Pulse-Shape-Based Analysis with Recurrent Neural Networks in the MAJORANA DEMONSTRATOR

The Majorana Demonstrator experiment was a search for neutrinoless double-beta decay (0νββ) in ⁷⁶Ge using p-type point contact high-purity germanium detectors. The data-taking for 0νββ by the DEMONSTRATOR has successfully completed in March 2021 and set a 0νββ half-life limit of T_1/2 > 8.3´10²⁵ yrs based on its full exposure. The MAJORANA Collaboration developed traditional pulse-shape-based approaches to discriminate different types of events, such as multi-site (MS) events and single-site (SS) events. The collaboration is also exploring machine learning (ML) tools for event discrimination, such as recurrent neural networks (RNN). In this talk, we will discuss MAJORANA ML efforts. For example, the RNN is used to tag pileup-event waveforms, due to both random coincidences in calibration data and real physics correlations. The talk will also discuss the interpretable ML models for SS and MS event discrimination where the attention mechanism is implemented to focus on the most important component of the waveform, and to enhance the network performance.