A new, AI-enhanced simulator can match the accuracy of conventional weather forecasts and can also extrapolate how much the atmosphere has warmed with climate change. The simulator could lead to the development of weather and climate modeling tools that require a small fraction of the computing power needed today, according to the Google-led research team behind it. The research was published Monday in Nature.

By using a hybrid approach combining standard physics-driven models with a machine-learning tool, the team avoided problems seen in experiments using only AI, said Stephan Hoyer, the Google researcher who leads the project. “We’ve really tried to try to pull apart the black box, instead of having just a pure AI model,” he said.

Weather and climate models are workhorses for everyone from local TV meteorologists to climate scientists investigating how much people may heat up the world. Today’s climate models are physics rendered as software, with the major parts of the Earth system set in action together: atmosphere, ocean, land and ice.

These models can capture large-scale climate and weather systems with more confidence than localized phenomena. Clouds, rainfall and tornadoes occur at such small scales that general equations can’t describe them. Scientists typically estimate these from real-world data, and program them into the models as “parameters.”

The experimental model, NeuralGCM, relies on existing general circulation models (GCMs) to simulate large-scale physics and uses a machine-learning approach, called a neural net, to estimate the smaller-scale features.

“Because of that, we’re able to build a model that is much more stable and gives much more reliable results when you run it out for longer periods of time into the future, even going out to years or decades,” Hoyer said.

R. Saravanan, a professor of atmospheric sciences at Texas A&M University who wasn’t involved in the research, called it “an important advance in atmospheric modeling and long-term weather prediction, but not necessarily a giant leap in climate prediction.”

The hybrid model has limitations. It works out rising temperatures only in the atmosphere, not in oceans, on land or on ice. Conventional models are able to simulate all the major elements of the Earth system. Nor does the new approach yet allow researchers to vary the level of greenhouse gases in the atmosphere, a central function of modern climate models. The researchers use sea-surface temperatures, not emissions, as their prompt for changing the atmosphere.

Saravanan said the work might prove particularly useful in predicting weather on sub-seasonal and seasonal scales. If the approach can be expanded to include the oceans, it might become useful to researchers studying El Niño and La Niña weather patterns, he said.

The researchers are developing a feature in NeuralGCM that generates year-ahead hurricane projections, which, if shown to be useful, could help people prepare for storms and build adaptation infrastructure, Hoyer said.

Machine-learning atmospheric models are much faster and demand less computing power than standard models. One of them, GraphCast, also developed by Alphabet Inc.’s Google, has 5,417 lines of code, compared with 376,578 lines for a US government model. Of Neural GCM, Hoyer said, “You can run it on a laptop.”

Still, machine learning isn’t a substitute for physics, one climate scientist cautioned about the new results. “There’s no path that gets you to the future of climate without the current climate models,” said Gavin Schmidt, director of NASA’s Goddard Institute for Space Studies.

Scientists usually estimate the amount of global heating likely to come from greenhouse gas pollution as a range, reflecting the chaotic nature of the climate. It’s the same in weather prediction, when meteorologists say there’s, for example, a 40% chance of rain. Physics-based models are able to hone in on that chaos and constrain how the world might respond to higher temperatures. But AI models, because they’re not directly computing the physics, have no way to capture the inherent, unavoidable fuzziness in prediction, said Schmidt.

The new model’s clearest advance may be over climate simulations that use only machine learning (ML).

“At first glance, the NeuralGCM sounds like a major advance in pure ML-based modeling,” Saravanan said. “Actually, it’s quite the opposite — the paper highlights the limitations of pure ML-based approaches.”

The project is part of Google’s larger AI push, and the authors note that their hybrid physics-AI approach could buoy other efforts in materials science, protein folding and engineering. AI is already having another effect at Google: Power-intensive AI computing pushed its greenhouse gas emissions up 48% in five years.

This article was written by Eric Roston from Bloomberg and was legally licensed through the DiveMarketplace by Industry Dive. Please direct all licensing questions to legal@industrydive.com.