AI Models Identify Ocean Patterns That Drive U.S. Winter Precipitation
Researchers at the University of Virginia used artificial intelligence to uncover which climate patterns predict winter precipitation across the United States months in advance. The work, published in Artificial Intelligence for the Earth Systems, combined deep learning with explainable AI to solve a longstanding challenge in climate science: seasonal precipitation forecasting.
The findings could help communities prepare for droughts, floods, and water shortages, particularly across the southern United States, where winter precipitation proved substantially more predictable than in northern regions.
Why researchers need to verify what AI actually learns
The study's central contribution was not prediction accuracy alone. It was trust.
Antonios Mamalakis, who led the research, said the critical question is whether an AI model "predicts correctly for the right reasons." This sits at the core of explainable AI-an emerging field focused on understanding how AI systems reach their decisions.
In climate science, this matters enormously. "When you're using an AI model for a climate task, especially in high-stakes settings like forecasting the trajectory of a hurricane, you need to make sure it hasn't learned shortcuts," Mamalakis said. "Because if a new event falls outside the distribution of events the model was trained on, those shortcuts will not apply anymore, and the model can derive significantly wrong predictions."
AI systems used in environmental forecasting must be evaluated on whether they rely on physically meaningful climate signals, not just statistical patterns that happen to work in historical data.
The tropical Pacific Ocean dominates winter weather in the South
The study found that Florida, Georgia, the Carolinas, and Virginia showed the strongest forecasting skills for winter precipitation. This pattern aligns with decades of climate research linking winter precipitation in the South to El Niño and La Niña events in the tropical Pacific Ocean.
"The signal of El Niño and La Niña events is much stronger over the southern U.S.," Mamalakis said. "During El Niño years, the jet stream tends to intensify and shift to the south, bringing more winter storms and wetter conditions."
Across nearly all AI systems tested, the tropical Pacific consistently emerged as the dominant source of predictive information. The models also identified important climate signals in the tropical Atlantic Ocean, suggesting multiple ocean basins influence seasonal precipitation patterns.
When AI models agree, they've likely learned something real
The researchers introduced a concept called "meta consensus"-the idea that when different AI systems independently reach similar conclusions about what drives precipitation, that agreement signals genuine physical understanding.
In this study, different models demonstrated the highest agreement during periods when climate conditions became more predictable, particularly during strong El Niño and La Niña years. Mamalakis described this as evidence that "the models have learned something physical."
This represents a shift in how AI is used in science. "We are entering a period where AI can become a scientific tool, not just a forecasting tool," he said.
The energy cost of climate AI
Mamalakis acknowledges a tension in scaling AI for climate research. "On the one hand, AI can help accelerate science and help us gain new knowledge," he said. "On the other hand, at large scales, especially in massive data centers, it can require ridiculous amounts of energy."
For this study, the models were relatively small and trained locally. But as AI systems scale up, the electricity and water demands of data centers become a significant consideration.
Still, the potential benefits are substantial. Reliable seasonal forecasts months ahead of time could help communities manage water resources, prepare for floods and droughts, and respond to climate extremes before they occur.
To develop expertise in applying AI to scientific problems, consider exploring AI for Science & Research learning paths that cover AI applications in scientific discovery and research validation.
Your membership also unlocks:
