| 数值预报AI气象大模型国际发展动态研究 |
| |
| 引用本文： | 黄小猛,林岩銮,熊巍,李佳皓,潘建成,周勇.数值预报AI气象大模型国际发展动态研究[J].大气科学学报,2024,47(1):46-54. |
| |
| 作者姓名： | 黄小猛 林岩銮 熊巍 李佳皓 潘建成 周勇 |
| |
| 作者单位： | 清华大学 地球系统科学系, 北京 100084;浙江工业大学 计算机科学与技术学院, 浙江 杭州 310023;中国气象局 气象发展与规划院, 北京 100081 |
| |
| 基金项目： | 国家杰出青年科学基金资助项目(42125503);国家重点研发计划资助项目(2022YFE0195900;2021YFC3101600;2020YFA0607900) |
| |
| 摘 要： | 数值预报是研究地球系统的重要工具,有助于加深科学家对大气、海洋、气候和环境等复杂系统之间相互作用和变化过程的理解,在防灾减灾、气候变化和环境治理等方面发挥着不可或缺的作用。随着模式复杂度和分辨率的提高,传统数值模式在气候变化研究和气候预测方面取得了迅速的进展,但也面临一些挑战,需要得到数据同化、集合耦合、高性能计算和不确定性分析等多方面的支持。而近年来,“AI+气象”的交叉研究在气象领域引起了广泛关注。基于多种深度学习架构的人工智能大模型,依托强大的计算资源和海量的数据进行训练,能够以新的科学范式进行高效数值预报。气象大模型不断涌现,一些科技公司如华为、英伟达、DeepMind、谷歌、微软等,以及国内外高校如清华大学、复旦大学、密歇根大学、莱斯大学等发布了多个涵盖临近预报、短时预报、中期预报和延伸期预报等不同领域的气象大模型。这标志着人工智能与气象领域的交叉融合已经达到新的高度。尽管气象大模型在现阶段取得了较大突破,但其发展仍然面临弱可解释性、泛化能力不足、极端事件预报强度偏低、智能预报结果过平滑、深度学习框架能力需要拓展等诸多挑战。
|
| 关 键 词： | 数值预报 地球系统模式 深度学习 气象大模型 |
| 收稿时间： | 2023/12/1 0:00:00 |
| 修稿时间： | 2023/12/18 0:00:00 |
Research on international developments of AI large meteorological models in numerical forecasting |
| |
| HUANG Xiaomeng,LIN Yanluan,XIONG Wei,LI Jiahao,PAN Jiancheng,ZHOU Yong.Research on international developments of AI large meteorological models in numerical forecasting[J].大气科学学报,2024,47(1):46-54. |
| |
| Authors: | HUANG Xiaomeng LIN Yanluan XIONG Wei LI Jiahao PAN Jiancheng ZHOU Yong |
| |
| Institution: | Department of Earth System Science, Tsinghua University, Beijing 100084, China;School of Computer Science and Technology, Zhejiang University of Technology, Hangzhou 310023, China; Meteorological Development and Planning Institute, China Meteorological Administration, Beijing 100081, China |
| |
| Abstract: | Earth System Models (ESM) are powerful tools for studying the earth system and play an indispensable role in conducting scientific research on disaster prevention and mitigation,climate change,and environmental governance.Traditional weather and climate models rapidly evolve towards ESM,including ocean,sea ice,biogeochemical,and atmospheric chemical processes.At the same time,an increasing number of applications are adopting ESM for weather,climate,and ecological prediction.The current international mainstream trend in developing numerical models is to achieve seamless simulation and prediction by constructing integrated models,simultaneously meeting the needs of weather-climate forecasts and predictions at varying temporal and spatial scales.With improved model complexity and resolution,traditional numerical weather models have rapidly progressed in climate change research and climate prediction.However,challenges remain regarding data assimilation,ensemble coupling,high-performance computing,and uncertainty analysis and evaluation.The combination of artificial intelligence (AI) and meteorology has recently attracted tremendous attention.Based on various deep learning architectures,deep learning models can be trained using powerful computing resources and massive data for weather forecasts in a new scientific paradigm independent of traditional numerical weather models.Some technology companies,such as Huawei,NVIDIA,DeepMind,Google,Microsoft,etc.,as well as domestic and international universities such as Tsinghua University,Fudan University,the University of Michigan,Rice University,etc.,have released several Large Weather Models (LWMs) covering from nowcasting,short-term forecast to medium-term forecast,and even extended-period forecast.For instance,FourCastNet,GraphCast,NowcastNet,Pangu Weather,Fengwu,Fuxi,etc.,show significant advantages and great potential in improving forecast accuracy and accelerating the forecast inference process.For accuracy,except in areas like extreme weather,LWMs have matched or even surpassed that of traditional numerical models.Moreover,with continuous development of deep learning methods,their forecasting precision is steadily increasing.For timeliness,LWMs,leveraging deep neural networks'' powerful generalization capabilities,far exceed traditional numerical models'' predictive abilities under the same resolution conditions.For computational speed,LWMs have significantly increased inference computation speed compared to traditional numerical models,gradually reduced the enormous computation times required by traditional numerical models.The emergence of LWMs signifies that the cross-fertilization between AI and meteorological fields has reached a new horizon.Although these LWMs have made significant breakthroughs at this stage,their development still faces many challenges,such as the interpretability problem,the generalization and migration challenge,and the over-smoothing problem.The advancement of numerical weather prediction is closely tied to developments in computational and data storage technology,as well as observational techniques.Its application requires interdisciplinary integration,combining insights from various scientific fields.A critical scientific challenge in this field is to foster a more profound integration of numerical weather prediction with emerging information technologies such as artificial intelligence,quantum computing,and digital twins.This challenge also involves tailoring complex and refined component models to meet diverse disciplinary demands and societal needs.Advancing numerical weather prediction within the broader context of earth system science requires a concerted effort to promote cross-disciplinary collaboration,addressing vital scientific questions at the intersection of multiple fields. |
| |
| Keywords: | numerical forecasting Earth System Models deep learning large weather models |
|
| 点击此处可从《大气科学学报》浏览原始摘要信息 |
| 点击此处可从《大气科学学报》下载免费的PDF全文 |
|
