Using machine learning to unravel chemical and meteorological effects on ground-level ozone: Insights for ozone-climate control strategies
Zhiyuan Li, Yifan Wang, Junling Liu, Junrui Xian
Abstract
In the context of climate change, various countries/regions across East Asia have witnessed severe ground-level ozone (O 3 ) pollution, which poses potential health risks to the public. The complex relationships between O 3 and its drivers, including the precursors and meteorological variables, are not yet fully understood. Revealing the impact of multiple drivers on O 3 is crucial for providing evidence-based information for pollution control. In the present study, we evaluated the influence of key chemical-aerosol (e.g., volatile organic compounds, PM 2.5 , NO x ) and meteorological drivers (e.g., air temperature, relative humidity) on ground-level O 3 pollution at Tucheng site in New Taipei, Northern Taiwan, using fine-resolution atmospheric composition measurements and machine learning. The developed random forest machine learning models performed well, with 10-fold cross-validation R 2 values above 0.867. The results reveal seasonal disparities on chemical and meteorological effects on ground-level O 3 between winter and summer. Aggregated SHAP values indicated that chemical (e.g., NO x and VOCs) and aerosol variables (i.e., PM 2.5 ) accounted for 82.4 % of the explained variance in winter O 3 predictions and 62.1 % in summer. Meteorological variables (e.g., temperature, relative humidity) contributed the remaining variance, highlighting seasonally shifting sensitivities. Across seasons, temperature, 1,2,3-Trimethylbenzene, NO x , t-2-Butene, and relative humidity were identified as the dominant drivers of ground-level O 3 predictions, reflecting their modelled associations with elevated O 3 concentrations. The machine learning-based modelling framework developed in this study can be easily adapted to new sampling sites with minor modifications if necessary.