Exploration of genetic algorithms to build a balanced neutron spectra dataset useful to train unfolding techniques based on artificial neural networks

Diverse domains need neutrons unfolding technics to assess the incident neutron energy spectrum. Examples are radiation protection, nuclear reactor physics or criticality safety. Traditionally, methods based on the Bayesian approach requires an initial guess of the solution which may significantly impact the unfolding result. This work proposes a novel method for neutron spectrum reconstruction using machine learning (ML) techniques trained on a large dataset. To ensure the ML algorithm to perform on a large domain of application particular attention has been paid to the dataset creation. We propose a comparison of two methods of building large dataset where the most adequate solution is obtained using a dynamic genetic algorithm (GA). This GA targets optimal combinations of 48 parameters to generate a variety of neutron spectra. The resulting dataset is then used to train a new convolutional neural network architecture for unfolding neutron spectra. Obtained performance metrics of the tested architecture show high efficiency and emphasize the added value of the built dataset.