A Machine Learning approach to aerosol thermodynamics embedded in aglobal chemistry-climate model

The chemical composition of the aerosol phase is still a major uncertainty in global chemistry climate models. One the one hand, aerosol thermodynamics  calculations are needed to determine the chemical composition of the inorganic fraction of the aerosol particles, on the other hand these calculations are computationally expensive. However, to properly describe the combined gas and aerosol phase composition, e.g., the reactive nitrogen budget including HNO3 or chlorine displacement from sea-salt aerosol, it is mandatory to have a reasonable description of the aerosol thermodynamics. To overcome the computational costs, but to still obtain a reasonable degree of proper process description, a machine learning  approach for the aerosol thermodynamics might offer opportunities in CCM modelling.
In this study, we embed a machine learning approach for the description of aerosol thermodynamics in the chemistry climate model EMAC to reduce computational load (compared to explicit thermodynamics calculations) and show the capabilities of a modern computing approach, implemented in a multi-modal aerosol scheme.
The new aerosol thermodynamics scheme is formulated as a machine learning neural network, which has been trained with the help of an explicit inorganic aerosol thermodynamics box model, i.e. ISORROPIA-2.

This study presents first results of global 3D simulations using the ML approach and compares the results to explicit calculations in terms of the spatio-temporal distribution of the aerosol chemical composition as well as the effective performance of the modelling system.