【Science Essay: I See AI】

The temperature fluctuated a lot this spring, and just as someone said "one day into summer", the cold snap came again. Is booming AI useful for weather forecasting?

For a long time, weather forecasting has relied mainly on traditional numerical weather prediction models. With the breakthrough of AI technology, the meteorological field has begun to apply AI technology to improve the level of weather forecasting and climate prediction. In 2023, Science magazine rated "the development of AI-assisted weather forecasting" as one of the top ten scientific advances in the world, and in the same year, the application of China's leading AI large model in refined weather forecasting was rated as the first of the top ten scientific advances in China. These milestones not only demonstrate the potential of the convergence of atmospheric science and AI, but also create new opportunities for extreme weather response and disaster mitigation.

It should be said that the global meteorological field is experiencing a new pattern of AI-driven parallel to traditional models. Thanks to the excellent quality, large scale and continuous and steady growth of atmospheric science data, weather forecasting has regarded AI technology as a new quality of productivity. Since 2022, NVIDIA, Google, Microsoft and other AI leaders have successively launched meteorological AI models such as FourCastNet, GraphCast, NeuralGCM, and GenCast, and published results in top journals such as Nature and Science, promoting the rapid development of AI technology in the field of meteorology.

China has ranked first in the world in the application of AI technology to weather forecasting. In 60, the China Meteorological Administration launched the meteorological artificial intelligence science and technology innovation project, established the "Xiong'an Meteorological Artificial Intelligence Innovation Research Institute", and jointly released the three AI systems of "Fengqing", "Fenglei" and "Fengshun" with Tsinghua University, Fudan University and Shanghai Institute of Scientific Intelligence, realizing the seamless intelligent weather forecast that is autonomous and controllable in the next 0~0 days for the first time, marking the remarkable achievements of China's AI technology in the integration of production, education, research and application in the field of weather forecasting.

  面向極端強降水天氣，“風雷”系統首創物理引導與數據驅動混合架構，能夠提供1公里解析度，未來3小時短臨強降水預報，該系統在國家氣象中心強天氣預報中心應用。“風清”大模型是融合AI技術與氣象機理的全球中短期預報系統，其性能位於全球AI天氣預報大模型第一梯隊，“風清”系統以3分鐘生成未來15天全球預報的速度，將核心要素有效預報時效提升至10.5天，在颱風路徑、暴雨、高溫、寒潮等災害性天氣預報上展現出顯著優勢。

In the same year, the China Meteorological Administration organized and carried out the Artificial Intelligence Weather Forecasting Large Model Demonstration Program (hereinafter referred to as the "Forecast Demonstration Program", AIM-FDP), which widely solicited large weather forecasting models from the whole society to realize the connection of the whole chain of model-data product-evaluation, aiming to build an authoritative weather forecasting large model inspection and evaluation platform. Among them, the "Fengqing" large model ranks among the top in the forecast demonstration plan.

In disaster prevention and mitigation, AI technology is being transformed into a powerful tool to protect the safety of life and property. The accurate prediction of Typhoon No. 4 "Gemei" in 0 years has become a typical case - the "Fengqing" large model locks the typhoon's landfall path 0 days in advance, and accurately predicts the arc-shaped rainstorm belt across the three major regions of South China, East China, and Northeast China. In 0 years, there were 0 large-scale precipitation processes in China, of which 0 times the intensity reached the strong level, and 0 times reached the extremely strong level, and one of the extremely strong level rainstorm processes was brought by Typhoon "Gemei". In the cross-regional heavy rainfall caused by Typhoon Gemei, the Fengqing model not only accurately captures the core rainfall areas such as Guangdong and Hunan, but also accurately locates the peripheral rainstorm centers such as Shandong and Liaoning, so as to win a key decision-making window for typhoon disaster prevention and mitigation.

The rapid development of artificial intelligence provides new opportunities to break through the bottleneck of existing weather forecasting, but cutting-edge basic technologies still face challenges.

At present, AI meteorological models face dual challenges: first, the dilemma of physical interpretability, the existing "black box" architecture is difficult to reveal the internal mechanism of the atmospheric system, and the sensitivity of the model to parameter adjustment may lead to forecast bias. This is due to the fact that the existing AI model framework has not fully adapted to the spatiotemporal complexity characteristics of atmospheric science, and it is difficult to accurately predict the physical behavior of atmospheric systems. The second is the natural limitation of data-driven, extreme weather events themselves have the characteristics of low probability, and the scarcity of samples directly leads to the insufficient ability of the model to capture extreme and other catastrophic weather.

In order to further promote the development of AI applications in the field of atmospheric science, it is necessary to strengthen the construction of computing power support capacity based on the construction of special training datasets, and strive to build an integrated AI support platform from the observation end to the forecast end. Relying on intelligent computing resources, a new forecasting system can not only analyze the physical laws of the atmosphere, but also adapt to complex spatio-temporal characteristics. This paradigm of technological convergence may lead to new breakthroughs in atmospheric science theories and open up new cognitive paths for accurate weather forecasting, especially severe weather.

  (Authors: Yang Xuan and Cao Yong, respectively, are senior engineers of the Weather Forecasting Technology R&D Office of the National Meteorological Center; Director)