MIT machine learning method improves two-to-six-week weather forecasts
Researchers from MIT and partner institutions have developed a machine-learning postprocessing technique that improved weather forecasts in the difficult two-to-six-week range. Applied to a leading forecasting system, the approach doubled predictive accuracy in that subseasonal window and took first place in a European Centre for Medium-Range Weather Forecasts (ECMWF) competition.
The work, published in Nature, addresses a period in which forecasts have often been no more useful than long-term climatological averages. The team reports that its method is the first to outperform those averages between two and six weeks ahead, while also delivering a statistically significant improvement in predicting extreme weather.
A correction layer for existing models
The model does not replace a weather simulator. It acts as a wrapper around an existing forecasting system, learning from historical mistakes and shifting future probabilistic forecasts to correct them. Judah Cohen, an MIT research scientist and co-author, describes the approach as postprocessing designed to quantify model errors and remove them.
Subseasonal forecasting sits between short-range weather prediction, which can draw heavily on current atmospheric conditions, and seasonal outlooks influenced by slower boundary forces. Cohen calls the interval a “predictability desert.” Dynamical models numerically simulate the atmosphere and oceans, while AI systems learn from historical data; the researchers note that AI models have improved short-range prediction but have performed worse than dynamical models over longer periods.
For training, the team used MIT Engaging, a supercomputing cluster run by MIT’s Office of Research Computing and Data. The completed model does not need a supercomputer to run. It can also adapt its correction as additional forecasts and data become available.
Competition results and extreme events
Last year, the researchers submitted three systems with the correction layer to ECMWF’s seasonal forecasting competition. The winning entry combined ECMWF’s forecasting model with another model before applying the machine-learning correction. It doubled the original model’s predictive power in the subseasonal range.
The paper also found that the technique improved land forecasts across most global regions. When applied to other models, it improved them as well, in some cases bringing less capable systems to the level of ECMWF forecasts. The researchers describe the add-on as open source and low cost.
Extreme events are a central part of the result. Traditional bias correction can tend toward the mean, which may weaken prediction of outcomes at the edges of a probability distribution. The team found that its AI-based correction significantly improved predictions of events including flooding and cold spells, whereas ECMWF’s non-AI error correction had worsened extreme-weather prediction in the cited comparison.
Planning value before an event arrives
The research is relevant to sectors that need lead time rather than only a near-term forecast. The 2021 Texas winter storms illustrate the operational stakes: the resulting energy infrastructure failure caused food, water, heat and electricity shortages, and at least 246 deaths. More advance warning could have allowed energy operators to position additional natural-gas trucks in disrupted areas.
The Republic of Kenya’s meteorological office has contacted the team about using the model to improve annual monsoon forecasting for agriculture. Cohen also points to decisions on fertilising around rain, protecting winter crops ahead of freezes, and preparing energy systems and emergency responses for an ice storm. For businesses, the practical implication is to evaluate corrected subseasonal probabilities alongside existing planning processes, where an extra two weeks may change what preparations remain feasible.