Asymptotic-preserving methods: a new paradigm for ...conditions. Forlinear multistep methods (LMM)if...

Introduction Implicit-explicit methods Remarks & Conclusions

Asymptotic-preserving methods: a new paradigm formultiscale differential problems

Lorenzo Pareschi

Department of Mathematics and Computer ScienceUniversity of Ferrara, Italy

http://lorenzopareschi.com

Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 1 / 36


Motivations

Motivations

Many applications involve multiscale differential problems, likeconvection-diffusion systems

Convection-Diffusion Equations

∂tU + ∂xF (U) = ∂x(K(U)∂xU), x ∈ R,

or hyperbolic balance laws with source terms of the form

Conservation Laws with Relaxation

∂tU + ∂xF (U) =1

εR(U), x ∈ R,

where U = U(x, t) ∈ RN , and ε > 0 is called relaxation parameter. Several kineticequations have the same structure with U = F (U) = f(x, v, t) ≥ 0, v ∈ R.

I The numerical discretization of such systems may be challenging in presence ofmultiple scales, typically when the diffusion or the source terms are stiff.



Motivations

Applications

Gas dynamics (gas with chemical reactions, extended thermodynamics)

Shallow water equations

Granular gases

Hydrodynamic models for semiconductors

Traffic flow models...

Kinetic equations (Fokker-Planck, Boltzmann-like)

Convection-diffusion-reaction equations (biology, environment)

0

0.5

1

1.5

2

2.5

3

x 104

2 4 6 8 10 12 14

1

2

3

4

5

6

7

8

9

10

11

x

y



Numerical approaches


Method of lines (MOL) approach

Discretize all spatial operators

Obtain a system of ODEs

U ′ = F(U) + G(U)

with F non stiff and G a stiff term.

Integrate ODEs system in time

Advantages

Spatial discretization and time integration are treated separately

Spatial discretization: easy to combine different schemes

Time integration: free to choose suitable method (Runge-Kutta, multi-step,etc.)





Fully explicit methods

Non stiff term: ∆t ≤ ρ(∇uF )∆x (CFL condition)

Stiff term: ∆t ≤ D−1(∆x)2 or ∆t ≤ Cε.Stability will require very small step-sizes for stiff sources, diffusion orrelaxation terms (ε small).

Fully implicit methods

For problems with shocks or steep gradients, implicit methods are not muchbetter than explicit ones (spurious shocks and wrong wave propagation speedwhen the CFL is violated).

For convection discretizations with slope limiters, the implicit relations arehard (expensive) to solve even for linear problems.

I Thus it is desirable to develop schemes which are Implicit in G(U) and Explicitin F(U) (IMEX).




Explicit vs Implicit Convection

Buckley-Leverett equation F (U) = U2

(U2+ 13(1−U)2)

with second order space discretization

(Van Leer limiter). Solution for explicit BDF2 (top) and implicit BDF2 (bottom) with

∆t/∆x = 1/8 (blue), 1/4 (magenta), 1/2 (red) at t = 0.25.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−0.2

0

0.2

0.4

0.6

0.8

1 extrap. BDF2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−0.2

0

0.2

0.4

0.6

0.8

1 impl. BDF2

Un+1 = 43Un − 1

3Un−1

+ 43∆tF(Un)

− 23∆tF(Un−1)

Un+1 = 43Un − 1

3Un−1

+ 23∆tF(Un+1)



Requirements on IMEX

Requirements on IMEX

The combination of the implicit and explicit method should satisfy suitable orderconditions. For linear multistep methods (LMM) if both methods are of order pthen the IMEX scheme has order p. For Runge-Kutta (RK) schemes we need tosatisfy additional mixed compatibility conditions.

Explicit methodThe stability region should be the largest possible.Monotonicity requirements

‖Un+1‖ ≤ ‖Un‖, ∆t ≤ ∆t∗

Strong Stability Preserving (SSP) property1.

Implicit methodStable for stiff systems, and good damping properties.The method should be Asymptotic Preserving (AP) namely it should beconsistent with the model reduction that may occur in very stiff regimes 2.

