Machine learning breaks two-week barrier, delivering accurate weather forecasts up to six weeks

A research team that includes scientists from the Massachusetts Institute of Technology (MIT), Harvard, the University of Toronto, the European Centre for Medium-Range Weather Forecasts (ECMWF), and Microsoft, published a paper in the journalNaturetitled “Enhancing Subseasonal Forecasts with Machine Learning”, it presents a long-awaited breakthrough: an automated correction model that doubles the accuracy of weather forecasts in the critical range of two to six weeks, a range that has been known for decades as the “predictability desert.”

Why were the two weeks a turning point?

In the 1970s, MIT meteorologist Edward Lorenz set a theoretical limit: after two weeks, atmospheric noise overwhelms any initial signal, making random guessing better than models. This belief “had a strong impact on the field and may have hampered exploration of forecasts beyond two weeks,” says MIT research scientistJudea Kochen. Kochen was not fully convinced: “Theoretically the atmosphere can generate an infinite number of patterns, but it tends to fall into preferred patterns repeatedly. When something happens, you can anticipate the next domino, and the one after that.”

Approach: a correction layer that learns from model errors

Instead of building a new climate model, the team builta post-processing layer(postprocessing wrapper) that sits atop any existing forecast model, whether dynamical or AI-based. The layer uses machine learning to measure the baseline model’s historical errors and correct them adaptively. The more forecasts and data become available, the correction improves automatically.

The subseasonal range usesprobabilistic forecasts(probabilistic forecasts) that provide probabilities for a range of outcomes rather than a single deterministic prediction. The correction layer operates on these probability distributions, not on individual values.

Real-world test: the ECMWF competition and the result

The team submitted three models that incorporated the correction layer toECMWF, the world’s leading forecasting centre, which holds seasonal model competitions. The winning model combined two ECMWF models and then applied the correction on top of them. The result:doubling the predictive skillof the original model in the subseasonal range, and for the first time a model outperformslong-term climatological averages(climatological averages) in this range.

“So far, no one has been able to demonstrate a model that predicts the weather better than long-term climatological averages for these subseasonal forecasts,” says Kochen. “They were almost all worse; you could flip a coin and get a better result. Our models consistently outperform the climatological averages.”

Extreme events: the most important improvement

The model showedstatistically significant improvement in predicting extreme weather, such as floods, cold spells, or others. When ECMWF developed its own error correction (not based on AI), its performance worsened on extreme events. Kochen explains: “Traditional correction always pulls toward the mean, whereas extreme weather lies on the tail of the probability distribution.”

Open-source, low-cost, portable

The model was trained on a high-performance computing cluster at MIT (MIT Engaging), butrunning does not require a supercomputer. It is an “add-on” (add-on wrapper) that works with any model, even those that have been corrected with another AI technique. The meteorological office inKenyais already in contact with the team to apply the model to improve seasonal wind forecasts and increase agricultural yields this winter.

Regional bet: from water security to energy security

The Gulf, Egypt, and the Levant live on the edge of decisions that rely on a two-week window: filling reservoirs before floods, arranging gas imports before a cold snap, preparing cooling networks before a heat wave. A model that extends the reliable decision window to six weeks changes the cost-risk equation. For water utilities in Riyadh, power-plant operators in Abu Dhabi, or agriculture ministries in Cairo, an extra two weeks of trustworthy forecast means better stock management, reduced emergency purchases, and saved lives. The open-source, low-cost model removes the entry barrier: no need for supercomputing infrastructure, just a decision to adopt the correction layer on top of existing operational models.

The honest takeaway

The paper does not discard dynamical weather models; it corrects them, not replaces them. The remaining challenge is operational: integrating the correction layer into the daily forecast pipelines of national centres and demonstrating performance over full seasons in the region’s hot, arid climate. Yet, as a peer-reviewed achievement in Nature, open-source and deployable on ordinary computers, this is the kind of AI research that moves from the lab to the control room within months rather than years.