58
On Energy Stable dG Approximation of the PML for the Wave Equation Monash Workshop on Numerical Differential Equations and Applications February 2020 Kenneth Duru Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation Monash Workshop on Numerical Differential Equations and Applications1 / 34

On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

On Energy Stable dG Approximation of the PML forthe Wave Equation

Monash Workshop on Numerical Differential Equations and Applications

February 2020

Kenneth Duru

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications1 / 34

Page 2: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Waves are everywhere

Simulations of seismic waves to quantify and assess earthquake risks and hazards.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications2 / 34

Page 3: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Wave propagation:

Forward Modeling

Efficient Time Domain Wave Propagation Tool

Accurate and stable volume discretizations

Efficient and reliable absorbing boundaries

Accurate source generation

Efficient and scalable implementation on modern HPC platforms.

My research has penetrated all components.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications3 / 34

Page 4: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Truncated Domain

Which boundary conditions ensure that numerical simulations converge to the infinite domainproblem? (old but relevant: Engquist and Majda (1977)).

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications4 / 34

Page 5: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Reflections from boundaries

A solution of the acoustic pressure with a point source.

0 2 4 6 8 10t[s]

-0.4

-0.2

0

0.2

0.4p[M

Pa]

∆x = 5/9∆x = 5/27

Analytical

Classical absorbing boundary condition

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications5 / 34

Page 6: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Absorbing Layer

Equations must be perfectly matched: J. P. Berenger (1994).

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications6 / 34

Page 7: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Anechoic chamber

”non-reflective”, ”non-echoing”, ”echo-free”.

A room designed to completely absorb reflections of either sound or electromagnetic waves.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications7 / 34

Page 8: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Complex coordinate stretching: Chew and Weedon (1994)

∂/∂x → 1/Sx∂/∂x , Sx := dx/dx = 1 + dx (x)/s, dx ≥ 0.Simplifies PML construction for hyperbolic systems

−2 −1 0 1 2 3−1

−0.5

0

0.5

1U

PM

L

x

PML

!

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications8 / 34

Page 9: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Acoustic wave equation in first order form

∂p∂t

+∇ · v = 0, ρ∂v∂t

+∇p = 0.

(x , y , z) ∈ Ω = [−1, 1]3,

BCs:1− rη

2Zvη ∓

1 + rη2

p = 0, |rη | ≤ 1, at η = ±1.

dEdxdydz

=12

1κ|p|2 + ρ

∑η=x,y,z

|vη|2 > 0, E(t) =

∫Ω

dE > 0.

ddt

E(t) = −∮∂Ω

p (n · v) dS = −∑

η=x,y,z

∫ 1

−1

∫ 1

−1BT(η) dydzdx

dη≤ 0,

BT(η) =1− |rη |2

Z|χ(−η)|2

∣∣∣η=−1

+1− |rη |2

Z|χ(+η)|2

∣∣∣η=1≥ 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications9 / 34

Page 10: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Acoustic wave equation in first order form

∂p∂t

+∇ · v = 0, ρ∂v∂t

+∇p = 0.

(x , y , z) ∈ Ω = [−1, 1]3,

BCs:1− rη

2Zvη ∓

1 + rη2

p = 0, |rη | ≤ 1, at η = ±1.

dEdxdydz

=12

1κ|p|2 + ρ

∑η=x,y,z

|vη|2 > 0, E(t) =

∫Ω

dE > 0.

ddt

E(t) = −∮∂Ω

p (n · v) dS = −∑

η=x,y,z

∫ 1

−1

∫ 1

−1BT(η) dydzdx

dη≤ 0,

BT(η) =1− |rη |2

Z|χ(−η)|2

∣∣∣η=−1

+1− |rη |2

Z|χ(+η)|2

∣∣∣η=1≥ 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications9 / 34

Page 11: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Acoustic wave equation in first order form

∂p∂t

+∇ · v = 0, ρ∂v∂t

+∇p = 0.

(x , y , z) ∈ Ω = [−1, 1]3,

BCs:1− rη

2Zvη ∓

1 + rη2

p = 0, |rη | ≤ 1, at η = ±1.

dEdxdydz

=12

1κ|p|2 + ρ

∑η=x,y,z

|vη|2 > 0, E(t) =

∫Ω

dE > 0.

ddt

E(t) = −∮∂Ω

p (n · v) dS = −∑

η=x,y,z

∫ 1

−1

∫ 1

−1BT(η) dydzdx

dη≤ 0,

BT(η) =1− |rη |2

Z|χ(−η)|2

∣∣∣η=−1

+1− |rη |2

Z|χ(+η)|2

∣∣∣η=1≥ 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications9 / 34

Page 12: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Derive PML in the Laplace domain

u(x , y , z, s) =

∫ ∞0

e−st u (x , y , z, t) dt, s = a + ib, Res = a > 0,

sp +∇ · v = 0, ρsv +∇p = 0.

PML : ∂/∂η → 1/Sη∂/∂η, Sη = 1 +dη(η)

s, dη(η) ≥ 0, η = x , y , z.

sp +∇d · v = 0, ρsv +∇d p = 0,

∇d = (1/Sx∂/∂x , 1/Sy∂/∂y , 1/Sz∂/∂z)T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications10 / 34

Page 13: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Derive PML in the Laplace domain

u(x , y , z, s) =

∫ ∞0

e−st u (x , y , z, t) dt, s = a + ib, Res = a > 0,

sp +∇ · v = 0, ρsv +∇p = 0.

PML : ∂/∂η → 1/Sη∂/∂η, Sη = 1 +dη(η)

s, dη(η) ≥ 0, η = x , y , z.

sp +∇d · v = 0, ρsv +∇d p = 0,

∇d = (1/Sx∂/∂x , 1/Sy∂/∂y , 1/Sz∂/∂z)T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications10 / 34

Page 14: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Derive PML in the Laplace domain

u(x , y , z, s) =

∫ ∞0

e−st u (x , y , z, t) dt, s = a + ib, Res = a > 0,

sp +∇ · v = 0, ρsv +∇p = 0.

