MIT is revolutionising weather forecasting with the MicroDuet model and machine learning
The AI model predicted a catastrophic cold snap in the eastern United States four weeks in advance – an event that traditional operational systems completely failed to detect
Key points
Weather modelling has reached a historic milestone. A research team led by the Massachusetts Institute of Technology (MIT) has developed a machine learning technique capable of improving the accuracy of forecasts over a time span of between two and six weeks. The study, published in *Nature*, introduces the Probabilistic Bias Correction (PBC) architecture, a system capable of quantifying and removing the systematic errors (bias) inherent in both traditional dynamic physical models and those based on artificial intelligence.
“For decades, sub-seasonal forecasts – those extending beyond 14 days – have been our Achilles’ heel, as we have struggled to beat historical averages,” explains Judah Cohen, a climatologist at MIT and co-author of the study. “With this approach, we are not only forecasting the weather more accurately, but we are also providing critical sectors with the necessary operational leeway to prepare for extreme weather events, transforming uncertainty into a manageable risk.”
Alerts up to 150 per cent more accurate and top performance in tests
When applied to the model of the European Centre for Medium-Range Weather Forecasts (ECMWF), the PBC correction doubled overall accuracy, improving estimates of pressure, temperatures and precipitation in over 90 per cent of the target areas. The excellence of the system (known as MicroDuet) earned it first place in the international AI Weather Quest 2025 competition, outperforming the official dynamic forecasts of six government meteorological centres worldwide (including NOAA and the ECMWF) and the models of 34 international teams.
The leap in quality is particularly evident on extreme events: the methodology has increased predictive accuracy by 51 per cent for heatwaves and by 39 per cent for abnormal pressure events. However, it is in the area of hydrogeological damage that the greatest impact is seen. Using the GDACS database, the model has increased predictive capability by 76–92 per cent for general flooding and by up to 150 per cent for severe flooding, providing valuable lead time where traditional systems effectively eliminated any warning capability.
The impact on the agricultural sector
The implications for the agri-food sector are immediate. The visibility of temperature and rainfall over a three- to six-week period enables a shift from reactive management to predictive planning. Farmers can optimise fertiliser application, schedule sowing and plan defences against sudden frosts. This versatility has already led to operational collaborations, such as the one with the meteorological department of Kenya for monsoon forecasting to protect local crops.
