Research on Single-Nucleon Separation Energy Based on Artificial Neural Networks

本研究基于三个具有代表性的理论模型：相对论连续谱Hartree-Bogoliubov (RCHB)理论，相对论平均场(RMF)理论，Skyrme-Hartree-Fock-Bogoliubov (SHFB)模型，首先介绍了人工神经网络(ANN)方法，计算出了三个模型的单核子分离能的理论预测值。随后利用神经网络对单核子分离能的理论值进行了优化训练，降低了单核子分离能的理论预测值与实验值之间的均方根偏差(RMSD)，并在此基础上进行了两种分区优化，分别为质子和中子的幻数分区，分区优化训练后进一步降低了RMSD。单核子分离能分区训练后的RMSD比整体直接训练的效果更好，特别能显著降低轻核区的RMSD，单中子分离能进行中子幻数分区训练的效果更好，单质子分离能进行质子幻数分区训练的效果更好。This research is based on three representative theoretical models: the Relativistic Continuum Hartree-Bogoliubov (RCHB) theory, Relativistic Mean Field (RMF) theory, and Skyrme-Hartree-Fock-Bogoliubov (SHFB) model. First, the Artificial Neural Network (ANN) method was introduced to calculate theoretical predictions of single-nucleon separation energies for these three models. Subsequently, the neural network was employed to optimize and train the theoretical values of single-nucleon separation energies, reducing the root mean square deviation (RMSD) between theoretical predictions and experimental values. Two partitioning optimization schemes were then implemented: proton magic number partitioning and neutron magic number partitioning. The partitioned optimization training further reduced RMSD values. The partitioned training of single-nucleon separation energies demonstrated better performance than direct global training, particularly in significantly reducing RMSD in the light nuclei region. Specifically, neutron magic number partitioning showed superior effectiveness for optimizing single-neutron separation energies, while proton magic number partitioning yielded better results for single-proton separation energies.