AI foundation model development for water forecasting

HydroORBIT: A Physics-Informed Multimodal AI Foundation Model for Coupled Surface-Groundwater Prediction in Energy-Critical Regions

This project seeks to improve prediction of surface and groundwater systems across the continental U.S. Predictive understanding of surface and groundwater processes remains fragmented across regions, despite decades of progress. Diverse hydrometeorology, heterogeneous subsurface conditions and uneven data availability make these forecasts a challenge. HydroORBIT, a multimodal AI foundation model trained on ORNL’s Frontier supercomputer, will enable physically consistent prediction of surface and subsurface water states for faster assessments of hydropower potential, groundwater availability for energy, and flood risk. This capability will be delivered through a modular workflow spanning data curation, model pretraining and fine-tuning, physics-based attribution, and customizable predictions to allow real-time decision-making. HydroORBIT will rely on the American Science Cloud for scalable data integration and codesign with the Transformational AI Models Consortium for distributed training and automated workflows. Results will be integrated into the Genesis platform for community use.