A new calibration approach based on machine learning: application to the ALADIN regional climate model

Summary

The physical parameterizations used in regional climate models, such as radiation or convection, require approximations that introduce a number of parameters that are generally not directly observable and therefore difficult to constrain, generating considerable uncertainties. The adjustment of these parameters, known as calibration or tuning, is a key step in climate modeling, aimed at building models that are more consistent with observations. Automatic calibration approaches have recently been developed for global models, but the issue of calibration is still rarely addressed in the regional climate modeling literature. The aim of this work is to use a new calibration approach based on machine learning, applied to the ALADIN regional climate model developed at CNRM.

In the presentation, we will first detail this approach, then present the results obtained for the ALADIN model over the Euro-CORDEX domain at a horizontal resolution of 12.5 km. We will compare past simulations (driven by ERA5, 1979-2020) with the existing version of ALADIN63 used in previous Euro-CORDEX simulations, and with the new, calibrated version.