Application of Deep Learning to Enhance the Computation of Phase Matrices of Nonspherical Atmospheric Particles Across All Size Parameters

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Xi, Yue, Bi, Lei, Lin, Wushao
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• DOI: 10.1029/2025JH000628
• ISSN: 2993-5210

Single‐scattering properties of nonspherical particles are essential to atmospheric radiative transfer and remote sensing studies. In particular, full scattering phase matrices are indispensable for simulating polarized radiative transfer and polarimetry‐based remote sensing. However, accurately computing these optical properties of nonspherical aerosols across a complete spectrum of size parameters ranging from Rayleigh to geometric optics domains poses a significant challenge. Prior investigations have demonstrated the efficacy of the machine learning approach in computing the optical properties (extinction efficiency, single‐scattering albedo, and asymmetry factor) of nonspherical particles based on the optical property database generated from a combination of exact numerical methods for smaller size parameters and physical‐geometric optics approximation methods for large size parameters. In this study, we successfully applied the machine learning method to compute six nonzero scattering matrices at scattering angles from 0 to 180° for a broad range of size parameters. We found that the deep learning method can significantly enhances the accuracy of the scattering phase matrix in a size parameter range that bridges the exact results computed from the invariant imbedding T‐matrix method and the approximate results from the improved geometric optics method. Furthermore, we have developed a deep neural network capable of effectively computing the scattering matrices of super‐spheroidal particles with a wide range of size parameters, shape parameters, and refractive indices, thereby facilitating their application in polarimetry‐based remote sensing and radiation transfer models. Accurately simulating how nonspherical aerosols scatter light is essential for atmospheric radiative transfer and remote sensing studies. However, calculating these scattering properties across a broad spectrum of particle sizes and wavelengths—spanning all possible size parameters—remains a significant challenge. To solve this, machine learning techniques have been explored in this study. By training a deep neural network model on a comprehensive database that merges precise numerical simulations and physical‐geometrical optics approximations, a new model can now predict the optical properties of nonspherical aerosols for any size parameters. The approach not only improves accuracy for mid‐sized parameters but also works well for particle parameters not in the training data. This machine learning‐based method enhance our ability to study radiative transfer and has potential applications in related fields. The deep neural network (DNN), trained using the phase matrices of small size parameters, showed exceptional extrapolation accuracy for medium‐sized parameters The DNN integrates invariant imbedding T‐matrix method and improved geometric optics method calculation results for small and large size parameters, respectively, ensuring smooth transition and outperforming the traditional method The DNN offers an accurate and efficient method for computing phase matrices of nonspherical particles across all size parameters