Google DeepMind has developed a machine learning algorithm that it says can predict the weather more accurately than current forecasting methods using supercomputers.
Google’s model, called GraphCast, generated a more accurate 10-day forecast than the High Resolution Forecast System (HRES) run by the European Center for Medium-Range Weather Forecasts (ECMWF), forecasting in minutes instead than in hours. Google DeepMind calls HRES the current benchmark weather simulation system.
GraphCast, which can run on a desktop computer, outperformed ECMWF on more than 99% of weather variables in 90% of the 1,300 regions tested, according to results published Nov. 14 in the journal. Science.
But researchers say it’s not perfect because the results are generated in a black box – meaning the AI can’t explain how it found a pattern or show how it works – and it should be used to complement rather than replace established tools.
Related: Is climate change making the weather worse?
Today, forecasting relies on integrating data into complex physical models and using supercomputers to run simulations. The accuracy of these predictions relies on granular details within the models, and they are energy-intensive and expensive to execute.
But machine learning weather models can run more cheaply because they require less computing power and operate faster. For the new AI model, researchers trained GraphCast on 38 years of global weather records through 2017. The algorithm established patterns between variables such as air pressure, temperature, wind and humidity that even researchers have not understood.
After this training, the model extrapolated forecasts from global weather estimates made in 2018 to make 10-day forecasts in less than a minute. By running GraphCast alongside ECMWF’s high-resolution forecasts, which use more conventional physical models to make predictions, the scientists found that GraphCast gave more accurate predictions on more than 90% of the 12,000 data points used.
GraphCast can also predict extreme weather events, such as heat waves, cold waves and tropical storms, and when Earth’s upper atmospheric layers have been removed leaving only the lowest level of the atmosphere , the troposphere, where weather events impacting humans are predominant, accuracy increased to over 99%.
“In September, a live version of our publicly available GraphCast model, deployed on the ECMWF website, accurately predicted approximately nine days in advance that Hurricane Lee would make landfall in Nova Scotia,” wrote Rémi Lam, research engineer at DeepMind, in a press release. statement. “In contrast, traditional forecasts had greater variability in where and when landfall would occur, and only predicted Nova Scotia about six days in advance.”
Despite the model’s impressive performance, scientists don’t see it supplanting currently used tools anytime soon. Regular forecasts are always necessary to verify and define the starting data for any prediction, and because machine learning algorithms produce results that they cannot explain, they can be prone to errors or “hallucinations.” .
Instead, AI models could complement other forecasting methods and generate faster forecasts, the researchers said. They can also help scientists observe changes in climate patterns over time and get a clearer picture of the bigger picture.
“The pioneering use of AI in weather forecasting will benefit billions of people in their daily lives. But our broader research is not just about anticipating the weather, it is also about understanding the broader patterns of our climate “Lam wrote. “By developing new tools and accelerating research, we hope AI can enable the global community to tackle our greatest environmental challenges.”