Automated Optimization of an ATLAS Search for Higgs Boson Pair Production at the LHC

Particle physics analyses are inherently complex and need to process large datasets through a chain of many interdependent steps. These analyses typically aim to optimize their sensitivity to potential signals. However, due to the stepwise nature, optimizations at several stages often need to rely on approximations to the final analysis sensitivity. This approach leads to extensive re-optimizations that require a careful balance between optimizing selection criteria, accounting for uncertainties, and maintaining a meaningful statistical analysis. This work explores neural end-to-end-optimized summary statistics (NEOS) as a novel, automated alternative to this procedure. For the first time, it is applied to a full-featured analysis in a search for boosted Higgs boson pair production via vector boson fusion, decaying into a four b-quark final state. A dedicated optimization framework, auTOMATed Optimization of Sensitivity (TOMATOS), is introduced, enabling a unified analysis optimization that targets the analysis sensitivity directly. Its performance is benchmarked against traditional methods, demonstrating comparable results. This thesis presents the latest expected ATLAS constraints on the $\kappa_\mathrm{2V}$ coupling using Run 2 data: $0.47 < \kappa_\mathrm{2V} < 1.55$.