Semi-visible jets, energy-based models, and self-supervision

Luigi Favaro, M. Krämer, Tanmoy Modak, Tilman Plehn, Jan Rüschkamp

Abstract

DarkCLR is a self-supervised framework designed to detect semi-visible jets—elusive signals of new physics—at the Large Hadron Collider (LHC) without requiring labeled signal data. By combining contrastive learning with a novel "anomalous augmentation" strategy, DarkCLR creates a representation space that is both physically invariant and highly sensitive to signals that standard classifiers often miss. This work demonstrates superior background rejection and robustness across various dark sector models compared to traditional anomaly detection methods.

Introduction: The Challenge of Semi-Visible Jets

In the search for physics beyond the Standard Model, researchers frequently analyze "jets"—sprays of particles produced in high-energy collisions. While many searches target specific predicted signals, "Hidden Valley" models suggest a dark sector that interacts weakly with the Standard Model. These models produce semi-visible jets: sprays where some particles are visible to detectors while others remain invisible.

Because these jets closely resemble standard Quantum Chromodynamics (QCD) background noise but contain "missing" components, they are notoriously difficult to distinguish. Traditional anomaly detection often relies on density estimation, which can be sensitive to data representation and coordinate transformations. DarkCLR aims to provide a model-agnostic tool that is robust, invariant to physical symmetries like rotations, and capable of identifying these "holey" jets without prior knowledge of their exact structure.

Methodology

The DarkCLR framework is built upon Contrastive Learning of Representations (CLR), utilizing a two-pronged approach to learn meaningful features from jet data.

Self-Supervision and Augmentations

Instead of relying on labels, the model learns by comparing different versions of the same data point.

  • Positive Pairs: The model is presented with a jet and a "physically augmented" version (e.g., rotated or translated). It learns to map these to similar representations, ensuring invariance to these symmetries.
  • Anomalous Augmentation: This is the core innovation. The authors create "anomalous" versions of background jets by randomly dropping constituents, simulating the invisible fraction characteristic of semi-visible jets. The model is trained to push these representations far away from standard background jets in the latent space.
Visualizing physical and anomalous augmentations used in DarkCLR training. The top row shows physical symmetries (rotation, translation), while the bottom row illustrates the "dropping" strategy used for anomalous augmentation.

Architecture

The model employs a Transformer encoder to process unordered sets of particles within a jet. This is followed by a projection head that maps the encoded features into a 512-dimensional latent space.

The DarkCLR architecture: a Transformer encoder handles constituent sets $(N_c, 3)$, followed by a feed-forward (FF) head to produce the latent representation $z$.

Anomaly Scoring

Once trained, two scores are used for signal detection:

  1. $s_{CLR}$: The norm of the representation vector in the latent space.
  2. $s_{NAE}$: The reconstruction error from a Normalized Autoencoder (NAE) trained on the learned representations. The NAE serves as a robust density estimator, flagging data points that deviate from the learned background manifold.

Experimental Setup and Datasets

The model was evaluated using several datasets:

  • Background: 100,000 QCD di-jet events representing standard light quarks.
  • Signal (Aachen Dataset): Simulated semi-visible jets from a Hidden Valley model ($m_{Z'} = 2$ TeV, $m_{q_d} = 500$ MeV) with an invisible fraction ($r_{inv}$) of 0.75. The dark sector includes dark pions ($m_{\pi} = 4$ GeV) and dark rhos ($m_{\rho} = 5$ GeV).
  • Robustness Benchmarks: Variations of the signal with different invisible fractions ($r_{inv} = 0.2, 0.5$) and dark meson masses (10 GeV, 20 GeV) to test generalization.

Simulation was performed using MadGraph5 for hard processes, Pythia 8.2 for showering, and Delphes 3 for detector simulation. Jets were reconstructed using the anti-$k_t$ algorithm with $R=0.8$.

Key Results and Insights

  • Superior Background Rejection: DarkCLR achieved a background rejection ($\epsilon_B^{-1}$) of 59 at a signal efficiency of 0.2. This significantly outperforms previous state-of-the-art methods, such as NAE on jet images (41) or Variational Autoencoders (36).
  • Stability and Robustness: The Normalized Autoencoder ($s_{NAE}$) provided more stable ROC curves than raw latent scores. Furthermore, while supervised classifiers often fail when signal parameters change, DarkCLR maintained high performance across various dark sector masses and invisible fractions, demonstrating its value as a general-purpose discovery tool.
  • Invariance by Design: By explicitly training for rotational and translational invariance, the model focuses on the intrinsic internal structure of the jet rather than uninteresting coordinate variations.

Conclusion and Future Directions

DarkCLR demonstrates that "breaking" background data through anomalous augmentation is a powerful proxy for learning the features of new physics. However, limitations remain. The current strategy assumes uniform constituent dropping, whereas real dark sector hadronization may be more complex. Future work will focus on non-uniform decays, accounting for more intricate detector effects, and eventually applying DarkCLR to real LHC data to bridge the gap between simulation and discovery.

Summary updated

Abstract

Dataset published with "Semi-visible jets, energy-based models, and self-supervision", Favaro L. et al. It consist of a QCD dijets background file "qcd_constit.h5" and a signal file "aachen_constit.h5". The signal model has been proposed in "Strongly interacting dark sectors in the early Universe and at the LHC through a simplified portal", Bernreuther E. et al. The signal model has a dark mediator with mass m_Z'=2TeV and dark quarks with mass m_qd=500MeV charged under dark SU(3)_d. The hadronization of the dark sector consist of dark pions with mass m_pi=4GeV and dark rhos with mass m_rho=5GeV. Simulation: Madgraph + Pythia + Delphes simulation; (E, px, py, pz) of the first 200 constituents, pT ordered; most jets have less than 70 constituents, the remaining ones are zero-padded; Fastjet reconstructed fat jets using anti-kt algorithm with R=0.8; kinematic cuts: 100<pT<300, |\eta<2| ; semivisible jets are matched to the parton level dark quark; pheno parameters: invisible fraction r_inv=0.75, m_pi=4GeV, m_rho=5GeV. Other publications which use this dataset: [1] "A Normalized Autoencoder for LHC triggers", Dillon B. et al., arXiv:2206.14225 [2] "What's anomalous in LHC jets", Buss T. et al., arXiv:2202.00686
BibTeX
@inproceedings{favaro2025semi,
  author = {Luigi Favaro and M. Krämer and Tanmoy Modak and Tilman Plehn and Jan Rüschkamp},
  title = {{Semi-visible jets, energy-based models, and self-supervision}},
  booktitle = {SciPost},
  year = {2025},
  doi = {10.21468/scipostphys.18.2.042},
}
Figure