FPGA implementation of muon momentum assignment with machine learning at the CMS level-1 trigger

With the advent of the High-Luminosity phase of the LHC (HL-LHC), the instantaneous luminosity of the Large Hadron Collider at CERN is expected to increase up to around 7.5 x 10^34 cm^-2s^-1. Therefore, new strategies for data acquisition and processing will be necessary, in preparation for the higher number of signals produced inside the detectors, that would eventually make the trigger and readout electronics currently in use at the LHC experiments obsolete. In the context of an upgrade of the trigger system of the Compact Muon Solenoid (CMS), new reconstruction algorithms, aiming for an improved performance, are being developed. For what concerns the online tracking of muons, one of the figures that is being improved is the accuracy of the transverse momentum (pT) measurement. Machine Learning techniques have already been considered as a promising solution for this problem, as they make possible, with the use of more information collected by the detector, to build models able to predict the pT with an improved precision. In this Master Thesis, a step further in increasing the performance of the pT assignment is taken by implementing such models onto a type of programmable processing unit called Field Programmable Gate Array (FPGA). FPGAs, indeed, allow a smaller latency with a relatively small loss in accuracy with respect to traditional inference algorithms running on a CPU, both important aspects for a trigger system. The analysis carried out in this work uses data obtained through Monte Carlo simulations of muons crossing the barrel region of the CMS muon chambers, and compare the results with the pT assigned by the current (Phase-1) CMS Level 1 Barrel Muon Track Finder (BMTF) trigger system. Together with the final results, the steps needed to create an accelerated inference machine for muon pT are also presented.