Identification of regional PM2.5 transport dynamics with an adaptive data-driven model

Abstract

Regional transport of pollutants imparts fluidity and non-local characteristics to PM_2.5 pollution, posing significant challenges for timely air quality management. Traditional approaches for identifying transport pathways rely on accurate and near-real-time wind field data, limiting their applicability for near-real-time analysis. In this study, a fully data-driven deep learning framework was developed to infer the major PM_2.5 transport channels without requiring precise wind field inputs. Applied to the Sichuan Basin, a region characterized by complex terrain and frequent haze events, the adaptive graph-generated adjacency structure learned from historical pollutant and meteorological data aligns closely with the prevailing directions of air mass movement. High-weighted stations are primarily located in upwind, neighboring, and downwind regions relative to the receptor site, consistent with observed diffusion patterns. The agreement with HYSPLIT results further confirms the model's ability to identify physically meaningful transport pathways. These findings indicate that, even in the absence of accurate wind field data, historical spatiotemporal variations in pollutant concentrations alone can reveal physically interpretable transport patterns. This approach enables near-real-time identification of regional transport dynamics and provides both methodological innovation and practical value for coordinated air quality management in complex basin environments.