Optimization of pulse shape discrimination based on one-class support vector machine for SiPM-based CLYC scintillator detector

The neutron-gamma pulse shape discrimination (PSD) algorithm is a critical algorithm for neutron detector applications. The effectiveness of PSD relies on analyzing the pulse characteristics of neutron and gamma within the detector and eliminating the interference of gamma by distinguishing the difference between neutron and gamma. However, practical challenges arise because neutron sources are scarce, whereas gamma sources are more readily available. Based on the idea of anomaly detection, this paper propose a one-class support vector machine (OCSVM) method and uses the pure gamma pulses of the detector reduced by principal component analysis (PCA) method as training data to verify the discrimination ability of Φ25.4 mm × 25.4 mm cylinder Cs 2 LiYCl 6 :Ce 3+ (CLYC) neutron detector coupled with SiPMs array. The results show that FoM-OCSVM = 2.56 compared with FoM-CCM = 1.49 which was obtained by the traditional charge comparison method (CCM). After PCA dimensionality reduction of 47-D pulse data to 2-D, FoM can be further improved to 32.23, and the Area Under Curve(AUC) of the Receiver Operating Characteristic(ROC) curve is 0.9999. Additionally, gamma pulses with different Signal-to-Noise Ratio(SNR) are generated by simulation, the results show that the false alarm rate of OCSVM decreases from 13.35% of CCM to 3.25% in the group of 5 dB SNR, and to 0% for SNRs between 10–20 dB. These findings confirm that the OCSVM-PCA framework significantly enhances neutron-gamma discrimination performance — even when trained exclusively on unlabeled gamma pulses, eliminating the need for neutron pulse labeling.