Particle physicists usually build a new machine learning model for every analysis they run at the Large Hadron Collider, and each one is trained on purpose-made simulations and checked from scratch. EVENET, described in a peer-reviewed paper in Nature Machine Intelligence, tries the approach that has worked for language and protein models: train one large model once, then adapt it cheaply to many tasks. The headline result is that it also works on real detector data, not only on simulation.
The model treats each collision as an unordered cloud of reconstructed objects, such as jets, electrons and muons, and passes it through a shared transformer backbone with swappable output heads for classifying the process, matching particles to the parent particles they came from, grouping them, and reconstructing missing ones such as neutrinos. It was pretrained in two stages on about 500 million simulated collisions, kept from roughly 3 billion generated: first by learning to fill in masked particles, then with labelled physics targets added. That took on the order of 100,000 GPU-hours on 512 Nvidia A100 chips at the Perlmutter supercomputer. Fine-tuning for a single task then needs a few to tens of GPU-hours.
The results are consistent across four tests. In a search for a hypothetical heavy particle, it improved the best signal sensitivity by about 20% over XGBoost, a standard tool in the field, and about 10% over TabPFN, a pretrained model for tabular data. In a search for exotic Higgs boson decays, with 5% of the reference training set it beat the best score the specialist SPANet model reached with 150%. In a measurement of quantum entanglement in top quark pairs, it reached 3.33% precision against a 5.26% earlier benchmark. The most persuasive test used 2016 open data from the CMS experiment: searching for the Upsilon meson, a known particle, in a messy region, it found the signal at a median significance of 7.6 sigma, above the 6.4 sigma previously published. A version trained without pretraining managed 1.5.
The wider point is about cost. A heavy one-off pretraining bill, paid back by reusing the model across analyses, is the same bargain that justifies general-purpose language models, and the paper shows it holding in a field with strict standards for evidence. Rediscovering a known particle is a sanity check rather than a discovery, but it is the right check for a model trained on simulation.
The authors are frank about what is missing. EVENET does not yet track detector and theory uncertainties in the rigorous way a physics result requires, its output heads are not forced to agree with each other, and it has not been stress-tested where simulation and real data disagree badly. Its pretraining data also used a simplified generic detector and no overlapping collisions. Adoption by the big LHC collaborations, which demand exactly those checks, will be the real test.