Unsupervised Machine Learning for Real-Time Anomaly Detection in the ATLAS Level-1 Trigger

High-energy physics experiments at modern colliders demand intelligent trigger systems capable of identifying rare or unconventional events within nanoseconds. We present NomAD (Nanosecond Anomaly Detection), an unsupervised machine learning algorithm developed for real-time event selection in the ATLAS Level-1 Topological (L1Topo) trigger. The algorithm combines a Variational Autoencoder (VAE) with Boosted Decision Tree (BDT) regression to compress and distill deep-learning inference into a firmware-compatible representation suitable for FPGA hardware. Trained on ATLAS Run-3 data using low-level L1 muon features, NomAD identifies anomalous dimuon events beyond standard trigger logic. The resulting anomaly score enables mass-agnostic and model-independent selection of low-momentum dimuons with clear signal–background separation. Firmware tests demonstrate sub-25 ns inference latency and successful implementation using standard FPGA design tools. This work presents the complete ML-to-firmware workflow, its performance on B-physics signals, and its implications for fast, inclusive triggering in high-throughput collider environments.