Weight-adjusted DG methods for elastic wave propagation in arbitrary heterogeneous media

Transcription

1 Weight-adjusted DG methods for elastic wave propagation in arbitrary heterogeneous media Jesse Chan Department of Computational and Applied Math, Rice University ICOSAHOM 2018 July 12, 2018 Chan (CAAM) WADG 6/12/18 1 / 27

2 Collaborators and acknowledgements Prof. T. Warburton: Virginia Tech, Department of Mathematics Dr. Russell Hewett: TOTAL Research and Technology USA Prof. Maarten de Hoop: Rice University, Department of Computational Mathematics Khemraj Shukla: Oklahoma State University, Dept. of Geophysics. Chan (CAAM) WADG 6/12/18 2 / 27

3 High order DG methods for wave propagation Unstructured (tetrahedral) meshes for geometric flexibility. Low numerical dissipation and dispersion. High order approximations: more accurate per unknown. Explicit time stepping: high performance on many-core. Figure courtesy of Axel Modave. Chan (CAAM) WADG 6/12/18 3 / 27

4 High order DG methods for wave propagation Unstructured (tetrahedral) meshes for geometric flexibility. Low numerical dissipation and dispersion. High order approximations: more accurate per unknown. Explicit time stepping: high -0.8 performance on many-core Fine linear approximation. Chan (CAAM) WADG 6/12/18 3 / 27

5 High order DG methods for wave propagation Unstructured (tetrahedral) meshes for geometric flexibility. Low numerical dissipation and dispersion. High order approximations: more accurate per unknown. Explicit time stepping: high -0.8 performance on many-core Coarse quadratic approximation. Chan (CAAM) WADG 6/12/18 3 / 27

6 High order DG methods for wave propagation Unstructured (tetrahedral) meshes for geometric flexibility. Low numerical dissipation and dispersion. High order approximations: more accurate per unknown. Explicit time stepping: high performance on many-core. Max errors vs. dofs. Chan (CAAM) WADG 6/12/18 3 / 27

7 High order DG methods for wave propagation Unstructured (tetrahedral) meshes for geometric flexibility. Low numerical dissipation and dispersion. High order approximations: more accurate per unknown. Explicit time stepping: high performance on many-core. Graphics processing units (GPU). Chan (CAAM) WADG 6/12/18 3 / 27

8 Outline Weight-adjusted DG methods: acoustics 1 Weight-adjusted DG methods: acoustics 2 Extension to elastic wave propagation 3 Acoustic-elastic coupling, poroelasticity Chan (CAAM) WADG 6/12/18 3 / 27

9 Outline Weight-adjusted DG methods: acoustics 1 Weight-adjusted DG methods: acoustics 2 Extension to elastic wave propagation 3 Acoustic-elastic coupling, poroelasticity Chan (CAAM) WADG 6/12/18 3 / 27

10 Weight-adjusted DG methods: acoustics High order methods and arbitrary heterogeneous media (a) Mesh and exact c 2 (b) Piecewise const. c 2 (c) High order c 2 Efficient implementations on triangular or tetrahedral meshes assume piecewise constant wavespeed c 2. High order vs constant approx. of media: spurious reflections. This talk: generalized mass lumping for DG with provable stability, high order accuracy, simple numerical fluxes (central flux + penalty). Chan (CAAM) WADG 6/12/18 4 / 27

11 Weight-adjusted DG methods: acoustics Energy stable discontinuous Galerkin formulations Model problem: acoustic wave equation (pressure-velocity system) 1 p c 2 (x) t = u, u t = p Local formulation on element D k 1 p D k c 2 (x) t q = uq + 1 ([[u]] n + τ p [[p]]) q D k 2 D k u t v = p v + 1 ([[p]] + τ u [[u]] n) v D k 2 D k D k Energy stability (penalty terms weakly enforce continuity conditions) ( ) Dk p 2 t c 2 (x) + u 2 = τ p [[p]] 2 + τ u [[u n]] 2 0. k k D k Chan (CAAM) WADG 6/12/18 5 / 27

