Incompressible MHD simulations

Transcription

1 Incompressible MHD simulations Felix Spanier 1 Lehrstuhl für Astronomie Universität Würzburg Simulation methods in astrophysics Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 1 / 20

2 Outline Where do we need incompressible MHD? Theory of HD&MHD Turbulence Numerical methods for HD Numerical methods for MHD Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 2 / 20

3 Heliosphere Closest available space plasma: The solar wind includes shocks and density fluctuation The large scale evolution seems to be compressible... Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 3 / 20

4 Heliosphere... but in the frame of the solar wind, we find a turbulent spectrum Alfvén waves travelling Sounds like incompressible turbulence! Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 3 / 20

5 Interstellar medium Infamous picture this is usually used to suggest turbulence it really shows density fluctuations Density fluctuations? That s compressible turbulence? Really? Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 4 / 20

6 / er g cm s -3-1 Interstellar medium -3-1 / er g cm s Damping rates 1 1e+10 1e-05 Viskos Joule Landau Neutral Viskos Joule Neutral 1 1e-10 1e-05 1e-15 1e-10 1e-15 1e-20 1e-20 1e-25 1e-25 1e-30 1e-17 1e-16 1e-15 1e-14 1e-13 1e-12 1e-11 1e-10 1e-09 1e-08 1e-07 1e-06-1 k / cm Alfvén wave damping 1e-30 1e-17 1e-16 1e-15 1e-14 1e-13 1e-12 1e-11 1e-10 1e-09 1e-08 1e-07 1e-06-1 k / cm Sound wave damping Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 4 / 20

7 Interstellar medium So is this incompressible MHD? Wave damping suggests that for large k only Alfvén waves survive Density fluctuations suggest there is compressible turbulence Two solutions: Alfvén turbulence with enslaved density fluctuations Not waves but shocks govern the ISM Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 4 / 20

8 Euler Equation Force on a volume element of fluid F = p df (1) Divergence theorem F = p dv (2) Newton s law for one volume element m du dt = F (3) Using density instead of total mass ρ du dv = dt p dv (4) ρ du dt = p (5) Lagrange picture Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 5 / 20

9 Euler Equation Changing to material derivative Euler s equation du dt du = u dt + (dr )u (1) t u t = u t Here the velocity is a function of space and time + (u )u (2) + (u )u = 1 p (3) ρ u = u(r, t) (4) This can now be applied to the momentum transport equation t (ρu i) = ρ u i t + u i ρ t (5) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 5 / 20

10 Euler Equation Partial derivative of u is known from Euler Π ik = stress tensor u i t u i = u k 1 p (1) x k ρ x i t (ρu i) = ρu k u i x k p x i u i ρu k x k (2) = p x i = δ ik p x k (ρu i u k ) x k (3) (ρu i u k ) x k (4) ρu i t = Π ik x k (5) Π ik = pδ ik ρu i u k (6) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 5 / 20

11 Continuity equation t ( ρ t + (ρu) ρ t For incompressible fluids (ρ=const) ρ dv = ρu dv (7) ) dv = 0 (8) + (ρu) = 0 (9) u = 0 (10) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 6 / 20

12 Viscid fluid Π ik = pδ ik + ρu i u k σ ik (11) σ ik = shear stress For Newtonian fluids this depends on derivatives of the velocity Coefficients have to fulfill σ ik = a u i x k + b u k x i no viscosity in uniform fluids(u=const) no viscosity in uniform rotating floew It follows a = b ( ui σ ik = η + u k x k x i 2 3 δ ik + c u l x l δ ik (12) ) u l u l + ζδ ik (13) x l x l Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 7 / 20

13 Viscid fluid Stress tensor for incompressible fluids u i = 0 (11) x i ( ui σ ik = η + u ) k (12) x k x i σ ik x k = η 2 u i x 2 k (13) And we find the Navier-Stokes-equation : u t + (u )u + p = ν u (14) ν = η ρ = Viscosity Pressure p = (u )u + ν u (15) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 7 / 20