PML : ∂/∂η → 1/Sη∂/∂η, Sη = 1 +dη(η)

s, dη(η) ≥ 0, η = x , y , z.

sp +∇d · v = 0, ρsv +∇d p = 0,

∇d = (1/Sx∂/∂x , 1/Sy∂/∂y , 1/Sz∂/∂z)T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications10 / 34

Page 15: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Derive PML in the Laplace domain

u(x , y , z, s) =

∫ ∞0

e−st u (x , y , z, t) dt, s = a + ib, Res = a > 0,

sp +∇ · v = 0, ρsv +∇p = 0.

PML : ∂/∂η → 1/Sη∂/∂η, Sη = 1 +dη(η)

s, dη(η) ≥ 0, η = x , y , z.

sp +∇d · v = 0, ρsv +∇d p = 0,

∇d = (1/Sx∂/∂x , 1/Sy∂/∂y , 1/Sz∂/∂z)T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications10 / 34

Page 16: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Time-domain PML

Auxiliary variables: sσ =

(dx − dy

Sx

Sy

)∂v∂y, sψ =

(dx − dz

Sx

Sz

)∂w∂z

.

Modified PDE:

(∂p∂t

+dx p)

+∇ · v−σ − ψ = 0, ρ

(∂v∂t

+dv)

+∇p = 0, d =

dx 0 00 dy 00 0 dz

.

Auxiliary differential equation: ODE(∂σ

∂t+ dyσ

)+ (dy − dx )

∂v∂y

= 0,(∂ψ

∂t+ dzψ

)+ (dz − dx )

∂w∂z

= 0,

BCs:1− rη

2Zvη ∓

1 + rη2

p = 0, at η = ±1.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications11 / 34

Page 17: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Time-domain PML

Auxiliary variables: sσ =

(dx − dy

Sx

Sy

)∂v∂y, sψ =

(dx − dz

Sx

Sz

)∂w∂z

.

Modified PDE:

(∂p∂t

+dx p)

+∇ · v−σ − ψ = 0, ρ

(∂v∂t

+dv)

+∇p = 0, d =

dx 0 00 dy 00 0 dz

.

Auxiliary differential equation: ODE(∂σ

∂t+ dyσ

)+ (dy − dx )

∂v∂y

= 0,(∂ψ

∂t+ dzψ

)+ (dz − dx )

∂w∂z

= 0,

BCs:1− rη

2Zvη ∓

1 + rη2

p = 0, at η = ±1.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications11 / 34

Page 18: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

2D: SBP finite difference approximation

0 1000 2000 3000 4000 500010

−3

10−2

10−1

100

time

‖E

z‖h

2nd−order

4th−order

6th−order

(b) Discrete PML

0 1000 2000 3000 4000 500010

−3

10−2

10−1

100

time

‖E

z‖h

2nd−order

4th−order

6th−order

(c) No PML (dx = 0).

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications12 / 34

Page 19: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

A nightmare for finite element and dG practitioners!

-60 -40 -20 0 20 40 60

x[km]

0

10

20

30

40

50

y[k

m]

t=150 s

0

0.02

0.04

0.06

0.08

0.1

GLL.

-60 -40 -20 0 20 40 60

x[km]

0

10

20

30

40

50y[k

m]

t=35 s

0

0.02

0.04

0.06

0.08

0.1

GL.

-60 -40 -20 0 20 40 60

x[km]

0

10

20

30

40

50

y[k

m]

t=10 s

0

0.02

0.04

0.06

0.08

0.1

GLR.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications13 / 34

Page 20: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Some pioneering work: Cauchy problem

Abarbanel and Gottlieb (1997 & 98), Hesthaven et al. (1999),

Collino and Tsogka (2001),

Becache et al. (2003): Geometric stability condition,

Appelo, Hagstrom and Kreiss (2006),

Diaz and Joly (2006), Halpern et al. (2011),

Duru and Kreiss (2012), ..., Duru (2016)

Skelton, et al. (2007): Destabilizing effects of boundary conditions

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications14 / 34

Page 21: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Summary of literature: IBVP

Assume constant coefficients and consider:

1. IVP PML : (x , y) ∈ (−∞,∞)× (−∞,∞). Geometric stability condition Becache et al.2003. Classical Fourier analysis

2. IBVP PML : PML is asymmetric.2a. Lower half–plane PML problem: (x, y) ∈ (−∞,∞)× (−∞, 0), with Ly U = 0, y = 0.2b . Left half–plane PML problem: (x, y) ∈ (−∞, 0)× (−∞,∞), with Lx U = 0, x = 0.

Assumption: In the absence of the PML, dx = 0, the IBVPs are stable in the sense of Kreiss.All eigenvalues are in the stable half of the complex plane.

What will happen when the PML is present, dx > 0?

TheoremIf the geometric stability condition is satisfied then the PML IBVPs with dx ≥ 0 areasymptotically stable. The PML damping dx > 0 will move all eigenvalues further into thestable complex plane.

Duru & Kreiss SINUM (2014), Duru SISC (2016), Duru et al. JCP (2016).

PML is asymptotically stable ! Too technical to be extended to numerical approximations inmultiple dimensions

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications15 / 34

Page 22: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Summary of literature: IBVP

Assume constant coefficients and consider:

1. IVP PML : (x , y) ∈ (−∞,∞)× (−∞,∞). Geometric stability condition Becache et al.2003. Classical Fourier analysis

2. IBVP PML : PML is asymmetric.2a. Lower half–plane PML problem: (x, y) ∈ (−∞,∞)× (−∞, 0), with Ly U = 0, y = 0.2b . Left half–plane PML problem: (x, y) ∈ (−∞, 0)× (−∞,∞), with Lx U = 0, x = 0.