1S.Gottlieb, C-W.Shu, E.Tadmor ’01, R.Spiteri, S.Ruth, ’022S.Jin ’99



The asymptotic-preserving (AP) property

A toy example

Consider the singularly perturbed problem3

Singularly perturbed problem

P ε :

{u′(t) = f(u, v),εv′(t) = g(u, v), ε > 0.

As ε→ 0 we get the index 1 differential algebraic equation (DAE)

u′(t) = f(u, v), 0 = g(u, v).

Assuming that g(u, v) = 0⇔ v = E(u) we obtain

P 0 : u′(t) = f(u,E(u)).

Explicit methods: restricted to ∆t ∼ ε.Implicit methods: require the numerical inversion of f(u, v) and g(u, v), and asε→ 0 must satisfy the algebraic condition g(u, v) = 0⇔ v = E(u).

3E.Hairer, C.Lubich, M.Roche ’89Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 8 / 36


The asymptotic-preserving (AP) property

The AP diagram

P ε

P ε∆t

∆t→ 0 ∆t→ 0

ε→ 0

ε→ 0

6

-

P 0

P 0∆t

6

-

In the diagram P ε is the original singular perturbation problem and P ε∆t itsnumerical approximation characterized by a discretization parameter ∆t.The asymptotic-preserving (AP) property corresponds to the request that P ε∆t is aconsistent discretization of P 0 as ε→ 0 independently of ∆t.



Space discretizations

Space discretizations

Consider the case of the single hyperbolic equation

Conservation Law

Ut + ∂xF (U) = 0.

We can use any finite difference/volume or spectral method to approximatethe spatial derivative, and use the standard (linear) stability analysis.

In presence of shocks and discontinuities this stability analysis is not sufficient(nonlinear problems can develop discontinuous solutions in finite time evenstarting from a smooth solution).

Build spatial discretizations which capture the shock structure and thatsatisfy some nonlinear stability properties. These methods include totalvariation diminishing (TVD) schemes and essentially non-oscillatory (ENO)or weighted ENO (WENO) schemes4.

4A. Harten ’87, T.Chan, X-D.Liu, S.Osher ’94, G-S.Jang, C-W.Shu ’95Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 10 / 36


Splitting methods

Implicit-explicit methodsSplitting methods

We consider the system of stiff ODE’s

System of stiff ODEs

U ′ = F(U) + G(U)

where F is non stiff and G is a stiff term.Splitting methods

Solve separately the advection problem and the stiff source problem

U ′ = F(U), t ∈ [0, T ] U ′ = G(U), t ∈ [0, T ].

Although it is only first order accurate (even if the two steps are exact, unlessthe operators commute), it is very popular due to its simple concept and thefreedom in choosing different solvers for advection and sources.

Higher order splitting (ex. Strang splitting) can be constructed but maypresent a loss of accuracy when the source term is highly stiff.



IMEX Runge-Kutta methods

IMEX Runge-Kutta methods5

IMEX Runge-Kutta

Ui = Un + ∆t

i−1∑j=1

aijF(t0 + cj∆t, Uj) + ∆t

ν∑j=1

aijG(t0 + cj∆t, Uj),

Un+1 = Un + ∆t

ν∑i=1

wiF(t0 + ci∆t, Ui) + ∆t

ν∑i=1

wiG(t0 + ci∆t, Ui).

A = (aij), aij = 0, j ≥ i and A = (aij): ν × ν matrices.The coefficient vectors are c = (c1, . . . , cν)T , w = (w1, . . . , wν)T ,c = (c1, . . . , cν)T , w = (w1, . . . , wν)T .I We restrict to diagonally implicit (DIRK) scheme, aij = 0, j > i since theyguarantee that F is evaluated explicitly.

5U.Ascher, S.Ruth, R.Spiteri ’97, L.P., G.Russo ’00Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 12 / 36



Order conditions

If wi = wi and ci = ci mixed conditions are automatically satisfied. This isnot true for higher that third order accuracy

IMEX-RK schemes are a particular case of additive Runge-Kutta (ARK)methods. Higher order conditions can be derived using a generalization ofButcher 1-trees to 2-trees.

The number of coupling conditions increase dramatically with the order ofthe schemes6.

