Impact of horizontal resolution on the robustness of radiation emulators
in a numerical weather prediction model
Abstract
Developing a radiation emulator based on machine learning in a weather
forecasting model is valuable because it can significantly improve the
computational speed of forecasting severe weather events. In order to
fully replace the radiation parameterization in the weather forecasting
model, the universal applicability of radiation emulator is essential,
indicating a transition from the research to the operational level. This
study addressed the universal issue of radiation emulators associated
with horizontal resolutions from the climate simulation scale (100 km)
to the cloud-resolving scale (0.25 km). All simulations were performed
using an emulator trained at 5 km simulation. In real-case simulations
(100–5 km), the forecast errors of radiative fluxes and precipitation
were reduced at coarse resolutions. The ideal-case simulations (5–0.25
km) also showed a similar feature with increased errors in heating rates
and fluxes at fine resolutions. However, all simulations maintained an
appropriate accuracy range compared with observations in real-case
simulations or the infrequent use of radiation parameterization in
ideal-case simulations. These findings demonstrate the feasibility of a
universal radiation emulator associated with different resolutions and
models and emphasize the importance of future development directions
toward the emulation of high-resolution modeling.