1 / 40

Monge Ampere Based Moving Mesh Methods for Weather Forecasting Chris Budd (Bath)

This talk discusses an r-adaptive moving mesh method for weather forecasting based on the Monge-Ampere equations, which allow for optimal placement and fast computation of mesh points. The method avoids mesh tangling and extreme skewness, works well in 2D and 3D, and can be easily coupled to CFD solvers and DA codes.

Download Presentation

Monge Ampere Based Moving Mesh Methods for Weather Forecasting Chris Budd (Bath)

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Monge Ampere Based Moving Mesh Methods for Weather Forecasting Chris Budd (Bath) Chiara Piccolo, Mike Cullen (Met Office) Emily Walsh (Bath,SFU), Phil Browne (Bath)

  2. In NWP often need to locally refine a mesh to capture small scales For accurate numerical computation eg. Storms For accurate assimilation of observed data to avoid spurious correlations • r-adaptive moving mesh methods aim to do this bythe ‘optimal placement’ of a fixed number of mesh points which move during the computation • There are many advantages with r-adaptivity • Constant data structures and mesh topology • Ease of coupling to CFD solvers • Capturing dynamical physics of the solution eg. Lagrangian behaviour, symmetries, conservation laws, self-similarity • Global control of mesh regularity Traditional problems with r-adaptivity: Mesh tangling, mesh skewness, implementation in 3D

  3. Talk will describe an r-adaptive moving mesh method for doing this based upon geometrical ideas: optimal transport theory and Monge Ampere equations • Will demonstrate that this leads to a fast, robust and effective moving mesh method for adaptive NWP • Which avoids mesh tangling and extreme skewness • Works well in 2D and 3D • Can be easily coupled to CFD solvers and DA codes H. Ceniceros and T Hou, J. Comp Phys (2001) CJB and J.F. Williams, Journal of Physics A, (2006) G.Delzano, L.Chacon, J. Finn, Y. Chung & G.Lapenta, J. Comp Phys (2008) CJB and J.F. Williams, SIAM J. Sci. Comp (2009) CJB, W-Z Huang and R.D. Russell, ACTA Numerica (2009) C.Kuhnlein, P.Smolarkiewicz and A.Dornbrack, J. Comp. Phys (2012) CJB, M.Cullen and E.Walsh, (2012) submitted

  4. Geometrical strategy Have a computational domain Physical domain Identify a map

  5. Determine F by Equidistribution Introduce a positive unit measureM(x,y,z,t) in the physical domain which controls the mesh density A: unit set in computational domain F(A,t) : image set Equidistribute integral with respect to this measure Equidistribution minimises the maximum value of this integral

  6. Differentiate to give: Basic, nonlinear, equidistribution mesh equation Choose M to concentrate points where needed without depleting points elsewhere: error/physics/scaling Note: All meshes equidistribute some function M [Radon-Nicodym]

  7. Choice of the monitor function M(X) • Physical reasoning • eg. Potentialvorticity, arc-length, curvature, • A-priori mathematical arguments • eg. Scaling, symmetry, error estimates (interpolation) • A-posteriori error estimates (primal-dual) • eg. Residuals, super-convergence • Data correlation estimates

  8. Mesh construction Problem: in two/three -dimensions equidistribution does NOT uniquely define a mesh! All have the same area Need additional conditions to define the mesh: GCL, Harmonic Maps, Tensor product equidistribution

  9. Optimally transported meshes F Want to avoid mesh tangling and long thin regions Argue: A good mesh for solving a pde is often one which is as close as possible to a uniform mesh

  10. Monge-Kantorovich optimal transport problem Minimise Subject to Also used in image registration,meteorology ….. Optimal transport helps to prevent small angles, reduce mesh skewness and prevent mesh tangling.

