Interpretable ensemble learning unveils main aerosol optical properties in predicting cloud condensation nuclei number concentration

Variations in cloud condensation nuclei number concentration (N_CCN) significantly influence cloud microphysics, yet direct N_CCN measurements remain challenging. Here, we present an N_CCN ensemble learning (NEL) model utilizing ensemble learning and interpretability analysis on aerosol optical parameters. Validated at two land sites, two ocean sites and one polar site within the Atmospheric Radiation Measurement program, the mean absolute percentage error range of the NEL model across different environments is from 12% to 36%, demonstrating high accuracy. Key findings reveal that aerosol optical parameters can serve as predictors for N_CCN. Aerosol scattering and backscattering coefficients, absorption coefficient, backscatter fraction (BSF), and Ångström exponent (AE) are positively correlated with N_CCN, while single scattering albedo shows negative correlations. N_CCN prediction at land sites is highly sensitive to BSF, largely driven by the backscattering coefficient, as fine particles dominate in these sites. At ocean sites, N_CCN prediction is more sensitive to AE, primarily influenced by the scattering coefficient, due to the higher proportion of larger particles. At the polar site, N_CCN prediction shows sensitivity to both BSF and AE, mainly driven by the scattering coefficient, as polar sites are cleaner and contain larger particles. These differences reflect the variation in particle size and number concentration across different environments.