A.I. Is Quietly Powering a Revolution in Weather Prediction

A.I. is spurring a revolution in the world of meteorology, allowing forecasts that once required huge teams of experts and massive supercomputers to be made on a laptop. Faster and cheaper forecast production means that poorer countries without their own government weather offices or supercomputer access should be able to produce their own custom forecasts. And researchers can delve into huge batches of forecasts to unpack how and why their results are so good: The A.I. may be seeing things that experts have missed, for example, revealing aspects of the physics of weather that aren’t coded into conventional models. 

But this approach also has potential drawbacks. Because it’s usually unclear how A.I. makes its decisions, its results can be hard to trust, meteorologists say. And as climate change increasingly pushes weather into previously uncharted territory, they add, the A.I. might fail. 

For now, meteorologists continue to need both kinds of forecasts: A.I. depends upon the comprehensive data generated by the slower supercomputers. But there are big questions about whether that process might change, and whether A.I. will eventually be all we need.

Though people tend to make fun of the reliability of weather forecasts, they have gotten very good over the past 50 years. In general, says Dueben, they’re improving about a day per decade: In other words, today’s six-day forecast is as accurate as the five-day forecast was 10 years ago.

In the 1800s, forecasters would chart weather systems and consider former weather behavior to try to predict the future. By the turn of the 20th century, researchers audaciously suggested it might be better to instead use physics-based models of fluid flow and thermodynamics to mathematically determine what air patterns were going to do. But it wasn’t until the invention of better computers in the 1960s that “numerical weather prediction” started to gain traction. In 1975 the ECMWF was set up, in the U.K., as an independent intergovernmental organization, pooling European resources to run a Cray computer and crunch out physics-based global forecasts for up to a few weeks in advance. Today it has centers in the U.K., Germany, and Italy. 

There are now more than 200 billion weather observations made each day around the world, by satellite or on-the-ground instruments. Computers have become far more powerful, and more detailed understandings of fluid flow have been worked into their algorithms. Today, national weather services take coarse global model outputs with pixels on the order of tens of kilometers and use them to help make finer-resolution, local, and shorter-term forecasts down to one kilometer or so. The U.K.’s Met Office, which uses its own supercomputers to make its own global models, is experimenting with regional models that get down to the 100-meter scale.

The resulting improvements have been called a “quiet revolution” in science, less noticeable than any one sudden breakthrough discovery but just as influential for society. Now, thanks to A.I., things are changing again. “This is the second revolution,” says Charlton-Perez.

Machine-learning A.I. systems, which have taken off since around 2010, are incredibly good at finding patterns in large amounts of data. In a sense, this means A.I. has brought weather forecasting back to its roots, says Charlton-Perez, reverting to the idea that the future can be predicted by looking at patterns of the past. 

Starting in 2022, several major technology companies and academics released A.I.-based weather forecasting systems: Notably, Google Deepmind released GraphCast; CalTech researchers published a system called FourCast; and the Chinese company Huawei developed Pangu-Weather. More have followed since. None are yet operationally producing daily forecasts (that is, for use by others). But plenty of researchers have compared these experimental systems with gold-standard numerical weather predictions and have discovered they are, in general, just as good. One September 2024 study looked at five stand-out A.I.-based weather forecasting systems that are “comparable, and in some cases, superior” to the ECMWF’s numerical prediction system — while being orders of magnitude more computationally efficient.

While the ECMWF’s A.I. forecasting system just went operational this February, the center has been testing it for years. “We really haven’t seen in the last two years where the machine learning model was catastrophically wrong,” says Dueben. “We are happy with the results,” he says, though they are three times coarser in resolution than the non-A.I. forecasts, which limits their accuracy for small-scale weather phenomena.

Of particular interest is whether these systems are still accurate when tackling extreme weather events. In general, A.I. systems are very good when they are operating within the bounds of their training data, but if you throw something entirely different at them, they can go off the rails. Yet predicting extreme weather events clearly is hugely important. 

