AI weather forecasting reduces costs but still needs physical constraints
In a lecture published by the WeChat public account 格致论道讲坛, Xia Jiangjiang of the Institute of Atmospheric Physics at the Chinese Academy of Sciences said AI weather forecasting had developed faster than expected. He cited the July 2023 Nature paper introducing Huawei’s Pangu Weather model and said AI models could reduce infrastructure costs from the scale of supercomputers and shorten forecast runtimes to seconds or minutes in some cases. He also reported that his team’s Yantian global forecasting model ran a global forecast in about two to three hours on a laptop costing 5,000 yuan. Xia cautioned that AI models had surpassed traditional numerical forecasting for only some meteorological elements, while numerical forecasting remained the core operational system and was more stable and reliable overall. He said data-driven models could produce or remove water vapor without physical basis and argued for incorporating physical rules such as water-vapor conservation, through collaboration between meteorologists and AI researchers.
The lecture presents AI weather forecasting as a way to lower computational costs and reduce runtimes while explaining why physical constraints and continued use of traditional numerical forecasting remain important.

2020年的时候,我预测做成这件事情需要10年,结果只过了3年就有人做出来了。夏江江·中国科学院大气物理研究所格致校园第66期|2026年7月24日北京大家好,我是夏江江,来自中国科学院大气物理研究所。我演讲的主题是“当天气预报遇见AI”。题目里边有“天气预报”这四个字,那我们就首先了解一下传统的天气预报是怎......
本文来自微信公众号: 格致论道讲坛 ,作者:夏江江,原文标题:《本以为AI预报天气还要等10年,没想到失算的竟是我们 | 夏江江》
2020年的时候,
我预测做成这件事情需要10年,
结果只过了3年就有人做出来了。
#Pangu Weather#Yantian Global Weather Forecasting Model#Huawei#Google#NVIDIA