DeepPE: Emulating Parameterization in Numerical Weather Forecast Model Through Bidirectional Network

Abstract

To make weather/climate modeling computationally affordable, subgrid-scale physical processes in the numerical models are usually represented by semi-empirical parameterization schemes. For example, planetary boundary layer (PBL) parameterizations are used in atmospheric models to represent the diurnal variation in the formation and collapse of the atmospheric boundary layer—the lowest part of the atmosphere. We consider the problem of developing an accurate alternative to physics-based PBL parameterizations for speeding up the operation of atmosphere modeling. Our contributions are twofold. The first contribution is to propose a deep neural network emulator, called DeepPE, that focuses on simulating nonlocal closures in the PBL to capture cross-layer large eddies. We also explore a transfer method to maintain accuracy when applying a trained model to systems with different external forcing. We provide a comparison with three data-driven approaches as well as multi-task fine-tuning in predicting the PBL vertical profiles outputted by the Yonsei University (YSU) parameterization in the Weather Research Forecast (WRF) climate model over 16 locations. The experiment results show that our method can better simulate the vertical profiles within the boundary layer of velocities, temperature, wind speed, and water vapor over the entire cycle. And they also indicate that it achieves a comparable generalization performance with less computational cost.