Learning nuclear mass systematics with hybrid quantum-classical mixtures of experts

Abstract

We introduce a mixture-of-experts framework combining sixteen hybrid quantum-classical neural networks with nucleus-dependent classical gating to learn residual corrections to the WS4 nuclear mass model. On a fixed AME2020 test set, the framework reduces the root-mean-square error from 456.22 to 179.83 keV and outperforms a parameter-matched classical counterpart. The learned corrections improve shell-gap systematics near the Sn and Pb shell closures, while analysis of the gating networks identifies shell-related descriptors as the leading inputs. For 66Ca, the model predicts Sn = 0.577 MeV and S2n = -0.252 MeV, allowing two-neutron emission while the one-neutron channel remains closed.