AI System Doubles Weather Forecast Accuracy Out to Six Weeks Ahead, Study in Nature Finds

Researchers from MIT, Harvard, the University of Toronto, Microsoft, Brazil's National Institute for Pure and Applied Mathematics (IMPA), and the European Centre for Medium-Range Weather Forecasts (ECMWF) published a study in the journal Nature showing that a machine-learning postprocessing layer can double weather forecast accuracy in the 2–6 week 'subseasonal' range — the period historically considered too distant for reliable daily forecasting. The system acts as a 'wrapper' that learns to detect and correct systematic errors in existing models, such as ECMWF's AI Forecasting System. Applied to that system, it improved accuracy for 91% of pressure targets, 92% of temperature targets, and 98% of precipitation targets. Critically, it beat climatological averages for the first time in this range, even for extreme-weather events — a benchmark that traditional bias-correction methods had never cleared. The model requires no supercomputer and can be applied to any forecasting system. The researchers said the improvement has major real-world applications: better 2–6 week warnings could help energy grid operators, farmers, and emergency managers prepare for severe weather. As an example, the team noted that adequate warning could have helped mitigate the 2021 Texas winter storms, which killed at least 246 people. The team won first place in ECMWF's international forecasting competition with the technique.