Using convolutional neural networks to predict the optical properties of coated black carbon

Black carbon (BC) is a type of light absorbing substances in atmosphere, which has a significant impact on regional and global radiation balance. When evaluating the climatic effects of BC, the complex morphology of BC poses a challenge for large-scale climate models to proceed with the calculations. In this study, we developed a deep convolutional neural network (CNN) that combines residual links with one-dimensional multi-scale dilated convolutions. Using the multiple sphere T-matrix (MSTM), the extinction efficiency (Qext), absorption efficiency (Qabs), scattering efficiency (Qsca), and asymmetry factor (g) for coated BC were evaluated under different fractal dimensions, monomer radii, refractive indices, and incident wavelengths, by considering the volume fraction of thinly coated BC and thickly coated BC in the range from 20 % to 60 % and 2 % to 10 %, respectively. The optical properties of small particles were then treated as the training set to train the CNN, and the trained CNN was used to output the optical properties of large particles. By comparing the Qext, Qabs, Qsca, and g predicted by the CNN with those obtained from the MSTM, we found that the CNN has superior predictive capabilities for the optical properties of coated BC, and the newly established CNN exhibited broad applicability in predicting the optical properties of coated BC. Although relative deviations are observed in predicting the optical properties of small particles using the CNN, the errors for large particle predictions are essentially <1 %, with the mean absolute errors and root mean square errors being lower than 0.029 and 0.043, respectively. This study demonstrates that the CNN has great potential for further development. Future research should focus on how to use less data to predict more accurate results for the range of computational parameters for BC.