Nonempirical shape dynamics of heavy nuclei with multitask deep learning

Abstract

A microscopic description of nuclear fission represents one of the most challenging problems in nuclear theory. While phenomenological coordinates, such as multipole moments, have often been employed to describe fission, it is not obvious whether these parameters fully reflect the shape dynamics of interest. We here propose a novel method to extract collective coordinates, which are free from phenomenology, based on multitask deep learning in conjunction with density functional theory (DFT). To this end, we first introduce randomly generated external fields to a Skyrme energy density functional (EDF) and construct a set of nuclear number densities and binding energies for deformed states of ²³⁶U around the ground state. By training a neural network on such a dataset with a combination of an autoencoder and supervised learning, we successfully identify a two-dimensional latent variables that accurately reproduce both the energies and the densities of the original Skyrme-EDF calculations, within a mean absolute error of 113 keV for the energies. In contrast, when multipole moments are used as latent variables for training in constructing the decoders, we find that the training data for the binding energies are reproduced only within 2 MeV. This implies that conventional multipole moments do not provide fully adequate variables for a shape dynamics of heavy nuclei.