Improving the efficiency of Calibration in Earth System Modelling with Short-term Simulations

Abstract

Calibration is an important process in developing Earth system models to reduce the deviations between model and observations and improve model performance. Current calibration process is quite resource demanding, requiring enormous computing resources to explore model responses to different parameter choices. In this study, we explored a method using short-term simulations for training emulators to predict model responses given parameter changes in calibration to potentially improve its efficiency. We generated two Perturbed Parameter Ensembles (PPEs) using the same model configuration and perturbed parameters values, but differ in the length of simulation. A significant linear relationship between short and long-term simulations in model response to parameters was found through EOF analysis. The contributions of each ensemble member in respond to parameter values were very similar between PPEs ran for short and long-term, which provided a support for using short-term simulations for emulator in calibration. The evaluation of Gaussian Process emulator trained with short-term simulations showed that the emulator trained with short-term simulations could capture major spatial variability in long-term simulations and generally has a good performance in predicting model responses to parameters. Nevertheless, there are more fragmented and chaotic components in model responses in short-term simulations, which may be caused by the larger internal variability in a shorter length of simulation. Such characteristics in short-term simulations may lead to regional biases when assess model responses in regional scales or related to natural variations using emulators trained with short-term simulations. Given that short-term simulations require much less computational time to run, these results implied that it is possible to use short-term simulations to improve the efficiency of calibration in the Earth system modelling, while the potential regional biases should be aware of when assessing model responses on regional bases.