A Physics-Constrained Deep-Learning Framework based on Long-Term Remote-Sensing Data for Retrieving Vertical Distribution of PM_2.5 Chemical Components

Abstract. The vertical distribution of PM_2.5 chemical components is crucial for identifying the causes of atmospheric pollution and its impact on climate change and extreme weather. By integrating long-term lidar measurements, deep-learning algorithms and a physics-constrained optimization method, this paper presents a novel lidar-based retrieval framework to obtain vertical mass concentration profiles of PM_2.5 chemical components for the first time. Identifiable components include sulfate (SO₄^2-), nitrate (NO₃^-), ammonium (NH₄⁺), organic matter (OM) and black carbon (BC), which extend beyond the component types that traditional remote-sensing retrievals can identify. A 1-year retrieved surface mass concentrations of these components closely aligned with the observations, with Pearson correlation coefficient values ranging from 0.91 to 0.98. The retrieval framework applied in varying non-training spatiotemporal scenarios also showed robust generalization capabilities. Tower and aircraft-based field campaigns indicate that the retrieved and observed vertical profiles of these components exhibited consistent patterns in mass concentrations and proportions. Subsequently, an explainable method was incorporated into the retrieval framework to quantify the multivariate driving effects on vertical profile retrieval. Results showed that the extinction coefficient and representative indicators within physiochemical processes contributed significantly to mass concentrations of these components. Finally, a dataset of vertical mass concentration profiles of these components over six years in a Chinese megacity was generated by the retrieval framework, revealing the dominant roles of OM and NO₃^- in PM_2.5 throughout the entire boundary layer across all seasons. Through implementing clean air policies, the reduction rates of these components in the megacity exhibited the highest reduction rate of 0.17–0.82 µg m^-3 a^-1 occurring at an altitude of ~300 m. Our retrieval framework offers a novel approach for acquiring vertical profiles of PM_2.5 chemical components, thereby providing a new perspective on elucidating the vertical evolution of atmospheric pollutants.