Weather forecasts usually stop being reliable after about two weeks. Microsoft researchers say they’ve found a cheap way to push useful forecasts out to six weeks. According to The Verge AI, the team published its method in the journal Nature and released it as open-source software. They didn’t build a new weather model. They built a tool that corrects the mistakes existing models already make.
The tool is called Probabilistic Bias Correction (PBC). What stands out is its size: it runs on ordinary CPUs. Supercomputers and racks of GPUs aren’t required.
Why forecasts fall apart after two weeks
There are two main kinds of weather model, and they work differently:
- Physics-based models try to simulate the Earth’s atmosphere directly. They need supercomputers.
- AI models look for patterns in historical weather data. They need powerful GPUs.
At longer ranges, both kinds hit the same problem. Each one fills gaps with approximations. Those gaps come from places with few weather stations and from parts of atmospheric physics scientists still don’t fully understand. Over time, those shortcuts turn into systematic bias, and the errors compound the further out a forecast goes.
Microsoft Research New England’s Lester Mackey gave The Verge a clear example. Fossil fuel emissions are warming the planet, but a model trained on older records can lean toward colder temperatures than the world actually has today. PBC uses machine learning to find biases like that and correct them.
The results
The researchers tested PBC on two forecasting systems from the European Centre for Medium-Range Weather Forecasts (ECMWF). One is ECMWF’s leading physics-based model and the other is its new AI system.
- ECMWF physics-based model, precipitation: better accuracy on 98% of targets
- ECMWF physics-based model, temperature: better accuracy on 92% of targets
- ECMWF AI forecasting system: accuracy doubled
The method also had outside recognition before the paper came out. ECMWF had already given the authors an award for the work in its AI Weather Quest competition. Researchers from Microsoft, MIT, ECMWF and other institutions co-wrote the paper.
Why the cheap part matters most
I think the real story is cost, not accuracy. Weather agencies in many developing countries can’t afford the compute that modern forecasting demands, so they keep missing out on new advances. A correction layer that runs on CPUs and sits on top of models those agencies already use is much easier to adopt.
“One of the things that’s really exciting to me about models like PBC is how lightweight they are, how easy to train,” co-author Genevieve Flaspohler told The Verge. She added that scientists “without huge pools of resources” can tune these models to work better for their own region.
That’s already being put to work. Flaspohler now runs the nonprofit Rhiza Research, which is working with the Kenya Meteorological Service Authority on Gates Foundation funding. They’re testing PBC and other methods to see whether AI can get earlier and more accurate warnings to people. The timing matters, because a record-breaking El Niño is expected to bring a more extreme rainy season.
What practitioners can take from this
- Correct before you rebuild. You don’t always need a new model to get better results. A small learned correction on top of a strong existing system can deliver big gains.
- Look for drift in your training data. The climate example applies far beyond weather. Any model trained on historical data can miss a world that has since changed.
- It’s open source. Weather agencies, agritech teams and climate-risk analysts can try PBC on their own forecasts today.
- Longer lead time means real planning. Farmers and emergency responders gain weeks to prepare, not days.
The caveats
ECMWF’s Matthew Chantry called the paper “a good example of new ideas being developed for the benefit and consideration of the community.” He also offered a fair warning: “turning ideas into reality [in] an operational process is a most challenging step.” Good results in a study aren’t the same as a system forecasters rely on every day. PBC also depends on the model underneath it. It can fix systematic bias, but it can’t create skill a base model never had.
Microsoft is entering a crowded field, with Google and Nvidia also building AI weather tools. Its approach is different, though. Instead of competing to build the biggest model, it makes the models agencies already have work better. If the Kenya pilot holds up this rainy season, expect other agencies to follow. Full details are in The Verge’s original report.