Combined Machine Learning Analysis for Antihelium Search with the AMS-02 Experiment

Heavy anti-nuclei, such as antihelium, are unlikely to be formed during cosmic rays (CRs) propagation, as supported by measurements from the PHENIX and ALICE collaborations. The precise measurement of cosmic-ray antinuclei is a powerful probe for possible identification of Dark Matter (DM) signatures, and in combination with measurements of other cosmic-ray species, can help constrain cosmic-ray production and propagation in the Galaxy. The first-time discovery of cosmic-ray antihelium has been identified as a potentially promising detection signature related to DM interactions or to a primordial origin, due to its extremely low astrophysical backgrounds, especially towards low energies. This thesis presents an antihelium search, performed using 12.5 years of data collected by the AMS-02 experiment. Initially, an event selection for events with |Z|=2 is applied, after requiring a negative reconstructed rigidity, this analysis exploits four Deep Neural Networks (DNNs), two supervised classifiers and two unsupervised autoencoders, to search for possible antihelium candidates in data. The DNNs are independent and orthogonal, allowing different working points to be combined to improve background suppression. The analysis follows a fully data-driven workflow while applying the same strategy independently to data and Monte Carlo (MC) simulation. The dominant background consists of charge-confused helium nuclei, which is resonant in the same mass region as the possible signal. As a consequence, no control region can be defined directly from data, and the background estimate relies entirely on MC simulation. The search for antihelium is performed by looking for an excess in the number of events with a reconstructed mass in the range [2,5] GeV/c², compatible with He-3 and He-4 masses, within Eₖ/N [0.65,31.4] GeV/N for anti-He-3 and Eₖ/N [0.41,23.3] GeV/N for anti-He-4. Up to 12 antihelium candidates have been found, resulting in a mild tension with the background-only hypothesis at the level of 1-2 σ, depending on the chosen working points on the DNNs outputs. No statistically significant excess has been found; therefore, two exclusion limits using the CLs method on the anti-He-3/He and anti-He-4/He ratios are presented and discussed.