CMS Experience with Anomaly Detection in the L1 Trigger

L1 scouting provides the possibility for unbiased HL-LHC data acquisition and storage for future analysis, but the resulting datasets would be prohibitively large, order 100-1000 PB per data-taking year depending on the kind of information saved. In this task, we propose to apply cutting-edge compression techniques, including nonlinear lossy compression with AI algorithms (e.g., autoencoders) to reduce the L1-scouting dataset size. Autoencoders are also a promising algorithm for anomaly detection, and so they will also be explored for that purpose, that can be already applicable to Run 3 data. For this task, we would also work on optimizing the hardware design of the algorithm (resource consumption and latency), to potentially run it as part of the main L1 trigger system to add scouting, both for Run3 and HL-LHC.