AI is learning to predict the weather

Using AI to predict weather has evolved over the past five years from an academic notion to operational tests at weather agencies in the U.S. and Europe, as well as at companies that provide intelligence to businesses.

In May, Microsoft released a forecast tool called Aurora that produces five-day global air- pollution predictions and 10-day weather forecasts 5,000 times faster than existing models run by the National Oceanic and Atmospheric Administration and the European Center for Medium-Range Weather Forecasts. Companies and labs around the country, including Villanova University, the University of Oklahoma and a California startup firm, are training new weather AIs.

Faster, more accurate forecasts are becoming ever more important. The world is warming, extreme weather has become more deadly, and storms are more costly. “We really need to get the forecasts of the weather really accurate," says Remi Lam, a research scientist at Google DeepMind, which introduced an AI-based weather model called GraphCast in November.

Beyond equations

For decades, meteorologists have derived weather forecasts using equations that describe the atmosphere, such as the relationship between air pressure and prevailing wind from one region to another, or how quickly temperatures change as cold fronts move through. They populate these equations with measurements of the atmosphere and ocean taken hourly by weather stations, high-altitude balloons, ocean buoys and satellites. The data is fed into supercomputers that produce what is known as numerical weather prediction.

The problem is that small errors in measuring the weather or in the calculations can lead to bigger forecast errors. What’s more, running complicated simulations of the Earth’s weather takes a lot of expensive computing time.

AI algorithms look for patterns in weather data, rather than solving equations as supercomputers do. The pattern-hunting algorithms are trained on decades of weather data to predict what will happen in the days ahead.

“All those AI tools do is recognize patterns," says Paris Perdikaris, a principal researcher on the Aurora project at Microsoft Research. “And they’re really good at doing that."

Researchers trained Aurora with a huge amount of historical weather data so it could make those predictions, about 16 times more data than the amount used to train the latest version of the AI-powered chatbot ChatGPT, according to Perdikaris.

Microsoft expects to make Aurora publicly available in coming months to allow more people, including researchers at weather-forecasting agencies, to give it a test drive.

“It is ultimately up to them to decide whether and when they will adopt AI models like Aurora into their operational forecasting workflow," Perdikaris says. “My personal estimation is that this will happen within the next two to five years."

Training on history

WindBorne Systems of Palo Alto, Calif., has developed its own AI forecasting model that uses data from its constellation of weather balloons. The balloons, launched from three continents, traverse oceans and circumnavigate the globe. The data is analyzed using similar AI techniques that power ChatGPT.

“You have a large neural network that you train on historical data, you input the current state of weather, it outputs a predictive future state, and nowhere inside do you ever program the laws of physics," says John Dean, co-founder and CEO of WindBorne.

NOAA has been testing WindBorne’s data, and a recent study found that a small amount of WindBorne balloon data improved the accuracy of NOAA’s storm ground track forecasts by up to 18%. The firm also recently was awarded contracts with the Navy and Air Force to further develop its AI forecasting model.

The Weather Company, an Atlanta-based forecasting and information-technology firm that owns the Weather Channel app and Weather Underground, recently signed a deal with chip maker Nvidia to develop a global AI weather forecast program that can predict extreme weather events such as tornadoes, hurricanes and thunderstorms.

The firm’s existing AI model gives a rough indication of an incoming storm’s strength, but not enough information about wind speed or rainfall needed to make decisions on the ground, says Thomas Hamill, head of innovation at the Weather Company.

“If you’re managing the airport of Dallas-Fort Worth, and you are seeing a blob of precipitation, that doesn’t tell you a lot of the detail that you really care about," Hamill says. “Is this going to be a squall line coming through, or is it an air mass of thunderstorms that come and go in a 20-minute period?"

Faster processing of data will make AI forecasts more accurate and more detailed, he says.

Next-generation AI forecasts will produce weather scenarios known as ensembles, Hamill says. Now made with numerical weather prediction, ensembles give meteorologists a range of future possibilities based on slight variations in initial conditions. “That would really help us, help our customers," Hamill said.

An earlier warning

Graduate students in Stephen Strader’s lab at Villanova University trained an AI program to identify the size and shape of storms—an important indicator of their strength and whether they might produce tornadoes or hail, for example. That task used to require students to review and classify pictures of storms one at a time, said Strader, an associate professor of geography and the environment.

“You train it again, train it again, and refine it," Strader said. “And it does really, really, really well."

Amy McGovern, professor of computer science and meteorology at the University of Oklahoma, and colleagues are developing AI forecasts to extend the warning time for tornado and hail warnings to an hour from the current 15 minutes.

“You certainly can’t move your farm, but you can get your livestock inside," says McGovern about the hail warning system. “And if you are a car dealer or an airplane company, you could move your stock under cover."

NOAA is also evaluating AI’s use to predict the track of tropical storms across the Atlantic, but the algorithms need to be more accurate before the agency can use them, according to Wallace Hogsett, science and operations officer at the National Hurricane Center in Miami.

“It’s really early in the game," he says.

Write to Eric Niiler at eric.niiler@wsj.com