12 Weight-adjusted DG methods: acoustics Semi-discrete formulation and weighted mass matrices Weighted mass matrix M 1/c 2: SPD, induces weighted norm on p ( M1/c 2) ij = 1 D k c 2 (x) φ j(x)φ i (x) dx. Semi-discrete form: face mass and weak derivative matrices M f, S i. dp M 1/c 2 dt = Must build and factorize M 1/c 2 d i=1 S i u i + faces M du i dt = S ip + M f F u. faces c 2 constant and ( M 1/c 2) 1 = c 2 M 1 ). M f F p, separately over each element (unless Chan (CAAM) WADG 6/12/18 6 / 27

13 Weight-adjusted DG methods: acoustics Weight-adjusted DG: generalized mass lumping Weight-adjusted DG (WADG): M ( M 1/w ) 1 M is SPD and an energy stable approximation of weighted mass matrix. du M w dt M ( ) 1 du M 1/w M dt Bypasses inverse of weighted matrix (M w ) 1 = right hand side. ( M ( M 1/w ) 1 M ) 1 = M 1 M 1/w M 1. Low storage matrix-free application of M 1 M 1/w using quadrature. M = V T q WV q, V q = eval. at quad. pts, W = quad. weights (M) 1 M 1/w RHS = M 1 V T q W }{{} P q diag (1/w) V q (RHS). Chan (CAAM) WADG 6/12/18 7 / 27

14 Weight-adjusted DG methods: acoustics WADG and high order accuracy Generates norm with same equivalence constants w min u T Mu u T M w u w max u T Mu High order accuracy of weighted projection P w, WADG projection P w u P w u C w h N+1 w L 2 W N+1, u W N+1,2 P w u P w u C w,nh N+2 w L 2 W N+1, u W N+1,2. WADG retains high order accuracy for moments: if v P M v T M w u v T M 1 M 1/w M 1 u C w w W N+1, h 2N+2 M u W N+1,2 v L 2. Chan (CAAM) WADG 6/12/18 8 / 27

15 Weight-adjusted DG methods: acoustics WADG and high order accuracy Generates norm with same equivalence constants w min u T Mu u T M 1 M 1/w M 1 u w max u T Mu High order accuracy of weighted projection P w, WADG projection P w u P w u C w h N+1 w L 2 W N+1, u W N+1,2 P w u P w u C w,nh N+2 w L 2 W N+1, u W N+1,2. WADG retains high order accuracy for moments: if v P M v T M w u v T M 1 M 1/w M 1 u C w w W N+1, h 2N+2 M u W N+1,2 v L 2. Chan (CAAM) WADG 6/12/18 8 / 27

16 Weight-adjusted DG methods: acoustics WADG and high order accuracy Generates norm with same equivalence constants w min u T Mu u T M 1 M 1/w M 1 u w max u T Mu High order accuracy of weighted projection P w, WADG projection P w u P w u C w h N+1 w L 2 W N+1, u W N+1,2 P w u P w u C w,nh N+2 w L 2 W N+1, u W N+1,2. WADG retains high order accuracy for moments: if v P M v T M w u v T M 1 M 1/w M 1 u C w w W N+1, h 2N+2 M u W N+1,2 v L 2. Chan (CAAM) WADG 6/12/18 8 / 27

17 Weight-adjusted DG methods: acoustics WADG and high order accuracy Generates norm with same equivalence constants w min u T Mu u T M 1 M 1/w M 1 u w max u T Mu High order accuracy of weighted projection P w, WADG projection P w u P w u C w h N+1 w L 2 W N+1, u W N+1,2 P w u P w u C w,nh N+2 w L 2 W N+1, u W N+1,2. WADG retains high order accuracy for moments: if v P M v T M w u v T M 1 M 1/w M 1 u C w w W N+1, h 2N+2 M u W N+1,2 v L 2. Chan (CAAM) WADG 6/12/18 8 / 27

18 Weight-adjusted DG methods: acoustics Acoustic wave equation: variable coefficients (a) Weight w = c 2 (x, y) (b) Standard DG Figure: Standard vs. weight-adjusted DG with spatially varying c 2 containing both smooth variations and a discontinuity. Chan (CAAM) WADG 6/12/18 9 / 27

