On October 15, 2026, CERN, the European Organization for Nuclear Research, and the French AI company Mistral announced that they had signed a letter of intent to pursue a long-term strategic partnership on developing foundation models for scientific research. At the center of the plan is the "Large Physics Model," which would be trained on billions of simulated proton-proton collisions of the kind produced at the Large Hadron Collider (LHC). The goal is to support tasks such as particle identification and inferring what happened in a collision with a single shared pretrained model, rather than the multiple specialized algorithms that have traditionally handled them. According to CERN's official announcement, the partnership looks ahead to the High-Luminosity LHC (HL-LHC), which is scheduled to begin operation around 2030. It does not announce the performance of a finished AI model. The plan is meant to lay the groundwork for using AI in particle physics so that ever-growing volumes of experimental data can be analyzed efficiently.
One Shared AI Model for the Many Steps of Reading Particle Traces
What the LHC's detectors record is not a picture of particles colliding. It is a collection of signals produced by the collisions, such as the positions where particles passed through and the energy they deposited in the detector. Researchers combine these signals to estimate which particles were produced, in which directions they flew, and with how much momentum. This process is called "reconstruction," and it is an essential step in analyzing particle physics experiments.
For example, to identify an electron in the LHC's CMS experiment, researchers must match the track recorded by the tracker on the inner side of the detector with the energy measured by the electromagnetic calorimeter on the outer side. Because a magnetic field bends an electron's path, its momentum can be determined from the curvature. However, electrons can also emit photons and lose energy as they pass through material inside the detector. That means analysts must consider not only the particle's track but also information from multiple detectors. According to CMS's technical explainer, machine learning is already used to integrate this information and improve the accuracy of energy measurements.
Current particle reconstruction relies on running algorithms, each specialized for a particular task, in sequence and passing the results on to the next step. By contrast, the Large Physics Model envisioned by CERN and Mistral aims to be a shared model that learns across the whole chain of reconstruction steps and can be applied to multiple analyses. Envisioned uses include not only reconstructing collisions and identifying particles but also anomaly detection to spot unexpected signals and evaluating the design of future detectors.
The defining feature of a foundation model is that it learns common features and structures from large amounts of data in advance, and that knowledge can then be applied to a variety of uses. If researchers no longer have to train a model from scratch for each analysis task, they may be able to build on shared learning results for their individual measurements and searches. However, the announcement did not disclose the model's specific architecture or how it would be adapted for different uses. How much of the conventional reconstruction pipeline can be folded into a shared model will also need to be tested through future research and development.
A 10-Fold Increase at the High-Luminosity LHC Makes Particle Identification More Complex
A major factor behind the Large Physics Model is the High-Luminosity LHC (HL-LHC), due to begin operation around 2030. By collecting far more collision data than before, the HL-LHC aims to increase opportunities to observe extremely rare phenomena. At the same time, as the number of collisions grows, the signals recorded by the detectors become more complex, and the computation and difficulty of analysis rise with them.
CERN has set a goal for the HL-LHC of increasing integrated luminosity tenfold relative to the LHC's original design. But this tenfold figure is a measure of the amount of collision data accumulated over a given period; it does not mean that the number of collisions in a single beam crossing will rise tenfold. Based on CERN's HL-LHC technology explainer, the difference between the two can be summarized as follows.
| Metric | Baseline and target given by CERN | Impact on analysis |
|---|---|---|
| Integrated luminosity | 10 times the LHC's original design performance | More collision data can be collected over the long term |
| Collisions per beam crossing | About 60 under the conventional conditions in the same document; a target of 140–200 for the HL-LHC | More signals overlap at the same time, making particle identification harder |
*The figure of about 60 is the value given as the conventional condition in CERN's technical document; it is not a measurement of operating conditions as of October 2026.
At the LHC, beams made up of bunches containing many protons are made to cross, but a single crossing does not necessarily involve just one colliding pair of protons. Because several proton-proton collisions occur almost simultaneously, the detector records overlapping particle tracks and energy signals from different collisions. This phenomenon is called "pileup," and it makes it harder to determine which tracks and energy deposits come from the same particle or the same collision. Because the HL-LHC will have more overlapping collisions than ever, it will require not just fast processing of large volumes of data but also the ability to accurately identify individual particles within densely overlapping signals.
The shared model that CERN and Mistral hope to develop is expected to address both the growth in data volume and the growing complexity of analysis. However, a tenfold increase in collision data is not the same thing as AI making processing speed or computational efficiency ten times better. How much speedup or accuracy improvement can be achieved cannot be judged until the model is actually built and compared with existing analysis methods.
How the Plan Differs From Existing Jet Foundation Models
Research on foundation models that can be applied to multiple analyses is already under way in particle physics. One example is "OmniLearned," by Wahid Bhimji and colleagues. The work was published in 2026 in the peer-reviewed journal Physical Review D, volume 113, article 032020, and examines the potential of using a shared pretrained model for particle physics analyses.
