Unsupervised Non-Parameterized Searches for Anomalous Events at CMS Using Wasserstein Normalizing Autoencoders

The Compact Muon Solenoid experiment relies on the Level-1 (L1) trigger to select the most interesting events from roughly one billion proton–proton collisions per second. Traditional L1 triggers apply fixed, parameterized thresholds on kinematic quantities such as energy or momentum, which, while effective for known Standard Model (SM) processes, limit sensitivity to rare or unexpected Beyond Standard Model (BSM) signatures. The AXOL1TL algorithm represents the first unsupervised, non-parameterized anomaly detection trigger at CMS, capable of identifying outlier events directly from data distributions in real time.

This work introduces the Wasserstein Normalizing Autoencoder (WNAE), a next-generation model for AXOL1TL that employs the Wasserstein distance (Earth Mover’s Distance) as its training objective to improve stability and the expressiveness of the learned SM manifold. The architecture is optimized to meet latency and FPGA resource constraints of the L1 trigger. Preliminary studies demonstrate enhanced anomaly detection efficiency and improved background reconstruction compared to the current VICReg–VAE implementation. Ongoing efforts focus on accelerating training through neural EMD estimators and developing firmware for deployment in the 2026 CMS data-taking run.