Scientific research draws on many kinds of information, from numerical datasets to physical simulations. Polymathic AI is an international initiative to train models across scientific fields, with the goal of giving researchers a flexible starting point for new analyses.
Building a foundation for scientific machine learning
Foundation models have mainly been used in image and language processing. Polymathic AI aims to apply that approach to scientific machine learning, where researchers often build models for particular problems from scratch.
The initiative’s premise is that a pre-trained model may be faster and more accurate than starting from nothing, even when its training data does not directly match the task at hand. Scientists could then refine the model for a specific application.
The work was announced alongside related scientific papers on arXiv.org, an open-access repository. The team includes researchers from the Simons Foundation and the Flatiron Institute, New York University, the University of Cambridge, Princeton University, and Lawrence Berkeley National Laboratory. Their expertise spans physics, astrophysics, mathematics, artificial intelligence, and neuroscience.
Working with scientific data in its own form
Scientific information does not share a single standard format like text does in natural language processing. Research datasets can differ from one another and come from distinct disciplines, making it difficult for a model to learn across them.
Polymathic AI treats numbers as real values rather than as characters, such as letters and punctuation. Its training data includes scientific datasets that capture the underlying physics of the cosmos, alongside numerical information and physical simulations from different fields.
The project is intended to support modeling phenomena such as giant stars and the Earth's climate. These examples illustrate the ambition: train a model to learn from scientific information in ways that can be useful beyond one narrowly defined task.
Connecting disciplines, with practical hurdles
Project director Shirley Ho, group leader at the Flatiron Institute's Center for Computational Astrophysics in New York, said the project would “completely change how people use AI and machine learning in science.” The team’s broader aim is to help researchers use knowledge drawn from several fields, rather than rely only on tools built for one use case or one dataset.
Co-initiator Siavash Golkar, a visiting scientist at the Flatiron Institute's Center for Computational Astrophysics, sees the effort as a way to help scientists discover connections between disciplines. A model that brings together information from different areas could help researchers navigate work beyond their own specialties.
That goal comes with a resource challenge. Miles Cranmer, a co-initiator and member of the Department of Applied Mathematics and Theoretical Physics and the Institute of Astronomy at the University of Cambridge, points to the computational cost of using foundation models in academic work. He says the collaboration with the Simons Foundation provides resources to test models for scientific research.
Openness and the limits of earlier efforts
Co-initiator François Lanusse, a cosmologist at the Centre national de la recherche scientifique (CNRS) in France, describes Polymathic AI as intended to draw on data from varied sources and domains, rather than being confined to particular use cases. The aim is to apply multidisciplinary knowledge to a wide range of scientific problems.
Ho has also emphasized transparency and openness, with the stated goal of democratizing AI for science. A pre-trained model that researchers can adapt could make it easier to explore different questions and domains, though the project’s intended benefits depend on how well its models work in practice.
The source article contrasts this effort with Meta's open-source AI model Galactica, which was optimized for scientific tasks but relied primarily on language for training. After researchers criticized misinformation it generated, Meta took it offline after a few days. Polymathic AI’s focus on numerical scientific data and simulations reflects a different approach to building AI for research.