As the planet warms up and oceans rise, extreme weather events are becoming the norm. Increasingly severe hurricanes bring wind damage and flooding when they make landfall. And just this week the world dealt with the three hottest days ever recorded.
Getting notified in time to prepare for a catastrophic hurricane or heat wave—like the recent scorcher in the southern and midwestern US, where daily temperatures soared up to 112 degrees F—could be the difference between life and death. The problem is that predicting the weather, even day-to-day events, can still be a gamble. AI can help.
A pair of studies published July 5 in the journal Nature described the usefulness of two AI models that could improve weather forecasting. The first AI-based system is called Pangu-Weather, and it was capable of predicting global weather a week in advance. The second, NowcastNet, creates accurate predictions for rainfall up to six hours ahead, which would allow meteorologists to better study weather patterns in real-time.
Pangu-Weather and other methods demonstrate AI’s potential for extreme weather warnings, especially for less developed countries, explains Lingxi Xie, a senior researcher at Huawei Cloud in China and a coauthor for one of the studies.”
A majority of countries use numerical weather prediction models, which use mathematical equations to create computer simulations of the atmosphere and oceans. When you look at AccuWeather or the weather app on your phone, data from numerical weather predictions is used to predict future weather. Russ Schumacher, a climatologist at Colorado State University who was not involved in both studies, hails these forecasting tools as a major scientific success story, decades in the making. “They have enabled major advances in forecasts and forecasts continue to get more accurate as a result of more data, improvements to these models, and more advanced computers.”
But Xie notes that “AI offers advantages in numerical weather prediction being orders of magnitudes faster than conventional, simulation-based models.” The numerical models often do not have the capacity to predict extreme weather hazards such as tornadoes or hail. What’s more, unlike AI systems, it takes a lot of computational power and hours to produce a single simulation.
[Related: Strong storms and strange weather patterns sweep the US]
To train the Pangu-Weather model, Xie and his colleagues fed 39 years of global weather data to the system, preparing it to forecast temperature, pressure, and wind speed. When compared to the numerical weather prediction method, Pangu-Weather was 10,000 times faster, and was no less accurate. Pangu-Weather also contains a 3D model, unlike past AI forecasting systems, that allows it to record atmospheric states at different pressure levels to further increase its accuracy.
Pangu-Weather can predict weather patterns five to seven days in advance. However, the AI model cannot forecast precipitation—which it would need to do to predict tornadoes and other extreme events. The second Nature study fills this gap with their model, NowcastNet.
NowcastNet, unlike Pangu-Weather, focuses on detailed, realistic descriptions of extreme rainfall patterns in local regions. NowcastNet uses radar observations from the US and China, as well as deep learning methods, to predict precipitation rates over a 1.6-million-square-mile region of the eastern and central US up to 3 hours in advance. Additionally, 62 meteorologists from China tested NowcastNet and ranked it first, out of four other leading weather forecasting methods, in reliably predicting heavy rain, which it did 71 percent of the time.
“All of these generative AI models are promising,” says Amy McGovern, the director of the National Science Foundation AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography, who was not affiliated with either study. But these AI models will need some refinement before they can fully replace current weather forecasting systems.
The first concern McGovern raises is the lack of physics-based mathematical equations. Accounting for the physics of moisture, air, and heat moving through the atmosphere would generate more accurate predictions. “These papers are still a proof-of-concept,” she says, “and don’t use the laws of physics to predict extreme weather.” A second concern, and major downside to AI tech in general, is coded bias. An AI is only as good as the data it is fed. If it is trained with low-quality data or with information that is non-representative of a certain region, the AI forecaster could be less accurate in one region while still being helpful in another.
As AI continues to expand into different facets of life, from art to medicine, meteorology won’t be left out. While the current AI systems require further development, McGovern is making her own prediction of the future: “Give it 5 to 10 years, we are going to be amazed at what these models can do.”
