Browsing by Author "Lieberman, Benjamin"

Now showing 1 - 1 of 1

The Use of Semi-Supervised Machine Learning Techniques in the Search for New Bosons with the ATLAS Detector
(University of the Witwatersrand, Johannesburg, 2024-06) Lieberman, Benjamin; Mellado, Bruce
Since the completion of the Standard Model, with the discovery of the Higgs Boson, there has been a significant shift in the exploration of new physics to explain deviations between model simulated data and that produced at the Large Hadron Collider. These investigations are greatly aided by the integration of advanced machine learning techniques. Machine learning methods offer powerful solutions for complex collider physics challenges. However, these models often depend on simulations that might not fully align with actual physical phenomena. In order to remove this model dependency, semi-supervised classifiers can be used. This solution, however, is not without challenges. In this thesis, an evaluation of the use and limitations of semi-supervised classifiers in particle physics is presented. This is achieved by using a well constrained di-lepton final state dataset to assess the efficacy and ability of the technique, compared to its supervised counterpart. Following this baseline study, the S(150) ! Zg final state, motivated by the multi-lepton anomalies at the LHC, is used to perform narrow resonance searches with semi-supervised Neural Network (NN) classifiers. This work details the methodology and outcomes of a frequentist study aimed at quantifying the extent of a trials factor introduced by semi-supervised NN classifiers. This involves an in-depth analysis of the rate of statistical fluctuations generated in the training of semi-supervised techniques. A secondary component of this study is the evaluation of machine learning based data generators, with emphasis on the Wasserstein Generative Adversarial Network (WGAN), to produce large quantities of realistic data for physics analyses. The results of this investigation into semi-supervised techniques, firstly validates the efficacy and ability of these techniques to classify particle events. This is followed by the frequentist study results, which substantiation that the trials factor remains suitably managed with the use of semi-supervised NN classifiers. The insights derived from this research pave the way for enhancing the reliability of upcoming resonance searches, underscoring the potential of semi-supervised learning in searches for narrow resonances.