EveNet AI Model Finds the Upsilon in CMS Open Data

EveNet, a foundation model for particle physics built by researchers in Taiwan, China, the US and Japan, has been pretrained on 500 million simulated collision events and then adapted to find a known particle in real detector data. The team described the model in Nature Machine Intelligence on 29 October and has released its code on GitHub and its pretrained weights on Hugging Face, both under the MIT licence.

The idea is the one behind large language models: train a single model once at great expense, then adapt it cheaply to many tasks. In collider physics, analysts usually train a new network for each search. EveNet is meant to be a shared starting point for those analyses.

In brief

EveNet was pretrained on 500 million simulated collision events on NERSC’s Perlmutter supercomputer, using about 100,000 GPU-hours.
Adapted to real 2016 CMS Open Data, it picked out the Upsilon meson with a median significance of 7.6 sigma, above a published 6.4 sigma benchmark.
The authors say it is not yet ready for rigorous statistical use, and that its real test will be deployment in LHC analyses.

How the EveNet foundation model was trained

The authors, from National Taiwan University, Shanghai Jiao Tong University, the SLAC National Accelerator Laboratory, the University of Washington, Stanford University, Nagoya University and National Tsing Hua University, treat each collision as a “particle cloud”: a variable-length set of the reconstructed objects in an event. A transformer builds a shared representation of that cloud, and swappable heads on top handle different jobs: classifying events, assigning particles to their parent decays, segmenting events and generating events with a diffusion model.

Pretraining ran in two stages. The model first learned by filling in masked parts of events (self-supervised learning), then trained on a mix of that objective and labels taken from the simulation, which the paper calls physics-informed supervision. Training used 512 NVIDIA A100 GPUs on Perlmutter at the US National Energy Research Scientific Computing Center. Model training alone took of the order of 100,000 GPU-hours, the paper says, not counting generating the simulated events. The pretraining data, about 180GB after processing, is public on Hugging Face.

Rediscovering the Upsilon in CMS Open Data

The key test was whether a model trained on simulation would work on real collisions. The team searched 2016 CMS Open Data, released by CERN’s Compact Muon Solenoid experiment, for Upsilon mesons decaying to two muons in a hard-to-study region where the muons are buried in other particle debris. The Upsilon is well known, so the exercise checks the method rather than claiming a discovery.

The fully pretrained EveNet reached a median local significance of 7.6 sigma, against 6.4 sigma in an earlier published analysis of the same kind. The same network trained from scratch reached only 1.5 sigma.

Searches and precision measurements

In a simulated search for exotic Higgs boson decays to four bottom quarks, a channel left out of pretraining, EveNet scored 4.1 on the paper’s sensitivity measure, against 1.6 for the same model trained from scratch and 1.4 for SPANet, an established specialist model. With only 5% of the reference training data it beat both baselines’ best results on 150%.

In a simulated measurement of quantum correlations between top quarks, it reached a precision of about 3.33%, which the authors put at roughly 37% better than a 5.26% benchmark. The biggest gains came where labelled data were scarce.

Limits the authors set out

EveNet’s results are strong, but the paper is careful about what they show. Only one benchmark, the Upsilon search, uses real collision data; the Higgs and top-quark results come from simulation. The authors write that EveNet “is not yet a fully uncertainty-aware framework”, meaning detector and theory uncertainties still have to be modelled separately before any result could support a formal physics claim. Its different heads are not forced to give mutually consistent answers. And they have not yet stress-tested the model where simulation and real data disagree badly, which is exactly where a model trained on simulation is most likely to fail. “Deployment in LHC analyses will provide a decisive test,” they write.

The case for EveNet is efficiency rather than a new capability. A single pretraining run costing around 100,000 GPU-hours only pays off if many analyses reuse it, and the paper’s evidence that fine-tuning needs far fewer training steps is the argument for doing so.

It also arrives as larger players move into the same space. CERN and Mistral this month signed a statement of intent to build a Large Physics Model, which CERN described as the first foundation model for collider physics; that model has not been built. EveNet is a smaller, working and openly licensed example from university and national-lab groups, with its data, code and weights published so other physicists can test it. That fits wider efforts to make research outputs reusable, such as Hugging Face’s ProtoStar project for public science data and the European Commission’s AI in Science Enablers initiative, which NVIDIA has joined. More stories like this are collected in TSN’s Tech for Good hub.