Neural-Network Based b-Tagging in the CMS Level 1 Trigger System

Abstract

The Phase II upgrade of the Compact Muon Solenoid (CMS) detector for the High Luminosity Large Hadron Collider will greatly expand the hardware capabilities of the Level 1 Trigger (L1T) system. With the advent of tracks in the L1T, these upgrades bring the idea of b jet identification within the L1T environment into the realm of feasibility for the first time at CMS. This thesis focuses upon the development of a neural network-based algorithm for the identification of jets originating from bottom quarks in the CMS detector. The algorithm is designed for the timing constraints of the L1T, using seeded cone jet reconstruction of PUPPI algorithm objects as inputs. We present the input data, architecture, and training of the network with performance on simulated tt-bar events. We then focus on the application of the network towards the HH to bbbb decay channel, where our neural network-based trigger displays a 12% increase in trigger efficiency for events useful for measuring Higgs self-coupling. Finally, results are presented for the ongoing progress on implementations of this b-tagging algorithm onto Field Programmable Gate Arrays suitable for the L1T environment. From these results we conclude that b-tagging is possible in the Phase II L1T, a novel achievement that opens the chance to study numerous new physics processes with the first instance of L1T heavy flavor tagging at the LHC.