Deep learning representation of the aerosol size distribution

Abstract

Aerosols influence Earth's radiative balance via the scattering and absorbing of solar radiation, affect cloud formation, and play important roles on precipitation, ocean seeding and human health. Accurate modeling of these effects requires knowledge of the chemical composition and size distribution of aerosol particles present in the atmosphere. Computationally intensive applications like remote sensing and weather forecasting commonly use simplified representations of aerosol microphysics, prescribing the aerosol size distribution (ASD), introducing uncertainty in climate predictions and aerosol retrievals. In this work, we develop a neural network model, MAMnet, to predict the ASD and mixing state for seven lognormal modes based on the bulk aerosol mass and the meteorological state. MAMnet is designed to operate with outputs from single-moment, mass-based aerosol schemes, making it compatible with existing models. We demonstrate that MAMnet can accurately reproduce the output of a two-moment modal aerosol scheme, and also agrees well with field measurements when driven by reanalysis data. Our model paves the way to improve the representation of aerosols in atmospheric models while maintaining the versatility and efficiency required in large scale applications.