A radiative transfer deep learning model coupled into WRF with a generic fortran torch adaptor

Abstract

Advances in deep learning have created new opportunities for improving traditional numerical models. As the radiation parameterization scheme is crucial and time-consuming in numerical models, researchers sought to replace it with deep learning emulators. However, progress has been hindered at the offline emulation stage due to the technical complexity of the implementation. Additionally, the performance of the emulators when coupled with large-scale numerical models has yet to be verified. In this paper, we have developed a new tool called the Fortran Torch Adaptor (FTA) to facilitate this process and coupled deep learning emulators into the WRF model with it. The performance of various structured AI models was tested in terms of accuracy, generalization ability, and efficiency in different weather forecasting scenarios. Our findings revealed that deep learning models outperformed ordinary feedforward neural networks (FNN), achieving greater accuracy both online and offline, and leading to better overall forecasting results. When it came to unusual extreme weather events, all models were affected to some extent, but deep learning models exhibited less susceptibility than other models. With the assistance of FTA, deep learning models on GPU could achieve significant acceleration, ranging from 50x to 300x depending on the parameterization scheme replacing strategy. In conclusion, this research is crucial for both the theoretical and practical development of radiation transfer deep learning emulators. It demonstrates the emerging potential for using deep learning-based parameterizations in operational forecasting models.