OmniLearned targets "jets," in which many particles originating from quarks or gluons fly off in a roughly common direction. According to the research paper, the team pretrained the model on more than one billion jets and evaluated its performance on several tasks, including identifying jets originating from top quarks. They also tested a task of identifying jets originating from b quarks using detailed simulations of the ATLAS detector. In addition, they tackled anomaly detection using real data collected by the CMS experiment, showing that the pretrained model can be applied to different analyses.
However, the "more than one billion" that OmniLearned learned from refers to the number of jets, whereas the "billions" cited for the Large Physics Model plan refers to simulated collisions. Because the two count different things, the scale of training or the relative performance of the models cannot be judged from a simple comparison of the numbers. Nor are the results from OmniLearned guaranteed to be reproduced in the Large Physics Model.
Still, the fact that a shared pretrained model could be applied to multiple uses in jet analysis is an important precedent for considering this plan. The idea of a foundation model has already been adopted in particle physics research. CERN positions the Large Physics Model as the first foundation model for physics research using particle colliders, but given the earlier work, this does not mean it is the first foundation model developed anywhere in physics. What stands out about the new plan is that it targets a large-scale shared model covering the entire chain of particle reconstruction, rather than being limited to specific particles or analysis tasks, and that it will be developed jointly with Mistral, which has expertise in large-scale AI training.
Learning From Billions of Simulated Collisions and Validating on Real Data
According to CERN, LHC experiments have so far accumulated about one exabyte of data, equivalent to roughly one million terabytes. However, what has been named as the training target for the Large Physics Model is billions of simulated particle collisions. It has not been announced that all of the accumulated experimental data will be used to train the AI, or that all of it will be provided to Mistral. How real collision data will be used may become clear only as development proceeds.
The advantage of training on simulated data is that the correspondence between the particles and physical processes that were generated and the signals that appear in the detector is known. Because researchers can train and evaluate the model with ground-truth information, it is easier to teach it particle identification and collision reconstruction. On the other hand, simulation cannot perfectly reproduce the real detector's response. Even if a model shows high accuracy on simulated data, it will not necessarily perform equally well on real data, so differences between the two must be handled appropriately.
Addressing this problem has long been an important challenge in particle physics. In reconstructing electrons and photons, CMS assumes that differences remain between simulation and real data and calibrates its measurements using well-known physical phenomena. For example, by using the process in which a Z boson decays into an electron and a positron, researchers can check whether a known phenomenon is reconstructed correctly and correct any measurement offsets. This is not a result demonstrating the performance of the Large Physics Model, but an example of a method that conventional particle analyses use to ensure accuracy.
When a shared foundation model is used for multiple analyses, one must also be careful that errors stemming from differences between simulation and real data could be carried into many analyses in common. What needs to be evaluated is not just the accuracy of simple particle identification. It is also important to determine how much bias arises in energy and momentum measurements and whether the uncertainty of measurements can be properly estimated. The more widely a shared model is used, the more it becomes necessary to understand not only individual analyses but also the systematic errors originating from the model as a whole.
For anomaly detection as well, it is necessary to distinguish between AI finding an unexpected signal and discovering a new particle or physical phenomenon. In the OmniLearned study, the anomaly detection method was evaluated through a task of rediscovering the known top quark, using data collected by CMS in 2016. After the events judged to be anomalous were extracted, a step was also included to test their statistical significance. The same applies when searching for unknown phenomena: researchers must examine in detail the background events from known physical processes and the detector response, and test whether an observed signal can be explained by chance statistical fluctuations. AI-based anomaly detection is a means of narrowing down candidates that could lead to a discovery; it is not in itself proof of a new physical phenomenon.
Accuracy on Real Data and Terms for Sharing the Model Will Be Key
The partnership combines CERN's expertise in particle physics and large-scale computing with Mistral's technology for training large AI models. The initial stage of development will focus mainly on high-energy physics, and the organizations say the work could be applied to other scientific fields in the future. However, what has been presented so far is a long-term policy toward joint development, and the model has not been completed. Expansion into other fields is likewise not yet at the stage of concrete results.
Enrica Porcari, CERN's Chief Information Officer (CIO), has expressed the aim of advancing AI technology that can be used and shared across the physics research community. However, the announcement did not say whether the model's weights or source code will be released, nor what licenses or terms of use would apply if they are. How freely researchers can use the shared model, and whether third parties can independently re-verify its performance and research results, will depend on the sharing conditions announced in the future.
No specific figures have been published for the Large Physics Model's expected completion date, development budget, or accuracy and processing speed compared with existing analysis methods. As development progresses, key evaluation criteria will include whether the model can reconstruct events accurately even under the heavily overlapping collision conditions expected at the HL-LHC, and whether it performs well not only on simulations but also on real data. The ability to quantitatively assess the uncertainty of analysis results will also be essential for use in scientific research.
The initiative aims to evolve the individual AI techniques that support particle physics analysis into a large-scale foundation model that can be shared across multiple research uses. Its value will be determined not only by the scale of the model but also by its reliability on real data and by whether an environment is in place in which researchers can verify and reuse the results. If those conditions are met, it could become a new research foundation for measuring known physical phenomena more precisely, and for searching for unknown ones, in the vast collision data of the High-Luminosity LHC that will begin around 2030.