19 Weight-adjusted DG methods: acoustics Acoustic wave equation: variable coefficients (a) Weight w = c 2 (x, y) (b) Weighted-adjusted DG Figure: Standard vs. weight-adjusted DG with spatially varying c 2 containing both smooth variations and a discontinuity. Chan (CAAM) WADG 6/12/18 9 / 27

20 Weight-adjusted DG methods: acoustics Low storage WADG: efficiency on GPUs N = 1 N = 2 N = 3 N = 4 N = 5 N = 6 N = 7 Mw WADG Speedup Time (ns) per element: storing/applying M 1 1/w vs WADG (deg. 2N quadrature). Efficiency on GPUs: reduce memory accesses and data movement. (Tuned) low storage WADG faster than storing and applying M 1 1/w! Chan (CAAM) WADG 6/12/18 10 / 27

21 Outline Extension to elastic wave propagation 1 Weight-adjusted DG methods: acoustics 2 Extension to elastic wave propagation 3 Acoustic-elastic coupling, poroelasticity Chan (CAAM) WADG 6/12/18 10 / 27

22 Extension to elastic wave propagation Matrix-valued weights and elastic wave propagation Symmetric hyperbolic system: velocity v, stress σ = vec(s) with σ = (σ xx, σ yy, σ zz, σ yz, σ xz, σ xy ) T. ρ v t = d i=1 A T i σ x i, 1 σ C t = d i=1 A i v x i. Factoring out the constitutive stiffness tensor C results in simple and spatially constant matrices A i µ + λ λ λ C = λ 2µ + λ λ λ λ 2µ + λ, A 1 = µi }{{} for isotropic media Hughes and Marsden Classical elastodynamics as a linear symmetric hyperbolic system. Chan (CAAM) WADG 6/12/18 11 / 27

23 Extension to elastic wave propagation DG formulation for elasticity: energy stability D k D k Analogous to acoustics: numerical fluxes independent of media! ( ) 1 σ T d v C q = A i q + 1 ( ) A n [[v]] + τ σ A n A T n [[σ]] q, t x i 2 ( ρ v ) T w = t D k D k i=1 d i=1 A T i Energy method: take q = σ, w = v D k ( C 1 σ D k D k σ w + 1 ( ) A T n [[σ]] + τ v A T n A n [[v]] w. x i 2 D k ) T σ = t t D ( k ρ v ) T v = t t ( ) σ(x) T C 1 (x)σ(x) D k ρ(x) v 2. Chan (CAAM) WADG 6/12/18 12 / 27

24 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) d v A i q + 1 ( ) A n [[v]] + τ σ A n A T n [[σ]] q, x i 2 D k i=1 d A T i D k i=1 σ x i w D k D k ( ) A T n [[σ]] + τ v A T n A n [[v]] w Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

25 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) d v A T q ( i + A n {{v}} + τ ) σ x i 2 A na T n [[σ]] q, D k i=1 d A T i D k i=1 σ x i w D k D k ( ) A T n [[σ]] + τ v A T n A n [[v]] w Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

26 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) d v A T σ ( i + A n {{v}} + τ ) σ x i 2 A na T n [[σ]] σ, D k i=1 d A T i D k i=1 σ x i v D k D k ( ) A T n [[σ]] + τ v A T n A n [[v]] v Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

27 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) D k Sum volume terms, which cancel to zero + ) σ ( A n {{v}} + τ σ 2 A na T n [[σ]] ( A T n [[σ]] + τ v A T n A n [[v]] ) v Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

28 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) D k Sum volume terms, which cancel to zero + ) σ ( A n {{v}} + τ σ 2 A na T n [[σ]] ( A T n [[σ]] + τ v A T n A n [[v]] ) v Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

29 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) Sum over all elements: central flux terms cancel + τ σ 2 A na T n [[σ]] σ + τ v 2 AT n A n [[v]] v k D k Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

30 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) Sum over all elements: combine penalty flux terms+ τ σ 2 AT n [[σ]] A T n σ + τ v 2 A n [[v]] A n v k D k Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

31 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) Sum over all elements: combine penalty flux terms+ τ σ 2 AT n [[σ]] A T n [[σ]] + τ v 2 A n [[v]] A n [[v]] k D k Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

