Microsoft opens tool for 2–6-week weather forecasts — The Verge
Microsoft has made its Probabilistic Bias Correction (PBC) machine-learning tool publicly available. It is designed to improve the accuracy of weather forecasts for periods from two to six weeks. The researchers published their findings in the journal Nature, The Verge reports.
PBC is not a standalone weather forecasting model. The method uses machine learning to identify and correct systematic errors in existing physical and AI models. According to the developers, this could help farmers and emergency services plan their actions more accurately in advance.
Correcting model errors
Physical forecasting models reproduce complex processes in Earth’s atmosphere, while artificial intelligence-based models look for patterns in historical data. Both approaches become less reliable when forecasting weather two weeks or more ahead. This is due to approximations needed to fill gaps in data, particularly in regions with too few weather stations or when modeling atmospheric phenomena that have not yet been fully studied.
Such errors can accumulate over time and reduce the accuracy of long-range forecasts. One example cited by the researchers is climate change: a model that relies heavily on historical indicators may be biased toward lower temperatures than those corresponding to the planet’s current warming.
More current news is available on the UA.News Telegram channel Telegram.
Testing results
In the study, PBC improved the accuracy of the European Centre for Medium-Range Weather Forecasts’ (ECMWF) leading physical model for 98% of precipitation target metrics and 92% of temperature target metrics. Within the analysis, the approach also doubled the accuracy of ECMWF’s new AI forecasting system.
Before the scientific paper was published, ECMWF had already recognized the PBC authors’ research in the AI Weather Quest competition. The project involved researchers from Microsoft, the Massachusetts Institute of Technology, ECMWF, and other institutions.
The developers note that the PBC tool can run on central processing units. One of the method’s authors, Genevieve Flaspohler, who leads the nonprofit organization Rhiza Research, is testing the approach in cooperation with the Kenya Meteorological Department. The project, funded by the Gates Foundation, is intended to determine whether the method will help improve early warnings for the public.
