Microsoft says machine learning keeps weather forecasts reliable up to six weeks out

What Microsoft announced

Microsoft researchers claim to have pushed dependable weather forecasts past the two-week barrier that has long constrained the field. According to The Verge, the team's findings were published in the journal Nature, and the software behind them has been released as open source.

The technique targets forecasts made two to six weeks in advance, a range where farmers and emergency responders currently have little to rely on. If the results hold up outside the study, longer-range planning decisions could rest on firmer ground.

Correcting existing models instead of building new ones

Rather than training another AI weather model from scratch, a route already taken by Google, Nvidia and others, Microsoft's approach attaches to models that already exist. Conventional simulations try to reproduce the physics of the atmosphere on supercomputers, while AI models mine historical data for patterns. Both depend on approximations where observations are sparse or the underlying science is incomplete, and those approximations harden into systematic bias. The errors compound the further ahead a forecast reaches, which is why skill drops off sharply beyond about two weeks.

The method, which Microsoft calls Probabilistic Bias Correction (PBC), uses machine learning to spot those systematic biases and adjust for them. Lester Mackey, senior principal researcher at Microsoft Research New England, offered one example to The Verge: a model drawing heavily on historical records can skew toward the cooler climate of the past, understating the warmer conditions the planet now experiences because of greenhouse gas emissions.

A notable practical detail is that PBC runs on ordinary CPUs, rather than the supercomputers behind physics simulations or the powerful GPUs behind generative AI. "I have a lot of hope both for the increased resource efficiency of weather forecasting and the extended reliability of forecasts, especially for extreme weather," Mackey told The Verge.

Measured against ECMWF systems

The researchers tested PBC against the European Centre for Medium-Range Weather Forecasts (ECMWF), a leading operator in the field. The corrections improved the accuracy of ECMWF's flagship physics-based model for 98 percent of precipitation targets and 92 percent of temperature targets, and doubled the accuracy of ECMWF's newer AI forecasting system in the study's analysis.

ECMWF had recognized the work before publication, awarding the authors in its AI Weather Quest competition. The paper's authors are affiliated with Microsoft, MIT, ECMWF and other institutions.

Matthew Chantry, ECMWF's strategic lead for machine learning, described the research in an emailed comment as a good example of new ideas developed for the benefit and consideration of the forecasting community. He added a caveat: turning ideas into an operational process is the most challenging step.

Built with smaller weather services in mind

The compute and infrastructure demands of modern forecasting have kept advanced tools out of reach for many weather agencies in developing economies. The paper's authors argue PBC changes that equation. "One of the things that's really exciting to me about models like PBC is how lightweight they are, how easy to train, that scientists without huge pools of resources can work on trying to configure [these] models and make them perform better in specific localities," said Genevieve Flaspohler, one of the authors.

Flaspohler developed PBC as an intern at Microsoft alongside Mackey and now serves as executive director of the nonprofit Rhiza Research. With funding from the Gates Foundation, Rhiza is working with the Kenya Meteorological Service Authority to test AI methods, including PBC, for delivering more accurate early warnings ahead of a rainy season expected to be unusually severe during a record-breaking El Niño.

Why it matters

Most consequential planning, from planting schedules to disaster preparation and water management, happens on timescales of weeks, not days. Sub-seasonal forecasting has been a stubborn limit for both physics-based and AI approaches, and a correction layer that improves any existing model without exotic hardware could shift what agencies and the public can expect from a forecast. Because the tool is open source and CPU-friendly, the gains are not confined to institutions with supercomputing budgets. The open question, as ECMWF's Chantry noted, is whether the research translates into day-to-day operations, and results from deployments such as the Kenya pilot will be the real test.