Predictions of unknown masses in the feedforward neural network

Abstract In this paper, a feedforward neural network (FNN) approach is employed to optimize three local mass models (GK, GKs, and GK+J). It is found that adding physical quantities related to pairing effect in the input layer can effectively improve the prediction accuracy of local models. For the known masses in AME2012, the FNN reduces the root-mean-square deviation between theory and experiment for the three mass models by 11 keV, 32 keV and 623 keV. Among them, the improvement effect of light mass region with mass number between 16 and 60 is better than that of medium and heavy mass regions. It also has good optimization results when extrapolating AME2012 to AME2020 and the latest measured masses after AME2020. Based on the improved mass data, the separation energies for single- and two-proton (neutron) emissions, and $\alpha$-decay energies are obtained, which agree well with the experiment.