Machine Learning Prediction of Long-term Variations in Cloud Condensation Nuclei in the upper boundary layer of North China

The scarcity of field observations of cloud condensation nuclei (CCN) limits effective constraints on aerosol–cloud interactions. While a small number of recent studies have explored machine learning approaches based on aerosol chemical and optical characteristics, even fewer have explicitly included particle number size distributions (PNSDs). Here, we developed an observation-driven model based on XGBoost to predict CCN number concentrations (N_CCN) by incorporating PNSDs and auxiliary variables. The model exhibits robust performance on the test dataset at supersaturations (SS) of 0.2%, 0.4%, and 1.0% (R² = 0.91–0.92; RMSE = 235–381 ppbv), demonstrating excellent capability in capturing the temporal variability of N_CCN. PNSDs are identified as the most influential features for N_CCN prediction using the SHapely Additive exPlanation (SHAP) approach, with the dominant size range shifting from 100–150 nm at SS ≤ 0.4% to 50–100 nm at 1.0% SS. The XGBoost model was further employed to reconstruct the long-term variations of N_CCN in the upper boundary layer over North China during 2007–2025. Our results show that N_CCN predominantly ranges from 866 to 2104 cm^-3, with higher values in spring and winter but enhanced activation ratios in summer and autumn. Interannual variability beyond seasonal influences indicates that N_CCN exhibits pronounced interannual fluctuations, largely driven by changes in highly oxidized particle sources. In contrast, overall aerosol hygroscopicity and activation ratio exhibit a gradual decline. The proposed XGBoost framework not only extends long-term N_CCN records but also provides new mechanistic insights into CCN activation, thereby reducing uncertainties in the assessments of aerosol-cloud interactions.