AI foundation model pretrained on 500 million collision events transforms particle physics analysis

Physicists have unveiled a foundation model for particle collider data that can be trained once and then adapted to a strikingly diverse set of analyses, from searches for new heavy particles to precision measurements of quantum correlations.

The motivation is a familiar bottleneck in high-energy physics. Data-intensive collider experiments increasingly require models that transfer across high-dimensional, partially observed measurements, yet most analyses still optimize separate models for each task. Every new search or measurement traditionally means building and training a bespoke classifier or reconstruction algorithm from scratch, often with limited simulated signal events. EVENET addresses this by constructing a shared particle-cloud representation from the variable-length set of physics objects in each event, such as jets, electrons and muons, together with their kinematic properties, flavour-tagging information, lepton identification and charge.

Architecturally, the model is a hybrid point–edge transformer. A transformer backbone provides a permutation-equivariant mechanism for modelling sets of physics objects through attention-based interactions, while k-nearest-neighbour message passing incorporates local geometric structure. The backbone was pretrained on roughly 3 billion simulated Standard Model events, of which about 500 million were retained after preselections, using a two-stage curriculum. In the first stage, only a self-supervised generative head was active: a diffusion-based objective reconstructs masked portions of the particle cloud from the unperturbed remainder, with the masking probability progressively increased from 30% to 100%. In the second stage, supervised objectives were added, including classification into one of ten physics-process categories and a supervised generative head that reconstructs invisible event components such as neutrino momenta, conditioned on the observed objects and the missing transverse momentum.

The pretraining bill was substantial. Training both stages required on the order of 100,000 GPU-hours on 512 NVIDIA A100 GPUs at the Perlmutter supercomputer at the National Energy Research Scientific Computing Centre, with the raw simulation corpus amounting to about 4 terabytes before preprocessing. The authors argue this one-time investment is amortized through reuse: downstream fine-tuning tasks typically required only a few to tens of GPU-hours, and pretrained variants consistently reached optimal validation loss with far fewer optimizer updates than models trained from scratch.

The first major test was generalization across a grid of signal hypotheses, a scenario typical of searches for new physics, where large parameter spaces must be explored with limited simulated data for each point. The benchmark considered a heavy particle X decaying into a lighter particle Y and a Standard Model Higgs boson, with Y decaying to two bottom quarks and the Higgs to a pair of W bosons, one of which decays leptonically. This topology is absent from the pretraining corpus, and the fully simulated CMS Open Data samples include pile-up and detector-response effects that the pretraining data, generated without pile-up, do not. Performance was quantified with the significance-improvement characteristic, SIC, the ratio of signal efficiency to the square root of background efficiency. Fine-tuned at each of 121 mass points, EVENET with full two-stage pretraining achieved the highest sensitivity across most of the mass grid, improving the maximum SIC by approximately 50% relative to the same architecture trained from scratch, 20% relative to XGBoost and 10% relative to TabPFN, a pretrained foundation model for tabular data. Pretraining also cut convergence from roughly 3,000 optimization steps to about 1,000.

The most striking demonstration is a direct transfer from simulation to real collision data. The team searched for anti-isolated upsilon meson decays to muon pairs in the 2016 CMS Open Data DoubleMuon dataset, a regime where the decay muons are embedded in substantial surrounding hadronic activity and standard isolation-based selections fail. Following the CATHODE anomaly-detection strategy, the model’s self-supervised generative head learned a background model from the dimuon-mass sidebands, generated background events as particle clouds, and an XGBoost classifier then flagged discrepancies in the signal region. In the primary opposite-sign configuration, the fully pretrained model reached a median local significance of 7.6 sigma, exceeding the published 6.4-sigma benchmark, while the same architecture trained from scratch managed only 1.5 sigma. The pretrained model also required far smaller kinematic calibration corrections, about 11.6% versus 74.4% for the scratch model, indicating more physically faithful generation. Control configurations using same-sign dimuon data, where no physical resonance can appear, were compatible with zero significance, providing no evidence of an artificially sculpted signal.

A third evaluation targeted an exotic Higgs decay, H to two pseudoscalar particles a, each decaying to a pair of bottom quarks, a channel absent from pretraining and plagued by large multijet backgrounds. Here EVENET was fine-tuned with classification, particle-assignment and particle-segmentation heads and compared against SPANet, a state-of-the-art transformer-based model for joint assignment and classification. At the representative pseudoscalar mass of 30 GeV, the fully pretrained model with the classification head alone reached a maximum SIC of 4.1, versus 1.6 for scratch training and 1.4 for SPANet. Remarkably, with only 5% of the reference 200,000-event training set, it reached an SIC of 1.7, exceeding both baselines’ peak values even when they used 150% of the reference data. Under simulated jet-energy-scale variations, the pretrained model’s sensitivity was also markedly more stable than that of the scratch model.

The fourth application pushed into precision measurement: quantum correlations in dileptonic top-quark pair events, where spin-correlation coefficients are combined into the entanglement-sensitive observable Delta, for which positive values indicate quantum entanglement. Reconstruction is the dominant limitation because two undetected neutrinos leave the kinematic system under-constrained. EVENET tackled this with an assignment head resolving combinatorial ambiguities and a supervised generative head inferring neutrino momenta. With only about 1.5% of a reference dataset of one million events, it achieved a measurement precision of 3.88%, compared with 6.20% for self-supervised-only pretraining and 5.58% for scratch training, and precision saturated near 3.33%, approximately 37% better than a published benchmark of 5.26%. Under jet-energy-scale and missing-transverse-momentum variations applied at evaluation time, the pretrained model’s precision shifts stayed within tight bounds, while scratch-model fluctuations were about five times larger.

The authors are candid about limitations. EVENET is not yet a fully uncertainty-aware framework, and explicit modelling and propagation of detector systematics and theoretical uncertainties remain necessary for rigorous statistical interpretation. Predictions from different task heads are not explicitly constrained to be mutually consistent, systematic stress tests under substantial simulation-to-data mismodelling have not yet been performed, and a detailed analysis of the latent-space organization and interpretability is deferred to future work. Deployment in actual LHC analyses is described as the decisive next test. Still, the pattern across all four downstream settings is consistent: a single pretrained backbone, once its substantial training cost is paid, can be adapted efficiently to heterogeneous inference tasks without loss of physics performance, positioning event-level foundation models as a potential shared substrate for the next generation of collider analyses.

Subject of Research: An event-level foundation model for particle collider data analysis

Article Title: Foundation model for particle collision data analysis

Article References: Hsu, T.-H., Zhou, B., Liu, Q., Xu, Y., Li, S., Hou, G. W.-S., Nachman, B., Hsu, S.-C., Mikuni, V., Chou, Y.-T., & Zhang, Y. (2026). Foundation model for particle collision data analysis. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01322-9

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01322-9

Keywords: particle physics, foundation models, machine learning, collider data, CMS Open Data, anomaly detection, Higgs boson, top quark, quantum entanglement, self-supervised learning, Nature Machine Intelligence, LHC