SUPPLEMENTARY INFORMATION DOI: 10.1038/NGEO2277 Robust Arctic sea-ice influence on the frequent Eurasian cold winters in past decades Masato Mori 1*, Masahiro Watanabe 1, Hideo Shiogama 2, Jun Inoue 3, & Masahide Kimoto 1 1: Atmosphere and Ocean Research Institute, the University of Tokyo, Chiba, Japan 2: Center for Global Environmental Research, National Institute for Environmental Studies, Ibaraki, Japan 3: National Institute of Polar Research, Tokyo, Japan Supplementary Information S1. Differences of ensemble-mean fields between AGCM and reanalysis data Despite similarity in the pattern of SAT anomalies between the reanalysis data and the AGCM ensemble simulation (Fig. 1a, 1b), there is a discernible difference in the associated circulation pattern over the Pacific Atlantic side (Fig. 1c, 1d). The major cause of this difference is that the composite field in reanalysis data contains considerable signals associated with the AO, which ought to be largely eliminated from the AGCM ensemble-mean field. In order to demonstrate this, an additional analysis was performed. Figure S5 shows the same composite as Figs. 1a and 1c, but the anomalies have been reconstructed only using EOF1 and EOF2 (Fig. 2a-b). A comparison between Fig. 1a, c, and Fig. S5a, b indicates that the composite fields in reanalysis can be reconstructed well from two EOFs. Furthermore, decomposition into each EOF (Fig. S5c-f) reveals that the AO has imprints over the North Atlantic and the polar region. This is in contrast to the ensemble-mean fields in the AGCM, which can be explained predominantly by the WACE (Fig. S6, compared with Fig. S5). These differences between the AGCM and the reanalysis are due to the small sample size in the composite of the reanalysis. NATURE GEOSCIENCE www.nature.com/naturegeoscience 1
S2. Relevance of the AO/NAO in the SIC-forced circulation response Several modeling studies have shown a negative AO/NAO-like response to Arctic sea-ice loss 2,13-16, 20-23. It should be noted that the pattern of ensemble-mean circulation response to sea-ice loss in our simulation (Fig. 1d) is very similar to ensemble mean responses shown in previous numerical experiments prescribing realistic sea-ice reduction in their AGCMs. Specifically, the pattern of Z500 responses shown in ref. 22 (their Fig. 7b) and in ref. 16 (their Fig. 6e) are similar to our Fig. 1d, although they referred to those responses as a negative AO/NAO-like pattern. In this study, we showed that the atmospheric response to sea-ice loss is detectable more clearly when it is decomposed into a canonical AO/NAO (not AO/NAO-like ) and the WACE pattern. We believe that the results in previous modeling studies would also be clearer if their ensembles were decomposed in a similar way. S3. Temperature anomaly associated with the WACE pattern The mid-latitude cold SAT anomalies, spreading over central to eastern Asia, observed from the WACE pattern (Fig. 2b, Supplementary Figs. S3b, S4b, S7b, and S7d), are induced primarily by a northerly cold-air advection associated with the anticyclonic circulation anomaly centred around 60 E, 60 N (Fig. 2b). Over central Eurasia, this circulation anomaly shows equivalent barotropic and baroclinic structure to the west and east, respectively (Supplementary Fig. S8). By contrast, the WACE accompanies warm SAT anomalies over northern Europe and western Russia. The influence of the WACE on SAT anomalies is insignificant over North America in both the reanalysis and the AGCM simulation (Supplementary Fig. S7), which indicates that the contribution of the WACE to the warm Arctic cold continents pattern 7 is restricted to the Eurasian continent. In this study, the WACE is defined by EOF2 for SAT over the domain of 0 180, 20 N 90 N. When the EOF is computed over the entire Northern Hemisphere north of 20 N, EOF2 shows features similar to Fig. 2b. However, its fractional variance is reduced and the spatial pattern contains significant signals over North America, even though the WACE does not in reality exhibit a 2
