Study region showing 30 drainage basins in Massachusetts, USA. The black dots are the precipitation gauging stations. Credit: Journal of Hydrology (2026). DOI: 10.1016/j.jhydrol.2025.134427

When engineers and planners design roads, bridges and dams, they rely on hydrological models intended to protect infrastructure and communities from 50- and 100-year floods. But as climate change increases the frequency and severity of floods, existing models are becoming less and less reliable, new Cornell research finds.

To combat this uncertainty, physics-based models should be supplemented with AI hydrological models, and creating regional flooding estimates, rather than relying on site-specific estimates, according to the study, published in the Journal of Hydrology.

"Models are simplified representations of the real world, so we validate them against past observations to ensure they work well under historical conditions," said first author Sandeep Poudel, a doctoral student in the lab of Scott Steinschneider, associate professor of biological and environmental engineering.

"However, climate change is making droughts and floods more frequent and severe. This means the future won’t look like the past, which leads to a crucial question: How much should we trust our models that were validated on historical data to make projections for the future? And which models are better suited for long-term water infrastructure planning?"

To answer those questions, the researchers first developed a "virtual hydrolab" experiment, composed of 1,000 years worth of synthetic data based on current and future climate conditions. This data included extreme events like floods and droughts, and daily values of climate and hydrologic conditions including air temperature, precipitation, soil moisture, evaporation and runoff. Then they tested six flood-prediction models against their virtual dataset to see which would come closest to predicting important outcomes, under both present and future conditions.

The models fit into three categories: traditional models that rely on physical equations about hydrological processes; AI-based models that take in all of that data and make predictions based on learned input and output relationships; and a hybrid model combining the traditional and AI models. They found that while there were very large uncertainties among all the models tested, the AI model performed best.

Steinschneider warned that overperformance by the AI model in this one, virtual case study should not be taken as a reason to throw out physics-based models, but rather to continue studying and refining both types of models.

"Rather than assuming we can precisely predict how floods will change in every watershed, we should acknowledge the limits of our models and look for patterns that persist across larger regions," he said. "That broader perspective gives planners a much more reliable guide for preparing infrastructure and protecting communities in a warming climate."

Regional predictions were significantly more stable than site-specific predictions, the researchers found. Making flood change projections at multiple river basins, then aggregating those projections and averaging them, created more robust and accurate findings.

But the most troubling finding, they said, was the models’ unreliability in predicting flooding under climate change conditions.

"This is concerning because these are the models and hydrologic data that we commonly use today to make decisions about how to design bridges, roads and water infrastructure into the future—and they are not good enough," Poudel said.

More information: Sandeep Poudel et al, Uncertainty in estimating the relative change of design floods under climate change: a stylized experiment with process-based, deep learning, and hybrid models, Journal of Hydrology (2026). DOI: 10.1016/j.jhydrol.2025.134427

Citation: AI improves flood projections under climate change (2026, January 8) retrieved 8 January 2026 from https://phys.org/news/2026-01-ai-climate.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.