efficient space and time discretization of the wave equation ......e cient space and time...
TRANSCRIPT
Efficient Space and Time Discretization of theWave Equation. Application to the Reverse Time
Migration
J. Diaz1,2
In collaboration with C. Baldassari1,2, H. Barucq1,2,H. Calandra3, B. Denel3
(1) INRIA Bordeaux Research Center, Project Team Magique3D(2) LMA, CNRS UMR 5142, Universite de Pau(3) TOTAL
08/06/2010
Outline
1 What is the Reverse Time Migration?
2 Appropriate space discretization of the wave equation
3 Improvement of the time discretization of the wave equation
The seismic reflection
Oil exploration by seismic reflection :
Why? Have a good image of the sub-surface andconsequently know where there is oil or not.
How? Seismic acquisition campaigns.
Seismic acquisition : principle
Seismic Imaging (summary)
1 Get an initial velocity model from acquisition campain:
Tomography step
2 Using same sources and receivers as for tomography
Solve the waveequation twice foreach source
(a) Propagate the sources into the velocitymodel
(b) Retropropagate the recorded waves(reverse time)
3 Cross-correlation of (a) and (b): imaging condition(Claerbout)
Produce an image of ground
4 Compare with velocity model: if different, go to 2 aftermodification of the velocity model.
Seismic Imaging (summary)
1 Get an initial velocity model from acquisition campain:
Tomography step
2 Using same sources and receivers as for tomography
Solve the waveequation twice foreach source
(a) Propagate the sources into the velocitymodel
(b) Retropropagate the recorded waves(reverse time)
3 Cross-correlation of (a) and (b): imaging condition(Claerbout)
Produce an image of ground
4 Compare with velocity model: if different, go to 2 aftermodification of the velocity model.
Seismic Imaging (example)
Seismogram from theacquisition campaign
Seismic Imaging (example)
Seismogram from theacquisition campaign
=⇒
Initial guess
Seismic Imaging (example)
Computational Domain
Seismic Imaging (example)
Computational Domain
=⇒
RTM Image
Numerical methods for solving the wave equation
∂2u
∂t2− c2∆u = f (t, x), (t, x) ∈ [0 ; T ]× Ω
• Finite difference methods: fast but not really effective inheterogeneous medium.
• Finite element methods with the choice criteria:
• Efficient to take the topography effects into account;
• Optimized computational burden;
• Limitation of the dispersion effects.
Numerical methods for solving the wave equation
M∂2U
∂t2+ KU = F
• Finite difference methods: fast but not really effective inheterogeneous medium.
• Finite element methods with the choice criteria:
• Efficient to take the topography effects into account;
• Optimized computational burden;
• Limitation of the dispersion effects.
Numerical methods for solving the wave equation
Most popular Finite Element Method (FEM) for geophysicalapplications:
Spectral Element Method (SEM) (Seriani, Priolo, Cohen, Komatitsch...)
• Explicit scheme (diagonal mass matrix)
• In general, FEM =⇒ implicit scheme;
• Combined with a lumping process: the order of convergence isdestroyed except for order one or SEM.
Numerical methods for solving the full-wave equation
Discontinuous Galerkin Method (DG)
• Quasi-explicit scheme: mass matrix is block-diagonal;
• Based on tetrahedras (easy to handle);
• Adapted to strongly varying velocities;
• Fast computational methods:
• Volume calculi are made locally;
• Communications between elements thanks to conditions on theedges only ((N − 1)D calculi);
• High-order elements to overcome dispersion effects.
SEM versus DG
1 The SEM method• Gauss Lobatto quadrature rule : diagonal mass matrix, do not
hamper the order of convergence.• Meshes made of quadrangles in 2D or hexahedra in 3D :
difficult to compute, not always suitable to complextopographies.
2 The DG method• Representation of the solution is quasi-explicit because the
mass matrix is block-diagonal without any approximation.• To compute easily its coefficients, we use an exact quadrature
formula which does not hamper the order of convergence.• Meshes made of triangles in 2D or tetrahedra in 3D. Thus the
topography of the computational domain is easily discretized.• Handle polynomial velocities inside each element.
