Generalizability of the UNet-based bias correction to other CNRM-CM6 models

Determine whether the UNet-based neural bias correction model trained to map climatology-removed CNRM-CM6-1-HR AOGCM sea surface temperature and dynamic sea level projections to ORAS5 reanalysis fields can be effectively applied to other members of the CNRM-CM6 model suite, preserving correction skill and stability across differing model configurations and forcings.

Background

The paper develops a UNet-based deep learning model to correct biases in CNRM-CM6-1-HR AOGCM projections of sea surface temperature (SST) and dynamic sea level (DSL) over the Bay of Bengal by learning a mapping from climatology-removed model outputs to ORAS5 reanalysis. The approach demonstrably reduces RMSE and improves pattern correlation compared with the EDCDF statistical method in test years.

In the concluding section, the authors explicitly indicate uncertainty regarding whether this trained correction model can be used for other members of the CNRM-CM6 suite, which differ in components, parameterizations, and forcings. Establishing generalizability would clarify the scope of applicability of the proposed neural operator across related climate models.