Microsoft says it has found a cheaper way to make weather forecasts more accurate from two to six weeks ahead, a window where both traditional forecasting models and newer AI systems typically become less reliable. The findings were published in the journal Nature, and the tool is now open-source software. The researchers say the longer-range predictions could help farmers and emergency responders plan with more confidence.
Microsoft joins Google, Nvidia and other companies developing AI tools for weather forecasting. But rather than building a new standalone weather model, the Microsoft team uses machine learning to correct bias in models that already exist. The researchers hope that approach will be easier for resource-strapped weather agencies to adopt.
Physics-based and AI-based models struggle for a similar reason. Traditional models try to simulate the complex physics of Earth’s atmosphere, while AI models look for patterns in historical data. Both rely on approximations to fill gaps, whether where weather stations are sparse or where atmospheric phenomena are still poorly understood. Those approximations can become a systematic bias, and the errors compound the further ahead a forecast reaches.
Microsoft calls its method Probabilistic Bias Correction, or PBC. The tool uses machine learning to identify systematic bias in a model and make corrections. Lester Mackey, a senior principal researcher at Microsoft Research New England, offered one example: climate change. Greenhouse gas emissions from fossil fuels are heating the planet, but a weather model built on historical records may pull its predictions toward colder temperatures than the world is experiencing now.
Mackey said the tool can run just on CPUs, which matters because physics-based modeling is a resource-intensive undertaking that runs on supercomputers, and generative AI depends on powerful GPUs. “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.
In the study’s analysis, PBC improved the accuracy of a leading physics-based model from the European Centre for Medium-Range Weather Forecasts (ECMWF) for 98 percent of precipitation targets and 92 percent of temperature targets. The machine learning framework also doubled the accuracy of ECMWF’s new AI forecasting system. The ECMWF had already awarded the authors’ PBC research in its AI Weather Quest competition before the paper was published.
The paper, produced by researchers at Microsoft, MIT, ECMWF and other institutions, drew praise from the agency. “It is a good example of new ideas being developed for the benefit and consideration of the community,” said Matthew Chantry, ECMWF’s strategic lead for machine learning, in an emailed comment to The Verge.
Chantry was candid about the hurdles that remain. “As always, turning ideas into reality [in] an operational process is a most challenging step,” he added.
The computing power and infrastructure needed for weather forecasting have often placed technological advances out of reach for agencies in economically developing countries. The authors believe PBC could be different. “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 paper’s authors.
Flaspohler worked on PBC as an intern at Microsoft with Mackey. She is now executive director of the nonprofit Rhiza Research, which is working with the Kenya Meteorological Service Authority, with funding from the Gates Foundation, to see whether AI can help deliver more accurate early warnings to people. PBC is one of the methods the group is testing ahead of an expected more extreme rainy season during a record-breaking El Niño.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.