news and events

An international team of scientists, in collaboration with the University of Cambridge, has embarked on a novel research endeavor leveraging the technology that underpins ChatGPT to develop an AI tool for scientific exploration. Unlike ChatGPT, which focuses on text, this AI will specialize in learning from numerical data and physics simulations, spanning a multitude of scientific disciplines to assist scientists in modeling complex phenomena, from supergiant stars to Earth's climate.
The project, dubbed Polymathic AI, was unveiled recently alongside the release of a series of associated research papers on the open-access repository, arXiv.

Polymathic AI's principal investigator, Shirley Ho, who is also a group leader at the Flatiron Institute's Center for Computational Astrophysics in New York City, commented, "This will revolutionize the use of AI and machine learning in scientific research."
The concept behind Polymathic AI draws parallels to language learning. Commencing with a substantial pre-trained model, referred to as a foundation model, often proves quicker and more accurate than crafting a scientific model from the ground up. This holds true even when the training data appears unrelated to the specific problem at hand.

Miles Cranmer, co-investigator from the University of Cambridge, expressed, "Undertaking academic research with full-scale foundation models has been challenging due to the extensive computing power required. Our collaboration with the Simons Foundation has provided us with invaluable resources to prototype these models for basic scientific use, with the potential for global adoption."
Polymathic AI will unveil commonalities and connections across diverse scientific fields, thereby fostering cross-disciplinary insights that might otherwise remain hidden.

The Polymathic AI team comprises experts from the Simons Foundation, the Flatiron Institute, New York University, the University of Cambridge, Princeton University, and the Lawrence Berkeley National Laboratory, encompassing a spectrum of fields such as physics, astrophysics, mathematics, artificial intelligence, and neuroscience.
While AI has been used in scientific research previously, these solutions have typically been purpose-built and trained using domain-specific data. Polymathic AI seeks to transcend these constraints by harnessing data from a range of sources within physics, astrophysics, and, eventually, fields like chemistry and genomics, aiming to apply its multifaceted expertise to a wide array of scientific problems.

Transparency and accessibility are core tenets of the project, as Ho emphasized, "We aspire to make all aspects of this project public. Our goal is to democratize AI for scientific research so that, in the coming years, we can provide the scientific community with a pre-trained model capable of enhancing scientific analyses across diverse problem domains and fields."
To enhance accuracy, Polymathic AI will avoid some of the common limitations of AI models like ChatGPT. For instance, it will treat numbers as numerical values, not just as characters, and will incorporate authentic scientific datasets capturing the fundamental physics underlying various natural phenomena.