Machine Learning for the Leptoquark Search Using CERN ATLAS Data

In this thesis, we improve the cross-section limit for pair production of third-generation scalar Leptoquark decaying into a top quark and a tau-lepton and design a method to predict its mass. Events are selected if they have two light leptons (electron or muon) of the same sign and exactly one hadronically decaying tau-lepton. Algorithms from two machine learning categories widely used for tabular data classification, gradient boosting decision trees and deep neural networks, are deployed to analyze simulated data for Leptoquark masses from 300 to 2000 GeV. The data for all available masses are combined to show that one universal classifier can be used for all Leptoquark mass cases. The dependence of the performance on the number of features and the size of the simulated data set is demonstrated. Finally, the TRExFitter program developed in CERN is used to achieve reliable results for cross-section limit calculation. Additionally, we study how to recognize Leptoquark mass using another connected classifier.