Development of four machine learning schemes used for moist physics parameterization in CMA-TRAMS

Moist physics parameterization in numerical weather prediction models has great influences on precipitation forecast. Conventional moist physics parameterization is complicated, computation-intensive and has great uncertainties. In this paper, four machine learning (ML) architectures, i.e., decision tree based light gradient-boosting machine (LightGBM), fully connected neural network (FC), convolutional neural network (CNN) and convolutional block attention module (CBAM), are developed for moist physics parameterization via extracting local information of model variables at each grid point. Simulation experiments are carried out for a typhoon process in the South China Sea. Results show that the four ML-based schemes can well simulate the thermal and moisture effects of the moist physics parameterization and can present the spiral structure of heat source and moisture sink related to typhoon convection. Large errors of potential temperature tendency and specific humidity tendency are found in the middle and lower troposphere respectively. All the four machine learning schemes deteriorate with forecast time. This paper provides a useful reference for the development of ML-based physical parameterization scheme.