Least-Squares Finite Element

Methods for Nonlinear Hyperbolic

PDEs

Hans De Sterck

Department of Applied Mathematics

University of Colorado at Boulder
(desterck@colorado.edu)

Lehrstuhl fuer Simulation, Universitaet Erlangen-Nuernberg

June 2004

< > -+
University of Colorado at Boulder
Erlangen p.1
 Outline

(1) Hyperbolic Conservation Laws: Introduction

(2) Least-Squares Finite Element Methods

(3) Fluid Dynamics Applications

< > -+
University of Colorado at Boulder
Erlangen p.2
(1) Numerical Simulation of Nonlinear Hyperbolic PDE Systems

Example application: gas dynamics

supersonic
flow of air
over sphere
(M=1.53)
bow shock
(An album of
fluid motion,
Van Dyke)

< > -+
University of Colorado at Boulder
Erlangen p.3
 Nonlinear Hyperbolic Conservation Laws

Euler equations of gas dynamics

~v

~v ~v + p I~

~v + =0
t
e ( e + p(, e) ) ~v

nonlinear hyperbolic PDE system

U
+ F~ (U ) = 0
t

conservation law
Z  I

U dV + ~n F~ (U ) dA = 0
t

< > -+
University of Colorado at Boulder
Erlangen p.4
 Model Problem: Scalar Inviscid Burgers Equation

scalar conservation law in 1D

u f (u)
+ =0
t x

model problem: inviscid Burgers equation

u u2 /2
+ =0
t x

< > -+
University of Colorado at Boulder
Erlangen p.5
 Burgers Equation: Model Flow

u u2 /2
+ =0
t x

hyperbolic PDE: information

propagates along charac- t
teristic curves u=1

u is constant on characteristic
u is slope of characteristic u=1 u=0
where characteristics cross:
shock formation (weak solu-
tion)
u=0 x

< > -+
University of Colorado at Boulder
Erlangen p.6
 Space-Time Formulation

u f (u)
+ =0
t x
define x,t = (x , t )
define f~x,t (u) = (f (u), u)
 

x,t f~x,t (u) = 0 R2

u=g I

 

conservation in space-time
I
~nx,t f~x,t (u) dl = 0

< > -+
University of Colorado at Boulder
Erlangen p.7
 Some Notation

L2 scalar product L2 norm

Z sZ
hf, gi0, = f g dxdt kf k0, = f 2 dxdt

space H(div, )
{ (u, v) L2 L2 | k (u, v)k20, < }

remark: (u, v) can be discontinuous, n

(u,v)1
with normal component continuous:
~n ((u, v)2 (u, v)1 ) = 0 (u,v)2

< > -+
University of Colorado at Boulder
Erlangen p.8
 Weak Solutions: Discontinuities
t

 
u=1

x,t f~x,t (u) = 0

u=1 u=0
u=g I

 

u=0 x

(1) Rankine-Hugoniot relations: ~nx,t (f~x,t (u2 ) f~x,t (u1 )) = 0

(2) equivalent: f~x,t (u) H(div, ) (solution regularity)

1
Burgers model flow: f~x,t (u) H(div, ) shock speed s =
2
< > -+
University of Colorado at Boulder
Erlangen p.9
 Numerical Approximation: Finite Differences
derivatives use truncated Taylor series expansion

u  ui ui1
= + O(x)
x  x i

u u uhi,n+1 uhi,n u h
i,n u h
i1,n
Burgers: +u =0 + uhi,n =0
t x t x

convergence to wrong solu-

tion!
1

0.8

reason: Taylor expansion not 0.6

u
valid at shock! 0.4
1
0.2

0 0.5
0
0.2
0.4
0.6
0.8 0 t
1
x

< > -+
University of Colorado at Boulder
Erlangen p.10
 Conservative Finite Difference Schemes
THEOREM. Lax-Wendroff (1960).

u f (u) uhi,n+1 uhi,n fi+1/2,n (uh ) fi1/2,n (uh )

+ =0 + =0
t x t x

theorem: conservative finite difference scheme guarantees convergence

to a correct weak solution (assuming convergence of uh to some u
)

f i1/2 f i+1/2

x
i1 i i+1

conservative form is a sufficient condition for convergence to a

weak solution (but it may not be necessary! . . . )

< > -+
University of Colorado at Boulder
Erlangen p.11
 Why the Name Conservative Scheme?

uhi,n+1 uhi,n fi+1/2,n (uh ) fi1/2,n (uh )

+ =0
t x

~nx,t (f(uh ), uh ) dl = 0
H
i
i
I
recall conservation in space-time ~nx,t f~x,t (u) dl = 0

exact discrete conservation in

every discrete cell i i

exact discrete conservation constrains the solution, s.t. convergence

to a solution with wrong shock speed cannot happen
< > -+
University of Colorado at Boulder
Erlangen p.12
 Lax-Wendroff Scheme
 2 !
1 u
i+1
2  u 2
i t ui + ui+1
fi+1/2 = + (ui+1 ui )
2 2 2 x 2

conservative

1.5
O(x2 ) (Taylor)
1
1
u

0.5 correct shock speed

0
1
0.5
0 t
0.5
1
0
. . . oscillations!
x

< > -+
University of Colorado at Boulder
Erlangen p.13
 Possible Remedy: Numerical Diffusion

add numerical diffusion

u f (u) 2u
+ = num
t x x2

num = O(x2 ), e.g.