Assumption: In the absence of the PML, dx = 0, the IBVPs are stable in the sense of Kreiss.All eigenvalues are in the stable half of the complex plane.

What will happen when the PML is present, dx > 0?

TheoremIf the geometric stability condition is satisfied then the PML IBVPs with dx ≥ 0 areasymptotically stable. The PML damping dx > 0 will move all eigenvalues further into thestable complex plane.

Duru & Kreiss SINUM (2014), Duru SISC (2016), Duru et al. JCP (2016).

PML is asymptotically stable ! Too technical to be extended to numerical approximations inmultiple dimensions

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications15 / 34

Page 23: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Summary of literature: IBVP

Assume constant coefficients and consider:

1. IVP PML : (x , y) ∈ (−∞,∞)× (−∞,∞). Geometric stability condition Becache et al.2003. Classical Fourier analysis

2. IBVP PML : PML is asymmetric.2a. Lower half–plane PML problem: (x, y) ∈ (−∞,∞)× (−∞, 0), with Ly U = 0, y = 0.2b . Left half–plane PML problem: (x, y) ∈ (−∞, 0)× (−∞,∞), with Lx U = 0, x = 0.

Assumption: In the absence of the PML, dx = 0, the IBVPs are stable in the sense of Kreiss.All eigenvalues are in the stable half of the complex plane.

What will happen when the PML is present, dx > 0?

TheoremIf the geometric stability condition is satisfied then the PML IBVPs with dx ≥ 0 areasymptotically stable. The PML damping dx > 0 will move all eigenvalues further into thestable complex plane.

Duru & Kreiss SINUM (2014), Duru SISC (2016), Duru et al. JCP (2016).

PML is asymptotically stable ! Too technical to be extended to numerical approximations inmultiple dimensions

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications15 / 34

Page 24: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Summary of literature: IBVP

Assume constant coefficients and consider:

1. IVP PML : (x , y) ∈ (−∞,∞)× (−∞,∞). Geometric stability condition Becache et al.2003. Classical Fourier analysis

2. IBVP PML : PML is asymmetric.2a. Lower half–plane PML problem: (x, y) ∈ (−∞,∞)× (−∞, 0), with Ly U = 0, y = 0.2b . Left half–plane PML problem: (x, y) ∈ (−∞, 0)× (−∞,∞), with Lx U = 0, x = 0.

Assumption: In the absence of the PML, dx = 0, the IBVPs are stable in the sense of Kreiss.All eigenvalues are in the stable half of the complex plane.

What will happen when the PML is present, dx > 0?

TheoremIf the geometric stability condition is satisfied then the PML IBVPs with dx ≥ 0 areasymptotically stable. The PML damping dx > 0 will move all eigenvalues further into thestable complex plane.

Duru & Kreiss SINUM (2014), Duru SISC (2016), Duru et al. JCP (2016).

PML is asymptotically stable ! Too technical to be extended to numerical approximations inmultiple dimensions

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications15 / 34

Page 25: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Need to extend theory

1 Energy estimates + useful for designing stable numerical methods.

2 Energy estimate must account for BCs.

3 PML is asymmetric. Energy estimate in the time-domain technically difficult.

4 Energy estimate in the Laplace space.

5 Numerical boundary/inter-element procedures.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications16 / 34

Page 26: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Need to extend theory

1 Energy estimates + useful for designing stable numerical methods.

2 Energy estimate must account for BCs.

3 PML is asymmetric. Energy estimate in the time-domain technically difficult.

4 Energy estimate in the Laplace space.

5 Numerical boundary/inter-element procedures.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications16 / 34

Page 27: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Need to extend theory

1 Energy estimates + useful for designing stable numerical methods.

2 Energy estimate must account for BCs.

3 PML is asymmetric. Energy estimate in the time-domain technically difficult.

4 Energy estimate in the Laplace space.

5 Numerical boundary/inter-element procedures.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications16 / 34

Page 28: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Need to extend theory

1 Energy estimates + useful for designing stable numerical methods.

2 Energy estimate must account for BCs.

3 PML is asymmetric. Energy estimate in the time-domain technically difficult.

4 Energy estimate in the Laplace space.

5 Numerical boundary/inter-element procedures.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications16 / 34

Page 29: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Energy estimate in the Laplace space

Forcing:(fp, f, fσ , fψ

)T with f = (fx , fy , fz ).

Laplace transform + Eliminate PML auxiliary variables

sp +∇d · v =1κ

Fp, ρsv +∇d p = ρf,

fη =1

Sηfη , Fp =

(1

Sxfp −

κ

sSy Sxfσ −

κ

sSzSxfψ

),

BCs:1− rη

2Z vη ∓

1 + rη2

p = 0, at η = ±1.

s2p −∇d ·(

1ρ∇d p

)=

Fp −∇d · f,

BCs:1 + rη

2Zsp ±

1− rη2

1Sη

∂p∂η

= 0, at η = ±1,

s∗

Sηρp∗∂p∂η≥ 0, η = −1;

s∗

Sηρp∗∂p∂η≤ 0, η = 1.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications17 / 34

Page 30: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Energy estimate in the Laplace space

Forcing:(fp, f, fσ , fψ

)T with f = (fx , fy , fz ).

Laplace transform + Eliminate PML auxiliary variables

sp +∇d · v =1κ

Fp, ρsv +∇d p = ρf,

fη =1

Sηfη , Fp =

(1

Sxfp −

κ

sSy Sxfσ −

κ

sSzSxfψ

),

BCs:1− rη

2Z vη ∓

1 + rη2

p = 0, at η = ±1.

s2p −∇d ·(

1ρ∇d p

)=

Fp −∇d · f,

BCs:1 + rη

2Zsp ±

1− rη2

1Sη

∂p∂η

= 0, at η = ±1,

s∗

Sηρp∗∂p∂η≥ 0, η = −1;

s∗

Sηρp∗∂p∂η≤ 0, η = 1.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications17 / 34

Page 31: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Energy estimate in the Laplace space

Forcing:(fp, f, fσ , fψ

)T with f = (fx , fy , fz ).

