Vander Vegt

Discontinuous Galerkin Finite Element Methods for
Compressible Flows
Jaap van der Vegt
Numerical Analysis and Computational Mechanics Group
Department of Applied Mathematics
University of Twente
Joint Work with

Sander Rhebergen, Onno Bokhove, Christiaan Klaij, Fedderik van der Bos
Toulouse, December 8, 2009

University of Twente - Chair Numerical Analysis and Computational Mechanics 1
Introduction
Motivation of research:
• Many important flow phenomena are modelled by nonconservative (hyperbolic)

partial differential equations, e.g. dispersed multiphase flows.
• Our aim is to develop space-time discontinuous Galerkin discretizations which are

suitable for both conservative and nonconservative partial differential equations
Introduction
Motivation of research:
• Nonconservative hyperbolic partial differential equations contain nonconservative

products
∂xu + A(u)∂x u = 0
• The essential feature of nonconservative products is that A != Df , hence A is
not the Jacobian matrix of a flux function f .
• This causes problems once the solution becomes discontinuous, because the weak
solution in the classical sense of distributions then does not exist.
• This also complicates the derivation of discontinuous Galerkin discretizations since

there is no direct link with a Riemann problem.
• Alternative: use the theory for nonconservative products from Dal Maso, LeFloch
and Murat (DLM)
Overview of Presentation
• Overview of main results of the theory of Dal Maso, LeFloch and Murat for
nonconservative products
• Space-time DG discretization of nonconservative hyperbolic partial differential

equations
• Extension to the compressible Navier-Stokes equations
• Multigrid techniques for space-time DG discretizations
• Conclusions
Nonconservative Products
• Consider the function u(x)
u(x) = uL + H(x − xd)(uR − uL), x, xd ∈]a, b[,
with H : R → R the Heaviside function.
• For any smooth function g : Rm → Rm the product g(u)∂x u is not defined at

x = xd since here |∂xu| → ∞.
• Introduce a smooth regularization uε of u. If the total variation of uε remains

uniformly bounded with respect to ε then Dal Maso, LeFloch and Murat (DLM)
showed that
du ! ε " duε
g(u) ≡ lim g u
dx ε→0 dx
gives a sense to the nonconservative product as a bounded measure.
Effect of Path on Nonconservative Product
The limit of the regularized nonconservative product depends in general on the path
used in the regularization.
• Introduce a Lipschitz continuous path φ : [0, 1] → Rm, satisfying φ(0) = uL

and φ(1) = uR , connecting uL and uR in Rm.
• The following regularization uε for u then emerges:



 uL,
 if x ∈]a, xd − ε[,
ε x−x +ε
u (x) = φ( 2εd ), if x ∈]xd − ε, xd + ε[, ε > 0.


 uR , if x ∈]xd + ε, b[
• When ε tends to zero, then:

ε ' 1
ε du dφ
g(u ) $ Cδxd , with C = g(φ(τ )) (τ ) dτ,
dx 0 dτ
weakly in the sense of measures on ]a, b[, where δxd is the Dirac measure at xd .
• The limit of g(uε )∂x uε depends on the path φ.
• There is one exception, namely if an q : Rm → R exists with g = ∂uq . In this

case C = q(uR ) − q(uL ).
DLM Theory
Dal Maso, LeFloch and Murat provided a general theory for nonconservative
hyperbolic pde’s.
• Introduce the Lipschitz continuous maps φ : [0, 1] × Rm × Rm → Rm which

satisfy the following properties:
(H1) φ(0; uL , uR ) = uL , φ(1; uL, uR ) = uR ,

(H2) φ(τ ; uL, uL ) = uL,
( ∂φ (
(H3) ∂τ (τ ; uL, uR )( ≤ K|uL − uR |, a.e. in [0, 1].
(
• Theorem (DLM). Let u :]a, b[→ Rm be a function of bounded variation and

g : Rm → Rm a continuous function. Then, there exists a unique real-valued
bounded Borel measure µ on ]a, b[ with:
1. If u is continuous on a Borel set B ⊂]a, b[, then

'
du
µ(B) = g(u)
B dx
2. If u is discontinuous at a point xd of ]a, b[, then

' 1
∂φ
µ({xd }) = g(φ(τ ; uL, uR )) (τ ; uL , uR ) dτ.
0 ∂τ
By definition, this) measure* µ is the nonconservative product of g(u) by ∂xu and

denoted by µ = g(u) du dx φ .
Rankine-Hugoniot Relations
• For conservative hyperbolic system of pde’s, ∂x u + ∂xf (u) = 0 the

Rankine-Hugoniot relations across a jump with uL and uR and velocity v are
equal to
R L R L
−v(u − u ) + f (u ) − f (u ) = 0.
• For a nonconservative hyperbolic pde ∂xu + A(u)∂x u = 0 the Rankine-Hugoniot

relations in the DLM theory are equal to
' 1
R L L R L R
−v(u − u ) + A(φD (s, u , u ))∂s φD (s; u , u )ds = 0
0
with φD a Lipschitz continuous path satisfying φD (0; uL , uR ) = uL and

φD (1; uL, uR ) = uR .
• The Rankine-Hugoniot relations are essential for the definition of the NCP flux
used in the DG discretization.
Space-Time Discontinuous Galerkin Method
t
K
Ω n+1
t n+1
In
tn
Ωn
Qn
t0 y
• A time-dependent problem is considered directly in four dimensional space, with

time as the fourth dimension
Key Features of Space-Time Discontinuous Galerkin Methods
• Simultaneous discretization in space and time: time is considered as a fourth

dimension.
• Discontinuous basis functions, both in space and time, with only a weak coupling
across element faces resulting in an extremely local, element based discretization.
• The space-time DG method is closely related to the Arbitrary Lagrangian Eulerian

