Explainable AI for Predicting Micropollutant Fate and Transport in Coupled Surface Water–Groundwater Systems

Abstract

Micropollutants move from rivers into aquifers through hyporheic exchange and riverbank filtration, where sorption, redox-dependent biodegradation, photolysis, travel time and dilution jointly control their fate. This study builds an explainable AI framework that predicts the fraction of a micropollutant surviving the river–aquifer–well pathway and audits its explanations against exact Shapley attributions of a process-based reference model. A coupled model of river transport, hyporheic exchange and oxic–anoxic bank filtration generated 12,000 cases for 80 virtual compounds. Under leave-compounds-out validation a multilayer perceptron reached R² = 0.96, against 0.81 for XGBoost and 0.60 for a travel-time screening model. SHAP explanations of the perceptron reproduced the true feature ranking almost exactly (Spearman ρ = 0.997), whereas TreeSHAP of the less accurate XGBoost model under-weighted compound charge three-fold (ρ = 0.94). Oxic half-life, aquifer conductivity and sorption dominated attenuation; compound and aquifer properties carried about 85% of the attribution and the river and hyporheic zone less than 10%. Accuracy improved most with more training compounds. Explanations inherit the errors of the model they explain, so their fidelity should be audited before XAI supports riverbank-filtration management.