Assessing the Influence of Sea Surface Temperature and Arctic Sea Ice Cover on the Uncertainty in the Northern Winter Future Climate Projections

1. School of Atmospheric Sciences & Guangdong Province Key Laboratory for Climate Change and Natural Disaster Studies, Sun Yat-sen University, Zhuhai, China 2. Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, China 3. Geophysical Institute, University of Bergen, Bergen, Norway 4. Bjerknes Centre for Climate Research, University of Bergen, Bergen, Norway 5. Nansen Environmental and Remote Sensing Center, Bergen, Norway 6. Swedish Meteorological and Hydrological Institute, Norrköping, Sweden 7. Bolin Centre for Climate Research, Stockholm University, Stockholm, Sweden 8. Danish Meteorological Institute, Copenhagen, Denmark 9. Nansen-Zhu International Research Centre, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China 10. Graduate School of Science, Hokkaido University, Japan 11. Guy Carpenter Asia-Pacific Climate Impact Centre, School of Energy and Environment,


23
There is great concern on if and how the intensity and location of large-scale near-surface 24 circulation features (e.g., sea level pressure, SLP, near the Aleutian low and the Icelandic low) 25 and their associated high-impact weather will change with global warming. Climate models 26 consistently simulate a significant and systematic increase in surface air temperature in the high 27 emission scenarios. Still, they show a large discrepancy in the magnitude of global warming 28 because of their different climate sensitivity (Andrew et  2) To what extent do AGCMs reproduce these SLP patterns when forced by these 96 corresponding SST and SIC patterns? Do the mechanisms simulated by the AGCMs 97 correspond to those of the CMIP5 inter-model differences, estimated from ten models 98 with at least three ensemble members? 99 We focus on the winter (DJF) climate projections from the CMIP5 RCP8.5 scenario, which 100 is the highest emission scenario and thus produces the strongest externally forced responses. 101 The projected climate change is considered for the late twenty-first century (2069/70-2098/99) 102 relative to the late twentieth century (1971/72-2000/01), which is referred to as "response" 103 hereafter. For example, the projected change of SLP is called the SLP response. 104 To identify uncertainties that are inherent to individual models requires large-ensemble 105 simulations (Deser et al. 2020). This is especially true for the extratropical circulation where 106 large internal climate variability can even mask the signals from strong external forcing (Deser 107 et al. 2012). Thus, we analyze simulations from ten CMIP5 coupled models that have at least 108 three ensemble members in the historical and RCP8.5 runs (Table 1); this is to try to minimally 109 limit noise from internal variability. This selection was limited by the availability of model 110 simulations at the time this study was started. (CSIRO-Mk-3.6.0 was also available, but is not 111 included because its SST and SIC biases were too large). We average across ensemble members 112 of each model to increase the signal-to-noise ratio for the external forcing, where the response 113 averaged across the ensemble members is called the "forced response". The MME forced 114 response is the unweighted average of the forced response from the ten CMIP5 models. 115 Section 2 presents analysis of the uncertainty in the forced response of DJF SLP from the 116 ten CMIP5 models. Section 3 describes the design of sensitivity experiments. Section 4 117 the ensemble mean is reduced by one over the square root of the ensemble size, and for the 140 three members is ~0.6. The selected ten CMIP5 models appear to largely capture the uncertainty 141 and internal variability in the response across the 34 available CMIP5 models: the uncertainty 142 pattern computed from all ensemble members of the ten CMIP5 models is very similar to that 143 computed from single ensemble members of the 34 available CMIP5 models (pattern 144 correlation of +0.94), with only slightly lower amplitude (Fig. S1). It should be stressed that the 145 inter-model spread of the 34 CMIP5 models does not provide a good estimate of the uncertainty 146 in the forced response over the northeastern Atlantic and high latitudes, where the internal 147 variability is large. 