Quantifying regional transport contributions and diagnosing ozone formation sensitivity: A trajectory informed machine learning study at background sites across China

Abstract. Surface ozone (O₃) pollution in China is affected by transported pollutants, air mass pathways, local pollution conditions, meteorology, temporal variability, and stratospheric inputs, but their relative roles remain difficult to distinguish at background sites. Here we developed a nationwide trajectory informed machine learning framework to quantify model explained processes controlling hourly O₃ at 76 background monitoring sites across China during 2015–2024. Hourly observations were integrated with 72 h HYSPLIT backward trajectories, high resolution gridded pollutant fields along trajectory endpoints, ERA5 meteorology, a CAMS stratospheric O₃ tracer, and temporal predictors. XGBoost models showed robust performance, with a median test R² of 0.853 across valid station year models. SHAP interpretation showed that trajectory pollutants were the dominant explanatory group, accounting for 56.4 % of the total model explained contribution, followed by ERA5 meteorology, local pollutants, temporal features, trajectory location and height, and stratospheric O₃. During MDA8 O₃ exceedance days (>160 mg m⁻³), trajectory pollutant contribution increased to 65.3 %, and further to 72.3 % during the peak 8 h window. This enhancement was mainly driven by trajectory O₃, whose share within the trajectory pollutant group increased from 75.6 % to 89.3 %. OMI HCHO/NO₂ ratios indicated that O₃ formation was mainly VOC limited or transitional. Diurnal analysis further showed that maximum O₃ was associated with local pollutant meteorology coupling, whereas higher minimum O₃ reflected transport pathway and background structure. These results highlight the importance of transported O₃ rich air masses in high O₃ episodes at Chinese background sites.