Researchers double an AI weather model’s forecast skill two to six weeks ahead

Forecasts two to six weeks ahead can help farmers and energy operators prepare—if they are useful. In a Nature study, researchers report that a machine-learning correction layer improved forecasts in this difficult window, including predictions of extreme weather. On one leading AI weather system, it doubled forecast skill from a modest starting point.

Learning from the forecasts that went wrong

The research, described in an October 14 MIT account republished by Phys.org, targets subseasonal forecasting: the period between two and six weeks ahead. This window sits between two sources of predictability. The influence of current atmospheric conditions fades, while slower-moving climate forces are more useful over longer periods.

The team’s response was not to build another forecasting engine from scratch. Instead, it developed probabilistic bias correction, or PBC, a machine-learning framework that learns systematic errors in historical forecasts and adjusts future output. Probabilistic forecasts describe the likelihood of different outcomes rather than giving one fixed prediction.

MIT research scientist and co-author Judah Cohen says Microsoft researcher Lester Mackey first proposed the idea. The correction layer learns from past forecast errors, adjusts predictions and adapts as new forecasts and observations come in. It can be added to existing systems, even those already corrected by another AI algorithm.

A competition result, then broader tests

In the European Centre for Medium-Range Weather Forecasts’ 2025 real-time competition, the team’s global forecasts placed first across all weather variables and lead times. A Microsoft Research publication page from April describes the entries as outperforming forecasting systems from six operational centers, an international model ensemble and 34 competing teams.

PBC doubled the modest subseasonal skill of ECMWF’s AI Forecasting System, according to the study. It also improved ECMWF’s physics-based model, which simulates how the atmosphere changes and already had an operational error correction. Gains covered most of the pressure, temperature and precipitation targets tested.

91%Pressure targets improved

For pressure targets, PBC improved forecast skill in 91% of cases in ECMWF’s operationally bias-corrected dynamical model.

92%Temperature targets improved

PBC improved forecast skill for 92% of temperature targets in the same model.

98%Precipitation targets improved

PBC improved forecast skill for 98% of precipitation targets in the same model.

Cohen says the corrected forecasts consistently beat climatology—the long-term weather averages that provide a basic benchmark. The Nature work also showed improvements in land forecasts across most regions globally and gains when the correction was applied to other forecasting models.

Importantly, the researchers reported a statistically significant improvement in extreme-weather prediction, not just better average forecasts. Cohen says ECMWF’s own non-AI correction had worsened extreme-event predictions. His explanation: conventional correction tends to pull predictions toward the average, while unusual events lie at the edges of the probability range.

From a research result to seasonal planning

Building the correction model required MIT’s Engaging supercomputing cluster. Running the resulting model does not require a supercomputer, according to the MIT account. Cohen describes the method as open source and low cost, making its computing requirements different from those of the systems used to develop it.

Kenya’s meteorological office has contacted the team about using the method for monsoon-season forecasts, Cohen says. That is interest in implementation, not a confirmed deployment. The intended agricultural uses include timing fertilizer around rainfall, protecting crops before freezes and preparing for drought.

For energy operators, the potential benefit is preparation time. Cohen says officials have told him that two days’ notice of an ice storm may be insufficient, while two weeks could allow adequate preparations. Whether interested agencies adopt the correction layer remains the next practical step beyond the research and competition results.