Microsoft researchers have released an open-source machine learning tool called Probabilistic Bias Correction that improves weather forecasts two to six weeks out by correcting systematic errors in existing models. The approach is designed to be lightweight, running on standard CPUs rather than the supercomputers or GPU clusters typically associated with weather AI.
The method addresses a common weakness: both physics-based and AI models lose accuracy beyond two weeks because approximations and data gaps create biases that compound over time. PBC learns to identify and correct those biases. In tests against a leading ECMWF physics model, it improved accuracy for 98% of precipitation and 92% of temperature targets, and it doubled the accuracy of ECMWF's new AI forecasting system.
For business and public-sector users, the significance is cost. Weather agencies in developing countries often lack the infrastructure for advanced forecasting. Because PBC is open-source and CPU-friendly, it can be adapted locally. The researchers are already testing it with Kenya's meteorological service ahead of an expected extreme rainy season. ECMWF, which awarded the research in its AI Weather Quest competition, cautioned that moving from research to operational use remains challenging.