Explainable Machine Learning Prediction of Long-Term Variations in Cloud Condensation Nuclei at a High Mountain Site in the North China Plain

Abstract

Cloud condensation nuclei (CCN) regulate cloud formation and climate, yet long-term variability remains poorly constrained due to sparse observations and complex aerosol–CCN relationships. Here, we developed an observation-driven framework based on the Extreme Gradient Boosting (XGBoost) model to reconstruct CCN number concentrations (NCCN) at a high mountain site in the North China Plain, from 2007 to 2025. Particle number size distribution (PNSD) was identified as the dominant driver of NCCN variability, accounting for 55.7%–66.3% of the predictive importance, with the most influential size range shifting from 100–200 nm at 0.2% and 0.4% supersaturation (SS) to 50–100 nm at 1.0% SS. Although the reconstructed NCCN showed no monotonic trend, both the bulk activation ratio and hygroscopicity parameter decreased during summer from 2009 to 2025, indicating weakened particle hygroscopicity and CCN activation potential. By integrating positive matrix factorization with XGBoost, we further attributed interannual CCN variability to source-driven changes. Photochemically aged particles remained a persistent contributor, while the influence of SO2 decreased substantially, reflecting the reduced importance of sulfur-related processes in controlling CCN variability. In contrast, the contributions associated with new particle formation and particle growth processes became increasingly important over time. This study provides a robust reconstruction of long-term CCN variability and helps reduce uncertainties in aerosol–cloud–climate interactions.