Reproduce “Geometry aware inference of steady state PDEs using Equivariant Neural Fields representations” on OpenResearch

About this paper

Giovanni Catalani, Michael Bauerheim, Frédéric Tost, Xavier Bertrand, Joseph Morlier

Developed by Airbus and ISAE-Supaero, enf2enf is an AI framework that leverages Equivariant Neural Field representations for geometry-aware inference of steady-state Partial Differential Equations (PDEs) across variable geometries. It preserves local geometric structure to achieve up to 5 orders of magnitude speedup over traditional CFD, along with superior accuracy on complex aerodynamic and material benchmarks, particularly for localized physical phenomena.

Key insights

• Spatially-anchored local latent representations, derived through meta-learning, effectively capture and preserve intricate local geometric details essential for accurate PDE solutions across variable geometries.
• The translation-equivariant neural field architecture facilitates true discretization invariance, enabling models trained on lower-resolution data to maintain high accuracy during inference on unseen, high-resolution meshes.
• Directly combining local geometric features with global operating conditions in the decoder, enhanced by latent self-attention, offers superior performance in predicting localized physical phenomena compared to approaches relying solely on global latent embeddings.

Summary by alphaXiv.