Integrating causal inference and interpretable automated machine learning to identify drivers of near-surface ozone concentration in eastern China

In recent years, near-surface ozone (O 3 ) pollution in Eastern China has intensified, with high concentrations, recurrent episodes, and pronounced spatiotemporal heterogeneity, underscoring the urgent need to elucidate its driving mechanisms. However, the complex nonlinear interactions among multiple factors limit the ability of traditional statistical methods or standalone machine learning models to provide accurate explanations. This study aims to systematically identify the key drivers of O 3 concentrations in Eastern China during 2022–2023 by developing an integrated framework of “automated modeling–feature interpretation–causal inference.” Specifically, the Extra Trees model is used for prediction, SHAP for feature interpretation, and causal inference methods (CF and DML) for causal estimation. The results indicate that meteorological variables dominate O 3 variation, with 2-m temperature (2 mT) exerting a significant positive causal effect, particularly in high-emission provinces such as Shandong and Jiangsu. Tropospheric O 3 column concentration (Tor_O 3 ), reflecting regional background transport, also shows a strong causal impact in central Jiangsu and northern Zhejiang. SHAP analysis further reveals that elevated temperature and strong radiation enhance photochemical reactions, thereby facilitating O 3 formation. The consistent causal directions and spatial distributions obtained from CF and DML confirm the robustness of the findings. Overall, this study not only advances empirical understanding of the mechanisms of O 3 pollution but also provides methodological guidance for improving regional air quality management. • Proposed a unified AutoML–SHAP–causal inference framework to identify drivers of near-surface O 3 . • Extra Trees yielded accurate near-surface O 3 predictions across Eastern China (R 2 > 0.97). • SHAP revealed surface temperature and radiation as key drivers with spatial variation. • CF and DML models confirmed and quantified meteorological causal impacts on near-surface O 3 . • The framework integrates interpretability and causality to inform adaptive control.