Geophysical Research Letters

Transcription

1 RESEARCH LETTER Key Points: Ocean thermohaline circulation variations drove recent decadal Arctic winter sea ice trends Initialized climate model ensembles can skillfully predict Arctic winter sea ice trends Decadal trends in Atlantic winter sea ice will be neutral or positive in the near future Supporting Information: Texts S1 S5 Figure S1 Figure S2 Figure S3 Figure S4 Figure S5 Figure S6 Figure S7 Figure S8 Figure S9 Figure S10 Figure S11 Figure S12 Correspondence to: S. G. Yeager, Citation: Yeager, S. G., A. R. Karspeck, and G. Danabasoglu (2015), Predicted slowdown in the rate of Atlantic sea ice loss, Geophys. Res. Lett., 42, 10,704 10,713, doi:. Received 13 JUL 2015 Accepted 30 NOV 2015 Accepted article online 8 DEC 2015 Published online 19 DEC American Geophysical Union. All Rights Reserved. Predicted slowdown in the rate of Atlantic sea ice loss Stephen G. Yeager 1, Alicia R. Karspeck 1, and Gokhan Danabasoglu 1 1 Climate and Global Dynamics Laboratory, National Center for Atmospheric Research, Boulder, Colorado, USA Abstract Coupled climate models initialized from historical climate states and subject to anthropogenic forcings can produce skillful decadal predictions of sea surface temperature change in the subpolar North Atlantic. The skill derives largely from initialization, which improves the representation of slow changes in ocean circulation and associated poleward heat transport. We show that skillful predictions of decadal trends in Arctic winter sea ice extent are also possible, particularly in the Atlantic sector. External radiative forcing contributes to the skill of retrospective decadal sea ice predictions, but the spatial and temporal accuracy is greatly enhanced by the more realistic representation of ocean heat transport anomalies afforded by initialization. Recent forecasts indicate that a spin-down of the thermohaline circulation that began near the turn of the century will continue, and this will result in near-neutral decadal trends in Atlantic winter sea ice extent in the coming years, with decadal growth in select regions. 1. Introduction There is little doubt that we will see a decline in Arctic sea ice cover in this century in response to anthropogenic warming, and yet internal climate variations and other external forcings could generate considerable spread in Arctic sea ice trends on decadal timescales [Kay et al., 2011; Swartetal., 2015]. Variations in the strength of the Atlantic Meridional Overturning Circulation (AMOC), in particular, may play an important role in modulating rates of Northern Hemisphere sea ice loss because of the associated variations in heat transport into the high-latitude North Atlantic [Mahajan et al., 2011; Day et al., 2012; Koenigk et al., 2012; Msadek et al., 2014]. It has recently been argued that anomalously strong AMOC-related northward heat transport could have been a significant contributor to the observed rapid decline in Arctic summer sea ice extent and that similar anomalies of opposite sign could result in a hiatus of Arctic sea ice loss [Zhang, 2015]. A slowdown in the rate of Arctic sea ice loss is a particularly relevant scenario to consider at present, because there are emerging signs that the AMOC is weakening [Robson et al., 2014a; Smeed et al., 2014; Hermanson et al., 2014]. Arctic sea ice predictability studies focused on seasonal-to-interannual timescales have generally found that initial-value predictability is limited to less than 3 years [e.g., Blanchard-Wrigglesworth et al., 2011; Tietsche et al., 2014; Day et al., 2014], with external climate forcing contributing to significant predictability at longer lead times [Blanchard-Wrigglesworth et al., 2011; Germe et al., 2014]. Both components of predictability are important for decadal climate prediction [Meehl et al., 2009, 2014]. To produce reliable forecasts of future decadal climate change, decadal prediction systems based on initialized coupled climate model simulations must demonstrate skill in retrospective predictions (of past climate variations). This demonstration is particularly challenging for climate fields such as sea ice for which the observational record is relatively short. The Community Earth System Model (CESM) decadal prediction (DP) system [Yeager et al., 2012; Karspeck et al., 2014] is one of several nascent DP efforts [e.g., Pohlmann et al., 2009; Smith et al., 2010; van Oldenborgh et al., 2012; Chikamoto et al., 2013; Doblas-Reyes et al., 2013] that have demonstrated multiyear forecast skill for sea surface temperature (SST) in the North Atlantic subpolar gyre (SPG) region. In this study, we show that the high skill of the CESM DP simulations in predicting ocean-driven SST variations in the subpolar North Atlantic translates into useful skill at predicting decadal trends in Arctic winter sea ice coverage. Retrospective predictions suggest that the extreme reduction in Arctic winter sea ice extent observed in the late 1990s may have been a predictable consequence of the preceding decade of persistent positive winter North Atlantic Oscillation (NAO) conditions and an associated spin-up of the thermohaline circulation (THC), which projects onto both AMOC and horizontal gyre flows [Yeager, 2015]. CESM DP forecasts indicate that relatively low rates of North Atlantic Deep Water (NADW) formation in recent years will result in a continuation of a THC spin-down that YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,704

