Streamlining hyperparameter optimization for radiation emulator training with automated Sherpa

Abstract

This study aimed to identify the optimal configuration for neural network (NN) emulators in numerical weather prediction, minimizing trial and error by comparing emulator performance across multiple hidden layers (1–5 layers), as automatically defined by the Sherpa library. Our findings revealed that Sherpa-applied emulators consistently demonstrated good results and stable performance with low errors in numerical simulations. The optimal configurations were observed with one and two hidden layers, improving results when two hidden layers were employed. The Sherpa-defined average neurons per hidden layer ranged between 153 and 440, resulting in a speedup relative to the CNT of 7–12 times. These results provide valuable insights for developing radiative physical NN emulators. Utilizing automatically determined hyperparameters can effectively reduce trial-and-error processes while maintaining stable outcomes. However, further experimentation is needed to establish the most suitable hyperparameter values that balance both speed and accuracy, as this study did not identify optimized values for all hyperparameters.