Piled-up neutron-gamma discrimination system for CLLB using convolutional neural network

Abstract

Abstract A piled-up neutron-gamma discrimination system is designed to discriminate single and piled-up events under high counting rate. The data acquired by a Cs2LiLaBr6:Ce (CLLB) detector and an Am-Be neutron source are used to train and test the model in the n-γ discrimination system. The charge comparison method is applied to discriminate the non-piled-up events in the experimental data and label the dataset of single events. As a result of the method, the figure-of-merit (FOM) value is 1.10, which indicates that the wrong labeling ratio is about 0.248%. A dataset of piled-up events is created by adding up waveforms and labels of the events in the single-pulse dataset. The discrimination system consists of three convolutional models, called Model_PulseNum, Model_OnePulse and Model_TwoPulses. All the models are trained and tested by the created dataset. Model_PulseNum is created and trained to define the number of pulses in the waveform of the event, with an accuracy of 99.94%. The other two models (Model_OnePulse and Model_TwoPulses) are created and trained to discriminate the particle types for non-piled-up and two-fold piled-up events with the accuracy of 99.5% and 98.6%, respectively. For the whole discrimination system, the accurcy for the particle identification is over 97% for each class (γ, n, γ + γ, γ + n, n + γ and n + n). These results indicate that CNN model can improve the performance of particle detection systems by effectively discriminate neutron and gamma for both piled-up and non-piled-up events under high counting rates.