China is intensifying the use of artificial intelligence (AI) in weather forecasting in response to the increasing frequency of extreme weather events. During the passage of Typhoon Dolphin, Chinese AI systems were implemented alongside traditional meteorological methods.
Several technologies are competing in this sector, including Fengwu, created by the Shanghai Artificial Intelligence Laboratory; Pangu, developed by Huawei; and Fuxi, from Fudan University. Researchers indicate that these tools can generate forecasts with greater speed and, in some analyses, match or surpass the results of conventional methods.
This advancement aims to provide faster and more accurate forecasts, helping governments and the public prepare for floods, evacuations, and potential transport disruptions. However, despite the advances, AI still faces challenges, such as difficulty in determining storm intensity, and does not completely replace established meteorological models.
In the days leading up to Dolphin's arrival in Chinese territory, the new systems operated side-by-side with physics-based model tools. This combination demonstrates the growing role of AI in improving severe weather forecasting.
Classical meteorology relies on supercomputers to simulate atmospheric behavior through complex calculations. AI, on the other hand, operates under a different logic: its models are trained on vast historical datasets of climatic conditions, allowing for the generation of new projections in a significantly shorter period.
Speed is a primary benefit of this technology. In situations involving typhoons in Southeast Asia, small improvements in route prediction can give authorities more time to organize protective measures, plan evacuations, and manage impacts on infrastructure and displacement.
One notable system is Fengwu. Its creators reported that this model demonstrated superior performance to Google's GraphCast in about 80% of the analyzed meteorological variables. Furthermore, the system expanded the horizon of useful global forecasts beyond ten days.
The practical application of Fengwu was notable during the Dolphin event. According to the company responsible for its industrial applications, five days before the typhoon's arrival, the system predicted the time and location of impact on the Chinese mainland with an error margin of approximately 30 minutes and 30 kilometers.
Even with these successes, there is a crucial distinction between anticipating a storm's movement and calculating its strength. The Fengwu team clarifies that AI models still fall short of traditional forecasts in this second aspect.
Limitations extend to large-scale climatic phenomena. The confidence required to predict occurrences months in advance, such as El Niño episodes or changes in sea surface temperature affecting fish reproductive cycles, still requires years of scientific investigation, according to the system's team assessment.
Beyond Chinese projects, there is international competition involving technologies from various companies and institutions. Among the mentioned models are GraphCast and GenCast, both from Google; FourCastNet, supported by Nvidia; and the AI forecasting system of the European Centre for Medium-Range Weather Forecasts, known as AIFS.
The trend observed in this scenario is one of coexistence between artificial intelligence and traditional meteorological models. The agility and lower computational cost of AI systems increase their potential as a complementary tool, although current restrictions prevent them from fully assuming the forecasting function.
