Transfer-learning-driven machine-learning model for α -decay half-life predictions

Alpha decay, a fundamental quantum mechanical phenomenon, serves as a vital tool for exploring nuclear decay properties and stability. In this study, we developed machine learning models based on Random Forest (RF), Support Vector Regression (SVR), and Deep Neural Network (DNN) techniques to predict alpha decay half-life (${T}_{1/2}$). A transfer learning (TL) framework is introduced, wherein a pretrained alpha decay energy (${Q}_{\ensuremath{\alpha}}$) model serves as an input for half-life prediction, improving both accuracy and computational efficiency. The performance of RF, SVR, and DNN models are rigorously evaluated using the ${R}^{2}$ test to confirm the absence of overfitting or underfitting, with RF emerging as the most robust and reliable model for half-life prediction model. Multiple RF models with varying feature sets are developed to assess the impact of input features, and SHAP (SHapley Additive exPlanations) values is utilized to quantify the influence of each feature on model output. The TL-RF model incorporating $Q\ensuremath{\alpha}$, neutron number (N), proton number (Z), pairing factor ($\ensuremath{\delta}$), and promiscuity factor (P) are identified as the optimal-feature space, achieving the lowest root mean square deviation ($\ensuremath{\sigma}=0.2$). Furthermore, the TL-RF model is extended to the superheavy region, revealing enhanced nuclear stability near $N=184$, supporting its potential as the next magic number beyond 126, also highlighting an additional zone of relative stability in the neutron range $N=158\ensuremath{-}162$. These findings underscore the effectiveness of transfer-learning-driven machine learning approaches in nuclear physics, offering valuable insights into nuclear decay properties and the stability of exotic nuclei.