑗 = ∇𝑖 𝑚𝑗 − 𝑑𝑖 𝑚𝑗 , and

1
𝐝≔ ∇𝜌 . (2.13)
𝜌
We use a generalization of the symmetric interior penalty (SIP) method, given by
the discrete bilinear operator
𝑎visc (𝐰, 𝐯) = ∑ ∫ 𝐿visc

𝑘𝑙 (𝐰) ∇𝑘 𝑣𝑙 + ∑ ∫ 𝜂𝐹 J𝐰K ⋅ J𝐯K
𝑇 𝐹
𝑇∈𝒯 𝐹∈ℱi ∪ℱD
(2.14)
− ∑ ∫ (J𝐯K ⋅ {𝐿visc (𝐰)} + J𝐰K ⋅ {𝐿visc (𝐯)}) ⋅ 𝐧𝐹
𝐹
𝐹∈ℱi ∪ℱD
and the linear operator
𝑙 visc (𝐯) = ∑ ∫ (𝜂𝐹 𝐦D ⋅ 𝐯 − 𝐦D ⋅ 𝐿visc (𝐯) ⋅ 𝐧𝐹 ) + ∫ 𝐟N ⋅ 𝐯 . (2.15)

𝐹 𝜕ΩN
𝐹∈ℱD
2.3. Discrete Momentum Equation 15
This reduces to the regular SIP method when substituting (𝜇/𝜌)∇𝑖 𝑤𝑗 for 𝐿visc 𝑖𝑗 (𝑤).
Compared to other interior penalty methods and the local DG method, the advan
tages of the SIP method are the optimal convergence rate for all polynomial orders,
its adjoint consistency, and its compact stencil [21].
If the viscosity is constant and the flow is incompressible (i.e., ∇⋅𝐮 = 0), then ∇⋅𝜏
can be simplified to ∇ ⋅ (𝜇∇𝐮) on the continuous level. But note that using a regular
2
SIP method for ∇ ⋅ (𝜇∇𝐮) would always be different from the above discretization
of ∇ ⋅ 𝜏. As one would expect, our numerical tests (not shown in this thesis) show
a negligible numerical difference between these discretizations when 𝜇 is constant
and ∇ ⋅ 𝐮 = 0.
The above discretization of the viscous term can be compared to what is usually
done for compressible flows, where the system of equations 1.1 is solved for a full
state vector 𝑈 ≔ [𝜌, 𝐦, 𝜌ℎ]. In that case all elliptic terms in Eqs. 1.1 can be written
as ∇ ⋅ (𝐺(𝑈)∇𝑈)), where 𝐺(𝑈) is a homogeneity tensor that does not contain any
gradients of the unknowns. (See, e.g., [21, 22].) This tensor is then kept fixed
during an iteration step, while ∇𝑈 can be treated in a timeimplicit manner. If the
density is constant, then that approach is equivalent to the above discretization.
If the density is variable (as it will be in later chapters), then using a homogeneity
tensor is subtly different from the current method in terms of which variables get
treated implicitly. Three of the six terms in Eq. 2.12 contain 𝐝 𝐦, which is a product
of 𝐦, 1/𝜌, and ∇𝜌. When using a homogeneity tensor, 𝐦 and (1/𝜌) are frozen
within an iteration step, and ∇𝜌 is solved for implicitly. In the current discretization,
(1/𝜌) and ∇𝜌 are frozen, and 𝐦 is treated implicitly.
Our approach also differs from that of Klein et al. [23], in that we treat all terms
in the viscous stress (Eq. 1.2) in a timeimplicit manner, whereas they only do this
for the first term (𝜇∇𝑖 𝑢𝑗 ). In our treatment the velocity components are coupled.
We have no a priori estimate for the difference in magnitude between the effects of
the first term (𝜇∇𝑖 𝑢𝑗 ) and its transpose (𝜇∇𝑗 𝑢𝑖 ) on the viscous force ∇ ⋅ 𝜏, especially
when the viscosity varies strongly in space. Note the gradients in the effective
viscosity will increase greatly when a large eddy simulation (LES) model is included
in chapter 5.
Following Hillewaert [7, p. 30], we set the penalty parameter to
‖𝐹‖leb
𝜂𝐹 = max (𝐶𝑇 card(ℱ𝑇 ) ) max (𝐾|𝑇 ) , (2.16)
𝑇∈𝒯𝐹 ‖𝑇‖leb 𝑇∈𝒯𝐹
where 𝐾 = 𝜇/𝜌 is the diffusion parameter, and card(ℱ𝑇 ) is the number of faces
of element 𝑇. The factor 𝐶𝑇 depends on the type of elements in the mesh: for a
polynomial order 𝒫, 𝐶𝑇 = (𝒫 + 1)2 for quadrilaterals and hexahedra, 𝐶𝑇 = (𝒫 +
1)(𝒫 + 2)/2 on triangles, and 𝐶𝑇 = (𝒫 + 1)(𝒫 + 3)/3 for tetrahedra.
We compute the penalty parameter in a pointwise manner, even though Hille
waert took the maximum value of the above expression on the face in his stability
analysis. Our experience suggests no difference in the stability, and taking a local,
pointwise (as opposed to faceaveraged) numerical flux seems more in the spirit of
the DG method, which can have an arbitrarily rich structure within an element. We
will encounter a similar situation for the stabilization term for the local LaxFriedrichs
flux in section 2.5.1.
Paradoxically Eq. 2.16 depends on the shape, size, and number of faces of the
neighbors, giving the impression that it is still not local to a particular point on a
face, even though we evaluate it in a pointwise manner. However, the norms of a
2 polynomial on a face and on an element are related by wellknown trace inequalities,
which underlie the above expression for the penalty parameter.
2.4. Discrete Enthalpy Equation

We solve for the specific enthalpy ℎ from the enthalpy transport equation in con
servative form. Given a solution space 𝑉ℎ for the enthalpy, the discrete weak form
of the stationary enthalpy equation 2.1c is
Find ℎ ∈ 𝑉ℎ , such that, for all 𝑣 ∈ 𝑉ℎ ,
(2.17)
𝑎conv (𝐦, ℎ, 𝑣) + 𝑎SIP (ℎ, 𝑣) = 𝑙 conv (𝐦, ℎD , 𝑣) + 𝑙 SIP (𝑣) + ∫ 𝑄 𝑣 ,
Ω
SIP SIP
where 𝑎 and 𝑙 are standard SIP bilinear and linear forms to discretize the
Fourier heat flux. The SIP penalty parameter is as in Eq. 2.16, with a diffusion
coefficient 𝐾 = 𝑘/𝑐𝑝 . Note that the convective discretization is the same as for the
mass flux, except that the convecting field is 𝐦, rather than 𝐮 = (1/𝜌)𝐦. (That is,
𝐛 = 𝐦 in Eqs. 2.20–2.21.) This is convenient, because we also solve for 𝐦.
2.5. Discretization of the Convection

To derive an expression for the convective bilinear and linear forms (𝑎conv and 𝑙 conv
in Eqs. 2.10 and 2.17), consider the timeindependent, purely convective problem
with a numerical (discontinuous) advecting field 𝐛, and an unknown generic scalar
𝜙:
∇ ⋅ (𝐛 𝜙) = 0 . (2.18)
Though this equation is linear (because 𝐛 is given), its analysis is also relevant to
the nonlinear convection in the momentum equation, which must be linearized with
some estimate for the convecting velocity field.
At 𝜕ΩD we have the Dirichlet values 𝐛D ⋅ 𝐧 and 𝜙D , whereas there may be
no numerical inflow (i.e., 𝐛 ⋅ 𝐧 < 0) at the Neumann boundary. Note that 𝜙D is
only defined where 𝐛D ⋅ 𝐧 < 0. This is an important point to which we will return
later: the inflow region is defined by the known value 𝐛D ⋅ 𝐧, not by the sign of the
numerical value 𝐛 ⋅ 𝐧.
The discrete weak form of Eq. 2.18 is
Find 𝜙 ∈ 𝑉𝜙 , such that, for all 𝑣 ∈ 𝑉𝜙 ,
(2.19)
𝑎conv (𝐛, 𝜙, 𝑣) = 𝑙 conv (𝐛, 𝜙D , 𝑣) ,
where 𝑉𝜙 is the solution space of 𝜙. The linear term is
𝑙 conv (𝐛, 𝜙D , 𝑣) = − ∫ 𝑣 min (0, 𝐛D ⋅ 𝐧) 𝜙D . (2.20)

𝜕ΩD
2.5. Discretization of the Convection 17
The bilinear term has the general form
𝑎conv (𝐛, 𝑤, 𝑣) = − ∑ ∫ 𝑤 𝐛 ⋅ ∇𝑣 + ∑ ∫ J𝑣K 𝐻𝐹 (𝐛, 𝑤)

𝑇 𝐹
𝑇∈𝒯 𝐹∈ℱi (2.21)
+∫ 𝑣 𝑤 𝐛⋅𝐧+∫ D
𝑣 𝑤 max (0, 𝐛 ⋅ 𝐧) .
2
𝜕ΩN 𝜕ΩD
Here 𝑤 is a scalar, 𝑣 is the test function, 𝐛 is the convective field, and 𝐻𝐹 is the
numerical flux function on a face 𝐹, which will be defined later.
It is well known that imposing a Dirichlet boundary condition for the velocity at
an outlet results in an illposed problem, and that it is numerically unstable for a
convectiondominated flow, and therefore we would normally have max(0, 𝐧⋅𝐛D ) =
0 on 𝜕ΩD . Here we nevertheless include this term in 𝑎conv , because we will use it
in the TaylorGreen vortex in section 3.3.1, as is standard practice for that laminar
benchmark case (e.g., [12, 24]).
In practice one may know the value 𝜙D at a Dirichlet outlet (𝐧 ⋅ 𝐛D > 0), such as
for the TaylorGreen vortex manufactured solution in section 3.3.1, but we nonethe
less use the internal value 𝑤 in the last term in Eq. 2.21, so as not to overconstrain
the problem. This theoretical point has little practical value, because Dirichlet out
lets are only viable for academic problems anyway.
If (the normal component of) 𝑏 is continuous at each face, then 𝜙 can be up
winded in an unambiguous manner, and there is only one correct discretization
(see, e.g., [21, p. 33]). In practice 𝐛 is a velocity (or a mass flux) that was ob
tained with a DG method, and therefore the flux is discontinuous at each face, and
there are multiple possible discretizations.
2.5.1. Choice of the Numerical Flux

The convective discretization is closed by defining a numerical flux function 𝐻𝐹
in Eq. 2.21 for an internal face 𝐹. It only depends on the basis function (𝑤)
and the normal component of the flow (𝐛 ⋅ 𝐧) on both sides of the face; there
is no interpolation between the elements. The numerical flux can therefore be
borrowed from finite volume methods for onedimensional hyperbolic conservation
laws. There are many monographs on this subject, including the standard works by
LeVeque [25, 26]. Toro [27] offers a particularly clear overview with an emphasis
on the Euler equations.
The most principled numerical flux can be obtained by solving the associated Rie
mann problem at the discontinuity exactly, which is known as Godunov’s method.
See, for example, [8, p. 105] and especially [3] for notes in a DG context. Go
dunov’s method is known to lead to the least amount of numerical dissipation
[2, 28]. Unfortunately the Riemann problem typically takes up a substantial part of
the total computation time. It is therefore more common to use an approximation
for the numerical flux.
We use the local LaxFriedrichs flux, which is cheap, but also known to be fairly
dissipative. Among the many other numerical fluxes, we mention the following.
• The Roe numerical flux, which has been compared favorably to the local Lax
Friedrichs flux by Winters et al. [29]. They focused on the interplay between
the numerical flux and their underintegrated nodal DG method that required
antialiasing, and so it is hard to say how well their results would translate to
our modal DG method with exact integration in the weak forms.
2
• The Vijayasundaram numerical flux. See, for example, the seminal paper [30]
for a general analysis, and [31] for a DG application. In a DG method this
numerical flux can be shown to be consistent with the numerical flux in the
continuity equation.
The Vijayasundaram flux has also been implemented in DGFlows; our experience
suggests no significant difference with respect to the local LaxFriedrichs flux.
The local LaxFriedrichs flux is given by
1
𝐻𝐹 (𝐛, 𝑤) = J𝑤K 𝛼 𝐹 + {𝑤𝐛} ⋅ 𝑛𝐹 , (2.22)
2
where 𝛼 𝐹 is a function of the normal component of the convecting field (i.e., 𝐛 ⋅ 𝐧).
For a general scalar hyperbolic system of the form ∇ ⋅ 𝐟(𝜙) = 0 with a given vector
valued function 𝐟, the parameter 𝛼 𝐹 is the maximum value of |𝐟′ (𝜙) ⋅ 𝐧𝐹 | on either
side of the face. If 𝜙 is an advected scalar, such as for the enthalpy in Eq. 2.17,
then 𝐟(𝑤) = 𝑤 𝐛, and thus |𝐟′ (𝜙) ⋅ 𝐧𝐹 | = |𝐛 ⋅ 𝐧𝐹 |.
The situation is different for the nonlinear convection in the momentum equation
2.10, where 𝜙 is itself equal to the convecting field 𝐛 (up to a factor of the density).
In that case, 𝜙 = 𝐦, 𝐛 = 𝐮 = (1/𝜌)𝐦, and Eq. 2.18 becomes a vector equation
of the form ∇ ⋅ 𝐹 (𝐦) = 0 with 𝐹𝑖𝑗 (𝐰) = 𝑤𝑖 𝑤𝑗 /𝜌. Now 𝛼 𝐹 is the maximum of the
spectral radii of the Jacobi matrix 𝜕/𝜕𝜙𝑖 (𝐹𝑗𝑘 (𝜙)𝑛𝑘𝐹 ) = 𝜕/𝜕𝑚𝑖 (𝑚𝑗 (𝐦 ⋅ 𝐧𝐹 )) /𝜌 =
(𝐮 ⋅ 𝐧𝐹 ) 𝛿𝑖𝑗 + 𝑛𝑖𝐹 𝑢𝑗 on either side of the face. Its eigenvectors are either parallel to
𝐧 (with eigenvalue 2(𝐮 ⋅ 𝐧𝐹 )), or perpendicular to 𝐮 (with eigenvalue (𝐮 ⋅ 𝐧𝐹 )). This
can be summarized as
𝛼 𝐹 = 𝜘 max |𝐛 ⋅ 𝐧𝐹 |𝑇 (2.23)
𝑇∈𝒯𝐹
with 𝜘 = 1 for advected scalars (such as the enthalpy in Eq. 2.17), and 𝜘 = 2 for
the nonlinear convection in the momentum equation 2.10 (see, e.g., [2, 12]).
Note that we evaluate 𝛼 𝐹 in a pointwise manner in the integral in Eq. 2.21,
which is the only right choice if 𝜙 is an advected scalar. Averaging 𝛼 𝐹 over the
face would be inconsistent, as can beq shown
y by considering the special case of
a continuous convecting field 𝐛̃ (i.e., 𝐛̃ = 0 on internal faces, and 𝐛̃ = 𝐛D on
𝜕ΩD ). With a test function that is 𝑣 = 1 on an element 𝑇 and 𝑣 = 0 elsewhere, the
discretization Eq. 2.19 can be written as
0= ∫ 𝑤 (𝐛̃ ⋅ 𝐧) + ∫ 𝑤up (𝐛D ⋅ 𝐧)

𝜕𝑇 ⋂ 𝜕ΩN 𝜕𝑇 ⋂ 𝜕ΩD
1 (2.24)
+ ∑ ∫ ((𝑤+ − 𝑤− ) 𝛼 𝐹 + (𝑤+ + 𝑤− ) 𝐛̃ ⋅ 𝐧𝑇 ) ,
𝐹 2
𝐹∈ℱ𝑇 ⋂ ℱi
where 𝐧𝑇 is the outward normal of element 𝑇, 𝑤+ is the value on 𝑇, 𝑤− is the value

on its neighbor, and
𝜙D , for 𝐧 ⋅ 𝐛D < 0,
𝑤up ≔ { (2.25)
𝑤, for 𝐧 ⋅ 𝐛D > 0
is the upstream value at a Dirichlet boundary. Since 𝐛̃ is continuous, this must 2
reduce to a pure upwind flux, so the last integrand must equal (𝐛̃ ⋅ 𝐧𝑇 ) 𝑤± for
±𝐛̃ ⋅ 𝐧𝑇 > 0. This only holds if 𝛼 𝐹 = |𝐛̃ ⋅ 𝐧𝑇 | everywhere on 𝐹.1
The above argument breaks down for the nonlinear convection term in the mo
mentum equation 2.10, because then 𝜙 and 𝐛 are equivalent, and so a continuous 𝐛
would imply a continuous 𝜙. Nevertheless our experience suggests that evaluating
𝛼 𝐹 in a pointwise manner sometimes results in noticeably lower errors, especially
for highorder numerical solutions, whereas the stability does not seem negatively
impacted. As mentioned in section 2.3.1, averaging a penalty parameter over a
face or an element runs counter to our intuitive understanding of a highorder DG
method.
This opinion does not appear universally shared in the literature. Cockburn and
Shu [2] have suggested using to the two element averages on the neighbors to
compute 𝛼 𝐹 for the nonlinear convection term. Shahbazi et al. [12] was perhaps
the first to put this into practice, and others have followed (e.g., [32]). More re
cently, de la Llave Plata et al. [33] have evaluated 𝛼 𝐹 in a pointwise manner with
underresolved DG large eddy simulations, and Tavelli and Dumbser [34] have done
the same for a spacetime DG method, where averaging over a spacetime element
would presumably have been more involved. None of the above authors have given
explicit reasons for their choices, and we are not aware of a systematic study on
averaging 𝛼 𝐹 or not.
2.5.2. Solution Spaces for the Enthalpy and the Pressure

If a scalar quantity 𝜙 is advected with a velocity field that was obtained with a
discontinuous Galerkin method, then the solution space of the scalar must be a
subset of the solution space of the pressure, that is, 𝒫𝜙 ≤ 𝒫𝑝 . Discretizations with
𝒫𝜙 > 𝒫𝑝 are inconsistent, and therefore often unstable, because the continuity
equation is weighed by the pressure basis functions, so that the numerical velocity
only satisfies the incompressibility constraint in a weak sense up to order 𝒫𝑝 . This
means that the convective discretization can only be consistent up to an order 𝒫𝑝
[1, 35].
To make this more precise, consider the discrete continuity equation in Eq. 2.7
with the divergence operator in the form of Eq. 2.5:
−𝑎stab (𝑝, 𝑞)+ ∑ ∫ 𝐦⋅∇𝑞 = ∑ ∫ J𝑞K {𝐦}⋅𝐧𝐹 +∫ 𝑞 𝐦⋅𝐧+∫ 𝑞 𝐦D ⋅𝐧 , (2.26)

𝑇 𝐹 𝜕ΩN 𝜕ΩD
𝑇∈𝒯ℎ 𝐹∈ℱi
for a test function 𝑞 that lies in the pressure solution space. This can be compared
to the advection discretization in Eq. 2.19 by substituting the continuous solution
1 This is also another way of showing that 𝜘 = 1 for a scalar 𝜙.
q y
𝜙 ← 𝜙,̃ which satisfies 𝜙̃ = 0 on internal faces and 𝜙̃ = 𝜙D on 𝜕ΩD . Every viable
numerical flux must reduce to the central flux for a continuous solution, that is,
̃ = 𝜙̃ {𝐛} ⋅ 𝐧𝐹 (see, e.g., [21]), so that Eqs. 2.19–2.21 become
𝐻𝐹 (𝐛, 𝜙)
∑ ∫ 𝜙̃ 𝐛 ⋅ ∇𝑣 = ∑ ∫ 𝜙̃ J𝑣K {𝐛} ⋅ 𝐧𝐹 + ∫ 𝜙̃ 𝑣 𝐛 ⋅ 𝐧 + ∫ 𝜙D 𝑣 𝐛D ⋅ 𝐧 . (2.27)

2 𝑇∈𝒯
𝑇
𝐹∈ℱi
𝐹 𝜕ΩN 𝜕ΩD
This can clearly only be consistent for a test function 𝑣 that is part of the test
space of the continuity equation. In the special case of a constant solution 𝜙,̃ the
convective term should vanish for all 𝑣, but 𝑎conv (𝐦, 𝜙,̃ 𝑣) = 𝜙̃ 𝑎div (𝐦, 𝑣), which
only vanishes if the test function 𝑣 lies in the pressure solution space, that is, if
𝒫𝜙 ≤ 𝒫𝑝 .
The requirement 𝒫𝜙 ≤ 𝒫𝑝 was not satisfied in some previous literature on mixed
order DG schemes. For example, Klein et al. [23] chose the same solution space
for the temperature as for the components of the velocity field. They probably
found good results because their tests were done at a low Prandtl number of 0.7,
whereas the problem with the solution spaces manifests itself when the convective
term dominates.
In theory the discretization of the convection of a scalar is consistent as long as
its solution space is a subset of the solution space of the pressure, but in practice the
enthalpy has the same solution space as the pressure, because a scalar transport
equation is much cheaper to solve than the momentum and pressure equations,
so there is little reason not to obtain the highest available spatial accuracy for the
enthalpy. In later chapters that address lowMach number flow, the density will be
a function of the temperature (and thus, the enthalpy). Then the choice of the en
thalpy solution space is no longer free: it must be the same as that of the pressure,
because the continuity equation and the enthalpy equation become coupled due to
the temperaturedependent density.
2.5.3. Proper Treatment of Dirichlet Boundary Conditions

Regardless of the numerical flux, setting 𝑣 = 1 in the discretization in Eqs. 2.19–
2.21 reveals the global conservation property
0=∫ 𝑤 𝐛⋅𝐧+∫ 𝑤up 𝐛D ⋅ 𝐧 (2.28)

𝜕ΩN 𝜕ΩD
with 𝑤up as in Eq. 2.25, as opposed to ∫𝜕Ω 𝑤 𝐛 ⋅ 𝐧 = 0, which one may have
guessed from the governing equation 2.18. This shows the effect of our treatment
of the Dirichlet boundary condition: the total inflow is determined by the known
D
boundary conditions (i.e., (𝜙𝐛) ⋅ 𝐧), rather than the internal value (i.e., 𝜙 𝐛 ⋅ 𝐧),
which generally contains a numerical error.
Another way of looking at this is to rewrite Eq. 2.27 for the continuous solution
𝜙̃ as
∑ ∫ 𝜙̃ 𝐛 ⋅ ∇𝑣 − ∑ ∫ 𝜙̃ J𝑣K {𝐛} ⋅ 𝐧𝐹 = ∫ 𝜙D 𝑣 (𝐛D − 𝐛) ⋅ 𝐧 . (2.29)

𝑇 𝐹 𝜕ΩD
𝑇∈𝒯 𝐹∈ℱi,N,D
The righthand side acts as a source to counterbalance the deviation of the numer
ical inflow from the imposed inflow at the boundary. Note that the inflow term in
the discrete continuity equation 2.7 must be consistent with the treatment of the
Dirichlet boundary in the convection discretization. Otherwise Eqs. 2.26 and 2.27
in the previous section would not have been consistent in the special case of a
constant continuous solution 𝜙.̃ 2
Eqs. 2.20–2.21 present the only correct treatment of the Dirichlet boundary,
though it is far from standard in recent literature. For example, Piatkowski et al.
[31] used a Vijayasundaram numerical flux 𝐻𝐹 (𝐛, 𝑤) = 𝑤+ max (0, {𝐛} ⋅ 𝐧𝐹 ) +
𝑤− min (0, {𝐛} ⋅ 𝐧𝐹 ) on internal faces, and extended this in a seemingly logical way
to Dirichlet faces by replacing 𝑤− with 𝜙𝐷 , so that the contribution of the Dirichlet
boundary became
∫ 𝑣 (𝑤 max (0, 𝐛 ⋅ 𝐧) + 𝜙𝐷 min (0, 𝐛 ⋅ 𝐧)) . (2.30)

𝜕ΩD
This is subtly wrong, because the numerical value 𝐛 ⋅ 𝐧 may have any sign, though
a value for 𝜙𝐷 should not be required at outlets and walls (where 𝐛D ⋅ 𝐧 = 0).
A similar objection can be made to the boundary treatment in Shahbazi et al.
[12], who used a local LaxFriedrichs flux as in Eq. 2.22, and substituted {𝑤 𝐛} ⟵
D
(𝜙𝐛 + (𝜙 𝐛) ) /2 and J𝑤K ⟵ 𝑤 − 𝜙D on 𝜕ΩD , giving a Dirichlet boundary contri
bution
1 D
∫ 𝑣 ((𝑤𝐛 + (𝜙𝐛) ) ⋅ 𝑛 + (𝑤 − 𝜙D ) 𝛼 𝐹 ) . (2.31)
𝜕Ω D 2
They based 𝛼 𝐹 on the average in the boundary element (𝐛avg ) and the Dirichlet
value 𝐛D . For walls, defined by 𝐛D ⋅ 𝐧 = 0, the above term becomes
1
∫ 𝑣 ( 𝑤 𝐛 ⋅ 𝐧 + (𝑤 − 𝜙D ) |𝐛avg ⋅ 𝐧|) . (2.32)
𝜕ΩD 2
They used this discretization for the nonlinear convective term, so that 𝐛 = 𝐮 and
𝜙 = 𝑢𝑖 for some direction 𝑖. Clearly this requires a Dirichlet value for all directions
of the velocity, while the nonnormal components of 𝐮 are not physically relevant
to the convection at a wall.
Many other authors have extended their numerical fluxes to the Dirichlet bound
ary in different but similar ways, with similar problems. Examples include [19,
33, 36] for the local LaxFriedrichs flux, [37] for the LesaintRaviar numerical flux,
and [38, 39] for Riemannsolved artificial compressibility flux.2 These papers have
treated Dirichlet faces like they are internal faces, except with the neighbor values
replaced by the Dirichlet values. This is motivated by an understandable desire to
impose the Dirichlet boundary condition weakly, but note that the present treat
ment is also weak, because we only use the moments ∫𝜕ΩD 𝑣 min (0, 𝐧 ⋅ 𝐛D ) 𝜙D for
all test functions 𝑣, not the inflow value at every point on 𝜕ΩD . The correct Eqs.
2 though in that last case it is possible that the boundary treatment is at least consistent between the
continuity and momentum equations
2.20–2.21 can also be found in some previous works, such as [13, pp. 27–28] and
[40].
Perhaps the inaccuracies described above have little impact on the numerical
solution in practical calculations, because most convected quantities are also dif
fused, which often weakly enforces the Dirichlet boundary condition for 𝜙 at 𝜕ΩD .
2 Superfluous convective boundary terms, such as Eqs. 2.30 and 2.31, would then
effectively change the penalty parameter in the SIP method at Dirichlet boundaries.
2.6. Implementation
All flow simulations are performed with an inhouse solver DGFlows. Its distin
guishing feature is that it can be coupled with another inhouse solver for particle
transport, called Phantom𝑆𝑁 . This has been used to solve the coupled flow and
neutron transport equations for modeling a theoretical nuclear reactor that is based
on a liquid fuel [41].
As mentioned in section 2.1.2, the basis functions are modal and hierarchical.
As is standard in finite element implementations, they are defined on a ‘local’ ref
erence element, which is mapped to the ‘global’, physical elements in the mesh.
Sometimes this localglobal mapping is affine, such as when the global elements
are triangles, tetrahedrons, rectangles, or rectangular parallelepipeds. In that case
the basis functions can be orthonormalized on the local element, which results in an
orthogonal basis on the global elements, so that the mass matrix is diagonal. This
can be a minor performance gain, especially if one were to use a numerical method
that requires frequent projections of the numerical solution onto lowerorder poly
nomial spaces, such as a 𝑝multigrid solver, or a dynamic large eddy simulation.
All integrals are evaluated with a quadrature set that is sufficiently accurate to
negate the polynomial aliasing effect that has plagued other DG solvers. (See, e.g.,
[42].) This is feasible because there are only (𝒫 + 𝑑)!/(𝑑! 𝒫!) degrees of freedom
in a 𝑑dimensional element with a polynomial order 𝒫. In the limit of large 𝒫 in
𝑑
three dimensions, this is 𝑑! = 6 times less than the (𝒫 + 1) degrees of freedom in
some nodal bases. The abscissa and the weights are taken from Solin et al. [43].
We store the values and derivatives of the basis functions on the quadrature set
for a fast evaluation of integrals and numerical solutions. All results in this thesis
remained unchanged when the accuracy of the quadrature was increased.
All meshes were generated with the opensource software tool Gmsh [44].
The linear systems are solved with the MPIbased software library PETSc [45,
46]. We use a conjugate gradient (CG) method for the pressure equation, and a
GMRES method for the enthalpy and momentum equations, which are asymmetric.
The computations are parallelized by partitioning the mesh with the software
package METIS [47]. Each core is assigned one partition. The parallel matrix
preconditioner is a standard block Jacobi method. We use PETSc’s implementa
tions of the pipelined Krylov methods, which require fewer inner products (meaning
fewer global MPI reductions) at the cost of more serial computations (see [48] for
pipelined CG, and [49] for pipelined GMRES). This provided better performance on
multinode computations.
2.7. Test Case: A Heated Backwardfacing Step 23
The preconditioner for the submatrix within a process depends on the specific
calculation. If the matrix remains constant for all time steps, then we use an incom
plete LU (or Cholesky) decomposition with zero fill (i.e., ILU(0) or ICC(0)). These
can be reused every time that the system is solved. If the matrix is different for
each time step, then the incomplete decompositions are too expensive, and we
instead use a block GaussSeidel method, where the degrees of freedom within an 2
element are treated as one block.
The ordering of the elements within a core is random. It would likely be more
efficient to renumber them based on the flow direction, so that a block GaussSeidel
method for the transport equations would correspond to the upwind direction. Fid
kowski et al. [50] and Diosady and Darmofal [51] have shown significantly increased
performance by forming lines of maximum coupling between elements, and solving
a blocktridiagonal system along each line.
Another obvious improvement is based on the hierarchical highorder solution
space, which strongly suggests a 𝑝multigrid method, based on Galerkin projection
into lowerorder polynomial spaces. Multigrid methods can greatly reduce the com
putational time for the pressure Poisson equation, which is often the most expensive
part of an incompressible flow solver. This idea has been tested successfully for DG
discretizations as early as 2005 [52, 53], and has since been applied to convection
advection problems by several other groups (e.g., [50, 54], [7, pp. 63–73]).
More recently, the high algorithmic intensity and minimal coupling between el
ements in a DG discretization have been exploited for efficient implementations on
graphical processing units (GPUs). Various groups have demonstrated huge poten
tial gains (e.g., [55–57]).
Unfortunately these ideas have not made it into generic numerical software
packages, and we have made no attempt to implement them into DGFLows, which
is purely a research code. Linear solvers are not studied in this thesis.
2.7. Test Case: A Heated Backwardfacing Step

