When Hurricane Beryl was rushing across the Atlantic basin in July, the weather forecasting tool made by Google Deepmind, the tech company’s artificial intelligence unit, saw something other models missed. Deepmind’s AI-driven program, called GraphCast, forecast the storm would take a sharp turn away from southern Mexico to southern Texas nearly a week earlier than conventional forecasts did – and it was right.
The dramatic prediction shows the promise of new AI-driven weather models amid a destructive hurricane season that also featured Helene and Milton. With the season winding down, meteorological agencies and tech companies are taking a look at how these new models stacked up against traditional ones. Early returns suggest AI is capable of eerily accurate forecast tracks even as models still need to improve their skill with other metrics.
Scientists have made incredible progress using models that rely on physics to make storm predictions. Three-day forecasts for hurricane tracks were off by an average of 520 miles in 1970. Today, it’s a tenth of that. Four- and five-day projections didn’t even begin until after 2000 and have also seen dramatic progress over the past two decades. But physics-driven models’ phenomenal rate of improvement is slowing down, just as climate change is speeding up.
“It’s becoming increasingly difficult to make advances in that field,” said Rémi Lam, a research scientist at Google DeepMind.
The Beryl projection is only one recent highlight of how AI is pushing the boundaries of hurricane forecasting. An analysis presented this week at a hurricane forecast improvement conference in Miami looked at how well GraphCast did from 2021 to 2024. It beat out the conventional models in both the Atlantic and Pacific hurricane basins over the first five days of a storm, said Ferran Alet, a research scientist at Google DeepMind. Its forecasts hit the mark 12 hours faster than the US Global Weather Forecast System.
Yet another AI-driven model from the European Centre for Medium-Range Weather Forecasts (ECMWF) projected that Hurricane Francine would hit Louisiana 10 days beforehand, well in advance of most other models.
While they’re good at charting where storms will go, the models are somewhat blind to other critical elements, particularly intensity. GraphCast is trained to resolve conflicting predictions by averaging them out, which tends to under-predict wind speed. AIs are discouraged from making mistakes, said Ryan Keisler, who authored a 2022 paper credited with kickstarting recent research advances, so they commonly forecast lower wind intensity estimates rather than return higher, rarer and potentially wrong results.
But the benefits of AI are becoming increasingly clear, even beyond how accurate their track forecasts are.
GraphCast can generate a 10-day weather prediction in under a minute on a machine Lam said is “bigger than a laptop, but you can hold it in your hands.” In comparison, it can take standard models about an hour to do the same on a supercomputer. GraphCast has also performed with more than 90% greater accuracy than the standard physics-based model put together by the ECMWF, considered the gold standard.
DeepMind’s approach has proven so compelling, ECMWF has borrowed it to produce its own AI model, which already edges out the group’s conventional one. Advanced-computing giant NVIDIA’s weather AI is called FourCastNet that belongs to a larger suite of tools available on the company’s Earth-2 platform, which it bills as a “climate digital twin” of the third planet from the Sun. Huawei’s Pangu-Weather project also outperforms the European medium-range standard on the variables tested, and Fudan University’s FuXi beats it on 15-day forecasts.
Leading weather models, like the ECMWF’s and the US National Oceanic and Atmospheric Administration’s, are enormous catalogs of physical equations governing the atmosphere. In comparison, AI models don’t know physics. Instead, they start with a neural network or other learning platform and are fed training data – lots of it for every point on the globe. Many rely on ECMWF’s historical weather simulation database and also observations.
ECMWF expects to use its model in general weather forecasting operations next year. But there will still be physics-based guardrails.
“We’re not looking at turning off the physical system,” said Matthew Chantry, ECMWF’s machine learning coordinator. “We’re looking to run two systems that, for the moment, have their own strengths and weaknesses.”
Ultimately, researchers want their AI products to help meteorologists issue forecasts that can better protect lives and property. Attendees at COP29 climate talks in Baku, Azerbaijan, are also discussing ways that AI can help deal with climate change, marking a shift from the usual discussions around AI’s emissions.
Amy McGovern, a weather and AI expert at the University of Oklahoma, is part of a drive to ensure that products can be tested against common benchmarks – a key trust-building exercise. “Extreme Weather Bench” is the name of a system she’s developing to do just that as part of startup Brightband, where she is the lead AI and meteorology strategist and Keisler works as chief scientist. When the platform launches early next year, she expects it to provide a standardized and open-source library of past extreme weather events, where people can test how AI models perform against each other.
“We’re trying to really make sure that what we’re generating is trustworthy,” McGovern said, and discovering “what it means to be trustworthy.”