How well do downscaled Global Climate Models represent SMB in Antarctica?

Surface Mass Balance (SMB) is a critical forcing of the long term contribution of the Antarctic Ice Sheet (AIS) to sea level. While most GCMs do not produce SMB as an output, several RCMs, including a new ensemble produced in the PolarRES project do and are used to provide forcing for ice sheet models. There are significant spatial variations between regional climate models forced with reanalysis. RCMs also inherit biases from forcing GCMs when run for historical and future climate pathways which may exacerbate or cancel out biases within the RCM. Since the previous intercomparison, there has been significant regional model development and an expansion of datasets that can be used for evaluating these. We assess the SMB estimates of downscaled GCM and reanalysis simulations over Antarctica, compared to a new updated observational database for the historical period. Our ensemble compares 19 SMB products, generated from 4 different GCMs downscaled by 6 RCMs and SMB models, to observations of SMB gathered in the 2024 SUMup dataset. As 8 datasets do not explicitly calculate SMB, we approximate SMB by subtracting evaporation and melt from precipitation with these models. The various simulations are all pan-Antarctic at resolutions from 11 to 27 km, and span periods from 24 to 64 years, between 1950 and 2023. Fidelity of models to observations varies from product to product. As would be expected given they assimilate observational data, reanalyses perform better overall, with minor biases, whereas climatological SMB from GCM-forced runs  are usually too dry (5/9 GCM, 5/19 total). This is a significant bias in GCMs that will have an impact on modelled future evolution of the Antarctic Ice Sheet. One notable exception is the HIRHAM RCM forced by UKESM. In this case opposing biases appear to cancel out, giving the lowest t-statistic and one of the highest correlation coefficients in the intercomparison, while having the most comparison points due to the length of the simulation. The mean yearly accumulation of the models is 2100 Gt/year on the grounded AIS (Zwally’s mask) with most models predicting about 2000 Gt/year and 3 potential outliers predicting over 2500 Gt/year. Our analysis demonstrates that assessing model performance based on reanalysis driven simulations may provide misleading evidence of model performance for future projections. There are still large divergences in the spatial variability of modelled SMB. We also show the need for observational data with a wide spread in time and space.