Interior Penalty Discontinuous Galerkin Method (IPDG)
• V hl :=
v ∈ L2 (Ω) : v|K ∈ Pl (K ) ∀K ∈ Th
• M is block-diagonal
• Kij =∑K∈Th
∫K
1
ρ∇v i · ∇v j dx −
∑F∈F
∫F
[[ v j ]] · 1
ρ∇v i dF
-∑F∈F
∫F
[[ v i ]] · 1
ρ∇v j dF +
∑F∈F
∫F
γ[[ v i ]] · [[ v j ]] dF
(Ainswoth et al., Grote et al.)
Interior Penalty Discontinuous Galerkin Method (IPDG)
• V hl :=
v ∈ L2 (Ω) : v|K ∈ Pl (K ) ∀K ∈ Th
• M is block-diagonal
• Kij =∑K∈Th
∫K
1
ρ∇v i · ∇v j dx −
∑F∈F
∫F
[[ v j ]] · 1
ρ∇v i dF
-∑F∈F
∫F
[[ v i ]] · 1
ρ∇v j dF
+∑F∈F
∫F
γ[[ v i ]] · [[ v j ]] dF
(Ainswoth et al., Grote et al.)
Interior Penalty Discontinuous Galerkin Method (IPDG)
• V hl :=
v ∈ L2 (Ω) : v|K ∈ Pl (K ) ∀K ∈ Th
• M is block-diagonal
• Kij =∑K∈Th
∫K
1
ρ∇v i · ∇v j dx −
∑F∈F
∫F
[[ v j ]] · 1
ρ∇v i dF
-∑F∈F
∫F
[[ v i ]] · 1
ρ∇v j dF +
∑F∈F
∫F
γ[[ v i ]] · [[ v j ]] dF
(Ainswoth et al., Grote et al.)
Interior Penalty Discontinuous Galerkin Method (IPDG)
• V hl :=
v ∈ L2 (Ω) : v|K ∈ Pl (K ) ∀K ∈ Th
• M is block-diagonal
• Kij =∑K∈Th
∫K
1
ρ∇v i · ∇v j dx −
∑F∈F
∫F
[[ v j ]] · 1
ρ∇v i dF
-∑F∈F
∫F
[[ v i ]] · 1
ρ∇v j dF +
∑F∈F
∫F
γ[[ v i ]] · [[ v j ]] dF
(Ainswoth et al., Grote et al.)
“Drawbacks” of the DG methods
• The solution to the wave equation is not discontinuous.
The solution is neither piecewise polynomial. Actually, thediscontinuities of the approximate solution are controlled bythe penalization parameter and they do not hamper theaccuracy of the method.
• DG requires more degrees of freedom than SEM.
That is true for a given mesh. But there is no reason to usethe same mesh for both methods.
• The stability condition (which restricts the time space of thescheme) is smaller for DG than for SEM.
That is true for a given mesh.
“Drawbacks” of the DG methods
• The solution to the wave equation is not discontinuous.
The solution is neither piecewise polynomial. Actually, thediscontinuities of the approximate solution are controlled bythe penalization parameter and they do not hamper theaccuracy of the method.
• DG requires more degrees of freedom than SEM.
That is true for a given mesh. But there is no reason to usethe same mesh for both methods.
• The stability condition (which restricts the time space of thescheme) is smaller for DG than for SEM.
That is true for a given mesh.
“Drawbacks” of the DG methods
• The solution to the wave equation is not discontinuous.
The solution is neither piecewise polynomial. Actually, thediscontinuities of the approximate solution are controlled bythe penalization parameter and they do not hamper theaccuracy of the method.
• DG requires more degrees of freedom than SEM.
That is true for a given mesh. But there is no reason to usethe same mesh for both methods.
• The stability condition (which restricts the time space of thescheme) is smaller for DG than for SEM.
That is true for a given mesh.
“Drawbacks” of the DG methods
• The solution to the wave equation is not discontinuous.
The solution is neither piecewise polynomial. Actually, thediscontinuities of the approximate solution are controlled bythe penalization parameter and they do not hamper theaccuracy of the method.
