Model Card for EveNet

EveNet is an event‑level foundation model for collider physics. It represents each collision event as a combination of a fixed‑length point cloud of reconstructed objects and a vector of global observables, and processes these inputs with a hybrid point–edge transformer (PET) backbone. The backbone encodes both local geometric relationships and global context and feeds several task‑specific heads that support classification, regression, combinatorial assignment, segmentation, and generative modelling. EveNet is pretrained on large‑scale Standard Model (SM) simulations using a hybrid of supervised and self‑supervised objectives and can be fine‑tuned for a broad range of downstream analyses.

Model Details

Model Description

EveNet treats a hadron‑collider event as a point‑cloud of up to 18 reconstructed objects (jets, electrons, muons, etc.) with seven features each (energy, transverse momentum (p_{\mathrm{T}}), pseudorapidity (\eta), azimuthal angle (\phi), b‑tag score, lepton flag and charge) plus optional per‑event scalars. Masks indicate which padded entries are valid. A separate normalisation layer handles global variables. The backbone consists of three stages:

1. Global embedding: a learned embedding of global scalars configured via Body.GlobalEmbedding.
2. PET body: a point–edge transformer with attention layers and k‑nearest‑neighbour edges, configurable by depth, number of heads and feature‑drop rate. A 10 % feature dropout was applied during pretraining to encourage robustness.
3. Object encoder: mixes the particle and global tokens, again using configurable attention layers.

Multiple task‑specific heads are attached to this shared representation. Discriminative heads include classification, regression, assignment (resonance matching) and segmentation. Generative heads implement diffusion‑based self‑supervised reconstruction of visible objects and supervised generation of invisible particles (such as neutrinos). Conditioning on the diffusion time unifies discriminative and generative tasks within a single latent space.

• Developed by: EveNet collaboration (Ting‑Hsiang Hsu, Bai‑Hong Zhou, Qibin Liu, Yue Xu, Shu Li, Wei‑Shu Hou, Benjamin Nachman, Shih‑Chieh Hsu, Vinicius Mikuni, Yuan‑Tang Chou, Yulei Zhang et al.).
• License: MIT.
• Model type: Hybrid point–edge transformer with diffusion‑based generative heads and discriminative heads.
• Languages: Not language‑based – inputs are numeric features of physics objects.
• Fine‑tuned from: Pretrained from scratch on SM simulations; this is the base foundation model.

Model Sources

• Repository: GitHub – EveNet‑Full
• Documentation: EveNet Documentation Portal
• Paper: “EveNet: A Foundation Model for Particle Collision Data Analysis”.
• Pretrained Weights: Available on Hugging Face.

Uses

Intended Uses

EveNet is designed as a multi‑task foundation model for collider‑physics analyses. It supports event‑level classification, combinatorial reconstruction, segmentation of final‑state particles and generative modeling for missing or masked objects. Intended uses include:

• Pretraining and fine‑tuning: Use the released checkpoints as a starting point for new analyses. The pretraining corpus consists of approximately 500 million proton–proton collision events at (\sqrt{s}=13,\text{TeV}) spanning QCD multijet, top‑quark, (W/Z)+jets and diboson processes. Events were simulated with MADGRAPH5_aMC@NLO for matrix elements, PYTHIA for parton showering and hadronisation, and Delphes for fast detector simulation. Hard objects were required to satisfy (p_{\mathrm{T}}>10,\text{GeV}) (20 GeV for jets) and central pseudorapidity |(\eta)|<2.5, and additional preselections ensured at least two jets for QCD events and appropriate multiplicities for other processes.
• Downstream analyses: The pretrained model can be fine‑tuned on smaller datasets for tasks such as heavy‑resonance searches, exotic Higgs decays ((H\to aa\to 4b)), quantum‑correlation measurements in (t\bar{t}) dilepton events and anomaly detection on collision data. Example datasets and configurations for these tasks are provided in the EveNet benchmark repositories.

