BOOSTING ONLINE RECALIBRATION OF PHYSICS OBJECTS FOR THE 40 MHZ SCOUTING SYSTEM AT CMS

This work explores the accelerator capabilities of a modern FPGA board for fast Neural Network inference. The physics application involves the analysis of raw data collected at the Level-1 Global Muon Trigger of CMS detector, with the goal of developing a real-time recalibration of the Level-1 muon primitives to match as closely as possible the parameters of the physics objects of interest. The application is deployed in the 40 MHz Level-1 Trigger Scouting demonstrator of the CMS experiment. The project is organized in two parts. As a first step, the dataset for training the network, the network architecture and the tools are presented. Depending on the configuration of the hyperparameters, multiple network models are created to optimize the recalibration performances, and tested emulating the hardware implementation. The second and last step focuses on the FPGA resource usage optimization and on the deployment of the network on a Xilinx VCU128 evaluation board. The validation of the implemented optimal design in a test setup is finally discussed