Table of Contents
CERN and Mistral have signed a statement of intent to establish a long-term AI research partnership, with plans to develop a Large Physics Model for analyzing particle collisions at the Large Hadron Collider.
The initiative aims to train a single model on billions of simulated proton-proton collisions. The proposed system as the first foundation model for collider physics, able to support several scientific tasks instead of a single narrow analysis.
How the Large Physics Model Would Work
The technical target is collision reconstruction, turning the signals a detector records into the particles and other objects researchers use in physics analysis. Today, that process runs through a sequence of specialized algorithms, with machine learning already used to improve individual tasks.
CERN and Mistral want to explore whether a foundation model can learn across that reconstruction chain. Rather than building a separate system for each step, the proposed approach would use one pretrained model as a basis for multiple tasks. Mistral's has identified billions of simulated collisions as the planned training material.
The intended applications include reconstructing collision events, identifying particles, detecting unexpected signals, and evaluating designs for future detectors. This spreads the project across analyzing collision data and building tools that could help researchers assess experimental hardware.
The training data and the experimental data differ. The stated training plan uses simulated proton-proton collisions, while CERN's existing archive of recorded LHC experiments gives context for the scale of the scientific challenge rather than a disclosed training dataset for this model.
Preparing for a Larger Collision Workload
CERN has stored around one exabyte of LHC experimental data, equivalent to one million terabytes. The laboratory expects collision numbers to increase tenfold with the High-Luminosity LHC, which should begin operating around 2030.
The proposed model would address how detector information is transformed into usable physics objects, and its value would lie in supporting the analysis itself across several related tasks.
Johannes Brandstetter, Mistral's vice president of AI for Engineering, said the work would explore what large-scale models can contribute to the growing complexity of particle physics, with the potential to support new discoveries.
