The Higgs, the Beauty and the Charm: Improving Jet Flavour-Tagging and Higgs Boson Measurements with Graph Neural Networks in the ATLAS Experiment at the LHC.

Testing the Yukawa couplings of the Higgs boson with fermions is essential to understanding the origin of fermion masses. Higgs boson decays to quark pairs are an important probe of these couplings, and of properties of the Higgs boson more generally. This thesis presents a novel measurement of the H -> bb and H -> cc decay rates in the VH production channel, where V stands for the vector bosons W and Z, using a dataset comprising 140 fb^−1 of proton-proton collision data collected by the ATLAS experiment at the Large Hadron Collider between 2015 and 2018. The measurement is expected to provide the most precise estimation of the VHbb process rate to date and the most stringent 95% CL upper limit on the VHcc process rate set by the ATAS Collaboration. Additionally, it is expected to mark the first observation of the WHbb process, and the first observation of the VZcc process by the ATLAS Collaboration. Measuring the H->bb and H->cc processes requires efficient algorithms, termed jet flavour-tagging algorithms, to identify hadronic jets containing b or c hadrons within the detector. In this thesis, GN2+Clus, an innovative jet flavour-tagging algorithm based on transformers, is introduced. By incorporating the clusters of silicon hits in addition to the reconstructed tracks produced by charged particles in the ATLAS Inner Detector, GN2+Clus improves the performance of the ATLAS state-of-the-art algorithm GN2 of significant factors, with peaks of ∼ 80%(∼ 10%) of additional rejection when used to identify b-jets (c-jets) with transverse momenta at the TeV scale. A novel technique to parametrize the efficiency of the jet flavour-tagging algorithms using Graph Neural Networks is also introduced and implemented in the context of the VH->bb/cc measurement. This technique, which enables better usage of the simulated statistics within the analysis, surpasses previous approaches based on 2-dimensional efficiency maps.