Identification of Hadronic Tau Lepton Decays with Domain Adaptation using Adversarial Machine Learning at CMS

This thesis reports improved machine learning-based techniques to discriminate genuine decays of tau leptons into hadrons and a neutrino against all main backgrounds at the CMS experiment. The deep convolutional neural network, DeepTau, used for tau identification by physics analyses of the 2016-2018 data-taking period at CMS, shows a sizeably different performance on collider data versus Monte Carlo simulations. This effect is particularly prominent for regions of parameter space that have high genuine tau purity. The effects of this mismodelling on discrimination against quark and gluon jets are reduced by introducing domain adaptation into the training workflow. This approach was validated by comparing the performance of the resulting network on proton-proton collision data and simulated events. The use of these adversarial machine learning techniques reduced the discrepancies from 13.3% to 0.80% in the region where the purity of hadronic taus is expected to be above 96%, while having no significant impact on tau identification performance.