Reproduce “GLOBE: Accurate and Generalizable PDE Surrogates using Domain-Inspired Architectures and Equivariances” on OpenResearch

About this paper

Peter Sharpe

GLOBE, a neural surrogate model, combines boundary-element methods and equivariant machine learning to accurately solve partial differential equations, achieving up to a 200-fold reduction in prediction error over baselines and maintaining robust performance even with highly imperfect input meshes. This model demonstrates strong generalization to unseen flow conditions while being computationally efficient for industrial engineering applications.

Key insights

• Structurally embedding exact physical constraints and symmetries into ML architectures greatly improves data efficiency and out-of-distribution generalization for physics-based problems.
• The boundary-element-inspired formulation provides inherent robustness to severe geometric imperfections and non-watertight meshes, simplifying data preparation and enabling direct use of CAD-derived geometries.
• A hybrid approach, integrating known physical structures with learned components, offers a more effective path for scientific machine learning than purely general-purpose methods.

Summary by alphaXiv.