Weather forecasts can reach up to 6 weeks with AI: the study

Develop weather forecasts that can have a time frame of up to six weeks? It could be possible, also thanks to a new technique developed by a team of researchers from the Massachusetts Institute of Technology (MIT) which would be able to significantly improve weather forecasts well beyond two weeks.

The developed model

Specifically, the researchers created a machine learning model capable of correcting errors in existing forecasting models, improving their accuracy over a period of between two and six weeks. Applied to a reference forecasting system, the model essentially made it possible to double the precision of forecasts in this time period. In a new article published in the scientific journal "Nature", the researchers described their approach and explained how this can contribute to obtaining forecasts that, for the first time ever, exceed the canonical climatological averages.

Extreme events

“This method not only improves time-averaged forecasts for this range, which is already very useful for farmers and energy operators trying to balance supply and demand, but also improves forecasts of extreme weather events, whether floods, cold spells or something else,” commented Judah Cohen, co-author and MIT researcher. "It is essential to predict both, but extreme events could be the most impactful aspect for society," he added. The MIT project, in fact, was able to highlight a statistically significant improvement in the ability to predict extreme meteorological events.

MIT Engaging and a specific test

To build the model in question, the researchers used MIT Engaging, a supercomputing cluster managed by MIT's Office of Research Computing and Data. With the particularity, however, that the resulting model does not require a supercomputer to function. To test their model, among other things, the team of researchers collaborated with the European Center for Medium-Range Weather Forecasts (Ecmwf), considered the best forecasting center in the world. Every season the organization offers prediction competitions between models and in one of these, in 2025, the MIT researchers' model ultimately stood out for its important predictive capabilities. For the first time, in particular, the model offered an improvement over the simple use of climatological averages. “Until now, no one had been able to demonstrate that a model predicted the weather better than long-term averages for these sub-seasonal forecasts,” Cohen continued. “Virtually all previous models turned out to be worse: you could literally flip a coin and get a better result. Some even questioned whether it was worth investing in models for this time frame. We, however, consistently beat climatological averages with our models.”