14 Vorticity Using u = 0 one can take the curl of the Navier-Stokes-equation, resulting in a PDE for ω = u ( u) (u ( u)) t = ν ( ( u)) (16) ω t = (ω )u + ν ω (17) ω is quasi-scalar for 2d-velocities We need the stream function Ψ to derive velocities ω = u (18) u = Ψ (19) Ψ = ω (20) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 8 / 20

15 The MHD equations Compressible MHD Solutions ρ v t ρ t + (ρv) = 0 + ρ(v )v = p 1 4π B B B t = (v B) Different wave modes are solutions to the MHD equations Alfvén modes (incompressible, aligned to the magnetic field) Fast magnetosonic (compressible) slow magnetosonic (compressible) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 9 / 20

16 The MHD equations Incompressible MHD Solutions ρ v t v = 0 + ρ(v )v = p 1 4π B B B t = (v B) Different wave modes are solutions to the MHD equations Alfvén modes (incompressible, aligned to the magnetic field) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 9 / 20

17 The MHD equations Elsasser variables w = v + b v A e z w + = v b + v A e z v = w + + w 2 b = w + + w 2v A e z 2 Solutions Different wave modes are solutions to the MHD equations Alfvén modes (incompressible, aligned to the magnetic field) w ± correspond to forward/backward moving waves Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 9 / 20

18 Short review of Kolmogorov Kolmogorov assumes energy transport which is scale invariant Time-scale depends on eddy-turnover times Detailed analysis of the units reveals 5/3-law Simplified picture but enough to illustrate what we are doing next E(k) Inertial- range k Productionrange Dissipationrange Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 10 / 20

19 Short review of Kolmogorov Kolmogorov assumes energy transport which is scale invariant Time-scale depends on eddy-turnover times Detailed analysis of the units reveals 5/3-law Simplified picture but enough to illustrate what we are doing next Dimensional Argument 5/3-law can be derived by assuming scale independent energy transport and dimensional arguments. This gives a unique solution. Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 10 / 20

20 Kraichnan-Iroshnikov General idea: Turbulence is governed by Alfvén waves Collision of Alfvén waves is mechanism of choice Eddy-turnover time is then replaced by Alfvén time scale Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 11 / 20

21 Kraichnan-Iroshnikov Energy E = (V 2 + B 2 )dx is conserved Cascading to smaller energies Dimensional arguments cannot be used to derive τ λ = λ/v λ Dimensionless factor v λ /v A can enter Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 11 / 20

22 Wave-packet interaction Kraichnan-Iroshnikov Solutions: w ± = f (r v A t) w = v + b v A e z w + = v b + v A e z Figure: Packets interact, energy is conserved, shape is not Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 11 / 20

23 Kraichnan-Iroshnikov Packet interaction Amplitudes δw + λ δw λ δv λ δb λ during one collision δv λ (δv 2 λ/λ)(λ/v A ) Number of collision required to change wave packet N (δv λ / δv λ ) 2 (v A /δv λ ) 2 τ IK Nλ/v A λ/δv λ (v A /δv λ ) E IK = δv 2 k k 2 k 3/2 Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 11 / 20

24 Goldreich-Sridhar 3/2 spectrum not that often observe Let s get a new theory We come back to Kraichnan-Iroshnikov: Alfvén waves govern the spectrum Turbulent energy is transported through multi-wave interaction Goldreich and Sridhar made up a number of articles dealing with this problem They distinguish between strong and weak turbulence Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

25 Goldreich-Sridhar Weak Turbulence What is weak turbulence? Weak turbulence describes systems, where waves propagate if one neglects non-linearities. If one includes non-linearities wave amplitude will change slowly over many wave periods. The non-linearity is derived pertubatevily from the interaction of several waves Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

26 Goldreich-Sridhar Weak Turbulence Goldreich-Sridhar claim that Kraichnan-Iroshnikov can be seen as three-wave interaction of Alfvén waves They also believe, this is a forbidden process, since resonance condition reads k z1 + k z2 = k z k z1 + k z2 = k z 4-wave coupling is now their choice! Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

27 Goldreich-Sridhar Weak Turbulence k 1 + k 2 = k 3 + k 4 ω 1 + ω 2 = ω 3 + ω 4 k 1z + k 2z = k 3z + k 4z k 1z k 2z = k 3z k 4z Parallel component unaltered 4-wave interaction just changes k of quasi-particles Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