2 began more than a decade ago [Robson et al., 2014a; Hermanson et al., 2014]. Consequently, projected 10 year trends in Arctic winter sea ice extent seem likely to be more positive than has recently been observed, with decadal climate variability substantially masking the effects of anthropogenic warming. Contrary to expected long-term trends, the CESM forecasts suggest that we should expect decadal sea ice growth in the Atlantic sector, particularly in the Barents Sea region, in the coming years. 2. Experiments and Methods We analyze simulations of the CESM version 1 [Gent et al., 2011] run in both coupled and uncoupled configurations. Both the atmosphere and the ocean models are run at nominal 1 horizontal resolution. The primary focus is on a set of CESM DP simulations (see Text S1 in the supporting information) submitted as part of the Coupled Model Intercomparison Project phase 5 (CMIP5). These consist of 10-member ensembles of the fully coupled CESM model initialized on 1 January of each year between 1955 and 2014 and integrated for 10 years (referred to as forecast years 1 10). This set of CESM DP experiments is an expansion of a previously documented set [Yeager et al., 2012] in which the start dates were limited to every 5 years. The historical initial conditions for the ocean and sea ice in the CESM DP runs are obtained from an uncoupled simulation of the CESM ocean and sea ice models forced at the surface with atmospheric reanalysis data (referred to as the Coordinated Ocean-ice Reference Experiments phase II, or CORE, hindcast simulation [see Danabasoglu et al., 2014]). The CORE hindcast compares favorably with various observational benchmarks in the Atlantic [Yeager et al., 2012; Danabasoglu et al., 2014; Yeager and Danabasoglu, 2014] even though there is no assimilation of ocean or sea ice observations in this surface-forced simulation. In particular, the climatological ( ) distribution of winter (JFM, January March) Arctic sea ice extent is well represented in the CORE simulation (Figure S1). A full field initialization approach is used for the CESM DP runs, and this necessitates a bias correction step prior to analysis (see Text S1). To quantify skill, we will use standard anomaly correlation and mean square skill score (MSSS) statistics [Boer et al., 2013; in what follows, r(a,b) denotes the correlation between time series A and B; MSSS(A,B) denotes the skill of time series A in replicating time series B using climatology as a reference forecast]. Note that MSSS values are 0 when the mean square error is equal to that of the reference forecast (climatology), and MSSS approaches 1 as the mean square error approaches 0. The significance of prediction skill scores is tested against a damped persistence null hypothesis using a parametric bootstrap approach (see Text S4). 3. Buoyancy-Forced Variations in Atlantic Circulation Variations in the winter buoyancy forcing of the subpolar Atlantic Ocean associated with the NAO are believed to generate significant decadal variance in AMOC strength [Häkkinen, 1999; Biastoch et al., 2008; Lohmann et al., 2009; Yeager and Danabasoglu, 2014]. The winter (December March) NAO index [Hurrell, 1995] was +2 on average over the 15 years from 1981 to 1995 (Figure 1a), and this persistent NAO+ forcing resulted in a predictable warming of the SPG region in the middle to late 1990s as a consequence of a strengthened THC [Yeager et al., 2012; Robson et al., 2012; Msadek et al., 2014]. In contrast, the average winter NAO index between 1996 and 2010 was slightly negative, punctuated by extreme NAO conditions in 1996 and Variations in the annual surface formation of North Atlantic Deep Water (NADW) computed from atmospheric reanalysis data and ocean surface observations (see Text S2 for the details of this computation; see Figure S1 for a map showing the specific regions referred to in the text) is highly correlated with the winter NAO index (Figure 1a; r(nadw,nao) =0.66). In the adiabatic limit, variations in the surface formation of NADW represent buoyancy-forced variations in the high-latitude overturning and gyre circulations of the Atlantic Ocean [Grist et al., 2009; Yeager, 2015], referred to here as THC. The upward trend in NADW formation from the mid-1960s to the mid-1990s therefore implies a multidecadal spin-up of the THC, and the neutral-to-weak NADW