Validation and extrapolation of atomic masses with a physics-informed fully connected neural network

Machine learning offers a powerful framework for validating and predicting atomic mass. We compare three improved neural network methods for representation and extrapolation for atomic mass prediction. The powerful method, adopting a macroscopic-microscopic approach and treating complex nuclear effects as output labels, achieves superior accuracy in AME2020, yielding a much lower root-mean-square deviation of 0.122 MeV in the test set, significantly lower than alternative methods. It also exhibits a better extrapolation performance when predicting AME2020 from AME2016 with a root-mean-square deviation of 0.191 MeV. We further conduct sensitivity analyses against the model inputs to verify interpretable alignment beyond statistical metrics. Incorporating theoretical predictions of magic numbers and masses, our fully connected neural networks reproduce key nuclear phenomena including nucleon pairing correlation and magic number effects. The extrapolation capability of the framework is discussed and the accuracy of predicting new mass measurements for isotope chains has also been tested.