Multi-objective machine learning for health-oriented O₃ and PM_2.5 control: Integrating VOC photochemical consumption...

Compound pollution of ozone (O₃) and fine particulate matter (PM_2.5) represents a critical challenge for air quality management in China. This study developed an integrated multi-objective analysis framework coupling Observation-Based Model (OBM), multi-task machine learning (Multi-CatBoost), and Positive Matrix Factorization (PMF) to systematically assess the synergistic formation mechanisms of O₃ and PM_2.5 in Shaoxing City, based on year-round measurements in 2023. An improved photochemical age parameterization method was employed to quantify VOC atmospheric consumption, revealing that 22.1 % of initial VOC concentrations were photochemically consumed during transport. The Multi-CatBoost model achieved reliable predictive performance (R² = 0.91 for both pollutants) while capturing cross-pollutant synergistic mechanisms through joint optimization, significantly enhancing the importance of cross-media precursors (NO₂ importance for PM_2.5 increased by 86.7 %). SHAP analysis identified concentration-dependent thresholds governing O₃-PM_2.5 formation trade-offs, with NO₂ exhibiting dual roles: promoting O₃ formation below 16 μg/m³ while favoring nitrate aerosol formation above 35.6 μg/m³. Health-oriented sensitivity analysis based on Air Quality Health Index (AQHI) demonstrated that VOC reduction is the optimal strategy across all pollution scenarios, with complete VOC reduction decreasing AQHI by 19.27 %-34.67 %, substantially exceeding NOx reduction effects (0.20 %-4.60 %). Source-specific analysis through the SHAP Interaction Redistribution Analysis (SIRA) framework identified industrial emissions and biomass burning as priority control targets, jointly contributing 55.8 % of total AQHI improvement potential. This study provides a novel paradigm for formulating health-oriented, scenario-specific emission reduction strategies for compound pollution control.