Real-time Anomaly Detection at the L1 Trigger of CMS Experiment

The Compact Muon Solenoid (CMS) experiment studies these collisions to uncover potential Beyond Standard Model (BSM) physics and precisely measure rare Standard Model (SM) processes [2]. While the high collision rate at the LHC increases the probability of producing and detecting rare processes, the nearly 100 million channels of the CMS detector also generate an enormous amount of data [10, 14]. Only a small fraction of the 40 MHz proton-proton collision events--around 1,000 per second--can be stored for detailed offline analysis. To meet this stringent data reduction, events are selected using a two-tiered trigger system. The first level (L1), composed of custom hardware processors built with field-programmable gate arrays (FPGAs), uses information from the calorimeters and muon detectors to select events at a rate of around 100 kHz within a fixed latency of 4 [14]. The second level, the high-level trigger (HLT), consists of a processor farm running optimized event reconstruction software, reducing the rate to around 1 kHz before storage [10].