Typhoon Dolphin gave China’s AI weather models a high-stakes opportunity to show what they can do outside a research paper. As the storm approached China’s east coast, systems developed by Shanghai AI Laboratory, Huawei, and Fudan University operated alongside conventional forecasting tools. Reuters reported that Fengwu predicted Dolphin’s mainland landfall time and location within 30 minutes and 30 kilometers five days before the storm arrived.
The result does not mean artificial intelligence has replaced traditional meteorology. Reuters makes clear that AI models still lag conventional systems in forecasting storm intensity and remain untested for some major climate developments. What Dolphin demonstrated is more specific: AI weather systems can generate useful track forecasts much faster than physics-based supercomputer models, making them a practical complement when authorities need to prepare for flooding, evacuations, and transport disruption.
Fengwu Forecast Typhoon Dolphin Five Days Before Landfall
Fengwu is one of three prominent Chinese systems named by Reuters, alongside Huawei’s Pangu and Fudan University’s Fuxi. The models use historical weather observations to identify patterns and produce forecasts in a fraction of the time required by conventional numerical weather prediction. Traditional systems simulate atmospheric physics on supercomputers; AI models learn relationships from large weather archives.
Sun Zhi, chief technology officer of Techwind, the company responsible for Fengwu’s industrial applications, told Reuters that the model identified the timing and place of Dolphin’s mainland landfall to within 30 minutes and 30 kilometers. The projection was made five days before the event. That kind of lead time can matter for local governments, ports, airlines, farmers, fishing fleets, and residents deciding whether to move equipment or change plans.
Sun described the broader value in practical terms. “With more extreme weather, people need information to make decisions,” he said. He identified local governments, national authorities, ordinary residents, farmers, and fishers as groups that can benefit from earlier and more precise forecasts.
Fengwu’s technical research provides context for the Dolphin result. Its arXiv paper describes a data-driven global medium-range forecasting system trained on 39 years of ERA5 reanalysis data. The researchers reported that Fengwu outperformed GraphCast on 80% of 880 evaluated forecast variables in 2018 hindcasts and extended skillful global medium-range forecasting to 10.75 days for a key atmospheric measure.
Those paper results are not the same as an operational typhoon forecast, but they help explain why Chinese weather agencies and companies are testing AI systems in live settings. A model that can generate a forecast quickly can be rerun as new observations arrive, giving forecasters an additional view of a developing storm.
AI Models Speed Up Weather Prediction but Do Not Replace Physics
The speed difference is the central attraction. Conventional numerical weather prediction requires supercomputers to calculate how the atmosphere will evolve using physical equations. That process remains indispensable, especially for estimating storm intensity and producing trusted long-range outlooks. But it is computationally expensive and can take time.
AI models take a different route. They infer patterns from historical data and can return an output much faster. That speed is the operational promise that has made AI weather forecasting attractive to meteorological agencies, particularly when forecasters need to refresh and compare projections as new observations arrive.
The advantage is not merely lower computing cost. Faster output means forecasters can compare more scenarios, refresh projections as conditions change, and use AI as a second system for checking conventional-model results. Reuters reports that Chinese researchers say the country’s AI systems can match or exceed traditional forecasts on some measures of accuracy. The useful phrase is “some measures.” Track prediction, pressure, precipitation, wind, and intensity are different tasks, and a model can be strong on one while weak on another.
Sun acknowledged those limits. He told Reuters that AI models still trail conventional systems on intensity forecasting. He also warned that people will not trust an 18-month climate prediction without years of scientific validation. That caution is significant. AI can offer speed and strong performance in certain weather tasks, but it does not remove the need for physical models, scientific review, or experienced human forecasters.
China’s Weather AI Moves From Research to Public Preparedness
China has been building weather-model capability for several years. Fengwu, Pangu, and Fuxi are now part of a global group of AI forecasting systems that includes Google’s GraphCast and GenCast, Nvidia-backed FourCastNet, and the European Centre for Medium-Range Weather Forecasts’ AI Forecasting System. The competition is increasingly about operational reliability rather than simply benchmark results.
Chinese developers are also extending AI into related environmental forecasts. EastFrontier covered Langya 2.0, China’s AI marine model, which was designed to forecast typhoons, storm surges, and Arctic sea ice. That earlier work shows how AI models can target different domains, from global atmospheric conditions to ocean and coastal risk.
The Dolphin test adds an immediate public-use case. Reuters reported that the typhoon brought heavy rain to Shanghai on August 10 and that the storm highlighted China’s emergence as a leading player in AI weather forecasting. When severe weather threatens dense urban areas, a track forecast can inform emergency preparation long before landfall. The benefit is not that an algorithm issues orders. It is that it can give human decision-makers more time and more evidence.
This application also connects China’s AI ambitions to a socially visible outcome. Weather forecasting is not a corporate pilot or a consumer novelty. Errors can affect safety, transport, agriculture, and disaster response. That creates a high standard for validation, but it also gives AI models a clear public purpose if they perform reliably.
The lesson from Typhoon Dolphin is therefore measured. Fengwu’s 30-minute and 30-kilometer forecast result is a meaningful operational data point. It does not settle the contest between AI and conventional weather forecasting. Instead, it shows why the future is likely to be hybrid: faster AI forecasts used alongside physics-based models, with meteorologists judging where each system is most trustworthy.