Machine Learning for Real-Time Flavour Tagging and Background Modelling in 4b Final States with the ATLAS Experiment

This thesis demonstrates the application of advanced machine learning techniques to improve the identification and analysis of b-jet final states within the ATLAS experiment at the Large Hadron Collider. Two complementary studies are presented, addressing challenges at both the online event selection stage and in offline searches for new physics. The first study develops graph neural network (GNN)-based algorithms for b-jet identification in the high-level trigger (HLT). Building on the DL1d tagger used in early Run 3 data taking, the GN1 tagger was deployed in 2023, achieving up to a factor of two improvement in light-jet rejection compared to Run 2 performance. This directly reduced trigger rates for HH –> 4b analyses while preserving signal efficiency. The second study presents a search for a new resonance X decaying into a scalar S and a Higgs boson, with both subsequently decaying to bb. Using 126.9 fb-1 of pp collision data at √s = 13 TeV from Run 2, the analysis focuses on the resolved 4b final state in the mass ranges mX ∈ [300, 3000] GeV and mS ∈ [70, 2500] GeV. To overcome challenges posed by large QCD multijet backgrounds, data-driven background modelling techniques based on normalizing flows and Gaussian processes are employed. No significant excess above the Standard Model expectation is observed, and 95% confidence-level upper limits on σ(pp –> X –> SH –> 4b) are set, ranging from 2.5 to 1978.9 fb. Together, these studies illustrate how modern machine learning approaches enhance real-time b-jet identification in the trigger system and enable robust analyses of complex b-jet signatures in the search for new physics.