• Ten machine learning algorithms were compared for predicting heavy metal fractions in tailings, with Extra Trees achieving the highest accuracy (R² = 0.946 on validation).
• Electronegativity emerged as the most influential feature affecting morphological fractions, highlighting its role in metal mobility.
• An ensemble of the top three models (GBDT, Extra Trees, CatBoost) further improved predictive performance, demonstrating the value of model stacking.
• The proposed framework offers a rapid, cost-effective alternative to traditional sequential extraction experiments, facilitating environmental risk assessment and resource recovery.