Researchers Use AI to Improve Weather Forecasts Two to Six Weeks Ahead, the Hardest Stretch to Predict
Scientists from MIT, Microsoft and Europe's leading weather forecasting center report in the journal Nature that an AI add-on makes long-range forecasts noticeably better by learning and correcting the mistakes forecasting programs keep making. It is free, runs on an ordinary computer, and won a forecasting competition last year, though outside experts say the gains are uneven around the world.
A team of researchers from MIT, Microsoft, the European Centre for Medium-Range Weather Forecasts (Europe's main international weather forecasting center) and several universities published a study today in Nature, one of the world's leading science journals. It describes an AI method that improves weather forecasts for two to six weeks ahead.
That stretch is famously hard. A forecast for tomorrow is usually reliable, and a forecast for a whole season can lean on slow-moving patterns like ocean temperatures. In between, small errors snowball, and for years the best computer forecasts did little better than simply saying what the weather usually is at that time of year. MIT calls this range a "predictability desert." Yet it is exactly the window farmers, power companies and emergency planners want: enough warning to prepare for a heat wave or a flood.
The new method doesn't replace any forecasting program. It sits on top of one. Every forecasting program has habits, such as running a little too warm in one place or too dry in another, and it repeats them. The AI add-on, which the team calls probabilistic bias correction, studies years of a program's past forecasts next to what actually happened, learns those habits, and corrects for them, including how confident each forecast should be. Once trained, it doesn't need a supercomputer, and the team has made it free for anyone to use.
The researchers report that it doubled the accuracy of the European center's own AI forecasting program in this range. On the center's traditional forecasting program, which works out the physics of the atmosphere, it improved 92% of the temperature forecasts and 98% of the rain and snow forecasts it was tested on. It also predicted extreme heat and heavy rain three weeks out better than before. A forecast using the method came first in the European center's 2025 real-time forecasting competition, ahead of national weather services and 34 other teams. Kenya's national weather service is testing it for forecasting the rainy season. "Just by correcting them, we can take an unskillful model and make it skillful," said Judah Cohen, an MIT climate scientist who helped build it.
Scientists not involved in the work were more measured. Froila Palmeiro, a researcher at Spain's national research council, said the benefit "is concentrated primarily in the tropical belt and remains quite limited in Europe." Others noted that it has been tested mostly against reconstructions of past weather rather than live forecasts, and that big percentage gains can look larger than they are when the starting accuracy was very low. Matthew Chantry, who leads machine learning at the European center, said moving methods like this into everyday forecasting is "a most challenging step." Even improved, these forecasts deal in odds, such as a better-than-usual chance of a hot week, not a day-by-day outlook.
Sources
- MIT News — Researchers improve weather forecasts looking two to six weeks outnews.mit.edu · primary
- Science Media Centre Spain — Expert reactions to the Nature studysciencemediacentre.es
- Gizmodo — AI just improved extreme weather warnings up to 6 weeks outgizmodo.com