Researchers at the Flatiron Institute, a New York–based scientific research institute, and Polymathic AI, which develops AI models for scientific research, have built two physics-driven foundation models. Much like large language models learn statistical structure in language by predicting missing words from text, these models learn the mathematical structure of physical systems by reconstructing missing or obscured scientific measurements. One model, AION-1, focuses on astronomy and learns from large collections of images and spectra of stars while the other, Walrus, focuses on fluid and fluid-like systems across many physical regimes. This enables knowledge learned in one physical context to be reused in another—for instance, Walrus can apply patterns learned from small-scale fluid systems, such as bacterial motion in a liquid, to infer the behavior of gas and plasma in extreme settings like stellar explosions, without training a new model from scratch.
Image Credits: Jeremy Thomas
