Microsoft researchers say they have developed a machine-learning method that can make weather predictions more accurate two to six weeks in advance, a range where both physics-based and AI forecasting systems tend to lose reliability. The Verge reported on October 14 that the method, called Probabilistic Bias Correction, was published in Nature and released as open-source software.

Rather than building another standalone weather model, the system is designed to improve models that already exist. It uses machine learning to detect systematic bias in their predictions and then applies corrections. That approach could make the research easier to adopt than systems that require agencies to replace their forecasting infrastructure.

Distorted weather symbols pass through a correction prism and emerge aligned.
Probabilistic Bias Correction adjusts systematic errors in existing weather models.

The underlying problem grows as forecasts stretch further into the future. Physics-based models simulate atmospheric processes, while AI models identify patterns in historical data, but both must use approximations when observations are incomplete or atmospheric behavior is difficult to represent. According to The Verge, those approximations can create persistent biases whose errors accumulate over longer forecast windows.

One example is a model trained on historical conditions that are colder than the present climate. As greenhouse-gas emissions warm the planet, that history can tilt predictions toward temperatures that no longer reflect current conditions. Probabilistic Bias Correction is intended to identify patterns like that and adjust the forecast rather than rebuild the entire model.

In the researchers’ analysis, the method improved results from a leading physics-based model operated by the European Centre for Medium-Range Weather Forecasts for 98 percent of the precipitation targets and 92 percent of the temperature targets examined. The Verge also reported that the framework doubled the accuracy of ECMWF’s newer AI forecasting system in the study. Those figures describe the paper’s evaluated targets; they do not mean every forecast improved by 98 or 92 percent.

A small processor projects clearer rain patterns over farms and emergency shelters.
The lightweight method is designed to run on CPUs and could help agencies with limited computing resources.

The project’s computing requirements are central to its appeal. Microsoft Research New England senior principal researcher Lester Mackey told The Verge that the tool can run on standard CPUs, unlike physics simulations that rely on supercomputers or generative AI systems that depend on powerful GPUs. The paper includes researchers from Microsoft, MIT, ECMWF, and other institutions, and the work had already received recognition in ECMWF’s AI Weather Quest competition.

The lighter footprint could matter most for weather agencies with limited computing resources. Genevieve Flaspohler, a paper author who worked on the project while interning at Microsoft, said models such as PBC are relatively easy to train and can be adapted by scientists for particular locations. Flaspohler now leads the nonprofit Rhiza Research, which is working with the Kenya Meteorological Service Authority, with Gates Foundation funding, to test whether AI methods including PBC can improve early warnings ahead of an expected extreme rainy season.

The research still faces a familiar gap between published results and operational forecasting. ECMWF strategic lead for machine learning Matthew Chantry called the paper a useful contribution while cautioning that turning an idea into a working operational process is the harder step. Microsoft’s method therefore represents a promising correction layer, not proof that long-range weather uncertainty has been eliminated. Its real value will depend on how well the open-source software performs across regions, forecast systems, and real warning operations.