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

Anima Anandkumar discusses her pioneering work in developing neural operators for physical systems, particularly weather forecasting. She highlights how traditional physics-based models require significant computational resources and are limited by decades of development. Her team created FourCastNet, an AI model that accurately predicts weather and climate phenomena far faster and more efficiently, using consumer-grade GPUs. Incorporating physical domain knowledge like spherical geometry significantly enhances stability and accuracy, enabling the model to project climate over longer timescales, thus democratizing access to high-fidelity weather predictions globally.

Anima explains the concept of neural operators, a mathematical framework blending data-driven learning with physical laws, allowing models to handle multi-scale inputs and outputs. She emphasizes that this approach revolutionizes modeling continuous systems like atmospheric and plasma physics by shifting from grid-based methods to function-based models. These models are capable of accommodating high-resolution data, physical priors, and complex geometries (such as the Earth's sphere), making it possible to simulate and predict complex phenomena like fusion plasma disruptions and fluid flows with remarkable speed and fidelity.

The discussion covers the integration of physics into AI models, for example through Fourier neural operators operating in the frequency domain, which capture non-local phenomena efficiently. Anima highlights tools like TorchLean, which enables formal verification of neural networks within proof assistants, critical for safety-critical applications such as controlling nuclear reactors or fusion devices. The conversation underscores the importance of principled AI that respects physical laws, especially when scaling models for tasks like climate prediction or complex geometric simulations.

Anima broadens the scope of neural operators to various scientific and engineering disciplines—ranging from material design and electromagnetic simulation to biological systems and quantum physics. She advocates for building 'foundation models' for physics that can generalize across multiple phenomena and enable inverse design and optimization. Emphasizing the need for structure and inductive biases in AI, she envisions future models capable of handling coupled multi-physics problems, thus advancing scientific discovery and engineering innovation.