Enhanced α/β discrimination in liquid scintillation via UMAP-HDBSCAN

In this study, an enhanced α/β discrimination method based on UMAP-HDBSCAN is proposed for liquid scintillation measurements. After waveform preprocessing of the acquired raw data, uniform manifold approximation and projection (UMAP) is employed to reduce the data dimensionality to three dimensions, enabling both feature extraction and visualization. Subsequently, hierarchical density-based clustering (HDBSCAN) is applied to classify α and β events. The resulting cluster labels are also used to train a random forest classifier, which not only facilitates analysis of waveform features contributing to discrimination performance but also effectively addresses class imbalance in α/β events when required. A mixed source of ²⁴¹Am and ⁹⁰Sr/⁹⁰Y is used for method validation, and the results demonstrate improved α/β separation performance compared to traditional pulse shape discrimination (PSD) approaches, as well as enhanced robustness with respect to input parameter variations.