Usefulness of Automatic Hyperparameter Optimization in Developing Radiation Emulator in a Numerical Weather Prediction Model

To improve the forecasting accuracy of a radiation emulator in weather prediction model over Korean peninsula, learning rate used neural network training was automatically optimized using Sherpa. The Sherpa experiment results were compared with two control simulation rates 0.0001 and 1 for different batch sizes (full to 500). In offline evaluation, showed significant improvements predicting longwave/shortwave heating fluxes lowest results, whereas highest relatively small because values by 0.4756–0.6656. online evaluation one month, which linked model, demonstrated usefulness on universal performance emulator. particular, at full size, contributed reducing one-week forecast errors fluxes, skin temperature, precipitation 39–125%, 137–159%, 24–26%, respectively, simulations. Considering widespread use parallel based batch, can contribute producing robust regardless developing emulators.