Structure-Preserving Learning Improves Geometry Generalization in Neural PDEs
Researchers have developed a new data-driven finite element method called General-Geometry Neural Whitney Forms (Geo-NeW) to improve the generalization capabilities of neural partial differential equation (PDE) solvers. This method jointly learns a differential operator and compatible finite element spaces tailored to specific geometries. By explicitly connecting geometry to the solution through a transformer-based encoding and learned spaces, Geo-NeW provides a strong inductive bias that enhances performance on unseen domains and preserves physical conservation laws. AI
IMPACT Introduces a novel approach for neural PDE solvers, potentially improving scientific and engineering simulations across various geometries.