Improved weather forecasting using neural network emulation for
radiation parameterization
Abstract
In this study, a neural network (NN) emulator for radiation
parameterization was developed for the use of an operational weather
forecasting model in the Korea Meteorological Administration. The
development of the NN emulator was based on large-scale training sets
and 96 categories (longwave–shortwave, months, land–ocean, and
clear–cloud). As the radiation parameterization was replaced by the NN
emulator, a 60-fold speedup for the radiation process was achieved, with
a decrease of 87.26% in the total computation time. The accuracy of the
NN emulator was strictly evaluated through comparison with the results
obtained from the infrequent use of the original radiation scheme with
the same computational cost. The mean errors of the NN radiation
emulator were significantly reduced by 21–34% compared with the
infrequent method. The combination of using the NN radiation emulator
and applying it infrequently provided an additional speedup of up to
36-fold, corresponding to 2180 times speedup compared with the control
run, without a significant reduction in accuracy. The optimized
structure for the radiation emulator designed in this study also showed
universal robustness even in the use of limited training sets with
incomplete coverage. In conclusion, the NN radiation emulator in this
study provides benefits regarding both accuracy and computational cost,
making it useful for improving weather forecasting modeling.