UNIVERSITY PARK, Pa. — A flash flood can devastate a community for years in a matter of minutes. These disasters account for roughly 85% of floods and cause thousands of deaths annually worldwide , with water levels capable of shooting up 30 feet and receding back to normal levels in the span of a single, unpredictable day, according to the National Weather Service. These catastrophic events are extremely hard to predict, but a solution to this serious problem could come from a silly place: the same artificial intelligence (AI) techniques used to generate images of weightlifting cats or cartoon versions of family photos.
Recently, a team of scientists, including researchers at Penn State, applied the same fundamental techniques behind AI models’ image generation to better predict hourly flash flood risk. Training their model with rainfall and streamflow data collected from over 500 river basins across the continental U.S. between 1990 and 2003, the team demonstrated that not only can their model parse through notoriously unreliable weather data to form accurate predictions, but it can also be easily updated with new data collected from water monitoring gauges around the country. They detailed their approach in a paper published in Water Resources Research .
For years, Chaopeng Shen , professor of civil and environmental engineering and corresponding author on the paper, and his team have leveraged machine learning and AI to build flood forecasting models. Although these national-scale models make accurate predictions using historical rainfall and streamflow — the rate at which water flows in a river — they have trouble capturing the severity of rapid weather swings that can lead to flash flood events. Many of the most devastating floods in recent memory, including the flooding that swept across Central Texas in the summer of 2025, swell and recede in a matter of hours.
“We’ve previously delivered a daily flood forecasting model, but it’s widely understood that the peaks of flash floods occur at an hourly scale,” explained Shen, who holds an additional affiliation at Penn State’s Institute of Energy and the Environment . “Water levels can rapidly rise and recede in just a matter of hours, meaning on a daily scale, they might still seem high, but not disaster-inducing. By calibrating the model for an hourly scale, we can more reliably capture the extraordinary peaks that can wreak havoc.”
To produce predictions, generative AI models must be primed with existing information called training data, making predictions based on patterns observed in previous data sets. The process is like giving a model an image or a sentence with a portion missing and asking it to predict content in that unknown region, Shen said.
“Before the age of AI, scientists had to manually assimilate data, using a lot of assumptions and mathematical gymnastics to make predictions,” Shen said, explaining that the explosion of AI science in the last decade has fundamentally changed how hydrologists, or scientist who study how water moves around the environment, make predictions. “With AI, you can do that much faster and easier by using generative AI trained on existing data.”
This work introduces a forecasting model that has fundamental differences from the large language models that power chatbots like ChatGPT, Shen explained. Their model incorporates diffusion, a training method most used in image generation, where a model receives a complete dataset or picture for training. After that clear data is processed, noise — randomized, unpredictable information — is introduced into the dataset. The AI model is then asked to remove the excess noise and reproduce an interpretable dataset or photo.
Shen said that this training approach is particularly effective for flood prediction, as weather data like rainfall can be incredibly unpredictable on an hourly basis. The model can quantify uncertainty in predictions, reducing doubts by assimilating recent gauge observations into the data set. The approach allows these recent observations to be assimilated into the model without the need for retraining — a process known as inpainting.
“Your basic training inputs include the hourly rainfall, your projected streamflow, all this data that is collected from previous patterns,” Shen said. “We can inject new, almost real-time information into our model, though, asking the model, ‘If the daily streamflow yesterday was a certain amount, and the hourly rainfall today is a certain amount, can you predict what the hourly peak for water level might be today?’”
The team trained the model on hourly streamflow records from 516 river basins across the U.S. from 1990 to 2003, then tested it on a window of data spanning 2009 to 2014, a period the model had never seen during training. Shen said their diffusion approach generally outperformed the best existing hourly deep-learning models, including at capturing how much water flows during high-flow periods, which are notoriously hard to predict. Additionally, the model can run in reverse, Shen explained. Given the streamflow recorded at a river gauge, it can estimate the hourly rainfall that most likely produced it, offering another layer of versatility.
The team plans to continue development, investigating the benefits of different hybrid approaches to training AI models, as well as building models that use more physical variables in predictions. For Shen, though, the goal isn’t about making the “best” model; he said it’s about understanding this technology and using it to help people.
“There isn’t one approach that stands out as the best — each model has their own aspect they are most comfortable with,” Shen said. “At the end of the day, our ultimate goal is to deliver a system that is robust, reliable and can help save people's lives, and I think we are making progress in that realm.”
Other co-authors at Penn State include Wencong Yang, a postdoctoral researcher; Yalan Song, assistant research professor of civil and environmental engineering; Kathryn Lawson, a research associate; and civil and environmental engineering doctoral candidates Haoyu Ji and Leo Lonzarich. Additionally, Ming Pan at Scripps Institution of Oceanography from the University of California, San Diego, contributed to this work.
This work was supported by the Cooperative Institute for Research to Operations in Hydrology (CIROH) and the National Oceanic and Atmospheric Administration (NOAA) Cooperative Agreement under award numbers A23-0271-S001, A24-0491-S001 and A25-0353-S019. Additional funding was provided by the California Department of Water Resources under award number 4600016152. The content is solely the responsibility of the authors and does not necessarily represent the official views of funders.
At Penn State, researchers are solving real problems that impact the health, safety and quality of life of people across the commonwealth, the nation and around the world.
For decades, federal support for research has fueled innovation that makes our country safer, our industries more competitive and our economy stronger. Recent federal funding cuts threaten this progress.
Learn more about the implications of federal funding cuts to our future at Research or Regress .
Water Resources Research
Computational simulation/modeling
Not applicable
Diffusion-Based Probabilistic Modeling for Hourly Streamflow Prediction and Assimilation
21-Jul-2026