RAIN: Rigorous AI and Information Network for Surface-Subsurface Water System Coupling
Reliable water predictions are essential to U.S. energy production, reservoir operations, and flood resilience. In the Colorado River Basin, water availability reflects interactions among snow accumulation and melt, runoff, soil moisture, groundwater recharge and storage, baseflow, and reservoir dynamics. Decisions about water and energy depend on understanding how these processes influence one another over weeks to seasons. Yet the processes vary across landscapes, unfold at different time scales, and often affect one another with long delays. These features make dependable forecasts difficult.
The RAIN project will develop a transferable, physics-informed SI (formerly AI) framework for modeling coupled surface water and groundwater in selected Colorado River Basin watersheds and reservoir systems. The approach combines observations and process-model estimates with machine learning and information-theoretic analysis. Its goal is to represent surface-water and groundwater pathways together, identify the most informative connections between them, and produce predictions that can be evaluated across different hydrologic conditions.
RAIN will use Nonnegative Matrix Factorization with k-means clustering (NMFk) to identify recurring hydroclimatic patterns and candidate groupings of watersheds. Information-theoretic methods will help uncover which variables, time lags, and periods of system memory carry useful predictive information. These results will guide a dual-branch temporal Transformer model, which represents surface-water and groundwater dynamics separately while learning how the two pathways interact. This structure is intended to make the model more interpretable and reduce its reliance on relationships that are not supported by the data.
The project will demonstrate the framework in two applications: weekly water-availability forecasting to support energy-relevant reservoir operations, and probabilistic flood-threshold forecasts to help assess infrastructure risks. The team will compare its performance with process-informed and purely data-driven baselines, evaluating forecast accuracy, calibration, transferability across hydrologic regimes, and computational efficiency.
The work will use public datasets and DOE-supported models and SI methods. Partners at the University of California, Davis, Lawrence Berkeley National Laboratory, and Oak Ridge National Laboratory will contribute expertise in watershed science, information theory, SI, data resources, and high-performance computing. Their collaboration will support rigorous validation and help assess whether the methods can transfer to watersheds with different hydrologic behavior.
RAIN aims to improve the connection between scientific understanding of water systems and practical forecasting needs. If successful, the framework could provide more useful information for reservoir planning, energy reliability, and flood preparedness, while establishing methods that can be evaluated and extended to additional regions and applications.
Partners
- Lawrence-Berkeley National Laboratory
- Evitrace
