Non-prompt background rejection with machine learning for soft lepton searches

Many different physics theories beyond the Standard Model are studied at the LHC. One branch of these theories is concerning supersymmetry and especially models with compressed mass spectra, which could be detected via production of two or three soft leptons and missing transverse momentum. An important background for the soft leptons comes from t¯t events. Lepton candidates in such events can arise either from prompt production, or from non-prompt production and misidentification. The non-prompt background is also referred to as "fakes". The main task of this project was to classify the prompt leptons from the fakes in the t¯t background in the context of new physics searches with the CMS experiment. This was done by training a neural network, which predicts the probability that the lepton is prompt. We found that the neural network succeeded very well in the task.