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

Caltech professor and researcher Anima Anandkumar developed FourCastNet, an open-source AI weather model that competes with the best traditional physics-based simulations while running on consumer-grade GPUs. While the artificial intelligence community has largely focused on language and text, modeling continuous physical systems like weather, fusion, and fluid flow has been neglected because it is chaotic, multi-scale, and lacks massive open-source datasets.

Unlike language models that rely entirely on massive text tokens and raw computing scale, physical modeling requires handling context lengths reaching the hundreds of billions of grid points. To overcome this limitation, Anandkumar pioneered neural operators, which combine data with actual physical laws to model functions that evolve across multiple scales rather than just processing basic grids. For example, using spherical harmonics to account for the earth being a sphere allows global weather models to remain stable and run far into the future instead of quickly breaking down.

Applying neural operators to other physical domains reveals that the physical world is surprisingly forgiving. In nuclear fusion, a few thousand samples are enough to predict plasma disruptions a million times faster than traditional simulations. Progress in this field comes from baking physical structure and inductive biases directly into neural networks rather than waiting for nonexistent training data, with the ultimate goal of creating a comprehensive foundation model for physics.

Beyond weather prediction and neural operators, Anandkumar has worked on integrating neural networks with automated proof techniques. This includes TorchLean, a framework for writing PyTorch-style networks inside the Lean proof assistant to formally verify them, which is critical for safety-critical applications like fusion reactor control loops. Anandkumar was also recently appointed to the United Nations Scientific Advisory Board to help bring evidence-based scientific viewpoints to global policy.