Machine Learning for Ozone Pollution Research: A Chemistry-Informed Framework for Model Selection and Interpretation

Abstract Ozone pollution remains a major challenge for air quality management, ecology and public health because of its nonlinear formation chemistry, strong meteorological modulation, multiscale transport and toxicity. Although chemical transport models and conventional statistical approaches are widely used, their ability to resolve high-dimensional interactions, regime-dependent behavior, and data heterogeneity is often limited. In this context, machine learning (ML) has emerged as a powerful complement, enabling high-resolution ozone prediction, multisource data integration, and data-driven attribution of key drivers. However, the rapid expansion of ML applications has outpaced systematic, chemistry-informed guidance on model selection, interpretability, and physical consistency. This review critically synthesizes recent advances in ML-based ozone research, spanning classical methods, tree-based ensemble models, deep learning architectures, and hybrid physical–ML frameworks. Rather than emphasizing predictive accuracy alone, we assess where different model families succeed or may mislead when applied to ozone problems characterized by chemical regime transitions, nonstationarity, and limited observations. By proposing a structured heuristic that links model selection to ozone research goals, this review aims to guide ML from a predictive tool toward a principled component of atmospheric chemistry analysis. Such a shift could enhance forecasting robustness, improve mechanistic understanding, and support policy-relevant assessments in ozone pollution research.