Predicting whether next season will bring a devastating flood or a crippling drought often comes down to a race against the clock. For climate scientists, forecasting the future movement of water requires some of the most complex—and computationally demanding—math on Earth.
Across the United States, water-related hazards threaten electrical grids, reservoirs and other critical infrastructure. Energy providers need localized, long-range forecasts to prepare for these risks, but current prediction systems are limited by scientific uncertainty and the enormous computing power required to run them.
Now, a University of Maryland-led team is using generative artificial intelligence to change that.
The researchers have been awarded $750,000 in Phase I funding from the U.S. Department of Energy's new Genesis Mission to develop an AI-powered framework that can generate faster, more detailed seasonal and multi-year water forecasts. The nine-month award includes approximately $470,000 for UMD, the lead institution.
The project, "An Agentic AI Framework for Seasonal-to-Interannual U.S. Water Prediction," was announced last month at the inaugural Genesis Mission Summit in Washington, D.C. It was one of 270 projects selected from more than 5,000 proposals submitted to the highly competitive national program. If chosen for Phase II, the team will expand its early prototypes into operational tools with broader real-world applications.
Led by Jinwoong Yoo, an assistant research scientist at UMD's Earth System Science Interdisciplinary Center, the project brings together experts in atmospheric science, artificial intelligence, high-performance computing and mathematics.
The UMD team includes four faculty members from the University of Maryland Institute for Advanced Computer Studies (UMIACS): Maria Molina, assistant professor of atmospheric and oceanic science; Chris Metzler, assistant professor of computer science; Abhinav Bhatele, professor of computer science; and Haizhao Yang, professor of mathematics.
Researchers from Oak Ridge National Laboratory, the University of Florida, the University of Colorado Boulder and Silurian AI are also collaborating on the effort.
For Molina, the partnership reflects a growing need to combine advances in AI with decades of expertise in Earth system science. The team's goal is to integrate cutting-edge AI forecasting with the Department of Energy's flagship Energy Exascale Earth System Model (E3SM), creating a system that can produce faster predictions without sacrificing scientific rigor.
“If we can get this framework right, we can give planners a clearer picture of the hazards their communities may face seasons to years in advance by revealing how event likelihood, intensity and local impacts may evolve,” Molina said.
Traditional physics-based Earth system models such as E3SM simulate interactions among the atmosphere, ocean, land and water cycle across the globe. Seasonal-to-multiyear forecasting requires many simulations to capture the range of possible outcomes, creating tradeoffs among resolution, forecast length, and ensemble size. Turning those forecasts into useful information at the community level remains difficult because predictability decreases at longer timescales, while local hazards depend on processes that global models cannot fully represent.
The UMD team's approach uses AI to accelerate nearly every step of that process.
The framework will first incorporate satellite observations of soil moisture and groundwater using machine learning to improve the model's starting conditions. Generative AI models trained on decades of E3SM simulations will then emulate the behavior of the original model, producing more than 100 possible forecast scenarios to estimate both likely outcomes and uncertainty more than 100 times faster than traditional methods.
Image-enhancement techniques will sharpen those regional forecasts into street-level flood maps with one-meter resolution. Behind the scenes, autonomous AI software agents will coordinate the complex workflow—from preparing model inputs to running simulations and analyzing results—allowing researchers to focus on evaluating forecast quality and scientific reliability.
To ensure the predictions remain grounded in physical science, Molina will lead the team's validation efforts. Researchers will test the framework against major hurricanes, evaluate whether it captures climate patterns such as El Niño and use explainable AI techniques to determine which factors drive the model's forecasts.
"The real bottleneck has always been computational cost," Metzler said. "By using AI to emulate these massive Earth system models, we can run simulations in minutes on supercomputers that used to take days. That speed is what makes real-time protection of local grids actually possible."
Before deploying the system on the Department of Energy's leadership-class supercomputers, the team will develop, train and evaluate its initial models using UMIACS' high-performance computing infrastructure.
Ultimately, the researchers hope their work will help utilities, water managers and emergency planners make better decisions before disasters strike.
For the team, success won't be measured simply by faster algorithms or more efficient computing. It will be measured by more accurate forecasts that help communities prepare for floods, protect critical infrastructure and better manage water resources before extreme conditions arrive.
—Story by Melissa Brachfeld, UMIACS communications group