Vertical distribution of atmospheric pollutants is critical for refined air quality management, yet continuous vertical observations and traditional numerical models remain limited in resolving vertical structures. Here, long-term vertical observations from a tower building and lidar measurements are integrated to develop a unified machine-learning framework for multipollutant vertical profiling and forecasting. Results show that meteorological factors substantially enhance 72 h O3 forecast performance, with an average increase in correlation coefficient of up to 0.29. Boundary layer height (BLH) and vertically integrated temperature contribute more than 36% in the upper layers. In the lower layer (0-0.7 km), O3 formation and dilution are dominated by near-surface temperature and BLH, whereas above 1.6 km, atmospheric thermal structure and vertical transport efficiency govern pollutant distributions. The multilayer interactive machine-learning framework further reveals pronounced asymmetric interlayer dependencies: the lower layer reduces mean absolute error in the middle and upper layers by 3.54% and 6.44%, respectively, while the middle layer introduces interference, and upper-layer information preferentially enhances middle-layer predictability but suppresses lower-layer forecasts. These findings provide new insights into meteorological regulation and interlayer coupling of pollutant vertical structures and offer a new framework for three-dimensional air quality forecasting and environment-oriented AI applications.
Lü et al. (2026) studied this question.