Dottorato in FISICA

Titolo della tesi: Search of Higgs Boson Pair Production in the $b\bar{b}\tau^+_{had}\tau^-_{had}$ Channel with Run $2$ and partial Run $3$ data at the ATLAS experiment and MEDI-Net: Development of Ultra-Fast Machine Learning Algorithms for Trigger Upgrade in the HL-LHC regime

The measurement of Higgs boson pair production constitutes a central objective of the Large Hadron Collider (LHC) physics programme, as it provides direct access to the Higgs boson self-coupling and to the structure of the Higgs scalar potential. Among the accessible final states, the $HH \rightarrow b\bar{b}\tau^+\tau^-$ channel, and in particular the $b\bar{b}\tau^+_{\mathrm{had}}\tau^-_{\mathrm{had}}$ final state, offers a favourable compromise between signal branching fraction and background controllability, but remains experimentally challenging due to its complex event topology and large reducible backgrounds. This thesis presents several developments aimed at improving the sensitivity of the $HH \rightarrow b\bar{b}\tau^+_{\mathrm{had}}\tau^-_{\mathrm{had}}$ analysis within the ATLAS experiment. On the physics analysis side, new trigger strategies are developed for the early Run~$3$ data-taking period and the corresponding analysis is updated to maximise their impact. Event selection and categorizatio is introduced, leading to a significant improvement in the expected sensitivity with respect to the previous Run~$2$-only analysis. The improvement exceeds that expected from the increased integrated luminosity alone by approximately $28\%$, and is attributed to the methodological enhancements implemented in the present work. In parallel, dedicated studies are performed in view of future iterations of the analysis using the full Run~$3$ dataset. These include the optimisation and evaluation of the GNtau algorithm as a candidate next-generation method for hadronic $\tau$-lepton identification in ATLAS, as well as the development of a new combined $b+\tau$ trigger configuration. The latter demonstrates a standalone efficiency comparable to that of the full trigger strategy adopted during early Run~3 operation, motivating its central role in the forthcoming $b\bar{b}\tau^+_{\mathrm{had}}\tau^-_{\mathrm{had}}$ analysis. The thesis also addresses challenges related to real-time event selection at the High-Luminosity LHC. A prototype hardware-oriented neural network (MEDI-Net) is developed as a candidate architecture for machine-learning-based triggering in the ATLAS Muon Spectrometer. While the achieved fake-muon rate does not yet satisfy the target requirement of $0.2\%$, the model exhibits competitive performance, with a latency of $347~\mathrm{ns}$, an FPGA resource occupancy of approximately $5\%$, and a plateau efficiency of about $99\%$. In addition, an exploratory research project investigates the use of memristive devices for analog in-memory computing as a potential solution for ultra-low-power inference at trigger level. As a proof of principle, a $3$-bit logic encoding is successfully implemented and characterised on memristor devices. Overall, this work contributes to both the methodological and technological aspects of Higgs boson pair searches at ATLAS, providing incremental improvements to current analyses and exploring solutions relevant for future high-luminosity operation.