Nuclear mass predictions with multi-hidden-layer feedforward neural network

Abstract

Based on Keras deep learning framework, the feedforward neural network (FNN) model is employed to improve the predictions of the liquid drop model (LDM). It is shown that the prediction ability of FNN can be significantly improved if multiple hidden layers are used with only four input parameters. The root-mean-square deviation (RMSD) of nuclear mass predicted by LDM can be reduced from 2.38 MeV to 196 keV with the multi-hidden-layer FNN model, which is only one third of the single-hidden-layer FNN model. The predictions of two-neutron separation energies (S2n) and single-neutron separation energies (Sn) indicate that the multi-hidden-layer FNN model gives a better description of the shell structure. In addition, the extrapolation capability of the model in the super-heavy nuclear region is studied, the results show that better extrapolation capability can be achieved if multiple hidden layers are employed.