关键词: 深度学习, 数据同化, 数值天气预报, 生成式模型, 人工智能气象大模型, 代理模型

Abstract: Data Assimilation (DA) is the core of modern Numerical Weather Prediction (NWP) systems. However, traditional methods are constrained by bottlenecks such as strong non-linear/non-Gaussian assumptions and prohibitive computational costs. Deep Learning (DL), with its powerful non-linear approximation and end-to-end optimization capabilities, offers a revolutionary paradigm to address these challenges. From a computer science perspective, this paper systematically reviews and analyzes a three-stage evolutionary framework for the application of DL in meteorological DA: “Component Replacement,” “Paradigm Shift,” and “System Reconfiguration”. First, it elaborates on the “Component Replacement” stage, where DL serves as surrogate models to replace traditional bottlenecks like observation operators and to correct model errors. Next, it provides an in-depth analysis of the “Paradigm Shift,” where generative models, represented by Variational Auto Encoders (VAE) and Diffusion Models, reframe the assimilation problem as a probabilistic generation task. Furthermore, it discusses the “System Reconfiguration” stage, represented by Artificial Intelligence（AI）weather models such as FengWu and FuXi, in which the assimilation module is deeply embedded to construct an integrated, closed-loop “assimilation-forecast” system. Finally, the paper explores the core challenges in this field, including physical consistency, data dependency, and operational deployment. It provides an outlook on future research directions, such as neuro-symbolic fusion and the development of open benchmarks.

Key words: deep learning, data assimilation, Numerical Weather Prediction (NWP)；generative model, artificial intelligence weather model, surrogate model