Characterization of electroluminescence signals from nuclear recoil events in the dual-phase argon Time Projection Chamber of the Red experiment with Convolutional Autoencoders

The nature of dark matter, which constitutes 26% of the mass-energy content of the Universe, remains a key mystery in astroparticle physics. Among the most compelling dark matter candidates are Weakly Interacting Massive Particles (WIMPs), whose thermal relic abundance from the Big Bang aligns with current observations. The Global Argon Dark Matter Collaboration (GADMC) seeks to directly detect WIMPs using DarkSide-20k, a multi-ton dual-phase liquid argon Time Projection Chamber (LAr TPC) under construction at INFN Laboratori Nazionali del Gran Sasso. This detector will identify WIMP-nucleon scattering events by detecting both prompt scintillation light and ionization charge, with the latter measured through delayed electroluminescence signals. While conventional WIMP searches focus on masses around hundreds of GeV/c², interest has recently expanded to lighter WIMPs, down to 1 GeV/c². The properties of liquid argon and LAr TPC technology make them highly suited for probing such candidates. However, detecting WIMP-induced nuclear recoils at these low masses, where deposited energies are only a few keV, presents unique challenges. These include difficulties in detecting scintillation signals and uncertainties in the ionization response at sub-5 keV recoil energies. Data from DarkSide-50 enabled the development of a preliminary ionization response model, but calibration in this low-energy regime remains incomplete due to significant ionization quenching fluctuations. The Recoil Directionality (ReD) experiment addresses this gap with a focused campaign to study the ionization response of argon at low recoil energies. ReD employs a miniaturized TPC irradiated with neutrons under two-body kinematics, enabling precise measurements of ionization quenching in this unexplored energy range. This thesis details my contributions to the ReD project, from commissioning and data acquisition to the development of an artificial intelligence-based analysis method. Specifically, I implemented a convolutional autoencoder (CAE) to classify electroluminescence signals recorded by silicon photomultipliers at cryogenic temperatures. By leveraging machine learning, the CAE efficiently identified patterns in experimental data, offering a novel, data-driven approach to nuclear recoil event tagging. This thesis also evaluates the performance of the CAE-based tagging method against the conventional signal-selection method. The document is structured as follows: Chapter 1 reviews dark matter and WIMP detection technologies, Chapter 2 details argon-based TPCs in the GADMC program, Chapter 3 focuses on the ReD experiment and its preliminary results, Chapter 4 presents the CAE-based tagging methodology, and Chapter 5 discusses the results and their comparison to traditional approaches.