correlation over that region (Supplementary Fig. S7). Therefore, the WACE could better be extracted using the EOF for SAT over the eastern hemisphere. S4. CMIP5 models analysed in this study We analysed monthly outputs of the CMIP5 historical and Representative Concentration Pathways (RCP) 4.5 experiments 24 for the period 1979 2098 from 22 coupled general circulation models: ACCESS1.0, BCC-CSM1.1, CanESM2, CCSM4, CNRM-CM5, CSIRO-Mk3.6.0, GFDL-CM3, GFDL-ESM2G, GFDL-ESM2M, GISS-E2-R, HadGEM2-CC, HadGEM2-ES, INM-CM4, IPSL-CM5A-LR, IPSL-CM5A-MR, MIROC-ESM, MIROC-ESM-CHEM, MIROC5, MPI-ESM-LR, MRI-CGCM3, NorESM1-M, and NorESM1-ME (see http://cmip-pcmdi.llnl.gov/cmip5/availability.html for more detail). These models were chosen based on the availability of SIC output. Before examining a future change in the frequency of cold winters, we checked dominant modes of SAT variability in CMIP5 climate model simulations in a similar manner to Figs. 2a and 2b. We found that the two leading EOFs for 22 CMIP5 simulations for 1979 2013 represent the AO and WACE (Supplementary Fig. S9). Associated SIC anomalies reveal that the WACE is strongly coupled with the sea-ice change in BKS compared to the AO (Supplementary Fig. S9). The above results demonstrate that the dominant WACE response to BKS sea-ice loss shown in our ensemble simulations is not an artifact due to a single model and reinforce our conclusions. S5. Projection uncertainty The results of analysis of simulation performed by 22 CMIP5 models suggest that the frequent occurrence of cold winters in the past decade may be a temporary phenomenon in a transitional phase of eventual global warming (Supplementary Fig. S10). However, for a robust estimation of sea-ice loss impact in future, we need to pay attention to projection uncertainty originating from at least two sources. The first source is the large ambiguity in 3
SIC decrease in the BKS (99 percentile ranging from 66 to 6% during 2089 2098, Supplementary Fig. S10), which can excite WACE responses of varying magnitudes. The second source is most relevant to our findings, arising from an insufficient number of ensemble members. Indeed, the single-member ensemble of 22 CMIP5 models failed to reproduce the recently observed cooling over central Eurasia (Supplementary Fig. S11). Based on our estimate of signal detection (Fig. 3), each model should have at least four members for robust estimation of SAT response to future sea-ice retreat in the BKS. Such a large ensemble projection would contribute to obtaining more concrete conclusions on the possible change in cold winter frequency in Eurasia. 4
Supplementary Figures Figure S1 a-b, The 10-year mean SAT differences between 1994 2003 and 2004 2013 in DJF (the recent minus the previous decade) for GISTEMP (a) and ERA-Interim (b). Stippling indicates regions exceeding 95% statistical confidence. Grey shading in a indicates missing-data grids. c, Time series of SAT anomalies averaged over central Eurasia (60 E 120 E, 40 N 60 N, indicated by rectangular box in b) for GISTEMP (green curve; anomaly relative to 1981 2013 mean) and ERA-Interim (black curve; anomaly relative to 1979 2013 mean). The severely warm (red dots) and cold (blue dots) winters are defined by the SAT anomaly above and below one standard deviation (dashed lines) for 1979 2013 in ERA-Interim.
