An international team of scientists has launched a new research collaboration that aims to develop an AI-powered tool for scientific discovery using the same technology that underpins ChatGPT. While ChatGPT deals with words and sentences, the new Polymathic AI initiative will learn from numerical data and physics simulations from different fields of science to help scientists model everything from supergiant stars to the Earth’s climate.
“This is going to revolutionise the way people use AI and machine learning in scientific research,” said Shirley Ho, a group leader at New York City’s Flatiron Institute’s Centre for Computational Astrophysics.
Ho compared the concept behind Polymathic AI to how learning a new language is made simpler when you already know five languages. It can be quicker and more accurate to build a scientific model from scratch rather than starting with a sizable, pre-trained foundation model.
That may be the case even if the training set’s connections to the issue at hand aren’t immediately apparent.
“Polymathic AI can reveal commonalities and connections between different fields that we may have missed,” said co-author Siavash Golkar, a guest researcher at the Flatiron Institute’s Centre for Computational Astrophysics.
The Polymathic AI team includes specialists in physics, astrophysics, mathematics, artificial intelligence, and neuroscience. The project from Polymathic AI, according to its developers, will learn using data from different places across physics and astrophysics (and eventually, fields like chemistry and genomics), and apply that multidisciplinary know-how to a variety of scientific issues.
Accuracy-wise, ChatGPT has well-known limitations. According to Ho, many of these pitfalls will be avoided by Polymathic AI’s project because it treats numbers as actual numbers rather than simply words on the same scale as letters and punctuation. Additionally, actual scientific datasets that depict the laws of physics underlying the cosmos will be used as training data.
According to Ho, the project places a high priority on transparency and openness. “We want to release all information. We want to democratise artificial intelligence (AI) for science so that, in a few years, we can offer the community a pre-trained model that can enhance scientific investigations throughout a wide range of problems and domains.