Out‑of‑Scope Uses

• Non‑HEP data: EveNet is tailored to collider physics; it is not a general machine‑learning model for images, language or tabular data outside particle‑collision analyses.
• Human or personal data: The model should not be applied to sensitive personal data. It is trained exclusively on simulated physics events and does not generalize to human‑related tasks.
• Uncalibrated inference on real data: While EveNet transfers well to collision data, users must validate model performance and calibrate outputs to ensure physical fidelity, particularly when applying the generative heads to unseen detectors or energy regimes.

Bias, Risks & Limitations

• Simulation bias: The pretraining dataset comprises simulated SM processes that include generator‑level cuts (minimum (p_{\mathrm{T}}), multiplicity requirements, etc.). These cuts bias the population toward energetic events; analyses of soft‑physics regimes or rare final states may require additional training data.
• Model scope: The model inputs are limited to reconstructed objects with seven features each. Analyses requiring detailed detector information (e.g., calorimeter cells) or more than 18 objects may need to modify the architecture or use custom preprocessing.
• Generalisation to non‑SM physics: Although EveNet demonstrates transfer to unseen processes, it has been pretrained on SM simulations. Performance on exotic signatures not represented in the pretraining corpus may be limited.

Recommendations

Users should:

• Provide domain‑appropriate training data for any new physics regime and re‑estimate normalisation statistics when modifying the input schema.
• Validate the model on collision data and apply calibration procedures when using generative outputs to enforce physical constraints.
• Be aware of class imbalances and provide appropriate per‑event weights to the classification head when training.

Training Details

Training Data

The foundation model was trained on a corpus of roughly 3 billion simulated SM events, from which about 500 million were retained after preselections. These simulations were generated with MADGRAPH5_aMC@NLO for hard processes, PYTHIA for parton showers and hadronisation, and Delphes for fast detector simulation. Cuts of (p_{\mathrm{T}}>10,\text{GeV}) for leptons and (p_{\mathrm{T}}>20,\text{GeV}) for jets and |(\eta)|<2.5 were imposed. For QCD events at least two jets were required.

Training Procedure

EveNet uses PyTorch Lightning with a Ray‑based distributed training pipeline. Pretraining combines:

• Supervised objectives: event‑level classification of physics processes, object assignment to resonances and segmentation of final‑state particles. Regression targets (momentum, masses) are normalised using mean and standard‑deviation tensors.
• Self‑supervised and supervised generative objectives: diffusion models reconstruct visible inputs and predict invisible neutrinos; both tasks are conditioned on a diffusion time (t) to unify generative and discriminative regimes.

An Exponential Moving Average (EMA) of the weights is maintained during training; training scripts allow resuming from checkpoints and replacing model weights with EMA weights before saving. The default configuration uses feature dropout to encourage generalisation. For fine‑tuning, the backbone learning rate may be reduced relative to task‑specific heads, and pretraining weights can be loaded safely with mismatched layers skipped.

Evaluation

The EveNet paper evaluates the pretrained model on four downstream tasks:

1. Heavy resonance search: fine‑tuning on simulated searches for a heavy resonance decaying to a light scalar and the SM Higgs. EveNet shows improved sensitivity compared with scratch and self‑supervised baselines.
2. Exotic Higgs decays ((H\to aa\to 4b)): improved performance relative to dedicated models.
3. Quantum correlations in (t\bar{t}) dilepton events: EveNet pretrained on 500 million events achieved a normalised uncertainty (\Delta D) of 1.61 % on the entanglement‑sensitive observable after fine‑tuning with 15 % of typical training statistics, outperforming scratch and self‑supervised baselines. It also reached 82 % pairing accuracy, several points above scratch (80 %) and SSL (79 %) models.
4. Anomaly detection on CMS Open Data: Generative diffusion heads were fine‑tuned on 2016 dimuon data to rediscover the Υ meson. EveNet replaced the conditional normalising flow baseline with a generative model that directly produces dimuon point clouds; after calibration, it achieved competitive or superior anomaly significance while maintaining physical fidelity.

Citation

If you use this model in your research, please cite:

@article{Hsu:2026sww,
  author        = {Ting-Hsiang Hsu and others},
  title         = {{EveNet: A Foundation Model for Particle Collision Data Analysis}},
  eprint        = {2601.17126},
  archivePrefix = {arXiv},
  primaryClass  = {hep-ex},
  month         = {1},
  year          = {2026}
}