Emulating Microphysics Parameterization in Unified Forecast System for Hurricane Application

Microphysics parameterization is a crucial component in numerical weather models. It represents the collective effect of microphysical processes that govern hydrometeor (e.g. cloud water, cloud ice, rain, snow, graupel, hail) formation, growth, aggregation, and dissipation on very small scales in the atmosphere. The microphysics processes play important roles in many weather phenomena, ranging from everyday thunderstorms to extreme weather (e.g. hurricane, tornado, derecho, snowstorm), especially for related precipitation and floods. Due to the complexity associated with many microphysical processes, microphysics parameterization scheme is computationally expensive. Therefore, most operational models choose to use a simple one-moment bulk, despite that many studies have demonstrated the superior performance of higher moment (i.e., more than one moment) bulk microphysical parameterization schemes and other non-bulk microphysical parameterization schemes (e.g. bin microphysics scheme) in simulating microphysical processes.

In this study, we use a neural network to emulate a physics based microphysics scheme to increase the efficiency. The neural network is built upon Residual Neural Network (ResNet) units and constrained by mass conservation. 1-D convolutional layer is used in ResNet units to capture the vertical coherent structure in the atmosphere. 250 hurricane simulations for cases in 2020 and 2021 are used as training data and 50 hurricane simulations for cases in 2022 are used as testing data. The preliminary results show that the emulator captures the collective effect of various microphysical processes. The correlation coefficient of predicted values from the emulator against the values from the simulations are greater than 0.9 for all variables except surface precipitation of ice and graupel due to their rare occurrences in the training data. Our results indicate that neural-network microphysics emulator is a promising approach through which a much more sophisticated microphysics scheme can be adopted in operational models.