Master thesis defence by Stefan Hasselgren – University of Copenhagen

Discovery Center > Discovery calendar > Calendar 2018 > Master thesis defence ...

Master thesis defence by Stefan Hasselgren

Title: Using machine learning for improving identification of electrons in the ATLAS experiment

Abstract : Identification of electrons, which are an essential part of the ATLAS physics program, is currently done using a likelihood (LH) based method. In this work, a machine learning based approach have been implemented, first in Monte Carlo (MC) and later in real data (Data).

The MC based classifiers are implemented in several versions, some based on the same variables as the LH and others based on extra variables. Sub-classifiers are trained in separate variable groupings in calorimeter, inner detector and isolation. The calorimeter and inner detector sub-classifiers are later combined into one classifier comparable to the LH. This approach yields an overall improvement of 323% over the LH.

These improvements decrease when testing the MC based classifiers in data. Based on these results, the MC based classifiers are used to purify data samples containing data from 2016 at √s = 13 TeV. The data samples contain mislabelled events, which hinders the training of data based classifiers, so the MC classifiers are used to remove mis-labelled events from the data samples, yielding data sample purities of > 99%. The Data based classifiers, trained on the purified data samples, yield an overall improvement of 197% over the LH.