New AI models trained on physics, not words, are driving scientific discovery (www.cam.ac.uk)

Researchers from the Polymathic AI collaboration have introduced two innovative AI models, Walrus and AION-1, which are trained on extensive scientific datasets rather than on textual or photographic data. These foundational models leverage principles from physics to solve diverse problems across disciplines, including astronomy and fluid dynamics. For instance, Walrus can analyze phenomena from exploding stars to Wi-Fi signals by applying learned knowledge from one context to entirely different scenarios. This cross-disciplinary approach promises to accelerate scientific discovery, particularly beneficial for researchers facing limited data or resources. The significance of these models lies in their ability to generalize knowledge across various fields, making them versatile tools for scientists. Walrus is trained on a massive fluid dynamics dataset, while AION-1 utilizes astronomical survey data to enhance understanding of celestial objects, even from low-resolution observations. With open-source access to their code and data, the Polymathic AI team hopes to empower researchers to adapt these models for specific applications, streamlining their workflow and elevating the standard of accuracy in physics-based simulations. Overall, the introduction of Walrus and AION-1 marks a substantial advancement in the integration of AI within scientific research, fostering collaboration across disciplines.