148 The dominant modes of the uncertainty in the forced SLP response are identified by 149 performing an inter-model empirical orthogonal function (EOF) analysis on the forced response 150 in DJF SLP from the ten CMIP5 models over the domain 20°-90°N (Fig. 2). In the EOF analysis, 151 the MME forced response is removed prior to computing the covariance matrix, which is 152 weighted by the square root of the cosine of latitude to account for the change in grid size. As 153 shown in Fig. 2a, the first mode (EOF1) has the strongest positive loading over the North Pacific 154 and mainly characterizes weakening or strengthening of the Aleutian low, but also has a positive 155 loading over the northeastern Atlantic and a negative loading near the Mediterranean. The 156 second mode (EOF2) has a strong positive loading over the southeastern flank of the Icelandic 157 low and a strong negative loading over the Hudson Bay (60º N, 90º W; Fig. 2b). The spatial 158 pattern of EOF2 over the North Pacific represents a shift in the Aleutian low, with a positive 159 loading in its eastern flank and a negative loading in its northwestern flank. Overall, EOF1 and 160 EOF2 separately capture the two major centers in the uncertainty in the forced SLP response 161 (Fig. 1b). 162 Previous studies have investigated the uncertainty in circulation responses over the North 163 Pacific (e.g., Gan et al. 2017) and the North Atlantic (e.g., Harvey et al. 2015) separately. 164 However, the uncertainty in the forced SLP response over the northeastern Pacific (40°-65°N  165 and 160°-135°W) and over the northeastern Atlantic (60°-85°N and 20°W-10°E) is strongly 166 correlated (+0.770, p < 0.05, Fig. 2d). These uncertainties are partly captured by EOF1 and 167 EOF2, which are not well separated according to the North's rule of thumb (North et al. 1982). 168 In other words, the EOF analysis is not able to isolate the coherence of the SLP response over 169 the Pacific and the Atlantic in the ten CMIP5 models. Thus, we adopt a simple approach to 170 adequately represent this coherence: we add EOF1 and EOF2 to form a Pacific-Atlantic SLP 171 uncertainty pattern (PASLP) 2 that captures both uncertainties (Fig. 2c) and accounts for 31% of 172 the total variance. The forced SLP response in the ten CMIP5 models is projected onto the 173 PASLP pattern to get an index (PAISLP = PC1+PC2) for the SLP uncertainty among the models. 174 In the experiments, we prescribe SST and SIC patterns (SSTSLP and SICSLP) that are 188 computed by linear regression of the forced response in SST and SIC against PAISLP across 189 the ten CMIP5 models; the patterns are computed for each calendar month and have an 190 amplitude that corresponds to one unit of the standardized index. To attribute a physical 191 meaning of SST and SIC patterns, we illustrate the inter-model regression patterns of DJF SST 192 and SIC; these are referred to as the DJF SSTSLP and SICSLP ( Fig. 3a-b). Note that SSTSLP 193 is global and SICSLP is restricted to the Northern Hemisphere. 194

Methods
The uncertainty in the forced SST response is generally larger at high latitudes (Fig. 3a). In 195 the Northern Hemisphere, it is especially large over the Barents-Kara Sea and the midlatitude 196 North Atlantic. The DJF SSTSLP (associated with positive PAISLP) captures these SST 197 uncertainties, representing a substantial weakening in the MME forced response of the SST 198 gradient between the tropics and the northern high-latitudes (Fig. 3c). SSTSLP also represents 199 a warmer Northern Hemisphere and a cooler Southern Hemisphere than the MME climate 200 response, suggesting that PASLP also co-varies with a reduced inter-hemispheric SST gradient 201 ( Fig. 3d). Moreover, SSTSLP is associated with weakening in the zonal SST gradient between 202 the equatorial western and eastern Pacific, where the climatological mean SST is higher in the 203 western Pacific (warm-pool region). 204 The uncertainty in the forced response of the Arctic SIC is largest in the Kara Sea (~60°-205 90°E) and has a secondary maximum north of the Laptev Sea (~120°E) (Fig. 3b). The DJF 206 SICSLP represents a decline in the entire Arctic with a larger decline in these two seas. 207 Therefore, PASLP co-varies with the total Arctic sea ice extent (Fig. 3e) in a consistent manner 208 with the high-latitude SST (Fig. 3a). 