Year: 2026
Type: preprint
Abstract: Machine learning (ML) based anomaly detection offers a promising, model-agnostic alternative to traditional analyses of LHC data, with supervised contrastive learning providing powerful low-dimensional neural embeddings of high dimensional physics observables. However, the propagation of systematic uncertainties through learned representations remains poorly understood, severely limiting the application of these methods to real analyses. We address this gap with a likelihood-based regularization that fine-tunes the structure of contrastive latent spaces to covary predictably with systematic distortions, returning a model of such distortions that can be directly exploited in a downstream profile-likelihood test for anomaly detection. Using an emulated CMS Level-1 trigger data, we demonstrate that our method successfully captures the effect of jet momentum scale uncertainty via a linear latent model while retaining sensitivity across signal benchmarks, improving interpretability, robustness, and scalability of the anomaly detection task.
Source: arXiv (Cornell University)
Authors: Tong, Shelley, Harris, Philip, Grosso, Gaia
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
Funder: National Science Foundation
OpenAlex ID: W7166853512