In this work, two popular modern machine learning (ML) algorithms, the light gradient boosting machine (LightGBM, a decision tree-based algorithm) and the sparse variational Gaussian process (SVGP, a Bayesian method), are used to assess their performance in predicting the nuclear charge radius $R_{\mathrm{C}}$ by learning available experimental data. The training set consists of 820 $R_{\mathrm{C}}$ data points measured before 2013, and the validation set includes 110 data points measured up to 2021, and the most recently measured 24 data points are used to further evaluate the trained ML algorithms. With the consideration of 11 input features, both ML algorithms achieve competitive accuracy on the validation set, with the standard root-mean-square deviation (RMSD) of 0.0118 fm for SVGP and 0.0125 fm for LightGBM. The present work confirms that the feature of handling outliers for Hg isotopes proposed in previous work can significantly improve the performance of the Bayesian method, but it has almost no effect on the performance of LightGBM. Furthermore, the RMSD values obtained on the 24 newly measured nuclei are 0.0142 fm for SVGP and 0.0121 fm for LightGBM, once again demonstrating that the present ML algorithms with proper consideration of physical features can achieve competitive accuracy in predicting nuclear charge radii.

Close

PrevNext

View figure in article

• Nuclear charge radii
• Nuclear radii
• Nuclear structure & decays
• Bayesian methods
• Machine learning