Interpretable spatiotemporal machine learning reveals hydrogeological and climatic controls on groundwater heavy metal transport in acid mine drainage.

Acid mine drainage (AMD) generated by sulfide-bearing waste rock dumps poses persistent risks to groundwater through acidic leachate and heavy-metal mobilization. Traditional solute-transport models are limited by the scarcity of fracture-network parameters, while temporal machine learning algorithms often overfit small datasets and underrepresent spatial heterogeneity. Here, we develop an interpretable spatiotemporal machine learning (ML) framework that integrates temporal encodings and climatic forcing with hydrogeological covariates to identify controls on groundwater heavy metal transport. The study focuses on groundwater in the area downstream of a valley-type AMD reservoir retaining waste rock leachate at the Dexing Copper Mine, China. From March 2021 to October 2022, a total of 160 groundwater samples were collected from eight monitoring wells, complemented by multiscale hydroclimatic controls, temporal encodings, and equivalent hydraulic conductivity (fracture-included permeability) derived from pumping tests and hydrogeological surveys. Out of five machine learning models, XGBoost achieved the best predictive performance. SHAP analysis identified fracture-included permeability and 7-day rainfall as dominant external controls. Sulfate (SO42-) acted as a conservative co-migrating tracer closely coupled with Ni transport, whereas the weak association between Mn and Sulfate is consistent with additional retention driven by redox conditions. An ablation study showed that simple temporal encodings improve model performance and that leaching triggered by rainfall exhibits a lagged response. Integrating ML-based evidence with hydrogeological interpretation further indicated that the mechanical nature of faults (opening or sealing) plays a critical role. This study advances mechanistic understanding of fracture-mediated contaminant transport and provides a robust framework for risk assessment and remediation in AMD-impacted aquifers.