Optimizing the Search for Charged Higgs Boson Decaying into a Neutral Higgs Boson Using the ATLAS Detector at CERN

The search for physics Beyond the Standard Model (BSM) is a primary focus of particle physics. A key target in these searches is the charged Higgs boson ($H^\pm$), a particle predicted by well-motivated theoretical frameworks such as the Two-Higgs-Doublet Model (2HDM). This thesis presents an optimized search for the $H^\pm$ produced in association with a top and a bottom quark ($tbH^\pm$), focusing on the decay channel $H^\pm \to W^\pm h$ in the complex $2\ell_{\text{SS}}1\tau_{\text{had}}$ final state. The primary challenge of this analysis is the discrimination of the rare signal signature from large and complex Standard Model backgrounds at the Large Hadron Collider.

A full-chain analysis was performed using 140 fb$^{-1}$ of proton-proton collision data at $\sqrt{s} = 13$ TeV recorded by the ATLAS detector. To enhance the signal-background separation, a novel, scale-based approach was developed, employing two distinct XGBoost models optimized for low-mass and high-mass signal hypotheses. The performance of these models was extensively compared against the attention-based TabNet deep learning architecture. Data-driven template fits were implemented to provide a robust estimate of the significant non-prompt lepton background.

The analysis is expected to set a 95% confidence level upper limit on the charged Higgs boson production cross-section, excluding masses below 800 GeV for the targeted 2HDM Type-II scenario with $\tan\beta = 20$. The optimized methods developed in this work yield a substantial gain in sensitivity, resulting in expected limits that are more than twice as stringent as those from the previous analysis using the same dataset. This demonstrates the powerful impact of the tailored machine learning strategy and establishes a new benchmark for sensitivity in this search channel.