formation since 1996 suggests that the THC has been weakening since the start of this century. This is consistent with the reasoning from other recent studies [e.g., Robson et al., 2014a; Hermanson et al., 2014]. Observed interannual to decadal changes in ocean density in the upper 1050 m of the central Labrador Sea [Yashayaev and Loder, 2009] are consistent with the NAO-related changes in the rate of surface formation of NADW estimated from historical air-sea flux data, showing a maximum in the early 1990s with a steep decline to anomalously negative values in recent years (Figure 1b, blue curve). The CORE simulation (forced with the same atmospheric reanalysis data used to compute NADW formation) shows broad agreement with the YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,705

3 Figure 1. Formation and propagation of buoyancy-forced water mass anomalies. (a) Annual rate (Sv (sverdrup); 1Sv=10 6 m 3 s 1 ) of surface formation of NADW (σ 0 > 27.6 kg m 3 ) over the subpolar North Atlantic (60 W 20 E; 50 N 90 N) diagnosed from observed atmospheric and oceanic surface fields (thick green curve) and the winter (DJFM, December March) NAO index (thin blue curve, right axis). The remaining panels show 3 year running mean anomalies from CORE (black curves), the CESM DP averaged over the 5 7 year forecast period (red curves and shading are ensemble mean and minimum/maximum range, respectively), and various observational time series (blue curves; see Text S3 in the supporting information for details). Apart from the winter NAO in Figure 1a, all time series are based on annual mean data. (b) Upper 1050 m density anomaly (σ 0 ;10 2 kg m 3 ) in the central Labrador Sea region (56 W 49 W; 56 N 61 N). Note that for these observations, the region of spatial averaging is ill defined because of the sparse measurements. (c) SSH (cm) in the central Labrador Sea, with satellite observations averaged over the same box region (note that the y axis is inverted). (d, e) Same as Figures 1b and 1c but for a region to the east of Grand Banks (50 W 35 W; 40 N 50 N). Anomalies are relative to the following climatologies: (green, black, and red curves), (blue curve in Figure 1b), and (blue curves in Figures 1c and 1e). Geographical regions are shown in Figure S1. YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,706

4 limited observations of Labrador Sea density (r(core,obs) =0.58), although it underestimates the density maximum of the early 1990s (Figure 1b, black curve). The CORE simulation also shows good agreement with satellite observations of sea surface height (SSH) over the Labrador Sea (Figure 1c; r(core,obs) =0.84), a field that reflects density variations over the full water column. The CORE upper ocean density and SSH fields further south in the shelf waters of the Grand Banks of Newfoundland (Figures 1d and 1e) also exhibit high lagged correlations with NADW formation (r(core density,nadw)=0.74 and r(core SSH,NADW)= 0.75 when NADW formation leads by 13 years), suggesting a slow southward propagation of decadal water mass anomalies generated at higher latitudes. The realism of CORE variability in the Grand Banks region is supported by a high correlation with observed SSH anomalies (r(core,obs)=0.77) that show a steep rise from the early 2000s to the present (note inverted scale in Figure 1e), presumably reflecting a sharp decrease in ocean density in that location. In the central Labrador Sea region, the CESM DP ensembles initialized from the CORE ocean and sea ice reconstruction (Figures 1b and 1c) exhibit low skill relative to either CORE (see Figure S4a for skill scores) or observations when averaged over the 5 7 year forecast period. Here and in what follows, we choose to focus on the 5 7 year forecast lead. This midrange forecast tends to yield skill scores that fall between the higher (lower) scores associated with shorter (longer) lead times and that are generally significant (see Text S4 and Figure S4; for present purposes, a significant score exceeds the 95% confidence level of a bootstrapped damped persistence null forecast). Low skill in the Labrador Sea essentially reflects the model s inability to predict winter NAO conditions (Figure S12) and associated surface formation of NADW. However, there is significant skill in forecasting upper ocean density (r(dp,core) =0.85) and SSH (r(dp,core) =0.82) variability in the western boundary region off Grand Banks more than 5 years in advance (Figures 1d, 1e, and S4b). This suggests that the CESM DP experiments accurately simulate the southward propagation of preformed (i.e., initialized) water mass anomalies into the western boundary region between the warm subtropical and cold subpolar gyres a region believed to play a key role in regulating ocean heat transport into the SPG [Tulloch and Marshall, 2012; Buckley et al., 2012]. Decadal variations in ocean density in the vicinity of the Grand Banks are associated