New AI technique promises to reduce errors in weather forecasts made weeks in advance; understand

A new artificial intelligence technique could help reduce errors in weather forecasts made two to six weeks in advance, a range in which meteorological models still lose a lot of accuracy.

The method, called probabilistic bias correction (PBC), was presented in a study published this Wednesday (14) in the journal Nature.

Instead of creating a new meteorological model, the tool works as a correction layer: it learns from errors in systems that already exist and tries to adjust them in subsequent forecasts.

Today, forecasts can maintain greater accuracy for up to about two weeks. After that, small errors accumulate and, combined with the chaotic nature of the atmosphere itself, make it more difficult to predict rainfall, temperature and other factors with confidence.

“Our new system takes the chosen model, learns how it tends to make mistakes and corrects those errors,” explained Lester Mackey, a researcher at Microsoft Research and one of the authors of the study.

In tests, PBC was applied to models from the European Center for Medium-Range Weather Forecasts (ECMWF).

In the institution's artificial intelligence system, the technique doubled, on average, the predictive capacity of sub-seasonal projections.

In a traditional model that was already undergoing corrections, there was an improvement in 98% of the points analyzed for precipitation, 92% for temperature and 91% for atmospheric pressure.

In practice, the system analyzes old predictions, compares the results with what actually happened and identifies error patterns.

It then uses this information to correct new projections, also taking into account the most recently observed conditions.

The researchers also observed gains in forecasting extreme events, such as intense rain, floods and cold spells.

In an analysis of floods recorded between 2016 and 2026, for example, PBC increased forecast skill by 76% to 92%, depending on the analyzed horizon.

The technique was also tested in an international competition organized by ECMWF in 2025. A solution that used PBC came in first place for all meteorological variables and periods evaluated, beating models from six operational centers and systems from 34 other teams.

One of the examples cited by the researchers occurred in the United States. The system was able to indicate two strong winter storms recorded at the beginning of 2026 19 to 26 days in advance.

This type of forecasting can be useful in areas such as agriculture, water and energy management, and disaster preparedness. In Kenya, the method is already being tested to help meteorologists improve rainfall projections used by farmers.

Experts interviewed by Science Media Center Spain, however, say that the results still need to be viewed with caution.

Froila M. Palmeiro, researcher at the Institute of Geosciences in Spain, says that the biggest gains appeared mainly in tropical regions and that the improvement was even more limited in areas such as Europe.

Another point is that PBC does not replace meteorological models.

“You don’t build a new atmospheric model from scratch,” said Robert Monjo, a researcher at the Foundation for Climate Research.

According to him, if the source model fails in a very rare or complex situation, the method may not be able to correct the problem.