D‐DNet: A Dual Deep Neural Network Framework for High‐Efficiency Operational PM2.5 and AOD550 Forecasting With Data Assimilation

Shengjuan Cai, Fangxin Fang, Vincent‐Henri Peuch, Mihai Alexe, Ionel Michael Navon, Yanghua Wang

Abstract Accurate forecasting of PM2.5 (particulate matter with diameter ≤2.5 μm) and AOD550 (aerosol optical depth at 550 nm) is crucial for air quality management, public health, and environmental policy. Traditional physics‐based forecasting systems, though robust, require computationally intensive simulations and data assimilation (DA) schemes that limit their flexibility for real‐time applications. Deep learning (DL) models offer a more efficient alternative but often suffer from error accumulation over extended forecasts. To address these limitations, we propose a dual deep neural network (D‐DNet), a unified closed‐loop framework that tightly couples DL‐based forecasting with DA. By combining a forecasting network with an observation‐driven correction module, D‐DNet assimilates satellite‐derived aerosol data to dynamically adjust forecasts and reduce long‐term drift. Evaluated on global PM2.5 and AOD550 forecasting for the full year of 2019, D‐DNet consistently produced forecasts that closely matched the EAC4 reanalysis. Compared to the physics‐based CAMS 4D‐Var system, D‐DNet achieved comparable or slightly better accuracy while being approximately 70,000 times faster in its end‐to‐end workflow. This computational efficiency makes D‐DNet well‐suited for ensemble forecasting, uncertainty quantification, and near‐real‐time air quality applications.