Neutron-gamma pulse shape discrimination for EJ301 liquid scintillator based on machine learning

In fast neutron multiplicity counting measurement, misclassification of γ signals and loss of neutrons introduce significant measurement errors. To address these problems, machine learning (ML) algorithms were employed to improve the n/γ discrimination of liquid scintillators. A dual-scintillator time-of-flight device combined with charge comparison (CC) method was used to select reliable datasets from the D-T neutron generator. Decision Tree, Random Forest, and Back-Propagation Neural Network (BPNN) were developed and compared with the CC method. The CC method and ML algorithms were validated using ¹³⁷Cs sources. The results showed that the ML algorithms had effective n/γ discrimination capabilities. The BPNN exhibited the highest DER_γ (1.26%) and DER_n (1.64%) discrimination performance, which reduced neutron loss and γ misclassification. In addition, the trained BPNN was used in practical measurement.