Oct 15 : European physics research centre CERN and Mistral AI have established a long-term strategic partnership to develop a foundation model for scientific discovery, the two organisations said on Thursday.
The statement of intent gives the French AI company access to one of the world's largest scientific datasets as it seeks to compete on frontier science.
Mistral Chief Executive Arthur Mensch wrote in an article published in Le Monde on Thursday that frontier models were "now indispensable to scientific progress" and criticised companies that keep their advantages to themselves.
CERN Chief Information Officer Enrica Porcari said the new partnership would "advance AI capabilities that can serve the whole physics community and be shared".
The partners aim to build a "Large Physics Model", called "Flare", which CERN said in a statement would be the first foundation model for collider physics, trained on billions of simulated particle collisions.
It will use the same idea as the large language models behind generative AI but be trained on collision data rather than text, a CERN spokesperson told Reuters.
It is meant to replace some specialised algorithms now used to reconstruct and identify particles.
The spokesperson said CERN has been using AI to turn raw data from experiments into pictures of the events for a decade, but with multiple tools used for each step.
CERN holds about one exabyte, or 1 million terabytes, of experimental data from the Large Hadron Collider, and expects the volume of collisions to increase tenfold after its ongoing upgrade, expected around 2030.
"High-energy physics provides a unique environment in which artificial intelligence can be developed and tested with exceptional rigour," Porcari said.
No timeline or budget was disclosed. Both sides said the models could later serve other sciences.
The Paris-based lab, broadly viewed as Europe's best chance to catch up with US and Chinese competitors, this month unveiled a new open-weight model it says can challenge its rivals. It will be released to the public on October 27.
Users can download open-weight models from platforms and use them on their own hardware.