209 3.2.2 "Time-slice" sensitivity experiments 210 We conducted three sets of AGCM "time-slice" sensitivity experiments forced by monthly-211 varying SST and SIC repeated for 60 annual cycles in CAM4 and 50 annual cycles in IFS (Table  212 2), such that the mean response in each model largely increases the signal-to-noise ratio. The 213 prescribed SST and SIC are based on the 2069-2098 monthly climatology in the CMIP5 214 RCP8.5 scenario (SSTMME and SICMME) and the monthly-varying SSTSLP and SICSLP (Section 215 3a; Fig. S2 and Fig. S3). Consistently, the prescribed radiative forcing (greenhouse gas 216 concentrations and aerosol concentrations) are the 2069-2098 monthly climatology from the 217 CMIP5 RCP8.5 scenario. 218 When the boundary conditions of SST and SIC in any set of sensitivity experiments were 219 modified, half the monthly-varying SSTSLP and SICSLP was added or subtracted. The 220 difference between the boundary conditions of two runs in one set of experiments is equivalent 221 to one unit of the inter-model regression patterns of SST and SIC. Thus, the magnitude of 222 atmospheric response can be directly compared to that of the inter-model regression pattern 223 from the ten CMIP5 models. 224 In the first set of experiments (run1 and run2), both SST and SIC were modified and they 225 are called SST+SIC perturbation runs. In run1, half the monthly-varying SSTSLP and SICSLP 226 was added to the 2069-2098 monthly climatology of SST and SIC respectively (SSTMME and 227 SICMME) to form the boundary conditions. Conversely, in run2, half the monthly-varying 228 SSTSLP and SICSLP was subtracted from SSTMME and SICMME to form the boundary conditions. 229 It is assumed that SST and SIC change coherently as the inter-model regression analysis (Fig.  230 S2 and Fig. S3), so we did not perform experiments with the boundary conditions 231 SSTMME+SSTSLP with SICMME−SICSLP, and SSTMME−SSTSLP with SICMME+SICSLP. In the 232 second set of experiments (run3 and run4), SST was changed as in run1 and run2 while SIC 233 was the SICMME; they are called SST perturbation runs. In the third set of experiments (run5 234 and run6), SIC was changed as in run1 and run2 while SST was the SSTMME; they are called 235 SIC perturbation runs. 236 We follow Screen et al. (2013)'s approach to ensure that SST and SIC are consistent with 237 each other in the second and third sets of experiments (runs 3-6). In particular, for grid cells 238 with sizeable SIC perturbation (|SICSLP|  +0.1 (fraction)), the boundary condition of SST was 239 modified to SSTMME in the SST perturbation runs (run3 and run4) and to SSTMME±0.5×SSTSLP 240 in the SIC perturbation runs (run5 and run6). Otherwise the boundary condition of SST was set 241 as in Table 2. The monthly-varying SIC boundary condition of run 1 and run 5 (larger SIC and 242 SST response) and the monthly-varying SIC boundary condition of run 2 and run 6 (smaller 243 SIC and SST response) are shown in Fig. S4 and Fig. S5, respectively. 244

Diagnostics 245
Because the SST perturbation represents a weaker meridional SST gradient between the 246 low and high latitudes and the SIC perturbation represents a stronger decline of pan-Arctic SIC, 247 we investigate if the response in the zonal-mean streamfunction is related to the SST and SIC where ζ is the absolute vorticity. The overbar denotes the basic state and the prime denotes the 256 inter-model regression of the forced response against PAI∆SLP across the ten CMIP5 models or 257 the response in the AGCM sensitivity experiments. On the R.H.S. of (1), the first two terms are 258 the contribution from the vortex stretching, where in the first (second) term a stronger 259 convergence of the wind response (climatology) would enhance the cyclonic Rossby wave 260 source, while stronger divergence would enhance the anticyclonic Rossby wave source. The 261 third and fourth terms are the contribution from the vorticity advection by the divergent wind, 262 where the region with a strong vorticity gradient would enhance or reduce the Rossby wave 263 source. When the response of Rossby wave source is compared to the pressure response, we 264 could deduce if the tropical-midlatitude interaction is crucial for the pressure response. 265 To further demonstrate if the midlatitude atmospheric response is a Rossby wave train, we 266 will present the horizontal component of stationary wave activity fluxes at the 250-hPa (Takaya 267 and Nakamura 2001): 268

Mechanisms underlying the CMIP5 SLP uncertainty pattern 276