Laplace transform + Eliminate PML auxiliary variables

sp +∇d · v =1κ

Fp, ρsv +∇d p = ρf,

fη =1

Sηfη , Fp =

(1

Sxfp −

κ

sSy Sxfσ −

κ

sSzSxfψ

),

BCs:1− rη

2Z vη ∓

1 + rη2

p = 0, at η = ±1.

s2p −∇d ·(

1ρ∇d p

)=

Fp −∇d · f,

BCs:1 + rη

2Zsp ±

1− rη2

1Sη

∂p∂η

= 0, at η = ±1,

s∗

Sηρp∗∂p∂η≥ 0, η = −1;

s∗

Sηρp∗∂p∂η≤ 0, η = 1.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications17 / 34

Page 32: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Energy estimate in the Laplace space

Re(

(sSη)∗

)= a + εη(s, dη), εη(s, dη) :=

2dηb2

|sSη |2≥ 0.

Define the energy

E2p (s, dη) =

∥∥∥sp∥∥∥2

1/κ+

∑η=x,y,z

∥∥∥ 1Sη

∂p∂η

∥∥∥2

1/ρ> 0,

E2f (s, dη) =

∥∥∥sFp

∥∥∥2

1/κ+

∑η=x,y,z

∥∥∥ 1Sη

∂ fη∂η

∥∥∥2

κ> 0.

TheoremConsider the PML IBVP with s 6= 0, Res = a > 0 and piecewise constant dη ≥ 0.

aE2p (s, dη) +

∑η=x,y,z

∥∥∥ 1Sη

∂p∂η

∥∥∥2

εη/ρ+ BT (s, dη) ≤ 2Ep (s, dη) Ef (s, dη),

BT (s, dη) =

∫ 1

−1

∫ 1

−1

∑η=x,y,z

(s∗

Sηρp∗∂p∂η

∣∣∣η=−1

−s∗

Sηρp∗∂p∂η

∣∣∣η=1

)dxdydz

dη≥ 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications18 / 34

Page 33: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Well-posedness & Stability

TheoremThe problem, with dη ≥ 0, is asymptotically stable in the sense that no exponentially growingsolutions are supported.

Proof:Make the ansatz Q(x , y , z, t) = est Q(x , y , z), with s = a + ib,

aE2p (s, dη) +

∑η=x,y,z

∥∥∥ 1Sη

∂p∂η

∥∥∥2

εη/ρ+ BT (s, dη) = 0.

With Res = a > 0 the expression in the left hand side is positive. The conclusion is thatQ(x , y , z) ≡ 0.

TheoremLet the energy norms E2

p (t , dη) > 0, E2f (t , dη) > 0. For any a > 0 and T > 0

∫ T

0e−2at E2

p (t , dη) dt ≤4a2

∫ T

0e−2at E2

f (t , dη) dt , a > 0.

Key: Construct a discrete approximation that, as far as possible, mimics the energy estimate.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications19 / 34

Page 34: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Well-posedness & Stability

TheoremThe problem, with dη ≥ 0, is asymptotically stable in the sense that no exponentially growingsolutions are supported.

Proof:Make the ansatz Q(x , y , z, t) = est Q(x , y , z), with s = a + ib,

aE2p (s, dη) +

∑η=x,y,z

∥∥∥ 1Sη

∂p∂η

∥∥∥2

εη/ρ+ BT (s, dη) = 0.

With Res = a > 0 the expression in the left hand side is positive. The conclusion is thatQ(x , y , z) ≡ 0.

TheoremLet the energy norms E2

p (t , dη) > 0, E2f (t , dη) > 0. For any a > 0 and T > 0

∫ T

0e−2at E2

p (t , dη) dt ≤4a2

∫ T

0e−2at E2

f (t , dη) dt , a > 0.

Key: Construct a discrete approximation that, as far as possible, mimics the energy estimate.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications19 / 34

Page 35: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Dicretize domain

Discretize the domain (x , y , z) ∈ Ω = ∪Ωlmn with Ωlmn = [xl , xl+1]× [ym, ym+1]× [zn, zn+1],

(x , y , z)←→ (q, r , s) ∈ Ω = [−1, 1]3:

x = xl +∆xl

2(1 + q) , y = ym +

∆ym

2(1 + r) , z = zn +

∆zn

2(1 + s) ,

with

J =∆xl

2∆ym

2∆zn

2> 0, ∆xl = xl+1 − xl , ∆ym = ym+1 − ym, ∆zn = yn+1 − yn.

∫Ω

f (x , y , z)dxdydz =K∑

k=1

L∑l=1

M∑m=1

∫Ωklm

f (x , y , z)dxdydz =K∑

k=1

L∑l=1

M∑m=1

∫Ω

f (q, r , s)Jdqdrds.

Physics based flux fluctuations:

F x :=Z2

(vx − vx

)+

12

(p − p

)= 0, x = xl ,

Gx :=Z2

(vx − vx

)−

12

(p − p

)= 0, x = xl+1.

(1)

vx , p encode the solution of the IBVP at element boundaries.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications20 / 34

Page 36: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Dicretize domain

Discretize the domain (x , y , z) ∈ Ω = ∪Ωlmn with Ωlmn = [xl , xl+1]× [ym, ym+1]× [zn, zn+1],

(x , y , z)←→ (q, r , s) ∈ Ω = [−1, 1]3:

x = xl +∆xl

2(1 + q) , y = ym +

∆ym

2(1 + r) , z = zn +

∆zn

2(1 + s) ,

with

J =∆xl

2∆ym

2∆zn

2> 0, ∆xl = xl+1 − xl , ∆ym = ym+1 − ym, ∆zn = yn+1 − yn.

