Systematic analysis of nuclear binding energy using XGBoost residual modeling

Understanding nuclear structure and astrophysical nucleosynthesis processes depends on the analysis of nuclear binding energy, which remains one of the challenging problems in nuclear physics. The macroscopic semi-empirical mass formula (SEMF) provides a fundamental description based on the liquid-drop model; however, it lacks the ability to characterize microscopic quantum mechanical phenomena, resulting in systematic discrepancies from the experimental data. In this paper, we propose a hybrid machine learning (ML) framework to bridge this gap by modeling the residual errors of the SEMF using an eXtreme Gradient Boosting (XGBoost) regression model. Trained on a large dataset of 2,147 nuclei from the AME2020, the system utilizes domain-specific variables like shell-closure proximity, pairing correlations, and isospin asymmetry to extract and learn the nonlinear correlations that drive nuclear stability. The proposed model achieves a testing root mean square error (RMSE) of 0.470 MeV and mean absolute error (MAE) of 0.335 MeV, which is a substantial improvement over the theoretical baseline. Strict 5-fold cross-validation is performed for the stability of the model (RMSE ≈ 0.510 M e V ≈ 0.510 MeV ) as well as its generalization capability, confirming this hybrid approach as a robust and unbiased tool for high-precision data evaluation.