Machine learning algorithm facilitates fast derivation of light-scattering properties of sea salt aerosol

Abstract

Accounting for the impacts of ambient relative humidity (RH) on the optical properties of sea salt aerosols usually requires complex and time-consuming light-scattering computations. To facilitate the process, this study employs machine learning algorithms to gain fast access to the single-scattering properties (including extinction efficiency, scattering efficiency, single-scattering albedo, and asymmetry factor) of sea salt aerosols at arbitrary particle sizes, RHs and wavelengths. The sea salt particles are modeled as either coated-spheres or homogeneous spheres depending on the ambient relative humidity, and their scattering properties are calculated using the invariant imbedding T-matrix method at selected sizes, wavelengths, and RHs to establish a data set for the training and validation purposes. Extensive tests using various machine learning methods and hyperparameter optimizations are implemented. It is demonstrated that the Gradient Boosting Decision Trees method optimized with Optuna tuning outperforms the other methods in predicting the scattering properties of sea salt aerosols. The developed model is promising for radiative transfer applications involving sea salt aerosols and similar approach could be potentially applied to the other scenarios where quick access to the aerosol optical properties is desirable.