Backwardfacing steps have a long history as benchmark cases for CFD methods.
We simulate a twodimensional case with an expansion ratio of 2, combined with
scalar transport. Fig. 2.1 shows the domain. The inlet velocity (at 𝑥 = −𝐿0 ) is
(in)
given by 𝑢2 = 0 and 𝑢1 = 𝑢1 ≔ −6𝑢(𝑦 ̄ − 𝑆)(𝑦 − 2𝑆)/𝑆2 , so that the average
2𝑆 (in)
inlet velocity is ∫𝑦=𝑆 𝑢1 /𝑆 = 𝑢.̄ The temperature is 𝑇0 at the inlet and 𝑇1 along the
bottom wall; the other walls are isolated.
The Reynolds number can be defined in many ways; here
2𝑆𝑢̄
Re ≔ . (2.33)
𝜈
Some other literature uses the maximum velocity, or the height before the expan
sion. Eq. 2.33 is used because 2𝑆 is the hydraulic diameter at the inlet, making
it consistent with the standard definition of the bulk Reynolds number for three
dimensional channel flow and pipe flow, which we will also see in Chapter 5.
T0
S
2S
2
y
x T1
L0 L
Figure 2.1: Geometry of the backwardfacing step (not to scale).
The quantities of interest are the dimensionless velocity and temperature gra
dients along the bottom wall. Define the local Darcy friction factor
(bottom)
8 𝜏w 8𝜈
𝑓D ≔ 2
=− 2
𝐧 ⋅ ∇𝑢1 , (2.34)
𝜌(𝑢/2)
̄ (𝑢/2)
̄
(bottom)
where 𝜏w = −𝜇 𝐧 ⋅ ∇𝑢1 is the wall shear stress along the bottom wall, and
(𝑢/2)
̄ is the bulk velocity after the expansion. The local Nusselt number is
𝑆
Nu ≔ 𝐧 ⋅ ∇𝑇 . (2.35)
𝑇0 − 𝑇1
The domain should be long enough in order for the outlet not to influence the
flow near the expansion. The present results were obtained with 𝐿/𝑆 = 70. The
domain length before the expansion has little impact on the laminar solution; here
we let 𝐿0 = 𝑆. The results did not change when we ran the same simulation with
𝐿/𝑆 = 55.
Of course this does not mean that the flow is fully developed near the outlet. Far
from the expansion, the velocity and temperature approach the analytical solutions
2
3 𝑢̄ 𝑦−𝑆
lim 𝑢1 = (1 − ( ) ) ,
𝑥/𝑆→∞ 2 2 𝑆 (2.36)
lim 𝑇 = 𝑇1 ,
𝑥/𝑆→∞
giving
𝜈/𝑆 96
lim 𝑓D = 24 = ,
𝑥/𝑆→∞ 𝑢/2
̄ Re (2.37)
lim Nu = 0 ,
𝑥/𝑆→∞
2.7. Test Case: A Heated Backwardfacing Step 25
Figure 2.2: Computational mesh for the backwardfacing step near the expansion edge.
(For comparison, the Darcy friction factor for laminar flow in a circular channel [i.e.,
HagenPoiseuille flow] is 𝑓D = 64/Re.) The results will show that 𝑓D and Nu are
not close to these values near our outlet.
Fig. 2.2 shows the structured mesh. The mesh is refined near the expansion
edge. There are approximately 35k elements, with a secondorder polynomial ap
proximation for the mass flux, pressure, and enthalpy, resulting in approximately
210k degrees of freedom per unknown.
An accurate solution depends on a sufficiently fine mesh near the expansion
edge. The wall shear stress at the inlet is
(in)
(in) 𝜕𝑢1 𝜇𝑢̄ 𝜇2
𝜏w =𝜇 | =6 = 3 Re 2 , (2.38)
𝜕𝑦 𝑆 𝜌𝑆
𝑦=𝑆
(in)
which can be used to define a wall shear velocity of 𝑢𝜏 ≔ √𝜏w /𝜌 = (𝜈/𝑆)√3 Re,
and a dimensionless wall distance of 𝑦 + ≔ 𝑦𝑢𝜏 /𝜈 = (𝑦/𝑆)√3 Re. The first element
at the wall is placed at 𝑦 + = 1.0.
Fig. 2.3 shows the results for Re = 1 400 and Pr = 0.7132. These conditions
were meant to reproduce one of the numerical test cases by Xie and Xi [58]3 ,
who performed an unsteady simulation with the initial condition 𝑢1 = 0 for 𝑦 < 𝑆,
3 The fact that Pr = 0.7132 in their calculations was obtained from our private correspondence.
0.2
fD lim fD
x/S→∞
Nu 5
Barkley et al.
0.1
4
2
3
Nu
0.0
fD
2
−0.1
−0.2
0
0 10 20 30 40 50
x/S
Figure 2.3: Steadystate local Darcy friction factor 𝑓D (Eq. 2.34) and local Nusselt number Nu (Eq.
2.35) for the backwardfacing step along the bottom wall. The cross (×) indicates the reattachment
length (where 𝑓D = 0) that was reported by Barkley et al. [61].
(in)
𝑢1 = 𝑢1 for 𝑦 > 𝑆, 𝑢2 = 0, and 𝑇 = 𝑇0 . They reported timeaveraged results.
We performed the same timedependent calculation (with the pressure correction
method that will be the topic of the next chapter), but we found that all transients
die out, and we reached a steady state. Furthermore, our results differ substantially
from Xie and Xi (not shown here). We nevertheless believe our results to be correct
for two reasons. First, several other previous studies have also found a steady
state in 2D at this Reynolds number (e.g., [59, 60]). Second, the location of our
reattachment length (i.e., the largest value of 𝑥 for which 𝑓D = 0) is at 𝑥/𝑆 = 15.342,
which agrees with the value of 𝑥/𝑆 = 15.358 that can be inferred by interpolating
the data from Fig. 5 in Barkley et al. [61].
2.8. Discussion and Conclusion

The discrete convection in section 2.5 contains two corrections on previous litera
ture that are essential to the consistency of the numerical method. First, the inflow
at a Dirichlet boundary does not depend on an internal numerical value (ref. section
2.5.3). This contradicts a majority of papers on the DG method that we have seen.
Second, the solution space of an advected quantity lies in the solution space of the
pressure (ref. section 2.5.2). These points guarantee that the advection discretiza
tion is satisfied by the continuous advected quantity, even when the advecting field
contains a numerical error. Phrased differently, if the Dirichlet boundary is not
treated correctly, then the discretization of the full system of transport equations
may be consistent, but the transport equations in isolation are not.
References 27
The numerical fluxes for the convection and diffusion discretizations are based
on the wellknown local LaxFriedrichs and interior penalty methods, but we differ
from the commonly accepted approach of averaging or maximizing the penalty pa
rameters over a region, instead evaluating them in a pointwise manner. The merit
of this is somewhat subjective, though we have argued that pointwise flux defini
tions are more in line with the local nature of the DG method. This is particularly 2
true for the penalty term in the local LaxFriedrichs flux, which is inconsistent for
linear advection when 𝛼 𝐹 is averaged (ref. section 2.5.1, Eq. 2.24). Penalty terms
can have a significant impact on stability, the solution quality, and the stiffness of
the linear system. A numerical comparison between pointwise and averaged or
maximized penalty parameters could therefore be interesting.
The penalty parameter in Eq. 2.16 could be an overestimate of the minimum
value that achieves coercivity of the SIP method, because it was developed for a
nodal basis. For the same minimum polynomial order, our modal basis functions
have a substantially lower maximum polynomial order than the nodal basis functions
(see section 2.1.2). We did not investigate alternative expressions for the penalty
parameter.
The timeindependent test case in section 2.7 shows good agreement with pre
vious literature. It is also fairly simple. The following chapters will feature more
challenging transient simulations, which double as further verification and validation
of the spatial discretization that is described here.
References
[1] A. Hennink, M. Tiberga, and D. Lathouwers, A Pressurebased solver for low
Mach number flow using a discontinuous Galerkin method, Journal of Compu
tational Physics 425, 109877 (2021).
[2] B. Cockburn and C.W. Shu, RungeKutta Discontinuous Galerkin Methods

for ConvectionDominated Problems, Journal of Scientific Computing 16, 173
(2001).
[3] M. Dumbser, D. S. Balsara, E. F. Toro, and C.D. Munz, A unified framework for
the construction of onestep finite volume and discontinuous Galerkin schemes
on unstructured meshes, Journal of Computational Physics 227, 8209 (2008).
[4] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini, Unified analysis of dis

continuous Galerkin methods for elliptic problems, SIAM Journal on Numerical
Analysis 39, 1749 (2002).
[5] K. Shahbazi, An explicit expression for the penalty parameter of the interior
penalty method, Journal of Computational Physics 205, 401 (2005).
[6] Y. Epshteyn and B. Rivière, Estimation of penalty parameters for symmet

ric interior penalty Galerkin methods, Journal of Computational and Applied
Mathematics 206, 843 (2007).
28 References
[7] K. Hillewaert, Development of the discontinuous Galerkin method for high

resolution, large scale CFD and acoustics in industrial geometries, Phd thesis,
Université catholique de Louvain (2013).
[8] D. A. di Pietro and A. Ern, Mathematical Aspects of Discontinuous Galerkin

2 Methods (Springer, 2012).
[9] B. Krank, M. Kronbichler, and W. A. Wall, Wall modeling via function enrich
ment within a highorder DG method for RANS simulations of incompressible
flow, International Journal for Numerical Methods in Fluids 86, 107 (2018).
[10] B. Krank, M. Kronbichler, and W. A. Wall, A multiscale approach to hybrid

RANS/LES wall modeling within a highorder discontinuous Galerkin scheme
using function enrichment, International Journal for Numerical Methods in Flu
ids 90, 81 (2019).
[11] J. S. Hesthaven and T. Warburton, Nodal Discontinuous Galerkin Methods

(Springer, New York, 2008).

[13] G. Kanschat, Discontinuous Galerkin Methods for Viscous Incompressible Flow,

1st ed. (Teubner Research: Deutscher UniversitätsVerlag, Wiesbaden, 2007).
[14] V. Girault, B. Rivière, and M. F. Wheeler, A Discontinuous Galerkin Method

with Nonoverlapping Domain Decomposition for the Stokes and NavierStokes
Problems, Mathematics of Computation 74, 53 (2005).
[15] V. John, Finite Element Methods for Incompressible Flow Problems, Springer
Series in Computational Mathematics (Springer, Cham, 2016).
[16] B. Cockburn, G. Kanschat, and D. Schötzau, An EqualOrder DG Method for

the Incompressible NavierStokes Equations, Journal of Scientific Computing
40, 188 (2009).
[17] L. Botti and D. A. Di Pietro, A pressurecorrection scheme for convection

dominated incompressible flows with discontinuous velocity and continuous
pressure, Journal of Computational Physics 230, 572 (2011).
[18] K. Shahbazi, A Parallel HighOrder Discontinuous Galerkin Solver For the Un
steady Incompressible NavierStokes Equations in Complex Geometries, Phd
thesis, University of Toronto (2007).
[19] B. Krank, N. Fehn, W. A. Wall, and M. Kronbichler, A highorder semiexplicit

discontinuous galerkin solver for 3d incompressible flow with application to
dns and les of turbulent channel flow, Journal of Computational Physics 348,
634 (2017).
References 29
[20] D. A. Di Pietro, Analysis of a discontinuous Galerkin approximation of the

Stokes problem based on an artificial compressibility flux, International Journal
for Numerical Methods in Fluids 55, 793 (2007).
[21] R. Hartmann, Numerical Analysis of Higher Order Discontinuous Galerkin Finite
Element Methods, (2008). 2
[22] C. M. Klaij, J. J. W. van der Vegt, and H. van der Ven, Spacetime discontin
uous Galerkin method for the compressible NavierStokes equations, Journal
of Computational Physics 217, 589 (2006).
[23] B. Klein, B. Müller, F. Kummer, and M. Oberlack, A highorder discontinuous
Galerkin solver for low Mach number flows, International Journal for Numerical
Methods in Fluids 81, 489 (2016).
[24] C. R. Ethier and D. a. Steinman, Exact fully 3D NavierStokes Solutions for
Benchmarking, International Journal for Numerical Methods in Fluids 19, 369
(1994).
[25] R. J. LeVeque, Numerical Methods for Conservation Laws, 2nd ed. (Birkhäuser,
Basel; Boston; Berlin, 1992) p. 214.
[26] R. J. LeVeque, Finite Volume Methods for Hyperbolic Problems, Cambridge
Texts in Applied Mathematics (Cambridge University Press, 2002).
[27] E. F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics, 3rd
ed. (Springer, Berlin, Heidelberg, 2009).
[28] M. E. VázquezCendón, Solving Hyperbolic Equations with Finite Volume Meth
ods, 1st ed. (Springer International Publishing, 2015) p. 188.
[29] A. R. Winters, R. C. Moura, G. Mengaldo, G. J. Gassner, S. Walch, J. Peiro, and
S. J. Sherwin, A comparative study on polynomial dealiasing and split form
discontinuous Galerkin schemes for underresolved turbulence computations,
Journal of Computational Physics 372, 1 (2018), arXiv:1711.10180 .
[30] G. Vijayasundaram, Transonic flow simulations using an upstream centered
scheme of Godunov in finite elements, Journal of Computational Physics 63,
416 (1986).
[31] M. Piatkowski, S. Müthing, and P. Bastian, A stable and highorder accurate
discontinuous Galerkin based splitting method for the incompressible Navier
Stokes equations, Journal of Computational Physics 356, 220 (2018).
[32] B. Klein, F. Kummer, and M. Oberlack, A SIMPLE based discontinuous Galerkin
solver for steady incompressible flows, Journal of Computational Physics 237
(2013), 10.1016/j.jcp.2012.11.051.
[33] M. de la Llave Plata, V. Couaillier, and M.C. le Pape, On the use of a highorder
discontinuous Galerkin method for DNS and LES of wallbounded turbulence,
30 References
[34] M. Tavelli and M. Dumbser, A pressurebased semiimplicit spacetime discon

tinuous Galerkin method on staggered unstructured meshes for the solution
of the compressible NavierStokes equations at all Mach numbers, Journal of
2 [35] M. Tiberga, A. Hennink, J. L. Kloosterman, and D. Lathouwers, A highorder

discontinuous Galerkin solver for the incompressible RANS equations coupled
to the 𝑘𝜖 turbulence model, Computers & Fluids 212, 104710 (2020).
[36] F. Zhang, J. Cheng, and T. Liu, A reconstructed discontinuous Galerkin method
for incompressible flows on arbitrary grids, Journal of Computational Physics
418, 109580 (2020).
[37] E. Ferrer and R. H. J. Willden, A high order Discontinuous Galerkin Finite El
ement solver for the incompressible NavierStokes equations, Computers &
Fluids 46, 224 (2011).
[38] F. Bassi, A. Crivellini, D. A. D. Pietro, and S. Rebay, An artificial compressibil

ity flux for the discontinuous Galerkin solution of the incompressible Navier
[39] M. Franciolini, A. Crivellini, and A. Nigro, On the efficiency of a matrixfree
linearly implicit time integration strategy for highorder Discontinuous Galerkin
solutions of incompressible turbulent flows, Computers & Fluids 159, 276
(2017).
[40] N. Emamy, F. Kummer, M. Mrosek, M. Karcher, and M. Oberlack, Implicit

explicit and explicit projection schemes for the unsteady incompressible
NavierStokes equations using a highorder dG method, Computers & Fluids
154, 285 (2017).
[41] M. Tiberga, Development of a highfidelity multiphysics simulation tool for

liquidfuel fast nuclear reactors, Phd thesis, Delft University of Technology
(2020).
[42] G. Mengaldo, D. De Grazia, D. Moxey, P. Vincent, and S. Sherwin, Dealiasing

techniques for highorder spectral element methods on regular and irregular
grids, Journal of Computational Physics 299, 56 (2015).
[43] P. Solin, K. Segeth, and I. Dolezel, HigherOrder Finite Element Methods, 1st
ed. (Chapman and Hall/CRC, New York, 2003) p. 408.
[44] C. Geuzaine and J. F. Remacle, Gmsh: A 3D finite element mesh genera
tor with builtin pre and postprocessing facilities, International Journal for
Numerical Methods in Engineering 79, 1309 (2009).
[45] S. Balay, S. Abhyankar, M. F. Adams, J. Brown, P. Brune, K. Buschelman,

L. Dalcin, V. Eijkhout, W. D. Gropp, D. Kaushik, M. G. Knepley, D. A. May, L. C.
McInnes, R. T. Mills, T. Munson, K. Rupp, P. Sanan, B. F. Smith, S. Zampini,
References 31
H. Zhang, and H. Zhang, PETSc Users Manual, Tech. Rep. ANL95/11 Revi
sion 3.9 (Argonne National Laboratory, 2018).
[46] S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith, Efficient management of

parallelism in object oriented numerical software libraries, in Modern Software
Tools in Scientific Computing, edited by E. Arge, A. M. Bruaset, and H. P. 2
Langtangen (Birkhäuser Press, 1997) pp. 163–202.
[47] G. Karypis and V. Kumar, A Fast and High Quality Multilevel Scheme for
Partitioning Irregular Graphs, SIAM Journal on Scientific Computing 20, 359
(1998).
[48] P. Ghysels and W. Vanroose, Hiding global synchronization latency in the pre
conditioned Conjugate Gradient algorithm, Parallel Computing 40, 224 (2014).
[49] P. Ghysels, T. J. Ashby, K. Meerbergen, and W. Vanroose, Hiding Global Com

munication Latency in the GMRES Algorithm on Massively Parallel Machines,
SIAM Journal on Scientific Computing 35, C48 (2013).
[50] K. J. Fidkowski, T. A. Oliver, J. Lu, and D. L. Darmofal, pMultigrid solution of

highorder discontinuous Galerkin discretizations of the compressible Navier
[51] L. Diosady and D. Darmofal, Discontinuous Galerkin Solutions of the Navier

Stokes Equations Using Linear Multigrid Preconditioning, in 18th AIAA Com
putational Fluid Dynamics Conference (Miami, FL, U.S.A., 2007).
[52] B. Helenbrook and H. Atkins, Application of PMultigrid to Discontinuous

Galerkin Formulations of the Poisson Equation, in 17th AIAA Computational
Fluid Dynamics Conference (Toronto, Ontario, 2005).
[53] H. L. Atkins and B. T. Helenbrook, Numerical Evaluation of PMultigrid Method

for the Solution of Discontinuous Galerkin Discretizations of Diffusive Equa
tions, in 17th AIAA Computational Fluid Dynamics Conference (Toronto, On
tario, 2005).
[54] H. Luo, J. D. Baum, and R. Löhner, Fast pMultigrid Discontinuous Galerkin

Method for Compressible Flows at All Speeds, AIAA Journal 46, 635 (2008).
[55] Y. Xia, J. Lou, H. Luo, J. Edwards, and F. Mueller, OpenACC acceleration of

an unstructured CFD solver based on a reconstructed discontinuous Galerkin
method for compressible flows, International Journal for Numerical Methods
in Fluids 78, 123 (2015).
[56] A. Karakus, N. Chalmers, K. Świrydowicz, and T. Warburton, A GPU acceler

ated discontinuous Galerkin incompressible flow solver, Journal of Computa
tional Physics 390, 380 (2019).
32 References
[57] A. C. Kirby and D. J. Mavriplis, GPUAccelerated Discontinuous Galerkin Meth

ods: 30x Speedup on 345 Billion Unknowns, Journal of Computational Physics
(2020), manuscript submitted for publication.
[58] W. A. Xie and G. N. Xi, Fluid flow and heat transfer characteristics of separation
2 and reattachment flow over a backwardfacing step, International Journal of
Refrigeration 74, 177 (2017).
[59] A. Fortin, M. Jardak, J. J. Gervais, and R. Pierre, Localization of Hopf bifur
cations in fluid flow problems, International Journal for Numerical Methods in
Fluids 24, 1185 (1997).
[60] L. Kaiktsis, G. Em Karniadakis, and S. A. Orszag, Unsteadiness and convective

instabilities in twodimensional flow over a backwardfacing step, Journal of
Fluid Mechanics 321, 157 (1996).
[61] D. Barkley, M. G. M. Gomes, and R. D. Henderson, Threedimensional insta
bility in flow over a backwardfacing step, Journal of Fluid Mechanics 473,
167 (2002).
3
Pressure Correction
3.1. Introduction
This chapter treats the temporal discretization of the transport equations 1.1a–
1.1c with a constant density. The test cases at the end of this chapter use the
discontinuous Galerkin method that was described in chapter 2, which provides
additional confidence in the spatial discretization. This chapter can nevertheless be
read independently.
We assume that the boundary of Ω does not depend on 𝑡, so that the spatial
and temporal discretizations can be entirely uncoupled. The more general moving
boundary problem has previously been approached with a spacetime discontinuous
Galerkin method [2–4]. Our independent variables lie in the spacetime cylinder
(𝑡, 𝐫) ∈ (0, 𝑇) × Ω, which suggests time stepping with a simple finite difference
method.
3.1.1. Fully Discrete Linear System

The temporal discretizations for the momentum and the enthalpy are based on stan
dard backwarddifference formulae (BDF). For the mass flux this is straightforward:
for a constant time step size 𝛿𝑡,
𝑞
𝜕𝐦 𝛾0 𝑛 𝛾𝑖
≈ 𝐦 + ∑ 𝐦𝑛−𝑖 , (3.1)
𝜕𝑡 𝛿𝑡 𝛿𝑡
𝑖=1
𝑞
where 𝐦𝑛 is the mass flux at time step 𝑛. The weights {𝛾𝑖 }𝑖=0 are listed in Table
3.1.
If the density depends on the temperature, then the temporal discretization of
the enthalpy equation becomes more involved, as will be explained in great detail in
Parts of this chapter have been published in [1].
33
34 3. Pressure Correction
Table 3.1: Coefficients for the backward difference formula of various orders.
𝛾0 𝛾1 𝛾2 𝛾3
BDF1 1 1
BDF2 3/2 2 1/2
BDF3 11/6 3 3/2 1/3
chapter 4. In this chapter 𝜌 is simply constant, and so the finite difference scheme
3 for 𝜕(𝜌ℎ)/𝜕𝑡 is straightforward.
The fully discrete transport equations can be written as
−𝑫𝒎𝑛 + 𝑪𝒑𝑛 = −𝒓 (3.2a)

𝛾0
𝑴 𝒎𝑛 = −𝑵𝒎𝑛 − 𝑫⊺ 𝒑𝑛 + 𝒇 (3.2b)
𝛿𝑡
𝛾0 𝑛 𝑛
𝑻 𝒉 = −𝑭𝒉 + 𝒒 , (3.2c)
𝛿𝑡
where 𝒑, 𝒎, and 𝒉 are the solution vectors, containing the coefficients of the basis
functions, and
• 𝑴 is the mass matrix (i.e., the Gram matrix of the basis functions);
• 𝑫 corresponds to the divergence operator in Eq. 2.5;
• 𝑪 corresponds to the pressure stabilization in Eq. 2.8, which is zero for mixed
order discretizations;
• 𝑵 contains the implicit terms of the convection and diffusion discretizations
of the momentum equation;
• 𝑭 contains the implicit terms of the convection and diffusion discretizations of
the enthalpy equation;
• (𝛾0 /𝛿𝑡)𝑻 corresponds to the implicit part of the BDF scheme for the enthalpy,
that is, the coefficient of ℎ𝑛 in
𝑞
𝜕(𝜌ℎ) 𝛾𝑖
≈ ∑ (𝜌ℎ)𝑛−𝑖 . (3.3)
𝜕𝑡 𝛿𝑡
𝑖=0
For the constantdensity flow in this chapter, 𝑻 = 𝜌𝑴. For the variabledensity
flow in the next chapter, 𝑻 depends on how (𝜌ℎ)𝑛 is approximated (see section
4.3);
• 𝒓 corresponds to the rhs of the discrete continuity equation (Eq. 2.7);
• 𝒇, and 𝒒 collect various explicit terms, including those from the temporal
discretization, and from the boundary conditions.
3.2. Pressure Correction Method 35
The matrices 𝑴 and 𝑫 and the vector 𝒓 do not depend on the unknowns (𝒑, 𝒎, 𝒉),
but the other terms depend on the fluid properties, and therefore on 𝒉. The most
important nonlinearity is due to 𝑵, which depends on the convective field (1/𝜌)𝑚.
The pressurebased linear system 3.2a–3.2b for the discrete continuity and mo
mentum equations is very stiff, because the equations form a saddle point problem.
Therefore these equations are approximated with a timesplitting method, as ex
plained in the following section.
We use a secondorder BDF2 scheme (𝑞 = 2 in Eqs. 3.1 and 3.3), thereby fol
lowing previous DG literature (e.g., [5–7]). Discontinuous Galerkin methods for the
NavierStokes equations have traditionally been associated with highorder tempo
3
ral accuracy, which can easily be achieved in densitybased formulations that solve
a coupled system of transport equations. However, splitting methods have an in
herent error of order 3/2 in the 𝐻1 norm [8], so that highorder BDF schemes have
no merit.
3.2. Pressure Correction Method

