Data Cleaning with Traditional and Artificial Intelligence Methods for the LEGEND Experiment

Neutrinoless double beta decay (0νββ) is a theorized radioactive disintegration process that, if observed, would violate the conservation of lepton number and indicate that neutrinos are their own antiparticles. Detecting (0νββ) would provide key insights into the nature of neutrinos and physics beyond the Standard Model. LEGEND is an international collaboration aiming to detect 0νββ using high-purity Germanium (HPGe) detectors enriched in 76Ge. LEGEND will operate in two phases, aiming to set lower limits of T^{0ν}_{1/2} > 10^{28} years after its completion. Signals captured by HPGe detectors are contain physical events caused by particle interactions and non-physical events caused by transient anomalies in the electronics chain. Data cleaning aims to distinguish between physical versus anomalous events, with the ultimate goal of removing non-physical signals from datasets. Traditional data cleaning consists of creating and fine-tuning parameters based on digital signal processing (DSP). A hierarchical data cleaning method based on aggregations of DSP parameters was employed in LEGEND-200. Data cleaning cuts were studied and validated prior to the unblinding of LEGEND-200 data for its first 0νββ result. Traditional data cleaning methods yielded efficiencies ranging from 95.3 - 98.6 %. This study presents a novel data cleaning mechanism based on artificial intelligence (AI). The model is based on a clustering algorithm, Affinity Propagation (AP), with human supervision plus a Support Vector Machine (SVM) for classification. AP-SVM provided a maximum physics event sacrifice of 0.024 + 0.004 − 0.003 % on data obtained from the Full Chain Test stand at UNC Chapel Hill. AP-SVM was used as the primary data cleaning tool for the LEGEND-200 secondary Julia-based analysis stack, supported by traditional parameters when necessary. The combination of both data cleaning methods yielded efficiencies ranging from 97.9 - 100 %. AP-SVM has been successfully employed in several experimental configurations, including the Full Chain Test, LEGEND-200 in all its configurations from integration to present data taking, and various HPGe characterization stands. The model is currently being tested on waveform data from silicon photomultiplier (SiPM) detectors, with promising preliminary results. The AP-SVM model provides an autonomous, scalable, versatile, and data-driven approach to perform data cleaning for systems yielding waveform data.