Artificial intelligence is changing how we predict river flow—but a new study led by researchers at the University of British Columbia shows that these models often get the right answers for the wrong reasons.
The researchers looked at deep-learning models called LSTMs, which are widely used to forecast how rivers respond to changes in rain, snow and temperature. These models are praised for their accuracy, but the study found that their internal "thinking" often clashes with basic science.
"Accuracy alone isn't enough," said Dr. Ali Ameli, assistant professor in UBC's Department of Earth, Ocean and Atmospheric Sciences, and corresponding author of the paper. "If a model predicts a rain-fed river spike during a heat wave without rain, that's a problem. We need forecasts that reflect the physics of water movement through the landscape."
Accurate river forecasts are critical for flood warnings, drought planning and water management. If AI models misinterpret how heat and evaporation affect rivers, communities could face poor decisions during extreme weather—such as unnecessary reservoir releases or missed flood alerts. As climate change brings more heat waves and shifts in snowmelt, getting these predictions right is essential for safety and sustainability.
The team analyzed more than 1,100 rivers across North America and focused on 672 where the AI achieved strong predictive accuracy. They built a new tool to peek inside the black box and see how the models link rain, temperature, and PET (short for potential evapotranspiration, the atmosphere's demand for water) to river flow. The results were clear: While the models handled rainfall correctly, they frequently misinterpreted the roles of temperature and evaporation.
Put simply, in rainy regions, the models sometimes assume that heat waves or dry air make rivers rise—even when no rain falls. In snowy regions, they often treat PET as the main trigger for snowmelt, instead of temperature. These patterns are not physically realistic.
The study also raises concerns about long-term climate projections made using state-of-the-art AI methods. As global temperatures rise, the relationships among rain, snow, temperature and evaporation will change. If models are built on flawed logic, their predictions for future water availability could be unreliable.
To uncover these issues, the researchers developed a hydrology-specific 'explainable AI' framework. This method isolates the effect of each factor—rain, temperature and PET—on river flow, showing whether the model's reasoning makes sense. Across most regions, the models consistently linked short-term temperature spikes and PET increases to higher river flow, even in places where that shouldn't happen.
"We're not saying abandon deep learning," Dr. Ameli said. "We're saying check it. Build safeguards so models learn the right physics. With better inputs and physical constraints, AI can be a powerful tool for water forecasting."
The team suggests practical fixes: Remove seasonal patterns that confuse the models, add missing physical factors like glacier melt, embed the physics of water movement in the AI algorithms, and use diagnostic tools to screen models before they're deployed. These steps can help ensure that strong accuracy reflects sound physics—not shortcuts in the data.
The study, published in Water Resources Research , was led by UBC PhD student Ara Bayati, with corresponding author Dr. Ali Ameli from UBC and co-author Dr. Saman Razavi from the University of Saskatchewan.