  11. Key result which makes everything work!!!!! • Theorem: [Brenier] • There exists a unique optimally transported mesh • (b) For such a mesh the map F is the gradient of a convex function P : Scalar mesh potential Map F is a Legendre Transformation

  12. It follows immediately in 2D that Hence the mesh equidistribution equation becomes (MA) Monge-Ampere equation: fully nonlinear elliptic PDE Global and local properties of the mesh can be deduced from the regularity of the solution of the MA equation

  13. Basic idea: Solve (MA) for P with appropriate (Neumann or Periodic) boundary conditions Good news: Equation has a unique solution Bad news: Equation is very hard to solve Good news: We don’t need to solve it exactly, and can instead use parabolic relaxation Q P Alternatively: Use Newton [Delzano et. al.] Use a variational approach [van Lent]

  14. Relaxation in n Dimensions: MMPDE [Russell] (PMA) Spatial smoothing [Hou] (Invert operator using a spectral method) Ensures right-hand-side scales like Q in nD to give global existence Averaged monitor Parabolic Monge-Ampere equation

  15. Solution Procedure If M is prescribed then the PMA equation can be discretised in the computational domain and solved using an explicit forward Euler method. This is a fast procedure: 5 mins for a 3D meteorological mesh • Applications • Image processing and image registration • Mesh generation for meteorological Data assimilation [Browne, CJB, Cullen, Piccolo] • Take M to be a scaled approximation of the Potential Vorticity of the 3D flow

  16. Because PMA is based on a geometric approach, it has a set of useful regularity properties 1.System invariant under translations, rotations, periodicity

  17. Lemma 1: CJB, EJW [2012] The solutions of the MA equation exactly align with global linear features Alignment follows from a close coupling between the local structure of the solution and the global structure. This is NOT a property of other mesh generation methods

  18. 2. Convergence properties of PMA • Lemma 2: [CJB,Williams 2006] • If M(x,t) = M(x) then PMA admits the solution • (b) This solution is locally stable/convergent and the mesh evolves to an equidistributed state Proof: Follows from the convexity of P which ensures that PMA behaves locally like the heat equation This result is important when initializing a mesh to the initial data for a PDE

  19. Lemma 3: [CJB,JFW 2006] If M(x,t) is slowly varying then the grid given by PMA is epsilon close to that given by solving the Monge Ampere equation. Lemma 4: [CJB,JFW 2006] The mapping is 1-1 and convex for all times: No mesh tangling or points crossing the boundary

  20. 4.For appropriate choices of M the coupled system is scale-invariant Lemma 5: [B,W 2005] Multi-scale property If there is a natural length scale L(t) then for careful choices of M the PMA inherits this scaling and admits solutions of the form Lemma 6: [B,W 2009] Multi-scale regularity The resulting scaled meshes have the same local regularity as a uniform mesh Extremely useful properties when working with PDEs which have natural scaling laws

  21. Coupling to a PDE • More usually M is a function of the solution of a PDE and we must couple mesh generation to a CFD calculation • Carefully discretise PDE & PMA in the computational domain:

  22. Solve the coupled mesh and PDE system either Method One As one large system(stiff!) Velocity based Lagrangian approach. Works well for parabolic blow-up type problems (JFW) Advantages: No need for interpolation Mesh and solution become one large dynamical system and can be studied as such Disadvantage: Equations are very hard to solve especially when the PDE is strongly advective

  23. Method 2: More suitable for metorology By alternating between evolving the PDE and mesh Time march the PDE on given mesh Evolve to new mesh by solving PMA to steady state Interpolate PDE solution onto the new mesh Repeat from 1. Advantages: Very flexible, can build in conservation laws and incompressibility at stage 2 Disadvantage: Interpolation is difficult and expensive

  24. Example 1: Parabolic finite time blow-up Length scale M is locally scale-invariant, concentrates points in the peak and keeps 50% of the points away from the peak

  25. Solve using Method 1 10 Solution: y x Mesh:

  26. Solution in the computational domain

  27. Example 2: Tropical storm formation (Eady problem [Cullen])

  28. Conjectured discontinuity singularity after t = 6.3 days

  29. M : Maximum eigenvalue of Potential Vorticity R • Solve using Method 2: • Finite difference method on a 60x60 Charney-Phillips mesh with pressure correction • 2nd order interpolation [Tang] with conservation law and geostrophic balancing • Update solution initially every 10 mins • Update mesh every hour • Reduce time step as singularity is approached

  30. Moving mesh gives good solution profiles

  31. Refining uniform mesh leads to solution oscillations

  32. Mesh profile

  33. Local mesh regularity is good

  34. Mesh Skewness is very good

  35. Conclusions • Optimal transport is a natural way to determine moving meshes in dimensions greater than one • It can be implemented using a fast relaxation process by using the PMA algorithm • Method works well for a variety of problems, and there are rigorous estimates about its behaviour • Looking good on meteorological problems • Lots of work to do to compare its effectiveness with tried and tested AMR procedures on standard test problems • Maybe a good project for the INI programme?

More Related