An international team of experts in different disciplines and belonging to different research centers have joined forces to develop a new artificial intelligence tool aimed at scientific discovery taking advantage of the same technology behind ChatGPT.
While OpenAI's chatbot works with words and phrases, these scientists' AI - called Polymathic AI - is designed to process numerical data and physical simulations from all fields of science. The goal is to allow researchers to model all types of phenomena, from supergiant stars to Earth's climate.
Polymathic AI principal investigator Shirley Ho, who works at the Center for Computational Astrophysics at the Flatiron Institute in New York, says that this tool will change the way artificial intelligence and machine learning are used in the world of scientific research. .
This is because because Polymathic AI handles numbers as values instead of characters, it will avoid the precision limitations associated with language-based models. Furthermore, since this system will learn using data from various sources in physics, astrophysics, and probably also chemistry and genomics, it will be able to apply that multidisciplinary knowledge to a wide range of scientific problems, overcoming the boundaries between disciplines that normally exist. in the world of research and in other artificial intelligence tools.
"Despite the rapid progress of machine learning in recent years in various scientific fields, in almost all cases machine learning solutions are developed for a specific use and trained on some very specific data; this creates boundaries both within and between disciplines, which means that scientists who use AI for their research do not benefit from information that may exist in a different format or in another field of research other than their own, explains Francois Lanusse, cosmologist at CNRS (France), in the statement with which the University of Cambridge has announced the Polymathic AI initiative.
And the intention of the team of scientists who build Polymathic AI is that their tool does allow leaps between disciplines. The project "will connect many seemingly disparate subfields into something greater than the sum of the parts," in the words of Mariel Pettee, a postdoctoral researcher at Lawrence Berkeley National Laboratory.
"Our goal is to accelerate the development of versatile basic models designed for numerical data sets and scientific machine learning tasks. The challenge we are taking on is to build AI models that leverage information from heterogeneous data sets and from different scientific fields that, to Unlike domains such as natural language processing, they do not share a unified representation (that is, text)", explained the promoters of the initiative when announcing their project.
And they noted that these models they are developing can be used "as solid baselines or scientists can refine them for specific applications; this approach has the potential to democratize AI in science by providing ready-to-use models that have more solid track records ( that is, prior knowledge) for general shared concepts like causality, measurement, signal processing, and even more specialized shared concepts like waves, something that would otherwise have to be learned from scratch.”
In this sense, one of the members of the group, Siavash Golkar, assures that the artificial intelligence they are training will be able to show common points or connections between different disciplines that could have been overlooked until now due to the growing specialization within each. scientific domain, with research groups increasingly focused on specific topics.
The Polymathic AI team includes expert researchers in physics, astrophysics, mathematics, artificial intelligence and neuroscience from the Simons Foundation and its Flatiron Institute, the universities of New York, Cambridge, Princeton and Lawrence Berkeley National Laboratory, among other institutions.
