Optimised displaced object identification and search for long-lived dark matter candidates in $H$ → $aa$ decays with the ATLAS detector

Many extensions to the Standard Model predict new long-lived particles (LLPs) whose existence can be probed by the Large Hadron Collider (LHC). LLPs produce unique detector signatures, and require novel search strategies and analysis techniques to aid in their discovery. This thesis presents a search for light, pseudoscalar LLPs, performed using 140 fb$^{−1}$ of proton-proton collision data with $\{sqrt{s}$ = 13 TeV recorded at the ATLAS detector between 2015 and 2018. The search targets hadronically decaying LLPs a with masses ranging from 5 to 55 GeV. Benchmark models with LLP pair production from exotic decays of the Higgs boson to LLPs $H$ → $aa$ and models featuring long-lived axion-like particles (ALPs) are considered. No excess over the expected background is observed. Upper limits are set on branching ratio of the Higgs boson to pairs of LLPs, the cross-section for ALPs produced in association with a vector boson, and, for the first time, on the branching ratio of the top quark to an ALP and a u/c quark. The limits for the $H$ → $aa$ model are the most stringent to date for LLP masses $m$ $_{a}$ < 40 GeV and lifetimes 10$^{−1}$ - 10$^{−3}$ m. Future LLP searches will rely on the efficient reconstruction and identification of displaced objects to maximise discovery potential - this thesis contains further research to that effect. Key performance studies of the large radius tracking algorithm in the ATLAS trigger are presented. Additionally, transformer-based graph neural networks are applied to the task of identifying jets produced in $H$ → $aa$ decays, and are found to improve background rejection up ∼ 10-15×, when compared to an existing approach. Further improvements made by incorporating conditioning mechanisms into the network architecture are also studied in depth and detailed.