Unraveling Drivers of Secondary Pollutant Formation Using Explainable Machine Learning: Insights for Urban Air Quality Management

Prof. Hwajin Kim,
Department of Environmental Health Sciences, Seoul National University

Severe haze events in urban environments are increasingly driven by secondary pollutants such as nitrate, secondary organic aerosols (SOA), and ozone. Their formation involves complex, nonlinear interactions among precursors, meteorology, and atmospheric chemistry, making it challenging to fully understand these processes. To address this, we apply explainable machine learning (XGBoost with SHapley Additive exPlanations [SHAP]) alongside high-resolution seasonal measurements to identify key drivers of secondary pollution.

Our models demonstrated strong predictive performance (R² > 0.90), revealing dominant factors and nonlinear effects. For nitrate, NO_₂ and solar radiation were the dominant wintertime drivers, contributing 36.7%. RH and temperature showed nonlinear effects, with nitrate formation peaking around 65% RH and persisting at lower RH, suggesting ongoing aqueous-phase processing influenced by phase transitions. Below freezing (~0°C), levels dropped sharply, likely due to inhibited heterogeneous reactions in semi-solid or glassy states. SOA formation exhibited a nonlinear relationship with O_x (O_₃ + NO_₂), suggesting possible fragmentation under extensive oxidation. Ozone formation was primarily driven by the HCHO/NO_x ratio (25.8%), reflecting the balance between photochemical production and NO_x titration. Additionally, the impact of VOCs on ozone formation varied across concentration ranges, indicating a complex interaction that requires further investigation.

These findings highlight the potential of machine learning to untangle the complexity of secondary pollutant formation, offering insights for improved air quality management. However, further investigation is needed to assess pollutant transport effects, particularly for sulfate and rural ozone, as well as other potential unknown factors, to enhance model applicability. Additionally, careful consideration is required when selecting input variables and interpreting results to prevent misrepresentation or overgeneralization.