We now identify the potential mechanisms explaining the relation between the SLP 277 uncertainty pattern (PASLP) and the SST and SIC uncertainty patterns in the ten CMIP5 models. 278 The regression of PAISLP against the forced DJF SLP response reveals increasing SLP in the 279 northeastern Pacific and the northeastern Atlantic (Fig. 4a). These positive SLP regions 280 coincide with positive 250-hPa height regions (Fig. 4b), suggesting an equivalent barotropic 281 high structure. Moreover, positive PAISLP is associated with increasing SLP over the tropical 282 Indo-Pacific and decreasing SLP in the sea ice region, including the Arctic and Northeast 283 America with a maximum around the Canadian Archipelago (Fig. 4a). It is also associated with 284 a dipole pattern over Eurasia, with increasing SLP near the latitude of the Icelandic low and 285 decreasing SLP in the Mediterranean and the Middle East (Fig. 4a). We will investigate how 286 these SLP uncertainties are related to the zonal-mean meridional cells, the air-sea interactions 287 and the midlatitude Rossby wavetrain. 288 From the zonal-mean perspective, SSTSLP represents anomalous warming in the Northern 289 Hemisphere and anomalous cooling in the Southern Hemisphere (Fig. 3d). This SST pattern is 290 associated with more zonal-mean precipitation in the northern tropics and less zonal-mean 291 precipitation in the southern tropics, i.e., northward shift of the inter-tropical convergence zone 292 (ITCZ; Fig. 5a). The northward shift of the ITCZ is associated with 10-20% weakening of the 293 Hadley circulation except in the upper troposphere (Fig. 5a), where the DJF climatological 294 Hadley cell represents a circulation from the Southern Hemisphere to the Northern Hemisphere. 295 The northern hemisphere Ferrel and polar cells also slightly weaken (Fig. 5a). These changes 296 are consistent with an expected weakened poleward heat transport in the Northern Hemisphere 297 (Kang et al. 2009;Schneider et al. 2014). Note that PAISLP is weakly correlated to the forced 298 responses in the zonal-mean upper-tropospheric temperature (<0.2°C; Fig. S6a) and the global-299 mean surface temperature. In other words, the tropical precipitation changes associated with 300 PASLP are not driven by the moist convective processes related to global warming. Consistently, 301 the circulation uncertainties described by PASLP are not strongly related to the climate 302 sensitivity. Hence, it is important to investigate the factors other than the climate sensitivity 303 contributing to the SLP uncertainties. 304 Regionally, SSTSLP over the tropical region is strongest in the eastern Pacific (Fig. 3a). The 305 stronger warming over the tropical eastern Pacific accompanies stronger local convection and 306 precipitation (~10°-15°N and 150°-130°W; Fig. 6a). These accompany lower velocity 307 potential and stronger divergent wind at the 250 hPa blowing northeastwards from the tropical 308 region to the midlatitudes in the eastern North Pacific (Fig. 6b). The convergent wind is 309 associated with a cyclonic (i.e., positive sign) Rossby wave source (Fig. 6c). Meanwhile, the 310 region with a large gradient in the velocity potential is associated with an anticyclonic (i.e., 311 negative sign) Rossby wave source and an anomalous anticyclone over the central North Pacific. 312 The anticyclonic Rossby wave source accompanies the emanation of a Rossby wavetrain from 313 the anomalous anticyclone. This wavetrain propagates eastward to an anomalous cyclone over 314 the North American west coast and then propagates southeastward to the Gulf of Mexico (Fig.  315   6c). It appears that the tropical-midlatitude interactions over the Pacific is essential for the 316 northeastern Pacific SLP uncertainty. Moreover, because the wavetrain does not propagate 317 further from North America to the northeastern Atlantic, the local air-sea interaction appears to 318 be important in the northeastern Atlantic SLP uncertainty. 319

4.2
