Bridging process-based simulation and deep learning: A location-aware framework for quantifying water diversion impacts on groundwater recovery

North China Plain, China We propose a hybrid MODFLOW–Location-Aware LSTM (MODFLOW-LA-LSTM) framework to improve groundwater storage change prediction and support intervention attribution. MODFLOW-NWT simulations provide process-informed predictors (water-balance components and simulated storage), while an LSTM learns residual corrections to storage. “Location-aware” capability is implemented by concatenating a categorical location encoding (province/city one-hot vector) with the hydrologic feature sequence, enabling a single model to learn region-specific response patterns across basin-to-municipal scales. We then apply a counterfactual trend–deviation analysis to quantify deviations from pre-transfer groundwater trajectories associated with SNWD. The hybrid framework better captures spatial heterogeneity in NCP groundwater dynamics across contrasting managed settings than process-only simulations, improving the representation of both depletion phases and subsequent stabilization. The counterfactual analysis provides a bounded estimate of SNWD-associated shifts in post-2014 groundwater trajectories at regional and subregional scales, highlighting stronger and more coherent recovery signals in major receiving metropolitan areas and more variable responses in agriculturally dominated regions where concurrent management and climate variability can modulate outcomes. These results offer spatially differentiated evidence to support adaptive groundwater regulation and surface-water substitution strategies in the NCP. • MODFLOW–LA–LSTM corrects groundwater storage changes across the North China Plain. • Forcing-only baseline and ±10% perturbations confirm robustness beyond MODFLOW. • Trend–deviation counterfactual gives bounded SNWD-associated trajectory shifts. • Recovery is heterogeneous: clearer in metros, more variable in farming provinces.