28 Goldreich-Sridhar Weak Turbulence d 2 v λ dt 2 δv λ d 2 v λ dt 2 (k zv A ) 2 d dt (k v 2 λ) k v λ dv λ dt since (v ) v λ k for Alfvén waves k 2 v 3 λ ( ) δv λ 2 ( ) 4 v λ k v k v N k z v λ k z v λ v 2 λ = E(k z, k )d 3 k constant rate of cascade E(k z, k ) ɛ(k z )v a k 10/3 This spectrum now hold, when the velocity change is small Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

29 Goldreich-Sridhar Strong Turbulence Assume perturbation v λ on scales λ kz 1 and λ k 1 ζ λ k v λ k z v λ Anisotropy parameter N ζ 4 λ isotropic excitationk k z L 1 ( va v L ) 4 (k L) 4/3 Everything is fine as long as ζ λ 1 GS assume frequency renormalization when this is not given Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

30 Goldreich-Sridhar Strong Turbulence When energy is injected isotropically ζ L 1 we have a critical balance k z v A k v λ Alfvén timescale and cascading time scale match For the critical balance and scale-independent cascade we find E(k, k z ) k z k 2/3 L 1/3 v v A (k L) 1/3 ( ) v 2 A k f 10/3 L1/3 k z L 1/3 k 2/3 Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

31 Goldreich-Sridhar Strong Turbulence Consequences A cutoff scale exists Eddies are elongated Figure: From Cho, Lazarian, Vishniac (2003) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 12 / 20

32 Turbulence Conclusion There is a number of turbulence models out there It is not yet clear which is correct All analytical turbulence models are incompressible The only thing we know for sure is that there is power law in the spectral energy distribution Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 13 / 20

33 Numerical methods Methods for the Navier-Stokes-equation Projection solve compressible equations, project on incompressible part Vorticity use vector potential to always fulfill divergence-free condition Spectral use Fourier space to do the same as above Methods are presented for the Navier-Stokes-equation but also apply to MHD Spectral methods will be presented for Elsässer variables in detail Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 14 / 20

34 Projection method Different possible schemes Here a scheme proposed by [1]. Step 1: Intermediate Solution We are looking for a solution u, which is not yet divergence free u u n t = ((u )u) n+1/2 p n 1/2 + ν u n (21) This requires some standard scheme to solve the Navier-Stokes-equation (cf. Pekka s talk) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 15 / 20

35 Projection method Step 2: Dissipation Use Crank-Nicholson for stability in the dissipation step yields ν u n 1 2 ν (un + u ) (21) u = u n t ((u )u + p) (22) ( 1 ν t ) ( 2 u = 1 + ν t ) 2 u (23) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 15 / 20

36 Projection method Step 3: Projection step To project u on the divergence-free part, an auxiliary field φ is calculated, which fulfills (u t φ) = 0 (21) u φ = t (22) u n+1 = u t φ n+1 (23) Here a Poisson equation has to be solved. This is the tricky part and is the most time consuming. Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 15 / 20

37 Projection method Step 4: Pressure gradient Finally the pressure is updated. This can be calculated using the divergence of the Navier-Stokes-equation p n+1/2 = p n 1/2 + φ n+1 ν t 2 φn+1 (21) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 15 / 20

38 Projection method Total scheme u u n t + p n+1/2 = [(u )u] n+1/2 + ν 2 2 (u n + u ) (21) t 2 φ n+1 = u (22) u n+1 = u t φ n+1 (23) p n+1/2 = p n 1/2 + φ n+1 ν t 2 2 φ n+1 (24) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 15 / 20

39 Vorticity-Streamline The natural formulation for divergence-free flows is the stream function u = Ψ Step 1: Calculate velocity u n = Ψ n (25) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 16 / 20

40 Vorticity-Streamline Step 2: Solve the Navier-Stokes-equation t u = (u ) u + ν u (25) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 16 / 20

41 Vorticity-Streamline Step 3: Calculate Vorticity t ω = ( t u) (25) ω n+1 = ω n + dt ω (26) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 16 / 20