32 Extension to elastic wave propagation DG formulation for elasticity: energy stability, cont. A i constant, integration by parts if quadrature exact for degree (2N 1) Sum over all elements: combine penalty flux terms+ τ σ 2 AT n [[σ]] A T n [[σ]] + τ v 2 A n [[v]] A n [[v]] k D k... and done! Energy stability (penalty weakly enforces continuity conditions): 1 ρ v 2 + σ T C 1 σ = 2 t D k D k τ v faces f 2 A n [[v]] 2 + τ σ A T n [[σ]] τ v, τ σ > 0 penalizes normal jumps [[v]] n 0, [[S]] n 0. Chan (CAAM) WADG 6/12/18 13 / 27

33 Extension to elastic wave propagation Semi-discrete system: matrix-valued weighted Matrix-weighted mass matrix: let W R d d be SPD with entries w ij M w11... M w1d M W =..... M wd1... M wdd Semi-discrete DG formulation involves C 1 -weighted mass matrix Σ M C 1 t = M ρi V t = d (A i S i ) V + ( ) I M f F σ, faces d ) (A T i S i Σ + ( ) I M f F v. faces i=1 i=1 Chan (CAAM) WADG 6/12/18 14 / 27

34 Extension to elastic wave propagation Weight-adjusted DG: matrix-valued weights Matrix-weighted mass matrix large, hard to invert M C 1... M 11 C 1 1d M C 1 =..... M C 1 d1... M C 1 dd Weight-adjusted approximation for C 1 decouples each component M 1 C 1 ( I M 1) M C ( I M 1 ). Evaluate RHS components at quadrature points, apply C(x i ) to component vectors at quadrature points, project back to polynomials. Recovers Kronecker product M 1 C 1 = C M 1 for constant C 1. Chan (CAAM) WADG 6/12/18 15 / 27

35 Extension to elastic wave propagation Energy stability and DG spectra (a) Central flux (τ v, τ σ = 0) (b) Penalty flux (τ v, τ σ = 1) Figure: DG spectra with random heterogeneities at each quadrature point. Guaranteed energy stability for both energy conservative and energy dissipative numerical interface fluxes. CFL can be improved by setting τ σ 1/ C, τ v ρ Chan (CAAM) WADG 6/12/18 16 / 27

36 Extension to elastic wave propagation Elastic wave propagation: convergence Convergence for harmonic oscillation, Rayleigh, Lamb, and Stoneley waves: between O(h N+1 ) and O(h N+1/2 ). σ error grows as C 1 (e.g. incompressible limit λ/µ ). Relative L 2 error Mesh size h (a) L 2 errors (Stoneley wave) v error σ error λ/µ (b) C 1, N = 3, h = 1/8. Chan (CAAM) WADG 6/12/18 17 / 27

37 Extension to elastic wave propagation Elastic wave propagation: stiff inclusion Figure: tr(σ) and σ xy for stiff inclusion with N = 5, h 1/50. Chan (CAAM) WADG 6/12/18 18 / 27

38 Extension to elastic wave propagation Elastic wave propagation: anisotropy Simple implementation for anisotropy - fluxes independent of C. (a) Vertical velocity, t = 30µs (b) Vertical velocity, t = 60µs Anisotropic heterogeneous media: transverse isotropy (x < 0), isotropy (x > 0). Komatitsch, Barnes, Tromp Simulation of anisotropic wave propagation based upon a spectral element method. Chan (CAAM) WADG 6/12/18 19 / 27

39 Extension to elastic wave propagation Curved meshes and heterogeneous media Map curved elements D k to reference element ˆD, introduce determinant of Jacobian of mapping J (varies of ˆD). RHS terms: discretize skew-symmetric form d v A T i ˆD i=1 d A T i ˆD i=1 q ( J + A n {{v}} + τ ) σ x i D k 2 A na T n [[σ]] q, σ x i wj D k ( ) A T n [[σ]] + τ v A T n A n [[v]] w. Time-derivative terms: J incorporated into matrix weighting. D k ( C ) 1 σ T ( q = JC t ˆD ) 1 σ T q. t Chan (CAAM) WADG 6/12/18 20 / 27

