海洋与大气领域大模型研究进展  

Recent developments in large model research in the ocean and atmospheric fields

作　　者：张子良 于华明 任姝彤[3,4] 张辰宇 ZHANG Ziliang;YU Huaming;REN Shutong;ZHANG Chenyu(College of Oceanic and Atmospheric Sciences,Ocean University of China,Qingdao 266100,China;Sanya Oceanographic Institution,Ocean University of China,Sanya 572025,China;National Marine Environmental Forecasting Center,Beijing 100081,China;Key Laboratory of Marine Hazards Forecasting,National Marine Environmental Forecasting Center,Ministry of Natural Resources,Beijing 100081,China;Qingdao Ekman Technology Co.,Ltd,Qingdao 266100,China)

机构地区：[1]中国海洋大学海洋与大气学院,山东青岛266100 [2]中国海洋大学三亚海洋研究院,海南三亚572025 [3]国家海洋环境预报中心,北京100081 [4]国家海洋环境预报中心自然资源部海洋灾害预报技术重点实验室,北京100081 [5]青岛埃克曼科技有限公司,山东青岛266100

出　　处：《海洋预报》2025年第2期109-122,共14页Marine Forecasts

基　　金：山东省高等学校青年创新团队计划(2024KJN017)。

摘　　要：详细探讨了近年来全球尺度上主要的海洋和大气深度学习预报模型,尤其是基于Transformer架构的自回归模型。按照模型的提出时间,重点介绍了FourCastNet、Pangu-Weather、ClimaX、GraphCast、FengWu、FuXi和AI-GOMS模型的基本架构、训练数据以及预报性能。通过比较这些模型,揭示其发展的主要趋势,即从简单的Transformer结构到引入AFNO、Swin Transformer、GNN等不同架构,模型在结构上不断创新;同时,模型在输入变量选择、训练策略和计算效率方面也不断优化,提升了预报的准确性和实用性。展望未来,随着图形处理器(GPU)技术的进步和跨学科合作的加强,预计模型将在提高预报技巧、延长预报时效和揭示新物理机制等方面取得进一步突破,同时模型也将进一步提高计算效率和降低资源需求。This review comprehensively examines the major deep learning models applied to global-scale ocean and atmospheric forecasting in recent years,with a particular emphasis on autoregressive models based on the Transformer architecture.We focus on analyzing models such as FourCastNet,Pangu-Weather,ClimaX,GraphCast,FengWu,FuXi,and AI-GOMS,chronologically detailing their fundamental architectures,training data,and forecasting performance.Through a comparative analysis of these models,this paper elucidates the primary trends in model development:from simple Transformer structures to the incorporation of diverse architectures such as AFNO,Swin Transformer,and GNN,demonstrating continuous structural innovation.Concurrently,models have been optimized in terms of input variable selection,training strategies,and computational efficiency to enhance forecast accuracy and practical applicability.Looking ahead,with advancements in GPU(Graphics Processing Unit)technology and strengthened interdisciplinary collaboration,we anticipate further breakthroughs in improving forecasting skills,extending prediction lead times,and uncovering new physical mechanisms,while simultaneously focusing on enhancing computational efficiency and reducing resource requirements.

关 键 词：深度学习 数值预报 TRANSFORMER 大模型 

分 类 号：P732.6[天文地球—海洋科学]