(ALE) method.
Benefits of Discontinuous Galerkin Methods
• Due to the extremely local discretization DG methods provide optimal flexibility for
! achieving higher order accuracy on unstructured meshes

! hp-mesh adaptation
! unstructured meshes containing different types of elements, such as tetrahedra,
hexahedra and prisms
! parallel computing
Benefits of Space-Time Discontinuous Galerkin Methods
• A conservative discretization is obtained on moving and deforming meshes.
• No data interpolation or extrapolation is necessary on dynamic meshes, at free

boundaries and after mesh adaptation.
Disadvantages of Space-(Time) Discontinuous Galerkin Methods
• Algorithms are generally rather complicated, in particular for elliptic and parabolic
partial differential equations
• On structured meshes DG methods are computationally more expensive than finite

difference and finite volume methods.
Space-Time Domain
• Consider an open domain: E ⊂ Rd.
• The flow domain Ω(t) at time t is defined as:
Ω(t) := {x ∈ E | x0 = t, t0 < t < T }
• The space-time domain boundary ∂E consists of the hypersurfaces:
Ω(t0) :={x ∈ ∂E | x0 = t0},

Ω(T ) :={x ∈ ∂E | x0 = T },
Q :={x ∈ ∂E | t0 < x0 < T }.
• The space-time domain is covered with a tessellation Th consisting of space-time

elements K.
Discontinuous Finite Element Approximation
• The finite element space associated with the tessellation Th is given by:
+ 2 m k m ,
Wh := W ∈ (L (Eh )) : W |K ◦ GK ∈ (P (K̂)) , ∀K ∈ Th
• The jump of f at an internal face S ∈ SIn in the direction k of a Cartesian

coordinate system is defined as:
[[f ]]k = f Ln̄L

k + f R R
n̄k ,
with n̄R L
k = −n̄k .
• The average of f at S ∈ SIn is defined as:

L R
{{f }} = 12 (f + f ).
Space-Time DG Formulation of Nonconservative Hyperbolic

PDE’s
• Consider the nonlinear hyperbolic system of partial differential equations in

nonconservative form in multi-dimensions:
∂Ui ∂Fik ∂Ur
+ + Gikr = 0, x̄ ∈ Ω ⊂ Rq , t > 0,
∂t ∂xk ∂xk
with U ∈ Rm, F ∈ Rm × Rq , G ∈ Rm × Rq × Rm
• These equations model for instance bubbly flows, granular flows, shallow water
equations and many other physical systems.
• Weak formulation for nonconservative hyperbolic system:
Find a U ∈ Vh , such that for V ∈ Vh the following relation is satisfied

- ' ! "
Vi Ui,0 + Fik,k + Gikr Ur,k dK
K∈Th K
- .' ' 0
+ /i (U R − U L ) dK −
V /i (U R − U L ) dK
V
i i i i
K∈Th K(t−
n+1 ) K(t+
n)
- ' .' 1 0
/i L R ∂φr L R L
+ V Gikr (φ(τ ; U , U )) (τ ; U , U ) dτ n̄k dS
S∈SI S 0 ∂τ
- '
− /i [[Fik − vk Ui ]]k dS = 0
V
S∈SI S
Relation with Space-Time DG Formulation of Conservative

Hyperbolic PDE’s
• Theorem 2. If the numerical flux V̂ for the test function V is defined as:

 {{V }} at S ∈ SI ,
/ =
V
0 at K(tn ) ⊂ Ωh(tn) ∀n ≥ 0,
then the DG formulation will reduce to the conservative space-time DG formulation

when there exists a Q, such that Gikr = ∂Qik /∂Ur .
• After the introduction of the numerical flux V̂ we obtain the weak formulation:
- ' ! "
− Vi,0Ui − Vi,k Fik + Vi Gikr Ur,k dK
K∈Th K
- .' ' 0
+ ViLUiL dK − ViLUiL dK
K∈Th K(t−
n+1 ) K(t+
n)
- ' L R L
+ (Vi − Vi ){{Fik − vk Ui }}n̄k dS
S∈SI S
- '
+ ViL(Fik
L
− vk UiL )n̄L
k dS
S∈SB S
- ' .' 1 0
L R ∂φr L R L
+ {{Vi }} Gikr (φ(τ ; U , U )) (τ ; U , U ) dτ n̄k dS = 0
S∈SI S 0 ∂τ
Numerical Fluxes
• The fluxes at the element faces do not contain any stabilizing terms yet, both for
the conservative and nonconservative part
• At the time faces, the numerical flux is selected such that causality in time is
ensured 1
L −
/ = U at K(t n+1 )
U .
U R at K(t+ n )
• The space-time DG formulation is stabilized using the NCP (Non-Conservative

Product) flux ! " L
/ nc
Pi = {{Fik − vk Ui }} + Pik n̄k
Nonconservative Product Flux
t
SL " SM SR
#
# $$ %% T
# $ %
# Ω2 $ Ω3 %
# $ %
# (U = UL∗ ) $ (U = UR ∗
) %
# $ %
# $ %
Ω1 # $ % Ω4
(U = UL ) # $ % (U = UR )
# $ %
# $ %
# $ %
# $ %
#$% ! x
xL xR
Wave pattern of the solution for the Riemann problem

NCP Flux
Main steps in derivation of NCP flux:
• Consider the nonconservative hyperbolic system:
∂tU + ∂xF (U ) + G(U )∂x U = 0,
∗
• Introduce the averaged exact solution ŪLR (T ) as:
' T SR
∗ 1
ŪLR (T ) = U (x, T ) dx.
T (SR − SL) T SL
• Apply the Gauss theorem over each subdomain Ω1, · · · , Ω4 and connect each
subdomain using the generalized Rankine-Hugoniot relations.
• The NCP-flux is then given by:

