Leveraging Machine Learning for Enhanced FPGA- Based Tau Triggering and Combined H ->( b b / c c ) Analysis in the ATLAS Experiment

The exceptionally good agreement between the predictions of the Standard Model of particle physics and the measurements made by the experiments at the Large Hadron Collider at CERN requires collecting more and more data to set increasingly tighter upper limits on observables in the search for physics beyond the Standard Model. Ongoing upgrades aimed at increasing the amount of data delivered to the experiments are pushing the limits of the currently used algorithms that specialize in separating signal and background in order to produce clean samples for physics analyses. Recent advances in machine learning technology allow the development of a new generation of these algorithms that, more often than not, outperform their heuristic-based counterparts in online, real-time triggering scenarios and offline data analysis. In particular, fast inference of trained machine learning models in FPGAs has been recently made achievable by the introduction of new tools.

This thesis explores these next-generation algorithms in the context of online and offline analysis. The online analysis concerns triggering on t leptons at the Level-1 ATLAS trigger using boosted decision trees, and the offline one explores the utilization of graph neural networks and boosted decision trees in the VHbb(cc) Legacy analysis, which sets the most stringent constraints to date on the coupling of the Higgs boson to b- and c-quarks.

The work presented in this thesis results in the first operational machine learning model used for data taking in the Level-1 trigger of the ATLAS experiment, based on a fully evaluated boosted decision tree, with no special hardware requirements, for the identification of t leptons, opening the gate for the next generation of trigger algorithms at the ATLAS experiment.