🔬“We have foundation models for language, not for physics” — Anima Anandkumar, Bren Professor of Computing

Integrating AI with physical world modeling through neural operators enables high-fidelity simulations that outperform traditional numerical methods. Neural operators generalize neural networks by learning mappings between function spaces, allowing for multi-scale resolution and efficient modeling of complex physical phenomena like fluid dynamics. Unlike physics-informed neural networks (PINNs) that often face difficult optimization landscapes, neural operators leverage data-driven training combined with physical constraints to achieve significant speedups. FourCastNet, a prominent example, democratizes weather forecasting by providing accurate, probabilistic predictions at a fraction of the computational cost required by supercomputers. Furthermore, incorporating domain-specific structures, such as spherical geometry for global weather, ensures stability in long-term rollouts. These advancements facilitate the creation of digital twins for systems like fusion reactors, moving the field toward foundation models capable of simulating, controlling, and optimizing diverse physical processes.