Data-Efficient PDE Pre-training via Geometry and Physics Drivers

A new pre-training framework for unstructured 3D neural PDE solvers eliminates the need for massive pre-computed datasets. It uses intrinsic shape descriptors for steady-state problems and online synthetic data generation for transient regimes to improve generalization and scalability.