EveNet: A Foundation Model for Particle Collision Data Analysis

summary

The gist

EveNet introduces an event-level foundation model for particle collision data analysis that unifies discriminative and generative objectives in a single physics-informed pretraining framework to…

In short

EveNet introduces an event-level foundation model for particle collision data analysis by unifying discriminative and generative learning in one framework. It uses a point-edge transformer to learn complex event structures, outperforming existing models across various physics tasks like searching for new particles and detecting anomalies, while showing high data efficiency.

Key concepts

Point-Edge Transformer (PET) Encoder
This is the core part of EveNet that acts as the backbone. It processes irregular, unordered point clouds representing reconstructed collider events to learn their internal structure. It creates a unified representation of an event, which can then be used for many different analysis tasks.
Hybrid Pretraining Strategy
EveNet uses a combined approach: supervised learning and self-supervised learning. This involves 'shared parameterisation over diffusion time' and 'particle-level denoising.' This strategy aligns the model's ability to classify data (discriminative) with its ability to generate new, realistic event data (generative) simultaneously.
Event-Level Foundation Model
This refers to a large model trained on entire collision events rather than individual particles. It learns a general physics-aware structure from the raw event data. This allows it to be flexible and effective for many different, diverse analysis goals in high-energy physics.
Unified Latent Geometry
The model maps all different types of tasks (classification, segmentation, generation) into a single shared latent space. This means that the features learned during pretraining are useful across all downstream applications, making the model highly versatile and efficient.

Terminology used across episodes

This episode discusses

The paper

Foundation model for particle collision data analysis · Read on arXiv

Department of Physics, National Taiwan University · Tsung-Dao Lee Institute, Shanghai Jiao Tong University · Fundamental Physics Directorate, SLAC National Accelerator Laboratory · Department of Physics, University of Washington · Department of Particle Physics and Astrophysics, Stanford University · Kobayashi-Maskawa Institute, Nagoya University

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. Here, to address this bottleneck, we introduce EVENET, an event-level foundation model pretrained on 500 million simulated collision events using self-supervised learning and physics-informed supervision. The model constructs a shared particle-cloud representation from the variable-length set of physics objects in each event. Across searches for heavy resonances and exotic Higgs decays, EVENET outperforms both established and state-of-the-art baselines, particularly in low-statistics regimes. We validate transfer to experimental data by rediscovering the Υ meson in CMS Open Data. The pretrained model also supports precise extraction of quantum-correlation observables that remain stable under the tested systematic variations. Together, these results establish a reusable backbone for diverse collider analyses, improving downstream adaptation after a substantial one-time pretraining investment.

DOI: 10.1038/s42256-026-01322-9

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "EveNet: A Foundation Model for Particle Collision Data Analysis".

Mira: Here's the gist of the paper. EveNet introduces an event-level foundation model for particle collision data analysis that unifies discriminative and generative objectives in a single physics-informed pretraining framework to accelerate discovery at current and future colliders The model outperforms state-of-the-art baselines across diverse tasks, including searches for heavy resonances and exotic Higgs decays, and demonstrates exceptional data efficiency in low-statistics regimes Model Architecture and Training Strategy EveNet is built upon a point-edge transformer (PET) encoder that serves as the backbone for learning the internal structure of irregular, unordered point clouds from reconstructed collider events The model leverages a hybrid pretraining strategy that combines supervised and self-supervised objectives, combining "shared parameterisation over diffusion time" and "particle-level denoising-based inpainting of masked or undetectable objects" to align discriminative and generative tasks within a single latent geometry The architecture features a shared PET encoder that produces a unified event representation, which is then consumed by modular task-specific heads for classification, assignment, segmentation, and generation Downstream Task Versatility The model was validated in four distinct analysis settings to assess its versatility and generalizability across different physics goals These tasks include: Search for a new physics scenario, where it often scans the vast parameter space (Task 1); (Out-of-distribution new-physics search scenario), where hypothetical particles predicted by a new physics model...

Kai: are absent from the pretraining dataset (Task 2); (In-distribution data to measure quantum correlations in top-quark pair production), where the synergy between assignment and the generation of undetectable objects...

Mira: is critical (Task 3); (Collision data anomaly detection), where it validates that the foundation model, trained solely on simulated events, retains strong performance on collision data (Task 4).

