Tuning a Climate Model with Machine-learning based Emulators and History Matching

Abstract. In climate model development, tuning refers to the important process of adjusting uncertain free parameters of subgrid-scale parameterizations to best match a set of Earth observations such as global radiation balance or global cloud cover. This is traditionally a computationally expensive step as it requires a large number of climate model simulations to create a Perturbed Parameter Ensemble (PPE), which is increasingly challenging with increasing spatial resolution and complexity of climate models. In addition, this manual tuning relies strongly on expert knowledge and is thus not independently reproducible. Here, we develop a Machine Learning (ML)-based tuning method with the goal to reduce subjectivity and computational demands. This method consists of three steps: (1) creating a PPE of limited size with randomly selected parameters, (2) fitting an ML-based emulator to the PPE and generate a large PPE with the emulator, and (3) shrinking the parameter space with history matching. We apply this method to the Icosahedral Nonhydrostatic Weather and Climate Model (ICON) for the atmosphere to tune for global radiative and cloud properties. With one iteration of this method, we achieve a model configuration yielding a global top-of-atmosphere net radiation budget in the range of 0,1 W/m2, and global radiation metrics and water vapor path consistent with the reference observations. Furthermore, the resulting ML-based emulator allows to identify the parameters that most impact the outputs that we target with tuning. The parameters that we identified as mostly influential for the physics output metrics are the critical relative humidity in the upper troposphere and the coefficient conversion from cloud water to rain, influencing the radiation metrics and global cloud cover, together with the coefficient of sedimentation velocity of cloud ice, having a strong non-linear influence on all the physics metrics. The existence of non-linear effects further motivates the use of ML-based approaches for parameter tuning in climate models.