Microsoft claims a new machine-learning tool can forecast weather two to six weeks out, a range where existing models fall apart.
Company researchers published the findings today in Nature, describing a machine-learning method they say is more accurate and cheaper to run than physics-based forecasting or existing AI weather models at that two-to-six-week range. Both traditional numerical forecasting and the newer wave of AI weather models, including ones from Google and Nvidia, tend to lose reliability past the two-week mark. Microsoft says its approach holds up further out. The company has released the tool as open-source software rather than keeping it proprietary.
A forecast window that stretches past two weeks is the difference between reacting to a storm and planning around one. Microsoft points to farmers and emergency responders as the people who would benefit most from a few extra weeks of lead time. That it's open-source also means other researchers can test, or poke holes in, the claims rather than taking Microsoft's word for it.
Every company building an AI weather model claims the breakthrough; this time the claim comes with a peer-reviewed paper attached.