The Application of Neural Networks in Nuclear Masses

本文采用贝叶斯神经网络方法对多个核质量模型进行了优化，包括宏观模型LDM、宏观–微观模型FRDM12，微观模型RMF等。基于AME2020核质量数据表，BNN方法有效降低了实验值与理论预测值之间的均方根误差，尤其在LDM模型和RMF模型中，均方根误差都降低了80%。通过对轻、中和重核的单中子分离能进行测试，结果显示BNN优化后的核质量模型的单中子分离能与实验数据能够较好地趋近，并且再现了奇偶交错现象。This paper employs the Bayesian neural network method to optimize multiple nuclear mass models, including the macroscopic model LDM, the macroscopic-microscopic model FRDM12, and the microscopic model RMF, etc. Based on the AME2020 nuclear mass data table, the BNN method effectively reduces the root mean square error between experimental values and theoretical predictions, especially in the LDM and RMF models, where the root mean square error is reduced by 80%. By testing the single-neutron separation energies of light, medium, and heavy nuclei, the results show that the single-neutron separation energies of the BNN-optimized nuclear mass models can better approach the experimental data and reproduce the odd-even staggering phenomenon.