 2

 F L L
n̄ − 1 1G ∂ φ̄r L

 ik k 2 0 ikr (φ̄(τ ; UL , UR )) ∂τ (τ ; UL , UR ) dτ n̄k



 if SL > v,



 {{F }}n̄L + 1 ! (S − v)Ū ∗ + (S − v)Ū ∗ − S U L − S U R )

nc L
/ (UL, UR , v, n̄ ) = ik k 2 R i L i L i R i
P i 
 if SL < v < SR ,



 2

 R L 1 1G ∂ φ̄r L

 F n̄
ik k + 2 0 ikr (φ̄(τ ; UL , UR )) ∂τ (τ ; UL , UR ) dτ n̄k


 if SR < v,
• Note, if G is the Jacobian of some flux function Q, then P/ nc (UL, UR , v, n̄L) is

exactly the HLL flux derived for moving grids in van der Vegt and van der Ven
(2002).
Efficient Solution of Nonlinear Algebraic System
• The space-time DG discretization results in a large system of nonlinear algebraic

equations:
L(Û n ; Û n−1) = 0
• This system is solved by marching to steady state using pseudo-time integration
and multigrid techniques:
∂ Û 1 n−1
=− L(Û ; Û )
∂τ ∆t
Depth averaged two-fluid model
• The dimensionless depth-averaged two fluid model of Pitman and Le, ignoring
source terms for simplicity, can be written as:
∂t U + ∂xF + G∂xU = 0,
where:
   
h(1 − α) h(1 − α)u
   hαv 
 hα   
 ,  hαv 2 + 1 ε(1 − ρ)α gh2 α
U = hαv  F = 2 xx 
 hu(1 − α)  2 1 2 
 hu + 2 εgh 
b 0
 
0 0 0 0 0
 0 0 0 0 0 
 
 εραgh εραgh 0 0 ε(1 − ρ)αxxghα + εραgh
G(U ) =  .
 2u2 α 2 

1−α − αu − εghα −εghα − αu2 u(α − 1) 2uα
uα − 1−α (1 − α)εgh 
0 0 0 0 0
1.5
h(1−α)+b,hα+b,h+b,b
h+b
h(1−α)+b
1
hα+b
b
0.5
0
0 5 10 15 20
x
Steady-state solution for a subcritical two-phase flow (320 cells).
Total flow height h + b, flow height due to the fluid phase h(1 − α), flow height due
to solids phase hα and the topography b.
STDGFEM
h(1 − α) + b hα + b
Ncells L2 error p Lmax error p L2 error p Lmax error p
40 0.8171 · 10−3 - 0.2308 · 10−2 - 0.1404 · 10−2 - 0.4194 · 10−2 -

80 0.2025 · 10−3 2.0 0.5584 · 10−3 2.0 0.3537 · 10−3 2.0 0.9903 · 10−3 2.1
160 0.4871 · 10−4 2.1 0.1322 · 10−3 2.1 0.8511 · 10−4 2.1 0.2306 · 10−3 2.1
320 0.9789 · 10−5 2.3 0.2651 · 10−4 2.3 0.1712 · 10−4 2.3 0.4597 · 10−4 2.3
hu(1 − α) hv(α)
Ncells L2 error p Lmax error p L2 error p Lmax error p
40 0.3672 · 10−4 - 0.1442 · 10−3 - 0.1212 · 10−4 - 0.3409 · 10−4 -

80 0.5911 · 10−5 2.6 0.3448 · 10−4 2.1 0.1791 · 10−5 2.8 0.8054 · 10−5 2.1
160 0.1049 · 10−5 2.5 0.8471 · 10−5 2.0 0.3807 · 10−6 2.2 0.2048 · 10−5 2.0
320 0.1723 · 10−6 2.6 0.2078 · 10−5 2.0 0.5115 · 10−7 2.9 0.4861 · 10−6 2.1
Error in h(1 − α) + b, hα + b, hu(1 − α) and hvα for subcritical flow over a bump.
Two-phase dam break problem
3.5 0.3 1
3 0.2
0.8
0.1
2.5
hu(1−α),hvα
h(1−α),hα,h
h(1−α) 10000 0 0.6

2 hα 10000 hu(1−α) 10000
h 10000 hvα 10000 α 10000
−0.1
α
h(1−α) 128 hu(1−α) 128 α 128
1.5 hα 128 hvα 128
h 128 −0.2 0.4
1
−0.3
0.2
0.5 −0.4
0 −0.5 0
0 0.5 1 0 0.5 1 0 0.2 0.4 0.6 0.8 1
x x x
(a) Solution of h(1 − α), hα, b (b) Solution of hu(1 − α) and hvα. (c) Solution of α.
and h.
Two-phase dam break problem at time t = 0.175; mesh with 128 elements compared
to mesh with 10000 elements.
Effect of Path
3.5
h(1−α) 1
h(1−α) 1
hα 1
3 hα 1
h1 3 h1
h(1−α) t1
h(1−α) t1
2.5 hα t1
hα t1
h t1
h t1
h(1−α),hα,b,h
h(1−α) t2
h(1−α),hα,h
2 hα t2 2.5 h(1−α) t2
hα t2
h t2
h t2
1.5 h(1−α) t3
h(1−α) t3
hα t3
hα t3
h t3 2
1 h(1−α) t4
h t3
h(1−α) t4
hα t4
hα t4
0.5 h t4
h t4
h(1−α) t5
h(1−α) t5
0 hα t5 1.5
hα t5
h t5
h t5
b
−0.5
0 0.2 0.4 0.6 0.8 1 0.095 0.1 0.105 0.11
x x
(d) The solution on the whole domain. (e) The solution zoomed in on the left shock wave.
Solution of h(1 − α), hα, b and h at time t = 0.175 calculated on a mesh with
1024 elements using the paths defined by Toumi.
Flow Through a Contraction
5 1
h
x1
0.5
10
0 x2
h: 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
15 -0.5
Flow depth h of water-sand mixture in a contraction

h/L = 0.01, ρf /ρs = 0.5, slope 10◦.
Compressible Navier-Stokes Equations
• Compressible Navier-Stokes equations in space-time domain E :
∂Ui ∂Fke(U ) ∂Fke(U, ∇U )

