Abstract. Self-organizing maps (SOMs; a feature-extracting technique based on an unsupervised machine learning algorithm) are used to classify atmospheric boundary layer (ABL) meteorology over Beijing through detecting topological relationships among the 5-year (2013-2017) radiosondebased virtual potential temperature profiles. The classified ABL types are then examined in relation to near-surface pollutant concentrations to understand the modulation effects of the changing ABL meteorology on Beijing's air quality. Nine ABL types (i.e., SOM nodes) are obtained through the SOM classification technique, and each is characterized by distinct dynamic and thermodynamic conditions. In general, the selforganized ABL types are able to distinguish between high and low loadings of near-surface pollutants. The average concentrations of PM 2.5 , NO 2 and CO dramatically increased from the near neutral (i.e., Node 1) to strong stable conditions (i.e., Node 9) during all seasons except for summer. Since extremely strong stability can isolate the near-surface observations from the influence of elevated SO 2 pollution layers, the highest average SO 2 concentrations are typically observed in Node 3 (a layer with strong stability in the upper ABL) rather than Node 9. In contrast, near-surface O 3 shows an opposite dependence on atmospheric stability, with the lowest average concentration in Node 9. Analysis of three typical pollution months (i.e., January 2013, December 2015 and December 2016 suggests that the ABL types are the primary drivers of day-to-day variations in Beijing's air quality. Assuming a fixed relationship between ABL type and PM 2.5 loading for different years, the relative (absolute) contributions of the ABL anomaly to elevated PM 2.5 levels are estimated to be