Microsoft open-sources PBC, an AI tool that fixes bias in weather models

Microsoft researchers on October 14, 2026 published a paper in the journal Nature describing Probabilistic Bias Correction (PBC), a machine learning method they say makes weather forecasts two to six weeks out more accurate. The company has released the tool as open-source software. Co-authors include researchers from MIT, the European Centre for Medium-Range Weather Forecasts (ECMWF) and other institutions.

Forecasts from both traditional physics-based systems and newer AI models tend to lose reliability past the two-week mark. Rather than build another standalone forecasting model, Microsoft's team set out to correct the errors already baked into existing ones, an approach it hopes will appeal to weather agencies with limited budgets.

Targeting systematic bias

All forecasting models lean on approximations to cover gaps, such as areas with few weather stations or atmospheric processes scientists do not yet fully understand. Those workarounds can harden into systematic bias, and the errors grow the further ahead a model looks. PBC uses machine learning to detect and adjust for that bias. Lester Mackey, a senior principal researcher at Microsoft Research New England, points to climate change as one example: a model trained on historical records may run colder than today's warming world.

According to the researchers, PBC improved a leading ECMWF physics-based model on 98 percent of precipitation targets and 92 percent of temperature targets. In the study's analysis it also doubled the accuracy of ECMWF's new AI forecasting system. ECMWF had already recognized the work in its AI Weather Quest competition before publication. Matthew Chantry, the centre's strategic lead for machine learning, welcomed the idea but added a caveat: "turning ideas into reality [in] an operational process is a most challenging step."

Lightweight enough for smaller agencies

Physics-based forecasting depends on supercomputers and generative AI on powerful GPUs, but Mackey says PBC can run on CPUs alone. Heavy compute demands have long kept many agencies in developing economies from adopting new forecasting techniques. Co-author Genevieve Flaspohler argues that because the model is light and easy to train, scientists without large resources can tune it for their own regions.

Flaspohler worked on PBC as a Microsoft intern and now leads the nonprofit Rhiza Research, which is partnering with the Kenya Meteorological Service Authority, with Gates Foundation funding, to test whether AI can improve early warnings. PBC is among the methods under evaluation ahead of a rainy season expected to be more extreme during a record-breaking El Niño. The work places Microsoft alongside Google, Nvidia and others building AI tools for weather prediction.