∫Ω

f (x , y , z)dxdydz =K∑

k=1

L∑l=1

M∑m=1

∫Ωklm

f (x , y , z)dxdydz =K∑

k=1

L∑l=1

M∑m=1

∫Ω

f (q, r , s)Jdqdrds.

Physics based flux fluctuations:

F x :=Z2

(vx − vx

)+

12

(p − p

)= 0, x = xl ,

Gx :=Z2

(vx − vx

)−

12

(p − p

)= 0, x = xl+1.

(1)

vx , p encode the solution of the IBVP at element boundaries.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications20 / 34

Page 37: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Integral form

∫Ω

φp

(1κ

(∂p∂t

+ dx p)

+∇ · v− σ − ψ)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φp

ZFη|η=−1 −

φp

ZGη|η=1

)dqdrds

dη,

∫Ω

θT(ρ

(∂v∂t

+ dv)

+∇p)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

T nFη|η=−1 + θT nGη|η=1

) dqdrdsdη

,

∫Ω

φσ

(∂σ

∂t+ dyσ + (dy − dx )

2∆y

∂vy

∂r

)Jdqdrds

= −ωy (dy − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φσ

Zny Fη|η=−1 −

φσ

Zny Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

,

∫Ω

φψ

(∂ψ

∂t+ dzψ + (dz − dx )

2∆z

∂vz

∂s

)Jdqdrds

= −ωz (dz − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φψ

Znz Fη|η=−1 −

φψ

Znz Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications21 / 34

Page 38: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Integral form

∫Ω

φp

(1κ

(∂p∂t

+ dx p)

+∇ · v− σ − ψ)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φp

ZFη|η=−1 −

φp

ZGη|η=1

)dqdrds

dη,

∫Ω

θT(ρ

(∂v∂t

+ dv)

+∇p)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

T nFη|η=−1 + θT nGη|η=1

) dqdrdsdη

,

∫Ω

φσ

(∂σ

∂t+ dyσ + (dy − dx )

2∆y

∂vy

∂r

)Jdqdrds

= −ωy (dy − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φσ

Zny Fη|η=−1 −

φσ

Zny Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

,

∫Ω

φψ

(∂ψ

∂t+ dzψ + (dz − dx )

2∆z

∂vz

∂s

)Jdqdrds

= −ωz (dz − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φψ

Znz Fη|η=−1 −

φψ

Znz Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications21 / 34

Page 39: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Integral form

∫Ω

φp

(1κ

(∂p∂t

+ dx p)

+∇ · v− σ − ψ)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φp

ZFη|η=−1 −

φp

ZGη|η=1

)dqdrds

dη,

∫Ω

θT(ρ

(∂v∂t

+ dv)

+∇p)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

T nFη|η=−1 + θT nGη|η=1

) dqdrdsdη

,

∫Ω

φσ

(∂σ

∂t+ dyσ + (dy − dx )

2∆y

∂vy

∂r

)Jdqdrds

= −ωy (dy − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φσ

Zny Fη|η=−1 −

φσ

Zny Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

,

∫Ω

φψ

(∂ψ

∂t+ dzψ + (dz − dx )

2∆z

∂vz

∂s

)Jdqdrds

= −ωz (dz − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φψ

Znz Fη|η=−1 −

φψ

Znz Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications21 / 34

Page 40: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Integral form

∫Ω

φp

(1κ

(∂p∂t

+ dx p)

+∇ · v− σ − ψ)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φp

ZFη|η=−1 −

φp

ZGη|η=1

)dqdrds

dη,

∫Ω

θT(ρ

(∂v∂t

+ dv)

+∇p)

Jdqdrds

= −∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

T nFη|η=−1 + θT nGη|η=1

) dqdrdsdη

,

∫Ω

φσ

(∂σ

∂t+ dyσ + (dy − dx )

2∆y

∂vy

∂r

)Jdqdrds

= −ωy (dy − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φσ

Zny Fη|η=−1 −

φσ

Zny Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

,

∫Ω

φψ

(∂ψ

∂t+ dzψ + (dz − dx )

2∆z

∂vz

∂s

)Jdqdrds

= −ωz (dz − dx )∑η=q,r,s

∫ 1

−1

∫ 1

−1J√η2

x + η2y + η2

z

(φψ

Znz Fη|η=−1 −

φψ

Znz Gη|η=1

)dqdrds

dη︸ ︷︷ ︸PML stabilizing flux fluctuation

.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications21 / 34

Page 41: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

The dG Approximation

Galerkin approximation + Polynomial interpolants + GL, GLL, GLR, Collocation nodes:

κ−1(

dp(t)dt

+ dx p(t))

+∇D · v(t) + σ(t) + ψ(t) = −∑

η=x,y,z

H−1η

(eη(−1)

ZFη −

eη(1)

ZGη),

ρ

(dv(t)

dt+ dv(t)

)+∇Dp(t) = −

∑η=x,y,z

H−1η

(eη(−1)

ZnFη −

eη(1)

ZnGη

),

(dσ(t)

dt+ dyσ(t)

)+ (dy − dx )Dy vy (t) = −ωy (dy − dx )

∑η=x,y,z

H−1η

(eη(−1)

Zny Fη −

eη(1)

Zny Gη

)︸ ︷︷ ︸

PML stabilizing flux fluctuation

,

(dψ(t)

dt+ dyψ(t)

)+ (dz − dx )Dz vz (t) = −ωz (dz − dx )

∑η=x,y,z

H−1η

(eη(−1)

Znz Fη −

eη(1)

Znz Gη

)︸ ︷︷ ︸

PML stabilizing flux fluctuation

.

