Using Machine Learning for Post-Simulation Diagnostics of Microphysical Process Rates with the ICON Model

Clouds play an important role in the hydrological cycle and significantly affect the Earth's radiative budget. Cloud microphysics describes the formation and interaction of individual cloud and precipitation particles. Its representation in numerical weather prediction and climate models remains challenging. Due to their sub-scale nature, microphysical processes need to be represented by parametrization schemes, which often rely on simplifying assumptions. Furthermore, the large number and variety of cloud particles and numerous nonlinear interactions thereof render cloud microphysics extremely complex. Many of its aspects are poorly understood, and a comprehensive theoretical description does not yet exist. Detailed information about microphysical process rates is essential in order to establish a profound understanding of the microphysical pathways, feedback loops and aerosol-cloud interactions. In the ICON model, cloud microphysics is often parametrized with the two-moment scheme, developed by Seifert and Beheng (2006), with six hydrometeor categories. However, due to the high number of processes, including the microphysical process rates in the model output results in approximately 20-50 additional three-dimensional output variables. If this output is generated in every time step of the model, this quickly requires immense storage capacities. Machine learning (ML) opens the possibility of generating on-demand offline diagnostics based on standard output variables as an alternative approach. Based on the two-moment bulk microphysics scheme, we trained a neural network to emulate the calculation of the process rates in the ICON model for warm-rain formation, reproducing earlier results of Seifert and Rasp (2020). As input, we use cloud and atmospheric state variables. We conducted simulations with the ICON model in a global configuration with 13 km grid spacing in order to generate training and validation datasets. In the initial stage of model development, this resolution seems sufficient, however, we plan on using a smaller grid spacing in a limited-area configuration to improve the accuracy of our results. We performed analyses using Mutual Information to unveil the dependencies between model variables and process rates and chose the predictors of the model accordingly. We compare different sets of predictors and activation functions in order to improve the model's predictiveness. Furthermore, we discuss the possibility of constructing a similar model for processes in mixed-phase and ice clouds.