Sensitivity Study for Di-Higgs Boson Production in the $1l3\tau$ Channel Using ATLAS Run-2 Data

This work presents a Machine Learning study of Higgs boson pair production in the $1l3\tau$ final state using ATLAS Monte Carlo samples produced in Release 22. Events are extracted from Run-2 CAF ntuples, processed with a dedicated preselection, and described using a set of kinematic features constructed from global event variables. A residual multilayer perceptron is trained and optimized with Optuna, and its output is incorporated into a full statistical analysis using TRExFitter. The classifier score serves as the final discriminant in the signal region, yielding an expected upper limit of 34.1 at 95% CL compared to the Standard Model expectation for the di-Higgs boson production cross-section. This work establishes an end-to-end pipeline for integrating Machine Learning techniques into the di-Higgs boson analysis in Release 22 and provides a foundation for future improvements using extended feature sets and refined modeling.