Discrepancies between observations and climate models of large-scale wind-driven Greenland melt influence sea-level rise projections

While climate models project that Greenland ice sheet (GrIS) melt will continue to accelerate with climate change, models exhibit limitations in capturing observed connections between GrIS melt and changes in high-latitude atmospheric circulation. Here we impose observed Arctic winds in a fully-coupled climate model with fixed anthropogenic forcing to quantify the influence of the rotational component of large-scale atmospheric circulation variability over the Arctic on the temperature field and the surface mass/energy balances through adiabatic processes. We show that recent changes involving mid-to-upper-tropospheric anticyclonic wind anomalies – linked with tropical forcing – explain half of the observed Greenland surface warming and ice loss acceleration since 1990, suggesting a pathway for large-scale winds to potentially enhance sea-level rise by ~0.2 mm/year per decade. We further reveal fingerprints of this observed teleconnection in paleo-reanalyses spanning the past 400 years, which heightens concern about model limitations to capture wind-driven adiabatic processes associated with GrIS melt.


Uncertainties around quantifying tropical teleconnections' role in GrIS surface changes
The leading internal MCA mode is also moderately and significantly correlated with the summer GSI from ERA5, which suggests that fingerprints of this teleconnection can be found in the summer season (rZ500;GSI = 0.56 (0.47); rSST;GSI = 0.47 (0.36), with detrended values in brackets), despite known uncertainties imposed by the summer monsoon, which affects its stability 64 . Based on the correlations between the GSI and the global SST/Z500 patterns ( Supplementary Fig. 9) we speculate that the tropically-driven part of the local GSI changes might be up to 25% in annual means and about 15% in summer; however, these estimates are limited by inadequate representation of tropical SST variability in climate models. We note, that the use of available Pacific pacemaker experiments with CMIP5/6 climate models would provide minimal added insight since those experiments are forced by both specified SST (or nudged SST) and anthropogenic forcing. Currently there are no available pacemaker control simulations with fixed anthropogenic forcing, thus we would not be able to isolate the response of the model to imposed tropical SST forcing only. Additionally, based on previous calculations by the authors both CESM1 and CESM2 models have limilatitions in simulating the observed tropical rainfall trends in the Pacific, which is essential to the formation of tropical-Arctic teleconnections.

On the robustness of the MCA pattern seen in ERA5
We repeat the MCA calculations using the extended ERA5 reanalysis for the period 1950-2018. This analysis showed very similar spatial patterns to those seen in Fig. 4 in the main text and the expansion coefficient time series of the MCA as well as strong temporal agreement based on the extent to which the extended reanalysis and the shorter  period are correlated >0.95. This suggests that the PARC mode is a robust mode of atmosphere-ocean coupled variability. We note that the expansion coefficient time series from the extended ERA5 reanalysis also are highly correlated with the expansion coefficients derived from the MCA on the EKF400 database (~0.7).

Nudging run is robust in annual means
In the main text we have focused on the melt season (summer). However, we estimate that ~52% of the annual mean MAR GrIS SAT trends between 1990-2012 may be attributable to changes in the nudging experiment derived annual mean GSI with a similar spatial pattern to both our summer estimates and to the year-to-year regression of the GSI onto GrIS SAT or SMB using annual means ( Supplementary Fig. 5d,e). Also, the year-to-year correlation between the GSI and both the EBAF summer ( Supplementary Fig. 4a) and annual mean SEB ( Supplementary Fig. 4c) show marked spatial patterns with significant positive correlations over the Baffin Bay and in west Greenland in agreement with regions showing the most significant SMB and SAT changes since 1980 (Supplementary Fig. 3b-d).

Insensitivity of the linear trends to the time period
Regarding the sensitivity of our results based on linear trends to the time period chosen, we note that using 1980-2018 or 1980-2019/2020 did not affect our conclusions that are based on the 1990-2012 period. The nudging experiment explains a similar amount (~50%) of the observed Greenland surface air temperature trends during 1980-2018, 1990-2012 or 1980-2019/2020. Furthermore, as seen in Fig. 1 in the main text, the nudging experiment derived Greenland SAT or SMB in Fig. 2f both well resembles the observed 2012 high-and 2013 low-melt years, which indicates that the wind-driven mechanisms is at play at both interannual and decadal time scales.

LMR2.1 to verify EKF400 results
To further assess the robustness of the observed and EKF400 simulated teleconnection to impact GrIS warming over centennial time scales we also use the Last Millennium Reanalysis 2 55 proxy-assimilated reanalysis. The correlations between the GSI derived from the EKF400 (using 200hPa horizontal winds as per data availability) and the global mean removed Z500 and SST spatial patterns in both the EKF400 (over AD 1602-2003) and LMR2.1 (over AD  indicate, that while the EKF400 (Supplementary Fig. 11ab) shows promising similarity to the correlations calculated between the ERA5 GSI and the reanalysis SST/Z500 (Supplementary Fig. 9) regarding both magnitudes and spatial patterns, the LMR2.1 exhibits smaller correlation magnitudes  and less resemblance to the PARC. As for the MCA neither of the first two LMR2.1 MCA modes show adequate similarity to the PARC teleconnection, unlike the EKF400 which remarkably resembles the MCA seen in the ERA5 reanalysis (see Fig. 5 in the main text). However, the regression of the Z500 and SST fields from the LMR2.1 onto the Z500/SST expansion coefficient time series obtained from the EKF400, (after removing the global mean time-series from LMR2.1) still resemble the features of both the observed and the EKF400 simulated hemispheric teleconnections, which indicates that the PARC mode exists in the LMR2.1, but is not as separable from other modes as in the EKF400 ( Supplementary Fig. 12). Supplementary Table 1. 30 ice core records across the GrIS used in the study and the correlation coefficients between the individual ice cores and the MCA(1) related expansion coefficient time-series of Z500 (r<Z500>) and the 200hPa Greenland streamfunction index (r<Y200>) in the EKF400 simulation. The source codes of the oceanic coral records are also shown in the rightmost column referring to the codes of the original datasets in ref. 75