with large changes in barotropic gyre strength that are related to changes in the strength and orientation of the heat-ferrying North Atlantic Current. Thus, the considerable skill at forecasting decadal changes in ocean density in the vicinity of Grand Banks translates into significant skill at forecasting the decadal changes in gyre circulation (Figures 2a and S4c; r(dp,core)=0.88) and poleward heat transport across 50 N (Figures 2b and S4d; r(dp,core)=0.88) that dominate the high-latitude variance in the CORE simulation. Decadal AMOC fluctuations are also predictable at this latitude (not shown), but we focus here on changes in the gyre circulation because much of the ocean s mean northward heat transport (and most of its variance) at subpolar latitudes is associated with the barotropic gyre circulation [Tiedje et al., 2012]. SST in the central SPG is strongly influenced by heat advection by ocean currents, and so the large-scale SPG circulation skill contributes to the significant skill in predicting the observed cooling of the SPG in the 1960s [Robson et al., 2014b; Hermanson et al., 2014; Hodson et al., 2014] and the large, abrupt warming of the SPG in the 1990s [Yeager et al., 2012; Robson et al., 2012; Msadek et al., 2014; Karspeck et al., 2014] (Figures 2c and S4e; r(dp,core) =0.88, r(dp,obs) =0.81). There has been a marked cooling of SPG SST since it peaked in 2006, and the recent downward trend appears to be related to decadal ocean density, circulation, and heat transport trends which can be traced back to the abrupt change in the rate of surface formation of NADW in the late 1990s. Observed decadal changes in the hydrography of the Nordic Seas region have been linked to decadal changes in the inflow of warm, salty Atlantic Water across the Iceland-Scotland ridge [Hátún et al., 2005; Holliday et al., 2008; Eldevik et al., 2009; Glessmer et al., 2014] that in turn have been associated with slow modulations in the strength, shape, and heat content of the SPG [Hátún et al., 2005; Nakanowatari et al., 2014]. Both the CORE and DP simulations exhibit realistic decadal variations in ocean heat transport, upper ocean heat content, and SST (Figure S7) that are largely meridionally coherent between SPG latitudes ( 50 N) and Nordic Sea latitudes (>65 N), with evidence of northward propagation into the Norwegian and Barents Seas (Figure S12). In particular, the observed rapid warming of the Nordic Seas between the late 1990s and early 2000s [Holliday et al., 2008; Eldevik et al., 2009] is evident in both CORE and DP (Figures S7 and S12) and appears to be linked to slightly earlier changes in the SPG via anomalous Atlantic Water inflow into the Nordic Seas (see Text S5 for a more in-depth discussion of Nordic Seas heat content variability and its relation to winter sea ice variability). YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,707

5 Figure 2. Variations in large-scale Atlantic circulation, poleward heat transport, and winter sea ice extent. As in Figure 1 but showing the following: (a) Barotropic stream function averaged over the Grand Banks region (note that more negative values indicate stronger cyclonic circulation); (b) ocean poleward heat transport across 50 N in the Atlantic; (c) SST in the central subpolar gyre region (45 W 10 W; 50 N 60 N); (d) Northern Hemisphere winter (JFM) sea ice area over the whole Arctic (40 N 82 N); (e) Northern Hemisphere winter (JFM) sea ice area over the Atlantic sector (90 W 90 E; 40 N 82 N). Anomalies are relative to the following climatologies: (Figures 2a 2c), and (Figures 2d and 2e). The purple dashed curves show the ensemble mean of the six-member uninitialized CESM 20C simulations. 4. Ocean-Driven Trends in Arctic Sea Ice Extent The CESM DP system offers evidence that the rapid Arctic sea ice loss observed between about 1997 and 2007 was related to the very anomalous ocean heat transport that contributed to the rapid mid-1990s warming of the SPG and the early 2000s warming of the Nordic Seas. The 5 7 year forecasts show significant skill at reproducing the accelerated rate of winter sea ice loss over this time period (Figures 2d and S4f; r(dp,obs)= 0.91, MSSS(DP,OBS) = 0.82), most of which occurred in the Atlantic sector (Figures 2e YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,708

