Enhanced neutron-gamma discrimination with fast signal output from SiPM arrays via deep neural network optimization

Abstract

High count rate neutron measurements require fast signal outputs and effective neutron-gamma discrimination techniques, yet the wide pulse widths generated by silicon photo-multipliers (SiPM) arrays often limit their efficiency in such settings. In prior work, we developed a compensation network (CN) to reduce pulse width and overshoot in SiPM fast outputs, improving pulse shape discrimination (PSD). However, further enhancement was needed to achieve optimal neutron-gamma discrimination. Conventional methods like charge comparison method (CCM) and frequency gradient analysis (FGA) typically analyze either time-domain or frequency-domain features independently. In this study, we employ a deep neural network (DNN) that integrates both frequency and time-domain features from ultra-fast pulse signals to improve discrimination accuracy. We optimized DNN model inputs through a systematic variable selection strategy that included separation ranking, correlation analysis, and recursive feature elimination (RFE), reducing the input set from 73 to 27 variables for a balance of simplicity and discriminative power. The neutron-gamma discrimination was then quantified with an equivalent figure of merit (FOM). Testing with the ²⁵²Cf source demonstrated the superior performance of the DNN-based approach, achieving an FOM of 0.96 (98.8% discrimination probability) compared to 0.73 (95.7%) for CCM and 0.63 (93.0%) for FGA. These findings underscore the potential of enhanced ultra-fast signal output systems for nuclear detection in high-count-rate applications.