Utveckling av djupa neurala nätverk för identifiering av elektroner med CMS-experimentet vid LHC

This report addresses the need for improved electron identification in CMS measurements, in particular in analyses such as H → ZZ → 4ℓ.A broad range of approaches is investigated, from traditional Boosted Decision Trees to modern Deep Neural Network architectures. Beyond the high-level detector observables employed in current CMS algorithms, this study explores the use of lowerlevel detector information, including calorimetric energy deposits processed with Convolutional Neural Networks and Particle Flow candidate information exploited with Graph Neural Networks to characterize the local environment of the electron. The performance of the different models is evaluated using Monte Carlo samples of electrons originating from Drell–Yan + jets events. Electrons from Z/γ∗ → e+e− decays are used to define the signal, while misidentified electrons predominantly arising from additional jets constitute the main background. The results demonstrate that several of the proposed machine learning approaches achieve improved electron identification performance with respect to the current CMS XGBoost model, both in the high and low pT regions.