Abstract
Radioactive cesium released from nuclear activities poses persistent threats to ecosystems and human health, necessitating low cost and highly efficient remediation techniques such as adsorption process. Machine learning has been considered to be a powerful tool to guide the design of novel adsorbents. This study developed a robust predictive framework by optimizing XGBoost, LightGBM, CatBoost, and Random Forest using Bayesian optimization and five-fold cross-validation. In addition, two Stacking ensemble strategies with Ridge and ElasticNet were developed to be meta-learners to compensate for the prediction errors of individual base models. It was found that the Stacking ensemble could substantially enhance the predictive performance, with the Ridge meta-model achieving the best outcomes, yielding a coefficient of determination (R2) of 0.9761 and a root-mean-square error (RMSE) of 10.9208, outperforming all base learners. SHAP analysis reveals that Stacking leverages the complementary strengths of individual base learners. This study implements a Stacking-based transfer learning strategy that addresses the prediction accuracy bottleneck for novel adsorbent systems under limited data, offering predictive support for the rational design of advanced cesium adsorbents with both scientific and practical value.