Machine learning-based source apportionment of particulate pollution aids real-time predictions and long-term emission regulations

Accurate source apportionment of particulate pollution is essential for timely emission control strategies to improve air quality and mitigate its health impacts. Traditional receptor models face challenges including extensive data dependency, technical complexities, and computational burden. Here we develop a machine learning (ML) based source apportionment model that leverages diverse aerosol composition measurements across a wide range of spatiotemporal scales. Particularly in the near-real time fashion, the model enables swift PM_2.5 source apportionment using high temporal resolution online measurements on certain sites, supporting agile policy responses. On the larger spatial and temporal scales, the model also accurately quantifies PM_2.5 sources for two decades in the Pearl River Delta region in China and the state of California in U.S., exhibiting strong generalization capability. For the megacities, the ML model results reveal that Shenzhen, China, experienced a significant decline in PM_2.5 over the past decade due to the successful control of anthropogenic sources, while Los Angeles, U.S., witnessed a flattened PM_2.5 trend under the joint effects of the reduced coal combustion and the exacerbated climate-driven wildfire pollution. This study highlights the potential of ML in air pollution research and policy-making toward environmental sustainability.