• DG requires more degrees of freedom than SEM.
That is true for a given mesh. But there is no reason to usethe same mesh for both methods.
• The stability condition (which restricts the time space of thescheme) is smaller for DG than for SEM.
That is true for a given mesh.
“Drawbacks” of the DG methods
• The solution to the wave equation is not discontinuous.
The solution is neither piecewise polynomial. Actually, thediscontinuities of the approximate solution are controlled bythe penalization parameter and they do not hamper theaccuracy of the method.
• DG requires more degrees of freedom than SEM.
That is true for a given mesh. But there is no reason to usethe same mesh for both methods.
• The stability condition (which restricts the time space of thescheme) is smaller for DG than for SEM.
That is true for a given mesh.
“Drawbacks” of the DG methods
• The solution to the wave equation is not discontinuous.
The solution is neither piecewise polynomial. Actually, thediscontinuities of the approximate solution are controlled bythe penalization parameter and they do not hamper theaccuracy of the method.
• DG requires more degrees of freedom than SEM.
That is true for a given mesh. But there is no reason to usethe same mesh for both methods.
• The stability condition (which restricts the time space of thescheme) is smaller for DG than for SEM.
That is true for a given mesh.
What are the good criteria to compare IPDG and SEM?
• Compare the computational burden for a given accuracy.
• Compare the accuracy for a given computational burden.
The computation cost of one iteration is directly related tothe number of degrees of freedom.=⇒ compare the accuracy for a given number of degrees of
freedom
What are the good criteria to compare IPDG and SEM?
• Compare the computational burden for a given accuracy.
• Compare the accuracy for a given computational burden.
The computation cost of one iteration is directly related tothe number of degrees of freedom.=⇒ compare the accuracy for a given number of degrees of
freedom
What are the good criteria to compare IPDG and SEM?
• Compare the computational burden for a given accuracy.
• Compare the accuracy for a given computational burden.The computation cost of one iteration is directly related tothe number of degrees of freedom.
=⇒ compare the accuracy for a given number of degrees offreedom
What are the good criteria to compare IPDG and SEM?
• Compare the computational burden for a given accuracy.
• Compare the accuracy for a given computational burden.The computation cost of one iteration is directly related tothe number of degrees of freedom.=⇒ compare the accuracy for a given number of degrees of
freedom
Comparison IPDG versus SEM
To obtain a given accuracy:
• The number of degrees of freedom and the number ofmultiplications required by IPDG are (slightly) smaller.
• The stability condition of IPDG is higher (and the number ofiterations is smaller).
IPDG performs as well as (and sometimes better than) SEM.
A more realistic experiment
A more realistic experiment
Improvements of the method
• Use P1 polynomials in the fine mesh and P3 polynomials inthe coarse mesh (p-adaptivity)
Improvements of the method
• Use P1 polynomials in the fine mesh and P3 polynomials inthe coarse mesh (p-adaptivity)
Improvements of the method
Improvements of the method
Improvements of the method
• Use P1 polynomials in the fine mesh and P3 polynomials inthe coarse mesh
• Use a local time stepping strategy
• Use a second order time scheme in the fine mesh and a fourthorder time scheme in the coarse mesh.
Improvements of the method
• Use P1 polynomials in the fine mesh and P3 polynomials inthe coarse mesh
• Use a local time stepping strategy
• Use a second order time scheme in the fine mesh and a fourthorder time scheme in the coarse mesh.
Improvements of the method
• Use P1 polynomials in the fine mesh and P3 polynomials inthe coarse mesh
• Use a local time stepping strategy
• Use a second order time scheme in the fine mesh and a fourthorder time scheme in the coarse mesh.
Improvements of the method
• Use P1 polynomials in the fine mesh and P3 polynomials inthe coarse mesh
• Use a local time stepping strategy
• Use a second order time scheme in the fine mesh and a fourthorder time scheme in the coarse mesh.
