A prototype hyper-resolution groundwater digital twin for the contiguous United States: integrating physics-based modeling, machine learning, and observations

To advance large-scale hyper-resolution groundwater modeling, we leverage existing physically-based simulation results and water table depth (WTD) observations to develop a prototype groundwater digital twin for the contiguous United States (CONUS). This framework represents a continuously updatable virtual representation that integrates observations with physics-based predictions to support operational decision making. An adjusted random forest model is trained to downscale 1 km simulation results from the integrated physically-based hydrologic model ParFlow-CLM to 1 arcsec (∼30 m) and bias-correct to observations, producing daily 1 arcsec WTD and associated uncertainties across the CONUS. Trained on water year 2003 (WY2003), the model reliably estimates temporal variations in WTD at most previously unseen grid cells, achieving a median Spearman’s ρ of 0.66. Over half of the grid cells that contain continuous daily records in WY2003 exhibit good performance, with ρ ≥ 0.5. At the subbasin scale, the digital twin captures more detailed groundwater variability than ParFlow-CLM, especially in areas with strong surface–groundwater interactions. During the future time period (WY2024), the model consistently outperforms ParFlow-CLM, increasing the median ρ by 0.13. Enabled by multi-GPU computing, the digital twin generates each daily 1-arcsec resolution WTD map in approximately 35 min of GPU time, providing insights into groundwater systems across multiple scales. The success of the physics-guided machine learning (ML) digital twin highlights the advantage of combining ML and physically-based modeling in groundwater applications. This groundwater digital twin demonstrates a path toward operational capability, enabling near-real-time monitoring, scenario exploration, and decision support at unprecedented spatial resolution.