CERN and Mistral Plan a Foundation Model for Particle Physics

CERN, the European Laboratory for Particle Physics, and the French AI developer Mistral plan to build a pretrained "Large Physics Model". According to CERN, it would be the first foundation model for collision physics.

Built like a language model, trained on collisions

The model is meant to be structured much like a large language model (LLM). The decisive difference lies in the training material: instead of text, it will learn from data recorded in particle collisions. The idea is to carry the pretraining approach that made chatbots possible over to a scientific domain whose raw material is measurement data rather than sentences.

In general, a foundation model is trained once on very large volumes of data and can then be adapted to many different tasks. CERN's description suggests that the Large Physics Model is meant to follow this principle for collision physics.

Why it matters

Particle physics produces enormous amounts of data, and collisions are its central source of insight. A model pretrained specifically on this kind of data would be a notable step for a field that has so far had no foundation model of its own, if CERN's claim of being first holds.

The partnership also pairs one of Europe's best-known research institutions with a European AI company, which fits the continent's push for its own AI capabilities.

Many questions remain open. Details on the timeline, the size of the model, the data to be used and the first planned applications are not yet available. The same goes for how researchers outside CERN could eventually use the model. For now, the announcement sets out the goal rather than a finished product.