∇D = (Dx ,Dy ,Dz )T, Dx =

2∆x

(D ⊗ I ⊗ I) , Hx =∆x2

(H ⊗ I ⊗ I) , ex (η) = (e(η)⊗ I ⊗ I) ,

D = H−1A ≈∂

∂q, H = diag[h1, h2, · · · , hP+1], Aij =

P+1∑m=1

hmLi (qm)L ′j (qm) =

∫ 1

−1Li (q)L ′j (q)dq,

e(η) = [Li (η),Li (η), · · · ,LP+1(η)]T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications22 / 34

Page 42: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

The dG Approximation

Galerkin approximation + Polynomial interpolants + GL, GLL, GLR, Collocation nodes:

κ−1(

dp(t)dt

+ dx p(t))

+∇D · v(t) + σ(t) + ψ(t) = −∑

η=x,y,z

H−1η

(eη(−1)

ZFη −

eη(1)

ZGη),

ρ

(dv(t)

dt+ dv(t)

)+∇Dp(t) = −

∑η=x,y,z

H−1η

(eη(−1)

ZnFη −

eη(1)

ZnGη

),

(dσ(t)

dt+ dyσ(t)

)+ (dy − dx )Dy vy (t) = −ωy (dy − dx )

∑η=x,y,z

H−1η

(eη(−1)

Zny Fη −

eη(1)

Zny Gη

)︸ ︷︷ ︸

PML stabilizing flux fluctuation

,

(dψ(t)

dt+ dyψ(t)

)+ (dz − dx )Dz vz (t) = −ωz (dz − dx )

∑η=x,y,z

H−1η

(eη(−1)

Znz Fη −

eη(1)

Znz Gη

)︸ ︷︷ ︸

PML stabilizing flux fluctuation

.

∇D = (Dx ,Dy ,Dz )T, Dx =

2∆x

(D ⊗ I ⊗ I) , Hx =∆x2

(H ⊗ I ⊗ I) , ex (η) = (e(η)⊗ I ⊗ I) ,

D = H−1A ≈∂

∂q, H = diag[h1, h2, · · · , hP+1], Aij =

P+1∑m=1

hmLi (qm)L ′j (qm) =

∫ 1

−1Li (q)L ′j (q)dq,

e(η) = [Li (η),Li (η), · · · ,LP+1(η)]T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications22 / 34

Page 43: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

The dG Approximation

Galerkin approximation + Polynomial interpolants + GL, GLL, GLR, Collocation nodes:

κ−1(

dp(t)dt

+ dx p(t))

+∇D · v(t) + σ(t) + ψ(t) = −∑

η=x,y,z

H−1η

(eη(−1)

ZFη −

eη(1)

ZGη),

ρ

(dv(t)

dt+ dv(t)

)+∇Dp(t) = −

∑η=x,y,z

H−1η

(eη(−1)

ZnFη −

eη(1)

ZnGη

),

(dσ(t)

dt+ dyσ(t)

)+ (dy − dx )Dy vy (t) = −ωy (dy − dx )

∑η=x,y,z

H−1η

(eη(−1)

Zny Fη −

eη(1)

Zny Gη

)︸ ︷︷ ︸

PML stabilizing flux fluctuation

,

(dψ(t)

dt+ dyψ(t)

)+ (dz − dx )Dz vz (t) = −ωz (dz − dx )

∑η=x,y,z

H−1η

(eη(−1)

Znz Fη −

eη(1)

Znz Gη

)︸ ︷︷ ︸

PML stabilizing flux fluctuation

.

∇D = (Dx ,Dy ,Dz )T, Dx =

2∆x

(D ⊗ I ⊗ I) , Hx =∆x2

(H ⊗ I ⊗ I) , ex (η) = (e(η)⊗ I ⊗ I) ,

D = H−1A ≈∂

∂q, H = diag[h1, h2, · · · , hP+1], Aij =

P+1∑m=1

hmLi (qm)L ′j (qm) =

∫ 1

−1Li (q)L ′j (q)dq,

e(η) = [Li (η),Li (η), · · · ,LP+1(η)]T .

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications22 / 34

Page 44: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Stability of the discrete undamped problem

Introduce the elemental energy density

dE (q, r , s, t) =12

1κ(q, r , s)

|p(q, r , s, t)|2 + ρ(q, r , s)∑

η=x,y,z

(|vη(q, r , s, t)|2

)and the corresponding semi-discrete energy

E (t) =∑l=1

∑m=1

∑n=1

P+1∑i=1

P+1∑j=1

P+1∑k=1

dE (qi , rj , sk , t)Jhi hj hk , J =∆xl

2∆ym

2∆zn

2.

TheoremConsider the semi-discrete approximation. When all the damping vanish, dη = 0, the solutionof the semi-discrete approximation satisfies the energy identity

d

dtE (t) = −

L∑l=1

M∑m=1

N∑n=1

∑η=q,r,s

N+1∑i=1

N+1∑j=1

(J√η2

x + η2y + η2

z

( 1

Z|Fη|2 +

1

Z|Gη|2 + BT(η)

))ij

hi hj

lmn

≤ 0.

The numerical approximation is asymptotically stable!

Result does not translate to the PML when dη > 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications23 / 34

Page 45: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Stability of the discrete undamped problem

Introduce the elemental energy density

dE (q, r , s, t) =12

1κ(q, r , s)

|p(q, r , s, t)|2 + ρ(q, r , s)∑

η=x,y,z

(|vη(q, r , s, t)|2

)and the corresponding semi-discrete energy

E (t) =∑l=1

∑m=1

∑n=1

P+1∑i=1

P+1∑j=1

P+1∑k=1

dE (qi , rj , sk , t)Jhi hj hk , J =∆xl

2∆ym

2∆zn

2.