IMEX-RK Number of coupling conditionsOrder General case wi = wi c = c c = c and wi = wi

1 0 0 0 02 2 0 0 03 12 3 2 04 56 21 12 25 252 110 54 156 1128 528 218 78

6M.Carpenter, C.Kennedy, ’03Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 13 / 36



Design of IMEX-RK

Start with a p-order explicit SSP method and find the p-order DIRK method thatmatches the order conditions with good damping properties (L-stability).

Second order SSP IMEX-RK

U1 = Un + γ∆tG(U1)

U2 = Un + ∆tF(Un) + (1− 2γ)∆tG(U1) + γ∆tG(U2)

Un+1 = Un +1

2∆t(F(Un) + F(U1)) +

1

2∆t(G(U1) + G(U2)),

with γ = (1−√

2)/2.Third order SSP IMEX-RK

U1 = Un + γ∆tG(U1)

U2 = Un + ∆tF(Un) + (1− 2γ)∆tG(U1) + γ∆tG(U2)

U3 = Un +1

4∆t(F(Un) + F(U1)) + (1/2− γ)∆tG(U1) + γ∆tG(U3)

Un+1 = Un +1

6∆t(F(Un) + F(U1) + 4F(U2)) +

1

6∆t(G(U1) + G(U2) + 4G(U3)),

with γ = (1−√

2)/2.Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 14 / 36


IMEX Linear Multistep Methods

IMEX Linear Multistep Methods7

IMEX Linear Multistep

Un+1 =

s−1∑j=0

ajUn−j + ∆t

s−1∑j=0

wjF(Un−j) + ∆t

s−1∑j=−1

wjG(Un−j),

with starting values U0, U1, . . . , Un.

Order p = s.

Schemes are a direct combination of an explicit and an implicit method. Noneed of compatibility conditions.

Stability constraints usually increase with the order of the schemes (rathersevere for SSP methods). A-stable schemes have accuracy p ≤ 2.

Off course they need to be initialized and the starting values imply additionalstorage requirements.

7U.Ascher, S.Ruth, B.Wetton ’95, W.Hundsdorfer, S.Ruth ’07Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 15 / 36


IMEX Linear Multistep Methods

Design of IMEX-LMM

As for IMEX-RK we can start from an explicit SSP method and find thecorresponding implicit method with good damping properties (A(α)-stability). Orwe can start from an implicit method (BDF) and use the corresponding explicitscheme.

Second order IMEX-BDF

Un+1 =4

3Un − 1

3Un−1 +

4

3∆tF(Un)− 2

3∆tF(Un−1) +

2

3∆tG(Un+1).

Third order SSP IMEX-LM

Un+1 =3909

2048Un − 1367

1024Un−1 +

873

2048Un−2

+18463

12288∆tF(Un)− 1271

768∆tF(Un−1) +

8233

12288∆tF(Un−2)

+1089

2048∆tG(Un+1)− 1139

12288∆tG(Un)− 367

6144∆tG(Un−1) +

1699

12288∆tG(Un−2).



Application to hyperbolic relaxation systems


Consider the case of hyperbolic relaxation systems8

Hyperbolic system with relaxation (Full model)

∂tU + ∂xF (U) =1

εR(U), x ∈ R.

R : RN → RN is a relaxation operator if there exists a n×N matrix Q withrank(Q) = n < N s.t. QR(U) = 0 ∀ U ∈ RN .This gives n conserved quantities u = QU that uniquely determine a local equilibriumU = E(u), s.t. R(E(u)) = 0, and satisfy

∂t(QU) + ∂x(QF (U)) = 0.

As ε→ 0 ⇒ R(U) = 0 ⇒ U = E(u) ⇒ (subcharacteristic condition on on f(u))

Equilibrium system (Reduced model)

∂tu+ ∂xf(u) = 0, f(u) = QF (E(u)).

8G.Chen, D.Levermore, T.P.Liu, ’94Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 17 / 36



A simple example

A simple prototype example of relaxation system is given by9

Jin-Xin relaxation system{∂tu+ ∂xv = 0,

∂tv + ∂xau = − 1

ε(v − f(u)),

where u = u(x, t), v = v(x, t), (x, t) ∈ R× R+.For small values of ε we get the local equilibrium

v = f(u)

and (subcharacteristic condition a > f ′(u)2) we obtain at O(ε)

∂tu+ ∂xf(u) = ε∂x((a− f ′(u)2)∂xu).