+ − =0
∂x0 ∂xk ∂xk
• Conservative variables U ∈ R5 and inviscid fluxes F e ∈ R5×3

   
ρ ρuk
U =  ρuj  , Fke =  ρuj uk + pδjk 
ρE ρhuk
• Viscous flux F v ∈ R5×3  

0
v
= Fk τjk 
τkj uj − qk
with the total stress tensor τ is defined as:
∂ui ∂uj ∂uk
τjk = λ δjk + µ( + )
∂xi ∂xk ∂xj
and heat flux vector q

* is defined as:
∂T
qk = −κ
∂xk
• The viscous flux F v is homogeneous with respect to the gradient of the

conservative variables ∇U :
v ∂Ur
Fik (U, ∇U ) = Aikrs(U )
∂xs
with the homogeneity tensor A ∈ R5×3×5×3 defined as:

v
∂Fik (U, ∇U )
Aikrs (U ) :=
∂(∇U )
• The system is closed using the equations of state for an ideal gas.
First Order System
• Rewrite the compressible Navier-Stokes equations as a first-order system using the

auxiliary variable Θ:
e
∂Ui ∂Fik (U ) ∂Θik (U )
+ − = 0,
∂x0 ∂xk ∂xk
∂Ur
Θik (U ) − Aikrs (U ) = 0.
∂xs
Weak Formulation
• Weak formulation for the compressible Navier-Stokes equations

Find a U ∈ Wh, Θ ∈ Vh , such that for all W ∈ Wh and V ∈ Vh, the following
holds:
- ' ! ∂Wi ∂Wi e "
− Ui + (Fik − Θik ) dK
K∈Th K
∂x0 ∂xk
- ' L /
+ Wi (U /e / L
i + Fik − Θik )nk d(∂K) = 0,
K∈Th ∂K
- ' - ' ∂Ur

Vik Θik dK = Vik Aikrs dK
K K ∂xs
K∈T n K∈T n
h h
- '
+ VikLAL / L L
ikrs (Ur − Ur )n̄s dQ
K∈T n Q
h
Transformation to Arbitrary Lagrangian Eulerian form
• The space-time normal vector on a grid moving with velocity *

v is:

 T
at K(t−
 (1, 0, 0, 0)
 n+1 ),
n = (−1, 0, 0, 0)T at K(t+ n ),


 (−vk n̄k , n̄)T at Qn.
• The boundary integral then transforms into:

- ' L /
Wi (U /e / L
i + Fik − Θik )nk d(∂K)
K∈Th ∂K
- 9' ' :
= WiLU
/i dK + WiL U
/i dK
K∈Th K(t−
n+1 ) K(t+
n)
- ' L /e L
+ Wi (F / /
ik − Ui vk − Θik )n̄k dQ
K∈Th Q
Numerical Fluxes
• The numerical flux U / at K(t− ) and K(t+) is defined as an upwind flux to

n+1 n
ensure causality in time:
1
/ U L at K(t−
n+1 ),
U =
U R at K(t+
n ),
• At the space-time faces Q we introduce the HLLC approximate Riemann solver as

numerical flux:
/e − U
n̄k (F / i vk )(U L , U R ) = H HLLC(U L , U R , v, n̄)
ik i
ALE Weak Formulation
• The ALE flux formulation of the compressible Navier-Stokes equations transforms

now into:
Find a U ∈ Wh, such that for all W ∈ Wh, the following holds:
- ' ! ∂Wi ∂Wi e "
n K ∂x0
K∈T
∂xk
h
- 9' L L
'
L R
:
+ Wi Ui dK − Wi Ui dK
K∈T n K(t−
n+1 ) K(t+
n)
h
- '
+ WiL(HiHLLC(U L , U R , v, n̄) − Θ
/ ik n̄L) dQ = 0.
k
K∈T n Q
h
Lifting Operator
• Introduce the global lifting operator R ∈ R5×3, defined in a weak sense as:
Find an R ∈ Vh, such that for all V ∈ Vh :
- ' - '
Vik Rik dK = {{Vik Aikrs }}[[Ur ]]s dS
K∈T n K S∈S n S
h I
- ' L L L b L
+ Vik Aikrs (Ur − Ur )n̄s dS.
S∈S n S
B
• The weak formulation for the auxiliary variable is now transformed into
- ' - ' ∂Ur
Vik Θik dK = Vik (Aikrs − Rik ) dK, ∀V ∈ Vh .
n K n K ∂x s
K∈T K∈T
h h
Θ Equation
• The primal formulation can be obtained by eliminating the auxiliary variable Θ

using
∂Ur n
Θik = Aikrs − Rik , a.e. in Eh
∂xs
• Note, this is possible since ∇hWh ⊂ Vh
ALE Weak Formulation for Primal Variables
• Recall the ALE flux formulation of the compressible Navier-Stokes equations:
Find a U ∈ Wh, such that for all W ∈ Wh, the following holds:
- ' ! ∂Wi ∂Wi e "
n K ∂x 0 ∂xk
K∈T
h
- 9' ' :
+ WiLUiL dK − WiL UiR dK
K∈T n K(t−
n+1 ) K(t+
n)
h
- '
+ WiL(HiHLLC(U L , U R , v, n̄) − Θ
/ ik n̄L) dQ = 0.
k
K∈T n Q
h
Numerical Fluxes for Θ
/ in the primary equation is defined following Brezzi as a

