Advanced machine learning algorithms for heavy flavour jets identification & study of the Higgs boson couplings to the charm and beauty quarks with the ATLAS experiment

dentifying the flavour of jets plays an essential role in many ATLAS analyses. This subject is extensively discussed in this thesis, with a complete review of the algorithmic developments carried out by the ATLAS Collaboration from 2020 to early 2024. Increasingly sophisticated machine learning models called taggers have been developed for this purpose. The classical approach relies on a hierarchical construction combining low-level physically motivated taggers with a deep set network or a recurrent neural network as inputs to a high-level network predicting the jet flavour. A more flexible design leveraging a single network to deliver state-of-the-art performance has recently been introduced. The core of this network is either a graph attention network or a transformer encoder unit. Expert knowledge is passed to the model by optimising multiple tasks, with different physics input types analysed in a multimodal framework. The design and training of these taggers are reviewed, as well as a hyperparameter optimisation study using the µP parametrisation and the µTransfer technique from the ML literature on large language models.

Following the discovery of the Higgs boson by the ATLAS and CMS Collaborations in 2012, increasingly refined measurements of the new particle have been performed. The leading production modes and the decay mode to third-generation fermions and gauge vector bosons of the Higgs have now all been measured. Attention is shifting to second-generation fermions, such as the c-quark, and precision differential cross section measurements. This thesis presents a combined search for the H → cc¯ coupled with a differential measurement of the H → bb¯ in the associated VH production mode. The analysis exploits the full 140 fb^-1 proton-proton collision luminosity collected in Run 2 by the ATLAS experiment at a centre-of-mass energy of 13 TeV. Combining the decay modes allows for a coherent joint analysis strategy that improves the constraining of the shared backgrounds. Flavour taggers are used to identify candidate b- and c-jets to reconstruct the Higgs. The full p_T spectrum is covered, with the two candidate jets resolved at low momentum and a single merged boosted signature at high momentum. Three leptonic channels are defined based on the number of electrons and muons in the final state. Dedicated Boosted Decision Tree discriminants are deployed to increase the signal sensitivity. The analysis, which is still blinded, is expected to yield a 95% CL_s upper limit for the VH(H → cc¯) signal strength of 11.1 times the Standard Model prediction. The VH(H → bb¯) expected signal strength is 7.9 σ over the background-only hypothesis, with the WH and ZH productions respectively measured with expected significances of 5.5 σ and 6.2 σ. A standard cross section template measurement is performed for VH(H → bb¯), in bins of p_T and number of additional jets.