Microsoft open-sources Probabilistic Bias Correction, improving weather forecasts on CPU alone

Microsoft Research, with MIT and ECMWF, published Probabilistic Bias Correction (PBC) in Nature. Rather than building a separate forecasting model, it uses machine learning to identify and correct systematic biases in existing models—for example, stopping a model from underestimating current warming because its historical record ran cold. In tests, PBC improved 98% of precipitation and 92% of temperature targets for ECMWF's mainstream physical model, and doubled the accuracy of ECMWF's new AI forecasting system. The tool is open-source and trains and runs on CPU alone. The impact is most direct for underfunded weather agencies: forecasting offices in less wealthy regions that previously couldn't adopt AI due to compute barriers can now fine-tune models for specific regions on local hardware. Genevieve Flaspohler, a paper author at nonprofit Rhiza Research, is working with the Kenya Meteorological Department to test PBC for early warnings ahead of a projected more extreme El Niño rainy season. Matthew Chantry, ECMWF's machine learning strategy lead, says the hardest step remains moving such new ideas into real operational use.