인공지능형 전지구 대기대순환 모형의 기술 개발 동향

Abstract

The rapid advancement of artificial intelligence technologies has led to the development of numerous atmospheric general circulation models (AGCMs) based on artificial neural networks within the past two to three years. NVIDIA’s FourCastNet introduced Operator Learning techniques, leveraging AFNO and SFNO, while Huawei Cloud’s Pangu-Weather presented a conceptually robust approach with 3DEST capable of processing three-dimensional data directly and implemented a hierarchical temporal aggregation technique through independent training at distinct temporal intervals. Google DeepMind’s GraphCast, employing the next-generation GNN, maintained spatial and vertical resolution parity with ERA5, while the Shanghai AI Laboratory’s FengWu model enhanced long-term prediction performance through variable-specific independent processing tailored to atmospheric variable characteristics. Microsoft’s ClimaX stands as an integrated model capable of both short- and long-term forecasting as well as climate prediction, with applications spanning global, regional, and high-resolution domains. FuXi, developed by Fudan University in China, achieved leading performance in extended-range forecasting, reaching up to 15 days, by independently handling prediction in three segments: 0~5, 5~10, and 10~15 days. Lastly, ECMWF’s AIFS incorporates an attention-based GNN by merging the characteristics of GraphCast’s GNN with Transformers. Remarkably, the predictive capabilities of these models either match or surpass those of ECMWF’s physics-based numerical model IFS, while significantly reducing resource consumption and computational time. This study presents a comparative analysis of these AI-driven AGCMs, examining their predictive capabilities, neural network architectures, and data characteristics. The analysis aims to illuminate development trajectories for AI-based prediction models in meteorological and climatological applications.