Anomaly Detection for CMS L1 trigger at HL-LHC

With the High Luminosity Large Hadron Collider (HL-LHC) set to produce over 140 million proton-to-proton collisions, there needs to be a fast and reliable method to filter for new physics signatures. This is the reason why the LHC is set to employ a trigger system which consists of the L1-trigger system (L1T) and the High-Level trigger (HLT) - where this work focuses on the L1T, which is set to employ autoencoders (AEs) on FPGAs for microsecond period inferencing. In this work, we explore the feasibility of both Convolutional and graph-based autoencoder architectures at the CMS experiment’s L1T for HL-LHC. We show that a Convolutional autoencoder produces great results with AUC scores as high as 92.3%. In order to fit the strict requirements of the L1T, we compress the model, which includes quantisation and pruning. We show that the performance of this compressed version of the Convolutional autoencoder is at the level of the regular, uncompressed, Convolutional autoencoder. We also demonstrate that a GarNet-Dense autoencoder architecture is not suitable for HL-LHC.