Taming nucleon density distributions with deep neural network

With the datasets of density distributions calculated by Skyrme functional theories, we elaborated deep neural networks to generate profile and provide a table related hyperparameters set for similar applications other structural models. In process machine learning with objective/target functions that normalized mean square error Kullback–Leibler divergence (cross entropy), there is turning point showing transition from Fermi-like distribution realistic distribution, while this property transcended when Pearson χ2 employed. A training program about 35 minutes only 5%−10% nuclei (200−300) sufficient describe nucleon all nuclear chart within 2% relative error. We obtain results employing different theories. further investigate extrapolation properties, which show an addition 15 nucleons acceptable. Based on results, propose mixed dataset approach retraining in order go beyond single physical structure model.