Analysis and trigger approaches for New Physics searches exploiting Machine Learning at the ATLAS detector

This thesis explores the application of Machine Learning (ML) techniques in two key areas of the ATLAS experiment: model-independent searches for New Physics and real-time triggering for the Muon Spectrometer upgrade. A Transformer Anomaly Detection (AD) model is developed to identify unexpected resonances in fully hadronic final states. The model is trained exclusively on QCD dijet background events and employs a feature transformation technique to ensure jet mass independence. The approach is validated on various Beyond the Standard Model benchmark signals and is integrated into an analytical fit over the background invariant mass spectrum for measuring upper limits on the signal production cross-sections. In parallel, ML methods are explored for the upgraded ATLAS Level 0 muon trigger system. Both Convolutional Neural Networks and GNN are studied for real-time muon reconstruction in the Resistive Plate Chambers trigger, with Quantization Aware Training and Knowledge Distillation applied to meet FPGA constraints. These results demonstrate the potential of ML in enhancing both NP searches and real-time triggering, contributing to the optimization of the ATLAS experiment for the High-Luminosity LHC era.