TheoremConsider the semi-discrete approximation. When all the damping vanish, dη = 0, the solutionof the semi-discrete approximation satisfies the energy identity

d

dtE (t) = −

L∑l=1

M∑m=1

N∑n=1

∑η=q,r,s

N+1∑i=1

N+1∑j=1

(J√η2

x + η2y + η2

z

( 1

Z|Fη|2 +

1

Z|Gη|2 + BT(η)

))ij

hi hj

lmn

≤ 0.

The numerical approximation is asymptotically stable!

Result does not translate to the PML when dη > 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications23 / 34

Page 46: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Stability of the discrete PML problem

Laplace transform + Eliminating the PML auxiliary variables:

sp + ∇D · v =1κ

fp −∑

η=x,y,z

1Sη

H−1η

(eη(−1)

ZFη −

eη(1)

ZGη)

+∑η=y,z

(1− ωη) (dη − dx )

sSηSxH−1η

(eη(−1)

ZFη −

eη(1)

ZGη),

ρsv + ∇D p = ρf−∑

η=x,y,z

1Sη

H−1η

(eη(−1)

ZnFη −

eη(1)

ZnGη

), ∇D = (1/Sx Dx , 1/Sy Dy , 1/Sz Dz )T

.

Consider a single element and introduce the modified discrete operators

Dη =1

(Dη +

1 + rη2

H−1η (Bη (−1,−1)− Bη (1, 1))

),

Hη = H(

I +(1− rη)c

2sSηH−1η (Bη(−1,−1) + Bη(1, 1))

)−1

.

Eliminate the velocity fields:

s∗Hsκ−1sp +∑η

(1

SηDη)†( (s∗S∗η)

ρSηHη

)(1

SηDη)

p

+ |s|2∑η

1 + rη2ZSη

HHη−1 (Bη(−1,−1) + Bη(1, 1)) p = s∗Hκ−1sFp −

∑η

s∗H(

1Sη

D0η

)fη.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications24 / 34

Page 47: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Stability of the discrete PML problem

Laplace transform + Eliminating the PML auxiliary variables:

sp + ∇D · v =1κ

fp −∑

η=x,y,z

1Sη

H−1η

(eη(−1)

ZFη −

eη(1)

ZGη)

+∑η=y,z

(1− ωη) (dη − dx )

sSηSxH−1η

(eη(−1)

ZFη −

eη(1)

ZGη),

ρsv + ∇D p = ρf−∑

η=x,y,z

1Sη

H−1η

(eη(−1)

ZnFη −

eη(1)

ZnGη

), ∇D = (1/Sx Dx , 1/Sy Dy , 1/Sz Dz )T

.

Consider a single element and introduce the modified discrete operators

Dη =1

(Dη +

1 + rη2

H−1η (Bη (−1,−1)− Bη (1, 1))

),

Hη = H(

I +(1− rη)c

2sSηH−1η (Bη(−1,−1) + Bη(1, 1))

)−1

.

Eliminate the velocity fields:

s∗Hsκ−1sp +∑η

(1

SηDη)†( (s∗S∗η)

ρSηHη

)(1

SηDη)

p

+ |s|2∑η

1 + rη2ZSη

HHη−1 (Bη(−1,−1) + Bη(1, 1)) p = s∗Hκ−1sFp −

∑η

s∗H(

1Sη

D0η

)fη.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications24 / 34

Page 48: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Stability of the discrete PML problem

Laplace transform + Eliminating the PML auxiliary variables:

sp + ∇D · v =1κ

fp −∑

η=x,y,z

1Sη

H−1η

(eη(−1)

ZFη −

eη(1)

ZGη)

+∑η=y,z

(1− ωη) (dη − dx )

sSηSxH−1η

(eη(−1)

ZFη −

eη(1)

ZGη),

ρsv + ∇D p = ρf−∑

η=x,y,z

1Sη

H−1η

(eη(−1)

ZnFη −

eη(1)

ZnGη

), ∇D = (1/Sx Dx , 1/Sy Dy , 1/Sz Dz )T

.

Consider a single element and introduce the modified discrete operators

Dη =1

(Dη +

1 + rη2

H−1η (Bη (−1,−1)− Bη (1, 1))

),

Hη = H(

I +(1− rη)c

2sSηH−1η (Bη(−1,−1) + Bη(1, 1))

)−1

.

Eliminate the velocity fields:

s∗Hsκ−1sp +∑η

(1

SηDη)†( (s∗S∗η)

ρSηHη

)(1

SηDη)

p

+ |s|2∑η

1 + rη2ZSη

HHη−1 (Bη(−1,−1) + Bη(1, 1)) p = s∗Hκ−1sFp −

∑η

s∗H(

1Sη

D0η

)fη.Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications24 / 34

Page 49: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Discrete energy estimate in the Laplace space

Introduce the discrete scalar product ⟨u, v⟩

H= v†Hu, (2)

and the energy norms

E 2p (s) =

⟨sp, sp

⟩H/κ

+∑

η=x,y,z

⟨ 1Sη

Dηp,1

SηDηp

⟩Hη/ρ

> 0, (3)

E 2f (s) =

⟨sFp, sFp

⟩H/κ

+∑

η=x,y,z

⟨ 1Sη

D0ηfη , 1

SηD0ηfη⟩

κH> 0. (4)

TheoremConsider the one element dG approximation of the PML in the Laplace space withRes = a > 0 and constant damping dη ≥ 0. If ωη = 1, then we have

aE 2p (s) +

∑η=x,y,z

⟨ 1Sη

Dη p,1

SηDη p

⟩εη Hη/ρ

+∑η

1− rηρ

BT(η)num + BT(s) ≤ 2Ep(s)Ef (s),

BT(s) = |s|2∑η

Re(

1Sη

)1 + rη

2Zp†[HH−1

η (Bη(−1,−1) + Bη(1, 1))]

p ≥ 0.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications25 / 34

Page 50: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

2D: Stable approximations

-60 -40 -20 0 20 40 60

x[km]

0

10

20

30

40

50

y[k

m]

t=500 s

0

1

2

3×10 -5

GLL.

