Nuanced understanding of aerosol effects on climate requires an adequate representation of aerosols in Earth System Models (ESM). Aerosol mixing state, which is critical for the representation of climate-relevant aerosols properties in ESM, describes the distribution of different aerosol species among the aerosol particles in a population. The most rigorous representation of aerosol mixing state can be attained with a Particle-resolved Monte Carlo (PartMC) model but at large computational cost. In contrast, ESM use simplified representations for mixing state, which may introduce errors in the estimation of the aerosol impact on climate.

The aim of our study is to develop and evaluate machine learning (ML) techniques to represent aerosol mixing state metrics in the US DOE Energy Exascale Earth System Model (E3SM) at the global scale. This will allow us to estimate where the current E3SM aerosol treatment introduces errors in the calculation of climate-relevant aerosol properties. Scenario libraries of particle-resolved simulations were used to create a large number of aerosol populations with different compositions and different mixing states, which we used as training and testing datasets. A subset of features, common to both E3SM and PartMC, were selected as inputs for developing ML-enabled models for estimating mixing state metrics.

Our data-driven ML model leverages deep learning and XGboost. We employed rigorous model selection techniques, in conjunction with a suitable model from deep learning, to develop an ensemble approach that enables us to estimate the global spatiotemporal distribution of aerosol mixing state metrics relevant to the E3SM Atmosphere Model (EAM) Version 1.0 at 1 degree nominal spatial resolution. Our framework is designed to further advance the fundamental understanding of the representation of aerosol at a global scale. The novel capabilities help overcome some of the current limitations in next-generation earth system modeling.