Separating Daily 1 km PM2.5 Inorganic Chemical Composition in China since 2000 via Deep Learning Integrating Ground, Satellite, and Model Data

Fine particulate matter (PM 2.5 ) chemical composition has strong and diverse impacts on the planetary environment, climate, and health. These effects are still not well understood due to limited surface observations and uncertainties in chemical model simulations. We developed a four-dimensional spatiotemporal deep forest (4D-STDF) model to estimate daily PM 2.5 chemical composition at a spatial resolution of 1 km in China since 2000 by integrating measurements of PM 2.5 species from a high-density observation network, satellite PM 2.5 retrievals, atmospheric reanalyses, and model simulations. Cross-validation results illustrate the reliability of sulfate (SO 4 2– ), nitrate (NO 3 – ), ammonium (NH 4 + ), and chloride (Cl – ) estimates, with high coefficients of determination (CV-R 2 ) with ground-based observations of 0.74, 0.75, 0.71, and 0.66, and average root-mean-square errors (RMSE) of 6.0, 6.6, 4.3, and 2.3 μg/m 3, respectively. The three components of secondary inorganic aerosols (SIAs) account for 21% (SO 4 2– ), 20% (NO 3 – ), and 14% (NH 4 + ) of the total PM 2.5 mass in eastern China; we observed significant reductions in the mass of inorganic components by 40–43% between 2013 and 2020, slowing down since 2018. Comparatively, the ratio of SIA to PM 2.5 increased by 7% across eastern China except in Beijing and nearby areas, accelerating in recent years. SO 4 2– has been the dominant SIA component in eastern China, although it was surpassed by NO 3 – in some areas, e.g., Beijing–Tianjin–Hebei region since 2016. SIA, accounting for nearly half (∼46%) of the PM 2.5 mass, drove the explosive formation of winter haze episodes in the North China Plain. A sharp decline in SIA concentrations and an increase in SIA-to-PM 2.5 ratios during the COVID-19 lockdown were also revealed, reflecting the enhanced atmospheric oxidation capacity and formation of secondary particles.