Advanced Neutron Discrimination With Diamond Detectors Using Machine Learning

Abstract

This work explores the application of machine learning (ML) techniques for neutron/ $\gamma $ -ray discrimination using transient current signals from a single-crystal (sc) chemical vapor deposition (CVD) diamond-based detector, with a focus on model robustness and hardware deployment in real-time systems. Two architectures, a multi-layer perceptron (MLP) and a 1-D convolutional neural network (1D-CNN), are evaluated across various signal domains and experimental conditions, including low-temperature operations and frequency-domain preprocessing. The performance of each architecture is evaluated using the figure of merit (FoM) method and compared to performance values obtained by classical pulse shape discrimination (PSD) techniques. In addition to assessing classification performance, the impact of processing signals with a reduced signal-to-noise ratio (SNR) caused by the operation of the detector at a lower temperature on model reliability is presented, along with an investigation of compression strategies to enable deployment on resource-constrained field-programmable gate array (FPGA) platforms. The study demonstrates that both architectures can be effectively adapted for embedded systems, balancing accuracy and efficiency for real-time signal classification tasks in radiation detection applications while achieving better FoM values than classical PSD methods.