9S.Jin, Z.Xin ’95Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 18 / 36



AP schemes

DefinitionAn IMEX scheme for an hyperbolic system with relaxation is asymptotic preserving(AP) if in the limit ε→ 0 the scheme becomes a consistent discretization of the limitsystem of conservation laws. We use the notation APk if the scheme is of order k inthe limit ε→ 0.

We can prove10

TheoremIf detA 6= 0 then in the limit ε→ 0, the IMEX scheme applied to an hyperbolic systemwith relaxation becomes the explicit RK scheme characterized by (A, w, c) applied tothe limit system of conservation laws.

The theorem guarantees that in the stiff limit the IMEX scheme becomes theexplicit RK scheme applied to the equilibrium system.

To satisfy detA 6= 0 it is necessary that c 6= c. The corresponding scheme may beinaccurate if the initial condition is not “well prepared” (initial layer).

10L.P., G.Russo, ’04Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 19 / 36


Stability

Stability

To study the A-stability of a IMEX-RK scheme, one may consider the problem11

Test problem

u′ = λu+ µu, u(0) = 1, λ, µ ∈ C.

This test problem characterizes the stability properties also for linear systems

U ′ = AU +B U, U(0) = U0

with U ∈ Rm, and A,B ∈ Rm×m if A and B are normal, commuting matrices.In general the two matrices do not share the same eigenvectors, and can not bediagonalized simultaneously. This makes the stability analysis for systemsextremely difficult. Only available results are for the case m = 2.

11L.P., G.Russo ’00, L.P. G.Russo ’08Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 20 / 36


Numerical examples

Numerical examples

Broadwell model

∂tρ+ ∂xm = 0,

∂tm+ ∂xz = 0,

∂tz + ∂xm =1

ε(ρ2 +m2 − 2ρz),

with ε is the mean free path. The dynamical variables ρ and m are the density and themomentum respectively, while z represents the flux of momentum.In the relaxation limit ε→ 0 we obtain

∂tρ+ ∂xm = 0

∂tm+1

2∂x

(ρ+

m2

ρ

)= 0

(1) Accuracy test for IMEX-RK schemes with smooth initial data and periodic b.c.(2) Shock test for IMEX-RK schemes.



Numerical examples

Convergence rates

ε 1.0 10−1 10−2 10−3 10−4 10−5 10−6

Scheme Convergence rates for ρ

ARS 2.018553 1.513026 1.159427 1.165810 1.165483 1.165396 1.165387IMEX-SSP2 2.042282 2.054974 2.051493 2.053945 2.043920 2.042585 2.042449

ARSF 2.044539 2.074181 2.007092 1.982191 2.042974 2.040301 2.040004IMEX-SSP2F 2.050203 2.064147 2.061288 2.065251 2.056496 2.055314 2.055192

Convergence rates for z

ARS 1.950904 1.438073 1.114938 1.121625 1.121979 1.121974 1.121972IMEX-SSP2 2.027090 2.045509 1.965834 1.501554 1.309825 1.302670 1.302573

ARSF 2.031510 2.174882 1.762384 1.596958 2.061907 2.040474 2.039201IMEX-SSP2F 2.036402 2.034042 2.038568 2.368099 2.127720 2.052993 2.051640

Convergence rates where extrapolation has been used in order to avoid degradation of accuracy.



Numerical examples

Convergence rates ε = 1

101

102

103

10−8

10−7

10−6

10−5

10−4

10−3

10−2

Relative error in density vs N, ε = 1, no initial layer

N

Err

ρ

SSP2−222

SSP2−322

SSP2−332

SSP3−332

SSP3−433

ARS−222

101

102

103

10−8

10−7

10−6

10−5

10−4

10−3

10−2

Relative error in density vs N, ε = 1, with initial layer

N

Err

ρ

SSP2−222

SSP2−322

SSP2−332

SSP3−332

SSP3−433

ARS−222

Relative error for different second and third order IMEX-RK schemes for the Broadwell

equations with ε = 1. Left: no initial layer. Right: initial layer.



