As climate change alters precipitation patterns and drives more frequent extreme weather, traditional tools for estimating future flood risk are showing increasing uncertainty, according to new research by Cornell University scientists. A study published in the Journal of Hydrology and highlighted in a recent Cornell press release finds that artificial intelligence (AI)‑based hydrological models can improve projections of how design floods may change under future climate conditions, particularly when paired with or used alongside existing physics-based approaches.
The research addresses a central challenge facing engineers, planners and water managers: floods that were once viewed as “50‑year” or “100‑year” events no longer occur on historically expected timelines. As flood frequency and severity increase with climate change, models validated on past conditions may be less reliable for anticipating future extremes. “Models are simplified representations of the real world, so we validate them against past observations to ensure they work well under historical conditions,” said Sandeep Poudel, the study’s first author and a doctoral student in biological and environmental engineering at Cornell. “However, climate change is making droughts and floods more frequent and severe. This means the future won’t look like the past…”
To explore how different modelling frameworks perform under changing climate conditions, the research team created a “virtual hydrolab”, which is a synthetic dataset spanning 1,000 years of simulated climate and hydrologic variables, including air temperature, precipitation, soil moisture, evaporation and runoff. The virtual setup enabled testing of six different flood‑prediction models under controlled scenarios that represent both current and future conditions.
Despite the strong performance of the AI model in this stylised experiment, the researchers cautioned against discarding physics-based approaches
These models fell into three categories: traditional process-based models that apply physical equations to represent hydrological processes; deep learning models that learn relationships directly from data; and hybrid models combining process-based frameworks with machine learning enhancements. When tested against the synthetic dataset, the AI model was found to outperform the other models in estimating relative changes in design floods under future climate scenarios, even though all models exhibited considerable uncertainty.
The peer-reviewed article highlights that structural uncertainty and equifinality, the tendency for different model structures and parameter sets to produce similar outputs, dominate the overall uncertainty in flood change projections. While deep learning models were effective in estimating change, methods that pool projections across multiple sites reduced variance and improved reliability.
Despite the strong performance of the AI model in this stylised experiment, the researchers cautioned against discarding physics-based approaches. Scott Steinschneider, an associate professor of biological and environmental engineering and a co-author, said 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.”
One key insight from the study is that projections made at broader spatial scales tend to be more stable and meaningful than site-specific estimates
One key insight from the study is that projections made at broader spatial scales tend to be more stable and meaningful than site-specific estimates. Regional predictions, which average outcomes across multiple river basins, showed less variability, suggesting that planners and engineers could derive more robust guidance by focusing on aggregated trends rather than individual watershed predictions. “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,” Steinschneider said.
Another troubling finding was how unreliable many models were when forced to project conditions under climate change. The press release noted that this unreliability is significant because these are the very tools commonly used to inform infrastructure design for bridges, roads, dams and other critical systems. “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.
The study’s authors contend that blending AI with traditional techniques and focusing on broad, regional patterns of change could provide a more reliable foundation for long-term planning in a warming world. They emphasise that AI models alone are not a panacea; rather, the research points to the value of incorporating machine learning as a complementary tool within a broader modelling toolkit.
