Semi-Supervised machine learning technique for neutron-gamma discrimination and generalized approach for figure of merit

Abstract The discrimination between neutron and gamma radiation pulses is crucial in mixed environment for neutron spectroscopy, particularly in fields such as nuclear science, nuclear safety, environmental monitoring, and radiation imaging. A quantitative measurement is essential to evaluate the discriminatory performance and a generalized yardstick is desirable for all the available methods. This study introduces a semi-supervised machine learning approach utilizing Multi-Layer Perceptron, Convolutional Neural Network, Long Short-Term Memory Network and Transformer encoder-based classifier to perform neutron-gamma pulse discrimination. The proposed model is applied to pulse signals acquired from a liquid scintillator BC501A coupled with a photomultiplier tube R4144, recognized for their high sensitivity and effectiveness in neutron-gamma discrimination tasks. The model's performance is rigorously evaluated against traditional analogue and digital charge comparison discrimination techniques. A generalized method is introduced in terms of figure of merit for equipollent discrimination performance comparison with existing analog and digital-based methods as well as various other machine learning based classification techniques.