Local Time Stepping : Bibliography
POems Team: Becache, Collino, Fouquet, Joly, Rodrıguez
• Conservation of energy
• Optimal stability condition
• Requires the introduction of a Lagrange Multiplier
• Implicit scheme on the interface
Piperno
• First-order Maxwell system
• Conservation of energy
• Optimal stability condition
• Explicit or implicit scheme on the interface
Local Time Stepping : Bibliography
POems Team: Becache, Collino, Fouquet, Joly, Rodrıguez
• Conservation of energy
• Optimal stability condition
• Requires the introduction of a Lagrange Multiplier
• Implicit scheme on the interface
Piperno
• First-order Maxwell system
• Conservation of energy
• Optimal stability condition
• Explicit or implicit scheme on the interface
Local Time Stepping : Bibliography
ADER Schemes, Kaser, Dumbser et al.
• High-order Explicit Time Schemes ;
• First order Systems ;
• No energy conservation ;
• Difficult to implement (one time step by element).
Local Time Stepping : Bibliography
Hairer, Lubich and Wanner (2002), Leimkuhler and Reich (2004) :Local Time Stepping for ODE’s (second order scheme)
Diaz, Grote (2009) : High-Order Local Time Stepping for theWave Equation.
• Conservation of energy
• Optimal stability condition
Local Time Stepping : Bibliography
Hairer, Lubich and Wanner (2002), Leimkuhler and Reich (2004) :Local Time Stepping for ODE’s (second order scheme)
Diaz, Grote (2009) : High-Order Local Time Stepping for theWave Equation.
• Conservation of energy
• Optimal stability condition
Local Time Stepping
Local Time Stepping
Conclusions
• The performances of IPDG and SEM are similar with highorder polynomials in 2D.
• It is much easier to use elements of various order with IPDG.
• The local time stepping strategy does not hamper theaccuracy of the method.
• Comparison of IPDG and SEM in 3D.
• Extension to elastodynamics.
Time discretization of the wave equation
Md2U
dt2+ KU = 0
We consider space discretization methods such that M and K aresymmetric positive matrices and M is (block-)diagonal (FEM withmass lumping or DG methods).
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
M12d2U
dt2+ M−
12KM−
12︸ ︷︷ ︸
A
M12U︸ ︷︷ ︸Y
= 0
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2=
d2Y
dt2(t) + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2=
d2Y
dt2(t) + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= AY (t) + O(∆t2)
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y n+1 − 2Y n + Y n−1
∆t2= −AY n.
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equation
d2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y n+1 − 2Y n + Y n−1
∆t2= −AY n.
Energy Conservation
En+ 12 =
⟨Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩+⟨AY n+1,Y n
⟩
Time discretization of the wave equationd2Y
dt2+ AY = 0
Classical Leap Frog Scheme:
Y n+1 − 2Y n + Y n−1
∆t2= −AY n.
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩
Time discretization of the wave equation
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩CFL Condition
The scheme is stable, if and only if :
I − ∆t2
4A and A are symmetric positive
Time discretization of the wave equation
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩CFL Condition
The scheme is stable, if and only if :
I − ∆t2
4A and A are symmetric positive
0 ≤ λA ≤4
∆t2
Time discretization of the wave equation
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩CFL Condition
The scheme is stable under the CFL condition :
∆t ≤ αLFh
Time discretization of the wave equation
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩CFL Condition
The scheme is stable under the CFL condition :
∆t ≤ αLFh
Time discretization of the wave equation
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩CFL Condition
We want the new scheme to satisfy:
∆tcoarse ≤ αLFhcoarse and ∆tfine ≤ αLFh
fine
≈ αLFhcoarse/p
Time discretization of the wave equation
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4A
)Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨AY n+1 + Y n
2,Y n+1 + Y n
2
⟩CFL Condition
We want the new scheme to satisfy:
∆tcoarse ≤ αLFhcoarse and ∆tfine ≤ αLFh
fine ≈ αLFhcoarse/p
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:.
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2=
d2Y
dt2(t) +
∆t2
12
d4Y
dt4(t) + O(∆t4).
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2=
d2Y
dt2(t) +
∆t2
12
d4Y
dt4(t) + O(∆t4).
