麻豆淫院

A new method for predicting how rainfall contributes to river flow across the entire US has been developed by an international team of scientists.

The technique, which combines physics knowledge with advanced artificial intelligence (AI), aims to help decision-makers better prepare for extreme weather and climate impacts.

By integrating deep learning with watershed physics, the team led by Clemson University in collaboration with Cardiff University and IHE Delft Institute for Water Education, developed interpretable, physics-guided AI models for rainfall鈥搑unoff simulation.

The model outperformed several traditional hydrologic approaches while also estimating the likelihood of a range of different river-flow events, helping scientists identify limitations and improve forecasts.

Their findings, in the journal Water Resources Research, could enhance river flow prediction, water management, and climate resilience across the US and beyond.

Lead author Dr. Vidya Samadi, Assistant Professor of Water Resources Engineering at Clemson University, said, "This work was prompted by the need for more accurate and understandable tools to predict how rainfall translates into river flow.

"Traditional hydrologic models often struggle with complex watershed behaviors and complex river flow simulation.

"By embedding physical constraints into modern AI architectures, we developed two probabilistic, physics-guided models that not only outperform existing methods but also offer clear insights into how river flow responds to rainfall. This helps researchers and other stakeholders identify errors and enhance river flow forecasting," added Dr. Samadi who has recently completed a Visiting Academic Fellowship at the University of Cambridge.

The AI models were trained on extensive observed and simulated rainfall鈥搑unoff data using advanced deep learning architectures, including transformers, to capture complex patterns over time.

"By embedding physical attributes from watershed, our models achieved both high accuracy and interpretability," Sadegh Sadeghi Tabas, another of the paper's authors.

"This enables scientists, water managers, and other stakeholders to better understand hydrological processes, identify potential vulnerabilities, and assess uncertainties in river flow forecasts."

Their novel technique also gives clear explanations of how predictions are made鈥攕omething traditional models often lack, explains co-author Professor Catherine Wilson from Cardiff University's School of Engineering.

"Our study shows how combining physics with advanced, interpretable AI models leads to more accurate and transparent predictions of river flow. This helps scientists and researchers better understand and, crucially, trust AI models, so that they can be deployed with confidence to predict water flow and manage water resources, which are critical for planning, agriculture, and drinking water supplies worldwide," says Wilson.

"Our goal was to build river-flow prediction tools that are not only more accurate but also easier to trust," explains Professor Biswa Bhattacharya from IHE Delft Institute for Water Education and another co-author of this research.

"By combining physics with AI, we can give scientists and water managers clearer insights into when and why rivers respond the way they do, which is critical for preparing communities for floods, droughts, and weather extremes."

The team plans to expand the model's capabilities by integrating additional environmental variables, such as soil moisture, land use changes, and climate projections, to improve predictive accuracy under diverse conditions.

They also aim to work closely with stakeholders including water managers, emergency responders, policymakers and others, to apply the AI in real-world decision-making, refine its interpretability, and support more resilient water resource management.