Comparative study of radial basis function and Bayesian neural network approaches in nuclear mass predictions

Abstract

The radial basis function (RBF) and Bayesian neural network (BNN) approaches have been widely used in nuclear mass predictions. In this work, the predictive powers of RBF and BNN approaches are systematically explored by investigating nuclear masses and single-nucleon separation energies. To improve the description of single-nucleon separation energies, the odd-even effects have to be considered for the RBF approach, named as RBFoe approach. It is found that both RBFoe and BNN approaches significantly improve the description of nuclear masses and single-neutron separation energies in or near the known region, while they show different predictions when extrapolated far away from the known region. Comparing with the BNN approach, the RBFoe approach achieves better interpolation ability, while the extrapolation abilities of RBFoe and BNN approaches depend on the employed nuclear mass models. This indicates we should employ different reconstructed functions of RBF and BNN approaches to achieve better extrapolation ability for different nuclear models. Since the BNN approach is very flexible in designing its input, hidden, and output layers, it is desirable to find the optimal BNN structure to get better extrapolation ability in the future.