Critical operational parameters for metal removal efficiency in acid mine drainage treated by constructed wetlands: An explainable machine learning approach

Constructed wetlands have long been recognized as a sustainable, effective and economical approach for treating acid mine drainage (AMD). The varying components of AMD at different locations impose significant site-specific constraints on the construction and maintenance of these wetlands. Herein, machine learning (ML) was utilized to predict and analyze multi-metal removal efficiencies, and address the complex interactions in constructed wetlands. Five ML models were developed, among which the XGBoost model achieved high apparent accuracy (R² > 0.8) for the removal efficiency of total iron, manganese, aluminum and zinc in the main pipeline. While model performance generally declined (R² decreased by approximately 0.2 overall) under leakage-safe out-of-fold evaluation and forward-chaining time-series tests with naive baselines, tree-based models remained dominant, providing conservative estimates. Detailed feature and sensitivity analyses identified operation Days and inflow Chemical Oxygen Demand as significant predictors of metal removal efficiency. Furthermore, the empirical categories for metal removal, ranked by importance, were inflow parameters in first place, followed by time series, and wetland properties in last place. Partial dependence plots revealed certain ranges of the significant predictors and systematically illustrated their interactions and contributions to the metal removal efficiencies. These findings support near-real-time monitoring and short-horizon operational decisions.