problem: need nonlinear limiters
problem: higher-order difficult

this stabilization by numerical diffusion approach is employed in

- upwind schemes
- finite volume schemes
- most existing finite element schemes

< > -+
University of Colorado at Boulder
Erlangen p.14
Alternative: Solution Control through Functional Minimization

minimize the error in a continuous norm

uh = arg min kx,t f~x,t (uh )k20,

uh U h

goal:
- control oscillations
- control convergence to weak solution
- control numerical stability (no need for time step limitation)
- higher-order finite elements

achieve through norm minimization

(remark: h = x)

< > -+
University of Colorado at Boulder
Erlangen p.15
(2) Least-Squares Finite Element (LSFEM) Discretizations

with Luke Olson, Tom Manteuffel, Steve McCormick, Applied Math CU Boulder

finite element method: approximate u U by uh U h

i
n
X
uh (x, t) = ui i (x, t)
i=1

abstract example: solve Lu = 0 (assume L linear PDE operator)

define the functional F(u) = kLuk20,

< > -+
University of Colorado at Boulder
Erlangen p.16
Least-Squares Finite Element (LSFEM) Discretizations

minimization:

uh = arg min kLuh k20, = arg min F(uh )

uh U h

condition for uh stationary point:

F(uh + v h )
|=0 = 0 vh U h

< > -+
University of Colorado at Boulder
Erlangen p.17
 Least-Squares Finite Element Discretizations

algebraic system of linear equations:

n
X
ui hLi , Lj i0, = 0
i=1

(n equations in n unknowns, A u = 0)
(actually, with boundary conditions, A u = f )

Symmetric Positive Definite (SPD) matrices A

< > -+
University of Colorado at Boulder
Erlangen p.18
H(div)-Conforming LSFEM for Hyperbolic Conservation Laws

' $
reformulate conservation law in terms of flux vector w:
~

x,t w
~ =0
x,t f~x,t (u) = 0
~ = f~x,t (u)
w
u=g I
~ = ~nx,t f~x,t (g) I
~nx,t w
u=g I

& %
functional

~ h , uh ; g) =kx,t w
F(w ~ h f~(uh )k20,
~ h k20, + kw

~ h f~(g))k20,I + kuh gk20,I

+ k~nx,t (w

Newton linearization: minimize functional with linearized equation

< > -+
University of Colorado at Boulder
Erlangen p.19
 Finite Element Spaces

weak solution: f~x,t H(div, )

~ h H(div, )
choose w

~ h are
Raviart-Thomas elements: the normal components of w
continuous
~ h H(div, )
w

H(div)-conforming LSFEM

< > -+
University of Colorado at Boulder
Erlangen p.20
 Numerical Results

shock flow: ulef t = 1.0, uright = 0.5, shock speed s = 0.75

convergence to correct weak solution with optimal order
no oscillations, correct shock speed, no CFL limit

1 1 1

0.75 0.75 0.75

0.5 0.5 0.5

t

t
0.25 0.25 0.25

0 0 0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x x x

< > -+
University of Colorado at Boulder
Erlangen p.21
 Linear Advection Higher-Order Elements
linear (k=1, h=1/24) quadratic (k=2, h=1/12)
1 1

.8 .8

.6 .6

.4 .4

.2 .2

0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1

cubic (k=3, 1/8) quartic (k=4, h=1/6)

1 1

.8 .8

.6 .6

.4 .4

.2 .2

0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1

order k = 1, 2, 3, 4: sharper shock for same dof

remark: also discontinuous finite elements for uh
(SIAM J. Sci. Comput., accepted)
< > -+
University of Colorado at Boulder
Erlangen p.22
 Solution-Adaptive Refinement
LS functional is sharp a posteriori error estimator:

F(uh ) = kLuh k20,

= kLuh Luexact k20,
= kL(uh uexact )k20,
= kLeh k20,

0.9

0.8

0.7

0.6
u

0.5 1.4
t

1
0.4 1.2

0.8
1
0.3
u

0.6
0.8
t
0.2
0.4
0.6
0.1
0.2
0.4

0 0 0.1 0.2 0.3 0.4 0.5 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.6 0.7 0.8 0.9 1
x x

< > -+
University of Colorado at Boulder
Erlangen p.23
 Numerical Conservation
we minimize

~ h , uh ; g) =kx,t w
F(w ~ h f~(uh )k20,
~ h k20, + kw

~ h f~(g))k20,I + kuh gk20,I

+ k~nx,t (w

our H(div)-conforming LSFEM does not satisfy the exact discrete

conservation property of Lax and Wendroff

3
x 10

1
div(w)

2
1 1
0.8 0.8
0.6 0.6
t
0.4 0.4 x

~h
0.2 0.2
0 0 w
< > -+
University of Colorado at Boulder
Erlangen p.24
 Numerical Conservation

~ h , uh ; g) =kx,t w
F(w ~ h f~(uh )k20,
~ h k20, + kw

~ h f~(g))k20,I + kuh gk20,I

+ k~nx,t (w

however, we can prove: (submitted to SIAM J. Sci. Comput.)

