Master thesis defense by Nicolas Palm Perez
Electron Identification Using Machine Learning in the ATLAS Experiment with 2016 Data
Abstract: The identification of electrons in the ATLAS experiment is done using a likelihood method, which is constructed from Monte Carlo simulations. In this work, several machine learning algorithms have been employed for electron identification, as these are expected to be more performant. Unfortunately, the simulations and data have minor inconsistencies which cause a loss in gained performance when ML algorithms are trained on simulation and applied on data.
Tag&Probe selection of electrons from Z->ee events provides purity levels in the range of 30-90%, which is insufficient for training of the ML algorithms. Therefore, a division of variables into two independent sub-classifiers increases the purity levels to above 99%, allowing ML training on data. The performances of the ML classifiers vary significantly (and surprisingly) with energy and direction and give combined improvements of 94-109%. Furthermore, an isolation classifier has been constructed yielding improvement in the range of 100-600%.