Learning Discrete Riemannian Metric for Problem Aware Meshing
Abstract To solve PDE efficiently, posterior based standard Adaptive Mesh Refinement (AMR) techniques are widely used, but it requires a decent initial mesh for the first “cold start” computation. For steady‐state elliptic PDEs on fixed curved surfaces with varying source terms or constraints, we propose a novel supervised learning method to infer a Riemannian metric field to guide the initial mesh generation. Without requiring knowledge of the solution, from the source term or constraints, our method is able to provide a high‐quality “warm start” for AMR based on the “prior” learned from the dataset. Representing metric by tensor coefficients brings dependence of tangent bases and hinders the generality. Therefore, we encode the metric field using discrete edge lengths. Besides, we propose a solver‐agnostic synthesis strategy to design a scalable dataset encoding the prior from pre‐solved PDE instances. Our method demonstrates significant improvements on two common PDE problems: in Poisson problems, predicted metric‐guided meshes reduce absolute discretization error by 27–43 % relative to uniform meshes at matched resolution across our three test domains; in Obstacle problems, we achieve 2–3 × faster convergence, confirming the benefit of integrating our “warm start” into existing AMR workflows.
Authors
- Xiaohu Zhang (ORCID: https://orcid.org/0009-0009-0988-3288)
- Hujun Bao (ORCID: https://orcid.org/0000-0002-2662-0334)
- Jin Huang (ORCID: https://orcid.org/0000-0002-2549-6810)
- Yiyang Jin (ORCID: https://orcid.org/0009-0001-9408-2662)
Institutions
- Zhejiang University (CN)
Publication Details
- Journal
- Computer Graphics Forum
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1111/cgf.70606
- Primary Topic
- Computational Geometry and Mesh Generation
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China