Analytical Prediction of Scattering Properties of Spheroidal Dust Particles With Machine Learning

Abstract

A neural network (NN) model is trained with a database widely used in the aerosol remote sensing community to rapidly predict the single-scattering optical properties of spheroidal dust particles. Analytical solutions for their Jacobians with respect to microphysical properties are derived based on the functional form of the NN. The Jacobian predictions are improved by adding Jacobians from a linearized T-matrix model into the training. Out-of-database testing implies that NN-based predictions perform better than the business-as-usual method that interpolates optical properties from the database. Independent validation further demonstrates the efficacy of the NN-based predictions by reducing computational costs while maintaining accuracy. This work represents the first use of machine learning-based function approximation to computationally expedite the application of the existing spheroidal dust properties database; the resultant NN model can be implemented in atmospheric models and satellite retrieval algorithms with high accuracy, computational efficiency, and the rigor of analytical solutions.