Perks and perils of objectively calibrating regional Earth system models

The dynamical downscaling of global climate scenarios with coupled regional Earth system models is one of the most important tools to deliver accurate and applicable predictions for communities. However, uncertainties related to model setups are often systematic and exhibit spatiotemporal structures. Calibrating such systems and quantifying systematic model uncertainties thus remain a central challenge, as bias corrections typically violate fundamental physical principles,  such as mass and energy conservation, and may lead to unrealistic local climate sensitivity.

Here, we apply a range of objective calibration strategies to identify optimal model setups for the individual components of the Baltic Region Earth System Model, IOW-ESM. In particular, model sensitivities to changes within a set of parameters are identified using perturbed parameter ensembles and various types of surrogate models. It is thus possible to optimize the parameter set regarding a chosen set of metrics.

Comparing a range of surrogate models and optimization techniques, we find that all optimization strategies have limits in their applicability. In particular, the danger of overfitting the parameter sensitivities is large, and global optimization algorithms in high-dimensional spaces tend to find non-optimal local error minima. Keeping these limits in mind, we were able to significantly reduce model biases of the downward-directed shortwave radiation at the surface and improve the 2-meter temperature in the Baltic Sea region.