40 Extension to elastic wave propagation Curved meshes and random heterogeneous media (b) Central flux (τ v, τ σ = 0) (c) Central flux (τ v, τ σ = 1) A skew-symmetric discretization guarantees discrete energy stability. Chan (CAAM) WADG 6/12/18 21 / 27

41 Extension to elastic wave propagation Elastic wave propagation: 3D isotropic media (a) Computational mesh (b) Homogeneous isotropic media Figure: tr(σ) with µ(x) = 1 + H(y) cos(3πx) cos(3πy) cos(3πz), N = 5. Chan (CAAM) WADG 6/12/18 22 / 27

42 Extension to elastic wave propagation Elastic wave propagation: 3D isotropic media (a) Computational mesh (b) Piecewise constant C(x) Figure: tr(σ) with µ(x) = 1 + H(y) cos(3πx) cos(3πy) cos(3πz), N = 5. Chan (CAAM) WADG 6/12/18 22 / 27

43 Extension to elastic wave propagation Elastic wave propagation: 3D isotropic media (a) Computational mesh (b) High order C(x) Figure: tr(σ) with µ(x) = 1 + H(y) cos(3πx) cos(3πy) cos(3πz), N = 5. Chan (CAAM) WADG 6/12/18 22 / 27

44 Outline Acoustic-elastic coupling, poroelasticity 1 Weight-adjusted DG methods: acoustics 2 Extension to elastic wave propagation 3 Acoustic-elastic coupling, poroelasticity Chan (CAAM) WADG 6/12/18 22 / 27

45 Acoustic-elastic coupling, poroelasticity Acoustic-elastic coupling Coupling conditions for fluid (acoustic) and solid (elastic) media: S n = pn, v n = u n. Replace [[p]] and [[σ]] with residuals of coupling conditions [[p]] = n (S + n ) p, A T n [[σ]] = [[S]] n = p + n S n, (acoustic side) (elastic side). Energy stable for arbitrary acoustic and elastic media. Chan (CAAM) WADG 6/12/18 23 / 27

46 Acoustic-elastic coupling, poroelasticity Acoustic-elastic coupling: arbitrary heterogeneous media (a) Low order c 2 (x), µ(x) (b) tr(σ) Figure: Acoustic-elastic waves from a Ricker pulse (N = 10, h = 1/16). Chan (CAAM) WADG 6/12/18 24 / 27

47 Acoustic-elastic coupling, poroelasticity Acoustic-elastic coupling: arbitrary heterogeneous media (a) High order c 2 (x), µ(x) (b) tr(σ) Figure: Acoustic-elastic waves from a Ricker pulse (N = 10, h = 1/16). Chan (CAAM) WADG 6/12/18 24 / 27

48 Acoustic-elastic coupling, poroelasticity Poroelasticity: low frequency Biot s system Biot s system adds pore pressure p and relative fluid velocity q. Symmetric first order system with augmented stress, velocity. σ = (σ xx, σ yy, σ zz, σ yz, σ xz, σ xy, p) T, ṽ d E v t = A T σ i Dṽ, x i i=1 ṽ = (v x, v y, v z, q x, q y, q z ) T E σ σ t = d i=1 A i ṽ x i. E s, E v SPD, A i spatially const. + sparse (a single ±1 per row). ρ ρ f ( E v = C ρ f m 1, E 1 C s = 1 ) α α T C 1 1 M + αt C 1, α Lemoine, Ou, LeVeque High-resolution finite volume modeling of wave propagation in orthotropic poroelastic media. Chan (CAAM) WADG 6/12/18 25 / 27

49 Acoustic-elastic coupling, poroelasticity Numerical results for Biot (with de Hoop, Shukla) (a) Orthotropic media, 1.56 ms (b) Epoxy-glass, 1.8 ms (c) Isotropic interface, 230 ms Inviscid Biot solution (mass particle velocity) showing fast P, slow S, and slow P waves. Simulation utilizes N = 3 and a uniform mesh of 128 elements per side. Systematic derivation of WADG formulations: can symmetrize a system by defining an appropriate convex entropy (energy function). Carcione Wave propagation in anisotropic, saturated porous media: plane wave theory and numerical simulation. Lemoine, Ou, LeVeque High-resolution finite volume modeling of wave propagation in orthotropic poroelastic media. Chan (CAAM) WADG 6/12/18 26 / 27

