Interpretable Anomaly Detection for Collider Physics Using Contrastive Learning

Anomaly detection in collider physics faces two significant challenges: interpreting the meaning of anomaly scores and preventing these scores from correlating with trivial factors like energy scale or object multiplicity. A new framework, Organized Representation via Contrastive learning for Anomaly detection (ORCA), addresses these issues by introducing a two-stage process. First, ORCA employs supervised contrastive learning to develop a highly organized embedding space. This space is learned across a diverse range of known physics processes, ensuring that distinct processes occupy separate, identifiable regions. In the second stage, a standard autoencoder is applied within this learned embedding space to generate event-level anomaly scores. This approach helps to disentangle true anomalies from background variations. Simulations mimicking conditions at the High-Luminosity Large Hadron Collider demonstrated that ORCA substantially improves both the breadth and depth of sensitivity to new physics signals compared to a baseline autoencoder. Beyond enhanced detection, the contrastive embedding makes the anomalous samples interpretable. By fitting templates of known processes to the embedding distributions, ORCA can attribute anomalous events to specific physics processes with quantified uncertainties, even recovering signals not included in the embedding's training. This capability establishes ORCA as a promising route for interpretable anomaly detection in collider experiments, providing higher-dimensional physics information for downstream statistical analysis.