The mass of the atomic nucleus, as one of the fundamental physical quantities of the atomic nucleus, plays an important role in understanding and researching the structure of the atomic nucleus and nuclear reactions, the basic interactions between nucleons. However, accurately predicting the mass of nuclei far from the β stability line remains a huge challenge. Based on the machine-learning-refined mass model, the newly measured atomic nucleus masses since 2022, the residual proton-neutron interaction ($\delta V_{pn}$), and the α-decay energy of heavy nucleus are studied. It is found that: (1) For the 23 newly measured atomic nuclei, the root mean square deviations obtained by the machine-learning-refined mass models are between 0.51 and 0.58 MeV, which are significantly lower than the 3.275, 1.058, 0.752, and 0.785 MeV given by the liquid droplet model (LDM), Weizsäcker-Skyrme-4 (WS4), finite-range droplet model (FRDM), and Duflo-Zucker (DZ), respectively. (2) The $\delta V_{pn}$ of the atomic nucleus with N = Z obtained from machine-learning-refined mass models is consistent with the latest experimental data. (3) The root mean square deviations of the α-decay energy of heavy nuclei obtained from the machine-learning-refined mass models have also been significantly reduced. Furthermore, by using the Bayesian model average approach to consider the results of different machine-learning-refined mass models, a more accurate prediction can be obtained. These results demonstrate that the machine-learning-refined mass models possess good extrapolation capabilities and can provide useful insight for further researches. The datasets presented in this paper, are openly available at https://doi.org/10.57760/sciencedb.j00213.00246.