• The numerical flux Θ
/ = {{Θ}}:
central flux Θ
1 S
L R {{Aikrs ∂Ur
∂xs − ηRik }
} for internal faces,
/ ik (U , U ) =
Θ ∂U b
Abikrs ∂xsr − ηRSik for boundary faces,
• The local lifting operator RS ∈ R5×3 is defined as follows:

Find an RS ∈ Vh, such that for all V ∈ Vh:
 '


 {{Vik Aikrs }}[[Ur ]]s dS for internal faces,
- ' 
S
Vik RSik dK = '
K 

K∈T n
h

 VikLAL L b
ikrs (Ur − Ur )n̄s dS for external faces.
S
Delta Wing Simulations
• Simulations of viscous flow about a delta wing with 85◦ sweep angle.
• Conditions
! Mach number M = 0.3

! Reynolds number Re = 40.000
! Angle of attack α = 12.5◦ .
! Fine grid mesh 1.600.000 elements, 40.000.000 degrees of freedom
! Adapted mesh, initial mesh 208.896 elements, after four adaptations 286.416
elements
Streaklines and vorticity contours in various cross-sections

Impression of the vorticity based mesh adaptation

Large eddy simulation of turbulent flow about a delta wing

Computational Efficiency
• Computational efficiency is the key factor limiting industrial applications of higher

order accurate discontinuous Galerkin methods in computational fluid dynamics.
• Multigrid methods are good candidates to increase computational efficiency, but

need significant improvements for higher order accurate DG discretizations.
Objectives
• Perform a theoretical analysis of multigrid performance for advection dominated

flows, in particular for higher order accurate DG discretizations.
• Improve multigrid performance using theoretical analysis tools.
• Test multigrid performance on realistic problems.

Main Components of h-Multigrid Algorithm
• Consider a finite sequence Nc of increasingly coarser meshes Mnh,

n ∈ {1, · · · , Nc}
• Define operators to connect data on the different meshes:
! restriction operators
mh
Rnh : Mnh → Mmh, 1 ≤ n < m ≤ Nc ,
! prolongation operators
nh
Pmh : Mmh → Mnh, 1 ≤ n < m ≤ Nc .
• Use iterative solvers Snh to approximately solve the system of algebraic equations
on the various grid levels
Lnhvh = fnh on Mnh
• Since, the main effect of the multigrid algorithm should be the damping of high
frequency error components Snh is called a smoothing operator.
• Choose a cycling strategy between the different meshes, e.g. V- or W-cycle.

Multigrid error transformation operator for linear problems
• In order to understand the performance of the multigrid algorithm we need to

investigate the multigrid error transformation operator.
• The multigrid error transformation operator shows how much the error in the
iterative solution of the algebraic system is reduced by one full multigrid cycle.
• We analyze three-level multigrid algorithms for 2D problems to obtain better

estimates for the convergence rate.
• Given an initial error eA D

h , the error eh after one full multigrid cycle with three grid
levels is given by the relation
D 3g A
eh = Mh eh
with
3g ν h γ −1 2h ν
Mh = Sh2 (Ih − P2h(I2h − M2hc )L2h Rh Lh)Sh1
and
ν 2h −1 4h ν
M2h = S2h4 (I2h − P4h L4h R2h L2h )S2h3 .
• The properties of the multigrid error transformation operator are analyzed using
discrete Fourier analysis.
Three-grid Fourier analysis
θ2
π/2
π/4
−π/4
−π/2
−π
−π −π/2 −π/4 π/4 π/2 π
0
θ1
Low, medium and high frequencies Fourier modes in three-level multigrid

Multigrid Error Transformation Operator
• The discrete Fourier transform of the error transformation operator for a three-level
; 3g (θ) ∈ C16m×16m which is equal to
multigrid cycle M h
! " 9 :! "
; 3g (θ)
M = / 3g (θ) ν2
S I 3g / 3g (θ)U
−P / 3g (θ; γc)Q
/ 3g (θ)R
/ 3g (θ)L
/ 3g (θ) / 3g (θ) ν1
S
h h h h h h h
∀θ ∈ Θ4h \ Ψ3g
/ 3g (θ; γc ) is equal to
The matrix U
! 2g "
; (2θβ ) γc .
/ 3g (θ; γc) = I 2g − M
U 2h
with
2g ! 2g " ν4 9 2g 2g −1 0 / 2g 2g
:!
2g " ν3
; /
M2h (2θβ ) = S2h (2θβ ) / / / /
I −P2h (2θβ )L4h (4θ0 )R2h (2θβ )L2h (2θβ ) S2h (2θβ )
Asymptotic Convergence Rate
• The asymptotic convergence factor per cycle is defined as

 m
1
(m)
 /eh /12 (G ) 
h
µ = lim  sup (0) 
m→∞
e !=0 /eh /12 (G )
(0)
h h
and can be rewritten into

µ = ρ(Mhng ).
Optimization of multigrid performance
• The spectral radius of the error transformation operator can be used to optimize
the multigrid algorithm.
• In particular, the smoother is a good candidate for optimization since in general it

contains a number of free parameters.
• We will consider a Runge-Kutta type smoother.

