Integrating machine learning and transfer learning for data-driven prediction of lithium and rubidium adsorption

The rising demand for lithium (Li) and rubidium (Rb) underscores the need for efficient recovery from aqueous systems. Inorganic ion-exchange materials offer a promising solution, with lithium-ion sieves (LIS) commonly used for Li recovery and potassium metal hexacyanoferrate (KMFC) showing strong selectivity for Rb. This study presents a machine learning (ML)-based framework for predicting Li and Rb adsorption performance using historical experimental data. Three ML models—Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Boosted Trees (BT)—were developed for Li adsorption, with RF and BT achieving superior predictive performance (R 2 = 0.96 ± 0.01 and 0.95 ± 0.01; RMSE = 3.71 ± 0.38 and 4.14 ± 0.78, respectively). SHAP analysis revealed that operational parameters, particularly solution pH, have the greatest influence on adsorption efficiency. Due to limited available data for Rb, transfer learning was employed to adapt the pre-trained Li model for Rb adsorption prediction. This approach significantly improved performance compared to a baseline model trained solely on Rb data, achieving an average R 2 of 0.88 ± 0.02 and RMSE of 22.75 ± 1.86 across multiple random seeds. In contrast, the baseline model yielded R 2 0 . 84 ± 0.03 and RMSE 26.74 ± 1.28. These results demonstrate that integrating transfer learning into a tree-based modelling framework offers a scalable and effective strategy for enhancing predictive accuracy in adsorption systems with limited experimental data. • A machine learning framework was developed to predict lithium and rubidium adsorption using metal-based adsorbents. • Random Forest and Boosted Trees models achieved high accuracy in lithium adsorption prediction (R 2 > 0.95). • SHAP analysis identified solution pH as the most influential factor affecting lithium adsorption performance. • Transfer learning from lithium data improved rubidium adsorption prediction accuracy and reduced variability (R 2 0 . 88 ± 0.02, RMSE 22.75 ± 1.86). • The integration of transfer learning in a tree-based framework offers a scalable solution for adsorption systems with limited experimental data.