Online calculation of aerosol optics in atmospheric models with machine learning

Aerosols affect weather and climate by absorbing and scattering radiation. Such effects strongly depend on the optical properties of aerosols that are mainly controlled by their other characteristics like size distribution, morphology and chemical composition. Chemistry and aerosol microphysics constantly modify these characteristics causing a large spatial and temporal variability. Most atmospheric models cannot account for this variability as they rely on look-up table to treat aerosol optics. This simplification can lead to large errors in weather and climate models when it comes to aerosol radiative impacts.  

This study presents a novel and computationally inexpensive machine learning approach for online representation of the aerosol optical properties. These properties are fully coupled with the chemical and microphysical variability of particles. Aerosol composition is considered with two ternary systems for solid (dust, soot and sea salt) and liquid (water, sulfate and organics) mixtures. Then Mie calculations are performed based on these aerosol compositions assuming core-shell and volume-average mixing states. The output of the Mie code is then used to train an artificial neural network. The results show that neural network model is able to predict the aerosol optical properties (extinction coefficient, single scattering albedo and asymmetry parameter) by R² >0.90 and O(10³ ) lower computational cost compared to Mie calculations. Potential applications of this approach for ICON-ART modeling system is discussed.