Numerical examples

Convergence rates ε = 10−3

101

102

103

10−7

10−6

10−5

10−4

10−3

10−2

Relative error in density vs N, ε = 10−3

, no initial layer

N

Err

ρ

SSP2−222

SSP2−322

SSP2−332

SSP3−332

SSP3−433

ARS−222

101

102

103

10−7

10−6

10−5

10−4

10−3

10−2


, with initial layer

N

Err

ρ

SSP2−222

SSP2−322

SSP2−332

SSP3−332

SSP3−433

ARS−222


equations with ε = 10−3. Left: no initial layer. Right: initial layer.



Numerical examples

Convergence rates ε = 10−6

101

102

103

10−7

10−6

10−5

10−4

10−3

10−2


, no initial layer

N

Err

ρ

SSP2−222

SSP2−322

SSP2−332

SSP3−332

SSP3−433

ARS−222

101

102

103

10−7

10−6

10−5

10−4

10−3

10−2


, with initial layer

N

Err

ρ

SSP2−222

SSP2−322

SSP2−332

SSP3−332

SSP3−433

ARS−222


equations with ε = 10−6. Left: no initial layer. Right: initial layer.



Numerical examples

Shock test ε = 1

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10

0.5

1

1.5

2

2.5

x

ρ(x

,t),

m(x

,t),

z(x

,t)

IMEX−SSP2−WENO, ε=1, t=0.5, N=200

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10

0.5

1

1.5

2

2.5

x

ρ(x

,t),

m(x

,t),

z(x

,t)

IMEX−SSP3−WENO, ε=1, t=0.5, N=200

Numerical solution for second and third order SSP IMEX-RK schemes for the Broadwell

equations with ε = 1



Numerical examples

Shock test ε = 10−3

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10

0.5

1

1.5

2

2.5

x

ρ(x

,t),

m(x

,t),

z(x

,t)

IMEX−SSP2−WENO, ε=0.02, t=0.5, N=200

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10

0.5

1

1.5

2

2.5

x

ρ(x

,t),

m(x

,t),

z(x

,t)

IMEX−SSP3−WENO, ε=0.02, t=0.5, N=200


equations with ε = 10−3



Numerical examples

Shock test ε = 10−6

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10

0.5

1

1.5

2

2.5

x

ρ(x

,t),

m(x

,t),

z(x

,t)

IMEX−SSP2−WENO, ε=10−8

, t=0.5, N=200

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10

0.5

1

1.5

2

2.5

x

ρ(x

,t),

m(x

,t),

z(x

,t)

IMEX−SSP3−WENO, ε=10−8

, t=0.5, N=200


equations with ε = 10−6



Penalized IMEX Runge-Kutta for the Boltzmann equation

Design principles for kinetic equations

AP schemes avoiding the implicit solution of the collision term Q(f, f)

Boltzmann-like equations

∂tf + ∂xf =1

εQ(f, f), x ∈ R,

where f = f(x, v, t) and Q(f, f) = 0 implies f = M [f ] the local Maxwellian.

When f ≈M [f ] the collision operator is well approximated by its linearcounterpart Q(M,f) or directly by a BGK relaxation operator µ(M [f ]− f).

If we denote by LP (f) the linear approximating operator we can write 12

Penalized setting

Q(f, f) = G(f)︸︷︷︸explicit

+ LP (f)︸︷︷︸implicit

, G(f) = Q(f, f)−LP (f).

12S.Jin, F.Filbet ’11Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 29 / 36



Penalized IMEX Runge-Kutta methods

In the sequel we assume LP (f) = µ(M [f ]− f), µ > 0. The IMEX-RK schemetake the form 13

Penalized IMEX-RK for Boltzmann

F = fn e+ ∆t A

(1

εG(F )− v · ∇xF

)+µ∆t

εA(M [F ]− F )

fn+1 = fn + ∆t wT(

1

εG(F )− v · ∇xF

)+µ∆t

εwT (M [F ]− F ).

Clearly the scheme being implicit only in the linear part, which can be easilyinverted and computed, can be implemented explicitly.

Note however that here the problem is stiff as a whole. The hope is thatapplying the same design principles we used for hyperbolic systems withrelaxation we get an AP-scheme for the full Boltzmann model.