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= AY (t) +
∆t2
12Ad2Y
dt2(t) + O(∆t4).
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= AY (t) +
∆t2
12Ad2Y
dt2(t) + O(∆t4).
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y (t + ∆t)− 2Y (t) + Y (t −∆t)
∆t2= AY (t) +
∆t2
12A2Y (t) + O(∆t4).
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y n+1 − 2Y n + Y n−1
∆t2= −AY n +
∆t2
12A2Y n.
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
d2Y
dt2+ AY = 0
Modified Equation Scheme:
Y n+1 − 2Y n + Y n−1
∆t2= −AY n +
∆t2
12A2Y n.
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
Higher Order Schemes, Global Time Stepping
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
CFL Condition
The scheme is stable, if and only if
I − ∆t2
4
(A− ∆t2
12A2
)and A− ∆t2
12A2 are symmetric positive
Higher Order Schemes, Global Time Stepping
Energy Conservation
En+ 12 =
⟨(I − ∆t2
4
(A− ∆t2
12A2
))Y n+1 − Y n
∆t,Y n+1 − Y n
∆t
⟩
+
⟨(A− ∆t2
12A2
)Y n+1 + Y n
2,Y n+1 + Y n
2
⟩
CFL Condition
The scheme is stable under the CFL condition
∆t ≤ αMEh =√
3αLFh
Higher Order Schemes, Global Time Stepping
CFL Condition
The scheme is stable under the CFL condition
∆t ≤ αMEh =√
3αLFh
Higher Order Schemes, Global Time Stepping
CFL Condition
We want the new scheme to satisfy
∆tcoarse ≤ αMEhcoarse and ∆tfine ≤ αMEh
fine
≈ αMEhcoarse/p
Higher Order Schemes, Global Time Stepping
CFL Condition
We want the new scheme to satisfy
∆tcoarse ≤ αMEhcoarse and ∆tfine ≤ αMEh
fine ≈ αMEhcoarse/p
Auxiliary Function
At each time step n we define an auxiliary function
Qn(τ) =Y (n∆t − τ) + Y (n∆t + τ)
2
for τ ∈ [−∆t ; ∆t].
This function is obviously even and satisfy:d2Qn
dτ2(τ) = −AQn(τ),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
After having solved this equation, Y ((n + 1)∆t) can be computedusing Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t)
Auxiliary Function
At each time step n we define an auxiliary function
Qn(τ) =Y (n∆t − τ) + Y (n∆t + τ)
2
for τ ∈ [−∆t ; ∆t].
This function is obviously even and satisfy:d2Qn
dτ2(τ) = −AQn(τ),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
After having solved this equation, Y ((n + 1)∆t) can be computedusing Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t)
Auxiliary Function
At each time step n we define an auxiliary function
Qn(τ) =Y (n∆t − τ) + Y (n∆t + τ)
2
for τ ∈ [−∆t ; ∆t].
This function is obviously even and satisfy:d2Qn
dτ2(τ) = −AQn(τ),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
After having solved this equation, Y ((n + 1)∆t) can be computedusing Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t)
Two different ways to solve the auxiliary equation
First Way
Solve d2Qn
dτ2(τ) = −AQn(τ),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
by a fourth order modified equation scheme of time step ∆t/p andcompute Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t).
Remark
Is is equivalent to solve the original equation by a fourth ordermodified equation scheme of time step ∆t/p.
Two different ways to solve the auxiliary equation
First Way
Solve d2Qn
dτ2(τ) = −AQn(τ),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
by a fourth order modified equation scheme of time step ∆t/p andcompute Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t).
Remark
Is is equivalent to solve the original equation by a fourth ordermodified equation scheme of time step ∆t/p.
Two different ways to solve the auxiliary equation
Second Way
Use a fourth order approximation of AQn(τ):
AQn(τ) ≈ AQn(0) +τ2
2Ad2Qn(0)
dτ2= AY (n∆t)− τ2
2AAY (n∆t),
then solved2Qn
dτ2(τ) = −AY (n∆t) +
τ2
2AAY (n∆t),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
by a fourth order modified equation scheme of time step ∆t/pand compute Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t).