Atmospheric impact of the SST+SIC uncertainty patterns 320 Next we use AGCM experiments to assess the extent to which the SST and SIC drive PASLP 321 and whether the mechanisms identified above hold. The SLP response of CAM4 in the 322 SST+SIC perturbation runs has a positive center over the midlatitude North Pacific and a dipole 323 pattern over the North Atlantic (Fig. 4c). The pattern has features similar to the CMIP5 inter-324 model regression although the positive and negative centers shift westward (Fig. 4c vs. Fig. 4a); 325 the possible reason causing the westward shift will be studied later in this section. The pattern 326 correlation between the CAM4 SLP response and the SLP inter-model regression pattern from 327 the ten CMIP5 models increases from +0.381 to +0.731 when the CAM4 SLP response is 328 shifted zonally by 35°. The CAM4 simulations also shows good correspondence between the 329 SLP pattern and the 250-hPa height pattern (Fig. 4c,d). On the other hand, the SLP and 250-330 hPa height response of IFS is generally weaker than CAM4, with a positive response over the 331 North Pacific and a dipole-like structure over the North Atlantic (Fig. 4e,f). In short, CAM4 332 and IFS tend to simulate coherent atmospheric responses over the oceans. Comparatively, the 333 CAM4 response is closer to the inter-model regression from the ten CMIP5 models, and the 334 IFS response is weaker, especially over the North Pacific 335 The zonal-mean response of the two AGCMs shows suppressed precipitation in the southern 336 tropics and enhanced precipitation in the northern tropics, and the ITCZ shifts towards the 337 northern tropics (Fig. 5b,c). These precipitation responses are unlikely driven by the moist 338 convective processes because the response in upper-tropospheric warming is weak (Fig. S6b,c). 339 Moreover, the two AGCMs robustly simulate substantial weakening in the Hadley cell between 340 10º S and 10º N (Fig. 5b,c), which accompanies weaker poleward heat transport in the Northern 341 Hemisphere (Kang et al. 2009;Schneider et al. 2014). The tropical zonal-mean circulation 342 responses of the two AGCMs are generally consistent with the inter-model regression from the 343 ten CMIP5 models. In mid-and high-latitudes, CAM4 simulates a weak response in Ferrel and 344 Polar cells, where the southern (northern) edge of the Ferrel cell is enhanced (weakened) (Fig.  345   5b). On the other hand, IFS simulates more substantial weakening in Ferrel and polar cells (Fig.  346  5c). The center of these responses is located at 45º N and 65º N respectively (Fig. 5c), which 347 coincides with the two centers of the dipole-like response in the Atlantic (Fig. 4e,f). Therefore, 348 IFS has a stronger response in the zonal-mean circulation than CAM4 and what is found in the 349 CMIP5 inter-model difference. 350 Regionally, in CAM4, the Rossby wavetrain response from the North Pacific to the North 351 Atlantic (Fig. 7e) is associated with strong tropical-midlatitude interaction over the North 352 Pacific (Fig. 7c), which is related to the enhanced tropical rainfall over the tropical eastern 353 Pacific (Fig. 7a). The wave activity fluxes propagate northeastward from the North Pacific to 354 North America and then the propagation turns eastward to the high-latitude North Atlantic (Fig.  355 7e). The Rossby wavetrain response emanates from the regions with an anticyclonic Rossby 356 wave source at 30°-40°N over the northwestern Pacific and at 15°N over the northeastern 357 Pacific, which are due to a larger gradient in the velocity potential (Fig. 7c). Compared to the 358 inter-model regression from the ten CMIP5 models, the Rossby wavetrain response in CAM4 359 shifts westward (Fig. 7e vs. Fig. 6c). Although the significant rainfall response in CAM4 is 360 extended westward from ~155°W to 180° (Fig. 7a vs. Fig. 6a), the associated divergent wind 361 response over the tropical Pacific is not shifted westward (Fig. 7c vs. Fig. 6b). The westward 362 extension of the precipitation response cannot explain the westward shift of the Rossby 363 wavetrain response. Indeed, the midlatitude Northwestern Pacific (where the Rossby wavetrain 364 emanates) has an easterly wind response, which can be explained by the non-linear forcing by 365 transient eddies that is not considered in Eq. (1). Specifically, convergence of the low-frequency 366 (8-day low-pass filtered) E vector propagating westward from the northeastern Pacific 367 corresponds to the easterly wind response (Hoskins et al. 1983 ; Fig. S7a). The high-frequency 368 eddy forcing is much weaker than the low-frequency eddy forcing (Fig. S7a,b). The westward 369 propagating low-frequency E vector over the midlatitude North Pacific may explain the 370 westward shift of the geopotential height response and the associated Rossby wavetrain 371 response. In short, the Rossby wavetrain response in CAM4 is related to the tropical-midlatitude 372 interactions and the low-frequency transient eddy forcing. Note that the local air-sea interaction 373 may also be important in the North Atlantic circulation response, because there is limited wave 374 propagation from the North Pacific to the North Atlantic (Fig. 7e). 