13G.Dimarco, L.P. ’13Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 30 / 36



AP-property

First let us point out that since the linear operator enjoys the same conservationproperty of the full Boltzmann operator we have the same associated momentscheme characterized by (A, w) of the explicit method∫

R3

Fφ(v) dv =

∫R3

fn eφ(v) dv −∆tA

∫R3

v · ∇xFφ(v) dv∫R3

fn+1φ(v) dv =

∫R3

fnφ(v) dv −∆t wT∫R3

v · ∇xFφ(v) dv.

Consider now an invertible matrix A and solve the IMEX scheme for (M [F ]− F )

∆t(M [F ]− F ) =ε

µA−1

[F − fn e+ ∆tA

(v · ∇xF −

1

εG(F )

)]Again as ε→ 0 we get

F (i) = M [F (i)], i = 1, . . . , ν.

In fact A is lower triangular with aii = 0 and we have a hierarchy of equations

G(F (i)) = Q(F (i), F (i))− µ(M [F (i)]− F (i)) = 0, i = 1, .., ν.




Further requirements

As opposite to the case of hyperbolic systems with relaxation, now the last levelstill depends on ε. After some manipulations it reads

fn+1 = fn(1− wTA−1e

)−∆t wT

(v · ∇xF −

1

εG(F )

)+ ∆t wTA−1A

(v · ∇xF −

1

εG(F )

)+ wTA−1F.

For small values of ε the scheme turns out to be unstable since fn+1 is notbounded. A remedy, is to consider globally stiffly accurate schemes for which

fn+1 = F (ν),

and so as ε→ 0F (ν) = M [F (ν)]⇒ fn+1 = M [fn+1].

I For the Boltzmann case the stiffly accurate property is required to have a stableAP and asymptotically accurate scheme.




Mixing regimes problem

Collision term approximated by the Fast Fourier-Galerkin method 14. Second andthird order WENO is used in space 15

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.2

0.4

0.6

0.8

1

1.2

1.4

ε(x)

x

Knudsen number value for the mixed regime test with ε0 = 10−4{ε = ε0 + 1

2 (tanh(16− 20x) + tanh(−4 + 20x)), x ≤ 0.7ε = ε0, x > 0.7

14L.P., B.Perthame ’96, C.Mouhot, L.P. ’0615C-W.Shu ’97




Mixing regimes: third order scheme

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.45

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9Density for the mixing regime problem

IMEX3Runge−Kutta 3

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10.6

0.7

0.8

0.9

1

1.1

1.2

1.3Temperature for the mixing regime problem

IMEX3Runge−Kutta 3

Density (left) and temperature (right) profiles for the mixing regime problem. Time

t = 0.5, Nx = 100 using third order WENO. Reference solution computed using a third

order Runge-Kutta. Here ∆tIMEX/∆tRK = 7.




Mixing regimes: second vs third order

0.7 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.80.45

0.5

0.55

Density for the mixing regime problem

IMEX3IMEX2Runge−Kutta 3

0.7 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.80.85

0.9

0.95

1

1.05

1.1

Temperature for the mixing regime problem

IMEX3IMEX2Runge−Kutta 3

Density (left) and temperature (right) profiles for the mixing regime problem at t = 0.5

for x ∈ [0.7, 0.8].



Conclusion

IMEX schemes represent a powerful tool for the time discretization of partialdifferential equations where convection and stiff sources/diffusion are present.

In principle they can be applied to any other PDE where there is the presenceof multiple time-scales.

However they are not a universal cure for all problems. It is not difficult toimagine a situation where a fully explicit (or implicit) method is preferable.

The most critical case is the application to (nonlinear) PDEs where the stiffscales originate a model reduction. In such cases AP methods are essential inorder to capture the correct physical behavior.

The AP-property can be generalized by expanding the solution in powers of εand matching higher order terms in the expansion with a given accuracy 16.

16S.Boscarino, ’08, S.Boscarino, G.Russo ’09Lorenzo Pareschi (University of Ferrara) AP methods for multiscale problems Roma, March 17, 2015 36 / 36

Asymptotic-preserving methods: a new paradigm for ...conditions. Forlinear multistep methods (LMM)if...

Documents

Transcript of Asymptotic-preserving methods: a new paradigm for ...conditions. Forlinear multistep methods (LMM)if...