Self-Supervised Anomaly Detection on Real LHC Data: Contrastive Learning with AspenOpenJets

We present the first application of contrastive self-supervised anomaly detection trained on real LHC jet constituent data. Using the AspenOpenJets dataset (∼178M jets from CMS Open Data at √s = 13 TeV), we train a transformer encoder with the NT-Xent objective and physics-informed augmentations, then score anomalies via k-nearest neighbor distance in the learned representation space. Evaluated on the LHC Olympics 2020 benchmark, the contrastive objective outperforms an autoencoder baseline by 0.060 AUC for the real-data model and 0.082 AUC for the simulation model on 2-prong BSM signals, and by 0.12 on 3-prong signals; this is the dominant design choice. Real CMS data provides substantially faster early-stage convergence than simulation training (0.065 AUC advantage at epoch 10), though simulation catches up asymptotically. On matched features, the real-data advantage is small but consistent (+0.013–0.016).Detector-specific features (impact parameters, charge) show a signal-topology-dependent trade-off, and pT smearing augmentation is rendered redundant by real detector noise. This work establishes the first discriminative application of AspenOpenJets and provides systematic comparison of real-data vs. simulation-trained contrastive representations for anomaly detection.