Machine learning techniques for β/γ discrimination in phoswich detectors

Abstract

Particle discrimination technology is widely used in multiple fields. Phoswich detectors are detectors based on pulse shape discrimination technology that combine two or more scintillators with different time characteristics to achieve particle discrimination. This study focuses on a phoswich detector composed of BGO/EJ-260 and uses machine learning algorithms to classify pulses to achieve β/γ classification. Experiments were conducted using the 137Cs radioactive source and three different models were trained: Gaussian mixture model, support vector machine, and convolutional neural network. The classification capabilities of the three models were tested and the results were discussed. The calculation results show that all three models achieved pulse data classification and accurately marked most pulses to the correct category. The classification ability of low-amplitude pulses by the Gaussian mixture model and support vector machine is limited by data processing, while the convolutional neural network model avoids this problem. For higher amplitude pulses, all three models showed that high classification accuracy, with the convolutional neural network model achieving a classification accuracy of 96.1% in the training set, achieves the expected goal.