Discrimination of neutron-$γ$ in the low energy regime using machine learning for an EJ-276D plastic scintillator

TL;DR

This work tackles neutron–γ discrimination in a plastic scintillator (EJ-276D) at low deposited energies where traditional PSD degrades. It applies supervised ML using a Multi-Layer Perceptron Bayesian Neural Network (MLPBNN) and a Support Vector Machine (SVM), leveraging head-tail waveform features including $R_{head}$, $ig\langle t \big angle_Q$, and $\kappa\sigma^2$, with data labels produced by a data-driven $(\mathrm{kSigma})$-based selection. The MLPBNN outperforms the SVM, and both show reasonable agreement with Time-of-Flight benchmarking, achieving about 0.80 accuracy in the 200–800 keV$_{ee}$ range and robust discrimination up to higher energies. The results demonstrate that ML can extend PSD performance in low-energy regimes, with practical implications for anti-neutrino experiments and neutron-background rejection, and point to future adoption of CNN/DNN approaches for further improvements.

Abstract

In this work, we present results for discrimination of neutron and $γ$ events using a plastic scintillator detector with pulse shape discrimination capabilities. Machine learning (ML) algorithms are used to improve the discriminatory power between neutron and $γ$ events at lower energy ranges which otherwise are not addressed by the conventional pulse shape discrimination techniques. The use of a multilayer perceptron with Bayesian inference (MLPBNN) and support vector machine (SVM) algorithms are studied using the recorded waveforms from the detector. Input variables are constructed for the ML algorithms, which captures the essence of the differences in the head and tail part of the neutron and $γ$ waveforms. A new variable, which utilizes the product of kurtosis and variance calculated from the waveform gives better ranking in terms of separation of neutron and $γ$ events. The training and the testing of the ML algorithms are done using an AmBe neutron source. In the lower energy region, the results obtained from the ML predictions are compared with the results obtained from a time of flight (ToF) technique to benchmark the overall performance of the ML algorithms. A reasonable agreement is observed between the results obtained from ML algorithm and the ToF experiment in the studied energy range. The MLPBNN gives better discriminatory power for the neutron and $γ$ events than the SVM algorithm.

Figures (16)

• Figure 1: (a) Normalized waveform from a EJ276D PS detector, (b) neutron (red dashed line), and $\gamma$ (blue solid line) waveforms.
• Figure 2: Integrated charge distribution of (a) ${}^{137}\textbf{Cs}$, (b) ${}^{22}\textbf{Na}$ radioactive source and (c) calibration obtained from the measured Compton edge energy deposition in the detector.
• Figure 3: (a) PSD variable as a function of integrated charge, (b) projection of PSD variable for charge range between 200 to 800, and (c) projection of PSD variable for charge range between 1500 to 4000.
• Figure 4: Input variables used for the ML algorithms are (a) $R_{head}$, (b) $\langle t \rangle_{Q}$ (in ns), and (c) $\kappa\sigma^2$. The distributions are shown for neutron (red dashed) and $\gamma$ (blue) events.
• Figure 5: PSD distribution from AmBe source in different integrated charge ranges.