ORCA makes particle physics anomalies interpretable with AI
New ORCA framework boosts anomaly detection sensitivity 10x while making results explainable.
Researchers Haoyi Jia, Sagar Addepalli, and Julia Gonski have developed ORCA (Organized Representation via Contrastive learning for Anomaly detection), a novel AI framework that revolutionizes anomaly detection in particle physics experiments.
ORCA employs a two-stage architecture: first, it uses supervised contrastive learning to organize particle collision events into physics-informed embeddings where similar processes cluster together. Then it applies autoencoders to these embeddings to generate anomaly scores. Crucially, this contrastive embedding space enables interpretable anomaly attribution - events can be attributed to specific known physics processes with quantified uncertainties through maximum-likelihood template fitting.
In tests on simulated High-Luminosity Large Hadron Collider data, ORCA demonstrated significant advantages over traditional autoencoder baselines, showing both improved sensitivity to new physics signals and enhanced interpretability of results. The framework successfully recovered injected signal yields even when those signals weren't included in training, and could characterize novel signals by their closest known physics analogs.
- ORCA uses supervised contrastive learning to create physics-aware embeddings for particle collision events
- Achieves 10x better sensitivity to new physics signals while making anomalies interpretable
- Successfully recovers injected signals excluded from training and characterizes novel physics
Why It Matters
ORCA turns black-box anomaly detection into explainable physics discovery, bridging AI innovation with fundamental research needs.