Analysis and Optimization of Runge-Kutta Smoothers for the

Advection-Diffusion Equation
• The advection-diffusion equation in the domain Ω is defined as
∂u
+ a · ∇u = ∇ · (A∇u))
∂t
• The equations are discretized with a higher order accurate space-time

discontinuous Galerkin finite element method.
• The space-time discretizations are solved using a multigrid algorithm with a

Runge-Kutta type smoother.
Pseudo-time Runge-Kutta smoothers
• Let the system of algebraic equations be denoted as

n n−1
L(û , û ) = 0.
• A pseudo time derivative is added to the system and integrated to steady-state in

pseudo-time
∂ û∗
M = −L(û∗ , ûn−1),
∂τ
• At steady state ûn = û∗ .
• For the pseudo-time integration we consider 4- and 5-stage Runge-Kutta methods.

• A four-stage Runge-Kutta scheme is given by:
1
(1 + β1 λI)V̂ 1 =V̂ 0 − α21λM̄−1L(V̂ 0 ) + β1 λV̂ 0
∆x∆y
2 0λM̄−1 ! 0 1 " 1
(1 + β2 λI)V̂ =V̂ − α31 L(V̂ ) + α32 L(V̂ ) + β2 λV̂
∆x∆y
3 0λM̄−1 ! "
(1 + β3 λI)V̂ =V̂ − α41 L(V̂ 0) + α42 L(V̂ 1 ) + α43L(V̂ 2) + β3 λV̂ 2
∆x∆y
4 0λM̄−1 ! 0 1 2 3 "
(1 + β4 λI)V̂ =V̂ − α51 L(V̂ ) + α52 L(V̂ ) + α53L(V̂ ) + α54 L(V̂ )
∆x∆y
3
+β4 λV̂
• We require from the 4- and 5-stage Runge-Kutta schemes that they are second
order accurate
Optimized Runge-Kutta smoothers for space-time DGFEM
• Optimization of the smoother coefficients for three-level multigrid algorithms
• Space-time DG discretization of the 2D advection-diffusion equation using

quadratic basis-functions
• For the optimization process, we fix the Reynolds numbers Rex = Rey = 100,
the flow angle γ f low = π/4 and the aspect ratio AR = 1.
• For steady flows we fix CF L∆t = 100, while for unsteady flows CF L∆t = 1.
• The optimization process has a big impact on the Runge-Kutta coefficients which
greatly differ per case.
• The Runge-Kutta coefficients are very different from commonly used schemes since
they are optimized for fast multigrid convergence and not for time accuracy.
• Finding a good balance between optimization and more generally applicable

algorithms is still an open question.
• In practice, the Runge-Kutta smoothers use local time stepping and this gives the
opportunity to apply locally the best smoother for the actual flow state.
Optimized coefficients for dRK5 and fRK5 smoothers for 3-level multigrid (steady flow).
dRK5 p = 1 fRK5 p = 1 dRK5 p = 2 fRK5 p = 2
α21 0.05768995298 0.0578331573 0.04865009589 0.04877436325
α31 - -0.0002051554736 - -0.0002188348438
α32 0.1405960888 0.1403808301 0.130316854 0.1300906122
α41 - 0.0003953470071 - 2.608884832e-05
α42 - -0.001195029164 - 2.444376496e-05
α43 0.267958213 0.2681810517 0.2729621396 0.2734805705
α51 - 0.0001441249202 - -0.001250385487
α52 - -0.0002608610327 - -0.0007838720635
α53 - -0.0003368070181 - -0.0004890887712
α54 0.5 0.8473374098 0.5 4.412139367
α61 - 0.4115573097 - 0.8097217358
α62 - -0.003144851878 - 0.08435089009
α63 - -0.0001096455683 - -0.01986799007
α64 - 0.001555741114 - 0.01359815476
α65 1.0 0.5901414466 1.0 0.1121972094
CFLτ 0.8 0.8 0.4 0.4
ρS 0.98812 0.98914 0.98974 0.9896
ρM G 0.89151 0.81762 0.90049 0.89903
ρEXI−M G 167.06 - 124.02 -
Stability domain of smoother Spectrum of L and stability domain of smoother

5 5
low
4 4 high
0.18
0.9
0.7
3 918
0.0. 3
0.7 0.06.5
0.60.5
0.4
2 0.3 0.1 2
0.4
5
00. .4 0.2
0.3
0. 6
0.
0.1 0.2
3
2
0.
0 .81
1 0.1 1
0.70.9
0.9
Im(z)
Im(z)
0.71
0.8
0 0
0.1
0.7
0.8 0.19
0.1
0.2
0.30.4
−1 0.2 −1
0.6
0. 0.3
5
0.5
−2 0.6 0.4 −2
.4
0.1 .2
00.3
0. 0.5 0
0.78 0.6
−3 0.
19 −3
0.7
0.801.9
−4 −4
−5 −5
−10 −8 −6 −4 −2 0 2 −10 −8 −6 −4 −2 0 2
Re(z) Re(z)
(a) Stability domain dRK5. (b) Spectrum of Lh and stability domain dRK5.
Stability domain of dRK5 smoother and spectrum of a space-time DG discretization of

the 2D advection-diffusion equation using quadratic basis functions.
Stability domain of smoother Spectrum of L and stability domain of smoother

5 5
low
4 4 high
10.
3 10.98 00.7.98 3
0.
0.7 0.60.5
0. 0 0.4
0.4
6 .5
2 0.3 2
0.03 .2 0.1
0..5
4
00.6
0. 0.2
00..89 1
0.3
2
1
1 1
0.
0.1
Im(z)
Im(z)
0.7
1
0.80.9 1
0 0
0.9
0.1
0.2
0.3
0.7
0.8
0.1
−1 −1
0.4 0
0.2
0.6
0.3
.
0.5
−2 −2
5
0.6 0.4 0.1
0.4
0.07.8 0.5 00.2.3
0.9 0.6
−3 1 0.0.8
7 −3
0.9 1
−4 −4
−5 −5
−10 −8 −6 −4 −2 0 2 −10 −8 −6 −4 −2 0 2
Re(z) Re(z)
(a) Stability domain fRK5. (b) Spectrum of Lh and stability domain fRK5.
Stability domain of the fRK5 smoother and spectrum of a space-time DG discretization

of the 2D advection-diffusion equation using quadratic basis functions.
Spectrum of smoother Spectrum of smoother

low low
0.8 high 0.8 high
0.6 0.6
0.4 0.4
0.2 0.2
Im
Im
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

Re Re
(a) dRK5 smoother. (b) fRK5 smoother.

