Integrating the Machine Learning Framework to Decipher Ozone Control Strategies in Coastal Regions: From Seasonal Management to Source Interaction Effects.

Ground-level ozone (O3) pollution poses escalating challenges to urban air quality management, yet the nonlinear mechanisms governing its formation in complex coastal environments remain insufficiently characterized. This study develops an interpretable machine learning framework integrating Extreme Gradient Boosting (XGBoost), SHapley Additive exPlanations (SHAP), and Positive Matrix Factorization (PMF) to conduct a comprehensive attribution analysis of O3 formation in Hong Kong from 2020 to 2024. Annual mean O3 concentrations increased by 8% over the study period, reaching a record high of 77.1 μg/m3 in 2024, with pronounced seasonal peaks in autumn (95.1 μg/m3 in October). SHAP-based attribution revealed that meteorological factors dominate O3 variability, contributing 62.2% of the total predictive importance, with surface solar radiation (SSR) identified as the primary modulator. A critical synergistic mechanism was identified: under high SSR level (> 18 MJ/m2), elevated particulate matter concentrations were associated with enhanced atmospheric oxidative capacity under light-enhanced heterogeneous reactions, corresponding to a nonlinear escalation of O3. Seasonal attribution further demonstrated that transport driven by wind governs spring O3 dynamics, while photochemical production dominates autumn exceedances and regional advection intensifies during winter. By integrating PMF-derived source profiles into the XGBoost model, this study overcomes the linear constraints of traditional ozone formation potential (OFP) methods. The SHAP framework attributed 26.8% greater importance to secondary volatile organic compounds (VOCs) formation than OFP estimates, demonstrating that primary-reactivity-based approaches systematically underestimate secondary transformation pathways in oxidatively active coastal atmospheres. Solvent usage was identified as the dominant anthropogenic VOC source (25% contribution). Therefore, priority should be given to reducing reactive solvent-related VOCs from coating, printing, cleaning, and industrial solvent-use processes, together with coordinated control of industrial emissions to suppress source-level synergies in O3 formation. These findings provide a scientifically robust, mechanistically interpretable foundation for precision ozone mitigation strategies in the Guangdong-Hong Kong-Macao Greater Bay Area.