Researchers from Google have built a new weather forecasting model that combines machine learning with more conventional techniques, yielding accurate predictions. The model, called NeuralGCM and described in an article in Nature today, bridges a gap that has grown between weather forecasters in recent years.
While new machine learning techniques that predict the weather by learning from past data are extremely fast and efficient, they have problems with long-term predictions. General circulation models, on the other hand, which have dominated weather forecasting for the past 50 years, use complex equations to model changes in the atmosphere and give accurate forecasts, but they are extremely slow and expensive to run.
So experts are divided on which tool will be more reliable in the future Here comes the new model from Google that attempts to combine both.
“It's not a physics vs. artificial intelligence kind of thing. It's natural and artificial intelligence combined," says Stephan Hoyer, AI researcher at Google Research and co-author of the study.
The system still uses a conventional model to process some of the large atmospheric changes needed to make a forecast. It then incorporates artificial intelligence, which tends to do well where these larger models fail — typically for predictions at scales smaller than about 25 kilometers, such as those involving cloud formations or regional microclimates (San Francisco fog, for example).
"That's where we add artificial intelligence very selectively to correct errors that accumulate at small scales," Hoyer says.
The result, the researchers say, is a model that can produce quality predictions faster and with less computing power. They say NeuralGCM is as accurate as one- to 15-day forecasts from the European Center for Medium-Range Weather Forecasts (ECMWF).
But the real promise of such technology isn't the best weather forecasts for your area, says Aaron Hill, an assistant professor in the School of Meteorology at the University of Oklahoma, who was not involved in this research.
Instead, it is on larger-scale climate events that are prohibitively expensive to model with conventional techniques.
The capabilities could range from forecasting tropical cyclones with alerts and modeling more complex climate changes that are years away.
"It's very computationally intensive to simulate the globe over and over again or for long periods of time," Hill says. This means that the best climate models are hampered by the high cost of computing power, which is a real barrier to research."
The researchers said NeuralGCM will be open source and run with fewer than 5.500 lines of code, compared to the nearly 377.000 lines required for the model by the National Oceanic and Atmospheric Administration (NOAA). ).
Kochkov, D., Yuval, J., Langmore, I. et al. Neural general circulation models for weather and climate. Nature (2024)
https://doi.org/10.1038/s41586-024-07744-y
