�鶹��Ѱ�����

Researcher using AI to improve and personalize flood prediction

When Zhi Li was young, flash floods regularly swept through his village in China.

Once while he was swimming in a river, the water rose suddenly, submerging 10-year-old Li and pushing him downstream. Though massive amounts of rain had fallen upstream, where Li swam the sky remained a clear, deceptive blue.

“It was so quick,” said Li, an assistant professor of civil engineering who joined �鶹��Ѱ�����Boulder in fall 2025. “I had no clue there would be a flood.

Fortunately, a stranger pulled him to safety. The experience inspired Li to devote his life to flood research. “Floods don’t necessarily happen where the rain falls,” Li said. “In my community, many lives were taken by floods. It was devastating.”

Improved simulations

In the United States, the National Weather Service issues weather forecasts and warnings based on expected rainfall and whether the precipitation is likely to overflow nearby rivers, streams or dams.

At Li’s Flood Lab at �鶹��Ѱ�����Boulder, his simulations draw on NWS rainfall predictions and detailed terrain data, including ground permeability, land use, soil type and vegetation cover. By combining U.S. Geological Survey terrain maps with AI, his models predict flooding with far greater precision and speed than the Federal Emergency Management Agency’s traditional physics-based models.

His approach can pinpoint where floodwater will pool at a one-meter resolution, detailed enough to show whether an individual building is likely to flood, compared with FEMA’s 10-meter resolution.

Li said achieving one-meter resolution using FEMA’s traditional approach would require extensive expertise and about a year of computation. Using AI, he can simulate an entire county in just two to three hours at a far lower cost.

“Drilling it down to single households is only possible once you have an AI product like this,” Li said. “It was unimaginable in the past.”

To improve model accuracy, Li checks his results with local experts who know which areas are most vulnerable to flooding. This human insight serves as a guardrail for AI, which can sometimes “hallucinate” false warnings.

Personalized warnings

Li’s research also involves changing how flood risk is communicated. Instead of the traditional top-down approach, issuing a single warning message for an entire county, he envisions a customized alert that starts with individual households and scales up to neighborhoods, communities and counties.

“There is evidence that during floods, some of the people with language barriers were unable to evacuate in time and lost their lives,” Li said. “Personalized warnings can help ensure critical information is clearly communicated so people can act quickly.”

His goal is to use data from the American Community Survey, an annual U.S. Census Bureau report on household demographics, to generate personalized warnings in each resident’s preferred language about whether their home will likely be affected. Evacuation guidance would also reflect a person’s mobility needs and access to transportation.

Flood prediction for all

Li is also developing an AI-powered assistant to democratize access to flood modeling. The approach aims to remove technical barriers that limit flood modeling to experts. Anyone would be able to see an area’s overall flood risk based on historical events. Users could interact with the model in a ChatGPT-style chat to explore flood risk maps by region.

The platform will eventually include real-time alerts and interactive simulations to empower communities to better understand their flood risks during an event without waiting for official warnings.

“I hope the model can be used anywhere in the world to reduce flood damage and provide accessible information to people at risk,” Li said. “That goal has always guided my research.”
��

Measuring the cost

Li also has a vision for measuring a flood’s total impact to human health, ecology, agriculture and the urban economy as a way to mitigate flood costs.

Comparing it to the total gross domestic product (GDP) that is used to evaluate countries’ prosperity, he envisions creating a “gross flood damage” value to help policymakers determine the amount of government aid for a community and how to reduce flood impact in future years.

Invisible costs, such as a loss of income due to flooded agricultural and other work spaces, would also be included.