42 Vorticity-Streamline Step 4: Calculate Streamfunction Solve Poisson equation Ψ n+1 = ω n+1 (25) Problem: This would result in a mesh-drift instability. Use staggered grid! Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 16 / 20

43 Vorticity-Streamline Complete scheme u = Ψ (25) (u x ) i,j+1/2 = Ψ i,j+1 Ψi, j (26) dy (u y ) i+1/2,j = Ψ i+1,j Ψi, j (27) ( dx ( ) ) 1/2 (ux ) u i+1/2,j+1/2 = i+1,j+1/2 + (u x ) i,j+1/2 1/2 ( ) (28) (u y ) i+1/2,j+1 + (u y ) i+1/2,j Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 16 / 20

44 Spectral methods Basic idea: Use Fourier transform to transform PDE into ODE Problem: This is complicated for the nonlinear terms. Spectral Navier-Stokes-equation ũ α t k α ũ α = 0 ( = ik γ δ αβ k ) αk β ) (ũβ k 2 u γ νk 2ũ α (29) Quantities with a tilde are fouriertransformed. What s so special about the red and green term? Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

45 Spectral methods The green term Start with the Navier-Stokes-equation we can identify u t + (u )u + p = 0 (29) ik γ δ αβ ũ β u γ = (u )u (30) and taking the divergence of the Navier-Stokes-equation p = (u )u (31) so the second term is the gradient of the pressure ( ) kα k β ) ik γ (ũβ k 2 u γ = p (32) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

46 Spectral methods Step 1: Fouriertransform Starting with ũ α the velocity has to be transformed into u α Then u α u β can be calculated (only 6 tensor elements are needed) This is transformed back to Fourier space ũ α u β Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

47 Spectral methods Step 2: Anti-Aliasing In the high wavenumber regime of u α we will find flawed data. Due to the fact, that not all data is correctly attributed for by the Fourier transform, aliased data has to be removed For k > 1/2k max set ũ α u β (k) = 0 Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

48 Spectral methods Step 3: Update velocity ũ α = ũ n α t(ik γ ( δ αβ k αk β k 2 ) (ũβ u γ ) νk 2ũ α ) (29) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

49 Spectral methods Step 4: Projection Project u α on its divergence free part u n+1 = u k(k u ) k 2 (29) Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

50 Spectral methods Why to use Spectral Methods? Pro Easy to implement Relies on fast FFT-algorithms Easy to parallelize Contra high aliasing-loss Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 17 / 20

51 Fourier Transforms Fourier transforms are available in large numbers on the market FFTW 2 ubiquitious library, allows parallel usage FFTW 3 the improved version, no parallel support Intel MKL includes DFT, much faster than FFTW, simple parallelized version, limited to Intel-like systems P3DFFT UC San Diego s Fortran based FFT, highly parallelizable, requires FFTW-3 Sandia FFT Sandia Lab s C based FFT, highly parallelizable, requires FFTW-2 Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 18 / 20

52 Parallel Computing Usually a parallel fluid simulation requires exchange of field borders between processors. Spectral methods are slightly different: The calculation of u α u β and the update step are completely local. They have no derivatives and do not need neighboring points Every non-local interaction is done in the FFT So the only thing we have to do, is using local coordinates and a parallel FFT Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 19 / 20

53 Extension to MHD Elsässer notation ( t v A k z ) wα = i k α k β k ( γ 2 k 2 w + β w γ + w β w + γ ik β wα w + β ν 2 k 2n wα ( t + v A k z ) w α + = i k α k β k ( ) γ 2 k 2 w + β w γ + w β w γ + ) ik β w + α w β ν 2 k 2n w + α The equations are treated similarly to the Navier-Stokes-equation. Two major changes: Two coupled equations Magnetic field introduced via v A k z term Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 20 / 20

54 [1] D. L. Brown, R. Cortez, and M. L. Minion. Accurate Projection Methods for the Incompressible Navier-Stokes Equations. Journal of Computational Physics, 168: , Apr Felix Spanier (Uni Würzburg) Simulation methods in astrophysics 20 / 20