Robust managed aquifer recharge (MAR) design aided by graph neural networks

Rapid population growth and agriculture development have led to unsustainable exploitation of groundwater, a trend likely to be exacerbated by future climate change. Managed Aquifer Recharge (MAR) emerges as a cost-effective strategy for replenishing depleted aquifers, thereby supporting long-term groundwater sustainability. However, the highly heterogeneous nature of groundwater processes necessitates fine spatial and temporal resolution models for designing and evaluating MAR. The high computational burden of physics-based model simulations constrains existing MAR studies to limited scenarios, missing opportunities to evaluate MAR potential across the full range of uncertainties from climate change, policy shifts, and infrastructure development. Recent advances in artificial intelligence offer a promising solution to the trade-off between high spatial-temporal precision and computational efficiency through surrogate models. In this study, we leverage recent advances in attention-based Graph Neural Networks (aGNN) to develop a surrogate model for MAR (GNN-MAR), which allows us to capture multi-scale network structures across river systems, groundwater flow, and MAR infrastructure. Trained on a high-resolution physics-based integrated surface water and groundwater model, GNN-MAR is tailored for two MAR approaches, i.e., in-channel recharge and agriculture MAR (Ag-MAR). We apply GNN-MAR to the Baoding Plain in the North China Plain, one of the world’s most severely groundwater depleted regions. The search for optimal MAR schemes is conducted within large ensembles generated under the XLRM (eXogenous uncertainties, policy Levers, Relationships, Measures) framework, which encompasses climate change scenarios, groundwater pumping policies (X), MAR schemes (L), GNN-MAR (R), and groundwater sustainability targets (M). The framework enables identification of MAR schemes robust to deep uncertainties. Our study provides valuable insights for the development of high-fidelity surrogate models for integrated surface-groundwater systems and demonstrate the potential of AI-based surrogate model for robust decision-making in groundwater recharge management.