Caltech Researchers Develop Framework to Adapt AI for Continuous Physical Systems

Caltech researchers led by Bren Professor Anima Anandkumar have published a framework in Nature Machine Intelligence that extends standard neural network architectures to handle continuous scientific problems. The work addresses a fundamental mismatch: most AI models were designed for discrete data like text or images, but scientific phenomena such as weather, fluid dynamics, and fusion plasma evolve continuously across space and time.

The framework transforms conventional neural networks into neural operators, which learn the underlying continuous functions that govern physical systems rather than relationships between fixed grid points. Anandkumar first introduced neural operators in 2020, and they have since been applied to problems ranging from quantum chemistry to black hole simulations and carbon sequestration modeling.

In the new study, the team demonstrated a systematic method to modify existing architectures — originally built for computer vision and language tasks — so they naturally align with continuous physical domains. Lead authors Julius Berner, now at NVIDIA, and Miguel Liu-Schiaffini, a Caltech undergraduate at the time, showed that the adapted models maintain accuracy across resolutions they never encountered during training.

When tested on standard fluid dynamics simulations, neural operators trained on 128×128 grids remained accurate at both lower (64×64) and higher (256×256) resolutions. Conventional neural networks typically lose accuracy when resolution changes, requiring retraining or discrete approximations that can introduce errors between grid points.

The framework also achieved a million-fold speedup over traditional numerical simulations for predicting fusion plasma disruptions. By learning continuous operators, the models avoid the inaccuracies that arise from discretizing continuous phenomena into fixed grids.

Anandkumar described the work as a blueprint for adapting AI advances to scientific domains without designing new architectures from scratch. The research was supported by the Bren endowed chair, the Office of Naval Research, and the AI2050 Senior Fellow program at Schmidt Sciences.

The paper, "Principled approaches for extending neural architectures to function spaces for operator learning," includes co-authors from Caltech, NVIDIA, and Project Prometheus. Liu-Schiaffini received support from a Mellon Mays Undergraduate Fellowship.