Ensemble learning physics-guided surrogates for efficient emulation of MODFLOW-USG transient simulations in a multilayered, heterogeneous, and highly dynamic groundwater system

• Hybrid surrogate framework integrates MODFLOW-USG with ensemble learning. • Surrogate models predict hydraulic heads 65 × faster than MODFLOW-USG. • Sensitivity analysis with surrogates is completed in seconds vs. hours traditionally. • The aquifer system shows depletion due to high pumping rates in the study area. • Surrogates enable real-time scenario analysis, optimization, and probabilistic forecasts. Physically based groundwater flow models such as MODFLOW-USG are widely used for simulating aquifer dynamics; however, their application to transient systems is often constrained by high computational cost and intensive data requirements. This study presents a hybrid modeling framework that integrates a physics-based MODFLOW-USG model with data-driven surrogate machine learning approaches to investigate the influence of hydrological stresses (pumping and recharge) on groundwater dynamics in the unconfined Quaternary aquifer system of Debrecen, Eastern Hungary. A transient MODFLOW-USG model was developed using an unstructured grid and calibrated through inverse modeling to estimate hydraulic conductivity and storage parameters across four hydrostratigraphical layers. The calibrated model was then used to generate training data for ensemble-based surrogate models, including Random Forest, Gradient Boosting, and XGBoost, incorporating hydrological stresses and water budget components as input features. Model performance was evaluated using statistical metrics, including the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE). Results show that the surrogate models accurately reproduced MODFLOW-USG hydraulic heads (R² = 0.97–0.99; RMSE as low as 0.09 m), effectively capturing both regional flow patterns and localized drawdown. Water budget analysis revealed that pumping exceeded recharge by nearly a factor of two during peak stress periods, leading to significant storage depletion. The surrogate models achieved substantial computational efficiency, reducing simulation time from approximately 1.2 hours to 18 minutes, while preserving physical consistency with the governing flow processes. This integrated approach provides a robust and efficient tool for rapid scenario analysis, uncertainty assessment, and sustainable groundwater management in complex aquifer systems.