Spectrum of three−level operator Spectrum of three−level operator
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
Im
Im
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

Re Re
(c) 3-level MG with dRK5. (d) 3-level MG with fRK5.
Eigenvalue spectra for space-time DG discretizations of the 2D advection-diffusion

equation using quadratic basis functions.
Testing multigrid performance
• In order to demonstrate the performance of the optimized algorithms we consider

the 2D advection-diffusion equation on a square.
• The exact steady state solution is
. 0
1 exp(a1 /ν1 ) − exp(a1 x1 /ν1) exp(a2 /ν2 ) − exp(a2 x2 /ν2)
u(x1 , x2 ) = + .
2 exp(a1 /ν1) − 1 exp(a2 /ν2 ) − 1
• The advection-diffusion equation is discretized using the space-time discontinuous

Galerkin discretization with quadratic basis functions.
• In the discretization we use a Shishkin mesh which is suitable for dealing with
boundary layers.
• In pseudo-time we apply a rescaling to reduce the effect of grid stretching.

• The parameters in the tests are:

√
! ∆t = 100, a = 2, νx = νy = 0.01, N1 = N2 = 32.
! Flow angle γ f low = π/4.
! Depending on the stability of the smoother, we use different CF Lτ and V N τ
numbers.
! The EXV scheme was used when Rei ≤ 1 and the EXI scheme was used
otherwise.
! For the multigrid computations, ν1 = ν2 = ν3 = ν4 = 1, νC = 4 and γ = 1.
10-1
dRK5 coarse approx.
fRK5 coarse approx.
EXI/EXV coarse approx.
dRK5 coarse exact
fRK5 coarse exact
10-2 EXI/EXV coarse exact
L / L0
-3
10
-4
10
500 1000 1500 2000 2500 3000
Work units
Convergence of three-level multigrid with different Runge-Kutta smoothers for second

order accurate space-time DG discretization.
10-1 dRK5 coarse approx.

fRK5 coarse approx.
EXI/EXV coarse approx.
dRK5 coarse exact
fRK5 coarse exact
EXI/EXV coarse exact
L / L0
-2
10
-3
10
-4
10
5000 10000 15000 20000
Work units
Convergence of three level multigrid with different Runge-Kutta smoothers for third
order accurate space-time DG discretization.
We can draw the following conclusions:
• In all cases using the optimized Runge-Kutta smoothers a big improvement is

obtained over the original EXI-EXV smoother.
• For linear basis functions the number of multigrid work units to obtain 4 orders of
reduction in the residual is reduced from 3300 to 380.
• For quadratic basis functions the number of multigrid work units reduces from
22000 to 184.
• For linear basis functions the difference in convergence rate between Runge-Kutta
smoothers with only non-zero diagonal terms versus Runge-Kutta smoothers with
all coefficients non-zero is negligible.
• For quadratic basis functions this difference is, however, significant.
• Using more Runge-Kutta coefficients enlarges the possibilities to optimize the

smoother, but the optimization process requires a significantly larger computing
time.
• In order to speed up the optimization process the coefficients of Runge-Kutta

schemes with only non-zero diagonal terms are used as initial values.
• The effect of solving the equations on the coarsest mesh is very large.
• In particular, for nonlinear problems it is tempting to solve the algebraic system on

the coarsest mesh only approximately, but the effect is non-negligible.
• The flow angle has a small effect on the convergence rate. In general if the flow
angle is close to one of the mesh lines the convergence rate is the slowest.
• It is important not to extrapolate the results for the advection-diffusion equation

directly to the Euler and Navier-Stokes equations.
Euler Equations
Amount of work per multigrid cycle differs single grid, p-, h-, and hp-multigrid:


 g p bp , SG,



 p p p−1 p−1 p p p−2 p−2 p

 (g b + g b )(ν 1 + ν 2 ) + g b νC , pMG,
! "
work per cycle = g p bp (ch + ch−1)(ν1h + ν2h) + ch−2νC h
, hMG,


 c h
(g p p
b + g p−1 p−1
b )(ν p
+ ν p


 1 2 )+

 g p−2 bp−2((ch + ch−1)(ν1h + ν2h) + ch−2νC h
), hpMG,
with g p and bp, respectively, the number of Gauss quadrature points and basis
functions in an element, and ch a weighting for the number of cells depending on the
grid-level h.
Coefficients: g p = 9, g p−1 = 4, g p−2 = 1, bp = 6, bp−1 = 4, bp−2 = 1, ch = 1,

ch−1 = 1/4 and ch−2 = 1/16.
p-levels h-levels #τ
CF L ρ work
1 1 1.0 0.99010 75006

3 1 1.0 0.74318 33370
3 3 1.0 0.73347 32091
1 3 1.0 0.66350 24098
1 3 0.5 0.80364 45360
Spectral radii and multigrid work units of different multigrid strategies.