The pressure correction method is used to split the continuity and the momentum
equations, so that they can be solved in a segregated way, which is much cheaper
than a coupled solver. This technique has been thoroughly analyzed in the context of
many spatial discretizations. See, for example, Saleri and Veneziani [9] for analyses
based on LUdecompositions of system of transport equations, and the review by
Guermond et al. [8] for a comparison with the alternatives to pressure correction,
namely velocity correction and consistent splitting methods.
The pressure correction method for an equalorder discontinuous Galerkin dis
cretization differs from the spatial discretizations in the above references, because
of the pressure stabilization term in the continuity equation 3.2a. This couples the
discrete continuity and momentum equations more tightly. This section summarizes
the timesplitting scheme with pressure stabilization.
The momentum equation is solved with a known vector 𝒑̂ instead of the un
known pressure 𝒑𝑛 to obtain a predictor 𝒎̂ for the mass flux:
𝛾0
( 𝑴 + 𝑵) 𝒎̂ = −𝑫⊺ 𝒑̂ + 𝒇 . (3.4)
𝛿𝑡
These are then corrected to find the solutions at the new time step:
𝒑𝑛 = 𝒑̂ + 𝛿𝒑 , (3.5)
𝑛
𝒎 = 𝒎̂ + 𝛿𝒎 . (3.6)
Subtracting Eq. 3.4 from Eq. 3.2b gives ((𝛾0 /𝛿𝑡)𝑴 + 𝑵) 𝛿𝒎 = −𝑫⊺ 𝛿𝒑. The idea
of the pressure correction method is that 𝑵 𝛿𝒎 is of a higher order in 𝛿𝑡 than the
other terms, and therefore it can be neglected, giving
𝛿𝑡 −1 ⊺
𝛿𝒎 = − 𝑴 𝑫 𝛿𝒑 . (3.7)
𝛾0
This approximation means that the momentum equation 3.2b is not exactly satis
fied by the solution pair (𝒑𝑛 , 𝒎𝑛 ), though it can be made to satisfy the continuity
equation 3.2a exactly. Leftmultiply Eq. 3.7 by 𝑫 and use Eq. 3.2a to eliminate
𝑫𝒎𝑛 , to find
𝛿𝑡
𝑪𝒑𝑛 + 𝒓 − 𝑫𝒎̂ = − 𝑫𝑴−1 𝑫⊺ 𝛿𝒑 , (3.8)
𝛾0
which can be rearranged to
𝛾0 𝛾0
3 (𝑨LDG +
𝛿𝑡
𝑪) 𝛿𝒑 =
𝛿𝑡
(−𝒓 + 𝑫𝒎̂ − 𝑪𝒑)
̂ (3.9)
with
𝑨LDG ≔ 𝑫𝑴−1 𝑫⊺ . (3.10)
In the seminal paper on this method, Chorin [10] simply had 𝒑̂ = 0, but it was
shown later that the incremental approach with 𝒑̂ = 𝒑𝑛−1 is more accurate [11], and
that is what we use. It may be tempting to use a higherorder approximation of 𝒑𝑛
for 𝒑,̂ but this is only conditionally stable [8]. If the convection and the diffusion of
the momentum are treated in a timeexplicit manner, then 𝑵 = 0, and the pressure
correction method yields an exact solution to Eqs. 3.2a–3.2b, regardless of the
choice of 𝒑.̂
Splitting the momentum and continuity equations has created the need for ar
tificial initial and boundary conditions for the pressure, which are not present in
the original system of equations (at least not on Dirichlet boundaries). The initial
condition is not a practical problem; the pressure can be inferred from the initial
condition for the mass flux, or the pressure correction method can be iterated within
the first time step. Often the initial transient behavior is not physically relevant, in
which case one can also just set 𝒑0 = 0.
The pressure boundary condition would have been explicit if the pressure correc
tion method had been derived at the differential level from the continuous equations
1.1a–1.1b. This would have resulted in a continuous Poisson equation (∇2 𝑝 = …)
that needs to be supplemented with boundary conditions before it can be dis
cretized. This could be seen as an argument in favor of ‘algebraic’ splitting methods,
which are derived from the coupled discrete linear system 3.2a–3.2b, as was done
above. Nevertheless, even the discrete Poissonlike pressure equation 3.9 contains
implicit artificial boundary conditions, which are inherent to the construction of the
discrete divergence operator 𝑫 in Eq. 2.5.
In fact, Shahbazi et al. [5] have pointed out that 𝑨LDG (Eq. 3.10) is effectively
a local discontinuous Galerkin (LDG) discretization for a diffusion operator with ho
mogeneous Neumann boundary conditions at the walls and the inlet, so it can be
replaced by an SIP diffusion operator
𝑨LDG ≈ 𝑨SIP , (3.11)
which has a smaller stencil (as was discussed in section 2.1.1). Note that 𝑨SIP
is equipped with explicit boundary conditions, and the timesplitting scheme is no
3.2. Pressure Correction Method 37
longer of the algebraic kind. A consequence of using this SIP pressure matrix is that
the continuity equation 3.2a does not hold exactly. In the test cases that follow, we
have not noticed a difference between using LDG and SIP pressure discretizations,
which is in line with previous findings (e.g., [12, pp. 33–45]).
It has long been known that the artificial boundary condition is one of the main
drawbacks of timesplitting methods [13]. It creates a numerical boundary layer
that prevents the pressure from converging with secondorder temporal accuracy
in the 𝐿2 norm. The velocity is secondorder accurate in time in the 𝐿2 norm, but
not in the 𝐻1 norm [14].
This problem of a numerical boundary layer can be remedied with the socalled
3
rotational pressure correction method that was introduced by Timmermans et al.
[15], and popularized by a rigorous error analysis due to Guermond and Shen [14].
They showed that the rotational correction results in a consistent pressure bound
ary condition, and that this improves the orders of convergence for the pressure in
the 𝐿2 norm, and the velocity in the 𝐻1 norm, with the exact orders depending on
the geometry and the type of boundary conditions. See also the extensive numer
ical tests and heuristic explanation in [16]. Piatkowski et al. [17] have combined
rotational pressure correction with a discontinuous Galerkin method for the spatial
discretization.
While these works have demonstrated a substantial improvement of the tempo
ral error, they are based on the assumption of a homogeneous density and viscosity.
Deteix and Yakoubi [18] have shown how to incorporate a variable viscosity, at the
expense of having to solve two poisson equations for every pressure correction
step. We have not extended the rotational pressure correction method to our DG
solver with a variable viscosity.
𝑛
In summary, our algorithm to find the solution vectors 𝒑𝑛 , 𝒎𝑛 , 𝒉 at a new time
step 𝑛 is as follows.
1. Obtain predictors for (𝑘/𝑐𝑝 )∗ , and 𝐦∗ with a secondorder extrapolation from

previous time steps:
(⋅)∗ = 2(⋅)𝑛−1 − (⋅)𝑛−2 . (3.12)
𝑛
2. Solve for the enthalpy 𝒉 at the new time step, using the above predictors
for the diffusion constant ((𝑘/𝑐𝑝 )∗ ) and the advecting field (𝐦∗ ).
If the density depends on the temperature, as in chapter 4, then the implicit
time term is also approximated with a predictor 𝜌∗ = 2𝜌𝑛−1 − 𝜌𝑛−2 , using
either of the methods that are explained in Section 4.3.
̂ The matrix 𝑵 depends on the fluid properties, which are

3. Solve Eq. 3.4 for 𝒎.
evaluated at the new time step as a function of ℎ𝑛 . The convective field is
estimated as (1/𝜌𝑛 )𝐦∗ .
4. Solve the pressure Poisson equation 3.9 for 𝛿𝒑, possibly replacing the LDG
discretization by an SIP method (Eq. 3.11), and correct the pressure and the
mass flux with Eqs. 3.5–3.7.
The manufactured solutions in section 3.3 will show full secondorder temporal
accuracy in the enthalpy and the mass flux, even if the fluid properties are non
trivial functions of the enthalpy.
3.3. Verification with Manufactured Solutions

The numerical method and its implementation are verified with two manufactured
solutions: the wellknown TaylorGreen vortex with constant fluid properties, and
a variableproperty manufactured solution with a wall and an outflow boundary.
3 Our experience with DGFlows has been that the convergence results depend
critically on a careful calculation of the error, which is defined in the 𝐿2 norm as
2
‖𝜙 − 𝜙ex ‖2 ∑𝑇∈𝒯 ∫𝑇 (𝜙 − 𝜙ex )
= √ 2 (3.13)
‖𝜙ex ‖2 ∫ (𝜙ex ) Ω
for a quantity 𝜙 with an exact solution 𝜙ex . Each integral in the numerator is evalu
ated with a numerical quadrature, resulting in a large sum over the squares of small
numbers. A naive implementation gives very large rounding errors. We therefore
perform the double summation over the elements and the quadrature points with
the Kahan summation algorithm [19] and a 128bit floating point number.
All integrals in the weak forms are evaluated with a quadrature set with the
usual polynomial accuracy of (3𝒫𝑚 − 1), and we verified that this is sufficient to
integrate up to machine precision by comparing the results with a higherorder
quadrature set of polynomial accuracy (3𝒫𝑚 + 10). This is not surprising, since the
TaylorGreen vortex solution in section 3.3.1 is smooth, whereas the exact solution
in section 3.3.2 is a polynomial, and so are the corresponding forcing terms.
3.3.1. TaylorGreen Vortex

The first manufactured solution is the TaylorGreen Vortex, which is incompressible
and has constant fluid properties. We include a passive scalar temperature field with
this wellknown analytical solution. The enthalpy is ℎ = 𝑐𝑝 𝑇. The exact solution is
− cos (𝑥)̃ sin (𝑦) ̃

𝐮ex = exp (−2𝑡)̃ [ ] ,
+ sin (𝑥)
̃ cos (𝑦)̃
𝜌 (3.14)
𝑝ex = − exp (−4𝑡)̃ (cos (2𝑥) ̃ + cos (2𝑦))̃ ,
4
𝑇ex = exp (−2𝑡/Pr)
̃ cos (𝑥) ̃ cos (𝑦) ̃ ,
on a domain 𝑥, 𝑦 ∈ [−𝐿, 𝐿] with Dirichlet boundary conditions and 0 < 𝑡 ≤ 1, where

𝜈𝑡 𝑥 𝑦
𝑡̃ ≔ 2 , 𝑥̃ ≔ , 𝑦̃ ≔ , (3.15)
(𝐿/(𝑛𝜋)) 𝐿/(𝑛𝜋) 𝐿/(𝑛𝜋)
and 𝑛 must be a positive integer in order for ∫Ω 𝑝 = 0. This solves the transport
equations with 𝐅 = [0, 0] and 𝑄 = 0.
3.3. Verification with Manufactured Solutions 39
Fig. 3.1 shows the temporal convergence for 𝐿 = 1, 𝑛 = 1, 𝜇 = 0.01, 𝜌 =

1, Pr = 100, and a fourthorder polynomial space for the mass flux (i.e., 𝒫𝑚 =
4). We performed the same numerical experiments by independently varying the
Prandtl number (to Pr = 1), and the polynomial order (to 𝒫𝑚 = 2), all of which
yielded similar results. All errors saturate at small time steps, where the spatial
discretization error dominates.
Table 3.2 shows the spatial convergence at the smallest time step that was
tested. It also includes equalorder and mixedorder results for a polynomial orders.
All quantities of polynomial order 𝒫 appear to converge as 𝒪 (𝓁𝒫+1 ), where 𝓁 ∝
1/𝑁𝑦 is the characteristic mesh length, except for the equalorder cases, where the 3
order of convergence for the pressure is in the range [𝒫 + 1/2, 𝒫 + 1]. The spatial
convergence rates for the highorder polynomial cases is harder to make out from
the available data, because the temporal error is still significant, as could also be
seen in Fig 3.1.
The TaylorGreen vortex is of course a strange test case, in that it does not fea
ture any scale separation: its Fourier transform is comprised of Dirac delta functions.
The solution is an eigenfunction of the diffusive terms, so there is no interaction
between the transport terms. Furthermore, it has Dirichlet boundary conditions
where there is outflow, so that the continuity equation is overconstrained. This
manifests itself in a stiff linear system for the pressure, though its ubiquity in the
literature suggests that the TaylorGreen vortex is very easy to simulate. The next
section features a more challenging manufactured solution.
3.3.2. Variableproperty Manufactured Solution

This section features a manufactured solution with temperaturedependent trans
port properties, so that the momentum and enthalpy transport equations are cou
pled. It has walls and an outflow boundary condition; contrary to the TaylorGreen
vortex, it is wellposed. The domain is (0, 𝐿)×(−1, 1); see Fig. 3.2. We let 𝐿 = 10 in
all calculations, and use square elements, so that there are 𝑁𝑥 = (𝐿/2)𝑁𝑦 elements
in the 𝑥direction. The inflow boundary at 𝑥 = 0 has Dirichlet boundary conditions.
The goal of the manufactured solutions is obviously not to model a particular phys
ical phenomenon, but our configuration is vaguely reminiscent of a pipe flow with
walls at 𝑦 = ±1 that is heated asymmetrically, resulting in skewed velocity profiles.
We use a Neumann boundary condition for the temperature at the outlet, be
cause this is the most common choice in practical applications. The imposed heat
flux (𝑞N ) follows from the known exact solution. We also tried imposing a Dirich
let boundary condition for the temperature, and found that it makes a negligible
difference in the numerical errors.
We choose the polynomial manufactured solution
1 𝑦 (2𝑥 3 /3 − 𝐿𝑥 2 ) 2
𝐦ex = (1 + 𝑡3 ) [ ]+[ ] ,
𝐿3 (𝑦 − 1)(𝑦 + 1)𝑥(𝐿 − 𝑥) 0
(3.16)
𝑝ex = (1 + 𝑡3 ) (𝐿 − 𝑥)3 ,
𝑇ex = (1 + 𝑡3 ) (2 − 𝑦)((𝑥 − 𝐿)/𝐿)2 /6
with 0 < 𝑡 ≤ 1, which satisfies 𝑚2 = 0 on 𝑦 = ±1, and ∇ ⋅ 𝐦 = 0. The addition
Pm = Pp = Ph = 4 Pm = 4, Pp = Ph = 3
10−1 10−1
10−2
kT − T exk2 / kT exk2
10−2
3 10−3 10−3
−4
10
10−4
10−5
10−5
−6
10
10−6
22 23 24 25 26 27 28 29 210 211 212 213 22 23 24 25 26 27 28 29 210 211 212 213
10−2 10−2
ku − uexk2 / kuexk2
10−4 N = 64 10−4
N = 32
N = 16
10−6 N =8 10−6
N =4
22 23 24 25 26 27 28 29 210 211 212 213 22 23 24 25 26 27 28 29 210 211 212 213

101
100
kp − pexk2 / kpexk2
10−1
10−2
10−3
10−4
10−5
22 23 24 25 26 27 28 29 210 211 212 213 22 23 24 25 26 27 28 29 210 211 212 213

1 / δt 1 / δt
Figure 3.1: Convergence toward the 2D Taylor vortex (Eq. 3.14) at time 𝑡̃ = 1 with temporal refinement
for meshes with 𝑁2 square elements. The black dashed lines indicate ideal secondorder convergence
in 𝛿𝑡.
3.3. Verification with Manufactured Solutions 41
Table 3.2: Convergence toward the TaylorGreen vortex in Eq. 3.14 with spatial refinement, keeping 𝛿𝑡
fixed. The data in the last two blocks are taken from the highest temporal refinement in Fig. 3.1.
temperature velocity pressure

𝑁𝑦 error conv error conv error conv 3
Equal order (𝒫𝑚 = 𝒫ℎ = 𝒫𝑝 = 1), 𝛿𝑡 = 2−12 :
22 2.93e1 1.13e0 1.17e0
23 1.02e1 1.53 5.25e1 1.10 1.34e0 0.19
24 2.73e2 1.90 1.59e1 1.73 5.84e1 1.20
25 6.72e3 2.02 4.14e2 1.94 1.80e1 1.69
26 1.66e3 2.02 1.04e2 1.99 5.34e2 1.76
27 4.11e4 2.01 2.60e3 2.00 1.69e2 1.66
28 1.02e4 2.01 6.50e4 2.00 6.05e3 1.48
Mixed order (𝒫𝑚 = 2, 𝒫ℎ = 𝒫𝑝 = 1), 𝛿𝑡 = 2−12 :
22 2.70e1 2.57e1 2.20e0
23 7.25e2 1.90 2.13e2 3.59 4.01e1 2.46
24 1.70e2 2.09 1.62e3 3.72 5.43e2 2.88
25 4.14e3 2.03 1.49e4 3.44 8.67e3 2.65
26 1.04e3 2.00 1.64e5 3.18 1.71e3 2.34
27 2.61e4 1.99 1.97e6 3.06 3.86e4 2.15
28 6.59e5 1.99 3.25e7 2.60 9.28e5 2.06
Mixed order (𝒫𝑚 = 4, 𝒫ℎ = 𝒫𝑝 = 3), 𝛿𝑡 = 2−13 :
22 1.54e2 3.79e3 1.13e1
23 1.09e3 3.82 9.87e5 5.26 3.82e3 4.88
24 6.17e5 4.15 2.83e6 5.12 1.37e4 4.80
25 3.75e6 4.04 1.04e7 4.77 6.40e6 4.42
26 4.98e7 2.91 5.80e8 0.84 1.57e6 2.02
Equal order (𝒫𝑚 = 𝒫𝑝 = 𝒫ℎ = 4), 𝛿𝑡 = 2−13 :
22 2.73e3 3.47e3 8.27e2
23 9.56e5 4.84 9.46e5 5.20 3.15e3 4.71
24 5.24e6 4.19 2.79e6 5.08 1.25e4 4.65
25 6.55e7 3.00 1.03e7 4.76 1.19e5 3.40
26 2.49e7 1.40 5.79e8 0.83 6.51e6 0.86
(0,1) (L,1)
Ny elements
y
x
(0,-1) (L,-1)
3 Figure 3.2: Domain of the manufactured solutions in Sections 3.3.2 and 4.5.1.
of the constant [2, 0] to 𝐦ex ensures that 𝑚1 > 0 everywhere, so that there is no
backflow at the outlet, and no outflow at any of the Dirichlet boundary conditions.
The density is 𝜌 = 1, but the transport properties are nontrivial: 𝜇 = 0.1 + 𝑇(1 − 𝑇)
and 𝑘 = 𝜇𝑐𝑝 /Pr, with Pr = 1. The solution is depicted in Fig. 3.3.
Fig. 3.4 displays the temporal convergence. We consider various meshes, vary
ing the number of elements and the spatial polynomial orders. The equalorder
case does not appear to suffer from the infsup instability for small 𝛿𝑡. The velocity
and the temperature converge with secondorder accuracy in 𝛿𝑡 until the error sat
urates when the spatial error starts to dominate the temporal error. The order of
convergence for the pressure is slightly lower, in the range [1.5,2.0]. These orders
of convergence for the velocity and pressure agree with what is found in previous
literature on constantproperty incompressible flows (e.g., [5], [20]).
The spatial rates of convergence are in Table 3.3. As the mesh is refined, the
mixedorder discretization displays 𝒪 (𝓁𝒫+1 ) convergence for all quantities of poly
nomial order 𝒫, though the convergence rate of the velocity saturates at high spa
tial refinement, as the temporal error starts to become significant. The equalorder
discretization has the same convergence rates, meaning that the velocity shows
hyperconvergence, with the error in 𝑢 behaving as 𝒪 (𝓁𝒫𝑢 +2 ).
3.4. Validation with Flow Past a Circular Obstacle

We computed laminar flow past a circular cylinder to validate our numerical method.
This features a Von Kármán vortex street in the wake of the obstacle. The results
of this wellknown benchmark case can be compared to experiments and to other
direct numerical simulations.
Fig. 3.5 shows the computational domain. The velocity is fixed at [𝑢∞ , 0] on
the left, top, and bottom parts of the domain. A subscript ∞ denotes a farfield
value. The right side (at 𝑥 = 𝐿) is an outlet with homogeneous Neumann boundary
conditions (i.e., 𝑞N = 0 and 𝐟N = 𝟎). The cylinder has a noslip boundary condition.
We use the farfield values to define a Reynolds number Re ≔ 𝐷 (𝜌𝑢/𝜇)∞ .
The initial condition requires special care. First, it is not easy to find an initial
velocity field that satisfies 𝐧⋅𝐮 = 𝐧⋅𝐮D at the walls, inlet, and cylinder. We therefore
initialize the velocity to [𝑢∞ , 0], and leave out the convective term in the first 10
time steps, thus essentially simulating Stokes flow, which does not have the same
3.4. Validation with Flow Past a Circular Obstacle 43
1.33 1.5 1.67 1.83 2 2.17 2.33 2.5 2.67
(a) 𝑢1 .
-0.05 -0.0438 -0.0375 -0.0312 -0.025 -0.0188 -0.0125 -0.00625 0
(b) 𝑢2 .
0 0.0312 0.0625 0.0937 0.125 0.156 0.187 0.219 0.25
(c) Kinematic viscosity (𝜈) and thermal diffusivity (𝛼). (Note that Pr = 1, so 𝛼 = 𝜈.)
Figure 3.3: Constantdensity manufactured solution in Eq. 3.16 at 𝑡 = 1.

Pm = Pp = Ph = 1 Pm = 2, Pp = Ph = 1
−1
10
10−1
10−2
10−2
3
10−3 10−3
10−4 10−4
10−5 10−5
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212
100 100
10−1 10−1
Ny = 64 10−2
10−2
Ny = 32
Ny = 16 10−3
10−3
Ny =8
Ny =4 10−4
10−4
Ny =2
10−5
10−5
2 3 4 5 6 7 8 9 10 11 12
2 2 2 2 2 2 2 2 2 2 2 22 23 24 25 26 27 28 29 210 211 212
10−1 10−1
10−2 10−2
10−3 10−3
10−4 10−4
10−5 10−5
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212

1 / δt 1 / δt
Figure 3.4: Convergence of the numerical solution toward the constantdensity manufactured solution
(Eq. 3.16) with temporal refinement. The characteristic element length is inversely proportional to 𝑁𝑦 .
The black dashed lines indicate ideal secondorder convergence in 𝛿𝑡.
3.4. Validation with Flow Past a Circular Obstacle 45
Table 3.3: Convergence toward the constantdensity manufactured solution in Eq. 3.16 (Fig. 3.3) with
spatial refinement and fixed 𝛿𝑡 = 2−12 .

𝑁𝑦 error conv error conv error conv
Equal order (𝒫𝑚 = 𝒫𝑝 = 𝒫ℎ = 1):
21 2.67e2 1.82e1 3.75e3 3
22 3.26e3 3.03 3.58e2 2.35 8.66e4 2.11
23 6.70e4 2.29 5.81e3 2.62 2.14e4 2.02
24 1.66e4 2.01 8.46e4 2.78 5.34e5 2.00
25 4.15e5 2.00 1.16e4 2.87 1.33e5 2.00
26 1.04e5 2.00 1.57e5 2.89 3.34e6 2.00
Mixed order (𝒫𝑚 = 2, 𝒫𝑝 = 𝒫ℎ = 1):
21 1.08e2 1.11e1 3.62e3
22 2.65e3 2.03 1.83e2 2.61 8.75e4 2.05
23 6.63e4 2.00 2.62e3 2.80 2.15e4 2.02
24 1.66e4 2.00 3.52e4 2.90 5.35e5 2.01
25 4.15e5 2.00 4.57e5 2.94 1.34e5 2.00
26 1.04e5 2.00 6.43e6 2.83 3.34e6 2.00
L0 L
H
y
D
H
Figure 3.5: Geometry of flow past a circular cylinder (not to scale). The cylindrical obstacle is centered
at the origin.
requirement for the initial condition at the boundaries. Due to the instant smooth
ing property of the viscous operator, the velocity satisfies the Dirichlet boundary
conditions when the convective term is ‘activated’ after the first 10 time steps. This
causes an instantaneous change in the pressure, but that is neither unphysical in
the incompressible flow limit, nor is it a problem for the pressure correction method.
Second, we found that the combination of a symmetrical mesh and a symmetrical
initial condition does not induce vortex shedding. This is addressed by letting the
cylinder rotate counterclockwise for the first 100 time steps.
We are interested in the force on the cylinder 𝑆, which is given by
3
𝐅(cyl) = − ∫ (𝜏 − 𝑝 𝐼) ⋅ 𝐧 , (3.17)
𝜕𝑆
where 𝐧 is the outward normal of the fluid, pointing into the cylinder. The drag and
the lift coefficients are
(cyl) (cyl)
2 𝐹1 2 𝐹2
𝐶D = and 𝐶L = (3.18)
𝐷 (𝜌𝑢2 )∞ 𝐷 (𝜌𝑢2 )∞
respectively. At our Reynolds number, the flow is laminar, and the force oscillates
in time in a smooth, deterministic manner. This makes it easy to determine the
frequency 𝑓 of the lift coefficient, and the corresponding Strouhal number St ≔
𝑓𝐷/𝑢∞ .
Our numerical experiments indicate that a small domain results in an overesti
mation of the Strouhal number. This likely explains the large discrepancy in the nu
merical predictions of the Strouhal number in previous literature; see Niroobakhsh
et al. [21] for an overview. Collis [22] showed that a domain of (𝐻, 𝐿0 , 𝐿) =
(30𝐷, 15𝐷, 30𝐷) was sufficient for isothermal flow at Re = 100 and a Mach num
ber of 0.2. The results presented here were obtained on a domain of (𝐻, 𝐿0 , 𝐿) =
(40𝐷, 20𝐷, 40𝐷). The domain could probably be substantially smaller if the upper
and lower part were connected with periodic boundary conditions, but DGFlows
did not support this yet when the calculations were done.
Fig. 3.6 shows the mesh. It has approximately 27k elements with 𝒫𝑚 = 2 and
𝒫𝑝 = 𝒫ℎ = 1, resulting in approximately 160k degrees of freedom per direction of
the mass flux, and 64k degrees of freedom for 𝑝 and ℎ. There are 120 boundary
elements at the cylinder, each with a width of 0.005𝐷. The time step is given by
𝐷/(𝛿𝑡 𝑢∞ ) = 82. We find St = 0.166. This number remained unchanged when we
decreased the number of elements to 19k, or when we doubled 𝛿𝑡. It compares
well with the experimental values of St = 0.165 in [23], St = 0.165 in [24], and
St = 0.167 in [25, p. 71].
3.5. Discussion
The tests with the manufactured solutions in section 3.3 show secondorder tempo
ral accuracy in the 𝐿2 norm for the velocity and the temperature. This observation
is based on simulations on many levels of spatial discretization, and time step sizes
3.5. Discussion 47
Z X
(a) overview (bottom not shown)
Y
Z
(b) Detail near the cylinder.
Figure 3.6: Mesh for flow past an obstacle. It is structured near the cylinder and most of its wake. The
rest of the mesh is unstructured to allow for large differences in the element size. It is symmetrical
about the axis 𝑦 = 0.
that span many orders of magnitude (ref. Figs. 3.1 and 3.4). Even the pressure is
not far from secondorder accurate in 𝛿𝑡.
An important and obvious improvement is the rotational formulation of the pres
sure correction method, as explained in section 3.2. Even though this requires two
poisson solves per iteration in the case of a nonhomogeneous viscosity, previous
works suggests that the improved accuracy would be worth the cost.
3.5.1. Pressure Correction with Equalorder Discretizations

3 The manufactured solutions show that the equalorder discretization is stable in the
limit of small time steps. To achieve this, the pressure Poisson equation 3.9 has
been modified with a penalty matrix 𝑪.
Unfortunately this stabilization makes the linear system more expensive. This
is in part because adding penalization raises the condition number, which can be
−1
substantial in turbulent flow, given the 𝛿𝑡 scaling in Eq. 3.9, and the 𝜈−1 scaling
of the penalty parameter in Eq. 2.9. Perhaps more importantly, 𝑪 depends on the
viscosity, and so the pressure matrix is generally not constant, which has a large
impact on the total computation time in our implementation DGFlows.
A large penalty parameter may be less detrimental to the efficiency of other
solvers if they are suitably preconditioned. For example, Fehn et al. [26] have
investigated several 𝑝multigrid methods for a DGSIP discretization of the poisson
equation with a matrixfree implementation. They found that the most effective
approach is to project the finest DG grid onto a continuous Galerkin space, which
can in turn be coarsened with lowerorder continuous solution spaces. The resulting
solver was robust with respect to the SIP penalty parameter.
It should also be noted that the pressure stabilization 𝑪 was developed for steady
Stokes flow, and it is overly restrictive for timedependent convecting flow. In
our transient calculations, there are already two effects that implicitly stabilize the
pressure. These could be used to reduce the penalty parameter in Eq. 2.9, or
possibly to leave it out entirely.
The first stabilizing effect is inherent to the pressure correction method. The
error due to timesplitting methods can be shown to be equivalent to perturbing the
continuity equation with a pressure diffusion term (∇2 𝑝) that is proportional to 𝛿𝑡.
This means that the infsup instability does not manifest itself at large time steps,
as has been well documented for several spatial discretizations (e.g., [8, 27, 28]).
Ferrer et al. [20, 29] have provided estimates for the minimal time step size in a
DG discretization, and showed that this can still be higher than the maximum time
step that satisfies the CFL condition.
The other source of pressure stability is due to the SIP discretization for the
pressure matrix (Eq. 3.11), which contains a penalty term with a penalty parameter
as in Eq. 2.16 with a diffusion parameter 𝐾 = 1. Shahbazi [30, pp. 48–65] has
successfully used the SIP pressure matrix with an equalorder discretization without
extra pressure stabilization (i.e., 𝑪 = 0). Our tests (not shown here) also indicate
that, for equalorder discretizations without pressure stabilization, the LDG pressure
matrix is unstable for all reasonable time steps, whereas using the SIP matrix is
feasible for a wide range of practical time step sizes, though it always becomes
References 49
unstable in the limit 𝛿𝑡 → 0.

This provides an extra incentive for replacing the pressure LDG matrix by an
SIP matrix (Eq. 3.11). It was originally motivated by its smaller stencil and lower
condition number [5], but is could well be that its greater stability is more important.
References
tational Physics 425, 109877 (2021). 3
[2] C. M. Klaij, J. J. W. van der Vegt, and H. van der Ven, Spacetime discontin
uous Galerkin method for the compressible NavierStokes equations, Journal
of Computational Physics 217, 589 (2006).

[4] V. Ambati and O. Bokhove, Spacetime discontinuous galerkin discretization

of rotating shallow water equations, Journal of Computational Physics 225,
1233 (2007).



