Petr Krysl* Eitan Grinspun, Peter Schröder Hierarchical Finite Element Mesh Refinement *Structural Engineering Department, University of California, San Diego Computer Science Department, California Institute of Technology
Adaptive Approximations Adjust spatial resolution by: Remeshing Local refinement (Adaptive Mesh Refinement) Split the finite elements, ensure compatibility via Constraints Lagrangian multipliers or penalty methods Irregular splitting of neighboring elements Major implementation effort!
Refinement for Subdivision State-of-the-art refinement not applicable to subdivision surfaces. Refinement should take advantage of the multiresolution nature of subdivision surfaces. Subdivision surface: overlap of two basis functions.
Conceptual Hierarchy Infinite globally-refined sequence Mesh is globally refined to form and so on… Strict nesting of
Refinement Equation Refinement relation Refined basis of Any linearly independent set of basis functions chosen from with
Adapted basis 1 Quasi-hierarchical basis: Some basis functions are removed: Nodes associated with active basis functions
Adapted basis 2 True hierarchical basis Details are added to coarser functions: Nodes associated with active basis functions
Multi-level approximation Approximation of a function on multiple mesh levels Literal interpretation of the refinement equation has a big advantage: genericity. = set of refined basis functions on level m
CHARMS C onforming H ierarchica l A daptive R efinement M ethod S Refinement equation: Naturally conforming, dimension and order independent. Multiresolution: True hierarchical basis: Functions N (j+1) add details. Quasi-hierarchical basis: Functions N (j+1) replace N (j). Adaptation: Refinement/coarsening intrinsic (prolongation and restriction).
CHARMS vs common AMR CHARMS Level 0 Level 1 Original basis on quadrilateral mesh Adapted basis on a refined mesh Common AMR w/ constraints True hierarchical basis Quasi-hierarchical basis
Refinement for Subdivision CHARMS apply to subdivision surfaces without any change. The multiresolution character of subdivision surfaces is taken advantage of quite naturally. …
Algorithms Field transfer: prolongation, restriction operators. Integration: single level vs. multiple-level. Algorithms: independent of order, dimensions: generic; easy to program, easy to debug. Multiscale approximation: hierarchical and multiresolution (quasi- hierarchical) basis; multigrid solvers.
2D Example Hierarchy of basis function sets; Red balls: the active functions. Solution painted on the integration cells. Poisson equation with homogeneous Dirichlet bc. Quasi hierarchical basis. True hierarchical basis.
3D Example 3-level grid (true hierarchical) Solution painted on the integration cells
Heat diffusion: Hierarchical Level 1Level 2Level 3 Solution displayed on the integration cells Grid hierarchy True hierarchical basis; Adaptive step 2: 5,000 degrees of freedom (~3,000 hierarchical)
Heat diffusion: Quasi-hier. Level 1Level 2Level 3 Solution displayed on the integration cells Grid hierarchy Quasi-hierarchical basis; Adaptive step 2: 3,900 degrees of freedom
Highlights Easy implementation: The adaptivity code was debugged in 1D. It then took a little over two hours to implement 2D and 3D mesh refinement: Clear evidence of the generic nature of the approach. Expanded options: True hierarchical basis and multiresolution basis implemented by the same code: It takes two lines of code to switch between those two bases.
Onwards to … Theoretical underpinnings. Links to AVI’s, model reduction, wavelets,... Multiresolution solvers. Countless applications.