While meteorologists were tracking the typhoon's path toward China, a new generation of AI models was spinning alongside conventional systems — not yet able to predict the intensity of storms.
The typhoon season in East Asia is being used as a life-size test bed for machine-learning-based weather models. As Typhoon Dolphin approached, China's Fengwu systems, developed by the Shanghai AI Laboratory, Huawei's Pangu and Fuxi from Fudan University ran in parallel with traditional forecasting chains. The latter have been based for decades on numerical models that simulate the physics of the atmosphere on supercomputers; AI models, on the other hand, learn patterns from huge archives of observations and produce a forecast in a fraction of the time.
The claimed results are not anecdotal. Fengwu came to prominence after its designers reported that it outperformed Google's GraphCast on about 80% of the variables evaluated, and extended the useful medium-term global forecast beyond 10 days. On Dolphin, five days before landing on mainland China, the model predicted the time to within 30 minutes and the location to about 20 miles, or about 30 kilometers, according to Sun Zhi, CTO of Techwind, the company in charge of Fengwu's industrial applications. In this field, China is asserting itself as one of the leading players, against Google's GraphCast and GenCastcast, Nvidia-backed FourCastNet and the European Center for Medium-Range Weather Forecasts' AIFS system.
The limits are assumed. These models lag behind conventional approaches to predicting the intensity of a storm, and have not been proven to work on major climate changes. "If we predict a climate change event eighteen months in advance, people won't believe us," observes Sun Zhi, who believes that years of research are needed before an announcement of El Niño or surface warming affecting the reproductive cycle of fish is taken seriously. The cohabitation of the two methods should therefore last.
Read the full article on the original source
Go to the source →