By Abdullahi Lukman
New research from Cornell University suggests that artificial intelligence–based hydrological models can strengthen projections of future flood risk as climate change disrupts historical rainfall and flooding patterns.
The study, published in the Journal of Hydrology, finds that AI models can better estimate how design floods may shift under future climate conditions, especially when used alongside traditional physics-based approaches.
The findings come as engineers and water managers face growing uncertainty: floods once considered rare, such as “50-year” or “100-year” events, are occurring more often. Models calibrated on past climate conditions are increasingly unreliable for predicting future extremes.
According to lead author Sandeep Poudel, a doctoral researcher at Cornell, climate change means “the future won’t look like the past,” undermining long-standing validation methods.
To test different modelling approaches, the researchers built a “virtual hydrolab,” a synthetic dataset representing 1,000 years of climate and hydrological variables, including precipitation, temperature and runoff.
Using this controlled environment, they evaluated six flood-prediction models under both present and future climate scenarios.
The models fell into three groups: traditional process-based models, deep learning models, and hybrid models combining physics and machine learning.
The AI model performed best at estimating relative changes in future design floods, although all approaches showed significant uncertainty. Structural uncertainty and equifinality—where different models produce similar results—were identified as major sources of error.
The study also found that regional-scale projections were more reliable than site-specific estimates. Averaging results across multiple river basins reduced variability, suggesting that planners may gain more robust guidance by focusing on broader patterns rather than individual watersheds.
Despite the strong performance of AI in the experiment, the researchers caution against abandoning physics-based models. Co-author Scott Steinschneider stressed that AI’s success in a virtual case study should be seen as a reason to refine and combine approaches, not replace existing methods.
The authors conclude that integrating AI with traditional hydrological models and emphasising regional trends could provide a more dependable basis for infrastructure planning in a warming climate, while acknowledging the limits of current prediction tools.