50 Acoustic-elastic coupling, poroelasticity Summary and acknowledgements Weight-adjusted DG (WADG) for acoustic and elastic wave propagation in heterogeneous media and on curved meshes. Generalized mass lumping: energy stability and high order accuracy. Significantly simpler formulations using symmetric forms of PDEs. This work is supported by DMS and DMS Thank you! Questions? Chan, Hewett, Warburton Weight-adjusted DG methods: wave propagation in heterogeneous media (SISC). Chan Weight-adjusted DG methods: matrix-valued weights and elastic wave prop. in heterogeneous media (IJNME). Chan (CAAM) WADG 6/12/18 27 / 27

51 Acoustic-elastic coupling, poroelasticity Additional slides Chan (CAAM) WADG 6/12/18 27 / 27

52 Acoustic-elastic coupling, poroelasticity Computational costs at high orders of approximation Note: WADG at high orders becomes expensive! Runtime (s) WADG runtimes Volume Surface Update Large dense matrices: O(N 6 ) work per tet. High orders usually use tensor-product elements: O(N 4 ) vs O(N 6 ) cost, but less geometric flexibility Degree N Idea: choose basis such that matrices are sparse. WADG runtimes for 50 timesteps, elements. Chan (CAAM) WADG 6/12/18 27 / 27

53 Acoustic-elastic coupling, poroelasticity BBDG: efficient volume, surface kernels Nodal DG Bernstein-Bezier DG 6 Volume Surface 6 Volume Surface Runtime (s) Degree N Degree N Bernstein-Bezier speedup requires constant coefficient RHS evaluation! Chan (CAAM) WADG 6/12/18 27 / 27

54 Acoustic-elastic coupling, poroelasticity BBDG: efficient volume, surface kernels BBDG speedup over nodal DG 8 Volume Surface 6 Speedup Degree N Bernstein-Bezier speedup requires constant coefficient RHS evaluation! Chan (CAAM) WADG 6/12/18 27 / 27

55 Acoustic-elastic coupling, poroelasticity BBWADG: polynomial multiplication and projection (a) Exact c 2 (b) M = 0 approximation (c) M = 1 approximation WADG reuses fast Bernstein volume and surface kernels. Fast O(N 3 ) Bernstein algorithms for polynomial multiplication: represent c 2 P M, p(x) P N, and construct c 2 (x)p(x) P M+N. Fast O(N 4 ) polynomial L 2 projection P N : P M+N P N. Chan (CAAM) WADG 6/12/18 27 / 27

56 Acoustic-elastic coupling, poroelasticity BBWADG: computational runtime for M = 1 Runtime (s) BBWADG BBWADG 10 6 N 4 N 4 WADG WADG 10 7 N 6 N (a) Degree N Degree N (b) Chan (CAAM) WADG 6/12/18 27 / 27

57 Acoustic-elastic coupling, poroelasticity BBWADG: update kernel speedup over WADG (acoustics) N = 3 N = 4 N = 5 N = 6 N = 7 N = 8 WADG 1.60e e e e e e-6 BBWADG 2.20e e e e e e-7 Speedup For N 8, quadrature (and WADG) becomes much more expensive. (a) N = 7 quadrature (b) N = 8 quadrature Chan (CAAM) WADG 6/12/18 27 / 27

58 Acoustic-elastic coupling, poroelasticity BBWADG: approximating c 2 and accuracy L 2 error piecewise constant c 2 2 piecewise linear c 2 4 quadratic c 2 5 WADG (N = 4) Mesh size h M = 0 M = 1 M = 2 Approximating smooth c 2 (x) using L 2 projection: O(h 2 ) for M = 0, O(h 4 ) for M = 1, O(h M+3 ) for 0 < M N 2. Chan (CAAM) WADG 6/12/18 27 / 27