Remark
Is is equivalent to solve the original equation by a fourth ordermodified equation scheme of time step ∆t, whatever is p.
Two different ways to solve the auxiliary equation
Second Way
Use a fourth order approximation of AQn(τ):
AQn(τ) ≈ AQn(0) +τ2
2Ad2Qn(0)
dτ2= AY (n∆t)− τ2
2AAY (n∆t),
then solved2Qn
dτ2(τ) = −AY (n∆t) +
τ2
2AAY (n∆t),
Qn(0) = Y (n∆t),dQn
dτ(0) = 0,
by a fourth order modified equation scheme of time step ∆t/pand compute Y ((n + 1)∆t) = −Y ((n − 1)∆t) + 2Qn(∆t).
Remark
Is is equivalent to solve the original equation by a fourth ordermodified equation scheme of time step ∆t, whatever is p.
Local Time-Stepping
d2Qn
dτ2+ AQn = 0
Local Time-Stepping
d2Qn
dτ2+ AQn = 0
Let us now split Qn in two parts :
Qn =
[Qcoarse
n
Qfinen
]
Local Time-Stepping
d2Qn
dτ2+ AQn = 0
Let us now split Qn in two parts :
Qn =
[Qcoarse
n
0
]+
[0
Qfinen
]
Local Time-Stepping
d2Qn
dτ2+ AQn = 0
Let us now split Qn in two parts :
Qn =
[Qcoarse
n
0
]+
[0
Qfinen
]= (I − P)Qn + PQn, with P2 = P
Local Time-Stepping
d2Qn
dτ2+ A(I − P)Qn + APQn = 0
Let us now split Qn in two parts :
Qn =
[Qcoarse
n
0
]+
[0
Qfinen
]= (I − P)Qn + PQn, with P2 = P
Local Time-Stepping
d2Qn
dτ2+ A(I − P)Qn + APQn = 0
Idea
Approximate only A(I − P)Qn(τ) by
A(I − P)Qn(τ) ≈ A(I − P)Qn(0) +τ2
2A(I − P)
d2Qn(0)
dτ2
Local Time-Stepping
d2Qn
dτ2+ A(I − P)Qn + APQn = 0
Idea
Approximate only A(I − P)Qn(τ) by
A(I − P)Qn(τ) ≈ A(I − P)Y (n∆t)− τ2
2A(I − P)AY (n∆t)
Local Time-Stepping
d2Qn
dτ2+ A(I − P)Qn + APQn = 0
Idea
Approximate only A(I − P)Qn(τ) by
A(I − P)Qn(τ) ≈ A(I − P)Y (n∆t)− τ2
2A(I − P)AY (n∆t)
So that Qn is the solution to∣∣∣∣∣∣∣∣∣d2Qn(τ)
dτ2+ A(I − P)Y (t)− τ2
2A(I − P)AY (n∆t) + APQn(τ) = 0
Qn(0) = Y (t)
Q ′n(0) = 0
Algorithm of the fourth order local time-stepping scheme
Computation of Q(∆t)
We solve
∣∣∣∣∣∣∣∣∣d2
dτ2Qn(τ) + A(I − P)Y n − τ2
2A(I − P)AY n + APQn(τ) = 0
Qn(0) = Y n
Q ′n(0) = 0from τ = 0 to τ = ∆t, using a fourth order Modified Equation Scheme
with a time step∆t
p.
Algorithm of the fourth order local time-stepping scheme
Computation of Q(∆t)
Q0n = Y n
V 1 = −A(I − P)Y n − APQ0n = −AY n
V 2 = A(I − P)AY n − APV 1
Q1pn = Q0
n +∆t2
2p2V 1 +
∆t4
24p4V 2
For i = 1..p − 1
V 1 = −A(I − P)Y n + 12
(i∆tp
)2A(I − P)AY n − APQ
ipn
V 2 = A(I − P)AY n − APV 1
Qi+1p
n = 2Qipn − Q
i−1p
n +∆t2
p2V 1 +
∆t4
12p4V 2
Endfor