Science and technology, UK (Commonwealth Union) – A new research collaboration, featuring scientists from the University of Cambridge and international partners, has initiated the development of an AI-powered tool for scientific exploration. Inspired by the technology driving ChatGPT, this endeavor, named Polymathic AI, diverges from language processing to harness numerical data and physics simulations across diverse scientific domains. The goal is to facilitate scientists in modeling phenomena ranging from supergiant stars to Earth’s climate as indicated by the researchers of the study.
Polymathic AI, launched this week, aims to revolutionize the application of AI and machine learning in scientific pursuits. Shirley Ho, the principal investigator of Polymathic AI and a group leader at the Flatiron Institute’s Center for Computational Astrophysics in New York City, emphasized that the initiative’s concept is akin to learning a new language more easily when equipped with knowledge of multiple languages.
The project begins with a foundational model, a large pre-trained model, making the process of constructing scientific models faster and more accurate compared to building them from the ground up. Even when the training data may not seem directly relevant, leveraging a pre-trained model offers efficiency. Overcoming the computational challenges associated with large-scale foundation models, Miles Cranmer, co-investigator from the University of Cambridge, Department of Applied Mathematics and Theoretical Physics and Institute of Astronomy, highlighted the collaboration with the Simons Foundation, providing unique resources for prototyping models applicable in basic science research worldwide. The prospect is deemed exciting, marking a significant leap in the integration of AI into scientific exploration.
“Polymathic AI can show us commonalities and connections between different fields that might have been missed,” explained co-investigator Siavash Golkar, a guest researcher at the Flatiron Institute’s Center for Computational Astrophysics. “In previous centuries, some of the most influential scientists were polymaths with a wide-ranging grasp of different fields. This allowed them to see connections that helped them get inspiration for their work. With each scientific domain becoming more and more specialised, it is increasingly challenging to stay at the forefront of multiple fields. I think this is a place where AI can help us by aggregating information from many disciplines.”
The Polymathic AI team comprises researchers from various institutions, including the Simons Foundation and its Flatiron Institute, New York University, the University of Cambridge, Princeton University, and the Lawrence Berkeley National Laboratory. This interdisciplinary team brings together experts in physics, astrophysics, mathematics, artificial intelligence, and neuroscience.
While AI tools have been utilized by scientists in the past, these tools have typically been purpose-built and trained on specific, relevant data. Francois Lanusse, a cosmologist at the Centre national de la recherche scientifique (CNRS) in France and co-investigator, highlighted the limitations of such approaches, indicating that despite rapid progress of machine learning in recent years in various scientific fields, in almost every cases, machine learning solutions are formed for specific use cases and trained on some very specific data. This compartmentalization creates barriers within and between disciplines, hindering scientists from leveraging information that may exist in different formats or fields.
In contrast, Polymathic AI’s project aims to break down these barriers by learning from diverse data sources across physics and astrophysics, with plans to expand into fields like chemistry and genomics. The project’s approach involves applying this multidisciplinary knowledge to address a broad spectrum of scientific problems. Mariel Pettee, a postdoctoral researcher at Lawrence Berkeley National Laboratory and a member of the project, expressed the goal of connecting seemingly disparate subfields into a cohesive framework that transcends the individual contributions, creating something greater than the sum of its parts.
The extent to which we can navigate these leaps between disciplines remains uncertain, according Ho remarked, further indicating that their objective is to actively pursue and facilitate these connections.
