Graph-Learning Based Particle Discrimination for Scintillator Neutron Detectors

Practical neutron scintillation detectors need to have high neutron detection efficiency and high neutron-gamma discrimination ratio (NGD ratio, or the ratio of neutron and gamma efficiencies). It is a tough requirement for scintillators such as Ce:GAGG and Li-6 glass (GS20) due to small difference in pulse-shape between neutron and gamma events. Graph-embedded Nonnegative Matrix Factorization (GNMF) and digital signal filters have been shown effective for NGD for GAGG and GS20 after they are learned using data collected under separated neutron and gamma sources. In these attempts, the event graph was established from noisy signals of selected important events, which may result in less accurate graph structure and the failure of graph-embedded methods in varied environments such as mixed neutron-gamma fields. To solve this issue, we proposed two strategies to improve the quality of graphs: first, a multi-layer graph neural network (GNN) is used to denoise the signals and the graph is constructed from smoothed signals; second, an event graph is learned under structure constraints including graph connectivity and sparsity. It has been shown that a 3-layer GNN could enhance the NGD ratio and neutron detection efficiency of GAGG in a mixed neutron-gamma field. The performance of sparsity-based graph learning is being assessed. The graph learning strategies potentially enhance the robustness of neutron detectors. The proposed algorithms may be applied not only for neutron detectors, but also for other radiation detectors.