Convolutional Neural Networks Facilitate Process Understanding of Megacity Ozone Temporal Variability.

Ozone pollution is profoundly modulated by meteorological features such as temperature, air pressure, wind, and humidity. While many studies have developed empirical models to elucidate the effects of meteorology on ozone variability, they predominantly focus on local weather conditions, overlooking the influences from high-altitude and broader regional meteorological patterns. Here, we employ convolutional neural networks (CNNs), a technique typically applied to image recognition, to investigate the influence of three-dimensional spatial variations in meteorological fields on the daily, seasonal, and interannual dynamics of ozone in Shenzhen, a major coastal urban center in China. Our optimized CNNs model, covering a 13° × 13° spatial domain, effectively explains over 70% of daily ozone variability, outperforming alternative empirical approaches by 7 to 62%. Model interpretations reveal the crucial roles of 2-m temperature and humidity as primary drivers, contributing 16% and 15% to daily ozone fluctuations, respectively. Regional wind fields account for up to 40% of ozone changes during the episodes. CNNs successfully replicate observed ozone temporal patterns, attributing -5-6 μg·m -3 of interannual ozone variability to weather anomalies. Our interpretable CNNs framework enables quantitative attribution of historical ozone fluctuations to nonlinear meteorological effects across spatiotemporal scales, offering vital process-based insights for managing megacity air quality amidst changing climate regimes.