Anomaly Detection with tree-based autoencoder on FPGAs at Level 1 Trigger of ATLAS

We present a decision tree-based implementation of autoencoder for anomaly detection. A novel algorithm is shown, in which a forest of decision trees is trained only on background and used as an anomaly detector. The fwX platform is used to implement the trained autoencoder on ATLAS’ FPGAs at the Large Hadron Collider. Firmware design with fwX allows for it to stay within the 25ns latency and resource usage constraints demanded by the Level 1 (hardware-based) Topological Trigger system of the ATLAS detector.