634 (2017).
[8] J. L. Guermond, P. Minev, and J. Shen, An overview of projection methods

for incompressible flows, Computer Methods in Applied Mechanics and Engi
neering 195, 6011 (2006).
[9] F. Saleri and A. Veneziani, Pressure Correction Algebraic Splitting Methods for
the Incompressible NavierStokes Equations, SIAM Journal on Numerical Anal
ysis 43, 174 (2005).
[10] A. J. Chorin, A numerical method for solving incompressible viscous flow prob
lems, Journal of Computational Physics 2, 12 (1967).
[11] M. O. Henriksen and J. Holmen, Algebraic Splitting for Incompressible Navier
Stokes Equations, Journal of Computational Physics 175, 438 (2002).
50 References
[12] B. Klein, A highorder Discontinuous Galerkin solver for incompressible and

lowMach number flows, Ph.D. thesis, Technische Universität, Darmstadt
(2015).
[13] P. M. Gresho and R. L. Sani, On pressure boundary conditions for the incom
pressible NavierStokes equations, International Journal for Numerical Meth
ods in Fluids 7, 1111 (1987).
[14] J. L. Guermond and J. Shen, On the error estimates for the rotational pressure
3 correction projection methods, Mathematics of Computation 73, 1719 (2004).
[15] L. J. P. Timmermans, P. D. Minev, and F. N. Van de Vosse, An approximate
projection scheme for incompressible flow using spectral elements, Interna
tional Journal for Numerical Methods in Fluids 22, 673 (1996).
[16] J. Aoussou, J. Lin, and P. F. J. Lermusiaux, Iterated pressurecorrection
projection methods for the unsteady incompressible NavierStokes equations,
[18] J. Deteix and D. Yakoubi, Improving the pressure accuracy in a projection
scheme for incompressible fluids with variable viscosity, Applied Mathematics
Letters 79, 111 (2018).
[19] Wikipedia contributors, Kahan summation algorithm, (2020), Wikipedia, The
Free Encyclopedia. [Online] Accessed on 20200914.
[20] E. Ferrer, D. Moxey, R. H. J. Willden, and S. J. Sherwin, Stability of Pro
jection Methods for Incompressible Flows Using High Order PressureVelocity
Pairs of Same Degree: Continuous and Discontinuous Galerkin Formulations,
Communications in Computational Physics 16, 817 (2014).
[21] Z. Niroobakhsh, N. Emamy, R. Mousavi, F. Kummer, and M. Oberlack, Numer
ical investigation of laminar vortex shedding applying a discontinuous Galerkin
finite element method, Progress in Computational Fluid Dynamics, An Inter
national Journal 17, 131 (2017).
[23] A.B. Wang, Z. Trávníček, and K.C. Chia, On the relationship of effective
Reynolds number and Strouhal number for the laminar vortex shedding of a
heated circular cylinder, Physics of Fluids 12, 1401 (2000).
[24] C. H. K. Williamson, Oblique and parallel modes of vortex shedding in the wake
of a circular cylinder at low Reynolds numbers, Journal of Fluid Mechanics 206,
579 (1989).
References 51
[25] A. Roshko, On the development of turbulent wakes from vortex streets, Phd
thesis, California Institute of Technology (1952).
[26] N. Fehn, P. Munch, W. A. Wall, and M. Kronbichler, Hybrid multigrid methods
for highorder discontinuous Galerkin discretizations, Journal of Computational
Physics 415, 109538 (2020).
[27] J.L. Guermond and L. Quartapelle, On stability and convergence of projec
tion methods based on pressure Poisson equation, International Journal for
Numerical Methods in Fluids 26, 1039 (1998). 3
[28] N. Fehn, W. A. Wall, and M. Kronbichler, On the stability of projection methods
for the incompressible NavierStokes equations based on highorder discon
tinuous Galerkin discretizations, Journal of Computational Physics 351, 392
(2017).
[29] E. Ferrer and R. H. J. Willden, A high order Discontinuous Galerkin Finite El
ement solver for the incompressible NavierStokes equations, Computers &
Fluids 46, 224 (2011).
[30] K. Shahbazi, A Parallel HighOrder Discontinuous Galerkin Solver For the Un
steady Incompressible NavierStokes Equations in Complex Geometries, Phd
thesis, University of Toronto (2007).
Handling the Enthalpy
4
Equation for LowMach
Number Flow
4.1. Introduction
4.1.1. In Between Compressible and Incompressible
Several lowspeed flows of practical importance are compressible, that is, the veloc
ity is not divergencefree. This can occur due to mixing, or due to a temperature
dependent density near a heat source. An example is heat transfer in lowMach
number flows of supercritical fluids, where all fluid properties vary strongly with the
temperature, but do not depend significantly on the pressure. Most flow solvers
use either a pressurebased approach and assume a divergencefree velocity field,
or a fully compressible (densitybased) formulation. Neither of these methods is
directly applicable to compressible flows in the lowMach number limit.
Densitybased solvers can be used to simulate zeroMach flows by approximating
the flow with a low, nonzero Mach number (e.g., [2], [3]). This has often been
used for heat transfer in supercritical fluids at low speeds (e.g., [4], [5]). This is
expensive for several reasons. First, the temporal discretization needs to resolve
acoustic effects, and the resulting linear systems tend to be very stiff. Second, the
system of transport equations is solved in a coupled way, which is more expensive
than using a timesplitting method, though the performance may be improved with
suitable preconditioning [6]. Finally, the fluid properties are evaluated as a function
of two thermodynamic variables (usually the density and the volumetric enthalpy),
so that a spline interpolation costs far more memory, thus complicating massively
parallel calculations [4].
Parts of this chapter have been published in [1]
53
54 4. Handling the Enthalpy Equation for LowMach Number Flow
There is also substantial experience with discontinuous Galerkin (DG) discretiza

tions for incompressible flows. These are either based on the introduction of artifi
cial compressibility (e.g. [7, 8]), or they solve for the pressure (e.g. [9–12]). The
artificial compressibility method can be more than secondorder accurate in time,
though it requires the system of transport equations to be solved in a coupled man
ner (e.g., [7, 13]). By choosing entropy variables as the unknowns, the DG method
can also be formulated in a general way for both compressible and incompressible
flows, at the cost of great complexity (e.g., [14]). There is, however, almost no
literature on solving the lowMach number equations with a pressurebased discon
tinuous Galerkin method, as is done in this thesis.
The only previous work of which we are aware is by Klein et al. [15, 16], who
used a SIMPLE scheme to march the transport equations forward in time, iterating
the equations within each time step. This required underrelaxation in order for
4 the iteration to converge. They solved for the velocity, so that a predictor for the
density is needed in the temporal derivative of the momentum equation.
We avoid this by solving for the mass flux rather than the velocity. Another
advantage of this approach is that the divergence term in the continuity equation
does not have to be weighed by the density, so that the divergence matrix does
not depend on the density. This makes the transport equations less tightly coupled,
and it simplifies the pressure correction method, because the pressure matrix is
constant for each time step.
4.1.2. Which Enthalpy Equation Should be Solved?

Another important question is which form of the enthalpy transport equation should
be solved (primitive or conservative), and for which variable (the primitive ℎ or
conserved 𝐻 = 𝜌ℎ).
Solving for the specific enthalpy ℎ from the primitive transport equation, that is,
Dℎ 𝜕ℎ 1 𝑘 1
≔ + 𝐮 ⋅ ∇ℎ = ∇ ⋅ ( ∇ℎ) + 𝑄 , (4.1)
D𝑡 𝜕𝑡 𝜌 𝑐𝑝 𝜌
would pose two possible problems. First, the diffusive term is not in the standard
form ∇ ⋅ (𝛼∇ℎ) where 𝛼 ≔ 𝑘/ (𝜌𝑐𝑝 ) is the thermal diffusivity. The difference is
1 𝑘 1 𝑘
∇ ⋅ ( ∇ℎ) − ∇ ⋅ (𝛼∇ℎ) = − (∇ ) ⋅ ∇ℎ = 𝛼 𝐝 ⋅ ∇ℎ , (4.2)
𝜌 𝑐𝑝 𝜌 𝑐𝑝
where
1 𝜌ℎ
𝐝≔ ∇𝜌 = ∇ℎ (4.3)
𝜌 𝜌
is the relative gradient of the density (as in Eq. 2.13). Perhaps the commutation
error in Eq. 4.2 can be neglected in many flows. It is also possible that the diffusion
term in Eq. 4.1 can be discretized by a standard interior penalty method with a
diffusion parameter of 𝑘/𝑐𝑝 , except that the test function is weighed by 1/𝜌.
A more fundamental problem with the primitive transport equation 4.1 is posed
by the advection term. It is not in conservative form, as in Eq. 2.18, and so we do
4.1. Introduction 55
not know how to discretize it with a DG method. Perhaps it could be split as
𝐮 ⋅ ∇ℎ = ∇ ⋅ (𝐮 ℎ) − ℎ∇ ⋅ 𝐮 . (4.4)
The first term on the right is in standard form, whereas the second term resem
bles the divergence of the mass flux in the continuity equation, but we are not
sure how to proceed from there. We therefore solve the transport equation in the
conservative form 1.1c.
For highMach number compressible flows, the unknown is normally taken to
be a conserved variable, such as the volumetric enthalpy 𝐻 ≔ 𝜌ℎ. When written in
that variable, the transport equation becomes
𝜕𝐻 1 𝑘 𝐻
+ ∇ ⋅ ( 𝐦 𝐻) = ∇ ⋅ ( ∇ ( )) + 𝑄 . (4.5)
𝜕𝑡 𝜌 𝑐𝑝 𝜌 4
The diffusion term is again not in the standard form (due to the factor of 1/𝜌), in
direct analogy to when the viscous stress tensor is expressed in terms of a conserved
variable (𝑚), as was discussed in section 2.3.1. The generalization of the symmetric
interior penalty method for the viscous stress could also be applied here. That is,
one starts from a standard discretization for ∇ ⋅ (𝛼∇𝐻), and wherever the stress
𝛼∇𝐻 would appear in the weak form, it is replaced by 𝛼 (∇𝐻 + 𝐻𝐝).
Incidentally, there is a curious alternative interpretation of the effect of the vari
able density on the thermal diffusion, namely that it results in an extra advective
flux. Substituting Eq. 4.2 into Eq. 4.1 gives
𝜕ℎ 1
+ 𝐮+
eff ⋅ ∇ℎ = ∇ ⋅ (𝛼∇ℎ) + 𝑄 , (4.6)
𝜕𝑡 𝜌
where
𝐮±
eff = 𝐮 ∓ 𝛼𝐝 = 𝐮 ± 𝛽𝛼∇𝑇 (4.7)
is the effective advecting velocity, and 𝛽 ≔ −(1/𝜌)𝜌𝑇 is the thermal expansibil
ity. Similarly, the conservative transport equation 4.5 has a diffusion term with
commutation error
𝑘 𝐻
∇⋅( ∇ ( )) − ∇ ⋅ (𝛼∇𝐻) = −∇ ⋅ (𝐻𝛼𝐝) = ∇ ⋅ (𝐻𝛼𝛽∇𝑇) , (4.8)
𝑐𝑝 𝜌
so that
𝜕𝐻
+ ∇ ⋅ (𝐮−
eff 𝐻) = ∇ ⋅ (𝛼∇𝐻) + 𝑄 . (4.9)
𝜕𝑡
(Note the different signs of the advection correction ±𝛼𝛽∇𝑇 in Eqs. 4.6 and 4.9.)
This observation is probably only of theoretical significance, though the equa
tions with an adjusted advecting velocity could also be used for the discretization.
The extra term 𝛽𝛼∇𝑇 can treated explicitly (i.e., extrapolated from previous time
steps). Since (𝐮eff 𝐻 − 𝛼∇𝐻) ⋅ 𝐧 = (𝐮 𝐻 − (𝑘/𝑐𝑝 ) ∇ℎ) ⋅ 𝐧, the boundary contributions
in the weak forms are the same as for the regular equation 1.1c. Our brief tests
(not shown here) suggest that this is feasible, and secondorder temporal accuracy
for the enthalpy is maintained in the 𝐿2 norm.
Coming back to the question of which equation to solve for which unknown,
the main problem with the volumetric enthalpy is that it is not unique to a par
ticular thermodynamic state, which will be a recurring theme in this chapter. The
combination of the specific enthalpy ℎ (or, equivalently, the temperature) and the
thermodynamic pressure 𝑝th fixes the thermodynamic state, but the pair (𝐻, 𝑝th )
does not always do this, making 𝐻 an inconvenient thermodynamic variable. This
problem is not encountered by compressible flow solvers at high Mach numbers,
because they solve a transport equation for both 𝐻 and 𝜌, which implies a specific
enthalpy ℎ from which the fluid properties can be derived.
For example, consider an ideal gas with a specific gas constant 𝑟:
4 𝑝th = 𝑟𝜌𝑇 , (4.10)
for which 𝜌𝑇 = −𝜌/𝑇, and thus 𝐻ℎ = 𝜌 + ℎ𝜌𝑇 /𝑐𝑝 = 𝜌 (1 − ℎ/ (𝑐𝑝 𝑇)). If ℎ = 𝑐𝑝 𝑇,

as is sometimes assumed, then 𝐻𝑇 = 𝑐𝑝 𝐻ℎ = 0, meaning that the value of 𝐻 says
nothing about the temperature.
Another example can be found in fluids at a supercritical temperature. If the
temperature is increased from a point close to a liquid state, then the volumetric
enthalpy will rise initially, but then it will drop when the socalled Widom line is
crossed, where the thermal expansibility peaks [17]. In other words 𝐻𝑇 switches
sign, and thus 𝐻 does not uniquely determine 𝑇.
Solving for the volumetric enthalpy would therefore require some special treat
ment to determine the thermodynamic state in lowMach number flows. One possi
bility is to offset the specific enthalpy by a suitable constant, which would produce a
onetoone relation between the temperature and the new volumetric enthalpy, as
will be explained in section 4.3 (see Fig. 4.1 in particular). Another idea would be
to obtain a separate predictor for the density, for example by solving the continuity
equation, so that the specific enthalpy can be estimated.
Peeters [18] has claimed to have taken this approach, though we note that
the nondimensionalization of the variables in that work also involved offsetting the
specific enthalpy by a constant. It is possible that this created a onetoone rela
tionship between the nondimensional ℎ and the nondimensional 𝐻, in which case
they would have inadvertently solved the nonuniqueness of 𝐻 even before the
density predictor was introduced.
We avoid these difficulties by solving for the specific enthalpy ℎ from the con
servative transport equation 1.1c, which poses another problem that needs to be
overcome. Solving a conservative transport equation for a primitive variable com
plicates the temporal derivative: the term 𝜕(𝜌ℎ)/𝜕𝑡 can be an important source of
error and instability when 𝜌 is a function of ℎ. A similar issue occurs in multispecies
transport, and Najm et al. [19] devised a widely used twostep iterative method
to stabilize the temporal scheme. This has subsequently been adapted to handle
the strong property variations in supercritical fluids [20]. One could also obtain a
density predictor by solving the continuity equation.
In sections 4.3 and 4.4 we present a new alternative method that does not use
4.2. The Temporal Density Gradient 57
any predictor solves or iterations to handle the unknown density at a new time step.
We will show that the error in our approximation can be made negligible compared
to the error in the finite difference scheme, and that the method can be made
unconditionally stable by offsetting the specific enthalpy with a constant.
4.2. The Temporal Density Gradient

The pressurecorrection method for incompressible flow in section 3.2 can be mod
ified for lowMach number flow in a straightforward manner. The semidiscrete
continuity equation 3.2a is adjusted to include the temporal derivative of the den
sity:
𝑛
𝜕𝜌 𝑛
− 𝑫𝒎𝑛 + 𝑪𝒑𝑛 = −𝒓 − 𝒢 [( ) ] ≕ −𝒓̃ , (4.11)
𝜕𝑡
4
where 𝒢[⋅] denotes the Galerkin projection onto the solution space. The temporal
derivative of the density is estimated with a secondorder backward finite difference
scheme:
𝑛
𝜕𝜌 1 3 𝑛 1
( ) ≈ ( 𝜌 − 2𝜌𝑛−1 + 𝜌𝑛−2 ) . (4.12)
𝜕𝑡 𝛿𝑡 2 2
The timesplitting scheme is exactly the same as in section 3.2, except that 𝒓 is
𝑛
replaced by 𝒓̃ .
This is a large advantage of solving for the mass flux. If we had instead solved
for the velocity (such as in [15]), then the density would have had to have been
incorporated into the divergence operator 𝐷 in Eq. 3.9, and into the mass matrix
𝑀 in Eq. 3.4.
This extension of the pressurecorrection method to compressible flows has
sometimes proved unstable in finite difference schemes that were applied to mixing
flows with large density ratios (of approximately more than a factor of 3), because
the continuity equation was not satisfied in the inviscid limit; see Nicoud [21]. It is
not certain whether the same instability would occur for the discontinuous Galerkin
method presented here; our experience so far has not exposed instabilities with
large density ratios. Nicoud suggested a different generalization of the pressure
correction method to variabledensity flows, where the density is incorporated into
the pressure matrix, rather than on the righthand side of Eq. 3.9.
The large advantage of the approach presented here is that the pressure matrix
is the same at all time steps.1 We can therefore assemble it once, and precom
pute the incomplete Cholesky preconditioner for the linear solver. Furthermore, the
condition number of the diffusion matrix 𝐴 worsens if it includes a variable coeffi
cient that depends on the density. For these reasons, the pressure solves are much
cheaper with a constant pressure matrix.
Irrenfried [22, 23] also claimed (without reference or demonstration) that using
Eq. 4.12 may result in a considerable numerical error for small time steps, and
1 exceptfor equalorder discretizations with a temperaturedependent kinematic viscosity, in which case
the pressure stabilization (Eqs. 2.8–2.9) depends on ℎ
therefore used an interesting alternative, based on

𝜕𝜌 𝜕𝑇
= 𝜌𝑇 . (4.13)
𝜕𝑡 𝜕𝑡
The idea is to replace 𝜕𝑇/𝜕𝑡 on the rhs by the spatial discretization of the tem
perature. This was probably straightforward to implement in their work, which is
based on a finite volume scheme for the temperature equation in primitive form
(i.e., 𝐷𝑇/𝐷𝑡 = …). The idea is certainly attractive: given the onetoone correspon
dence between the density and the temperature (or the enthalpy), it seems natural
to use the same spatial discretization for their temporal derivatives.
It is less clear how the equivalent approach could be used here for at least
two reasons. First, (𝜌ℎ)𝜌 is not constant, and we use a finite element method, so
we cannot simply multiply the solution vector by the value of a fluid property in a
4 pointwise manner. This would probably have to be done with a Galerkin projection
instead. A more fundamental problem is that we solve for the enthalpy equation
in conservative form, so we have a spatial discretization for 𝜕(𝜌ℎ)/𝜕𝑡, not 𝜕ℎ/𝜕𝑡.
Unfortunately (𝜌ℎ) does not uniquely determine 𝜌 in many fluids, and therefore the
relation
𝜕𝜌 1 𝜕(𝜌ℎ)
= (4.14)
𝜕𝑡 (𝜌ℎ)𝜌 𝜕𝑡
could be undefined, or very badly conditioned, since |(𝜌ℎ)𝜌 | can be arbitrarily small.
This problem of a nonunique volumetric enthalpy (𝜌ℎ) will return later in this chap
ter. Perhaps Eq. 4.14 could be used if (𝜌ℎ)𝜌 is made nonzero by a suitable enthalpy
offset, as in section 4.3.2.
4.3. Linearizing 𝜕(𝜌ℎ)/𝜕𝑡

As mentioned in the introduction, solving for a primitive variable (ℎ) with the en
thalpy equation in conservative form complicates the temporal derivative of the
enthalpy, because it is weighed by the temperaturedependent density. This sec
tion and the next study the stability and convergence of the time stepping scheme
in detail.
The analysis is simplified by considering a spaceindependent enthalpy equation:
d (𝜌ℎ)
= −𝜆ℎ + 𝑄 , (4.15)
d𝑡
where 𝜆 is a constant, and 𝑄 = 𝑄(𝑡). Using an implicit finite difference scheme,
the enthalpy and the corresponding density can be estimated at a time step 𝑛 by
𝑞
𝛾0 𝑛 𝛾𝑖 𝑛−𝑖
(𝜌ℎ) + ∑ (𝜌ℎ) = −𝜆ℎ𝑛 + 𝑄𝑛 . (4.16)
𝛿𝑡 𝛿𝑡
𝑖=1
Due to the variable density, this equation is not linear in the unknown ℎ𝑛 . We
therefore consider two linearizations in ℎ𝑛 , which we term method #1 and method
#2.
4.3. Linearizing the Temporal Derivative of the Enthalpy 59
Both of these methods use a predictor ℎ∗ and a corresponding 𝜌∗ that are close to
ℎ and 𝜌𝑛 . This predictor can be obtained in several ways, such as by extrapolating
𝑛
from previous time steps. When solving the full system 1.1a1.1c, a predictor for
𝜌𝑛 can also be obtained by solving the continuity equation. The analyses in this
section are for a general (ℎ∗ , 𝜌∗ ), though we will make the reasonable assumption
that (ℎ∗ − ℎ𝑛 ) is at least firstorder accurate in 𝛿𝑡.
The two linearization methods are as follows.
Method #1 is perhaps the most obvious approach: let 𝜌𝑛 ≈ 𝜌∗ , resulting in an
approximation ℎ[1] ≈ ℎ𝑛 that is given by
𝑞
𝛾0 ∗ [1] 𝛾𝑖 𝑛−𝑖
𝜌 ℎ +∑ (𝜌ℎ) = −𝜆ℎ[1] + 𝑄𝑛 . (4.17)
𝛿𝑡 𝛿𝑡
𝑖=1
𝑛
4
Method #2 is based on a Taylor expansion of (𝜌ℎ) about the predictor:
∗ ∗ ∗
(𝜌ℎ)𝑛 ≈ (𝜌ℎ)∗ + (𝜌ℎ)ℎ (ℎ𝑛 − ℎ∗ ) = (𝜌ℎ)ℎ ℎ𝑛 − (ℎ2 𝜌ℎ ) . (4.18)
Substituting this into Eq. 4.16 yields an approximation ℎ[2] ≈ ℎ𝑛 , given by
𝑞
𝛾0 ∗ 𝛾𝑖 𝑛−𝑖 𝛾0 2 ∗
(𝜌ℎ)ℎ ℎ[2] + ∑ (𝜌ℎ) = (ℎ 𝜌ℎ ) − 𝜆ℎ[2] + 𝑄𝑛 . (4.19)
𝛿𝑡 𝛿𝑡 𝛿𝑡
𝑖=1
Note that method #1 is effectively a single step in a fixedpoint iteration, whereas

method #2 is a single step in a Newton iteration.
4.3.1. Error Estimates and Stability

The errors and the stability of methods #1 and #2 can be analyzed by using a
Taylor series for 𝜌𝑛 about the predictor, that is,
∞ ∗
1 𝜕𝑘 𝜌 𝑘
𝜌 = ∑ ( 𝑘 ) (ℎ𝑛 − ℎ∗ ) .
𝑛
(4.20)
𝑘! 𝜕ℎ
𝑘=0
Define the following deviations from the nonlinear finite difference equation 4.16:
error in the predictor: 𝜖 ∗ ≔ ℎ∗ − ℎ𝑛 ,
linearization error in method #1: 𝜖 [1] ≔ ℎ[1] − ℎ𝑛 , (4.21)
[2] [2] 𝑛
linearization error in method #2: 𝜖 ≔ℎ −ℎ .
The derivations are tedious, and deferred the appendix at the end of this chapter.
Here we summarize the main theoretical results.
The first result is an a priori error estimate. Appendix A shows that both Eq.
4.17 and Eq. 4.19 can be rewritten as
𝑞
𝛾0 𝑛 𝛾𝑖 𝑛−𝑖 𝑛
(𝜌ℎ) + ∑ (𝜌ℎ) = −𝜆eff ℎ𝑛 + 𝑄eff , (4.22)
𝛿𝑡 𝛿𝑡
𝑖=1
where
2 𝑛 2
𝜆eff = 𝜆 + 𝒪 (𝜖 ∗ /𝛿𝑡) and 𝑄eff = 𝑄𝑛 + 𝒪 (𝜖 ∗ /𝛿𝑡) for method #1, (4.23a)
and
3 𝑛 3
𝜆eff = 𝜆 + 𝒪 (𝜖 ∗ /𝛿𝑡) and 𝑄eff = 𝑄𝑛 + 𝒪 (𝜖 ∗ /𝛿𝑡) for method #2. (4.23b)
That is, the approximations in method #1 and #2 are equivalent to the original Eq.
4.16, except that 𝜆 and 𝑄𝑛 are replaced by their effective values, which are related
to the error in the predictor.
A second important result regards the stability of the linearization methods.
Appendix A.1 shows that the error for method #1 is related to the error in the
predictor as
4 𝜖 [1]
= −(
𝜌ℎ ℎ
∗
) + 𝒪 (𝜖 ∗ )
𝜖∗ 𝜌 + 𝜆 (𝛿𝑡/𝛾0 )
∗ (4.24)
𝜌ℎ ℎ
= −( ) + 𝒪 (𝛿𝑡) + 𝒪 (𝜖 ∗ )
𝜌
Note that 𝜖 [1] /𝜖 ∗ vanishes up to first order as ℎ∗ → 0. Eq. 4.24 also suggests that
method #1 cannot always be made stable by iterating within a time step, that is,
by calculating a new predictor 𝜌∗ from the estimate ℎ[1] , and repeating Eq. 4.17.
Stability of the iteration is guaranteed if the error in the new approximation is always
smaller than the error in the predictor, that is, |𝜖 [1] | < |𝜖 ∗ |. This condition is only
met if ∗ ∗
𝜌 𝑐𝑝
|ℎ∗ | < |− | = ( ) , (4.25)
𝜌ℎ 𝛽
where we have made the reasonable assumptions that 𝜖 ∗ = 𝒪 (𝛿𝑡), and that 𝜌ℎ < 0.
If Eq. 4.25 is not met at every step in the iteration, then it may not converge.
Similarly, iterating method #2 within a time step is stable if |𝜖 [2] | < |𝜖 ∗ |. Ap
pendix A.2 shows that
∗
∗ 𝛿𝑡 𝜖 [2] 1 2
((𝜌ℎ)ℎ + 𝜆) ∗ = (𝜌ℎ + 𝜌ℎℎ ℎ) 𝜖 ∗ + 𝒪 (𝜖 ∗ ) . (4.26)
𝛾0 𝜖 2
Since we can reasonably expect that the error in the predictor (𝜖 ∗ ) is at least first
order accurate in 𝛿𝑡, we always have |𝜖 [2] | < |𝜖 ∗ | (and therefore a stable iteration)
in the limit 𝛿𝑡 → 0, provided that
∗
(𝜌ℎ)ℎ ≠ 0 . (4.27)
In other words, the volumetric enthalpy (𝜌ℎ) must be a strictly monotonic function
of the specific enthalpy ℎ.
This restriction for method #2 also follows more directly from Eq. 4.19 in the
limit of small time steps, because the coefficient of ℎ[2] cannot vanish. In practice
one will want to satisfy the stronger relation
(𝜌ℎ)ℎ > 0 (4.28)
4.3. Linearizing the Temporal Derivative of the Enthalpy 61
to ensure that the enthalpy discretization is positive definite.