0.001
SG
hMG
hMG exact
pMG
0.0001 hpMG
1e-05
1e-06
L2 total
1e-07
1e-08
1e-09
1e-10
0 200 400 600 800 1000 1200 1400 1600
Work Units
Convergence history of different multigrid techniques for space-time DG discretization

for inviscid flow around a NACA0012 airfoil (MTC 1 test case, α = 2◦, M a = 0.5,
448 × 64 elements).
Conclusions
• A space-time DG discretization for nonconservative hyperbolic pde’s using the

DLM theory and the compressible Navier-Stokes has been developed.
• A new numerical flux for nonconservative hyperbolic pde’s has been developed,
which reduces to the HLLC flux for conservative pde’s.
• The effect of the choice of the path in phase space is in practice for nearly all cases
negligible.
• The algorithm has been successfully tested on a depth averaged two-phase flow
model and compressible flow simulations.
Conclusions
• A detailed three-level multigrid analysis has been conducted for h-, p- and
hp-multigrid algorithms for the advection-diffusion and the linearized Euler
equations.
• The analysis provides the asymptotic convergence rate of the multigrid algorithms
which is used to optimize the multigrid smoothers.
• For the advection-diffusion equation the optimization results in a significant

improvement in the convergence rate in actual computations using the h-multigrid
algorithm.
• For the Euler equations the optimized Runge-Kutta smoothers for h-, p- and
hp-multigrid show initially a significant improvement in convergence rate for
inviscid flow about a NACA 0012 airfoil, but asymptotically most algorithms have
the same convergence rate.
• Using an exact solution of the algebraic system on the coarsest mesh has a major
impact ion the convergence rate for the advection-diffusion equation, but not for
the Euler equations.

Vander Vegt

Uploaded by

Document Information

Original Description:

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

Vander Vegt

Uploaded by

Copyright:

Available Formats

Discontinuous Galerkin Finite Element Methods for

Joint Work with

Toulouse, December 8, 2009

• Many important flow phenomena are modelled by nonconservative (hyperbolic)

• Our aim is to develop space-time discontinuous Galerkin discretizations which are

• Nonconservative hyperbolic partial differential equations contain nonconservative

• This also complicates the derivation of discontinuous Galerkin discretizations since

• Space-time DG discretization of nonconservative hyperbolic partial differential

• Extension to the compressible Navier-Stokes equations

• Multigrid techniques for space-time DG discretizations

• Consider the function u(x)

u(x) = uL + H(x − xd)(uR − uL), x, xd ∈]a, b[,

with H : R → R the Heaviside function.

• For any smooth function g : Rm → Rm the product g(u)∂x u is not defined at

• Introduce a smooth regularization uε of u. If the total variation of uε remains

Effect of Path on Nonconservative Product

• Introduce a Lipschitz continuous path φ : [0, 1] → Rm, satisfying φ(0) = uL

• The following regularization uε for u then emerges:

• When ε tends to zero, then:

• The limit of g(uε )∂x uε depends on the path φ.

• There is one exception, namely if an q : Rm → R exists with g = ∂uq . In this

• Introduce the Lipschitz continuous maps φ : [0, 1] × Rm × Rm → Rm which

(H1) φ(0; uL , uR ) = uL , φ(1; uL, uR ) = uR ,

• Theorem (DLM). Let u :]a, b[→ Rm be a function of bounded variation and

1. If u is continuous on a Borel set B ⊂]a, b[, then

2. If u is discontinuous at a point xd of ]a, b[, then

By definition, this) measure* µ is the nonconservative product of g(u) by ∂xu and

• For conservative hyperbolic system of pde’s, ∂x u + ∂xf (u) = 0 the

• For a nonconservative hyperbolic pde ∂xu + A(u)∂x u = 0 the Rankine-Hugoniot

with φD a Lipschitz continuous path satisfying φD (0; uL , uR ) = uL and

Space-Time Discontinuous Galerkin Method

• A time-dependent problem is considered directly in four dimensional space, with

Key Features of Space-Time Discontinuous Galerkin Methods

• Simultaneous discretization in space and time: time is considered as a fourth

• The space-time DG method is closely related to the Arbitrary Lagrangian Eulerian

Benefits of Discontinuous Galerkin Methods

! achieving higher order accuracy on unstructured meshes

Benefits of Space-Time Discontinuous Galerkin Methods

• A conservative discretization is obtained on moving and deforming meshes.

• No data interpolation or extrapolation is necessary on dynamic meshes, at free

Disadvantages of Space-(Time) Discontinuous Galerkin Methods

• On structured meshes DG methods are computationally more expensive than finite

• Consider an open domain: E ⊂ Rd.

• The flow domain Ω(t) at time t is defined as:

Ω(t) := {x ∈ E | x0 = t, t0 < t < T }

• The space-time domain boundary ∂E consists of the hypersurfaces:

Ω(t0) :={x ∈ ∂E | x0 = t0},

• The space-time domain is covered with a tessellation Th consisting of space-time

Discontinuous Finite Element Approximation

• The jump of f at an internal face S ∈ SIn in the direction k of a Cartesian

[[f ]]k = f Ln̄L

• The average of f at S ∈ SIn is defined as:

Space-Time DG Formulation of Nonconservative Hyperbolic

• Consider the nonlinear hyperbolic system of partial differential equations in

• Weak formulation for nonconservative hyperbolic system:

Find a U ∈ Vh , such that for V ∈ Vh the following relation is satisfied

Relation with Space-Time DG Formulation of Conservative

then the DG formulation will reduce to the conservative space-time DG formulation

• The space-time DG formulation is stabilized using the NCP (Non-Conservative

Nonconservative Product Flux

Wave pattern of the solution for the Riemann problem

Main steps in derivation of NCP flux:

• Consider the nonconservative hyperbolic system: