Lepton identification for the CMS experiment with recurrent neural networks

Abstract

From the collision of two protons at the large hadron collider (LHC) massive vector bosons, W and Z bosons, that are sensitive probes of the electroweak interaction, can emerge and subsequently decay into lepton pairs. Among the leptons in such collisions, electrons and muons are efficiently detected by the CMS experiment at the LHC. The determination of the origin of these leptons is an important target of the event reconstruction.This thesis describes an approach to lepton identification using Deep Learning techniques that incorporate the information from all reconstructed particles in the vicinity of the lepton candidate. This leads to a better identification efficiency, as the surrounding particles contain information about the leptons origin. Deep neural networks were trained for electron and muon identification and their performance was evaluated. The training and performance evaluation were done separately for the LHC data taking periods of the years 2016, 2017, and 2018.For optimal performance, the fully connected deep neural networks were supplemented by long-short term memory units. The results show that this approach has a significantly better classification efficiency than traditional approaches, especially for low pT lepton candidates. The new classifier is subsequently used to improve the background suppression in a search for supersymmetry with low-energetic leptons, improving the limit on the masses of superpartners, on average by about 5 GeV.Finally, the relative importance of the input variables was determined and the timing and memory usage of different network configurations were compared.