We conjecture that the stability requirements (Eq. 4.25 for method #1; Eq.
4.27 for method #2) must always be satisfied in the limit of small time steps, even
when the linearization is not iterated within a time step. It seems reasonable to
expect that a stable numerical method can be iterated without diverging. This is
supported by the numerical tests in Section 4.4
4.3.2. Proper Scaling of the Enthalpy Equation

Curiously, the analyses in the previous subsection have led to stability requirements
(Eqs. 4.25, 4.27) that depend on the fluid properties, and they are not satisfied
for all fluids. For example, the volumetric enthalpy in supercritical fluids can either
increase or decrease with the temperature due to the strong thermal expansion,
thereby violating Eq. 4.27.
This problem can be addressed by solving for a different variable
4
ℎ̃ ≔ ℎ − ℎ0 . (4.29)
Eq. 1.1c then becomes
̃
𝜕(𝜌ℎ) 𝑘
ℎ0 𝑅 + + ∇ ⋅ (𝐦 ℎ) ̃ +𝑄 ,
̃ = ∇ ⋅ ( ∇ℎ) (4.30)
𝜕𝑡 𝑐𝑝
where 𝑅 ≔ 𝜕𝜌/𝜕𝑡 + ∇ ⋅ 𝐦 = 0 is the residual of the continuity equation 1.1a. Thus ℎ̃

satisfies the same transport equation as ℎ, and it can be discretized in the same way.
This does not affect the diffusion (since ∇ℎ̃ = ∇ℎ), but it does change the convection
and the temporal derivative, which now has a different stability guarantee.
In particular, Eq. 4.25 for method #1 becomes
∗
𝑐𝑝
|ℎ̃ ∗ | = |ℎ∗ − ℎ0 | < ( ) . (4.31)
𝛽
We cannot know in advance whether this will be satisfied at all time steps. There
is therefore no a priori value for ℎ0 that guarantees stability, though it seems that
ℎ0 is best chosen such that ℎ ≈ ℎ0 at the average temperature.
Conversely, we can find an a priori lower bound for ℎ0 when using method #2.
The stability guarantee (Eq. 4.28) becomes
𝑐𝑝
̃ = (𝜌(ℎ − ℎ0 )) = (𝜌ℎ) − ℎ0 𝜌ℎ > 0
(𝜌ℎ) ⇔ ℎ0 > (𝜌ℎ)𝜌 = ℎ − , (4.32)
ℎ̃ ℎ ℎ 𝛽
and so method #2 can be made unconditionally stable by choosing ℎ0 sufficiently

large. In particular, if the temperature is known to lie in a range [𝑇min , 𝑇max ], then
a theoretical lower bound for stable values for ℎ0 is
𝑐𝑝
ℎ0min = max (𝜌ℎ)𝜌 = max (ℎ − ) , (4.33)
𝑇min ≤𝑇≤𝑇max 𝑇min ≤𝑇≤𝑇max 𝛽
h0 = hmin
0 − 0.7 hmin
0 h0 = hmin
0
h0 = hmin
0 − 0.3 hmin
0 h0 = hmin
0 + 0.3 hmin
0
125
1.0
100
0.8
ρ(h − h0 ) (MJ/m3 )
75
ρ(h − h0 ) (–)
0.6
50
25 0.4
0 0.2
4 −25
0.0
302 303 304 305 306 307 302 303 304 305 306 307
T (K) T (K)
Figure 4.1: Top: rescaled volumetric enthalpy (𝜌(ℎ − ℎ0 )) of carbon dioxide at the supercritical pressure
of 7.5 MPa, as a function of the temperature for various choices of ℎ0 . Bottom: same data, but with
each line scaled to [0, 1].
The function increases monotonically for ℎ0 ≥ ℎ0min . The thermodynamic reference point is placed at (1
bar, 0 °C). The data are based on [24], accessed through the CoolProp software library [25].
which can of course be negative. Fig. 4.1 shows an example of the rescaled
volumetric enthalpy (𝜌(ℎ − ℎ0 )) for various choices of ℎ0 for a real fluid in the
temperature range (302K, 307K). There is a practical limit on the magnitude of
ℎ0 , because we are solving a transport equation for 𝜌(ℎ − ℎ0 ), which becomes
equivalent to the density 𝜌 for very large values of |ℎ0 |.
On a heuristic level, the change of variables in Eq. 4.29 can be thought of
as a way to discretize something in between the conservative and the primitive
transport equations. The conservative transport equation is merely based on the
conservation of enthalpy, whereas deriving the primitive transport also requires the
conservation of mass. The primitive equation can be obtained by subtracting ℎ𝑅
from the conservative equation. Comparing Eq. 4.30, we are subtracting ℎ0 𝑅 from
the conservative equation, and ℎ0 determines how much of the continuity equation
is used for the enthalpy transport.
Of course a rescaling of the unknowns, such as in Eq. 4.29, is not unknown in
CFD literature, but it has usually been presented as a mere numerical convenience
(e.g., [18]). The above analyses show that the accuracy and stability of the numer
ical scheme depend critically on a proper choice of ℎ0 . In practice this may require
some trial and error, though these analyses offer useful guidelines.
4.3.3. Special Case of an Ideal Gas

Since many fluids are accurately described by the ideal gas law (Eq. 4.10), it is
worth specializing the above analyses to this particular case. The specific heat
4.4. Test Case for the Spaceindependent Enthalpy Equation 63
capacity is usually approximately constant in an ideal gas, so that ℎ = 𝑐𝑝 𝑇 − ℎ0 for

some constant ℎ0 . As already mentioned in section 4.1.2, solving the conservative
enthalpy transport equation does not make sense when ℎ = 𝑐𝑝 𝑇 (i.e., ℎ0 = 0),
since the volumetric enthalpy would be constant. However, both method #1 and
method #2 can be stable for a proper choice of a nonzero constant ℎ0 .
First we consider the relationship between the linearization error (𝜖 [1] ) and the
predictor error (𝜖 ∗ ) for method #1, given by Eq. 4.24. From the equation of state
we have 𝜌ℎ = −𝜌/ (𝑐𝑝 𝑇). Eq. 4.24 becomes
𝜖 [1] ℎ∗
= + 𝒪 (𝛿𝑡) + 𝒪 (𝜖 ∗ ) . (4.34)
𝜖∗ ℎ∗ + ℎ0
Stability for method #1 is guaranteed if |𝜖 [1] /𝜖 ∗ | < 1 for all time steps, which is
equivalent to ℎ∗ /ℎ0 > 1/2. 4
For method #2, the volumetric enthalpy should be a strictly monotonically in
creasing function of the temperature. For an ideal gas, 𝐻𝑇 = 𝜌 (𝑐𝑝 − ℎ/𝑇). In the
case ℎ = 𝑐𝑝 𝑇 − ℎ0 this becomes 𝐻𝑇 = (𝜌/𝑇)ℎ0 , meaning that we must set ℎ0 > 0.
4.4. Test Case for the Spaceindependent Enthalpy

Equation2
Before solving the full system of transport equations, we clarify the theoretical
results for the spaceindependent enthalpy equation 4.15 in Section 4.3 with a
numerical example that is based on a manufactured solution. Omitting the units of
measurement, the exact temperature is
𝑇ex (𝑡) = 0.5 + 0.1 sin(2𝜋𝑡) (4.35)
with 0 ≤ 𝑡 ≤ 1. The equation of state is
𝜌 = 𝜌0 𝑇 + 𝜌1 (1 − 𝑇) , (4.36)
and the specific heat capacity is kept constant, so that
ℎ = 𝑐𝑝 𝑇 − ℎ0 . (4.37)
The required source term 𝑄(𝑡) follows from Eq. 4.15. For the numerical tests we
let 𝜌0 = 0.5, 𝜌1 = 2, 𝑐𝑝 = 1, and 𝜆 = 0.1. The results presented here were all
obtained at 𝛿𝑡 = 2−11 to investigate the limit of small time steps. We have checked
that lowering the time step size to 𝛿𝑡 = 2−14 does not affect whether the numerical
method is stable. To ensure that rounding errors did not play a significant role with
these tiny time steps, all calculations in this section were performed with 128bit
floating point precision.
The numerical schemes were tested with various orders of the BDF time stepping
scheme. The predictor ℎ∗ is obtained with an 𝑠th order extrapolation from previous
2 Thecode for the finite difference method for the spaceindependent enthalpy equation can be found
on GitHub [26]. It can be used to reproduce the results in this section.
Table 4.1: Coefficients for extrapolation from previous time steps. (See Eq. 4.38).
𝛼1 𝛼2 𝛼3 𝛼4
EX1 1
EX2 2 1
EX3 3 3 1
EX4 4 6 4 1
Table 4.2: Order of extrapolation for the enthalpy predictor (Eq. 4.38) for the linearization methods
described in Section 4.3. The minimum values satisfy Eq. 4.39; from the maximum value onward, Eq.
4.39 holds with strict inequality.
finite difference Method #1 Method #2

4 coefficients
min max min max
BDF1 EX2 EX2 EX1 EX2
time steps (denoted by EX𝑠), and the corresponding 𝜌∗ is determined from the
equation of state. Specifically,
𝑠
ℎ = ∑ 𝛼𝑖 ℎ𝑛−𝑖 = ℎ𝑛 + 𝒪 (𝛿𝑡 𝑠 ) .
∗
(4.38)
𝑖=1
The weights are in Table 4.1.

There are two numerical errors in each time step: (i) the BDF error, which is
inherent in the finite difference scheme, and (ii) the linearization error in going from
Eq. 4.16 to either Eq. 4.17 or Eq. 4.19 when using method #1 or method #2. If
the EX𝑠 coefficients are used to obtain a predictor, then the error in the predictor
is 𝜖 ∗ ≔ ℎ∗ − ℎ𝑛 = 𝒪 (𝛿𝑡 𝑠 ). For method #1, Eq. 4.23a then implies a linearization
2
error of 𝒪 (𝜖 ∗ /𝛿𝑡) = 𝒪 (𝛿𝑡2𝑠−1 ). A BDF𝑞 scheme makes an 𝒪 (𝛿𝑡𝑞+1 ) error per
time step, so that the overall order of accuracy is min(2𝑠 − 1, 𝑞 + 1). Similarly,
Eq. 4.23b implies that the overall error per time step for method #2 is of order
min(3𝑠 − 1, 𝑞 + 1). The order of extrapolation should therefore satisfy
(𝑞 + 2)/2 for method #1,

𝑠≥{ (4.39)
(𝑞 + 2)/3 for method #2,
or else the linearization error dominates, and the usual order of convergence of
the BDF scheme cannot be achieved. If strict inequality in Eq. 4.39 is satisfied,
then the linearization error is negligible, and increasing the order of extrapolation
is pointless. Table 4.2 lists the range of reasonable extrapolation orders.
Fig. 4.2 shows the error in the numerical temperature as a function of ℎ0 , by
using method #1 (Eq. 4.17). Note how the calculations diverge when ℎ0 is either
4.5. Test Cases with LowMach Number Flow 65
too small or too large. The stability guarantee in Eq. 4.25 cannot be determined a
priori, because ℎ∗ is not known before the calculation. For low extrapolation order,
method #1 sometimes converges even when Eq. 4.25 is not met at all time steps.
According to Table 4.2, the linearization error is negligible compared to the BFD
error for extrapolation orders of at least 2, 3, and 3 for the BDF1, BDF2, and BDF3
schemes. For these cases Eq. 4.25 becomes a strict requirement for stability. Note
that the range of stable values for ℎ0 decreases with higherorder extrapolations,
but all simulations converge with ℎ0 = 0.5, which is the value for which ℎ is closest
to zero.
Fig. 4.3 shows the equivalent error plots for method #2 (Eq. 4.19). For the
current equation of state, we have an explicit, a priori expression for the stability
criterion in Eq. 4.27:
(𝜌ℎ)𝑇 ≠ 0 ⇔ ℎ0 /𝑐𝑝 ≠ 2𝑇 − 𝜌1 /(𝜌1 − 𝜌0 ) . (4.40) 4

The tests show that the numerical scheme is stable if and only if this criterion is
satisfied everywhere in the domain, regardless of the order of the timestepping
scheme, or the order of extrapolation for the predictor. Furthermore, the results
show that the minimal extrapolation orders in Table 4.2 need to be reached in order
to achieve the lowest errors, but higher orders of extrapolation have no effect.
4.5. Test Cases with LowMach Number Flow

4.5.1. Variabledensity Manufactured Solution
As in the previous chapter, a manufactured solution is used to verify the numerical
scheme and its implementation. The exact solution constructed by working back
ward from the exact mass flux and pressure, which can be chosen arbitrarily. The
choice of the pressure is of little consequence, though we make sure that both 𝑚
and 𝑝 vary nonlinearly in time, and that they do not lie in the numerical solution
space. Integrating the continuity equation over time then gives the density, which
in turn determines the temperature and the enthalpy. The transport properties (𝜇
and 𝑘) are arbitrary functions of the temperature. The external force and heat
source follow from Eqs. 1.1b and 1.1c.
It is surprising that we could not find previous work with a manufactured solution
that is (i) compressible, (ii) uses temperaturedependent transport properties, and
(iii) satisfies the unmodified continuity equation (without an artificial mass source).
Most previous work has focussed on finding clever analytical solutions to the sys
tem 1.1 with a variable density (see, e.g., [27], and the references therein). Others
(e.g., [28]) have included a source term in the continuity equation, but this appears
less suitable for a timesplitting method, where the continuity plays a central role in
the discretization (as in Section 4.3.2), and we do not want to adapt the numerical
scheme to conform with the manufactured solution. Perhaps the current approach
was not taken before because it results in nontrivial source terms (𝐹 and 𝑄). We
handle these calculations symbolically with the Python SymPy library. The man
ufactured solution is made up of polynomials to keep these symbolic calculations
feasible.
BDF1 BDF2 BDF3

10−1
10−4
EX1
10−7
4
10−1
10−4
EX2
10−7
10−1
10−4
EX3
10−7
10−1
10−4
EX4
10−7
-1 0 1 2 3 4 -1 0 1 2 3 4 -1 0 1 2 3 4
h0 h0 h0
Figure 4.2: Error (|𝑇 − 𝑇ex | /𝑇ex ) at 𝑡 = 1 for the test case in Section 4.4 as a function of the enthalpy
offset ℎ0 , using method #1. The red vertical dotted lines bound the values for which Eq. 4.25 held at
all time steps.
BDF1 BDF2 BDF3
10−3
EX1
10−6
10−9
4
10−3
EX2
10−6
10−9
10−3
EX3
10−6
10−9
10−3
EX4
10−6
10−9
-1 0 1 2 -1 0 1 2 -1 0 1 2
h0 h0 h0
Figure 4.3: Error (|𝑇 − 𝑇ex | /𝑇ex ) at 𝑡 = 1 for the test case in Section 4.4 as a function of the enthalpy
offset ℎ0 , using method #2. The red vertical dotted lines bound the values for which the stability criterion
in Eq. 4.40 is violated at some time 𝑡.
The exact solution is

3
1 (𝑥/𝐿 − 1) 3/2
𝐦ex = (1 + 𝑡3 ) [ 2 ]+[ ] ,
4 (𝑥/𝐿 − 1) (𝑦 − 1)(𝑦 + 1) 0 (4.41)
ex 3 3
𝑝 = (1 + 𝑡 ) (𝐿 − 𝑥)
with 0 < 𝑡 ≤ 1, and the domain is as in section 3.3.2, Fig. 3.2. The addition of
[3/2, 0]⊺ ensures that 𝑚ex
1 > 0, so that there is no outflow at the Dirichlet boundary
condition, and no inflow at the outlet. The density is determined by integrating
(−∇ ⋅ 𝐦) over 𝑡, to find
1 1
𝜌ex = − ( 𝑡4 + 𝑡) (𝑥/𝐿 − 1)2 (2𝑦 + 3/𝐿) + 3 , (4.42)
4 4
4
where the addition of the constant 3 ensures that 𝜌 > 0 everywhere. The specific
heat capacity is constant, so that ℎ = 𝑐𝑝 𝑇 − ℎ0 with 𝑐𝑝 = 1. We use a nonaffine
equation of state: 𝑇 = ((𝜌1 − 𝜌)/(𝜌1 − 𝜌0 ))2 , where 𝜌0 = 2 and 𝜌1 = 4 are lower
and upper bounds for 𝜌, so that the temperature is between 0 and 1. As in Section
3.3.2, the viscosity and conductivity are 𝜇 = 0.1 + 𝑇(1 − 𝑇) and 𝑘 = 𝜇𝑐𝑝 /Pr with
Pr = 1. The solution is depicted in Fig. 4.4.
We base the enthalpy timestepping scheme on a linearization of (𝜌ℎ)𝑛 about
a predictor for ℎ𝑛 , that is, method #2 in Section 4.3. We let ℎ0 = 0.2, so that Eq.
4.28 is satisfied everywhere. Increasing ℎ0 had no noticeable effect. The predictor
ℎ∗ is obtained with a secondorder extrapolation from previous time steps (using
the EX2 weights in Table 4.1). We found that increasing the extrapolation order for
ℎ∗ had no noticeable effect on the stability or the errors, which is in line with the
tests for the BDF2 scheme in Fig. 4.3.
Fig. 4.5 shows the convergence with temporal refinement. The velocity and the
temperature converge with second order, just like for the constantdensity results in
Fig 3.4, though in this case the pressure also shows 𝒪 (𝛿𝑡2 ) behavior. Note that the
mixedorder cases remain fully stable, even for very small 𝛿𝑡, despite the possible
small𝛿𝑡 instability for variabledensity flows that was discussed in Section 4.3.
Fig. 4.6 shows two other examples of temporal convergence with a mixedorder
scheme, but with less effective enthalpy treatments. We found that method #1 was
stable for all ℎ0 ≥ 0, and that the precise value of ℎ0 is of little consequence to the
L2 errors in the final answer. It is clear that the temperature does not converge with
second order when method #1 is used for the enthalpy treatment. The right column
in Fig. 4.6 shows that method #2 can also be unstable when the enthalpy offset is
not sufficiently large (here it is 0.0 instead of 0.2): all quantities still converge with
secondorder accuracy, but the error diverges at small time steps. Some calculations
failed due to numerical backflow at the outlet, which our simple outlet boundary
condition cannot handle.
Table 4.3 collects the spatial convergence rates that were obtained with method
#2 and ℎ0 = 0.2 (i.e., the same conditions are for the temporal convergence in Fig.
4.5). As the mesh is refined, the mixedorder discretization displays 𝒪 (𝓁𝒫+1 ) be
havior for quantities with a polynomial order 𝒫, where 𝓁 ∝ 1/𝑁𝑦 is the characteristic
0.27 0.29 0.32 0.34 0.36 0.39 0.41 0.43 0.46 0.48 0.5
(a) 𝑢1
-0.16 -0.15 -0.13 -0.11 -0.098 -0.082 -0.065 -0.049 -0.033 -0.016 0
(b) 𝑢2
0.021 0.026 0.032 0.037 0.042 0.048 0.053 0.058 0.063 0.069 0.074
(c) Kinematic viscosity (𝜈) and thermal diffusivity (𝛼). (Note that Pr = 1, so 𝛼 = 𝜈.)
2.5 2.6 2.7 2.8 3 3.1 3.2 3.3 3.5 3.6 3.7
(d) 𝜌
Figure 4.4: Variabledensity manufactured solution in Eqs. 4.414.42 at 𝑡 = 1.

Pm = Pp = Ph = 1 Pm = 2, Pp = Ph = 1
10−2
10−2
10−3 10−3
10−4 10−4
4
10−5
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212
100 100
10−1 10−1
Ny = 64
Ny = 32 10−2
10−2
Ny = 16
Ny =8 10−3
10−3
Ny =4
Ny =2 10−4
10−4
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212
10−2 10−2
10−3
10−3
10−4
10−4
10−5
10−5
10−6
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212

1 / δt 1 / δt
Figure 4.5: Convergence of the numerical solution toward the variabledensity manufactured solution
(Eq. 4.41) with temporal refinement, using method #2 with ℎ0 = 0.2. The characteristic element length
is inversely proportional to 𝑁𝑦 . The black dashed lines indicate ideal secondorder convergence in 𝛿𝑡.
method #1, h0 = 0.0 method #2, h0 = 0.0

10−1 10−1
10−2 10−2
10−3 10−3
10−4 10−4
4
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212
100 100
10−1 10−1
Ny = 32
10−2 Ny = 16 10−2
Ny =8
Ny =4
10−3 10−3
Ny =2
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212
10−2 10−1
10−2
10−3
10−3
10−4
10−4
10−5 10−5
22 23 24 25 26 27 28 29 210 211 212 22 23 24 25 26 27 28 29 210 211 212
1 / δt 1 / δt
Figure 4.6: Equivalent of the mixedorder case (𝒫𝑚 = 2, 𝒫𝑝 = 𝒫ℎ = 1) in Fig. 4.5, but with a different
value of ℎ0 , and comparing method #1 to method #2. The missing values in the figures in the right
column indicate failed calculations.
Table 4.3: Convergence toward the variabledensity manufactured solution in Eqs. 4.41–4.42 (Fig. 4.4)
with spatial refinement and fixed 𝛿𝑡 = 2−12 .

𝑁𝑦 error conv error conv error conv
Mixed order (𝒫𝑚 = 2, 𝒫ℎ = 𝒫𝑝 = 1):
21 1.56e2 6.11e1 3.57e3
22 3.92e3 2.00 1.11e1 2.47 8.86e4 2.01
23 9.80e4 2.00 1.69e2 2.71 2.18e4 2.02
24 2.46e4 2.00 2.32e3 2.87 5.43e5 2.01
25 6.17e5 1.99 3.05e4 2.92 1.36e5 2.00
26 1.55e5 1.99 4.03e5 2.92 3.39e6 2.00
4 Equal order (𝒫𝑚 = 𝒫𝑝 = 𝒫ℎ = 1):
21 3.82e2 8.96e1 2.85e3
22 4.54e3 3.07 1.93e1 2.21 4.17e4 2.01
23 8.41e4 2.43 3.54e2 2.45 6.46e5 2.02
24 2.07e4 2.02 5.30e3 2.74 1.08e5 2.01
25 5.20e5 1.99 7.26e4 2.87 2.46e6 2.00
26 1.31e5 1.99 9.59e5 2.92 6.17e7 2.00
mesh length. For the equalorder case, the velocity shows 𝒪 (𝓁𝒫𝑢 +2 ) hyperconver
gence in 𝐮.
Paradoxically, the errors in the pressure are much lower for the equalorder
case, despite its a reduced solution space. This could perhaps be explained by
the error in the 𝜕𝜌/𝜕𝑡 term, which is extrapolated from previous time steps (Eq.
4.12). This might induce discontinuities in 𝑝, which would be suppressed by the
pressure stabilization, which is only present for the equalorder discretization. This
explanation is supported by the fact that this phenomenon did not occur for the
constantdensity manufactured solution in section 3.3.2 (Table 3.3). Note that the
orders of convergence are the same for constantdensity and the variabledensity
test cases.
4.5.2. Validation with Flow Past a Heated Circular Obstacle

To validate the numerical method with a variabledensity flow, we replicate a nu
merical test case by Shi et al. [29], who used a specialized cylindrical finite volume
scheme to handle the circular geometry. The geometry is the same as in section 3.4,
see Figs. 3.5. The temperature is fixed at 𝑇w at the cylinder, and 𝑇∞ at the inlet, top
and bottom parts of the domain. The outlet has a homogeneous Neumann bound
ary condition. The temperatures are 𝑇∞ = 20°C and 𝑇w = 1.5𝑇∞ = 166.575°C,
resulting in Pr = 0.7146. Shi et al. solved for the temperature, approximating the
material properties as
𝜑 = 𝑎0 + 𝑎1 (𝑇 − 𝑇𝐹 ) + 𝑎2 (𝑇 − 𝑇𝐹 )2 , (4.43)
4.6. Discussion and Conclusion 73
Table 4.4: Coefficients for the material properties in Eq. 4.43. Reproduced from [29].
𝜌 (kg 𝑚−3 ) 𝜇 (kg m−1 s−1 ) 𝑘 (W m−1 K−1 ) 𝑐𝑝 (m2 s−2 K−1 )
𝑎0 1.268672727 1.7254e05 2.4195e2 1.00620979e3
𝑎1 (K−1 ) 4.08741e03 4.95611e08 7.5234e5 1.4522145e2
𝑎2 (K−2 ) 7.23864e06 2.7214e11 3.2588e8 4.13753e4
where 𝜑 is one of (𝜌, 𝜇, 𝑘, 𝑐𝑝 ), 𝑇𝐹 = 0°C, and the coefficients 𝑎𝑖 are in Table 4.4.
Since 𝑐𝑝 ≔ ℎ𝑇 is a secondorder polynomial in 𝑇, we need to find the root of the
thirdorder polynomial ℎ = ℎ(𝑇) to map from ℎ to a fluid property. This minor
inconvenience permits a better comparison with the results in Shi et al. .
We obtain our results on the same mesh as for the isothermal case (Fig. 3.6).
Fig. 4.7 shows an example of instantaneous flow fields. Fig. 4.8 shows the lift and
4
drag coefficients and the Nusselt number, which is defined as
𝐷 1
Nu = ∫ 𝐧 ⋅ ∇𝑇 , (4.44)
𝑇w − 𝑇∞ ‖𝜕𝑆‖leb 𝜕𝑆
where ‖𝜕𝑆‖leb = 𝜋𝐷 is the circumference of the circular obstacle. Recall that the
dimensionless shedding frequency was approximately St = 0.166 for the isothermal
case in section 3.4. Here we find St = 0.1536 for the heated cylinder, which differs
by 1% from the value of St = 0.152 in Shi et al. [29] and the experimental value
of St = 0.152 in [30].
This result was obtained by linearizing (𝜌ℎ)𝑛 with method #2 for the temporal
derivative of the enthalpy. We subtracted an offset ℎ0 from the enthalpy (as ex
plained in Section 4.3.2), such that the maximum value of ℎ was zero. Interestingly,
there is no noticeable change in the results when we use an equalorder scheme,
even when 𝛿𝑡 is decreased by a factor of 10 or 100. For the mixedorder calcula
tions, we found no difference between using an SIP or an LDG pressure matrix. We
repeated the calculation using method #1, setting the enthalpy offset to a value
ℎ0∗ , such that ℎ = 0 at 𝑇 = (𝑇w + 𝑇∞ )/2. This resulted in almost exactly the same
shedding frequency (St = 0.1537).
For all test cases, we found that some values for the enthalpy offset result in
unstable schemes, yielding oscillatory pressure fields. For method #1 this happens
when ℎ0 is far from ℎ0∗ ; for method #2 this happens when ℎ0 is too small. When op
erating in the range of stable ℎ0 values, the exact enthalpy offset has no noticeable
impact on the shedding frequency.
4.6. Discussion and Conclusion

Since the density is a function of the specific enthalpy, the temporal finite difference
scheme requires that the volumetric enthalpy (𝜌ℎ) be linearized in ℎ, and we have
analyzed two methods for doing this, both of which need a predictor for the enthalpy
at the new time step. This led to theoretical stability requirements in case the
enthalpy equation is iterated within a time step, or, equivalently, in the limit of
num-m_abs_MathEval
0 0.158 0.316 0.475 0.633 0.791 0.949 1.11 1.27
(a) |𝑚|/(𝜌𝑢)∞
0.66 0.7 0.73 0.76 0.8 0.83 0.87 0.9 0.93 0.97 1
(b) Colored field of 𝜌/𝜌∞ , overlaid with isolines of curl(𝑢)𝐷/𝑢∞ .
Figure 4.7: Instantaneous fields for flow past a heated circular cylinder.
small time steps. The specific enthalpy is shifted with an offset ℎ0 , so that solving
for the new unknown ℎ − ℎ0 satisfies these stability requirements. These results
were verified with simple spaceindependent tests in Section 4.4.
Method #1 is basically what has been done in all previous literature we have
seen (e.g., [20], [18], [15]), but for this approach we cannot determine the range of
stable ℎ0 values a priori. Another disadvantage is that full accuracy with a second
order timestepping scheme (BDF2) can only be achieved if the predictor is third
order accurate. This necessitates storing three instead of two previous time steps
for the enthalpy, and the thirdorder extrapolation negatively affects the stability
(see Fig. 4.2).
The other linearization of (𝜌ℎ) (method #2) does provide full accuracy when the
predictor is extrapolated from two previous time steps, in which case the error in the
linearization is negligible in the limit of small time steps. Furthermore, the stability
requirement is simply that the volumetric enthalpy be a monotonic function of the
temperature (or, equivalently, the specific enthalpy), and this leads to a range of
stable ℎ0 values that can be determined a priori. This results in a stable scheme, in
which the error of the linearization of (𝜌ℎ) becomes negligible in the limit of small
time steps. The manufactured solutions demonstrate full secondorder temporal
accuracy, without any predictor steps.
In a real flow simulation the exact range of stable values for ℎ0 cannot be de
termined a priori due the coupling of the transport equations. Note that a discon
4.6. Discussion and Conclusion 75
0.2
0.1
CL
0.0
−0.1
−0.2
1.3950
4
1.3925
1.3900
1.3875
CD
1.3850
1.3825
1.3800
3.808
3.806
Nu
3.804
3.802
40 60 80 100 120 140 160

tu∞ /D
Figure 4.8: Temporal behavior of the lift and drag coefficients (Eqs. 3.18), and Nusselt number (Eq.
4.44) for flow past a heated circular cylinder. The crosses and circles mark local minima and maxima.
The dashed horizontal lines indicate the averages over the last five periods.
tinuous Galerkin discretizations does not guarantee that the extreme values of the
numerical solution lie within physically acceptable bounds, so that the numerical
temperature range is not known beforehand. Nevertheless, there are useful guide
lines: for method #1, ℎ − ℎ0 should be close to zero; for method #2, ℎ0 should
be sufficiently large. Once a stable ℎ0 value is found, our numerical experiments
in Section 4.5 suggest that the exact value of ℎ0 has little bearing on the overall
accuracy.
Comparing the present manufactured solution to the one from the previous
chapter suggests that the temporal finite difference approximation for 𝜕𝜌/𝜕𝑡 (Eq.
4.12) is a significant source of error (as explained at the end of section 4.5.1),
which would be in line with previous literature. An elegant improvement could be
to approximate 𝜕𝜌/𝜕𝑡 by the spatial discretization of 𝜕(𝜌ℎ)/𝜕𝑡, as mentioned in
section 4.2.
4 We are now finally ready to look at the bigger picture, and reconsider which
combination of transport equation (primitive or conservative) and unknown (ℎ or 𝐻)
is most suitable for discretization in lowMach number flows. Table 4.5 summarizes
the findings in this chapter. Solving for the volumetric enthalpy would also have
been a reasonable option. Given how much depends on the enthalpy offset, it is
surprising that we have not seen this notion mentioned in previous literature.
A. Derivations of the Results in Section 4.3.1

The error estimates that were presented in Section 4.3.1 are derived here. To ease
the notation, let
1 ∗
𝐾 ≔ 𝜌ℎ∗ and 𝐵 ≔ 𝜌ℎℎ , (4.45)
2
so that the Taylor expansion in Eq. 4.20 becomes
2 3
𝜌𝑛 = 𝜌∗ − 𝐾𝜖 ∗ + 𝐵𝜖 ∗ + 𝒪 (𝜖 ∗ ) . (4.46)
A.1. Derivations for Method #1

To derive Eq. 4.24, subtract Eq. 4.16 from Eq. 4.17 to get
𝛿𝑡 [1]
𝜌∗ ℎ[1] − 𝜌𝑛 ℎ𝑛 = − 𝜆𝜖 . (4.47)
𝛾0
Substituting Eq. 4.46 gives
2 𝛿𝑡 [1]
𝜌∗ 𝜖 [1] + (𝐾𝜖 ∗ + 𝒪 (𝜖 ∗ )) ℎ𝑛 = − 𝜆𝜖 , (4.48)
𝛾0
∗
𝜖 [1] −𝐾
∗
=( ) ℎ𝑛 + 𝒪 (𝜖 ∗ ) . (4.49)
𝜖 𝜌 + 𝜆 (𝛿𝑡/𝛾0 )
Since ℎ𝑛 = ℎ∗ − 𝜖 ∗ , this is indeed equivalent to Eq. 4.24.