Kai: Performance in New Physics Searches In the search for a heavy scalar decaying into a light scalar and the SM Higgs boson (X → Y HSM), EveNet–Full achieved the highest sensitivity across essentially the entire mass grid, consistently attaining the largest maximum SIC The model's performance in low-statistics regions is particularly notable, as it converges significantly faster than scratch models, reaching optimal performance within approximately 1 k steps compared to ∼3 k steps for scratch models Furthermore, EveNet–Full demonstrates strong generalisation to out-of-distribution exotic Higgs decay topologies with limited data (5% of the full dataset), where it outperforms SPANet and training from scratch Robustness and Stability in Precision Measurements EveNet–Full consistently exhibits enhanced stability against detector uncertainties when evaluated on systematic variations, such as jet-energy-scale (JES) shifts and soft missing transverse energy (MET) fluctuations, compared to the model trained from scratch For the quantum-correlation benchmark, EveNet–Full remains highly stable under both JES and MET variations with a spread of precision shift confined within 5 × 10−4 at 68% confidence level, whereas the scratch model exhibits fluctuations that are approximately five times larger The study concludes that EveNet demonstrates a potential event-level foundation model as a flexible and effective framework for a broad range of analyses in HEP The emergent physics-aware structure of the latent space creates new opportunities for interpretable representations and anomaly-sensitive embeddings, positioning EveNet as a concrete instantiation of the event-level foundation model approach The work suggests that adopting this foundation model paradigm can drive a paradigm shift in data-intensive scientific discovery by consolidating traditionally multistage analyses into a single backbone The key contributions of this work are: (1) we introduce EveNet, which is the event-level foundation model for HEP to unify discriminative and generative objectives in a single physics-informed pretraining framework, enabling transfer across heterogeneous downstream tasks (2) we provide the first comprehensive evaluation of a foundation model on CMS Open Data, demonstrating robust generalisation from simulation to experimental dataset (3) we present a systematic ablation study that quantifies the impact of different pretraining strategies and task combinations (4) we release a fully pretrained, ready-to-use EveNet checkpoint to serve as a powerful and versatile starting point for future HEP analyses Data Availability and Code Release The datasets used for pretraining and finetuning, together with the pretrained EveNet model weights, are publicly accessible on at https://huggingface.co/datasets/Avencast/EveNet The source code for EveNet and the scripts for training baseline models are openly available at https://github.com/EveNet-HEP Pretraining configurations can be provided upon reasonable request Acknowledgements We gratefully acknowledge Rikab Gambhir and Radha Mastandrea for their helpful discussions and constructive feedback on the anomaly-detection analysis VM is supported by JST EXPERT-J, Japan Grant Number JPMJEX2509 T.-H.

Paper summary: Kai: So we're talking about how EveNet works across these four settings, where they test its versatility and generalizability.

Mira: They first tested it in a search for a new physics scenario, which is often when you have to scan the whole parameter space.

Lev: That task requires training on and generating hundreds of thousands of datasets just to cover the full parameter space for things like the mass of new particles, which is impractical and usually forces you into compromises <ref:2601.17126#pg2>.

Kai: The paper says that this type of analysis often needs that much data, so training on it becomes something you can't practically do in a standard setup.

Mira: That's why EveNet aims to change how we approach those searches by making the model more efficient at handling that vast parameter space.

Lev: They also looked at an out-of-distribution new physics search scenario, where hypothetical particles predicted by a new physics model aren't even in the pretraining dataset.

Kai: That’s a tough one because you’re asking it to find something completely new without seeing examples of it during training.

Mira: They claim that EveNet still performs well in this scenario, even when the particles are totally outside what it saw before, outperforming models like SPANet and training from scratch <ref:2601.17126#pg2>.

Lev: It seems like the shared representation is key here because it lets the model learn some underlying structure that helps it generalize beyond just memorizing the training examples.

Kai: And then they had this task involving measuring quantum correlations in top-quark pair production, where they found that the synergy between assignment and generating undetectable objects was critical <ref:2601.17126#pg2>.

Mira: That specific task really showed how useful it is when you need both the assignment capability and the ability to generate stuff that isn't visible.

Lev: Then there was collision data anomaly detection, where they validated that this foundation model, trained only on simulated events, still holds strong performance when applied to actual collision data <ref:2601.17126#pg2>.

Kai: So, the paper is showing this model can handle searching for new physics and quantum correlations while still being robust enough to work on real collision data.

Mira: It’s about showing that this event-level foundation model approach can be flexible enough for a broad range of physics goals.

Conclusion: Kai: So we've covered how EveNet is built, how it performs across those various tasks, and what that means for the discovery process in particle physics.

Mira: The title itself suggests this is a foundation model approach applied directly to collision data analysis, aiming to create a single backbone for many different physics questions.

Lev: Essentially, they've consolidated what used to be several separate deep learning analyses into one single framework based on this unified latent geometry.

Kai: This means that instead of training separate models for every new search scenario, you can use this foundation model as a starting point and fine tune it for specific tasks.

Mira: The implication is that we might see a significant acceleration in how quickly physicists can discover new physics because the model handles the complexity of the input data better.

Lev: It’s about making the development process for deep learning-based analyses much more streamlined by having this unified starting point <ref:2601.17126#pg3>.

Kai: The authors also released a fully pretrained, ready-to-use EveNet checkpoint, which is a powerful starting point for future HEP analyses.

Mira: It’s about giving the community a concrete instantiation of the event-level foundation model approach to try out.

Lev: They also pointed out that this emergent physics-aware structure in the latent space creates new opportunities for interpretable representations and anomaly sensitive embeddings <ref:2601.17126#pg3>.

Kai: It positions EveNet as a concrete way to move toward an event-level foundation model framework in high energy physics.

More episodes

← Home