6 and S4g; r(dp,obs)= 0.94, MSSS(DP,OBS) = 0.67). A six-member ensemble of uninitialized CESM twentieth century simulations (CESM 20C) appears to reproduce the observed pan-arctic decline between 1979 and 2015 (Figure 2d; r(20c,obs)= 0.90, MSSS(20C,OBS) = 0.70) that is often ascribed to anthropogenic forcing. However, the 20C ensemble mean shows weak variability in the Atlantic sector during the satellite era (Figure 2e; r(20c,obs)= 0.53; MSSS(20C,OBS) = 0.15), indicating that the pan-arctic time series from 20C is masking compensating errors in the spatial distribution of sea ice loss. The observed late twentieth century rapid decline in Arctic winter sea ice extent was associated with a retraction of the ice edge throughout the Labrador, Greenland, Irminger, and Barents Seas that occurred in tandem with an abrupt warming of the SPG and Nordic Sea regions (see Figures S5 and S7). The large observed variance in winter ice extent in the Barents Sea, in particular, has been strongly linked to ocean heat content change in the marginal ice zone [Schlichtholz, 2011; Årthun et al., 2012; Onarheim et al., 2015]. While the magnitudes and spatial structures of the multidecadal sea ice and SST trends between 1983 and 2013 are realistic in the CORE and CESM DP simulations, the 20C ensemble shows only a very weak warming of the SPG and negligible sea ice loss in the Atlantic sector over this time period. The apparently high skill of 20C relative to observations in Figure 2d is therefore misleading, because it hides substantial differences in mechanism. It is an artifact of the loss of Pacific (instead of Atlantic) sea ice in 20C (Figure S5). Some of the DP hindcast skill comes from the persistence of anomalous initial conditions that can be dominated by long-term trends (this is the case for Arctic sea ice). Evaluating the tendency (or short-term trend) predicted by individual DP ensembles is a way to gauge skill that is not simply imparted by long-term trends in the initial conditions used for DP. The heat budget analysis in Yeager et al. [2012] showed that the skill of CESM DP in predicting upper ocean heat content (and SST) in the Atlantic SPG (Figure 2c) derives in large part from skillful predictions of heat content tendency associated with anomalous ocean heat advection, and this would also appear to explain the high DP skill in predicting decadal upper ocean heat content tendencies in the marginal ice zones of the Atlantic (see Figure S11 and related discussion in Text S5). All of the CESM DP ensembles initialized in the middle to late 1990s, when the SPG circulation and heat transport were at their strongest levels of the past half century (Figures 2a and 2b), predict positive trends in SST and negative trends in winter sea ice extent in the North Atlantic. The predicted 10 year trend patterns for winter sea ice and SST from the 1998 DP ensemble (Figures 3c and 3g), for example, compare well with both observations (Figures 3a and 3e) and CORE (Figures 3b and 3f) in the Atlantic sector, suggesting that the rapid sea ice loss there between 1997 and 2007 was predictable in advance and that it was largely driven by buoyancy-forced ocean dynamics. The pronounced 1990s spin-up of THC and heat transport is absent in the uninitialized 20C ensemble (Figures 2a and 2b), and so external forcing alone yields a relatively weak SPG warming (Figure 2c) and minimal Atlantic sea ice loss during (Figures 3d and 3h). Observed 10 year trends in Arctic winter sea ice extent have varied considerably over the satellite era, with the most rapid sea ice loss occurring in the decade (Figure 4). The CORE reconstruction exhibits excellent agreement with the observed decadal trends of JFM sea ice extent over the whole Arctic (Figure 4a; r(core,obs) = 0.95, MSSS(CORE,OBS) = 0.90) as well as in the Atlantic sector (Figure 4b; r(core,obs) = 0.95, MSSS(CORE,OBS) = 0.85), and as shown above, the associated spatial patterns are realistic (see also trends in Figure S6). The CESM DP ensembles are able to reproduce the observed record of the 10 year winter sea ice extent trends with high skill scores (Figure 4c; see plot for scores), particularly in the Atlantic (Figure 4d). The DP skill scores relative to observations are significant in the Atlantic sector but not for the whole Arctic (Figures S4h and S4i). To the extent that the CORE sea ice reconstruction represents a reasonable observation-based proxy for sea ice variability prior to 1979, the DP skill at predicting decadal trends can be evaluated relative to CORE over a longer time period ( , the overlap period of the CORE and DP curves in Figure 4) than the observed record. This results in even higher DP skill scores that are significant for both pan-arctic and Atlantic regions (Figures S4h and S4i). As discussed above, the skill of uninitialized 20C ensembles in replicating the pan-arctic time series (Figure 4a) is suspect because of unrealistic spatial patterns of variability, but there does appear to be a component of externally forced variability in winter sea ice trends that results in some 20C skill in the Atlantic (Figure 4b; see also Barents Sea region in Figures 3d, 3h, and S10). The significant DP skill in the Atlantic (Figure 4d) therefore comes from a baseline level of skill associated with external forcing that is enhanced by initialization. The increasingly rapid Atlantic sea ice loss between the early 1980s and the late 1990s is well captured by the CESM DP (Figure 4d), and it appears to be linked to the concomitant increase in ocean heat transport YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,709

