Enhancing 72-Hour air quality forecasting with an observation-driven deep learning chemistry transport model

Abstract

Real-time air quality forecasting with atmospheric chemistry transport models (CTMs) has long been hindered by the inaccessibility of in-time updates for crucial inputs (e.g., emissions) and chemical mechanism, posing a significant obstacle to designing effective control strategies for protecting human health. Here we present a study that leveraging multiple observations combined with corresponding meteorological conditions to overcome these long-standing limitations in high-resolution air quality forecasting. Specifically, we developed a novel observation-driven deep learning-based atmospheric chemistry forecasting (DeepFC), which effectively integrates abundant near-real-time satellite and ground-based observations. Using a nine-year observation dataset (2013–2021) with CTM simulation training over a 27 km × 27 km resolution domain in China, DeepFC driven entirely by observations significantly improves the accuracy of traditional numerical models in forecasting the concentrations of two major pollutants, PM_2.5 and O₃, for the following 72 h across the country. Specifically, it enhances R² from 0.2 to 0.6 and reduces RMSE by 50 % for PM_2.5 and 20 % for O₃, primarily due to the effective of fusing historical observation data and a more effective prediction strategy. Moreover, the newly developed observation-driven DeepFC provides deep insights into key factors for effective policy design, including source-receptor relationships, emission-response dynamics, and the separation of meteorology- and emission-driven variations. While consistent with traditional CTMs, DeepFC achieves these insights with significantly greater efficiency. These results highlight its strong potential in supporting global efforts to combat air pollution for better protection of human health and ecosystem.