Prediction of CCN spectra parameters in the North China Plain using a random forest model

Abstract

Cloud condensation nuclei (CCN) spectra, that indicate the dependence of CCN number concentrations (NCCN) on supersaturation, are important factors in aerosol–cloud interactions. Owing to the lack of direct measurements of NCCN, calculating the CCN spectra based on conventional observational data of aerosols is important to obtain NCCN; however, this is challenging owing to the complex relationship between aerosol properties and CCN activity. Machine learning techniques have recently been applied to estimate NCCN and found to be a promising method for CCN prediction. In this study, the random forest (RF) model was applied to predict CCN spectral parameters using the observation data measured in four campaigns in the North China Plain, and the effects of chemical and optical properties of aerosol on the estimation of CCN spectra were investigated. The results show that the RF model trained with the data of one campaign can be used to estimate CCN spectral parameters in another campaign, with the coefficient of determination between the estimated and measured CCN spectral parameters being approximately 0.5. The deviations of the estimation by the RF model may result from the difference in both the aerosol properties and measurement uncertainties among different campaigns, whose influence on the deviations can be further magnified by overfitting the RF model. Further analysis revealed that the major aerosol properties among the input variables of the RF model were the mass concentration of black carbon and aerosol hemispheric backscattering fraction at a wavelength of 450 nm. In addition, the roles of chemical compositions of aerosol in estimating CCN spectra parameters are different among different campaigns because the poor correlation between aerosol chemical compositions and CCN spectra parameter can affect the performance of the RF model. For estimating CCN spectra, using aerosol optical properties, including the hemispheric backscattering fraction and absorption coefficient, as model inputs is more recommended than using aerosol chemical properties.