Energy-Flow Pretraining: Light-Ray Operator Targets for Self-Supervised LHC Foundations

Conformal collider physics is now directly testable at the LHC, with Lee et al. (2025) deriving a factorization theorem for light-ray density matrices and computing scaling of multipoint energy correlators against CMS data. We can leverage these rigorous QCD structures as self-supervised pretext tasks: pretrain networks to predict k-point energy correlators, their NLL scaling exponents, and twist-2 spin-J light-ray operator expectation values for high-pT jets, given masked or permuted tokens (in the spirit of Visive et al. 2025). This differs from existing token reconstruction approaches by anchoring representation learning to EFT-controlled quantities rather than generic reconstruction losses. After pretraining, we fine-tune for model-independent anomaly detection (e.g., contrastive embeddings as in Metzger et al. 2025, or likelihood-ratio-free GOF tests). Because the pretext tasks encode QCD’s Lorentzian dynamics and OPE structure, the learned representation should better disentangle “expected” QCD fluctuations from genuine beyond-SM distortions, improving discovery sensitivity while providing interpretable axes tied to operator dimensions. Impact: physics-grounded foundation models for collider data that carry built-in interpretability and calibratability, opening a path to precision-informed anomaly searches and operator-resolved deviations.

References:

1. Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC. Kyle Metzger, Lana Xu, Mia Sodini, T. K. Årrestad, K. Govorkova, Gaia Grosso, Philip Harris (2025).
2. Event Tokenization and Next-Token Prediction for Anomaly Detection at the Large Hadron Collider. A. Visive, P. Moskvitina, C. Nellist, R. D. Austri, Sascha Caron Institute of Physics, U. Amsterdam, Amsterdam, The Netherlands, Nikhef, Dutch National Institute for Subatomic Physics, High Energy Physics, R. University, Nijmegen, Instituto de F'isica Corpuscular, IFIC-UVCSIC, Paterna, Spain. (2025).
3. Conformal collider physics meets LHC data. Kyle Lee, Bianka Meçaj, I. Moult (2022). Physical Review D.