A. Derivations of the Results in Section 4.3.1 77
Table 4.5: Overview of problems that must be overcome in various areas, depending on the form of the
transport equation (primitive or conservative), and the choice of the unknown variable (primitive ℎ or
conserved 𝐻 ≔ 𝜌ℎ). A blank cell indicates that the treatment is straightforward.
conservative equation primitive equation

solve for 𝐻 solve for ℎ solve for ℎ
temporal term
d/d𝑡
can be uncon
ditionally stable
4
with a negligible
error*
advection term unclear, non

trivial with FEM
diffusion term needs extra term** :

modification** 𝛼𝐝 ⋅ ∇ℎ
mapping from requires special

unknown to care†
fluid properties
replacing 𝜕𝜌/𝜕𝑡 depends on en depends on en

in continuity thalpy offset ℎ0 thalpy offset ℎ0
equation by (untested) (untested)
enthalpy
discretization‡
*
As shown in sections 4.3 and 4.4, using method #2, and depending on the
enthalpy offset ℎ0 .
**
See the discussion in section 4.1.2.
†
Can be done with a suitable enthalpy offset ℎ0 (ref. Fig. 4.1), or with a separate
density predictor, as discussed in section 4.1.2.
‡
As suggested by Irrenfried [22, 23], see section 4.2.
This result can now be used to derive the error estimates in Eq. 4.23a. Using
the fact that ℎ[1] = ℎ𝑛 + 𝜖 [1] , we can write
𝜌∗ ℎ[1] = 𝜌𝑛 ℎ𝑛 − (𝜌𝑛 − 𝜌∗ ) ℎ𝑛 + 𝜌∗ 𝜖 [1] , (4.50)
which, upon substitution into Eq. 4.17, gives

𝑞
(𝜌ℎ) + ∑ (𝜌ℎ)
𝛿𝑡 𝛿𝑡
𝑖=1 (4.51)
𝛾0 𝑛 𝛾0 𝛿𝑡
= − (𝜆 − (𝜌 − 𝜌∗ )) ℎ𝑛 + 𝑄𝑛 − (𝜌∗ + 𝜆) 𝜖 [1] .
𝛿𝑡 𝛿𝑡 𝛾0
4 Using the error estimate in Eq. 4.49,

𝑞
(𝜌ℎ) + ∑ (𝜌ℎ)
𝛿𝑡 𝛿𝑡
𝑖=1
𝛾0 𝑛 2
(4.52)
= − (𝜆 − (𝜌 − 𝜌∗ + 𝜌ℎ∗ 𝜖 ∗ )) ℎ𝑛 + 𝑄𝑛 + 𝒪 (𝜖 ∗ /𝛿𝑡)
𝛿𝑡
2 2
= − (𝜆 + 𝒪 (𝜖 ∗ /𝛿𝑡)) ℎ𝑛 + 𝑄𝑛 + 𝒪 (𝜖 ∗ /𝛿𝑡) ,
where the second equality follows from the Taylor expansion in Eq. 4.20.
A.2. Derivations for Method #2

In analogy with the previous subsection, Eq. 4.26 is derived by subtracting Eq.
4.16 from Eq. 4.19 to get
∗ ∗ 𝛿𝑡 [2]
(𝜌ℎ)ℎ ℎ[2] − 𝜌𝑛 ℎ𝑛 = (ℎ2 𝜌ℎ ) − 𝜆𝜖 . (4.53)
𝛾0
Upon substituting Eq. 4.46, this becomes
∗ 2 3 ∗ 𝛿𝑡 [2]
(𝜌ℎ)ℎ ℎ[2] − ℎ𝑛 𝜌∗ + 𝐾ℎ𝑛 𝜖 ∗ − 𝐵ℎ𝑛 𝜖 ∗ + 𝒪 (𝜖 ∗ ) = (ℎ2 𝜌ℎ ) − 𝜆𝜖 . (4.54)
𝛾0
∗
Substituting (𝜌ℎ)ℎ = 𝜌 + 𝐾ℎ on the lefthand side, and (ℎ2 𝜌ℎ ) = 𝐾ℎ∗ (ℎ𝑛 + 𝜖 ∗ ) on
the righthand side, gives
∗ 2 3 𝛿𝑡 [2]
(𝜌 + 𝐾ℎ) ℎ[2] − ℎ𝑛 𝜌∗ + 𝐾ℎ𝑛 𝜖 ∗ − 𝐵ℎ𝑛 𝜖 ∗ + 𝒪 (𝜖 ∗ ) = 𝐾ℎ∗ (ℎ𝑛 + 𝜖 ∗ ) − 𝜆𝜖 .
𝛾0
(4.55)
This can be rearranged to
2 𝛿𝑡 [2] 3
𝜌∗ (ℎ[2] − ℎ𝑛 ) + 𝐾ℎ∗ (ℎ[2] − ℎ𝑛 ) − 𝐾(ℎ∗ − ℎ𝑛 )𝜖 ∗ − 𝐵ℎ𝑛 𝜖 ∗ + 𝜆 𝜖 = 𝒪 (𝜖 ∗ ) .
𝛾0
(4.56)
A. Derivations of the Results in Section 4.3.1 79
Using the definitions for 𝜖 [2] and 𝜖 ∗ , this becomes

∗
𝛿𝑡 2 3
(𝜌 + 𝐾ℎ + 𝜆) 𝜖 [2] − (𝐾 + 𝐵ℎ𝑛 ) 𝜖 ∗ = 𝒪 (𝜖 ∗ ) . (4.57)
𝛾0
Recalling ℎ𝑛 = ℎ∗ − 𝜖 ∗ , this is indeed equivalent to Eq. 4.26.

This result can now be used to derive the error estimates in Eq. 4.23b. Upon
substituting
∗
(𝜌ℎ)ℎ ℎ[2] = 𝜌∗ ℎ[2] + 𝐾ℎ∗ ℎ[2] = 𝜌𝑛 ℎ𝑛 − (𝜌𝑛 − 𝜌∗ ) ℎ𝑛 + 𝜌∗ 𝜖 [2] + 𝐾ℎ∗ ℎ[2] , (4.58)
Eq. 4.19 becomes

𝑞
𝛾0 𝑛 𝑛
(𝜌 ℎ − (𝜌𝑛 − 𝜌∗ ) ℎ𝑛 + 𝜌∗ 𝜖 [2] + 𝐾ℎ∗ ℎ[2] ) + ∑
𝛾𝑖
(𝜌ℎ)
𝑛−𝑖 4
𝛿𝑡 𝛿𝑡
𝑖=1
𝛾0 2 (4.59)
= 𝐾ℎ∗ − 𝜆ℎ[2] + 𝑄𝑛
𝛿𝑡
𝛾0
= 𝐾ℎ∗ (ℎ[2] − (𝜖 [2] − 𝜖 ∗ )) − 𝜆 (ℎ𝑛 + 𝜖 [2] ) + 𝑄𝑛 ,
𝛿𝑡
𝑞
(𝜌ℎ) + ∑ (𝜌ℎ)
𝛿𝑡 𝛿𝑡
𝑖=1 (4.60)
𝛾0 𝛿𝑡
= ((𝜌𝑛 − 𝜌∗ ) ℎ𝑛 − 𝜌∗ 𝜖 [2] + 𝐾ℎ∗ (𝜖 ∗ − 𝜖 [2] ) − 𝜆𝜖 [2] ) − 𝜆ℎ𝑛 + 𝑄𝑛 .
𝛿𝑡 𝛾0
Eq. 4.57 can now be used to eliminate 𝜖 [2] :

𝑞
(𝜌ℎ) + ∑ (𝜌ℎ)
𝛿𝑡 𝛿𝑡
𝑖=1
𝛾0 2 3
= ((𝜌𝑛 − 𝜌∗ ) ℎ𝑛 + 𝐾ℎ∗ 𝜖 ∗ − (𝐾 + 𝐵ℎ𝑛 ) 𝜖 ∗ + 𝒪 (𝜖 ∗ )) − 𝜆ℎ𝑛 + 𝑄𝑛 (4.61)
𝛿𝑡
𝛾0 2 3
= ((𝜌𝑛 − 𝜌∗ − 𝐵𝜖 ∗ ) ℎ𝑛 + 𝐾𝜖 ∗ (ℎ∗ − 𝜖 ∗ ) + 𝒪 (𝜖 ∗ )) − 𝜆ℎ𝑛 + 𝑄𝑛
𝛿𝑡
𝛾0 2 3
= ((𝜌𝑛 − 𝜌∗ − 𝐵𝜖 ∗ + 𝐾𝜖 ∗ ) ℎ𝑛 + 𝒪 (𝜖 ∗ )) − 𝜆ℎ𝑛 + 𝑄𝑛 .
𝛿𝑡
Finally, using the Taylor series in Eq. 4.46,
𝑞
𝛾0 𝑛 𝛾𝑖 𝑛−𝑖 𝛾0 3 3
(𝜌ℎ) + ∑ (𝜌ℎ) = (𝒪 (𝜖 ∗ ) ℎ𝑛 + 𝒪 (𝜖 ∗ )) − 𝜆ℎ𝑛 + 𝑄𝑛 , (4.62)
𝛿𝑡 𝛿𝑡 𝛿𝑡
𝑖=1
which implies Eq. 4.22 with Eq. 4.23b.

80 References
References
tational Physics 425, 109877 (2021).
[2] G. Volpe, Performance of compressible flow codes at low Mach numbers, AIAA
Journal 31, 49 (1993).
[4] F. Föll, S. Pandey, X. Chu, C.D. Munz, E. Laurien, and B. Weigand, High
Fidelity Direct Numerical Simulation of Supercritical Channel Flow Using Dis
continuous Galerkin Spectral Element Method, in High Performance Comput
4 ing in Science and Engineering ’ 18 (Springer International Publishing, Cham,
2019) pp. 275–289.
[5] R. Barney, R. Nourgaliev, J. P. Delplanque, and R. McCallen, FullyImplicit,
HighOrder, Reconstructed Discontinuous Galerkin Method for Supercritical
Fluid Flows, in Proceedings of the International Conference on Mathemat
ics and Computational Methods applied to Nuclear Science and Engineering
(M&C) (Portland, OR, USA, 2019) pp. 377–386.
[6] A. Nigro, C. De Bartolo, R. Hartmann, and F. Bassi, Discontinuous Galerkin
solution of preconditioned Euler equations for very low Mach number flows,
International Journal for Numerical Methods in Fluids 63, 449 (2010).
[7] F. Bassi, A. Crivellini, D. A. D. Pietro, and S. Rebay, An artificial compressibil
ity flux for the discontinuous Galerkin solution of the incompressible Navier
[8] A. Crivellini, V. D’Alessandro, and F. Bassi, A SpalartAllmaras turbulence model
implementation in a discontinuous Galerkin solver for incompressible flows,
[10] L. Botti and D. A. Di Pietro, A pressurecorrection scheme for convection
dominated incompressible flows with discontinuous velocity and continuous
pressure, Journal of Computational Physics 230, 572 (2011).
References 81
[13] F. Bassi, L. Botti, A. Colombo, A. Ghidoni, and F. Massa, Linearly implicit

Rosenbrocktype RungeKutta schemes applied to the Discontinuous Galerkin
solution of compressible and incompressible unsteady flows, Computers & Flu
ids 118, 305 (2015).
[14] L. Pesch and J. van der Vegt, A discontinuous Galerkin finite element dis
cretization of the Euler equations for compressible and incompressible fluids,
[16] B. Klein, A highorder Discontinuous Galerkin solver for incompressible and
lowMach number flows, Ph.D. thesis, Technische Universität, Darmstadt 4
(2015).
[17] D. T. Banuti, Crossing the Widomline – Supercritical pseudoboiling, The Jour
nal of Supercritical Fluids 98, 12 (2015).
[18] J. W. R. Peeters, Turbulence and turbulent heat transfer at supercritical pres
sure, doctoral thesis, Delft University of Technology (2016).
[19] H. N. Najm, P. S. Wyckoff, and O. M. Knio, A Semiimplicit Numerical Scheme
for Reacting Flow: I. Stiff Chemistry, Journal of Computational Physics 143,
381 (1998).
[20] H. Nemati, Direct numerical simulation of turbulent heat transfer to fluids at
supercritical pressures, doctoral thesis, TU Delft (2016).
[21] F. Nicoud, Conservative HighOrder FiniteDifference Schemes for LowMach
Number Flows, Journal of Computational Physics 158, 71 (2000).
[22] C. Irrenfried, DNS and experimentally based modelling of convective turbulent
near wall heat transfer at high Prandtl numbers, Doctoral thesis, Graz Univer
sity of Technology, Fakultät für Maschinenbau und Wirtschaftswissenschaften,
Institut für Strömungslehre und Wärmeübertragung (2019).
[23] C. Irrenfried, Convective turbulent near wall heat transfer at high Prandtl num
bers. Monograph Series: Computation in Engineering and Science, Vol. 37
(Technischen Universität Graz, 2020).
[24] R. Span and W. Wagner, A New Equation of State for Carbon Dioxide Covering
the Fluid Region from the TriplePoint Temperature to 1100 K at Pressures
up to 800 MPa, Journal of Physical and Chemical Reference Data 25, 1509
(1996).
[25] I. H. Bell, J. Wronski, S. Quoilin, and V. Lemort, Pure and Pseudopure
Fluid Thermophysical Property Evaluation and the OpenSource Thermophys
ical Property Library CoolProp, Industrial & Engineering Chemistry Research
53, 2498 (2014).
82 References
[26] A. Hennink, Finite Difference Methods for the Nonlinear Enthalpy Equation,
GitHub. [Online] Accessed on 20191105 (2019).
[27] L. Shunn, F. Ham, and P. Moin, Verification of variabledensity flow solvers
using manufactured solutions, Journal of Computational Physics 231, 3801
(2012).
[28] K. Salari and P. Knupp, Code Verification by the Method of Manufactured So
lutions, Tech. Rep. (Sandia National Labs., Albuquerque, NM (US), Livermore,
CA (US), 2000).
[29] J.M. Shi, D. Gerlach, M. Breuer, G. Biswas, and F. Durst, Heating effect on
steady and unsteady horizontal laminar flow of air past a circular cylinder,
Physics of Fluids 16, 4331 (2004).
4
[30] A.B. Wang, Z. Trávníček, and K.C. Chia, On the relationship of effective
Reynolds number and Strouhal number for the laminar vortex shedding of a
heated circular cylinder, Physics of Fluids 12, 1401 (2000).
5
Channel Flow and
Large Eddy Simulation
5.1. Introduction and Governing Equations

The idea of a large eddy simulation (LES) is to take spatially filtered quantities as
the unknowns, in which the smallest scales of the flow have been removed. These
filtered quantities are denoted by an overline (e.g., 𝑝, 𝑚1 , …). On a conceptual
level, a filtered quantity 𝜙 can be thought of as a moving average of 𝜙, or a more
general convolution of 𝜙 with a lowpass filter, though this is not always explicitly
computed. The goal is to approximate the filtered quantities by solving some form
of the transport equations in less detail, not down to the smallest length scales.
We solve for the variables 𝐦, 𝑝, and ℎ ̃ ≔ 𝜌ℎ/𝜌, which are therefore known
as the resolved quantities. For the specific enthalpy we use a Favre average (i.e.,
weighed by the density before filtering), denoted by a tilde. This is more suitable for
primitive (nonconserved) variables. The fluid properties (̂ 𝜌, 𝜇̂, 𝛼
̂) are determined
from ℎ.̃ A hat denotes a computable variable, meaning that it is based on the
resolved quantities. The computable velocity is 𝐮 ̂ = 𝐦/̂ 𝜌, which forms the basis for
the resolved rate of strain
𝑆̂𝑖𝑗 = 𝐿 (∇̂
𝐮) , (5.1)
where the operator
1 2
𝐿 (𝐴) ≔ (𝐴 + 𝐴⊺ − trace (𝐴) 𝐼) (5.2)
2 3
takes the symmetric, deviatoric part of a matrix in three dimensions. The resolved
̂ )∇ℎ
̂ = −(𝑘/𝑐
Fourier heat flux is 𝐪 ̃ = −𝜌𝛼 ̃
̂ ∇ℎ.
𝑝
Filtering does not commute with multiplication, and the commutation errors
arise as extra terms in the transport equations. In particular, for the convective
83
84 5. Channel Flow and Large Eddy Simulation
term in the momentum equation, the difference between what we can compute
and the filtered original term is
1 1 1 1
̂ 𝐦 − 𝐮𝐦 =
𝐮 ̃ = ( − ) 𝐦𝐦 − 𝜌 𝜏SFS .
𝐦𝐦 − 𝜌𝐮𝐮 = 𝐦𝐦 − 𝜌𝐮𝐮 (5.3)
𝜌̂ 𝜌̂ 𝜌̂ 𝜌
The term
𝜏SFS ≔ 𝐮𝐮
̃ −𝐮
̃𝐮̃ (5.4)
is known as the sub filter scale stress tensor. As the name suggests, it acts in
a similar manner to the viscous stress, and it is usually modeled in terms of the
resolved rate of strain. Similarly, for the advection term in the enthalpy equation,
ℎ𝐦 ̃𝐮 − ℎ𝐮)
̃ − ℎ𝐦 = 𝜌 (ℎ̃ ̃ ≕ 𝐪SFS , (5.5)
which is called the sub filter scale heat flux1 , and it is usually modeled in terms of
the resolved Fourier heat flux. The subfilter stress (𝜏SFS ) and heat flux (𝐪SFS ) play
a central role in large eddy simulations of both incompressible and compressible
5 flows.
Under the assumption that filtering commutes with spatial and temporal deriva
tives, the filtered transport equations are
𝜕̂
𝜌
+ ∇ ⋅ 𝐦 =𝑅(0) , (5.6a)
𝜕𝑡
𝜕𝐦
+ ∇ ⋅ (̂
𝐮 𝐦) − ∇ ⋅ (2̂ ̂ + ∇𝑝 − 𝐅 = − ∇ ⋅ (𝜌𝜏SFS ) + ∇ ⋅ 𝑅(1)
𝜇𝑆) (5.6b)
𝜕𝑡
+ 2∇ ⋅ (𝑅visc + 𝐴visc + 𝐵visc ) ,
𝜕̂𝜌ℎ̃
̃ 𝐦) − ∇ ⋅ (𝜌𝛼
+ ∇ ⋅ (ℎ ̃ − 𝑄 =∇ ⋅ 𝐪SFS + 𝑅(2) + ∇ ⋅ (𝐀fhf + 𝐁 fhf ) . (5.6c)
̂ ∇ℎ)
𝜕𝑡
All noncomputable (subfilter) terms are gathered on the righthand sides:
𝜕
𝑅(0) ≔ (̂
𝜌 − 𝜌) , (5.7a)
𝜕𝑡
1 1
𝑅(1) ≔ ( − ) 𝐦𝐦 , (5.7b)
𝜌̂ 𝜌
𝜕
𝑅(2) ≔ ((̂ ̃ ,
𝜌 − 𝜌) ℎ) (5.7c)
𝜕𝑡
𝑅visc ≔ 𝜇̂ (𝑆̃ − 𝑆)
̂ , (5.7d)
𝐴 visc
≔ 𝜌 (𝜈̃𝑆 − 𝜈̃ 𝑆)
̃ , (5.7e)
𝐵 visc
≔ (𝜇 − 𝜇̂) 𝑆̃ , (5.7f)
𝐀 fhf ̃ −𝛼
≔ 𝜌 (𝛼∇ℎ ̃ ,
̃ ∇ℎ) (5.7g)
𝐁 fhf
≔ (𝜌𝛼 − 𝜌𝛼)
̂ ∇ℎ ̃. (5.7h)
1 Most ̃𝐮
other literature refers to 𝜌 (𝑇 ̃ as the subfilter heat flux, or the subfilter temperature flux.
̃ − 𝑇𝐮)
That definition works well when working with the ideal gas law.
5.1. Introduction and Governing Equations 85
The subfilter terms due to the viscous stress sum to
𝑅visc + 𝐴visc + 𝐵visc = 𝜇̂ (𝑆̃ − 𝑆)

̂ + 𝜌𝜈𝑆 − 𝜌𝜈𝑆̃ + (𝜇 − 𝜇̂) 𝑆̃
(5.8)
= 𝜇𝑆 − 𝜇̂𝑆̂
to account for the difference between the filtered original ∇ ⋅ (𝜇𝑆) and the com
putable ∇ ⋅ (̂ ̂ Similarly, the subfilter terms due to the Fourier heat flux (‘fhf’)
𝜇𝑆).
sum to
𝐀fhf + 𝐁 fhf = 𝜌𝛼∇ℎ − 𝜌𝛼∇
̂ ℎ̃. (5.9)
The terms in Eqs. 5.4, 5.5, and 5.7a–5.7h differ strongly in magnitude, so some
can be neglected.
Most other authors have considered the righthand side of Eq. 5.8 (resp. Eq.
5.9) as a single subfilter term, rather than splitting it into several terms as is done
here. The present decomposition shows more explicitly that each term in Eqs.
5.7a–5.7h falls into one of the following categories.
1. 𝐴visc and 𝐀fhf are commutation errors between the Favre filter, and multi 5
plication by a transport property. These occur in LES with inhomogeneous
transport properties. It is standard practice to neglect them (see, e.g., the
discussions in [1, 2]).
2. 𝐵visc and 𝐁 fhf are due to the difference between the filtered and the com
putable transport properties. It is standard practice to neglect them (e.g.,
[3]).
3. The terms 𝑅0 , 𝑅1 , 𝑅2 , and 𝑅visc result from the difference between the com
putable density and the filtered density. For 𝑅visc this can be made clear by
rewriting it as
1 1
𝑅visc = 𝜇̂ 𝐿 [∇ (̃ ̂ )] = 𝜇̂ 𝐿 [∇ ((
𝐮−𝐮 − ) 𝐦)] . (5.10)
𝜌 𝜌̂
These terms are particular to LES of lowMach flows; they do not occur in
incompressible flows (where 𝜌 is usually constant), or in highMach com
pressible flows (where 𝜌 is computable as one the resolved quantities).
We strongly suspect that these terms can be neglected, in analogy to those
in category 2.
4. The turbulent stress tensor 𝜏SFS and the subfilter temperature flux 𝐪SFS are
both commutation errors between the Favre filter, and multiplication by an
advecting field.
In other words, all terms are neglected in lowMach flows, except those in the last
category, which also dominate in incompressible flows with constant fluid proper
ties.
5.2. Subfilter Scale models

The oldest and most common form of LES is to model the effect of the subfilter
stress tensor 𝜏SFS (Eq. 5.4) as a function of the resolved quantities, which Sagaut
[4] calls a ‘functional’ model. In particular, this is usually done in terms of the
resolved rate of strain, so that
̂ .
− ∇ ⋅ (𝜌𝜏SFS ) ≈ ∇ ⋅ (2𝜇sfs 𝑆) (5.11)
This has the same form as the viscous stress, and is therefore easy to implement:
the momentum equation is the same as for direct numerical simulation, except that
the kinematic viscosity becomes
𝜈 = 𝜈molec + 𝜈sfs , (5.12)
where 𝜈molec is the ‘molecular’ (i.e., actual, physical) viscosity, and 𝜈sfs is the extra
viscous effect (‘eddy viscosity’) due to flow structures that are smaller than the LES
filter width.
The main challenge is to model 𝜈sfs in terms of computable quantities. Obvi
5 ously it cannot be constant, because that would be come down to simulating the
equivalent flow at a lower Reynolds number. Instead 𝜈sfs is expressed in terms
of the velocity gradient. Since the fluid is assumed to have no memory, 𝜈sfs only
depends on the local value of ∇𝐮. As there is no clear best choice, we mention
several models below.
The subfilter viscosity depends on the symmetric part of the computable veloc
ity gradient, given by
1
𝒮𝑖𝑗 ≔ (∇𝑖 𝑢̂𝑗 + ∇𝑗 𝑢
̂𝑖 ) . (5.13)
2
This is not tracefree, but otherwise it is equivalent to the resolved rate of strain
𝑆̂ (Eq. 5.1). The antisymmetric part of the velocity gradient corresponds to rigid
rotation, in which the relative position of the fluid particles does not change, and
which is assumed not to result in turbulent dissipation, and therefore not to affect
𝜈sfs .
Since the subfilter stress is supposed to model the physical phenomenon of
turbulent dissipation, it stands to reason that 𝜈sfs be a function of the invariants of
𝒮. The main invariants of an arbitrary rank2 tensor 𝐴 in three dimensions with
eigenvalues 𝜆𝑖 are trace (𝐴) = 𝜆1 + 𝜆2 + 𝜆3 , trace (𝐴2 ) = 𝜆21 + 𝜆22 + 𝜆23 , and
trace (𝐴3 ) = 𝜆31 + 𝜆32 + 𝜆33 . The numerical velocity in an incompressible flow is
either exactly solenoidal, or very close to solenoidal (depending on the spatial dis
cretization). Even in lowMach number compressible flows, we can assume that
trace (𝑆) = ∇ ⋅ 𝐮 has a much smaller impact on the turbulent dissipation than the
deviatoric part of 𝒮, and in any case most LES models for compressible flows are
straightforward extensions of incompressible models, so the first invariant does not
play a role.
The oldest and simplest model therefore expresses 𝜈sfs in terms of the second
main invariant of 𝒮:
sfs
𝜈smag = 𝐴2smag √2 trace (𝒮 2 ) = 𝐴2smag √2𝒮𝑘𝑙 𝒮𝑘𝑙 = 𝐴2smag √2 ‖𝒮‖Frob . (5.14)
5.2. Subfilter Scale models 87
Here ‖⋅‖Frob is the Frobenius norm. This is known as the Smagorinsky model, and
it the most well known, the most researched, and probably still the most commonly
used.
The factor 𝐴smag still needs to be determined. It has the dimension of length,
and it is directly related to the filter width: larger values of 𝐴smag damp more of the
small scales of the flow, lowering the maximum frequency in the Fourier transform
of the computed velocity. This is made explicit by writing
𝐴smag = 𝐶smag Δ , (5.15)
where Δ is the LES filter width, and 𝐶smag is known as the Smagorinsky constant.
In practice Δ is usually coupled to the resolution of the spatial discretization. Given
a characteristic element length 𝓁, the universally accepted engineering practice is
to let Δ = 2𝓁. This typically results in both a modelling error (due to the subfilter
flow structures, which are smaller than Δ), and a discretization error (because 𝓁 is
not much smaller than Δ).
A theoretical analysis of isotropic turbulence suggests that the Smagorinsky con
stant is indeed the same for all flows [5]. Berselli et al. [6, pp. 71–77] estimate 5
it at 𝐶smag ≈ 0.17, whereas Sagaut [4, pp. 113–124] gives 𝐶smag ≈ 0.148 or
𝐶smag ≈ 0.18. Extensive numerical experience has shown that these values are too
large for almost all flows and discretizations; it is too dissipative. Typically 𝐶smag
is lowered to approximately 0.1 to achieve the right amount of total kinetic energy
removal, though this does not guarantee an optimal local structure of turbulent
dissipation.
Another shortcoming of the Smagorinsky model is that the dissipation does not
automatically vanish in the laminar layer near the wall. One possible solution is
to multiply 𝐴smag by a scaling factor. By far the most common is the Van Driest
damping function
+
𝑓 (𝑦 + ) = 1 − 𝑒−𝑦 /25 , (5.16)
where 𝑦 + is the dimensionless wall distance, based on the wall shear stress (as
defined in Eq. 5.32 below). Another way of looking at this is that the LES method
implies a spatial filter through convolution with a filtering kernel, but a symmetric
kernel is not possible at the wall. The filter width is therefore gradually reduced to
zero near the wall.
3
It can be shown theoretically that 𝜈sfs ∝ (𝑦 + ) , though the above Van Driest
2
damping produces 𝜈sfs ∝ (𝑦 + ) . Piomelli et al. [7] suggested a different damping
function that achieves the correct asymptotic behavior near the wall, but found
no significant improvement over Van Driest damping. See Inagaki [8] for a more
recent review.
Nicoud and Ducros [9] have argued that the Smagorinsky model is fundamen
tally flawed because it does not include the effect of rotation, and proposed the
alternative WALE (walladaptive local eddy) viscosity model, which is based on the
square of the velocity gradient, that is,
(𝐵2 )𝑖𝑗 ≔ (∇𝑖 𝑢𝑘 ) (∇𝑘 𝑢𝑗 ) . (5.17)

Denoting
𝐹 ≔ 𝐿 [𝐵2 ] , (5.18)
we have
3/2 3
(𝐹𝑘𝑙 𝐹𝑘𝑙 ) ‖𝐹‖Frob
sfs
𝜈WALE = 𝐴2WALE 5/2 5/4
= 𝐴2WALE 5 5/2
. (5.19)
(𝒮𝑘𝑙 𝒮𝑘𝑙 ) + (𝐹𝑘𝑙 𝐹𝑘𝑙 ) ‖𝒮‖Frob + ‖𝐹‖Frob
The idea is that ‖𝐹‖Frob and ‖𝑆‖Frob scale differently near the wall, so it is possible
to choose the exponents in Eq. 5.19 in such a way that we get the appropriate
5/2
sfs
𝜈WALE = 𝒪((𝑦 + )3 ) behavior. The term ‖𝐹‖Frob is only there to ensure that the
denominator is nonzero, giving the expression a somewhat artificial look, though
it has performed well in practice. (See [10] for an example with a discontinuous
Galerkin method.)
In analogy to Eq. 5.15, the prefactor is related to the filter width:
𝐴WALE = 𝐶WALE Δ . (5.20)

