Microsoft's PBC promises better 2–6 week forecasts, now open source

Microsoft's latest work tackles a long-standing constraint in weather forecasting: physics-based models run on supercomputers, and generative AI models rely on powerful GPUs, making both costly to operate. The company's approach is not to build another standalone model but to layer machine learning on top of existing ones to correct systematic bias. This matters because bias in weather models compounds over time, making forecasts less precise beyond two weeks.

The method, called Probabilistic Bias Correction (PBC), was developed with researchers from MIT, ECMWF, and other institutions. One cited example is correcting for climate change: a model trained on historical records might predict colder temperatures than the world is actually experiencing. Before publication, the ECMWF had already recognized the authors' work in its AI Weather Quest competition.

According to Microsoft Research New England senior principal researcher Lester Mackey, PBC can run just on CPUs, which could make it accessible to weather agencies with limited resources. The paper's authors, including Genevieve Flaspohler, now executive director of the nonprofit Rhiza Research, believe PBC's lightweight nature allows scientists without huge pools of resources to configure it for local needs. Rhiza Research is working with the Kenya Meteorological Service Authority, funded by the Gates Foundation, to test whether AI can improve early warnings ahead of an expected extreme rainy season during a record-breaking El Niño.