As extreme weather intensifies, University of Minnesota researchers have developed a knowledge-guided AI model that could make flood forecasts faster and more accurate. The hybrid system blends physics and machine learning to support forecasters during high-stakes emergencies.
As extreme storms and record-breaking floods become more common, a team at the University of Minnesota Twin Cities has developed an artificial intelligence model that could give communities more time to prepare.
In paired studies published in Water Resources Research and the Proceedings of the IEEE International Conference on Data Mining, researchers showed that a new type of machine-learning system can improve how scientists predict river flows and flood levels. The work, done in collaboration with Pennsylvania State University, suggests that blending traditional physics-based models with modern AI could reshape flood forecasting across the United States.
Today, forecasters at the National Weather Service rely heavily on physics-based hydrologic models. These tools simulate how rain and snow move through a watershed, but they often need constant, hands-on tuning. Experts must manually adjust the models in real time based on stream gauges, rainfall reports and other field observations.
That process is time-consuming and difficult to scale when a major storm is unfolding over many counties at once.
The team’s new model takes a different approach. It combines the strengths of those long-trusted physics models with machine-learning techniques that automatically learn the current state of a river’s watershed from observed data. In practice, that means the system can update itself as new information comes in, without the labor-intensive recalibration that forecasters now perform by hand.
According to the researchers, this hybrid method can predict streamflow and flood levels more accurately than the standard tools now used nationwide.
The key is what the team calls a “knowledge-guided” form of AI, according to Vipin Kumar, a Regents Professor in the Department of Computer Science and Engineering and a senior author on the papers.
“The knowledge-guided approach allows the model to learn from real-world data while still respecting the fundamental laws of hydrology,” Kumar said in a news release. “This is not just about improving statistical accuracy. It is about providing reliable, actionable forecasts that emergency managers and forecasters can trust when making high-stakes decisions.”
Purely data-driven models, which ignore the underlying physics of how water moves through soil, rivers and reservoirs, have been tested before. But in many cases, traditional physics-based models still outperform those black-box AI systems, especially when conditions change or when data are limited.
The Minnesota group’s “knowledge-guided” strategy is designed to bridge that gap. By encoding hydrologic principles into the AI model, they aim to get the best of both worlds: the adaptability and pattern-recognition power of machine learning, plus the stability and physical realism of established hydrologic theory.
The research comes at a time when many communities are already feeling the effects of more frequent and intense flooding.
“We have already seen increasing floods within the last few decades in many parts of Minnesota, including several flood records set within the last couple of years,” added co-author Zac McEachran, a research hydrologist with the University of Minnesota Climate Adaptation Partnership. “It’s vital that we improve our ability to predict these events so we can protect lives and infrastructure.”
Better forecasts can help emergency managers decide when to issue evacuation orders, where to place sandbags and temporary flood barriers, and how to protect critical infrastructure such as bridges, wastewater plants and power stations. More accurate predictions, delivered earlier, can also guide long-term planning decisions about where to build homes, roads and levees.
Beyond Minnesota, the approach could be applied to river basins around the country and potentially worldwide, especially in regions where climate change is altering rainfall patterns and snowmelt in ways that past data alone cannot fully capture.
Co-authors of the studies include Rahul Ghosh, Arvind Renganathan, Somya Sharma, Kelly Lindsay and Michael Steinbach from the University of Minnesota’s Department of Computer Science and Engineering and John Nieber from the Department of Bioproducts and Biosystems Engineering, as well as Christopher Duffy from the Department of Civil and Environmental Engineering at Pennsylvania State University.
The work builds on a broader effort at the university to use advanced data science to tackle environmental challenges.
Next, the researchers plan to refine the model and work toward making it operational, with the goal of putting these tools directly into the hands of forecasters. That would allow agencies to test the system in real-time conditions and explore how it performs during actual flood events, not just in retrospective studies.
If successful, the knowledge-guided AI model could become part of the next generation of forecasting systems that help communities adapt to a wetter, more volatile climate, turning cutting-edge computer science into practical protection on the ground.
Source: University of Minnesota