Performance investigation of machine learning for optical modeling of black carbon aerosols with different coatings and humidities

Abstract

Optical properties of black carbon (BC) aerosols are essential for assessing atmospheric radiation and climate change, but fast and accurate optical calculation for BC with complex structures is still a challenge. In this study, a quick optical modeling method for BC coated by different materials under various humidities is developed based on the multiple-sphere T-matrix (MSTM) simulation and support-vector-machine (SVM) algorithm. The typical closed-cell, coated-aggregate, and partially-coated BC models with BC monomers ranging from 50 to 2000 and BC volume fractions ranging from 0.05 to 0.70 are employed, and the relative errors (REs) and determination coefficients (R²) are used to investigate the prediction performance of optical properties based on SVM. Results show that the SVM predicted optical properties, such as optical efficiency, asymmetry factor, single scattering albedo, and lidar ratio, overall agree well with the properties simulated using MSTM. The prediction performance could be influenced by coating structure and morphological parameters, most of the values of REs and R² were smaller than about 20 % and larger than about 0.85, respectively. The relative humidities (RHs) and coating materials slightly deteriorate the performance of SVM. With RHs increasing from 0 % to 95 %, or the coatings become non-absorbing organic carbon (OC) and brown carbon (BrC), the largest REs increase to over 30 %, while most of the R² is still larger than 0.85. This study presented a promising optical modeling method for BC aerosols, and it could be further improved with the development of machine learning.