Figure S2 a, Interannual change of SIC in September averaged over the Barents Kara Sea (30 E 90 E, 65 N 85 N; black rectangular region in b) with area weight, based on HadISST (for 1979 2012). Blue and red marks indicate the bottom and top 10 years of the time series, which define low- and high-ice years, respectively. b-c, Composite-SIC differences between the low- and high-ice years (former minus latter) in SON (b) and DJF (c). d, The 10-year mean SIC differences between 1994 2003 and 2004 2013 (former minus latter) in DJF. 6
Figure S3 a-b, As in Fig. 2a-b, but for SAT computed from GISTEMP during the 1979 2013 period. EOF1 (a) and EOF2 (b) account for 39 and 20% of the total variance, respectively. Grey shading indicates data missing grids. c-d, PC1 (c) and PC2 (d) of the leading two modes (indicated by blue curves). Grey curves indicate PCs obtained from ERA-Interim, shown in Fig. 2c-d. 7
Figure S4 a-b, As in Fig. 2a-b, but for the combined 200-member ensemble of LICE and HICE simulations. EOF1 (a) and EOF2 (b) account for 20 and 17% of the total variance, respectively. c-d, PDFs of the corresponding PC1 (c) and PC2 (d), calculated separately for LICE (blue curves) and HICE (red curves). 8
Figure S5 a-b, As in Figs. 1a and 1c, but anomalies, taken from ERA-Interim, have been reconstructed only using EOF1 and EOF2 of SAT shown in Figs. 2a-b. c-d, As in a-b, but anomalies have been reconstructed only using EOF1, e-f, As in a-b, but anomalies have been reconstructed only using EOF2. 9
Figure S6 a-b, As in Fig. S5, but for the AGCM ensemble simulation. The contour interval is 3 hpa in a, c, and e. Note that the contour interval and colour scales are different from that in Figs. 1b and 1d. 10
Figure S7 a-b, DJF-mean SAT anomalies regressed on PC1 (a) and PC2 (b) normalised by one standard deviation, associated with two leading EOFs to winter SAT anomalies over 0 180, 20 N 90 N, taken from ERA-Interim (1979 2013). c-d, As in a-b, but for the combined 200-member ensemble of LICE and HICE simulations. Stippling indicates regions exceeding 95% statistical confidence. 11
Figure S8 Meridionally (40 N 60 N) averaged vertical structure of the temperature (colour) and geo-potential height (contours) anomalies. a-b, Difference of composite fields between low- and high-ice years (former minus latter), taken from ERA-Interim (a) and from the 100-member ensemble-mean of the AGCM experiments (b). The contour interval is 4 m in a and 1 m in b, with negative contours dotted. c-d, Anomalies regressed on PC1 (c) and PC2 (d) shown in Fig. 2c-d, taken from ERA-Interim. e-f, As in c-d, but for combined 200-member ensemble simulation. Note that polarity of AO is reversed in c and e (i.e. negative phase is shown). Contour interval is 3 m in c-f. Stippling indicates regions exceeding 95% statistical confidence. Grey shading indicates data missing grids. 12
Figure S9 a-b, DJF-mean SAT anomalies regressed on PC1 (a) and PC2 (b) normalised by one standard deviation, associated with two leading EOFs for DJF-mean SAT anomalies over 0 180, 20 N 90 N, taken from combined 22 CMIP5 historical and RCP4.5 simulations during the 1979 2013 period (748 winters are analysed). SAT anomalies are defined by subtracting DJF climatology (1979 2013) in each model. EOF1 and EOF2 account for 24 and 16% of the total variance, respectively. c-d, As in a-b, but for DJF-mean SIC anomalies regressed on PC1 (c) and PC2 (d). Stippling in a-b indicates regions exceeding 95% statistical confidence. 13
Figure S10 Long-term changes in DJF-mean values in CMIP5 historical and RCP4.5 simulations. a, SAT anomaly (curves) and frequency of severe winters (bar) over central Eurasia. The severe winters are defined by the averaged SAT anomaly below one standard deviation computed each model during 1979 2013. b-c, Projection coefficients of SAT anomalies on EOF1 (b) and EOF2 (c) shown in Fig. 2a-b, respectively. Scale of the SAT anomaly associated with EOF1 and EOF2 over central Eurasia is shown at the right of each panel (note that the axis is reversed in c). d, SIC anomaly in the BKS. The black curves are derived from ERA-Interim in a-c and from HadISST in d. Thin and thick curves show individual models and the multi model ensemble (MME) mean, respectively. The anomalies are defined by deviations from the 1979 2013 mean, for observations and for each model. 14
Figure S11 a, As in Fig. S1a, but for MME mean of the 22 CMIP5 models. b, Regional mean values over central Eurasia, indicated by the rectangular box in a (60 E 120 E, 40 N 60 N). Grey dots indicate individual models, red dot indicates MME mean, and the black dot indicates ERA-Interim. 15