Researchers have developed an artificial intelligence (AI) approach that significantly improves the accuracy of rainfall predictions, helping scientists, water managers, and communities better prepare for droughts, floods, and other climate-related risks.
In a study recently published in Environmental Data Science, a team of researchers addressed a persistent challenge in climate modeling known as "drizzle bias," a problem in which computer models incorrectly predict light rainfall in places that should remain dry. While seemingly minor, these errors accumulate and distort assessments used for water management, agriculture, and hazard planning.
Led by researchers and UVA Environmental Institute affiliates Shivam Singh (also a Climate Fellow), Tom Hartvigsen, and Antonios Mamalakis, the team used generative artificial intelligence to teach climate models to better distinguish between wet and dry areas. Their approach combines deep learning with physical constraints, producing rainfall patterns that more closely match real-world observations.
"Knowing whether rain falls at all can be just as important as knowing how much falls," the researchers noted. "Accurately identifying dry periods and storm boundaries is essential for many environmental and societal applications."
Most global climate models operate at relatively coarse scales, making it difficult to capture local rainfall patterns. Scientists use a process called downscaling to translate those large-scale forecasts into finer-resolution information. However, many existing methods struggle to represent dry conditions accurately, often generating unrealistic low-level precipitation.
The UVA team's new framework uses a type of generative AI known as a Wasserstein Generative Adversarial Network, or WGAN, to predict where rainfall should and should not occur. The improvement could have broad implications for sectors that rely on precise precipitation information.
"We wanted to tackle a problem that climate scientists have recognized for years,” said Shivam Singh, “and that’s the tendency for models to create rain where none should exist. By combining advances in generative AI with climate science, we were able to develop a solution that produces more realistic precipitation patterns and ultimately improves confidence in future projections."
As climate change intensifies both heavy rainfall events and prolonged dry spells, researchers say improving the realism of precipitation forecasts is becoming increasingly important.