A Machine Learning approach for Heavy Neutral Leptons search from Ds meson decays in the CMS experiment

This thesis presents an application of Machine Learning (ML) techniques to search for Heavy Neutral Leptons (HNL) in Dₛ meson decays in the CMS experiment. The specific decay under study is Dₛ⁺ → N(→ μ π) μ⁺, where N is the HNL and final state muons can have the same charge, allowing for lepton number violation. The signal signature comes from the displaced N → μ π vertex while the background comes from accidental combination of muons and a track into a common vertex. This work is carried out in the context of an ongoing analysis, that relies on a cut−based event selection, and its aim is to explore different ML approaches to improve the event selection. The training of the ML models relies on Monte Carlo generated samples for background and different signal mass hypotheses. These samples reproduce the data−taking condition of the CMS experiment during 2018 Run 2 with √s = 13 TeV. Several ML models have been trained with three different algorithms: boosted decision trees (BDT), gradient boosted decision trees (XGB) and artificial neural networks (ANN) and their performances have been evaluated. The best performance is obtained by the XGB algorithm. The improvement in the event selection achieved with this ML approach translates into higher significances, in a range from 14 to 20%, over the cut−based approach. The results of this work show that the developed ML−based strategy is an effective contribution to the ongoing CMS analysis.