Reproduce “Physics-Informed Geometry-Aware Neural Operator” on OpenResearch

About this paper

Weiheng Zhong, Hadi Meidani

The Physics-Informed Geometry-Aware Neural Operator (PI-GANO) provides a method for learning solutions to parametric partial differential equations that simultaneously generalizes across variable physical parameters and diverse domain geometries. It achieves this without requiring high-fidelity training data, demonstrating improved accuracy and efficiency over prior physics-informed approaches that addressed only one type of variability.

Key insights

• Encoding geometry using collocation points only from variable boundary segments, combined with concatenating global and local embeddings, proved most effective for accuracy in the geometry-aware module.
• The novel geometry-coordinate encoder design facilitates more stable and accurate physics-informed training by ensuring a direct derivative path for local coordinates that bypasses pooling layers.
• Previous physics-informed neural operators could either generalize across varying PDE parameters or varying domain geometries, but not effectively both simultaneously in a data-free manner, identifying a critical research gap.

Summary by alphaXiv.