Duo‐AttnOPNets: Advancing Global Operational Forecasting for Atmospheric Carbon Monoxide With AI‐Empowered 4D‐Var

Abstract The accuracy of global forecasting of atmospheric composition is essential for protecting public health and advancing climate research. Carbon monoxide (CO), a key pollutant with indirect greenhouse effects, requires timely and accurate prediction. We propose Duo‐AttnOPNets, an operational framework that combines a deep learning‐based forecasting module Duo‐AttnForeNet with a training‐free data assimilation module Duo‐AttnVarNet. Duo‐AttnForeNet employs CSLSTM blocks—built from ConvLSTM cells and dual attention mechanisms—to effectively capture spatio‐temporal dynamics. Duo‐AttnVarNet leverages automatic differentiation and GPU acceleration to enable efficient four‐dimensional variational (4D‐Var) assimilation without extensive manual adjoint coding. We evaluate Duo‐AttnOPNets against the Integrated Forecasting System for Atmospheric Composition (C‐IFS). Results show that Duo‐AttnOPNets achieves comparable or superior accuracy in both forecasting and assimilation, while generating 5‐day forecasts within seconds on a single GPU. These findings demonstrate its potential for real‐time, scalable, and accurate CO forecasting, marking a promising advance in integrating deep learning with traditional variational methods for operational atmospheric modeling.