Emulating Surface Mass Balance in the Antarctic Peninsula Under Changing Forcings: A Transferability Assessment

Regional climate models (RCM) with good snow schemes can provide high-resolution surface mass balance (SMB) estimates over Antarctica, but come with high computational costs limiting the number of realizations across Earth System Models (ESM) and shared socioeconomic pathways (SSP). SMB’s great local variability as observable in regions characterized by complex topography such as the Antarctic Peninsula, requires high resolution. Recent machine learning approaches have shown promising downscaling results as cost efficient alternative to RCMs. However, their limited transferability across forcings remains a drawback for practical application.

Here, we investigate transferability of a Convolutional Neural Network (CNN) based emulator predicting SMB over the Antarctic Peninsula utilizing multiple SSPs. Two training scenarios are compared: a) a perfect model setup where training and prediction are performed under the same SSP, and b) an imperfect model setup in which the emulator is trained on a higher emission scenario and applied to a lower emission scenario. The latter represents an interpolation along SSPs, therefore comparable performance to the perfect model setup is expected.

Preliminary analyses suggest consistency in large-scale statistics over the full domain with benchmarks set in earlier studies. However, pronounced local variability in model performance is observable, particularly in regions of high melt or precipitation. Ongoing work aims to quantify these differences, detect potential causes, and their implications for predictions in the imperfect model setup.

Positive results could enable the generation of additional SSP-runs from existing RCM simulations, therefore substantially reducing total computational cost relative to the number of predictions.