Machine Learning Emulation of a Quasi Two-Moment Microphysical Scheme with Diagnosed Particle Sizes

In atmospheric numerical models, the dynamics of clouds are intricately linked to the microphysical schemes employed, with significant reliance on the assumptions underlying these schemes. This dependence is particularly pronounced in bulk microphysical schemes, which not only exhibit sensitivity to various assumptions but also struggle to accurately represent particle-size-dependent processes such as size sorting, coalescence, and evaporation. While statistical parameterizations derived through data fitting also carry uncertainties, they have the benefit of being able to readily incorporate data from diverse sources.

In this study, we develop and investigate a machine learning (ML) microphysical emulator based on two microphysical schemes. The first microphysical scheme is the two-moment Morrison scheme, while the second one is a Stochastic Bin Model (SBM) rooted in the Hebrew University Cloud Model. While the ML emulation of the Morrison scheme is rather straightforward, the emulation of an SBM is challenging and computationally intensive. However, the development of an SBM ML diagnostic model that predicts the mass weighted diameters of hydrometeors from their mixing ratios and other non-hydrometeor related variables is straightforward. We develop such a diagnostic model and combine it with an ML emulator of the Morrison microphysical scheme. The two-component ML microphysical emulator is incorporated in WRF and its performance is systematically investigated through the derivation of reflectivity Contoured frequency by Altitude Diagrams (CFADs) and joint distributions of Particle Size Distribution (PSD) parameters such as the generalized intercept and mass mean diameter. The simulated CFADs and intercept-size distributions are compared against their equivalents derived from real observations. The ML emulator's performance is also evaluated against that of the two-moment Morrison scheme.

Ultimately, this study contributes to a deeper understanding of the potential for machine learning techniques to emulate complex microphysical processes within numerical weather prediction models. By investigating the efficacy of ML-based emulators against established schemes, we pave the way for more accurate and efficient representations of cloud dynamics in meteorological simulations.