Hodge decomposition of vector fields in Cartesian grids

Guo-Wei Wei, Zhe Su, Yiying Tong, Guowei Wei 0001Siggraph Asia 2024 · DOI

TL;DR

This paper introduces a novel method for performing a 5-component Hodge decomposition of vector fields directly on Cartesian grids, avoiding the need for mesh generation. This approach simplifies the process and maintains topological consistency, making it more efficient and applicable to a wider range of problems in geometric modeling, fluid dynamics, and scientific visualization.

What Problem Does This Paper Address?

The Hodge decomposition is a powerful mathematical tool that breaks down vector fields into fundamental components (e.g., divergence-free and curl-free). Traditionally, this decomposition has been applied to domains represented by meshes (like triangles or tetrahedra). However, applying it to domains defined by implicit surfaces (like level-set functions) on Cartesian grids presents challenges. Specifically, projecting vector fields onto the interior or boundary of the domain becomes computationally complex without explicit mesh structures. Existing methods often rely on simplicial meshes, requiring costly tessellation and data conversion. This paper addresses the gap of efficiently performing a complete Hodge decomposition on Cartesian grids for implicitly defined domains, without the overhead of mesh generation.

Methods

The core contribution of this paper is a new framework for Hodge decomposition on Cartesian grids. Instead of relying on simplicial meshes, the authors leverage the structure of regular grids and adapt the principles of discrete exterior calculus (DEC).

Key aspects of their method include:

  • Direct Discretization on Grids: They define discrete forms (0-, 1-, 2-, and 3-forms) directly on grid cells, edges, and vertices, using integrals of vector fields over these elements.
  • Discrete Differential Operators: They construct discrete versions of the exterior derivative (d) and codifferential (δ) operators that act on these grid-based forms. These operators are represented by sparse matrices.
  • Discrete Hodge Star: A discrete Hodge star operator is defined to relate primal and dual grids, enabling the construction of discrete Laplacians.
  • Handling Boundary Conditions: A novel approach to enforcing normal and tangential boundary conditions is introduced using "compact supports" (normal and tangential supports) and projection matrices. This avoids explicit tessellation of boundary cells.
  • Modified Hodge Stars and Laplacians: To accommodate the different supports and boundary conditions, modified discrete Hodge stars and Laplacians are developed.
  • 5-Component Decomposition: The framework allows for the computation of the five orthogonal components of the Hodge decomposition: normal gradient, tangential curl, normal harmonic, tangential harmonic, and curly gradient fields. This is achieved through a series of linear system solves and projections.

Datasets

The paper uses synthetic datasets representing various geometric shapes (e.g., bunny, kitty, kettlebell, Buddha, figure-8) defined by level-set functions on regular Cartesian grids. These domains are used to demonstrate the effectiveness of the proposed Hodge decomposition method. No new benchmarks are introduced; rather, existing geometric models are utilized.

Key Insights and Ideas

  • Cartesian Grids are Sufficient: The paper demonstrates that a full Hodge decomposition can be achieved directly on Cartesian grids, eliminating the need for complex mesh generation and conversion. This is a significant simplification for many applications.
  • Topology is Preserved: The discrete exterior calculus framework, adapted for Cartesian grids, rigorously preserves the topological properties of the domain, which is crucial for the correctness of the Hodge decomposition.
  • Boundary Conditions Simplified: The concept of "normal" and "tangential" supports, along with modified Hodge stars, provides an elegant way to handle boundary conditions without explicit boundary meshing.
  • Efficiency Gains: By avoiding mesh generation and working directly with grid structures, the method offers potential for significant computational efficiency, especially for dynamic or evolving domains.

Key Results

The paper successfully demonstrates a 5-component Hodge decomposition of vector fields on various 2D and 3D domains embedded in Cartesian grids. Numerical experiments show that the decomposed components are pairwise orthogonal (within numerical precision) and accurately reflect the topology of the underlying domains. The authors report that their algorithm runs efficiently, with runtimes within seconds for moderately sized grids. The effectiveness is visually demonstrated through figures showing the decomposition of vector fields on complex shapes.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Domain Representation: The current framework is limited to domains represented by level-set functions on Cartesian grids. It cannot capture sharp features unless the grid resolution is sufficiently high.
  • Periodic Domains and Obstacles: Periodic domains with obstacles are not yet implemented.
  • Mixed Boundary Conditions: The current approach considers normal or tangential boundary conditions separately. Future work could generalize this to mixed boundary conditions.
  • Higher Dimensions: Extending the framework to higher dimensions is suggested as a relatively straightforward future step.
  • Adaptive Data Structures: To handle geometric details better, incorporating adaptive data structures like octrees is proposed.
  • High-Order Hodge Stars: Exploring high-order Galerkin-type Hodge stars instead of diagonal ones could be investigated for potential improvements in convergence rate.
  • Optimization: Investigating speedups for saddle point problems, such as through Schur complement reduction, is suggested.

Summary updated

BibTeX
@inproceedings{wei2024hodge,
  author = {Guo-Wei Wei and Zhe Su and Yiying Tong and Guowei Wei},
  title = {{Hodge decomposition of vector fields in Cartesian grids}},
  booktitle = {Siggraph Asia},
  year = {2024},
  doi = {10.1145/3680528.3687602},
}
Figure