Predicting Groundwater Storage from Seasonal Managed Aquifer Recharge: Insights from Machine Learning and Interpretable AI Techniques

Abstract Managed Aquifer Recharge (MAR) is widely used to enhance groundwater storage and support sustainable water use. To support site selection and planning, machine learning (ML) models are increasingly used as computationally efficient surrogates for traditional numerical models. While ML has shown promise for steady-state simulations, capturing transient responses remains a challenge, yet these are essential for understanding how recharged water is retained through dry periods. In this study, we use ML to model transient MAR effects by decomposing the groundwater storage time series after recharge ceases into two components: the MAR-response and a decay coefficient, assuming exponential storage decline. This simplified representation captures long-term storage dynamics, with U-Net and XGBoost accurately predict these components (R2 > 0.82) for the Baakse Beek catchment in the sandy, drought-sensitive soils of the Netherlands. The trained models are computationally efficient, enabling high-resolution prediction across 1.6 million sites in under 20 seconds, making large-scale scenario testing and optimization feasible. Explainable AI techniques, specifically SHAP values, were used to identify key site management decisions and surface water properties that control MAR effectiveness. Emphasizing model interpretability enhances transparency and fosters trust among hydrologists and stakeholders, improving the potential for broader application and generalizability.