Quantifying the drivers of ozone formation using localized photochemical reactivity and double machine learning

Surface ozone (O₃) pollution remains severe in many Chinese cities despite declining precursor emissions, reflecting nonlinear interactions among emissions, meteorology, chemistry, and transport. Here, we integrated hourly observations from Yuncheng in the Fenwei Plain during 2020-2022 with a localized observation-constrained box model to develop a photochemically constrained double machine learning (DML) framework. From 2020 to 2022, the annual mean O₃ concentration in Yuncheng increased from 76.2 to 79.9μg m^-3. Compared with clean days, O₃-polluted days had stronger southeasterly winds, higher temperatures, lower humidity. Box-model calculations indicated that O₃ formation in Yuncheng was VOC-limited during the pollution season. Using MCM v3.3.1, we derived a localized maximum incremental reactivity scale for Yuncheng (MIR-YC), with reactivity decreasing in the order isoprene > anthropogenic alkenes > aromatics > oxygenated VOCs > alkanes. The resulting localized ozone formation potentials were subsequently used in place of measured VOC concentrations to constrain the machine-learning analysis. The optimized random forest model achieved a test R² of 0.923, and SHAP analysis attributed 44.6% and 26.4% of predictive importance to local precursors and photochemical drivers, respectively. DML then quantified conditional marginal responses beyond predictive attribution, associating a 1 °C temperature increase with a 4.04μg m^-3 increase in O₃ after adjustment for observed covariates. This integrated framework links localized chemical sensitivity with observation-based atmospheric responses, providing a quantitative basis for locally tailored VOC-NOx control strategies.