So far, forecasting the weather has been a complex challenge that demands maximum performance from mainframe computers. Two research teams have now independently developed methods that use artificial intelligence to deliver predictions that are as precise as classic models – and with significantly less computational effort. One model forecasts the weather up to a week in advance, the other specializes in short-term forecasts of extreme precipitation. But extreme weather events in particular could push the AIs to their limits, experts warn.
Air pressure, temperature, wind speed and water vapour: These and numerous other physical factors have so far been considered when meteorologists use complex algorithms on large computers to calculate how the weather is likely to develop in the coming days. However, the predictions of these so-called numerical models take a long time and require a great deal of computing effort. As an alternative, various research teams are working on using artificial intelligence for weather forecasting. However, previous models were too imprecise for practical use.
Accurate at multiple speeds
Now two research teams have independently developed AI systems that can compete with classic models in terms of accuracy and far exceed them in terms of speed. The first model, Pangu-Weather, comes from a team led by Kaifeng Bi from Huawei Cloud in Shenzhen, China. "By training the models on 39 years of global weather data, Pangu-Weather forecasts better than the world's best numerical system and is much faster at the same time," reports the research team.
Unlike previous models, Pangu-Weather does not include physical parameters. "Instead of making predictions based on physical knowledge, the AI predicts weather patterns that are statistically plausible based on historical measurements," explain Imme Ebert-Uphoff and Kyle Hilburn of Colorado State University, who were not involved in the study, in a accompanying commentary, also published in the journal Nature. "The AI model provides predictions about 10,000 times faster than numerical models with the same spatial resolution and comparable accuracy." Because Pangu-Weather includes a 3D model, it also provides reliable values for different elevations.
Forecast of extreme precipitation
The second model, NowcastNet, was developed by a team led by Yuchen Zhang from Tsinghua University in Beijing and focuses on an area with which classic weather forecast models have previously had problems: the short-term forecast of extreme precipitation. "Extreme precipitation is a significant contributor to meteorological catastrophes, and there is a great need to mitigate their socioeconomic impacts through skillful forecasts with high resolution, long lead times, and local detail," the team writes. "Current methods are very error-prone: Physically-based numerical methods struggle to capture the chaotic dynamics involved in these events, and previous data-driven learning methods do not follow the laws of physics."
For NowcastNet, the research team combined physical equations with machine learning methods. "Based on radar observations from the USA and China, our model generates physically plausible precipitation forecasts with lead times of up to three hours," the researchers explain. The model was tested by 62 professional meteorologists from different parts of China. NowcastNet was superior to the previously leading methods in 71 percent of the cases.
chances and risks
From the point of view of Ebert-Uphoff and Hilburn, the new AI models offer great opportunities on the one hand. "These approaches are so promising that they could lead to a paradigm shift, with AI models completely replacing numerical approaches," they write. At the same time, however, they urge caution. Because the artificial intelligences are trained with historical data, they may not be sufficiently accurate if man-made climate change leads to an unprecedented accumulation of extreme weather events.
"Given the potential benefits and risks associated with AI models for weather forecasting, it is time for meteorologists to get involved to ensure that AI-based weather forecasting models are well suited for their tasks," said Ebert- Uphoff and Hilburn. "Moreover, meteorologists need to learn how to interpret their forecasts, because AI models behave differently than physical models, so understanding their forecasts requires special training."
Sources: Kaifeng Bi (Huawei Cloud, Shenzhen, China) et al., Nature, doi: 10.1038/s41586-023-06185-3; Yuchen Zhang (Tsinghua University, Beijing, China) et al., Nature, doi: 10.1038/s41586-023-06184-4