Interpretable hybrid nuclear mass prediction based on term-by-term model discrepancies

Abstract

Various theoretical mass models have consistently achieved impressive accuracy in reproducing experimental masses. However, their predictions in unmeasured neutron-rich regions exhibit noticeable model dependence. In this study, we systematically investigate the differences in model predictions by comparing the liquid-drop mass terms of two representative models. Using two widely used models, Weizsacker-Skyrme-type (WS4) and Duflo-Zuker-type (DZ10), as representative examples, we find that the differences in an isotope chain gradually become more remarkable with increasing neutron number, not only for total binding energies but also for individual mass terms. The most noticeable difference appears in the volume-symmetry energy. By leveraging the term-by-term differences between the liquid-drop energy components of WS4 and DZ10, we introduce a machine learning gating network that adaptively combines the two models to improve predictive accuracy. This conditional hybrid model achieves a lower root-mean-square deviation (rmsd) than either model alone, reducing the overall rmsd from 0.284 MeV (WS4) and 0.560 MeV (DZ10) to 0.232 MeV. In the future, this term-by-term comparison strategy can be extended to models based on density-functional theories.