Automatic and Real-Time Identification of Radionuclides in Gamma-Ray Spectra: A New Method Based on Convolutional Neural Network Trained With Synthetic Data Set

Abstract

Automatic and fast identification of gamma-ray-emitting radionuclides is a challenge in the field of nuclear safety, especially in case of emergency, since it requires complex calculations and often the knowledge of experts to interpret the data. We present a development of an automatic identification method based on convolutional neural networks (CNNs) as a new tool to analyze gamma-ray spectra in real time, which uses not only photoelectric peaks but also extracts all discriminant features in the spectrum, such as Compton structures, for instance. The original approach relies on the training of the CNN with a fully synthetic database, built by means of a Monte Carlo simulation with Geant4 combined with a detailed analytical detector response model. The algorithm and training method are evaluated to identify radionuclides in measurements of the mixtures of sources acquired with Caliste, a fine-pitch CdTe imaging spectrometer. The neural network is able to discriminate each element in an arbitrary mixture very quickly with high accuracy.