Earth-o1: A Purely Observations-driven Atmospheric World Model

完全基于观测驱动的通用大气世界模型

Junchao Gong1,3† Kaiyi Xu1,5† Wangxu Wei1,6† Siwei Tu1,7† Jingyi Xu1,7† Zili Liu1 Hang Fan8 Zhiwang Zhou1 Tao Han1 Yi Xiao1 Xinyu Gu1 Zhangrui Li1 Wenlong Zhang1 Hao Chen1 Yaqiang Wang9 Lijing Cheng6 Pierre Gentine8 Wanli Ouyang2 Zhe-Min Tan10 Feng Zhang4 Fenghua Ling1*; Ben Fei2*; Lei Bai1*;
(† equal contribution)
(*; corresponding authors)
1 Shanghai Artificial Intelligence Laboratory, Shanghai, China  
2 Department of Information Engineering, The Chinese University of Hong Kong, Hong Kong, China
3 School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China  
4 Department of Atmospheric and Oceanic Sciences, Fudan University, Shanghai, China 
5 School of Information Science and Technology, University of Science and Technology of China, Anhui, China 
6 State Key Laboratory of Earth System Numerical Modeling and Application, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China 
7 College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China  
8 Department of Earth and Environmental Engineering, Columbia University, New York, NY, USA  
9 Chinese Academy of Meteorological Sciences, Beijing, China  
10 School of Atmospheric Sciences, Nanjing University, Nanjing, China  
Paper Code Checkpoints Dataset (upcoming)

📰 News & Updates

  • 🚀

    December 5, 2025

    The Project Page are now live!

Abstract

The growing abundance of multi-source atmospheric observations provides unprecedented opportunities for real-time weather prediction and monitoring. Data assimilation (DA), the primary framework for integrating these diverse observations by incorporating them into the model space, currently exploits only a fraction of available observations. Similarly, forecast models initialized with DA initial fields are constrained by fixed grids, which further impairs the reconstruction and short-term prediction of local hazard. Here, we present a purely observation-driven atmospheric world model (Earth-o1), a next-generation, AI-based platform founded on a unified multimodal latent observation space that integrates diverse observational streams from satellites (both geostationary instruments and polar orbiters), and in-situ networks. Within this space, Earth-o1 learns the intrinsic dynamics of atmospheric evolution and supports flexible, arbitrary-resolution, multi-model, and multi-product forecasting and retrieval, enabling natural coupling across Earth subsystems. Earth-o1 delivers real-time atmospheric forecasts, high-resolution hazard monitoring, and cross-sensor retrieval products, while providing more diverse and timely diagnostic variables than ERA5 reanalysis and achieving forecast accuracy for surface in-situ observations comparable to IFS. This framework represents a step toward a living, closed-loop digital mirror of Earth, transforming heterogeneous observations into actionable intelligence for science and decision-making.

Overview of Earth-o1

Teaser.

Overview of Earth-o1. After tokenization, multi-modal heterogeneous data from LEOs, GEOs and in-situ platforms are trained via a Masked Autoencoder, where representations are mutually integrated to generate modality observation tokens. Analysis: these latent observation tokens support multi-sensor analysis tasks including microwave and infrared remote sensing. Forecast: facilitate the forecasting of observation signals. Inversion: enable the inversion of downstream geophysical products; as illustrated in the figure, high-value products such as sea ice concentration and CO concentration are inverted.