VOC observations improve transferable ozone prediction using interpretable machine learning

Abstract Surface ozone variability is controlled by complex interactions among emissions, chemistry, and meteorology, yet the extent to which machine learning models capture chemically meaningful relationships remains uncertain. Here we apply an automated machine learning framework to develop optimized tree-based ensemble models using four years of meteorological, primary pollutant, and volatile organic compound (VOC) observations collected in Shanghai. Interpretable machine learning analysis indicates that meteorological variables and primary pollutants account for much of the predictive importance in the model. However, models constructed solely from these predictors exhibit limited ability to generalize across different environments. Incorporating VOC observations substantially improves model transferability, as shown during the COVID-19 emission perturbation and in independent evaluations at a rural site in northern China. The improved extrapolation of VOC-based models suggests that VOC predictors provide transferable statistical constraints on ozone variability that are less dependent on site-specific conditions. Comparison with box-model simulations suggests that while the relationships learned by the machine learning models do not explicitly represent photochemical mechanisms, they reflect statistical associations between VOC concentrations and ozone formation. These findings demonstrate the importance of VOC observations for developing machine learning models that produce interpretable and transferable predictions of surface ozone under varying atmospheric regimes.