MIT-led AI post-processing doubles subseasonal forecast accuracy, tops ECMWF contest
A team including MIT and Microsoft researchers published a Nature paper describing a machine-learning "post-processing" wrapper that learns a forecasting system's systemic errors from historical data and corrects its output, doubling the predictive skill of a gold-standard ECMWF model in the two-to-six-week subseasonal range and, for the first time, beating climatological averages there. In ECMWF's 2025 real-time forecasting competition the approach produced the most accurate global forecasts across all variables and lead times, outperforming numerical models from six operational centers, ECMWF's own AI model and systems from 34 competing teams, while also showing a statistically significant improvement in extreme-weather prediction. The researchers describe the tool as model-agnostic, open source and low cost, and say it can be layered onto other models, including ones already bias-corrected by AI; Kenya's meteorological office has already asked about using it for monsoon forecasting.
If such a cheap post-processing layer is widely adopted, power grids, farmers and emergency managers could get weeks rather than days of warning ahead of extreme weather.
In 2021, Texas suffered the worst energy infrastructure failure in state history, following three severe winter storms in February. At least 246 people were killed as a result of the disruptions, which led to food, water, heat and electricity shortages.
If energy officials had had more time to prepare for the storms, for example by bringing in more trucks of natural gas to disrupted areas, damage from the event could have been mitigated. Unfortunately, weather forecasts beyond two weeks are notoriously unreliable, making for a tighter window for energy operators, farmers, and emergency response officials around the globe.
A new technique by a team including researchers at MIT significantly improves weather forecasting beyond two weeks. The researchers developed a machine-learning model that learns to correct errors in existing forecasting models to improve their output over the two-to-six-week time frame. When applied to a gold-standard forecasting model, it doubled forecasting accuracy over that time frame, placing first in a prominent international forecasting competition.
In a new paper in Nature, the researchers describe their approach and show how it enables forecasts that for the first time beat climatologic averages between two and six weeks.
#MIT#Microsoft#European Centre for Medium-Range Weather Forecasts#Nature#DTN#Google DeepMind