Efficient machine learning frameworks for prediciting China’s future air quality trends

Fine particulate matter (PM_2.5) and ozone pollution cause a significant number of premature deaths in China each year. Evaluating the effects of emission mitigation and climate change on air quality with climate-chemistry models is computationally expensive. In this study, we develop two machine learning models — a deep learning framework based on U-Net image segmentation and long short-term memory (LSTM) neural networks, and an extreme value model — to predict China’s future air quality trends. These models are trained using model simulation results from 2014 to 2022 under high to low anthropogenic emission scenarios. The deep learning model yields promising results for PM_2.5, with an R² of 0.79 and root mean squared error (RMSE) of 12.58 ug/m³. The extreme value model for ozone episode exceedance rates achieves an R² of 0.97 and RMSE of 1.43%. This work demonstrates the potential of deep learning and extreme value models in efficiently modeling air quality, offering robust tools for future air quality assessments.