Weather forecasts past two weeks are unreliable, whether they simulate atmospheric physics or learn patterns from historical data. Microsoft researchers say they have found a cheaper way to extend that horizon, publishing the method in Nature and releasing the tool as open-source software.
Rather than build another standalone model, the team used machine learning to correct bias in models that already exist. They call it Probabilistic Bias Correction.
Fixing the gaps that compound
Both forecasting approaches rely on approximations to fill holes in the data, whether from regions with sparse weather stations or atmospheric behavior scientists still do not fully understand. Those workarounds introduce systematic bias, and the errors compound the further ahead a model looks.
PBC identifies that bias and adjusts for it. Senior principal researcher Lester Mackey offers a climate example: emissions are heating the planet, but a model trained on historical records leans toward temperatures colder than current reality.
The researchers found the method improved accuracy at two to six weeks out, a window that matters for farmers planning harvests and emergency responders staging resources ahead of extreme weather.
The efficiency claim is the practical draw. Physics-based forecasting runs on supercomputers and generative AI needs powerful GPUs, but Mackey says the new tool runs on ordinary CPUs. That puts better long-range forecasting within reach of agencies that cannot afford the usual infrastructure.
Microsoft joins Google, Nvidia, and others building AI tools for weather prediction, though its angle is augmentation rather than replacement, an approach that may travel further in underfunded meteorological services worldwide.