5
Nicoud and Ducros [9] chose the constant 𝐶WALE such that Eqs. 5.14 and 5.19
predict the same amount of turbulent dissipation in homogeneous, isotropic tur
bulence. This calibration gave 𝐶WALE ≈ 0.5. Garnier et al. [3, p. 88] later
2
found (𝐶WALE /𝐶smag ) ≈ 10.6, which comes down to 𝐶WALE ∈ (0.326, 0.59) for
𝐶smag ∈ (0.1, 0.18). The widely used ANSYSFluent [11] software package sets
𝐶WALE = 0.325 by default.
Finally, we consider the QR model, which has been introduced far more recently
than the Smagorinsky and the WALE models by Verstappen [12] in 2011. Rather
than focussing on physical reasoning, he assumed that the subfilter scale flow
should not influence the larger structures, and then looked for the minimal value of
𝜈sfs that meets this constraint. This involved an interesting analysis of the filtered
NavierStokes equations, resulting in2
|𝒮𝑘𝑙 𝒮𝑙𝑚 𝒮𝑚𝑘 | trace (𝒮 3 )

sfs
𝜈QR = 𝐴2QR = 𝐴2QR | | , (5.21)
𝒮𝑘𝑙 𝒮𝑘𝑙 trace (𝒮 2 )
making it the only model in this section that involves the third invariant of 𝒮. As
before, we can write
𝐴QR = 𝐶QR Δ . (5.22)
Verstappen [12] finds 𝐶QR = 1/𝜋 = 0.32. The subfilter viscosity has proper
𝒪((𝑦 + )3 ) behavior near the wall for constant 𝐶QR .
The prefactors (𝐴smag , 𝐴WALE , 𝐴QR ) in Eqs. 5.14, 5.19, 5.21 can also be deter
mined from the simulation itself, rather than depending on userdefined parame
ters. This is done with the socalled ‘dynamic model’, originally introduced for the
Smagorinsky model by Germano et al. [13], who projected the numerical solution
2 Notethat the trace is the sum of the eigenvalues and 𝒮 is symmetric, so trace (𝒮 2 ) ≥ 0. The denomi
sfs
nator could theoretically vanish, but 𝜈QR is well behaved in the limit 𝑆𝑖𝑗 → 0.
5.3. Numerical Simulation 89
onto a coarser solution space (e.g., a coarser grid). The LES model (in their case Eq.
5.14) is assumed to be valid on both the fine and coarse grids. Since the prefactor
𝐴 is assumed not to change between these two flow scales, it can be computed by
comparing the coarse and fine solutions. The dynamic Smagorinsky model displays
3
the correct 𝒪 ((𝑦 + ) ) asymptotic behavior, and has widely been found to yield far
better results than the ‘constant’ Smagorinsky model (i.e., with a userdefined con
stant). (See, e.g., the comparisons and the very clear discussion in Vreman [14],
who also studied compressible flows.)
In the context of a highorder finite element discretization, the projection onto a
coarse grid could be replaced by a projection onto a lowerorder polynomial space.
This has been done with a discontinuous Galerkin method by Abbà et al. [15],
though it is not clear whether this is better than projecting onto larger elements.
It has also been implemented in DGFlows, but we do not use it in this chapter.
These different types of projections are reminiscent of 𝑝multigrid and ℎmultigrid
solvers, which were mentioned in section 2.6.
A dynamic model also makes it easier to use a symmetric positive definite matrix
instead of a scalar for 𝜇sfs in Eq. 5.11, which could in principal handle anisotropic 5
turbulence better. Abbà et al. [15] found that an anisotropic 𝜇sfs tensor indeed
gives superior results for in wallbounded flow. Nevertheless using a scalar 𝜇sfs is
far more common.
For the subfilter heat flux we only consider the simplest model, which is based
on a turbulent Prandtl number Prt that relates the effect of 𝐪SFS to 𝛕SFS . First we
assume that the subfilter heat flux results in a net turbulent dissipation, analogous
to Eq. 5.11:
sfs ̂
∇ ⋅ 𝐪SFS ≈ ∇ ⋅ ((𝜌𝛼) ∇ℎ) . (5.23)
sfs
Then we express (𝜌𝛼) in terms of 𝜈sfs with a turbulent Prandtl number:
sfs 𝜌̂ 𝜈sfs
(𝜌𝛼) = , (5.24)
Prt
where we have used the computable density 𝜌̂ because that is the only directly
available value in a lowMach number simulation. This is easy to implement: com
pared to direct numerical simulation, only the fluid property 𝑘/𝑐𝑝 = 𝜌𝛼 needs to be
adjusted to
molec sfs
𝜌𝛼 = (𝜌𝛼) + (𝜌𝛼) , (5.25)
molec
where (𝜌𝛼) is the ‘molecular’ (physical) fluid property.
The turbulent Prandtl number is usually chosen in the range (0.3, 0.9) [3, p. 84].
Lesieur [16] derived a theoretical value of Prt = 0.6. It can also be estimated with
a dynamic procedure.
5.3. Numerical Simulation

5.3.1. Discretization
The spatial discretization is as described in the previous chapters, except for the
pressure stabilization for equalorder polynomial spaces for the pressure and the
mass flux. The problem with the penalty term in the pressure equation (𝑎stab in Eq.
2.7) is that it makes the linear system too stiff, and our implementation DGFLows
is not wellequipped to deal with this. See section 3.5.1 for a discussion of this
problem.
The calculations with equalorder discretizations in this chapter are therefore
stabilized by penalizing violations of the continuity equation, as has been suggested
by Krank et al. [17]. (See also the remarks in section 2.2.) We incorporate their
penalty terms into the momentum equation, so that we do not require a separate
projection step for the momentum, as they do. That is, the terms
𝑎∗ (𝐰, 𝐯) = ∑ ∫ 𝜃 𝓁𝑇 (∇ ⋅ 𝐰) (∇ ⋅ 𝐯) + ∑ ∫ {𝜃} J𝐧𝐹 ⋅ 𝐰K J𝐧𝐹 ⋅ 𝐯K (5.26)

𝑇 𝐹
𝑇∈𝒯 𝐹∈ℱD,i
and
𝑙 ∗ (𝐯) = ∑ ∫ 𝜃 (𝐧 ⋅ 𝐯) (𝐧 ⋅ 𝐦D ) (5.27)
𝐹
𝐹∈ℱD
5
are added to the bilinear and linear operators in the discrete momentum equation
(given by Eq. 2.10 for timeindependent flows). Here 𝓁𝑇 is a characteristic length
of the element; we use the value
1/𝑑
‖𝑇‖leb
𝓁𝑇 = (5.28)
1 + 𝒫𝑚
for a 𝑑dimensional element with polynomial order 𝒫𝑚 for the mass flux, and do
not investigate other estimates.
Note that 𝑎∗ and 𝑙 ∗ couple the directions of the mass flux. Previously they
were only coupled implicitly in the pressure projection (and, less importantly, in the
viscous stress). We solve the momentum equation for all directions simultaneously.
The penalty parameter 𝜃 has the same dimension as the velocity, and Krank
et al. [17] used the natural 𝜃 = |𝐮| ≔ √𝐮 ⋅ 𝐮. This makes 𝑎∗ and 𝑙 ∗ nonlinear in
𝐮, which could presumably be handled efficiently by basing 𝜃 on an extrapolation
of 𝐮 from previous time steps. More problematically, the term 𝑎∗ would need to be
reassembled at every time step. In our implementation this takes a significant part
of the total calculation time.
We therefore use the alternative form
𝜃 = ⟨|𝐮|⟩ , (5.29)
where ⟨⋅⟩ denotes the average over time and over the homogenous directions. This
means that 𝜃 is timeindependent, so that 𝑎∗ and 𝑙 ∗ can be precomputed. In Eq.
5.26 the penalty parameter is an average of the neighbors of the faces, but taking
the maximum value would also have seemed reasonable to us. For the present
calculations it makes no difference, since the above form of 𝜃 is virtually continuous
at the element boundaries.
5.3. Numerical Simulation 91
5.3.2. Including a Variable Density

This chapter will only present a simulation with a constant density. As explained in
section 5.1, LES of lowMach number flow requires models for the same subfilter
scale terms as for constantdensity flows. The results for incompressible flows are
therefore directly relevant for variabledensity flows.
It would nonetheless have been interesting to include a variabledensity test
case. A lowMach number flow would pose new challenges, such as a stronger
coupling between the subfilter heat flux and the momentum equation, due to the
temperaturedependent density. Unfortunately our limited computational resources
do not allow for this.
The problem is that the transport equations in the form of Eqs. 1.1a–1.1c with
a nonconstant 𝜌 are only wellposed when there is an outflow boundary condi
tion. This would require an inlet with an expensive turbulent inflow generator.
The simulation must furthermore describe a developing flow, which requires a far
larger domain than when there are periodic boundary conditions. Nemati [18] has
nevertheless simulated developing lowMach number flow. He used a specialized
implementation that depends on the specific geometry of pipe flows, which have
one main flow direction, and where solving the Poisson equation might have been
5
sped up with a fast Fourier transform.
Of course a variabledensity flow can also be simulated with a compressible
solver at a low, nonzero Mach number, though this is probably far more expensive
than using the transport equations in the lowMach number limit. Examples include
[19], [20], and [21], who all required generic massively parallel solvers with some
of the world’s largest supercomputers. These authors performed direct numerical
simulation of CO2 at a supercritical pressure.
A computationally more efficient approach is to introduce a variable thermody
namic pressure 𝑝th . The density then becomes a function of both the temperature
𝑇 and 𝑝th , which makes the governing equations well posed with periodic boundary
conditions (i.e., a closed domain). The enthalpy equation in lowMach number limit
gets an extra term d𝑝th /d𝑡, as explained in the seminal paper by Rehm and Baum
[22]. The thermodynamic pressure is updated after each time step. Nicoud [23]
was one of the first to present a direct numerical simulation of variabledensity flow
in this manner. See also the review in Knikker [24]. Avila et al. [25] have also
included an LES model with the variable density.
These authors have all assumed an ideal gas. This is almost always done in the
literature, though a notable exception is the work by Accary et al. [26, 27], who
used a more general Van der Waals equation. We are not aware of similar papers
with an arbitrary equation of state.
A dubious alternative approach to cheap variabledensity flow calculations was
taken by surprisingly many authors [28–30], who simulated supercritical fluids in
closed domains, but ignored the effect of the thermodynamic pressure on the den
sity. That is, they left out the d𝑝th /d𝑡 term in the enthalpy equation, and computed
the equation of state at a fixed 𝑝th , so that 𝜌 was merely a function of ℎ. This re
sults in a mathematically illposed problem, because the continuity equation implies
a constant the total fluid mass (due to the closed domain), whereas the enthalpy
Twall
Ly = 2δ
‹ux› Twall
Lz = 3δ
Lx = 6δ
5 Figure 5.1: Geometry of the infinite plane channel.
equation implies a fluctuating total mass (due to the imposed heat flux and the fact
that 𝜌 = 𝜌(ℎ)). The results in these papers are therefore fundamentally irrepro
ducible.
The authors have in common that they focused on physical interpretations of
the CFD results, rather than on the numerical method. Their computations were
probably helped by the fact that they used finite difference schemes, which do not
conserve the global mass. In one case [30], we know from a private conversation
with one of the authors that the simulations were kickstarted with an artificial mass
sink in the continuity equation, to allow for the average temperature to rise after
the initial condition. It is possible that the conclusions in [30] hold despite the
mathematical inconsistency, because they are based on local fluctuations, which
are perhaps not strongly influenced by the global conservation of mass.
5.4. Test Case: Infinite Plane Channel Flow

This section describes a numerical simulation of flow between two planes. The
geometry is shown in Fig. 5.1. The origin is placed between the two planes, so that
−𝛿 < 𝑦 < 𝛿. The boundaries are periodic in the 𝑥 and 𝑧directions to approximate
flow between infinite planes.
The flow is driven by a homogeneous, constant volumetric force 𝐅 = [𝐹𝑥 , 0, 0]
and volumetric heat source 𝑄. The top and bottom planes have Dirichlet boundary
conditions for the velocity ( 𝐮|𝑦=±𝛿 = 𝟎) and the temperature ( 𝑇|𝑦=±𝛿 = 𝑇wall ). The
setup and the timeaveraged quantities are therefore symmetrical about 𝑦 = 0, and
homogeneous in the 𝑥 and 𝑧directions.
5.4. Test Case: Infinite Plane Channel Flow 93
5.4.1. Dimensionless Analysis

The following dimensionless quantities are defined in the usual way, based on the
average shear stress 𝜏wall at the wall. The characteristic friction velocity at the wall
is 𝑢𝜏 ≔ √𝜏wall /𝜌, and from that follows
the friction Reynolds number: Re𝜏 ≔ 𝑢𝜏 𝑅h /𝜈 = 𝑢𝜏 𝛿/𝜈 ; (5.30)
the turnover time: 𝑡+ ≔ 𝑡𝑢𝜏 /𝛿 ; (5.31)
+ ∗ ∗
the wall distance: 𝑦 ≔ 𝑦 𝑢𝜏 /𝜈 = (𝑦 /𝛿)Re𝜏 , (5.32)
where
flow area
𝑅h ≔ =𝛿 (5.33)
wetted perimeter
is the hydraulic radius, and 𝑦 ∗ = 𝛿 − abs(𝑦) is the dimensional distance from the
wall.
The desired Re𝜏 is chosen as a free parameter. Since the average wall friction
𝜏wall balances the volumetric force, this fixes the volumetric force at 𝐹𝑥 = 𝜏wall /𝑅h =
𝜏wall /𝛿. The average velocity in the domain is therefore a result of the computation,
not an input parameter. 5
Many authors have taken the opposite approach by choosing a bulk Reynolds
number
𝐷h 𝑢bulk 4𝑅h 𝑢bulk
Rebulk ≔ = , (5.34)
𝜈 𝜈
which is based on the hydraulic diameter 𝐷h ≔ 4𝑅h and the timeaveraged bulk
velocity
1 1
𝑢bulk ≔ ∫ 𝑢1 = ∫ 𝑢 . (5.35)
‖Ω‖leb Ω 2𝛿 |𝑦|<𝛿 1
The volumetric force is then adjusted dynamically during the simulation to get the
desired Rebulk . The disadvantage of that approach is that the results are less
reproducible, because the actual imposed force 𝐹𝑥 is unknown.
The two Reynolds numbers Re𝜏 and Rebulk are related by the DarcyWeisbach
equation
𝜌
𝐹𝑥 = 𝑓D 𝑢2 , (5.36)
2𝐷h bulk
where 𝑓D is the Darcy friction factor. Dimensionless analysis shows that 𝑓D is only
a function of Rebulk (e.g., [31, pp. 177–184]). The friction Reynolds number can
be expressed in terms of the volumetric driving force as
𝑅h 𝜌 𝑅h 𝜌 𝜏wall 𝑅h 𝜌 𝑅h 𝐹𝑥
Re𝜏 ≔ 𝑢 = √ = √ . (5.37)
𝜇 𝜏 𝜇 𝜌 𝜇 𝜌
Substituting the DarcyWeisbach equation and the definition for the bulk Reynolds
number gives Re𝜏 = Rebulk (𝑅h /𝐷h ) √(𝑅h /𝐷h ) (𝑓D /2), or
128 11.3
Rebulk = √ Re𝜏 ≈ Re𝜏 . (5.38)
𝑓D √𝑓D
We have surprisingly not seen this relation in previous literature, although it holds
for all channel geometries.
The friction factor can be estimated by using measurement data from turbulent
pipe flow. If the Reynolds number is sufficiently large, then the curvature of the wall
is negligible compared to the size of the flow structures near the wall, which jus
tifies a comparison between channels of different geometries. The semiempirical
ColebrookWhite correlation for smooth circular pipes is3
1 𝐴 𝐴/√128
= −2 log10 ( ) = −2 log10 ( ) (5.39)
√𝑓D √𝑓D Rebulk Re𝜏
with 𝐴 = 2.51. This is valid for Rebulk > 4.0 ⋅ 103 , or, equivalently, Re𝜏 > 71.
5.4.2. Initial Condition

We are interested in turbulent flow, which is not obtained with all initial conditions.
A turbulent channel flow can only be sustained above a certain critical Reynolds
number, whereas the laminar solution
5
3 𝑦 2
⎧𝐮laminar =𝑢bulk [ (1 − ( ) ) , 0, 0]
⎪ 2 𝛿
(5.40)
⎨ laminar 𝛿2 𝑄 𝑦 2
⎪𝑇 =𝑇wall + (1 − ( ) )
⎩ 2𝑘 𝛿
is stable at all Reynolds numbers, making it unsuitable as an initial condition.

A standard approach is to perturb the laminar solution vector with a random
number generator. For the present discretization, such a perturbation is quickly
damped by the viscous stress, because many degrees of freedom correspond to
highorder basis functions. This phenomenon of laminarization after a highfrequency
perturbation has been well known in the CFD community for a long time (e.g., [33]),
and has more recently also been established experimentally [34].
The laminar velocity field is therefore perturbed by both a random solution vec
tor, and the arbitrarily chosen largescale solenoidal structure
3 𝑦 2 sin(2𝜋𝑧) sin(𝑦/𝛿)
0.10 𝑢bulk (1 − ( ) ) [ 0 ] , (5.41)
2 𝛿 sin(2𝜋𝑥)
which induces a turbulent solution.
5.4.3. Domain Size and Mesh

The domain should be large enough to encompass the largest physical structures
of the flow between two infinite planes. This can be checked a posteriori by inves
tigating the correlation between the turbulent fluctuations at various points in the
3 Thereare many other commonly used correlations, but they have no independent value, as they are
designed to approximate the ColebrookWhite correlation. See, e.g., the discussion in Zigrang and
Sylvester [32].
5.4. Test Case: Infinite Plane Channel Flow 95
domain. For any two points that are half a domain size apart in the 𝑥 or 𝑧direction,
the correlation should be negligible.
Our domain size is based in part on the correlation coefficients that are reported
in [35] and [36]. Piomelli et al. [37] investigated various domain sizes more sys
tematically, and concluded that (𝐿𝑥 , 𝐿𝑦 , 𝐿𝑧 ) = (6𝛿, 2𝛿, 3𝛿) is ‘marginally sufficient’
for secondorder turbulent statistics at our Reynolds number, so this is what we
use.
Our setup is meant to reproduce one of the test cases in Patel et al. [30],
who performed direct numerical simulation (DNS) at Re𝜏 = 395. This is a popular
Reynolds number, presumably because it was used in the landmark paper by Moser
et al. [36]. Following Patel et al. , the fluid properties are kept constant, and Pr = 1.
The element sizes in the inhomogeneous 𝑦direction are based on a tanh grid
𝑁
spacing. More precisely, the element boundaries are located at 𝑦/𝛿 ∈ {𝜉𝑖 }𝑖=0 with
2𝑖
tanh (𝛾 ( − 1))
𝑁
𝜉𝑖 = . (5.42)
tanh 𝛾
Here 𝛾 > 0 is a stretching parameter, with 𝛾 → 0 corresponding to a uniform element

5
size. The parameters 𝛾 and 𝑁 are fixed implicitly by choosing the minimum and
maximum element widths (Δ𝑦 min and Δ𝑦 max ) at the wall and in the center of the
channel.4
We place the first element boundary at 𝑦 + = 2. Other authors have suggested
putting the first grid point at approximately 𝑦 + = 1 or 𝑦 + = 2 when using a high
order finitedifference scheme, which normally cannot attain its full order of accu
racy near the wall. It is unclear whether a discontinuous Galerkin method permits
a larger wall element width, though de Wiart et al. [38] have obtained accurate re
sults with wall elements at 𝑦 + = 2.5, using a thirdorder tensorproduct polynomial
space. We set 𝑁 = 46 and 𝛾 = 2.2132, so that the first element boundary is at
𝑦 + = 2.00, and the maximum element width is 38.8 wall units.
The element widths in the homogeneous 𝑥 and 𝑧directions are constant through
out the domain. This is by far the most common in the literature, though Collis
[39] has argued that the geometric flexibility of the discontinuous Galerkin method
should be used to refine the elements at the wall in the 𝑥 and 𝑧directions, so
that their aspect ratios do not become too large. Previous authors have used many
4 Thisentails solving the nonlinear coupled equations 𝜉1 − 𝜉0 = Δ𝑦min and 𝜉𝑚 − 𝜉𝑚−1 = Δ𝑦max with
𝑚 ≔ ⌈𝑁/2⌉ for (𝛾, 𝑁). By assuming 𝑁 ≫ 1 (which is normally the case), the index 𝑖 can be treated as
a continuous variable, giving the estimates
𝜕𝜉𝑖 2𝛾/𝑁 1 𝜕𝜉𝑖 2𝛾/𝑁
Δ𝑦min ≈ | = and Δ𝑦max ≈ | = , (5.43)
𝜕𝑖 𝑖=0 tanh 𝛾 cosh2 𝛾 𝜕𝑖 𝑖=𝑁/2 tanh 𝛾
resulting in a ratio of the element widths of

𝑦max 2 1
= cosh 𝛾 + 𝒪 ( ) . (5.44)
𝑦min 𝑁
We found that 𝛾 ≈ cosh−1 √𝑦max /𝑦min is very close to the actual solution for all (𝑦min , 𝑦max ), and
that a fixedpoint iteration always converges with this initial guess.
different aspect ratios for the minimum elements in the bulk; we use Δ𝑧/Δ𝑦 = 1
and Δ𝑥/Δ𝑧 = 2, resulting in element widths in the 𝑥 and 𝑧directions of 77.6 and
38.8, as measured in wall units.
5.4.4. Results
We are interested in the ensemble averages of various quantities over many re
alizations of the flow, which are called the Reynolds averages, denoted by angle
brackets (e.g., ⟨𝑇⟩, ⟨𝑚1 ⟩, …). The turbulent fluctuation is defined as the instanta
neous deviation from the Reynolds average, that is,
𝜙′ ≔ 𝜙 − ⟨𝜙⟩ . (5.45)
In practice the ensemble average is assumed equal to the time average, which can
be estimated from a single computation or experiment that runs for a sufficiently
long time.
Besides the firstorder statistics ⟨𝑢𝑖 ⟩ and ⟨𝑇⟩, we are also interested in the
Reynolds stress
5
⟨𝑢𝑖′ 𝑢𝑗′ ⟩ = ⟨𝑢𝑖 𝑢𝑗 ⟩ − ⟨𝑢𝑖 ⟩ ⟨𝑢𝑗 ⟩ , (5.46)
and the turbulent heat flux
⟨𝑢𝑖′ 𝑇′ ⟩ = ⟨𝑢𝑖 𝑇⟩ − ⟨𝑢𝑖 ⟩ ⟨𝑇⟩ . (5.47)
These secondorder statistics are commutation errors between the Reynolds aver
age and multiplication, so they come up when Reynoldsaveraging the convective
terms in the incompressible NavierStokes equation.
The temperature can be made dimensionless in various ways; we use the ref
erence value
2
−1 𝛿 𝑄
𝑇ref ≔ Re𝜏 . (5.48)
𝑘
The group 𝛿2 𝑄/𝑘 contains diffusionrelated quantities, and can also be seen in the
−1
laminar solution (Eq. 5.40). The factor Re𝜏 takes convective heat transfer into
ref
account, and is meant to let ⟨𝑇/𝑇 ⟩ depend less on the Reynolds number, since
⟨𝑇⟩ decreases with increasing Re𝜏 .
For our test case the quantities can also be averaged over the homogeneous
directions, making the averages converge far more quickly. The quantities of in
terest are sampled every 20 time steps. They are projected onto the eighthorder
polynomial space in the 𝑦 coordinate in each element, so that no information is lost.
The averages are based on approximately 10 turnover times.
We also average the two values in the lower and upper halves of the domain
(i.e., where 𝑦 < 0 or 𝑦 > 0). This does not work for the stresses involving 𝑢2 (i.e.,
⟨𝑢1′ 𝑢2′ ⟩ and ⟨𝑇′ 𝑢2′ ⟩), because they are antisymmetric about 𝑦 = 0. Therefore we will
instead report values based on
𝑢2↑ ≔ 𝑢2 sign(𝑦) = ±𝑢2 for ±𝑦 > 0, (5.49)

5.5. Discussion 97
which is the component of the velocity toward the closest wall. This yields sym
metric stresses ⟨𝑢1′ 𝑢2↑
′ ′
⟩ and ⟨𝑇′ 𝑢2↑ ⟩. These could also be interpreted as the usual
′ ′ ′ ′
stresses ⟨𝑢1 𝑢2 ⟩ and ⟨𝑇 𝑢2 ⟩ in the upper half of the domain.
Figure 5.2 shows the error in the average velocity profiles for a modelfree high
order simulation (uDNS32 ), a coarser modelfree lowerorder simulation (uDNS22 ),
and two LES simulations based on the coarser discretization (WALE and Smag.). The
time step size is given by (𝑢𝜏 /𝛿)𝛿𝑡 = 10−3 , resulting in a maximum CFL number of
approximately 0.7 in the flow direction on the coarse grid. The error is based on
the DNS data in Moser et al. [36]. The figure also shows that the difference with
the DNS data in Patel et al. [30] is negligible.
For comparison, the results of the LES by Singh, You, and Bose [40] are also
shown (indicated by SYB). They varied the spatial discretization independently from
the LES filter width, resulting in a gridindependent (i.e., fully resolved) LES. They
used a dynamic Smagorinsky model with LES filter widths in the 𝑥, 𝑦, and 𝑧
directions of (78, 0.53–60, 39) (measured in wall units). Ours are comparable:
approximately (72, 1.9–36, 36). We find that their results are approximately as ac
curate as our coarsegrid calculations, regardless of whether we use an LES model,
and despite the fact that we do have a discretization error. 5
Figure 5.3 shows the averaged subfilter effective viscosities for the three LES
models that were discussed in section 5.2. These were obtained from a calculation
sfs
without an LES model. (Using, e.g., a WALE model might dampen 𝜈WALE more than
the other subfilter scale viscosities, which would have distorted the figure.) The
subfilter terms display clear discontinuities, showing that the eddy viscosity is not
just a physical model, but that it is closely linked to the discretization.
One of the most important quantities for engineering purposes is the bulk ve
locity, given by Eq. 5.35. Table 5.1 lists the results. For comparison, the bulk
velocity can also be determined from the Darcy friction factor. From Eq. 5.38 and
the definitions of the Reynolds numbers (Eqs. 5.30 and 5.34),
𝑢bulk Rebulk /4 8
= =√ . (5.50)
𝑢𝜏 Re𝜏 𝑓D
The bulk velocity is overestimated in all calculations except the DNS.

