Rapid emulation of Regional Climate Models via deep learning for impact studies: evaluating inter-GCM generalization

Climate change impact studies (e.g. hydrology, agriculture, health) systematically require high-resolution meteorological variables. However, these disciplines rarely need the full suite of physical variables resolved by computationally expensive dynamical Regional Climate Models (RCMs). Their primary need lies in obtaining reliable projections and, crucially, a wide diversity of simulations to properly characterize climate uncertainty. To address this challenge, statistical emulation via deep learning emerges as a computationally efficient alternative.

This study proposes the use of UNet-based neural architectures as direct emulators of EURO-CORDEX RCMs for high-resolution temperature (EUR-11). The central research question is highly operational: is it possible to train an emulator using a single historical GCM-RCM pair and rely on its capacity to regionalize future projections driven by Global Climate Models (GCMs) different from the one used during its training?

To answer this, the model undergoes a rigorous generalization test against the CORDEX ensemble. Once the network is trained specifically using the historical MOHC-HadGEM2-ES and DMI-HIRHAM5 pair, the emulator is fed with the boundary conditions of all other GCMs available for the DMI-HIRHAM5 regional model in the European repository. The validation assesses whether the emulator's outputs, when forced by these new GCMs, deviate significantly from the original dynamical CORDEX projections for those exact pairs.

Preliminary results show that the dispersion (uncertainty spread) generated by the emulator's inference ensemble is equivalent to that of the original CORDEX dynamical ensemble. Nevertheless, zero-shot cross-evaluations reveal that the emulator systematically deviates from its dynamical "mirror" pair. These findings help define the viability of using neural networks to generate on-demand climate ensembles, highlighting both their potential for ultra-fast climate variability reproduction and the limitations associated with transfer biases between global models.

 

Acknowledgements: This work was supported by the ARUBA (PID2023-149080OB-I00/MCIN/EI/10.13039/501100011033), and the INSIEME (FSRM/10.13039/100007801) projects.