Machine Learning in Particle Physics Quark versus Gluon Jet Discrimination

This is a multidisciplinary project, adapting, applying, and assessing modern data analysis techniques and advanced machine learning in the field of Particle Physics. The objective is to derive a new technique to perform quark versus gluon jet discrimination with the ATLAS detector, one of the main experiments on the Large Hadron Collider at CERN. Different machine learning architectures targeting generated level simulations (not accounting for detector effects) have recently been explored in the literature with promising results. Among them, the Junipr approach learns a probabilistic model scaffolded on a leading-order model of Physics (jet factorisation algorithm) by recurrent learning and deep neural network mapping. It was proven in 2019 to lead to state-of-the-art quark versus gluon jet discrimination on generated level simulations, with the added quality that the model thus learnt is interpretable. The project aims to implement and adapt this method into an efficient discriminator for the ATLAS collaboration, using the particular conditions of its detector and simulations comparable to the data collected (the reconstructed level). An important constraint added to these restrictions is to work with low-level data to pave the way fo future research on a possible trigger application. The results show that a Junipr approach given low-level information succeeds in outperforming boosted decision tree and deep neural network models using high-level information, in a specific energy region. The adapted framework is also shown to obtained similar performance to that of the original paper while restraining to low-level variables and fully reconstructed simulations. Finally, Junipr stays globally competitive compared to the benchmarking models in all energy regions of interest, suggesting a trigger implementation should be considered for the ATLAS experiment.