-60 -40 -20 0 20 40 60

x[km]

0

10

20

30

40

50y[k

m]

t=500 s

0

1

2

3×10 -5

GL.

-60 -40 -20 0 20 40 60

x[km]

0

10

20

30

40

50

y[k

m]

t=500 s

0

1

2

3×10 -5

GLR.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications26 / 34

Page 51: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

2D: Stable approximations

0 100 200 300 400 500

t[s]

10-4

10-3

10-2

10-1

100

101

102

103

104

105

L∞-norm

ω = 0

ω = 1

GLL.

0 100 200 300 400 500

t[s]

10-4

10-3

10-2

10-1

100

101

102

103

104

105

L∞-norm

ω = 0

ω = 1

GL.

0 100 200 300 400 500

t[s]

10-4

10-3

10-2

10-1

100

101

102

103

104

105

L∞-norm

ω = 0

ω = 1

GLR.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications27 / 34

Page 52: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

h- and p-convergence

2 4 6 8 10

∆x

10-10

10-8

10-6

10-4

10-2

error

GLL

GL

GLR

C∆x5

h-convergence.

2 4 6 8

polynomial degree

10-10

10-8

10-6

10-4

10-2

error

GLL

GL

GLR

p-convergence.

dx (x) =

0 if |x | ≤ 50 km,

d0

(|x|−50δ

)3if |x | ≥ 50 km,

(5)

d0 =4c2δ

ln1tol

, tol = C0

[1δ

∆xP + 1

]P+1. (6)

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications28 / 34

Page 53: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Stable PML boundaries

1

2

3

4

X Axis

4Y Axis

4

Y Axis

1

2

3

4

Z Axis

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Z Axis

1 2 3 4

X Axis

t = 1.7 s

1

2

3

4

X Axis

4Y Axis

4

Y Axis

1

2

3

4

Z Axis

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Z Axis

1 2 3 4

X Axis

t = 2.5 s

1

2

3

4

X Axis

4Y Axis

4

Y Axis

1

2

3

4

Z Axis

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Z Axis

1 2 3 4

X Axis

t = 3.0 s

0 2 4 6 8 10t[s]

-0.4

-0.2

0

0.2

0.4

p[M

Pa]

∆x = 5/9∆x = 5/27

Analytical

Absorbing boundary condition

0 2 4 6 8 10t[s]

-0.4

-0.2

0

0.2

0.4

p[M

Pa]

∆x = 5/9∆x = 5/27

Analytical

PML

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications29 / 34

Page 54: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Extensions to linear elasticity

0 1 2 3

time [s]

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

velo

city [m

/s]

PML

ABC

analytical

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications30 / 34

Page 55: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

LOH1: A Seismology benchmark problem

time [s]15

10

5

0

5

10

15

v x [m

/s]

analyticalABCPML

15

10

5

0

5

10

15

v y [m

/s]

0 1 2 3 4 5 6 7 8 9time [s]

7.5

5.0

2.5

0.0

2.5

5.0

7.5

v z [m

/s]

Receiver 4

time [s]

4

2

0

2

4

v x [m

/s]

analyticalABCPML

4

2

0

2

4

v y [m

/s]

0 1 2 3 4 5 6 7 8 9time [s]

2

1

0

1

2

v z [m

/s]

Receiver 5

time [s]

0.5

0.0

0.5

v x [m

/s]

analyticalABCPML

0.5

0.0

0.5v y

[m/s

]

0 1 2 3 4 5 6 7 8 9time [s]

1.0

0.5

0.0

0.5

1.0

v z [m

/s]

Receiver 6

time [s]1.5

1.0

0.5

0.0

0.5

1.0

1.5

v x [m

/s]

analyticalABCPML

1.5

1.0

0.5

0.0

0.5

1.0

1.5

v y [m

/s]

0 1 2 3 4 5 6 7 8 9time [s]

1.0

0.5

0.0

0.5

1.0

v z [m

/s]

Receiver 9Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications31 / 34

Page 56: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

ExaHyPE: Exa-scale Hyperbolic PDE simulation Engine

ExascaleSpacetreeADER-DG

TUM TRE

DUR

Seismic

Astro

Software for next generation super computers (1018 flops/sec),Exascale HPC scalability,Energy efficiency, etc.

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications32 / 34

Page 57: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Summary

PML for acoustics equation is well-posed and asymptotically stable.

Numerical flux procedures can introduce instability.

Stable numerical flux procedures can be designed by mimicking continuous energyestimate.

Ideas has been extended to linear elasticity.

Initial ideas (2D SBP FDM): K. Duru SIAM J. Sci. Comput., 38(2016), A1171-A1194.

K. Duru, A.-A. Gabriel and G. Kreiss, Computer Methods in Applied Mechanics andEngineering, 350(2019), 898–937.

K. Duru, et al., Extensions to linear elastodynamics, Submitted to Numerische Mathematik(2019).

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications33 / 34

Page 58: On Energy Stable dG Approximation of the PML for the Wave ...users.monash.edu/~jdroniou/MWNDEA/slides/slides-duru.pdf · -2 -1 0 1 2 3-1-0.5 0 0.5 1 U PML x PML! Kenneth Duru: On

Support & Acknowledgements

Prof. Gunilla Kreiss, Uppsala University Sweden.

Prof. Eric M. Dunham, Stanford University, CA.

Dr. Alice-Agnes Gabriel, LMU Munich.

ExaHyPE team:

Heinz-Otto Kreiss

Kenneth Duru: On Energy Stable dG Approximation of the PML for the Wave Equation— Monash Workshop on Numerical Differential Equations and Applications34 / 34