Machine learning-based disentanglement of meteorological and anthropogenic drivers of ozone pollution in the Chengdu-Chongqing urban agglomeration.

To clarify the relative contributions of meteorological conditions and anthropogenic activities to ozone (O3) pollution in the Chengdu-Chongqing urban agglomeration, this study systematically quantified their impacts on the daily maximum 8-h average ozone concentration (O3-8 h) during 2015-2024. Using ground-based observations and ERA5-Land reanalysis data, we combined the Light Gradient Boosting Machine (LightGBM) machine learning model, the SHAP (SHapley Additive exPlanations) interpretability framework, a meteorological normalization approach, and controlled-variable perturbation experiments to investigate the drivers of O₃ pollution and compare the differences among interpretative methods. The results showed that the LightGBM model reproduced O₃-8 h well, with a cross-validated R2 of 0.88. Meteorological normalization analysis indicated that meteorological factors dominated the interannual variability of O3-8 h, with an average contribution of 70.5%, while anthropogenic emissions contributed 29.5%. Meteorological conditions generally suppressed O3 formation during 2015-2018 but shifted to a promoting effect after 2019, with a markedly enhanced influence during 2022-2024. SHAP analysis identified surface solar radiation (ssrd), 2-m air temperature (t2m), and soil moisture (swvl1) as the key meteorological drivers, with contribution rates of 34.0%, 16.4%, and 6.0%, respectively. In contrast, controlled-variable perturbation experiments showed that O3-8 h was most sensitive to temperature, whereas its response to solar radiation was weaker and nonlinear. In addition, satellite-based formaldehyde-to-nitrogen dioxide ratio analysis indicated a shift in ozone formation sensitivity from a VOCs-limited regime toward a NOx-limited regime during 2019-2024. These findings provide support for regional ozone control strategies.