A Kohn-Sham scheme based neural network for nuclear systems

Abstract

A Kohn-Sham scheme based multi-task neural network is elaborated for the supervised learning of nuclear shell evolution. The training set is composed of the single-particle wave functions and occupation probabilities of 320 nuclei, calculated by the Skyrme density functional theory. It is found that the deduced density distributions, momentum distributions, and charge radii are in good agreements with the benchmarking results for the untrained nuclei. In particular, accomplishing shell evolution leads to a remarkable improvement in the extrapolation of nuclear density. After a further charge-radius-based calibration, the network evolves a stronger predictive capability. This opens the possibility to infer correlations among observables by combining experimental data for nuclear complex systems.