Use of a Neural Network with a Perturbation Forecast Model to Emulate the Effect of Moist Physics Parameterization.

To develop a novel moist physics parameterization scheme, this study analyzed Typhoon Mujigae in the South China Sea. The China Meteorological Administration's Tropical Region Atmospheric Model System, a regional numerical weather prediction model, was run using alternately activated and deactivated conventional moist physics parameterization schemes. The difference between the outputs of these runs formed a dataset used to train a fully connected neural network. This network predicts the temporal tendencies of potential temperature and specific humidity, representing the heating and drying effects of moist physical processes. A perturbation forecast approach was employed to isolate these moist physical effects from the influence of large-scale dynamical processes on heat and moisture transport. The results demonstrate that the trained neural network scheme successfully replicates the heating and drying features, primarily latent heat release, around the typhoon center. It exhibited spatial distributions of heat sources and moisture sinks comparable to those of the conventional scheme. The analysis revealed a key characteristic of typhoon convection: heat sources correspond to moisture sinks. Vertically averaged moisture sinks exceed the heat sources, indicating an excess latent heat release that necessitates balancing by radiative cooling. This study confirmed that a deep-learning moist physics scheme can effectively emulate traditional parameterization schemes, particularly for typhoons. [ABSTRACT FROM AUTHOR]