Machine learning (ML) enhanced cloud and aerosol representations in Energy Exascale Earth System Model (E3SM)

Representations of cloud and aerosol processes, along with aerosol-cloud interactions, are critical yet challenging aspects of Earth system models (ESMs). Due to the complexity and computational demands of these processes, ESMs often rely on crude parameterizations or simplified assumptions, leading to significant uncertainties, particularly in climate projections where aerosol-cloud interactions are the largest source of uncertainty. In the Energy Exascale Earth System Model (E3SM), simple relationships are employed for the hydrometeor collision-coalescence process, and a uniform mixing assumption is used for aerosol mixing states. Explicit simulations of these processes are computationally prohibitive in ESMs.

Machine learning (ML) offers a solution by emulating the detailed process model results in a computationally efficient manner. In this presentation, we will present our work in developing and incorporating emulators for liquid drop collision processes and aerosol mixing state to E3SM. The ML emulator for liquid drop collision processes is built with XGBoost based on detailed calculations with the bin approach (size-resolving). This emulator uniquely unifies four processes: autoconversion to form rain, rain accretion, droplet and rain self-collections and are applicable to both 2-moment and 3-moment bulk microphysics schemes. We apply this emulator to the Particle Predicted Properties (P3) scheme in E3SM v3 and present notable findings. To address the overly simplistic aerosol mixing state assumption in E3SM, we have implemented a pre-trained XGBoost ML model, developed using the PartMC-MOSAIC particle-resolved model, to predict aerosol mixing states. This approach results in more realistic aerosol properties, including hygroscopicity, which have significant implications for both direct and indirect aerosol forcing.