Physics meets machine learning in a flood model that outperforms both alone
When a river starts to rise, the people who forecast floods face a maddening constraint: their best physics-based models need constant hand-tuning. As conditions change, forecasters at the National Weather Service must manually adjust model parameters based on real-time field observations - a process that is labor-intensive under normal circumstances and nearly impossible to scale during a multi-basin weather emergency, when dozens of rivers may be flooding simultaneously.
Machine learning, the obvious alternative, has its own problem. Pure data-driven models can find patterns in historical streamflow data, but they have consistently underperformed the traditional physics-based tools that forecasters actually rely on. The patterns they learn are only as good as the data they trained on, and when an unprecedented event arrives - the kind of flood that breaks records - a model trained on past norms may be exactly the wrong tool for the job.
Bridging two approaches
Two paired studies from the University of Minnesota Twin Cities, published in Water Resources Research and the Proceedings of the IEEE International Conference on Data Mining, describe a hybrid approach that outperforms both methods. Called knowledge-guided machine learning (KGML), the framework embeds the fundamental laws of hydrology directly into the architecture of a machine learning model, then lets the model learn from observed data within those physical constraints.
The result is a system that automatically learns the state of a river's watershed from observed data - eliminating the need for manual recalibration - while still respecting the physics of how water moves through landscapes. It predicts streamflow and flood levels with greater accuracy than the tools currently deployed across the United States.
Not just better statistics
Vipin Kumar, Regents Professor in the Department of Computer Science and Engineering and a senior author on both papers, draws a distinction between statistical improvement and operational reliability. The value of the KGML approach is not simply that it generates better predictions on a test dataset. It is that the predictions remain physically plausible even in extreme conditions, because the model cannot violate hydrological constraints.
Pure machine learning models lack this guardrail. They can produce predictions that are statistically reasonable but physically impossible - forecasting negative streamflow, for instance, or violating water balance constraints. When emergency managers are making evacuation decisions based on model output, physical plausibility is not optional.
KGML, an approach pioneered by researchers at the University of Minnesota, is designed to bridge that gap. By encoding physical laws as hard constraints within the model architecture, rather than treating them as optional regularization terms, the system achieves higher accuracy without sacrificing the interpretability and reliability that operational forecasters need.
Minnesota's rising waters
The research has particular urgency in the researchers' home state. Zac McEachran, a coauthor and research hydrologist with the University of Minnesota Climate Adaptation Partnership, notes that Minnesota has experienced increasing floods over recent decades, with several river-level records broken in just the last couple of years. As extreme precipitation events become more frequent - a trend projected to continue under most climate scenarios - the demand for accurate, fast, and scalable flood prediction tools will only grow.
The current reliance on manual recalibration creates a bottleneck. A single forecaster can adjust model parameters for one or two river basins during a storm event. When multiple basins flood simultaneously - as happened during several recent Midwest weather emergencies - the system simply cannot keep up. An automated model that maintains accuracy without manual intervention would fundamentally change the operational capacity of flood forecasting agencies.
From research to operations
The research team is now working to make the model operational, with the goal of putting KGML tools directly into the hands of National Weather Service forecasters for real-time risk assessment. That transition from research prototype to operational deployment involves substantial engineering challenges - integration with existing data pipelines, real-time computational performance, and validation against the full diversity of hydrological conditions across the US.
The collaboration spans the University of Minnesota's College of Science and Engineering and College of Food, Agricultural, and Natural Resource Sciences, along with Pennsylvania State University. The work was supported by the National Science Foundation, the State of Minnesota Weather Ready Extension, and the Minnesota Pollution Control Agency.
The fundamental insight - that combining physics knowledge with machine learning produces better results than either alone - extends well beyond flood forecasting. The KGML framework has potential applications across environmental science, from drought prediction to water quality monitoring to ecosystem modeling. But in a world where flood damages are measured in billions of dollars annually and rising, the immediate application may be the most consequential one.