Charlton-Perez decided to track these systems’ performance on Storm Ciarán, a “bomb” cyclone that shut down power in Northern Europe on Halloween 2023. He and his colleagues at the University of Reading chose this storm as it was timely (they wrote up the paper within a month of the storm) and it was outside of the A.I.’s experience, both in terms of the timeframe the A.I. had trained on and the training data’s scope: It was the strongest tornado reported in the British Isles since 1954. Again, the A.I. systems did surprisingly well. “We were really impressed,” says Charlton-Perez. It forecast the storm in about the right place, with about the right central pressure and strength. “Where we found they didn’t quite get the right responses is in the detailed structure and strength of the winds near to the ground,” he says.

Since then, even better A.I. systems have been developed. Google, for example, in late 2024 released the first ensemble A.I. weather forecasting system, GenCast. Importantly, this produces a collection of 50 or more different A.I. forecasts in order to get a probabilistic sense of how likely an event will be. Again, it beats out the ECMWF’s physics-based ensemble system for accuracy. The center is also working on its own A.I.-based ensemble forecasts, but this isn’t operational yet.

All these A.I. weather-forecasting systems are trained on a key dataset called ERA5, a massive “reanalysis” dataset made by the ECMWF by taking observational data and mashing it together with earlier numerical prediction forecasts. This approach helps to create a smooth and gapless dataset with quantities like temperature and pressure supplied for every grid point on Earth. 

This means that even though the A.I. forecasts are really good, the labor-intensive numerical models are still needed — to feed into the AI. This is one reason why the ECMWF now releases both products. 

But that could eventually change. A groundbreaking A.I. model called Aardvark, created by researchers at the University of Cambridge and colleagues, takes a different approach: It uses real observational weather data alone to spit out forecasts. The ECMWF is also working on its own product, which similarly ditches the model input and uses only observational data. For now, though, they are even coarser in resolution than the global models. “It’s really impressive, but there’s a lot of work left to be done there,” says Steven Ramsdale, one of the chief meteorologists at the U.K.’s Met Office in Exeter.

If such systems are improved, they could possibly unseat physics-based forecasts. “It’s a million-dollar question,” says Dueben. “There are some people who would say we definitely will run a physical model in 50 years. And others will say, ‘Why would you?’”

Vital to all these futures is continued access to basic observational weather data all around the world. In February, the Trump administration fired more than 800 employees at the National Oceanic and Atmospheric Administration, which operates the National Weather Service. This has left some, including NOAA officials, worried that the agency’s national data collection and forecasts will suffer as a result. 

When forecasts are built up from basic physics, it’s easy to understand their logic and have faith in them. With statistical A.I. forecasts, which lack transparent underpinnings, that’s no longer true, says Charlton-Perez. The solution to this problem is the continued testing of these models over time, experts say, while also developing techniques called “explainable A.I.”’ (X.A.I.). These are computer science methods for breaking open the black box of A.I. and reverse engineering what’s going on inside. “Explainability is a huge thing,” says Ramsdale. 

Charlton-Perez says he is working with the Met Office on X.A.I. projects that are a kind of experimental brain surgery for A.I. “If I take my scalpel and remove this part of the prior weather conditions, is that the key trigger that led to the storm being destructive? If you ping a particular neuron, what does it do?” Such experiments might reveal novel aspects of physics that are missing from the numerical models, he says. Ramdsale agrees. “We haven’t put in our understanding of physics; [the A.I. systems have] learned their own, effectively,” he says. That means they might “find new relationships that we haven’t seen before, because they’ve had to learn something from scratch.”

So how good will weather forecasts ultimately get? This is hard to put a number on. “I can’t tell you where it’s going to go,” says Charlton-Perez, but “[w]e’re never going to have perfect weather forecasts because it’s a chaotic system.” 

Overall, A.I. is just one more tool that will help make the forecast data ever better and easier to produce, opening the door for nations that can’t afford their own supercomputers. But A.I. is unlikely to replace physical models in the near future, says Ramsdale and, he says, it will never replace human expertise and accountability: “We still need people to turn the data into a usable piece of advice.”

Correction, April 17, 2025: An earlier version of this article incorrectly stated that Storm Ciarán occurred in 2024. It occurred in 2023.