China’s AI weather models are challenging the old forecasting order
Fengwu, Pangu and Fuxi are among a new generation of AI systems that can produce forecasts in a fraction of the time required by conventional models, though important limitations remain.
As meteorologists tracked Typhoon Dolphin’s path toward China in recent days, a new generation of artificial intelligence weather models worked alongside traditional forecasting systems, highlighting China’s emergence as a leading player in the race to improve weather prediction.
Chinese-developed systems, including Fengwu, developed by the Shanghai AI Laboratory, Huawei’s Pangu and Fudan University’s Fuxi, are among a handful of AI forecasting models that researchers say can generate forecasts much faster than conventional systems while matching or surpassing them on some measures of accuracy.
For decades, weather prediction has relied on numerical models running on supercomputers that simulate atmospheric physics. AI models take a different approach, learning patterns from vast archives of historical weather observations and producing forecasts in a fraction of the time. The technology is increasingly being tested during typhoon season in East Asia, where even small improvements in track forecasts can help authorities better prepare for flooding, organize evacuations and manage potential transportation disruptions.
The rise of AI weather forecasting has created a new arena of competition among technology companies, research institutes and meteorological agencies, with China emerging as one of the field’s leading players.
Among the best-known AI forecasting systems globally are Google’s GraphCast and GenCast, Nvidia-backed FourCastNet and the European Centre for Medium-Range Weather Forecasts’ AI Forecasting System, known as AIFS.
Fengwu attracted attention after its developers reported that it outperformed GraphCast across roughly 80% of evaluated weather variables and extended skillful global medium-range forecasts beyond 10 days.
“With more extreme weather, people need information to make decisions — both local governments and the national government, but also the average person, farmers and fishermen,” said Sun Zhi, CTO of Techwind, the company responsible for Fengwu’s industrial applications. “So we want to help provide better information so people can make decisions.”
While AI systems are becoming an increasingly important complement to conventional forecasting because of their speed and lower computing costs, they are unlikely to fully replace traditional weather models in the near future.
According to Sun, AI models are already capable of predicting the paths of typhoons. Five days before Dolphin’s landfall, Fengwu predicted the time and location at which the storm would hit mainland China to within 30 minutes and 30 kilometers (19 miles), he said.
But AI models still lag conventional forecasting systems in predicting storm intensity, Sun said, and remain largely untested when it comes to major climate developments.
“If we predict a climate change event 18 months in advance, people won’t believe it,” Sun said. “They need to know it’s reliable. We need to do years of scientific research before people trust us when we say there will be an El Niño event or that changes in sea-surface temperatures will affect the breeding cycle of fish.”
As AI weather models become more sophisticated, the parallel use of AI and conventional forecasting systems is likely to continue for some time.
