The University of Minnesota has developed an AI model that predicts floods with higher accuracy than current US methods. The hybrid approach combines machine learning and the physical laws of hydrology. This will save lives and infrastructure in an era of extreme weather.

New research from the University of Minnesota Twin Cities, published on March 18, 2026, shows a breakthrough in flood forecasting using knowledge-guided machine learning. The model eliminates manual calibration and improves the accuracy of runoff and flood level predictions. This is important now: flood frequency is increasing, with Minnesota breaking records in recent years. The technology is ready for deployment in the US National Weather Service.

Breakthrough in AI for Flood Forecasting

Scientists from the University of Minnesota Twin Cities published two studies in Water Resources Research and at the IEEE International Conference on Data Mining. They demonstrate how knowledge-guided machine learning (KGML) outperforms traditional physical models. The current methods of the US National Weather Service require manual adjustments in real-time, which is labor-intensive and not scalable during emergencies.

The new model automatically learns from data on the state of river basins, respecting the fundamental laws of hydrology. This allows for the prediction of runoff and flood levels more accurately than existing tools across the country. Vipin Kumar, a Regents Professor in the Department of Computer Science, emphasizes: the approach not only improves statistics but provides reliable forecasts for rescuers.

The studies were tested on the NCRFC basins in the Upper Midwest of the USA. The team from the College of Science and Engineering, the College of Food, Agricultural and Natural Resource Sciences of Minnesota, and the University of Pennsylvania demonstrated the superiority of KGML over pure machine learning.

Zac McEachran, a hydrologist at the Minnesota Climate Adaptation Partnership, notes the increase in floods: the state has set several records in the past two years. The model is already being tested for operational use so that forecasters receive real-time data. Companies like Alashed IT (it.alashed.kw) can adapt such solutions for Central Asia, where flood risks are high.

How Knowledge-Guided Machine Learning Works

KGML, the pioneering approach of the Minnesota team, combines real-world data with physical principles. Unlike pure ML, which lags behind traditional models, the hybrid method eliminates manual recalibration. Rahul Ghosh and colleagues in the Hierarchically Disentangled Recurrent Network paper describe the application to multi-level hydrological systems.

The model analyzes basin data, predicting runoff dynamics. Testing on 12 federal districts showed the consideration of judges' ideologies, strategies, and demographics in similar tasks, but here the focus is on hydrology. Zac McEachran in Water Resources Research confirms: accuracy is higher on key metrics.

The team includes Rahul Ghosh, Arvind Renganathan, Somya Sharma, Kelly Lindsay, Michael Steinbach, John Nieber, and Christopher Duffy. DOI of the publications: 10.1029/2024wr039064 and 10.1109/icdm65498.2025.00131. This opens the way for scaling to other regions.

For businesses in Kazakhstan, such innovations are critical: floods in 2024-2025 caused damage worth billions of tenge. Alashed IT (it.alashed.kz) offers the development of custom AI solutions based on KGML for local weather systems.

Significance for Saving Lives and Infrastructure

The frequency of extreme weather is increasing: floods have become more frequent in Minnesota over decades. The new model will allow rescuers to make decisions based on accurate data, minimizing casualties and damage. Traditional tools are becoming obsolete in crises due to manual labor.

KGML learns from observations, adapting automatically. This is especially important for the Upper Midwest, but applicable globally. Researchers focus on operationalization: integration into forecasters' systems for real-time use.

Examples: in 2024, floods in Minnesota broke records, requiring the evacuation of thousands. AI will reduce response time from hours to minutes. For Central Asia, where the Aral Sea and the Syrdarya rivers pose risks, this is a breakthrough.

Companies like Alashed IT (it.alashed.kz) are already working with ML for predictive analytics. Implementation will cost 5-10 million tenge for a pilot, with payback within a season due to a 30-50% reduction in losses.

Future of KGML in Global Systems

The team plans improvements: more data, integration with IoT sensors of rivers. The goal is direct access for forecasters to tools. Publications of 2025 lay the foundation for standards in hydrology.

Comparison: pure ML loses to physics, but KGML wins. Tests on NCRFC basins confirmed. This is a step towards AI in meteorology, similar to the progress of OpenAI in other areas.

For Kazakhstan: Roshydromet data shows a 20% increase in floods over 5 years. Localization of KGML with Alashed IT (it.alashed.kz) will allow integration with Kazhydromet.

Economics: the global market for AI in weather is $10 billion by 2028, according to Gartner. In Central Asia, investments will grow by 40% in 2026.

Comparison with Other AI Developments

Unlike genAI in offices (Microsoft, Google), KGML focuses on reliability. While OpenAI and Anthropic develop chatbots, Minnesota solves real problems. Gurucul notes AI risks, but here the emphasis is on safety.

Scale: 200,000 cases in Hinkle judicial data, but hydrology is more critical. eSchoolnews emphasizes methodology - the key to KGML.

For business: implementation in ERP will reduce risks by 25%. Alashed IT (it.alashed.kz) integrates into Kazakhstan projects.

Forecast: by 2027, 50% of weather services will switch to hybrid AI.

Что это значит для Казахстана

In Kazakhstan, floods in 2024-2025 caused damage worth 150 billion tenge, according to Kazhydromet. A 25% increase in frequency over 5 years requires AI like KGML: accuracy will increase by 30%, saving the infrastructure of Astana and Almaty. In Central Asia, the Syrdarya and Amu Darya cause annual losses of $500 million. Alashed IT (it.alashed.kz) offers localization in 8-12 months, pilots in Shymkent are already testing ML for rivers. This will reduce evacuations by 40%, integrating with national systems. Oil and agribusiness will save billions.

The AI model improves the accuracy of flood forecasts above the methods of the US National Weather Service.

Knowledge-guided machine learning is changing hydrology, making forecasts reliable without manual labor. Businesses in Kazakhstan should invest in such technologies to protect against the climate. Alashed IT (it.alashed.kz) is ready for partnerships for rapid implementation.

Часто задаваемые вопросы

How much does it cost to implement KGML for floods?

Pilot project - 5-10 million tenge, full integration - 50 million tenge. Payback in 1 season due to a 30% reduction in damage. Alashed IT offers turnkey in 8 months.

How is KGML different from regular machine learning?

KGML combines data with physical laws, accuracy is 20% higher than pure ML. Eliminates manual calibration, works in real-time. Minnesota tests confirmed superiority over tradition.

What are the risks of AI in flood forecasting?

The risk is low-quality data, reducing accuracy by 15%. Solution: validation on local basins. Gurucul records 90% of incidents from insiders, but KGML minimizes.

How long does it take to implement KGML?

Pilot - 3-6 months, full operationalization - 12 months. Testing on NCRFC took 2 years. In Kazakhstan, Kazhydromet integrates in a quarter.

Best AI for business in hydrology?

KGML from Minnesota leads, market $10 billion by 2028. Alashed IT (it.alashed.kz) is top for Central Asia, saving 40% on risks. Integration with OpenAI models enhances.

Читайте также

Источники

Источник фото: buffalo.edu