The doubling is a correction layer, not a new set of atmospheric equations
MIT News on October 14 reports that a team including MIT researchers built a machine-learning postprocessing layer to correct existing forecasts from two to six weeks out. The page time is 15:00 UTC, 23:00 in Beijing. Applied to a model they call a gold standard, it doubled forecasting accuracy over that window and placed first in an international forecasting competition. The paper is in Nature. Authors include researchers at Microsoft and at the European Centre for Medium-Range Weather Forecasts.
Beyond two weeks is often called a predictability desert: short range uses initial conditions, seasonal range uses slow boundaries such as ocean temperature, and two to six weeks gets neither. Cohen says forecasting there is, in most cases, like rolling dice.
[1]
They say this is the first time the forecast beat the climate average
The article says the method produces forecasts that for the first time beat climatological averages between two and six weeks. The winning submission combined the European centre's model with another model, then applied the correction, doubling the original model's predictive power in the subseasonal range. Cohen says that before this, models in that window were generally worse than the long-term average.
On extreme weather, the article says only that the improvement was statistically significant. It gives no percentage. Cohen also says the European centre's earlier, non-AI error correction got worse at extremes, because he thinks traditional bias correction pulls forecasts toward the mean. That is his comparison, not a new table in this article.
[1]It sits on top of other models
The correction is described as a wrapper. It does not replace the atmospheric model. The article says it can be used with any model, that it still helps models already bias-corrected by another AI method, and that it can bring weaker models in line with the European centre. Land forecasts improved in most regions. Training used MIT's Engaging cluster. Running the result does not require a supercomputer.
The article quotes Cohen calling it open source and low cost.
[1]Kenya is still at the inquiry stage
Kenya's meteorological office has already been in touch about using the model for the annual monsoon, to improve agricultural yields this winter. The article does not say it is deployed, and it does not say how much monsoon skill would rise. Cohen's examples are fertilizing around rain, protecting a winter harvest from a freeze, and energy and emergency planning: two days of warning for an ice storm is not enough time to move supplies, and two weeks might be. That is a need he is relaying, not the result of a dispatch trial.
The 2021 Texas winter storms killed at least 246 people. The article uses that to explain why lead time matters. It is not a backtest of this model. Lorenz argued in the 1970s for a theoretical limit around two weeks. Cohen thinks the atmosphere falls into repeated patterns, so a signal can remain past that line. That is his judgment.
[1]