Constructing A-spline weight functions for stable WEB-spline finite element methods

Chandrajit L. Bajaj, Radhakrishna Bettadapura, Na Lei, Alex Mollere, Chao Peng, Alexander RandSymposium on Solid and Physical Modeling 2010 · DOI

TL;DR

This paper introduces a novel method for constructing "weight functions" that are crucial for a powerful meshless numerical technique called the Weighted Extended B-spline (WEB-spline) finite element method. By leveraging algebraic splines (A-splines), they provide an explicit and computationally efficient way to generate these functions for complex 2D and 3D domains, leading to more stable and accurate simulations.

What Problem Does This Paper Address?

Traditional finite element methods (FEM) rely on discretizing complex geometries into a mesh of simpler elements. This meshing process can be incredibly time-consuming and computationally expensive, especially for intricate shapes. The WEB-spline method offers an alternative by decoupling the function space from the geometry. Instead of a mesh, it uses tensor-product B-splines on a grid to define the function space and a "weight function" to represent the domain's geometry. A key requirement for these weight functions is that they must be smooth, positive, and vanish at the domain's boundary at a specific rate. While previous methods for generating these weight functions exist, they often have limitations in terms of computational cost, complexity, or applicability to general domains. This paper aims to provide a more practical and generalizable approach to constructing these essential weight functions.

Methods

The core contribution of this paper is the development of a new method for constructing weight functions using algebraic splines (A-splines). The approach involves defining "scaffold regions" around the boundary of the domain. These regions are parameterized using local coordinates and are associated with boundary points and their inward-facing normal vectors, along with "scaffold sizes" that control the extent of these regions.

For 2D domains, the weight function is defined within these scaffold regions using cubic Hermite polynomials. The function is designed to smoothly transition from zero at the boundary to one within the domain. The method ensures that the resulting weight function is continuously differentiable (C¹).

For 3D domains, the method extends to handle quadrilateral and triangular surface meshes that define the boundary. Similar to the 2D case, scaffold regions are defined around the mesh elements. The weight function is then constructed by interpolating values and gradients along the edges of the mesh elements and blending them together using A-spline properties. This ensures a C¹ continuous weight function even for complex 3D surfaces.

Datasets

Not discussed in the paper. The paper focuses on a theoretical construction and validation using synthetic examples, not real-world datasets.

Key Insights and Ideas

  • A-splines as a Geometric Tool for FEM: The paper demonstrates that A-splines, traditionally used for curve and surface modeling, can be effectively adapted to construct crucial components for numerical methods like FEM. This opens up new avenues for leveraging geometric modeling techniques in scientific computing.
  • Explicit and Local Weight Function Construction: The proposed method provides an explicit formula for weight functions that only depend on local boundary information (nearby points, normals, and scaffold sizes). This is a significant advantage over methods that might require global information or complex iterative solvers.
  • C¹ Continuity is Key: The emphasis on constructing continuously differentiable (C¹) weight functions is crucial. This smoothness is shown to be important for achieving optimal convergence rates in the WEB-spline method and for avoiding numerical issues with quadrature.
  • Scaffold Size Matters: The paper highlights that the "scaffold size" (hs) plays a significant role in the accuracy of the solution. While a larger scaffold size can lead to better geometric approximation, it can also increase computational cost. Finding an appropriate balance is important.

Key Results

The paper presents a method for constructing C¹ continuous weight functions for general 2D and 3D domains. Through numerical examples, they demonstrate that their constructed weight functions, when used in the WEB-spline FEM, can achieve optimal convergence rates (O(h) for H¹ error) when the boundary mesh size (hb) and scaffold size (hs) are appropriately chosen relative to the grid size (hg). They show that their C¹ weight functions significantly outperform C⁰ weight functions in terms of error reduction. The paper also visually presents contour plots of the generated 2D and 3D weight functions, illustrating their smooth nature and how they adapt to the domain's boundary.

Limitations and Future Work

The authors acknowledge that the geometric error introduced by approximating the boundary with the weight function's zero-level-set is a factor in the overall solution error. While their C¹ weight functions improve upon C⁰ functions, the accuracy of the domain representation is still limited by the resolution of the boundary mesh (hb).

Future work could explore:

  • Higher-Order Continuity: Investigating methods to achieve C² or higher continuity for weight functions, which might further improve accuracy and stability.
  • Adaptive Scaffold Sizes: Developing strategies for adaptively choosing scaffold sizes based on local boundary curvature or other geometric features to optimize accuracy and efficiency.
  • Application to More Complex Problems: Testing the proposed weight function construction on a wider range of PDEs and more complex geometries to fully assess its robustness and applicability.
  • Relationship between Boundary Mesh and Scaffold Size: Further theoretical and empirical analysis to precisely quantify the optimal relationship between hb and hs for different types of domains and problems.

Summary updated

BibTeX
@inproceedings{bajaj2010constructing,
  author = {Chandrajit L. Bajaj and Radhakrishna Bettadapura and Na Lei and Alex Mollere and Chao Peng and Alexander Rand},
  title = {{Constructing A-spline weight functions for stable WEB-spline finite element methods}},
  booktitle = {Symposium on Solid and Physical Modeling},
  year = {2010},
  doi = {10.1145/1839778.1839800},
}
Figure