Rémi Lam had heard about San Francisco’s microclimates, but he didn’t realize how idiosyncratic they could be until he moved there this year. “The street I live in can be foggy, and it’s sunny two blocks down,” he says. Weather forecasts for the city can be wildly incorrect depending on the location. Even state-of-the-art weather forecasts can’t predict the city’s microclimates and how they will vary.
Lam has spent a lot of time thinking about weather and how to forecast it. As a researcher at Google DeepMind, the artificial intelligence (AI) firm based in London, Lam has been pioneering the use of machine learning to improve weather prediction. This field has made rapid advances in the past few years, and Lam and his colleagues have been at the forefront of these efforts.
They’re not alone. A number of groups are racing to develop AI-aided weather forecasts, including those at Microsoft, Nvidia, Huawei and the European Centre for Medium-Range Weather Forecasts (ECMWF) in Reading, UK. But for much of this year, the leading AI in terms of accuracy was a project called GraphCast, led by Lam (R. Lam et al. Science 382, 1416–1421; 2023).
“GraphCast raised the bar up in terms of skill of forecasting,” says Matthew Chantry, who leads research on AI-based weather prediction at the ECMWF.
Conventional weather forecasts are sophisticated programs that simulate the evolution of Earth’s atmosphere on the basis of known physics of how air, heat and water vapour move around the planet. GraphCast is an artificial neural network that is shaped like a grid covering the globe. Lam and his collaborators ‘trained’ it with data based on real atmospheric measurements, but without giving it any explicit knowledge of physical laws. Still, the AI forecasts were by many measures better than the conventional ones. “I was surprised it outperformed the physics-based forecasts so quickly — I thought it would take longer,” says Lam.
And although the training is computationally intensive, the forecasts take less than a minute on an advanced desktop computer — versus the hours of supercomputer running time for conventional ones.
Lam was born in a suburb of Paris in 1988, and trained as an aerospace engineer in France and the United States. He then realized that his understanding of the statistical modelling of fluid mechanics could be helpful to those using AI. DeepMind, with a culture focused on solving scientific problems, turned out to be an ideal fit. “There’s just no better place to do machine learning,” he says.
Maria Molina, an atmospheric scientist who applies AI to weather and climate modelling at the University of Maryland in College Park, gives credit to corporations such as Google for making their weather models available for anyone to download and run on their computers — at least so far. “At some point, when does that goodwill run out?” It could be worrying if those companies some day came to monopolize the best-available forecasts, she adds, especially when it comes to extreme weather events. “We should never expect the public to pay for access to life-saving information.”