375 In IFS, the Rossby wavetrain response over the North Atlantic is separated from the 376 response over the North Pacific. The Rossby wavetrain response over the North Pacific in IFS 377 is much weaker than that in CAM4 (Fig. 7f vs. Fig. 7e), which is related to weaker responses 378 of tropical precipitation and tropical-midlatitude interaction (Fig. 7b,d vs. Fig. 7a,c). The wave 379 activities fluxes over the North Atlantic are emanated from the region with anticyclonic Rossby 380 wave source (Fig. 7f). This is associated with enhanced precipitation and stronger divergent 381 wind at 250-hPa. This appears to be related to strengthening of the local air-sea interactions at 382 the mid-latitude North Atlantic. The role of transient eddies is not investigated due to the lack 383 of data availability. The above results suggest that the responses of CAM4 and IFS to the 384 SST+SIC perturbation have different dynamical mechanisms, where the response of CAM4 is 385 closer to the CMIP5 inter-model difference. 386

Separate impact of the SST and SIC uncertainty patterns 387
The SLP response in the SST+SIC perturbation runs of the AGCMs is not fully consistent 388 with the forced SLP response from ten CMIP5 models. The CAM4 response has a spatial 389 pattern similar to the CMIP5 inter-model difference albeit with a zonal shift, whereas the IFS 390 response is generally weaker than CAM4 and is associated with different dynamic mechanisms. 391 The difference in the SLP response between the AGCMs and CMIP5 models, as well as the 392 difference between the two AGCMs, could be contributed by the SST perturbation and/or the 393 SIC perturbation. Hence, it is instructive to know the influence of the SST and SIC perturbations 394 on the SLP response in the SST+SIC perturbation runs (Figs. 8a-d and 9a-d), and to know 395 whether the responses to these perturbations are linear (Figs. 8e,f and 9e,f). The impact is 396 predominantly linear when the SST+SIC response is equal to the sum of the separate responses 397 of SST and SIC. 398 For CAM4, the linear sum of pressure responses in the SST perturbation runs and the SIC 399 perturbation runs (shading in Fig. 8e,f) is broadly similar to the pressure responses in the 400 SST+SIC perturbation runs except those over the Eurasian continent (contour in Fig. 8e,f), 401 suggesting that the pressure responses of CAM4 in the SST+SIC perturbations can be mainly 402 linearly explained by their SST and SIC components. Specifically, the pressure responses of 403 CAM4 in the SST+SIC perturbations are largely explained by the SST perturbation, except for 404 the region with sea ice (Fig. 8a vs. Fig. 8e). The positive pressure response over the high-latitude 405 Euro-Atlantic region in the SST+SIC perturbation runs (contour in Fig. 8e,f) is split into two 406 centers over the Baffin Bay (~75°N, 70°W) and the Barents-Kara Sea in the SST perturbation 407 runs (Fig. 8a,b). The positive pressure response near Greenland (part of the dipole-like response) 408 is contributed by both SST and SIC perturbations (Fig. 8a,c). The SIC perturbation is the 409 dominant factor of the Arctic SLP response, and it contributes to the negative SLP response 410 over the Canadian Archipelago (Fig. 8c). Note that the pressure responses to the SIC 411 perturbation have smaller signal-to-noise ratios than those responses to the SST perturbation. 412 For IFS, the pressure response pattern in the SST perturbation runs is closer to that in the 413 SST+SIC perturbation runs than the SIC perturbation runs, but the patterns differ (top panel in 414 Fig. 9 vs. contour in the bottom panel in Fig. 9). The weak pressure responses over the Arctic 415 and the North Pacific in the SST+SIC perturbation runs (contour in Fig. 9e,f) are the result of 416 the opposing effect of the SST perturbation (Fig. 9a,b) and the SIC perturbation (Fig. 9c,d). 