Research on the influencing factors of PM2.5 in China at different spatial scales based on machine learning algorithm

• Machine learning model improves PM 2.5 prediction with R 2 up to 0.98 at grid level. • SHAP analysis shows that pollutant factors significantly impact urban PM 2.5 levels. • Meteorological factors dominate PM 2.5 influences in non-urban areas. • Tailored interventions by region can effectively reduce PM 2.5 pollution. PM 2.5 pollution is one of the prominent environmental issues currently faced in China, influenced by various factors and showed significant spatial differences. In this study, the Light Gradient Boosting Machine (LightGBM) model was employed in combination with SHapley Additive exPlanation (SHAP) methods to explore the key impact factors (precursor emissions, meteorological conditions, geographical features and socioeconomic factors) on average annual PM 2.5 levels from 2015 to 2022 at both city and grid levels in China. The results show that incorporating pollutant concentration into the model enhances its performance, with R 2 improving significantly from 0.79 to 0.93, which underscores the importance of pollutant concentration and the outstanding predictive performance of the LightGBM algorithm. Further, after increasing the spatial resolution and applying a grid-level model, R 2 further improves to 0.96 ∼ 0.99. SHAP analysis revealed that PM 2.5 levels in urban areas are significantly influenced by pollutant concentration such as NO 2 , CO, and SO 2 , accounting for 49.3 % of the total impact. In contrast, the grid-based model highlights the dominant role of meteorological factors such as temperature and precipitation influencing PM 2.5 levels in non-urban areas. Moreover, the model results also suggested that the PM 2.5 pollution in Yangtze River Delta (YRD) and Pearl River Delta (PRD) are mainly controlled by primary emissions, while in Beijing-Tianjin-Hebei (BTH), Fenwei Plain (FWP) and Sichuan Basin (SCB), atmospheric oxidation capacity is a limiting factor. This study underscores the potential of machine learning in atmospheric pollution control and offers insights for developing more effective and region-specific PM 2.5 control policies.