Abstract Nuclear masses are fundamental observables that reflect nuclear structure and stability, playing a key role in nuclear physics and astrophysical processes. Most existing neural network studies focus on predicting either binding energies or neutron/proton separation energies individually, with limited attention to the physical correlations between these observables. Based on the relativistic point-coupling model PCF-PK1, a physics-informed artificial neural network (ANN) was developed to systematically predict nuclear binding energies along with single- and double-neutron/proton separation energies, while preserving the physical self-consistency of the predictions. To assess the impact of incorporating separation-energy constraints, networks were trained with varying loss function weight combinations, enabling a comparison between networks without separation-energy constraints (e.g., ANN1) and those including such constraints (e.g., ANN3).
The neural network significantly improves the overall prediction accuracy of binding energies compared with the PCF-PK1 model. Without separation-energy constraints, ANN1 already achieves high precision for binding energies (RMSE ≈ 0.147 MeV) and separation energies (RMSE ≈ 0.158– 0.185 MeV). Incorporating separation-energy constraints in ANN3 results in a slight improvement in overall prediction accuracy. The binding energy predictions improve by approximately 4.6%, while the separation energy predictions increase by 8.9–12.0%. The improvement is particularly noticeable for nuclei where the deviations of ANN1 predictions from experimental values exceed 0.2 MeV. Supporting datasets are publicly accessible at the Science Data Bank (https://doi.org/10.57760/sciencedb.j00213.00239). To facilitate the review process, a private access link is provided for reviewers during the review period (https://www.scidb.cn/s/bqyemq).