Machine learning-driven high-precision model for α-decay energy and half-life prediction of superheavy nuclei

We develop a physics-informed machine-learning framework for predicting α-decay energies and half-lives across a broad range of nuclei. The approach is based on an eXtreme Gradient Boosting (XGBoost) regression model and incorporates physically motivated nuclear descriptors constructed from evaluated nuclear data and deformation tables. For half-life prediction, key structure-related features—including magic-number proximity, minimum orbital angular-momentum transfer, isospin asymmetry, and quadrupole deformation—are introduced to represent the dominant mechanisms governing α decay within a data-driven framework. Model performance for both the Qα and half-life prediction tasks is evaluated using five-fold cross-validation, which demonstrates strong predictive accuracy without evident overfitting. Benchmark comparisons with widely used empirical relations, including the Royer formula and the universal Decay Law (UDL), show that the XGBoost model achieves systematically lower prediction errors while preserving the established systematics of α-decay observables. SHapley Additive exPlanations (SHAP) analysis further reveals that the leading contributions from decay-energy information, centrifugal hindrance, and shell-related effects follow physically consistent trends, supporting the interpretability of the learned relationships. These results indicate that gradient-boosting models equipped with physics-guided features provide an accurate and robust framework for describing α-decay systematics and for predicting half-lives in regions where experimental data remain scarce.