CERN and Mistral plan Flare, a foundation model for collider physics

CERN and Mistral signed a statement of intent on October 15 to build a long-term partnership around foundation models for science. The centerpiece is a single pretrained "Large Physics Model," billed as the first foundation model for collider physics, trained on billions of simulated proton-proton collisions at the Large Hadron Collider. France 24 reports it is named Flare.

The model is meant to reconstruct collisions, identify particles, flag unexpected signals and evaluate future detector designs. Today, CERN says, reconstruction runs on a chain of specialized algorithms. A CERN spokesperson told Reuters that Flare uses the same idea as the large language models behind generative AI, trained on collision data instead of text, and is meant to replace some of those algorithms.

The data problem is real: CERN holds around one exabyte of LHC data, and collisions are set to increase tenfold when the High-Luminosity LHC starts around 2030.

What the announcement does not include: a timeline, a budget (Yahoo Tech notes neither was disclosed), or any benchmark. Mistral builds open-weight models, but nothing says whether Flare's weights will be shared. Context: Mistral bought physics AI startup Emmi AI earlier this year, and Mistral's Johannes Brandstetter co-founded it, according to TNW.

My take: a model trained on simulated collisions is only as trustworthy as the simulation. The press release promises rigor. The eval will have to deliver it.