Interpretable Boosted Decision Tree Model for Multi-Site Background Rejection in LEGEND

LEGEND is a next-generation experiment searching for neutrinoless double-beta decay (0vBB) with high-purity germanium detectors. LEGEND aims to achieve the highest discovery sensitivity of any previous 0vBB search. To accomplish this, the many background reduction and veto methods integrated into LEGEND will be accompanied by software-based background rejection techniques. Machine learning methods allow for background discrimination models to be trained from real data. This work explores the development of a boosted decision tree model, a supervised machine learning method, for the rejection of multi-site background events using pulse shape characteristics. The model also incorporates an explainability study to better understand the origin of the machine’s classification methods. The resultant model outperforms the traditional, un-tuned discrimination cut and is a quick and e↵ective software solution for simple analysis of LEGEND data which has great potential for use by the LEGEND collaboration.