Artificial gamma ray spectra simulation using Generative Adversarial Networks (GANs) and Supervised Generative Networks (SGNs)

Abstract

X-ray and gamma-ray spectral analysis is classically based on models of expected spectra that are compared with acquired data. The usual approach is to define a mathematical model of an instrument in a stable and constant environment. However, when the operating conditions are varied, unstable, when the instrument has multiple configurations or when the environment itself has an impact on the measurement, it becomes extremely difficult to make accurate models to represent those different conditions. In this context, new approaches based on Deep Neural Networks are developed for nuclear, medical and astrophysical applications to perform spectra analysis, with the advantage that they can handle a wide variety of acquisition conditions, provided that they are represented in the training database. Therefore, in order to train these models, it is utterly important to have high quality and representative data, usually based on simulation models. The main problem with these models is the trade-off between the quality of the simulated data used in training the network and the simulation time to generate that data. The problem lies in the time spent during Monte Carlo (MC) simulations, for instance using GEANT4 software, which can take weeks to simulate the complete instrumental response in a specific situation. In this paper, we address this issue by presenting an approach based on Deep Learning algorithms to generate MC data. We develop two deep neural network models, one based on Generative Adversarial Networks (GANs) and the other based on Supervised Generative Networks (SGNs). Once the networks are trained, we are able to generate in milliseconds, realistic spectral signatures of different radioelements and mixtures with the specifics response of our Caliste-HD detection system. The objective of this work is to study the feasibility of using this approach to provide a faster but still reliable alternative form of gamma spectra generation in order to train other models dedicated to Deep Neural spectra identification with high quality data, improving their performance and extending their domain of validity.