7 Geophysical Research Letters Figure 3. Spatial distributions of select 10 year trends in winter sea ice fraction and SST. Ten year linear trends in winter (JFM) sea ice fraction (fraction/decade) and annual mean SST ( C/decade) over the periods (a h) and (i l) from observations (OBS), the CORE simulation (CORE), the CESM DP 10-member ensemble mean (DP), and the CESM 20C 6-member ensemble mean (20C). The trends in Figures 3c and 3g (Figures 3i and 3j) are computed from the single DP ensemble initialized on 1 January 1998 (2008). Refer to Text S1 for details of how CESM DP trends are computed. into the SPG and subsequently into the marginal ice zones of the Labrador, Greenland, and Barents Seas (see Text S5 for a regional assessment of DP skill). Conversely, the recent slowdown in the decadal rate of winter sea ice loss that has been observed since 1998, also skillfully predicted, coincides with a rapid decline in THC strength and pan-atlantic ocean heat transport from the highs of the late 1990s (Figure 2b). The prediction of a positive trend in Atlantic winter sea ice extent between 2005 and 2015 (Figure 4d, red curve at 2005 which is based on the DP ensemble initialized on 1 January 2006) has now been verified by recent observations that extend through March 2015 (Figure 4d, blue dot at 2005). In contrast to 20C projections, initialized predictions suggest that we should expect growth or near-neutral maintenance of Atlantic winter sea ice extent in the coming years. The sea ice growth predicted for the period ( km2 /decade; see Figure 4d) is associated with SST cooling throughout the Labrador, Irminger, and Nordic Seas and a systematic expansion of the winter sea ice edge in the Atlantic sector, particularly in the Barents Sea (Figures 3i and 3j). The trends expected from external forcing alone are very different (Figures 3k and 3l). The most recent CESM DP prediction (for the period) shows a neutral trend for the Atlantic sector as a whole (Figure 4d, red curve at 2013), but a continued rebound of winter sea ice in the Barents Sea (Figures S6i and S6j). YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,710

8 Figure 4. Decadal trends in Arctic winter sea ice extent. Ten year linear trends in winter (JFM) sea ice extent (10 6 km 2 /decade) computed over (a, c) the whole Arctic (40 N 82 N) and (b, d) the Atlantic sector (90 W 90 E; 40 N 82 N). Trends are plotted on the x axis at the start year (e.g., the trend is plotted at 1997). Refer to Text S1 for details on how CESM DP trends are computed. Shading gives the minimum/maximum range of the ensemble. Refer to Figure S10 for a breakdown of Atlantic sector trends by subregion. 5. Conclusions Recently observed decadal trends in Arctic winter sea ice extent are not well explained by external forcing alone. The particularly rapid sea ice loss from 1997 to 2007 was related to extreme ocean conditions that drove a sustained warming of the surface waters throughout the subpolar Atlantic and Nordic Seas. Ongoing adjustment of the ocean THC is now contributing to a cooling trend in the subpolar Atlantic and an associated slowdown in the rate of Arctic winter sea ice retreat. Uninitialized simulations of the twentieth century driven by anthropogenic and other external forcings lackthe ocean heat transport variations that contributed to the magnitude and spatial patterns of observed sea ice loss. However, initialized prediction ensembles using CESM can skillfully predict low-frequency modulations in the decadal trends of Arctic sea ice, and the significant skill scores for Atlantic sector sea ice extent, in particular, suggest that CESM DP future forecasts merit serious consideration. In theory, the record high (positive) trend in Atlantic winter sea extent observed between 2005 and 2015 could have been predicted back in Future forecasts from the CESM DP indicate that we should expect a pause in decadal Atlantic winter sea ice loss over the next 5 to 10 years. The late 1990s warming of the subpolar Atlantic (and associated accelerated sea ice loss) and the current cooling of the subpolar Atlantic (and associated pause in sea ice loss) are related to decadal variations in Atlantic THC strength YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,711

