CERN already holds about one exabyte of data from the Large Hadron Collider, roughly a million terabytes of recorded collisions. On October 15 the laboratory signed a letter of intent with the French company Mistral AI to use a portion of that record, and the simulations built around it, to train a foundation model for particle physics.
The model is called Flare. It will be trained on billions of simulated particle collisions. The idea follows the structure of a language model: the same architecture, the same training objective, with collision data standing in for text. The goal is to replace some of the specialized algorithms that physicists currently use to reconstruct and identify particles from the signals detectors record.
CERN’s chief information officer, Enrica Porcari, said high-energy physics offers a unique environment in which artificial intelligence can be rigorously tested. The claim rests on the nature of the data. Physics measurements come with uncertainties that can be calculated, and a model’s output can be compared against theory rather than against another model. Few domains offer that kind of ground truth.
The reconstruction problem Flare is meant to address is a bottleneck that has grown with the collider’s power. When protons collide, detectors record hits in layers of silicon and gas, and software has to infer which hits belong to which particle and what kind of particle it was. The algorithms that do this are fast and well understood, and they are reaching the limits of what hand-written code can achieve as collision rates rise.
The upgrade program is what makes the timing urgent. CERN expects the collider’s collision rate to increase tenfold after its next upgrade, and the computing budget cannot rise by the same factor. Something has to give, and machine learning is the candidate that physicists have been developing for years, with applications in tracking, calorimetry and anomaly detection already in production.
Mistral brings expertise in training large models efficiently, which is the discipline the project needs. The laboratory has world-class computing and a data set no one else possesses, and it lacks the industrial practice of building models at scale. A partnership of that kind is how European institutions have approached AI, pairing public data and infrastructure with private model expertise.
The training corpus is where the project differs from its language-model analogues. Simulated collisions are generated with software that encodes physicists’ current understanding of the particles and their interactions, which means a model trained on them learns the simulation rather than nature. Physicists call this the simulation gap, and closing it requires techniques that mix simulated data with the real events recorded in the detector.
The scientific payoff is measured in sensitivity rather than in speed. A model that identifies rare decays slightly better than existing methods increases the number of events a search can examine, and searches for rare processes are limited by statistics. Improving the analysis improves the physics, and each improvement compounds over years of data taking.
There is also the question of what a physics foundation model can do that a task-specific network cannot. The argument for a large pretrained model is transfer: a system that has learned the structure of collision data can be adapted to new detectors, new analyses and new questions with a fraction of the labeled data. CERN runs multiple experiments, and each one rebuilds its software stack.
The institutional dimension matters in Europe. CERN is an intergovernmental organization with 23 member states, and its computing investments have historically supported open-source software used across the field. Partnering with a French company rather than an American one keeps the work inside a European framework at a time when European governments are funding domestic AI capacity.
The two sides did not announce a timetable or a budget, which is normal for a letter of intent and also a reminder that intent is not a commitment. Training a model of this kind requires computing resources, staff and access to data that CERN controls under its own policies, and each of those has to be arranged separately.
Data access is the constraint that will govern the project’s pace. CERN’s experiments are run by large international collaborations, and each has its own governance over how its recorded data may be used and by whom. A model trained across experiment boundaries has to satisfy every collaboration involved, and that negotiation is slower than the engineering.
The laboratory also runs one of the largest computing infrastructures in publicly funded science, built originally to distribute analysis across institutes worldwide. Training a foundation model requires a different pattern of computing, concentrated rather than distributed, and adapting the existing grid to that workload is part of what the partnership will have to solve.
Analysts said the project is a test of whether foundation-model methods generalize beyond language and images to domains where the data is structured by physical law. If it works, the pattern applies to astronomy, climate and materials science. If it does not, the field will have learned that the gains come from task-specific design rather than from scale.
