Understanding of the Wintertime Atmospheric Aerosol Properties with Explainable Machine Learning

Winter atmospheric aerosols, marked by formation under dynamic and complex conditions due to distinct environments, haze events and regional transportation, is greatly challenging to investigate. To address this, we employed XGBoost models integrated with SHapley Additive exPlanations (SHAP) to explore the meteorological and chemical drivers, as well as the impact of transportation, on aerosol characteristics.

We conducted measurements using a high-resolution time-of-flight aerosol mass spectrometer (HR-ToF-AMS) during the 2018 winter (Jan 17th to Feb 22nd) in urban Seoul, Korea. Our analysis included various PM components (nitrate, sulfate, chloride, ammonium, and organics) and sources of organic aerosols (OA), such as more-oxidized oxygenated OA (MO-OOA), less-oxidized OOA (LO-OOA), cooking OA (COA), hydrocarbon-like OA (HOA), and biomass burning OA (BBOA), using positive matrix factorization (PMF). The models demonstrated high predictive accuracy (R>0.90) for all species and sources.

Notably, nitrate formation was found to be significantly influenced by CO concentration and relative humidity (RH), highlighting the role of local sources and aqueous-phase formation. For sulfate, RH was identified as the dominant factor. Organic components, constituting 42.4% of total PM mass, were analyzed for their diverse sources. Temperature and RH were the major drivers for MO-OOA (O/C=0.94) formation, with a critical temperature threshold near 0 °C identified for differentiating formation conditions. Specifically, temperature above the ice point and high RH significantly enhanced MO-OOA formation, and it is likely related to the availability of liquid water for aqueous-phase oxidations to occur. LO-OOA (O/C=0.77) was controlled by CO concentration, suggesting its local formed feature being the same line with nitrate.

Primary OAs, HOA (O/C=0.09) and BBOA (O/C=0.39) were dominated combustion sources (CO concentration), while BBOA (O/C=0.39) was closely linked to temperature. Interestingly, BLH showed a greater impact on COA (O/C=0.20) ,likely due to accumulation during early shrinkage of the boundary layer in winter.

This novel approach effectively identified distinct drivers of aerosol formation and emission features in winter, offering new insights compared to traditional methods. However, the models showed limitation in defining the strong influence of transportation impacts from upwind areas, as found in various research, possibly due to constraints in cluster ID input, which could not distinguish high and low loading clusters. The limitation indicates an area for further investigation.