Ensemble Machine Learning for Metals Recovery Prediction of Lithium Battery Bioleaching Process

This study successfully demonstrated that the Ensemble Machine Learning (stacking) method significantly enhances metal recovery prediction in spent Lithium-ion battery bioleaching. Single models showed limited performance, with XGBoost achieving R 2 =88.5% and KNN at R 2 =66.9%. The best stacking model, XGBoost-XGBoost, attained the highest prediction accuracy ( R 2 = 94.9%) and robust cross-validated performance ( R 2 = 93.10%). SHAP analysis identified pulp density (optimal 10%), initial pH (optimal 2-3), particle size, and temperature as the most critical process parameters. Furthermore, Aspergillus niger highly promotes the recovery of Lithium and Copper, while Acidithiobacillus ferrooxidans is more effective for Nickel and Cobalt, consistent with their distinct bioleaching mechanisms. In addition, the ML model demonstrated its use in inverse prediction and process optimization: in this case, it successfully predicted that increasing the pulp density from 0.5% to 5% with Acidithiobacillus thiooxidans raised the metal recovery rate from approximately 60% to 99%.