For further validation, figures 5.4 and 5.5 compares the averaged stresses for
the modelfree highorder simulation (uDNS32 ) to the DNS data from [36] and [30],
which are practically exact solutions. Our firstorder statistic (⟨𝑢1 ⟩) is more accu
rate than the LES data (SYB), whereas the secondorder statistics (the Reynolds
stresses) are less accurate. Possible explanations include their dynamic LES model,
and our small domain. Overall the results are satisfactory.
5.5. Discussion
This chapter has presented a large eddy simulation of turbulent plane channel flow.
The results are comparable to previous literature.
Interestingly, the nomodel approach (i.e., ‘unresolved DNS’, denoted by uDNS)
yields more accurate results than the WALE model. This could be because the locally
PBP
0.25
hu1 i − hu1 iMKM / hu1 iMKM SYB
uDNS32
0.20
uDNS22
0.15
WALE

Smag.
0.10
0.05
0.00
−0.05
100 101 102
5 y+
Figure 5.2: Relative deviation of ⟨𝑢1 ⟩ from the DNS reference data ⟨𝑢1 ⟩MKM in Moser, Kim, and Mansour
[36]. See Table 5.6 for the meaning of the labels.
0.8
0.6
ν molec
0.4
0.2
D E
sfs
νsmag sfs
νWALE sfs
νQR
0.0
2.0 7.3 15 26 42 64 93 133 184 247 318
4.4 11 20 33 52 77 112 157 214 281 356
y+
sfs in Eq. 5.14,
Figure 5.3: Average values of the suffilter scale (SFS) viscosity for various LES models (𝜈smag
sfs sfs
𝜈WALE in Eq. 5.19, and 𝜈QR in Eq. 5.21), obtained from the uDNS32 calculation (see Table 5.6).
Scaled such that each element looks the same size; the vertical lines indicate the element boundaries.
5.5. Discussion 99
hu1 i /uτ
20
15
10
SYB
5 uDNS32
PBP
0
0 50 100 150 200 250 300 350
p y+ p
hu02
1 i/uτ hu02
2 i/uτ
1.0
SYB
2.5
uDNS32 5
0.8
PBP
2.0
0.6
1.5
0.4
1.0
SYB
0.5 0.2 uDNS32
PBP
0.0 0.0
p q
hu02
3 i/uτ u01 u02↑ /uτ
1.2 uDNS32
0.8 MKM
1.0
0.6
0.8
0.6 0.4
0.4 SYB
uDNS32 0.2
0.2
PBP
0.0 0.0
0 100 200 300 0 100 200 300
y+ y+
Figure 5.4: Reynolds average ⟨𝑢1 ⟩ and Reynolds stress components (Eq. 5.46), compared to other LES
and DNS results at Re𝜏 = 395. See Table 5.6 for the labels.
p
(hT i − Twall ) /T ref hT 02 i/T ref
0.05 uDNS32
0.006 PBP
0.04
0.03
0.004
0.02
0.002
0.01 uDNS32
PBP
0.00 0.000
hT 0 u01 i /(T ref uτ ) T 0 u02↑ /(T ref uτ )
uDNS32 0.0020 uDNS32

0.015
PBP PBP
5 0.0015
0.010
0.0010
0.005 0.0005
0.000 0.0000
0 100 200 300 0 100 200 300
y+ y+
Figure 5.5: Reynolds average ⟨𝑇⟩ and turbulent heat flux (Eq. 5.47), at Re𝜏 = 395, compared to DNS
results. See Table 5.6 for the labels.
𝑇ref is given by Eq. 5.48. Note that √⟨𝑇′2 ⟩/𝑇ref equals the rootmeansquare value of (⟨𝑇⟩ − 𝑇wall ) /𝑇ref .
Figure 5.6: Abbreviations used in Figs. 5.4–5.3 and Table 5.1.
PBP DNS reference data from Patel, Boersma, and Pecnik [30], accessed
through [41]
MKM DNS reference data from Moser, Kim, and Mansour [36], accessed
through [42]
SYB LES reference data from Singh, You, and Bose [40].
Fully resolved; only contains LES model error
uDNS22 present simulation without an LES model (i.e., unresolved DNS) with
𝒫𝑚 = 𝒫𝑝 = 𝒫ℎ = 2.
WALE same as uDNS22 , but with an WALE LES model
Smag. same as uDNS22 , but with a Smagorinsky LES model
uDNS32 same as uDNS22 , but with a higher order of approximation for 𝑚:
𝒫𝑚 = 3 and 𝒫𝑝 = 𝒫ℎ = 2
References 101
Table 5.1: Bulk velocity implied by various simulations. See Table 5.6 for the labels. The ColebrookWhite
estimate is based on Eqs. 5.39 and 5.50.
𝑢bulk /𝑢𝜏 deviation from MKM

MKM 17.6
PBP 17.5 0.10%
SYB 20.0 14%
uDNS22 19.9 13%
uDNS32 18.5 5.1%
WALE 21.0 20%
Smag. 20.7 18%
ColebrookWhite 18.4 4.7%
conservative spatial discretization acts as an implicit LES (‘iLES’) model, analogous

to iLES models for finite volume discretizations (e.g., [43]). Recent research indeed
indicates that a DG discretization with a highorder polynomial space is an lowpass
filter (e.g., [44–46]). Furthermore, note that ⟨𝑢1′2 ⟩, ⟨𝑢2′2 ⟩, and ⟨𝑢3′2 ⟩ are not equal in
5
the bulk of the fluid, at 𝑦 = 0, that is, 𝑦 + = 𝑅𝑒𝜏 = 395, as can be seen in Fig. 5.4.
The anisotropic Reynolds stress means that the effect of the walls is present in the
whole fluid. It is therefore not surprising that a fixedcoefficient (i.e., nondynamic)
isotropic LES model does not model the dissipation well.
We also found that lowering the time step size degrades the solution quality. This
could be explained by implicit temporal filtering due to the temporal discretization
error, which might damp spurious oscillations. This can be compared to the findings
by Meyers and Sagaut [47], who observed that coarser grids can sometimes result
in more accurate predictions of the mean flow.
Meyers and Sagaut [47] also emphasize that, while a turbulent channel flow
is easy to handle with some discretizations, it is not an easy LES test case. This
is because the turbulent flow structures are generated at the walls, whereas LES
models have traditionally been better at capturing the physics in isotropic flow in
the bulk.
We stress that the discontinuous Galerkin method is fundamentally an unstruc
tured discretization. It is ‘blind’ to any geometrical symmetries, and does not change
its approximation scheme depending on the direction, as is common for finite dif
ference in channel geometries. Infinite plane channel flow is therefore a relatively
difficult test case with our discretization.
References
[1] B. Vreman, B. Geurts, and H. Kuerten, Subgridmodelling in LES of compress
ible flow, Applied Scientific Research 54, 191 (1995).
[2] B. Lessani, J. Ramboer, and C. Lacor, Efficient Largeeddy Simulations of Low

Mach Number Flows Using Preconditioning and Multigrid, International Journal
of Computational Fluid Dynamics 18, 221 (2004).
102 References
[3] E. Garnier, N. Adams, and P. Sagaut, Large Eddy Simulation for Compressible
Flows, 1st ed., Scientific Computation (Springer, Dordrecht, 2009).
[4] P. Sagaut, Large Eddy Simulation for Incompressible Flows: An Introduction,

3rd ed. (SpringerVerlag Berlin Heidelberg, 2006) p. 558.
[5] S. B. Pope, Turbulent Flows, 1st ed. (Cambridge University Press, Cambridge,
UK, 2000) p. 771.
[6] L. C. Berselli, T. Iliescu, and W. J. Layton, Mathematics of Large Eddy Simu

lation of Turbulent Flows, 1st ed. (Springer, Berlin, Heidelberg, 2006).
[7] U. Piomelli, P. Moin, and J. H. Ferziger, Model consistency in large eddy sim
ulation of turbulent channel flows, Physics of Fluids 31, 1884 (1984).
[8] M. Inagaki, A new walldamping function for large eddy simulation employing
kolmogorov velocity scale, International Journal of Heat and Fluid Flow 32,
26 (2011).
5
[9] F. Nicoud and F. Ducros, SubgridScale Stress Modelling Based on the Square
of the Velocity Gradient Tensor, Flow, Turbulence and Combustion 62, 183
(1999).
[10] M. de la Llave Plata, V. Couaillier, and M.C. le Pape, On the use of a highorder
discontinuous Galerkin method for DNS and LES of wallbounded turbulence,
[11] ANSYSFluent, ANSYS FLUENT 12.0/12.1 Documentation, [Online] Accessed

on 20190109 (2009).
[12] R. Verstappen, When Does Eddy Viscosity Damp Subfilter Scales Sufficiently?
Journal of Scientific Computing 49, 94 (2011).
[13] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot, A dynamic subgridscale

eddy viscosity model, Physics of Fluids A: Fluid Dynamics 3, 1760 (1991).
[14] A. W. Vreman, Direct and LargeEddy Simulation of the Compressible Turbulent

Mixing Layer, Phd thesis, University of Twente (1995).
[15] A. Abbà, L. Bonaventura, M. Nini, and M. Restelli, Dynamic models for Large
Eddy Simulation of compressible flows with a high order DG method, Comput
ers & Fluids 122, 209 (2015), arXiv:1407.6591 .
[16] M. Lesieur, Turbulence in Fluids, 4th ed. (Springer, Dordrecht, 2008) p. 563.

634 (2017).
References 103
[18] H. Nemati, Direct numerical simulation of turbulent heat transfer to fluids at

supercritical pressures, doctoral thesis, TU Delft (2016).
[19] C. Yang, J. Xu, X. Wang, and W. Zhang, Mixed convective flow and heat
transfer of supercritical CO2 in circular tubes at various inclination angles,
International Journal of Heat and Mass Transfer 64, 212 (2013).
[20] X. Chu and E. Laurien, Flow stratification of supercritical CO2 in a heated
horizontal pipe, The Journal of Supercritical Fluids 116, 172 (2016).
[21] F. Föll, S. Pandey, X. Chu, C.D. Munz, E. Laurien, and B. Weigand, High
Fidelity Direct Numerical Simulation of Supercritical Channel Flow Using Dis
continuous Galerkin Spectral Element Method, in High Performance Comput
ing in Science and Engineering ’ 18 (Springer International Publishing, Cham,
2019) pp. 275–289.
[22] R. G. Rehm and H. R. Baum, Equations of Motion for Thermally Driven, Buoy
ant Flows, Journal of Research of the National Bureau of Standards 83, 297
(1978). 5
[23] F. C. Nicoud, Numerical study of a channel flow with variable properties, in
Annual Research Briefs (Center for Turbulence Research, 1998) pp. 289–310.
[24] R. Knikker, A comparative study of highorder variableproperty segregated
algorithms for unsteady low Mach number flows, International Journal for Nu
merical Methods in Fluids 66, 403 (2011).
[25] M. Avila, R. Codina, and J. Principe, Large eddy simulation of low Mach number
flows using dynamic and orthogonal subgrid scales, Computers & Fluids 99,
44 (2014).
[26] G. Accary, P. Bontoux, and B. Zappoli, Turbulent RayleighBénard convection
in a nearcritical fluid by threedimensional direct numerical simulation, Journal
of Fluid Mechanics 619, 127 (2009).
[27] G. Accary and I. Raspo, A 3D finite volume method for the prediction of a
supercritical fluid buoyant flow in a differentially heated cavity, Computers &
Fluids 35, 1316 (2006).
[28] W. Wang and S. He, Direct numerical simulation of fluid flow at supercritical
pressure in a vertical channel, in International Topical Meeting on Nuclear
Reactor Thermal Hydraulics 2015, NURETH 2015, Vol. 3 (Chicago, USA, 2015)
pp. 2334–2347.
[29] J. W. R. Peeters, R. Pecnik, M. Rohde, T. H. J. J. van der Hagen, and B. J.
Boersma, Turbulence attenuation in simultaneously heated and cooled annular
flows at supercritical pressure, Journal of Fluid Mechanics 799, 505 (2016).
[30] A. Patel, B. J. Boersma, and R. Pecnik, Scalar statistics in variable property
turbulent channel flows, Physical Review Fluids 2, 084604 (2017).
104 References
[31] R. B. Bird, W. E. Stewart, and E. N. Lightfoot, Transport Phenomena, 2nd ed.,

edited by P. Kulek (John Wiley & Sons, Inc., New York, NY, 2002) p. 914.
[32] D. J. Zigrang and N. D. Sylvester, Explicit approximations to the solution of

Colebrook’s friction factor equation, AIChE Journal 28, 514 (1982).
[33] J. Eggels, Direct and Large Eddy Simulation of Turbulent Flow in a Cylindrical
Pipe Geometry, doctoral thesis, Delft University of Technology (1994).
[34] J. Kühnen, B. Song, D. Scarselli, N. B. Budanur, M. Riedl, A. P. Willis, M. Avila,
and B. Hof, Destabilizing turbulence in pipe flow, Nature Physics 14, 386
(2018).
[35] H. Abe, H. Kawamura, and Y. Matsuo, Surface heatflux fluctuations in a

turbulent channel flow up to Re𝜏 =1020 with Pr=0.025 and 0.71, International
Journal of Heat and Fluid Flow 25, 404 (2004).
[36] R. D. Moser, J. Kim, and N. N. Mansour, Direct numerical simulation of tur

5 bulent channel flow up to Re𝜏=590, Physics of Fluids 11, 943 (1999).
[37] U. Piomelli, C. Benocci, and J. P. A. J. Van Beek, Large Eddy Simulation:
Theory and Simulation, (2016), lectures notes.
[38] C. C. de Wiart, K. Hillewaert, L. Bricteux, and G. Winckelmans, Implicit
LES of free and wallbounded turbulent flows based on the discontinuous
Galerkin/symmetric interior penalty method, International Journal for Numer
ical Methods in Fluids 78, 335 (2015).
[39] S. S. Collis, The DG/VMS Method for Unified Turbulence Simulation, in 32nd
AIAA Fluid Dynamics Conference and Exhibit (Houston, Texas, 2002).
[40] S. Singh, D. You, and S. T. Bose, Largeeddy simulation of turbulent channel
flow using explicit filtering and dynamic mixed models, Physics of Fluids 24,
85105 (2012).
[41] R. Pecnik, DNS data from variable property channel flows, Github. [Online]
Accessed on 20191009 (2019).
[42] R. D. Moser, J. Kim, and N. N. Mansour, DNS Data for Turbulent Channel Flow,
University of Texas. [Online] Accessed on 20210502 (2007).
[43] S. Hickel, A. Devesa, and N. A. Adams, Implicit Turbulence Modeling by Fi

nite Volume Methods, in Numerical Simulation of Turbulent Flows and Noise
Generation: Results of the DFG/CNRS Research Groups FOR 507 and FOR
508, edited by C. Brun, D. Juvé, M. Manhart, and C.D. Munz (Springer Berlin
Heidelberg, Berlin, Heidelberg, 2009) pp. 149–173.
[44] E. Ferrer, An interior penalty stabilised incompressible discontinuous Galerkin

Fourier solver for implicit large eddy simulations, Journal of Computational
Physics 348, 754 (2017).
References 105
[45] R. C. Moura, G. Mengaldo, J. Peiró, and S. J. Sherwin, On the eddyresolving

capability of highorder discontinuous Galerkin approaches to implicit LES /
underresolved DNS of Euler turbulence, Journal of Computational Physics
330 (2017), 10.1016/j.jcp.2016.10.056.
[46] A. R. Winters, R. C. Moura, G. Mengaldo, G. J. Gassner, S. Walch, J. Peiro, and

S. J. Sherwin, A comparative study on polynomial dealiasing and split form
discontinuous Galerkin schemes for underresolved turbulence computations,
Journal of Computational Physics 372, 1 (2018), arXiv:1711.10180 .
[47] J. Meyers and P. Sagaut, Is planechannel flow a friendly case for the testing
of largeeddy simulation subgridscale models? Physics of Fluids 19, 48105
(2007).
5
6
Conclusion
The previous chapters have treated various discretizations in computational fluid
dynamics without an overarching theme. The conclusions and their relevance are
therefore best described by considering the chapters separately.
Chapter 2 has explained our discontinuous Galerkin method, which differs from
most other literature on two minor points. First, the penalty parameters (for the
interior penalty and LaxFriedrichs numerical fluxes) are evaluated in a pointwise
manner, rather than being averaged over the face, or a neighboring element. Sec
ond, the convection at Dirichlet boundaries is treated correctly, which is not the
case in most of the listed references. In practice these points might not make
much difference to the numerical outcomes.
The pressure correction method in chapter 3 is almost completely standard.
The one peculiarity is that the projection step (Eq. 3.10) contains a penalty ma
trix, which is a consequence of using an equalorder finite element discretization.
As explained in section 3.5.1, this incurs a substantial computational cost at high
Reynolds numbers and small time steps, though this problem could be particular to
our implementation and solver.
Chapter 4 treats the variable density in great detail. We pay particular attention
to the most convenient form of the enthalpy equation (primitive or conservative) and
whether to solve for the specific enthalpy or the volumetric enthalpy. Each approach
has its own challenges; it is not completely clear which is best. We decided to solve
for the specific enthalpy from the conservative form of the transport equation.
The biggest potential problem with this choice is that it complicates the temporal
discretization of the enthalpy, which can destabilize the calculation, or degrade the
order of convergence. It was shown that these problems can be addressed by a
simple modification of the finite difference scheme (‘method #2’), and by ‘offsetting’
the specific enthalpy with a constant ℎ0 . The choice of this enthalpy offset was
central to stabilizing the coupled transport equations with a temperaturedependent
density.
107
108 6. Conclusion
The value of ℎ0 would also have been important if we had solved for the vol
umetric enthalpy (as explained in section 4.1.2), because it determines whether
there is a onetoone relation between the volumetric enthalpy and the density. It
is therefore surprising that we have not seen this mentioned previous literature.
Hopefully the discussions in chapter 4 are useful to the development of other nu
merical methods for lowMach number flows.
Chapter 5 features turbulent flow simulations between two infinite planes. This
geometry was chosen for the abundance of highquality reference data and its pe
riodic boundary conditions, which reduce the computational cost, because there
are fewer walls that require grid refinement. The discontinuous Galerkin method
and our solver are fundamentally unstructured, and do not exploit the simplicity
of the geometry. The test case is therefore no less challenging than if the mesh
were ostensibly more complicated. Our LES models have also been used by others
with similar discretizations, so our results add little to the existing body of litera
ture, though they can be considered further validation of the spatial and temporal
discretizations in chapters 2–3 and their implementation.
6
Acknowledgements
Sections 2.1.2–2.4, 2.7, 3.1–3.2, and 4.1–4.4 are largely based on work that I did
at TU Delft, where most of the code DGFlows was also written. I am grateful to
Danny Lathouwers for having meant to help with the scientific aspect of my work,
and I am happy that he has always kept his sense of humor. Danny suggested
method #2 as an improvement over method #1 in section 4.3. I would like to
thank Marco Tiberga for having been a pleasant colleague, and for his help with
the code development. Among other things, Marco wrote the initial implementation
of the penalty terms in Eq. 5.26.
My employment as a PhD candidate at TU Delft was funded by the sCO2 HeRo
project that has received funding from the European research and training program
2014–2018 (grant agreement ID 662116).
I am grateful to prof. dr. Laurien Eckart and prof. dr. Jörg Starflinger for having
given me the opportunity to work at the Institut für Kernenergetik und Energiesys
teme (IKE) at the University of Stuttgart, without which this thesis would not have
been realized. I was graciously given access to the HazelHen supercomputer at the
HighPerformance Computing Center Stuttgart (HLRS), which greatly sped up the
development and implementation of the large eddy simulation models. Sandeep
Pandey contributed significantly with practical tips on the turbulent channel flow
simulations.
My stay at IKE in Stuttgart was partially supported by the ENEN+ project that
has received funding from the Euratom research and training Work Programme
2016–2017 — 1 #755576.
I would like to thank Rene Pecnik for the discussion on the discretization of
the enthalpy equation with a variable density. Rene also provided access to the
Reynolds computer cluster at the TU Delft, which was used for some of the calcu
lations in section 5.4.
Ten slotte ben ik erg gesteund door het geduld en meeleven van Matteo. Ik ben
je onbeschrijvelijk dankbaar voor het vertrouwen, het lachen, de herinneringen en
de liefde die je me de afgelopen jaren gegeven hebt en zie ernaar uit om elkaars
levens daar nog lang mee te vullen.
109
Curriculum Vitæ
Aldo Hennink
1990 Born in Amsterdam, The Netherlands.
Education
2012 Msc. in Physics
Technical University of Delft
2015 Msc. in Physics

2021 PhD. in Physics

Thesis: LowMach Number Flow and the Discontinuous
Galerkin Method
Promotors: dr. D. Lathouwers
Prof. dr. J. L. Kloosterman
111
List of Publications
3. A. Hennink, M. Tiberga, D. Lathouwers, A Pressurebased solver for lowMach num
ber flow using a discontinuous Galerkin method, Journal of Computational Physics,
425, 109877 (2021).
2. M. Tiberga, A. Hennink, J. L. Kloosterman, D. Lathouwers, A highorder discontinuous

Galerkin solver for the incompressible RANS equations coupled to the 𝑘𝜖 turbulence
model, Computers & Fluids, 212, 104710 (2020).
1. A. Hennink, D. Lathouwers, A discontinuous Galerkin method for the monoenergetic

FokkerPlanck equation based on a spherical interior penalty formulation, Journal of
Computational and Applied Mathematics 330, 253267 (2018).
113

PHD Aldo Hennink

Uploaded by

Copyright:

Available Formats

PHD Aldo Hennink

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PHD Aldo Hennink

Uploaded by

Copyright:

Available Formats

Low­Mach Number Flow and the

Discontinuous Galerkin Method

for the purpose of obtaining the degree of doctor

Master of Science in Applied Physics, Delft University of Technology, The Netherlands

Composition of the doctoral committee:

Keywords: Low­Mach Number Flow, Pressure Correction, Discontinuous

Copyright © 2022 by A. Hennink

1.1. Mathematical Problem Setting

The heat flux is

• outflow boundaries, denoted by 𝜕ΩN , where

1.2. Overview of the Numerical Method

𝑨𝑖𝑗 = 𝑎(𝜉𝑖 , 𝜉𝑗 ) and 𝒃𝑖 = 𝑏(𝜉𝑖 ) . (1.6)

Find 𝜙 ∈ 𝑉𝜙 , such that, for all 𝑣 ∈ 𝑉𝜙 ,

1.2.1. Why Use a Discontinuous Galerkin Method?

targeting’. This makes it difficult to compare the efficacy of DG methods to the

1.3. Thesis Overview

[10] Y. Bazilevs and T. Hughes, Weak imposition of Dirichlet boundary conditions

[16] M. Drosson, Development of a high­order interior penalty discontinuous

[17] J.­B. Chapelier, M. De La Llave Plata, F. Renac, and E. Lamballais, DNS of

[18] H. L. Atkins and B. T. Helenbrook, Numerical Evaluation of P­Multigrid Method

[19] B. Helenbrook and H. Atkins, Application of P­Multigrid to Discontinuous

[21] K. Hillewaert, Development of the discontinuous Galerkin method for high­

[22] K. A. Cliffe, E. J. C. Hall, and P. Houston, Adaptive Discontinuous Galerkin

[25] S. Rhebergen, B. Cockburn, and J. J. W. van der Vegt, A space­time dis­

[28] R. Moura, S. Sherwin, and J. Peiró, Linear dispersion­diffusion analysis and

2.1.1. Overview of DG discretizations

2.1.2. Construction of the Solution Space

direction. On boundary faces, 𝐧𝐹 coincides with the outward normal of Ω.

where 𝐧𝑇 is the outward normal of element 𝑇.

Modal vs. Nodal Basis Functions

2.2. Discrete Continuity Equation

𝑎div (𝐯, 𝑞) = − ∑ ∫ 𝑞∇ ⋅ 𝐯 + ∑ ∫ {𝑞} J𝐯K ⋅ 𝐧𝐹 , (2.5)

so that Eq. 2.6 becomes

Find 𝑝 ∈ 𝑉𝑝 , such that, for all 𝑞 ∈ 𝑉𝑝 ,

for equal­order discretizations, and 𝑎stab (𝑝, 𝑞) = 0 for mixed­order discretizations.

2.3. Discrete Momentum Equation

𝑎div (𝐯, 𝑞) = ∑ ∫ 𝐯 ⋅ ∇𝑞 − ∑ ∫ J𝑞K {𝐯} ⋅ 𝐧𝐹 , (2.11)

2.3.1. Discretization of the Viscous Stress

is a linear operator, with 𝐵𝑖𝑗 = 𝜌∇𝑖 𝑢𝑗 = ∇𝑖 𝑚𝑗 − 𝑑𝑖 𝑚𝑗 , and

𝑎visc (𝐰, 𝐯) = ∑ ∫ 𝐿visc

and the linear operator

𝑙 visc (𝐯) = ∑ ∫ (𝜂𝐹 𝐦D ⋅ 𝐯 − 𝐦D ⋅ 𝐿visc (𝐯) ⋅ 𝐧𝐹 ) + ∫ 𝐟N ⋅ 𝐯 . (2.15)

2.4. Discrete Enthalpy Equation

2.5. Discretization of the Convection

𝑙 conv (𝐛, 𝜙D , 𝑣) = − ∫ 𝑣 min (0, 𝐛D ⋅ 𝐧) 𝜙D . (2.20)

The bilinear term has the general form

𝑎conv (𝐛, 𝑤, 𝑣) = − ∑ ∫ 𝑤 𝐛 ⋅ ∇𝑣 + ∑ ∫ J𝑣K 𝐻𝐹 (𝐛, 𝑤)

2.5.1. Choice of the Numerical Flux

0= ∫ 𝑤 (𝐛̃ ⋅ 𝐧) + ∫ 𝑤up (𝐛D ⋅ 𝐧)

where 𝐧𝑇 is the outward normal of element 𝑇, 𝑤+ is the value on 𝑇, 𝑤− is the value

2.5.2. Solution Spaces for the Enthalpy and the Pressure

−𝑎stab (𝑝, 𝑞)+ ∑ ∫ 𝐦⋅∇𝑞 = ∑ ∫ J𝑞K {𝐦}⋅𝐧𝐹 +∫ 𝑞 𝐦⋅𝐧+∫ 𝑞 𝐦D ⋅𝐧 , (2.26)

∑ ∫ 𝜙̃ 𝐛 ⋅ ∇𝑣 = ∑ ∫ 𝜙̃ J𝑣K {𝐛} ⋅ 𝐧𝐹 + ∫ 𝜙̃ 𝑣 𝐛 ⋅ 𝐧 + ∫ 𝜙D 𝑣 𝐛D ⋅ 𝐧 . (2.27)

2.5.3. Proper Treatment of Dirichlet Boundary Conditions

0=∫ 𝑤 𝐛⋅𝐧+∫ 𝑤up 𝐛D ⋅ 𝐧 (2.28)

∑ ∫ 𝜙̃ 𝐛 ⋅ ∇𝑣 − ∑ ∫ 𝜙̃ J𝑣K {𝐛} ⋅ 𝐧𝐹 = ∫ 𝜙D 𝑣 (𝐛D − 𝐛) ⋅ 𝐧 . (2.29)

∫ 𝑣 (𝑤 max (0, 𝐛 ⋅ 𝐧) + 𝜙𝐷 min (0, 𝐛 ⋅ 𝐧)) . (2.30)

2.7. Test Case: A Heated Backward­facing Step

2.8. Discussion and Conclusion

LowMach Number Flow and the

Keywords: LowMach Number Flow, Pressure Correction, Discontinuous

[16] M. Drosson, Development of a highorder interior penalty discontinuous

[17] J.B. Chapelier, M. De La Llave Plata, F. Renac, and E. Lamballais, DNS of

[18] H. L. Atkins and B. T. Helenbrook, Numerical Evaluation of PMultigrid Method

[19] B. Helenbrook and H. Atkins, Application of PMultigrid to Discontinuous

[21] K. Hillewaert, Development of the discontinuous Galerkin method for high

[25] S. Rhebergen, B. Cockburn, and J. J. W. van der Vegt, A spacetime dis

[28] R. Moura, S. Sherwin, and J. Peiró, Linear dispersiondiffusion analysis and

for equalorder discretizations, and 𝑎stab (𝑝, 𝑞) = 0 for mixedorder discretizations.

2.7. Test Case: A Heated Backwardfacing Step

[2] B. Cockburn and C.W. Shu, RungeKutta Discontinuous Galerkin Methods

[4] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini, Unified analysis of dis

[6] Y. Epshteyn and B. Rivière, Estimation of penalty parameters for symmet

[7] K. Hillewaert, Development of the discontinuous Galerkin method for high

[12] K. Shahbazi, P. F. Fischer, and C. R. Ethier, A highorder discontinuous Galerkin

[16] B. Cockburn, G. Kanschat, and D. Schötzau, An EqualOrder DG Method for

[17] L. Botti and D. A. Di Pietro, A pressurecorrection scheme for convection

[19] B. Krank, N. Fehn, W. A. Wall, and M. Kronbichler, A highorder semiexplicit

[34] M. Tavelli and M. Dumbser, A pressurebased semiimplicit spacetime discon

2 [35] M. Tiberga, A. Hennink, J. L. Kloosterman, and D. Lathouwers, A highorder

[38] F. Bassi, A. Crivellini, D. A. D. Pietro, and S. Rebay, An artificial compressibil

[40] N. Emamy, F. Kummer, M. Mrosek, M. Karcher, and M. Oberlack, Implicit

[41] M. Tiberga, Development of a highfidelity multiphysics simulation tool for

[49] P. Ghysels, T. J. Ashby, K. Meerbergen, and W. Vanroose, Hiding Global Com

[50] K. J. Fidkowski, T. A. Oliver, J. Lu, and D. L. Darmofal, pMultigrid solution of

[51] L. Diosady and D. Darmofal, Discontinuous Galerkin Solutions of the Navier

[52] B. Helenbrook and H. Atkins, Application of PMultigrid to Discontinuous

[53] H. L. Atkins and B. T. Helenbrook, Numerical Evaluation of PMultigrid Method

[54] H. Luo, J. D. Baum, and R. Löhner, Fast pMultigrid Discontinuous Galerkin

[56] A. Karakus, N. Chalmers, K. Świrydowicz, and T. Warburton, A GPU acceler

[57] A. C. Kirby and D. J. Mavriplis, GPUAccelerated Discontinuous Galerkin Meth

1. Obtain predictors for (𝑘/𝑐𝑝 )∗ , and 𝐦∗ with a secondorder extrapolation from

3.3.1. TaylorGreen Vortex

3.3.2. Variableproperty Manufactured Solution

Figure 3.3: Constantdensity manufactured solution in Eq. 3.16 at 𝑡 = 1.

3.5.1. Pressure Correction with Equalorder Discretizations

[3] S. Rhebergen, B. Cockburn, and J. J. W. van der Vegt, A spacetime dis

[4] V. Ambati and O. Bokhove, Spacetime discontinuous galerkin discretization

[5] K. Shahbazi, P. F. Fischer, and C. R. Ethier, A highorder discontinuous Galerkin

[6] B. Klein, B. Müller, F. Kummer, and M. Oberlack, A highorder discontinuous

[7] B. Krank, N. Fehn, W. A. Wall, and M. Kronbichler, A highorder semiexplicit

[12] B. Klein, A highorder Discontinuous Galerkin solver for incompressible and

There is also substantial experience with discontinuous Galerkin (DG) discretiza