Explainable machine learning reveals the meteorological drivers of ozone exceedances in a high-emission industrial park

Abstract Ground-level ozone exceedances in industrial parks remain poorly understood, as the drivers identified for urban atmospheres have rarely been quantified at the precursor levels typical of these parks. We analyzed a 56-day record of hourly observations from six monitoring sites in a fine-chemical industrial park on Hangzhou Bay, eastern China, and contrasted an eight-day ozone exceedance episode that emerged in early September 2020 with eight non-exceedance control days. The episode coincided with the peripheral subsidence of the successive typhoons Maysak and Haishen, which sustained the clear, dry and stable conditions under which ozone built to exceedance levels. A convolutional neural network trained on the monitored variables reproduces 88% of hourly ozone variability. By combining DeepSHAP attribution with controlled perturbations of the model, we find that meteorological variables, led by relative humidity and temperature, account for 65.9% of the model-attributed ozone variance during the episode versus 54.8% on control days. Perturbation analysis within the observed temperature–humidity domain further shows that the suppression of ozone by water vapor steepens beyond an absolute humidity of 19 g·m −3 , a threshold that reappears in observation-level regressions. Together, these results quantify the contributions of the meteorological variables and precursor concentrations to ozone at the precursor levels typical of high-emission industrial parks, and identify the absolute-humidity threshold beyond which the ozone response changes, providing a quantitative basis for ozone management in such parks.