Integrated diagnostic framework for resolving distinct ozone formation mechanisms via observations, modeling, and machine learning

Ozone (O3) pollution in relatively high-latitude industrial cities shows spatiotemporal patterns associated with coupled meteorological, transport, and chemical processes, yet its dominant formation mechanisms remain unresolved. This study developed an interpretable diagnostic framework integrating process-resolved Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) simulations, observations, and Extreme Gradient Boosting (XGBoost) interpreted using Shapley additive explanations (SHAP) for Changchun, Northeast China (September 2023–August 2024). Surface daily maximum 8-h average (MDA8) O3 exceedances occurred from April to July, peaking above 200 μg/m3. Lidar monthly means at 0.3–1.0 km peaked from April to June (108.9–111.4 μg/m3), while the 1.0–2.0 and 2.0–3.0 km layers peaked in May (97.3 and 76.4 μg/m3), indicating that lower-layer intensity and vertical extent peaked asynchronously. Across five analysis windows (E1–E5), WRF-Chem diagnosed vertical mixing as the largest positive contribution in E1 and E3 (+19.6 and +12.2 µg/m3), concurrent positive contributions in E2 led by vertical mixing (+41.3 µg/m3) and chemistry (+21.1 µg/m3), and chemistry as the largest positive contribution in E5 (+21.8 µg/m3). In E4, upper-quartile O3 coincided with predominantly negative contributions, particularly vertical advection (−24.0 µg/m3), showing that high concentrations and hourly O3 changes were asynchronous. XGBoost–SHAP identified 2-m temperature (T2) and NO2 as the leading nonlinear predictors: T2-related SHAP values became positive near 18.5 °C, whereas the NO2-related values shifted from positive at 10–20 μg/m3 to predominantly negative above 26 μg/m3. The framework integrates vertical observations, statistical attribution, and modeled tendencies to support event-scale O3 diagnosis.