Understanding on prediction differences among theoretical mass models with machine learning techniques

Nuclear mass is fundamental in both nuclear physics and astrophysics. Different theoretical mass models have been proposed based on distinct assumptions, leading to varied mass predictions. In this work, we employ a Gaussian process regression (GPR) machine learning (ML) model to systematically investigate the prediction differences among four theoretical mass models, namely, $\mathrm{Weizs}\stackrel{\ifmmode \ddot{}\else \"{}\fi{}}{a}\text{cker-Skyrme--type}$, finite-range droplet model--type, covariant density functional theory--type, and Bethe-$\mathrm{Weizs}\stackrel{\ifmmode \ddot{}\else \"{}\fi{}}{a}\text{cker--type}$ models. We use three feature sets with different numbers of features to study their impacts on the accuracy of the GPR model. The fact is confirmed that the accuracy can be improved by increasing the number of features. With the optimal feature set, the GPR model achieves an overall accuracy of approximately 0.2 MeV for these four mass evaluations. Besides, the prediction differences among these mass models are also systematically investigated by using interpretable ML techniques, which can provide insights into the underlying physical factors responsible for their prediction differences as well as more comprehensions on the nuclear mass evaluations.