Compound Parameterization to Improve the Accuracy of Radiation Emulator in a Numerical Weather Prediction Model

• Published in: Geophysical research letters Vol. 48; no. 20
• Main Authors: Song, Hwan‐Jin, Roh, Soonyoung, Park, Hyesook
• Format: Journal Article
• Language: English
• Published: Wiley 28.10.2021
• Subjects: emulator; neural network; radiation; RRTMG; stability; WRF
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2021GL095043

To improve the numerical weather prediction model over Korea using a neural network (NN) radiation emulator, two types of compound parameterization (CP) were developed. Although the CP returning to the original parameterization causes a considerable increase in the computational time, this increase can be compensated by the infrequent use of the radiation scheme, thus maintaining the 60‐fold speedup of the radiation process with the NN emulator. The first CP is based on the prediction of the heating rate error using the additional NN for all given input variables. In contrast, the second CP uses the cloud fraction to estimate the uncertainty of the NN emulator. As a result of model simulations for independent cases, including extreme flood events, the first CP was the most effective for passive use, whereas the second was useful in active use and exhibited the lowest error. Thus, these CP methods can help improve weather forecasting. Plain Language Summary Neural network emulators for radiation parameterization have been widely developed under numerical climate and weather prediction models, concerning ten times faster computation cost. The uncertainty resulting from extrapolating extreme data outside the training sets is a severe problem for the emulator. This uncertainty can cause a blowup of the entire model after long‐term simulation. This study developed two compound parameterizations to mitigate the uncertainty based on the return to the original scheme in case a significant error was expected. The developed compound parameterizations enabled significant reduction of the forecast error from the numerical weather prediction model based on the emulator. A discussion on the sensitivity results for the passive and active uses of compound parameterization is provided. Key Points Two methods were developed to complement the uncertainty of neural network radiation emulator The first method was based on the prediction of heating rate error and was effective in its passive use The second method based on cloud fraction showed the highest accuracy in its active use