9 that can be traced back to decadal changes in the NAO and its imprint on the production of the abyssal waters of the Atlantic Ocean. The spread in forecasted 10 year trends conveys the uncertainty associated with unpredictable elements of the climate system, and the discrepancy between the model simulations (both CORE and CESM DP) and observed reality over the time period where they overlap makes it clear that the models are imperfect. Nevertheless, this combined analysis of observations, forced model simulation, and initialized coupled model predictions suggests that there is good reason to expect lower rates of winter sea ice loss in the Arctic over the next 5 to 10 years than were observed in the late 1990s. Acknowledgments This work was supported by the National Oceanic and Atmospheric Administration (NOAA) Climate Program Office under Climate Variability and Predictability Program grants NA09OAR and NA13OAR , by the National Science Foundation (NSF) Collaborative Research EaSM2 grant OCE , and by the NSF through its sponsorship of the National Center for Atmospheric Research. The decadal prediction experiments were run by Haiyan Teng who was supported by the Regional and Global Climate Modeling Program (RGCM) of the U.S. Department of Energy s Office of Science (BER), cooperative agreement DE-FC02-97ER This work used computing resources of the National Energy Research Scientific Computing Center (NERSC) which is supported by the BER under contract DE-AC02-05CH11231, as well as resources provided by NCAR s Computational and Information Systems Laboratory (CISL). We thank Igor Yashayaev for generously providing us with his observational time series from the Labrador Sea. Finally, we gratefully acknowledge the effort and dedication of the team of scientists and software engineers who developed the Community Earth System Model (CESM) used in this research. References Årthun, M., T. Eldevik, L. H. Smedsrud, Ø. Skagseth, and R. B. Ingvaldsen (2012), Quantifying the influence of Atlantic heat on Barents Sea ice variability and retreat, J. Clim., 25, , doi: /jcli-d Biastoch, A., C. Böning, J. Getzlaff, J.-M. Molines, and G. Madec (2008), Causes of interannual-decadal variability in the meridional overturning circulation of the midlatitude North Atlantic Ocean, J. Clim., 21, Blanchard-Wrigglesworth, E., C. M. Bitz, and M. M. Holland (2011), Influence of initial conditions and climate forcing on predicting Arctic sea ice, Geophys. Res. Lett., 38, L18503, doi: /2011gl Boer, G. J., V. V. Kharin, and W. J. Merryfield (2013), Decadal predictability and forecast skill, Clim. Dyn., 41, , doi: /s Buckley, M. W., D. Ferreira, J.-M. Campin, J. Marshall, and R. Tulloch (2012), On the relationship between decadal buoyancy anomalies and variability of the Atlantic meridional overturning circulation, J. Clim., 25, , doi: /jcli-d Chikamoto, Y., et al. (2013), An overview of decadal climate predictability in a multi-model ensemble by climate model MIROC, Clim. Dyn., 40, , doi: /s y. Danabasoglu, G., et al. (2014), North Atlantic simulations in Coordinated Ocean-ice Reference Experiments phase II (CORE-II). Part I: Mean states, Ocean Model., 73, , doi: /j.ocemod Day, J. J., J. C. Hargreaves, J. D. Annan, and A. Abe-Ouchi (2012), Sources of multi-decadal variability in Arctic sea ice extent, Environ. Res. Lett., 7, 1 6, doi: / /7/3/ Day, J. J., S. Tietsche, and E. Hawkins (2014), Pan-Arctic and regional sea ice predictability: Initialization month dependence, J. Clim., 27, , doi: /jcli-d Doblas-Reyes, F. J., I. Andreu-Burillo, Y. Chikamoto, J. García-Serrano, V. Guemas, M. Kimoto, T. Mochizuki, L. R. L. Rodrigues, and G. J.van Oldenborgh (2013), Initialized near-term regional climate change prediction,nat. Commun., 4, 1715, doi: /ncomms2704. Eldevik, T., J. E. Ø. Nilsen, D. Iovino, K. A. Olsson, A. B. Sandø, and H. Drange (2009), Observed sources and variability of Nordic seas overflow, Nat. Geosci., 2, , doi: /ngeo518. Gent, P. R., et al. (2011), The community climate system model version 4, J. Clim., 24, , doi: /2011jcli Germe, A., M. Chevallier, y Mélia D. S., E. Sanchez-Gomez, and C. Cassou (2014), Interannual predictability of Arctic sea ice in a global climate model:regional contrasts and temporal evolution,clim. Dyn., 43, , doi: /s Glessmer, M. S., T. Eldevik, K. Våge, J. E. Ø. Nilsen, and E. Behrens (2014), Atlantic origin of observed and modelled freshwater anomalies in the Nordic Seas, Nat. Geosci., 7, , doi: /ngeo2259. Grist, J. P., R. Marsh, and S. A. Josey (2009), On the relationship between the North Atlantic Meridional Overturning Circulation and the surface-forced overturning streamfunction, J. Clim., 22, , doi: /2009jcli Häkkinen, S. (1999), Variability of the simulated meridional heat transport in the North Atlantic for the period , J. Geophys. Res., 104, 10,991 11,007. Hátún, H., A. B. Sandø, H. Drange, B. Hansen, and H. Valdimarsson (2005), Influence of the Atlantic subpolar gyre on the thermohaline circulation, Science, 309, Hermanson, L., R. Eade, N. H. Robinson, N. J. Dunstone, M. B. Andrews, J. R. Knight, A. A. Scaife, and D. M. Smith (2014), Forecast cooling of the Atlantic subpolar gyre and associated impacts, Geophys. Res. Lett., 41, , doi: /2014gl Hodson, D. L. R., J. I. Robson, and R. T. Sutton (2014), An anatomy of the cooling of the North Atlantic Ocean in the 1960s and 1970s, J. Clim., 27, , doi: /jcli-d Holliday, N. P., S. L. Hughes, S. Bacon, A. Beszcynska-Möller, B. Hansen, A. Lavín, H. Loeng, K. A. Mork, S. Østerhus, T. Sherwin, and W. Walczowski (2008), Reversal of the 1960s to 1990s freshening trend in the northeast North Atlantic and Nordic Seas, Geophys. Res. Lett., 35, L03614, doi: /2007gl Hurrell, J. W. (1995), Decadal trends in the North Atlantic Oscillation: Regional temperatures and precipitation, Science, 269, Karspeck, A., S. Yeager, G. Danabasoglu, and H. Teng (2014), An evaluation of experimental decadal predictions using CCSM4, Clim. Dyn., 44, , doi: /s Kay, J. E., M. M. Holland, and A. Jahn (2011), Inter-annual to multi-decadal Arctic sea ice extent trends in a warming world, Geophys. Res. Lett., 38, L15708, doi: /2011gl Koenigk, T., C. K. Beatty, M. Caian, R. Döscher, and K. Wyser (2012), Potential decadal predictability and its sensitivity to sea ice albedo parameterization in a global coupled model, Clim. Dyn., 38, , doi: /s z. Lohmann, K., H. Drange, and M. Bentsen (2009), Response of the North Atlantic subpolar gyre to persistent North Atlantic Oscillation like forcing, Clim. Dyn., 32, , doi: /s Mahajan, S., R. Zhang, and T. L. Delworth (2011), Impact of the Atlantic Meridional Overturning Circulation (AMOC) on Arctic surface air temperature and sea ice variability, J. Clim., 24, , doi: /2011jcli Meehl,J., et al. (2009), Decadal prediction: Can it be skillful?, Bull. Am. Meteorol. Soc., 90, Meehl, J., et al. (2014), Decadal prediction: An update from the trenches, Bull. Am. Meteorol. Soc., 95, , doi: / BAMS-D Msadek, R., et al. (2014), Predicting a decadal shift in North Atlantic climate variability using the GFDL forecast system, J. Clim., 27, , doi: /jcli-d Nakanowatari, T., K. Sato, and J. Inoue (2014), Predictability of the Barents Sea ice in early winter: Remote effects of oceanic and atmospheric thermal conditions from the North Atlantic, J. Clim., 27, , doi: /jcli-d Onarheim, I. H., T. Eldevik, M. Årthun, R. B. Ingvaldsen, and L. H. Smedsrud (2015), Skillful prediction of Barents Sea ice cover, Geophys. Res. Lett., 42, , doi: /2015gl YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,712

10 Pohlmann, H., J. H. Jungclaus, A. Köhl, D. Stammer, and J. Marotzke (2009), Initializing decadal climate predictions with the GECCO oceanic synthesis: Effects on the North Atlantic, J. Clim., 22, Robson, J. I., R. Sutton, K. Lohmann, D. Smith, and M. D. Palmer (2012), Causes of the rapid warming of the North Atlantic Ocean in the mid-1990s,j. Clim., 25, , doi: /jcli-d Robson, J.I., D. Hodson,E. Hawkins, and R. T. Sutton (2014a), Atlantic overturning in decline?, Nat. Geosci., 7, 2 3. Robson, J. I., R. T. Sutton, and D. Smith (2014b), Decadal predictions of the cooling and freshening of the North Atlantic in the 1960s and the role of ocean circulation, Clim. Dyn., 42, , doi: /s Schlichtholz, P. (2011), Influence of oceanic heat variability on sea ice anomalies in the Nordic Seas, Geophys. Res. Lett., 38, L05705, doi: /2010gl Smeed, D. A., et al. (2014), Observed decline of the Atlantic meridional overturning circulation , Ocean Sci., 10, 29 38, doi: /os Smith, D. M., R. Eade, N. J. Dunstone, D. Fereday, J. M. Murphy, H. Pohlmann, and A. A. Scaife (2010), Skilful multi-year predictions of Atlantic hurricane frequency, Nat. Geosci., 3, , doi: /ngeo1004. Swart, N. C., J. C. Fyfe, E. Hawkins, J. E. Kay, and A. Jahn (2015), Influence of internal variability on Arctic sea-ice trends, Nat. Clim. Change, 5, 86 89, doi: /nclimate2483. Tiedje, B., A. Köhl, and J. Baehr (2012), Potential predictability of the North Atlantic heat transport based on an oceanic state estimate, J. Clim., 25, , doi: /jcli-d Tietsche, S., J. J. Day, V. Guemas, W. J. Hurlin, S. P. E. Keeley, D. Matei, R. Msadek, M. Collins, and E. Hawkins (2014), Seasonal to interannual Arctic sea ice predictability in current global climate models, Geophys. Res. Lett., 41, , doi: /2013gl Tulloch, R., and J. Marshall (2012), Exploring mechanisms of variability and predictability of Atlantic meridional overturning circulation in two coupled climate models,j. Clim., 25, , doi: /jcli-d van Oldenborgh, G. J., F. J. Doblas-Reyes, B. Wouters, and W. Hazeleger (2012), Decadal prediction skill in a multi-model ensemble, Clim. Dyn., 38, , doi: /s Yashayaev, I., and J. W. Loder (2009), Enhanced production of Labrador Sea water in 2008, Geophys. Res. Lett., 36, L01606, doi: /2008gl Yeager, S. (2015), Topographic coupling of the Atlantic overturning and gyre circulations, J. Phys. Oceanogr., 45, , doi: /jpo-d Yeager, S., and G. Danabasoglu (2014), The origins of late twentieth century variations in the large-scale North Atlantic circulation, J. Clim., 27, , doi: /jcli-d Yeager, S., A. Karspeck, G. Danabasoglu, J. Tribbia, and H. Teng (2012), A decadal prediction case study: Late 20th century North Atlantic ocean heat content, J. Clim., 25, , doi: /jcli-d Zhang, R. (2015), Mechanisms for low-frequency variability of summer Arctic sea ice extent, Proc. Natl. Acad. Sci. U.S.A., 112, , doi: /pnas YEAGER ET AL. PREDICTED RATE OF SEA ICE LOSS 10,713