Model-repair capabilities of tree-based machine-learning algorithms applied to theoretical nuclear mass models

In this project, we systematically investigate and evaluate the performance and predictive power of refined nuclear-mass models with tree-based machine learning methods, including the model-repair capabilities applied to the ``bare'' mass-model calculations. Combining six theoretical models with five machine-learning algorithms, the refined model accuracies are illustrated based on available experimental masses and a physics-based feature space. The root-mean-square (rms) deviations between the physics-model calculations and the available experimental data range from the maximum, about 2.46 MeV, to the minimum, about 0.29 MeV. Significant improvements can be obtained for the refined theoretical models in the interpolation predictions of nuclear masses. The rms values of describing experimentally known nuclei can reach a high level (e.g., falling to about 0.16 MeV, crossing the accuracy of 0.20 MeV) in the combinations between theoretical mass models and machine learning algorithms. It is found that the model-prediction capabilities can be rearranged after they are refined by different machine learning algorithms, indicating the different abilities of picking up the missing ``physics.'' The residual between the ``bare'' mass model and experimental data usually presents a nonrandom pattern, revealing the model deficiency to some extent, while, after refining the theoretical model by machine learning, the residuals exhibt a uniform distribution. The model-repair capabilities of five data-driven machine-learning algorithms applied to six physically based nuclear-mass models are evaluated. In addition, based on optimal combinations of bare mass-models and machine-learning algorithms, the extrapolation predictions and model-repair abilities are validated and briefly discussed.