Explainable machine learning framework for predicting cobalt ion removal by natural hematite

Abstract

Cobalt contamination in aquatic systems presents considerable environmental and public health concerns. A hybrid artificial intelligence framework was developed for accurate modeling, and this study investigated natural hematite (α-Fe₂O₃) as an economical and sustainable adsorbent for the extraction of cobalt ions (Co²⁺). A maximum removal effectiveness of 89.7% was achieved at 100 ppm during batch adsorption tests conducted under various operational conditions, including contact time (10-120 min), adsorbent dosage (0.01-0.10 g), initial cobalt concentration (100-1000 ppm), and temperature (34-90 °C). Statistical analyses indicated that temperature and contact time were the most influential parameters (p < 0.01). Among the evaluated models, the contrastive learning XGBoost approach demonstrated superior predictive accuracy, achieving R² = 0.987 ± 0.0137 (95% CI: 0.9748-0.9845), RMSE = 1.8676 ± 0.8509 (95% CI: 1.2590-3.4763), MAE = 0.0500 ± 0.0200 (95% CI: 0.0100-0.0300), and MAPE = 1.6753 ± 0.5802 (95% CI: 1.2602-2.0904). The proposed framework's stability and robustness were confirmed by confidence interval analysis, and Shapley Additive Explanations (SHAP) indicated that temperature and contact time were the primary factors influencing removal efficiency.