THEOREM. [Conservation for H(div)-conforming LSFEM]

If finite element approximation uh converges in the L2 sense to u
as
h 0, then u is a weak solution of the conservation law.

exact discrete conservation is not a necessary condition for

numerical conservation!
(can be replaced by minimization in a suitable continuous norm)

< > -+
University of Colorado at Boulder
Erlangen p.25
 Numerical conservation

0.9

0.8

0.7

0.6

0.5
t

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
x

< > -+
University of Colorado at Boulder
Erlangen p.26
 LSFEM for Nonlinear Hyperbolic PDEs: Status

Burgers equation:
- nonlinear
- scalar
- 2D domains

extensions, in progress:
- systems of equations
- higher-dimensional domains

need efficient solvers for A u = f

< > -+
University of Colorado at Boulder
Erlangen p.27
 (3) Fluid Dynamics Applications
(A) Soil Sedimentation (Civil Engineering)
with Gert Bartholomeeusen, Mechanical Engineering, University of Oxford
t=0 t=a t=b







































































































































































































































settling column experiments: soil particles settle
nonlinear waves, modeled by
u f (u)
+ =0
t x

< > -+
University of Colorado at Boulder
Erlangen p.28
 Soil Sedimentation
experimental determination of flux function f (u), nonconvex

u f (u)
+ =0
t x

0.05
Surf. M
Flux, [mm/min] 0.04 Csh2
0.03 Csh3
Csh4 I
0.02 Fit

0.01
0
0.84 0.88 0.92 0.96 1
Porosity, [-]
(kaolinite soil suspension)

< > -+
University of Colorado at Boulder
Erlangen p.29
 Soil Sedimentation

simulation using flux function

observation of compound shock waves = shock + sonic rarefaction
new theory for transition between sedimentation and consolidation
(Proceedings of the 2002 Conference on Hyperbolic Systems)

< > -+
University of Colorado at Boulder
Erlangen p.30
 (B) Bow Shock Flows in Solar-Terrestrial Plasmas

supersonic solar wind plasma induces quasi-steady bow shock in

front of earths magnetosphere
plasma = gas + magnetic field B
described by Magnetohydrodynamics (MHD), hyperbolic system

< > -+
University of Colorado at Boulder
Erlangen p.31
 Recall: Gas Dynamics Bow Shock

< > -+
University of Colorado at Boulder
Erlangen p.32
 Bow Shock Flows in Solar-Terrestrial Plasmas

simulation: reason: MHD has mul-

tiple waves
for large upstream B:
also: compound
multiple shock fronts!
shocks (like in
"rho"
Z sedimentation appli-
2.63673
2.43265
X
cation)
2.22857
2.02449 Y
1.82041
1.61633 (Phys. Rev. Lett. 2000)
1.41224
1.20816
1.00408
0.8 predictive result:
- not observed yet
- confirmed in several
other MHD codes
- new spacecraft may al-
low observation

< > -+
University of Colorado at Boulder
Erlangen p.33
 (C) Supersonic Outflow from Exoplanet Atmospheres
with Feng Tian, Brian Toon, Alex Pavlov, PAOS, CU Boulder

extrasolar planets, as of 13 January 2004

- 104 planetary systems - gas giants (hot Jupiters)

- 119 planets - very close to star ( 0.05 AU)

- 13 multiple planet systems supersonic hydrogen escape

(like the solar wind), Euler
< > -+
University of Colorado at Boulder
Erlangen p.34
 Supersonic Outflow from Exoplanet Atmospheres

< > -+
University of Colorado at Boulder
Erlangen p.35
 Supersonic Outflow from Exoplanet Atmospheres

< > -+
University of Colorado at Boulder
Erlangen p.36
 Supersonic Outflow from Exoplanet Atmospheres

planet around HD209458

- 0.67 Jupiter masses, 0.05 AU
- hydrogen atmosphere and escape observed
(Vidal-Madjar, Nature March 2003)
Fengs simulations show:
- extent and temperature of Hydrogen atmosphere
consistent with observations
- atmosphere is stable (1% mass loss in 10 billion years)
Mercury-type planet with gas atmosphere would lose
10% of mass in 8.5 million years

< > -+
University of Colorado at Boulder
Erlangen p.37
 Collaborators

LSFEM for Hyperbolic PDEs

Luke Olson, Tom Manteuffel, Steve McCormick
Applied Math, CU Boulder

Fluid Dynamics Applications

Gert Bartholomeeusen
Oxford
Feng Tian, Brian Toon, Alex Pavlov
PAOS, CU Boulder

< > -+
University of Colorado at Boulder
Erlangen p.38