Deep learning inference on FPGAs for the CMS Level-1 Trigger Scouting at the LHC

The LHC delivers proton-proton collisions to the CMS detector at a 40 MHz bunch crossing frequency; the CMS Level-1 trigger drastically reduces this rate to 100 kHz to read out and store events effectively. The CMS Level-1 Scouting is a novel system, parallel to the trigger, that instead captures the Level-1 intermediate data at the full crossing rate and carries out online analyses based on these limited-resolution data. The 40 MHz Scouting system can provide high statistics samples for trigger diagnostics and enable the study of otherwise inaccessible signatures. In this context, deep learning algorithms implemented on FPGAs can be used to perform close-to real-time analysis, which could help identify potential physics signals. This thesis studies the performance of deep neural networks to infer and recalibrate the Level-1 Trigger quantities, using the reconstructed offline parameters as the target, and concurrently classify the trigger muons pairs. The Micron Deep Learning Accelerator (MDLA) is used to run models on the Micron Technology's SB-852 FPGA-based processing board. The strategies adopted and results are presented together with a description of the Scouting demonstrator and the framework used.