Interpretable machine learning identifies chemical factors associated with deweathered ozone pollution in Suzhou, Yangtze River Delta.

Surface ozone is affected by both atmospheric chemistry and weather, making it difficult to identify the chemical factors related to ozone pollution from direct observations alone. Here we combined observations from 2015 to 2022 with interpretable machine learning to examine key chemical factors associated with deweathered surface ozone at an urban site in Suzhou, Yangtze River Delta. The Extreme Gradient Boosting (XGBoost) algorithm was used in two steps. First, a model based on meteorological variables was trained to reduce the influence of weather on observed O3. Second, deweathered O3 was related to gaseous pollutants, grouped volatile organic compounds and inorganic aerosol components, and the results were interpreted using Shapley additive explanations (SHAP). The deweathering step substantially reduced the variability of O3, with the standard deviation decreasing by about 39%. The model also showed good performance, with a test R2 of 0.76. In the SHAP analysis, OVOCs, CO and alkenes ranked as the three most important predictors, accounting for 15.1%, 11.7%, and 10.2% of the normalized mean absolute SHAP contribution, respectively. OVOCs showed a clear positive relationship with deweathered O3, while high levels of alkenes and CO were more often associated with lower deweathered O3. The response of NO2 is consistent with a NOx-rich or VOC-sensitive chemical environment in Suzhou. Inorganic aerosol components, including nitrate, sulfate and ammonium, were more likely to indicate active secondary formation and oxidizing conditions than to act as direct ozone precursors. These results show that deweathering combined with interpretable machine learning provides a practical framework for reducing meteorological variability and interpreting chemical associations in long-term ozone observations.