Quantifying source-specific ozone concentrations is essential for formulating precursor emission control policies and assessing associated health risks. However, the formulation of targeted O3 policies is currently impeded by the limited accuracy of traditional chemical transport models at fine scales and the inherent inability of machine learning methods to provide source-specific attribution. To address this, we developed an integrated framework coupling the deep learning model with a source-oriented Community Multiscale Air Quality model to reconstruct O3 concentrations over 2005–2020 and apportion them at milestone years. Our results identify background contribution, industry, and transportation as the primary O3 contributors, exhibiting pronounced spatiotemporal heterogeneity. Notably, while background O3 accounts for ∼75.0% of the warm-season (May–September) concentration, industry and transportation contribute a disproportionate 28.1% of O3-attributable premature deaths despite representing only 14.0% of concentrations. This disparity stems from the strong spatial overlap between intensive anthropogenic emissions and densely populated areas. Furthermore, decomposition analysis revealed that spatiotemporal variations in O3 concentrations acted as the dominant driver (74.2%) of the increase in O3-related premature mortality during the study period, followed by changes in baseline mortality rates (14.9%) and population dynamics (10.9%). These findings provide a critical scientific basis for China’s transition from total concentration control to health-oriented, source-specific O3 management strategies.
Wang et al. (2026) studied this question.