417 However, the linear sum of these pressure responses (shading in Fig. 9e,f) does not resemble 418 the circulation response in the SST+SIC perturbation runs (contour in Fig. 9e,f). Neither SST 419 perturbation nor SIC perturbation can reproduce the dipole-like pressure response over the 420 North Atlantic and the pressure response over the high-latitude Eurasia (Fig. 9a-d). The non-421 linear effect of the pressure responses to the SST and SIC perturbations is also strong in the 422 high-latitude Eurasia, East Asia and the Arctic upper troposphere. The above results suggest 423 that the pressure response in the SST+SIC perturbation runs of IFS is due to the non-linear 424 effect of the SST and SIC perturbations. 425 The SLP responses of CAM4 and IFS in the SST perturbation runs are consistent in the 426 North Pacific (Fig. 8a and Fig. 9a), and match the sign of the inter-model regression pattern 427 from the ten CMIP5 models albeit with a westward shift (Fig. 4a). The SST perturbation causes 428 substantial weakening in the Hadley cell and a stronger ITCZ in the northern tropics (Fig. 10). 429 Regionally, the rainfall is enhanced in the northern tropical Pacific (Fig. 11a,b). This 430 accompanies stronger divergence over the tropics and stronger convergence over the 431 midlatitudes at the 250 hPa, representing stronger tropical-midlatitude interaction over the 432 North Pacific (Fig. 11c,d). The anticyclonic Rossby wave source is enhanced at the region with 433 a larger gradient in the velocity potential at 15°N (Fig. 11c,d), which is associated with 434 emanation of the wave activity fluxes from the Pacific to North America (Fig. 11e,f). 435 Indeed, the wave activity fluxes are also emanated from the mid-latitude region with 436 positive eddy height response (Fig. 11e,f). The increase in the eddy height is partly related to 437 the tropical midlatitude interactions near the central North Pacific, where stronger convergent 438 wind corresponds to a cyclonic Rossby wave source (Fig. 11c,d). In CAM4, part of the increase 439 in the eddy height is related the low-frequency transient eddy forcing, where the convergence 440 of E vector near 30°N,180° corresponds to an easterly wind anomaly associated with the 441 positive height anomaly (Fig. S6c). In IFS, the Rossby wavetrain propagates equatorward from 442 the northwestern Pacific (Fig. 11f), which is different from the results in its SST+SIC 443 perturbation runs (Fig. 7f) and the CAM4 runs (Fig. 11e). This again suggests that the dynamics 444 for the responses of CAM4 and IFS are different. Despite that, the above results suggest that 445 the SST perturbation from the North Pacific contributes to the high-pressure response over the 446 North Pacific and the response of Rossby wave propagation from the North Pacific to North 447

America. 448
In addition to the positive SLP response over the North Pacific, the two AGCMs simulate 449 consistently a negative SLP response over the mid-latitude North Atlantic in the SST 450 perturbation runs (Fig. 8a and Fig. 9a). However, their responses in the SST perturbation runs 451 are different over the continents and the Arctic. On one hand, IFS has a stronger response of 452 the zonal-mean mass streamfunction to the SST perturbation than CAM4, where the Ferrel and 453 polar cells weaken and shift northward albeit not statistically significant (Fig. 10b). The 454 associated weakening of the rising motion in the subpolar latitudes around 70°N is accompanied 455 by an increase in pressure over the subpolar region across the Canadian Archipelago and the 456 North Atlantic (Fig. 9a,b). On the other hand, CAM4 does not have responses in the Ferrel and 457 polar cells to the SST perturbation (Fig. 10a); its midlatitude circulation response is mainly 458 explained by the wavetrain response (Fig. 11e). 459 In response to the SIC perturbation, the two AGCMs simulate consistently a negative SLP 460 response over the Arctic (Fig. 8e and Fig. 9e). However, their response outside the Arctic 461 diverges, including over the North Pacific, Scandinavia and the Mediterranean. That is, the 462 difference in these SLP responses between the two AGCMs is statistically significant. There is 463 no consistent response in the meridional circulations (figures not shown). Because SIC does not 464 robustly influence the midlatitude circulation, we do not analyze in-depth the regional 465 circulation features over the mid-latitudes and the associated dynamics. perturbations. The CAM4 response and its associated dynamics are closer to the CMIP5 inter-501 model regression, suggesting that the forced SLP uncertainty over the northeastern Pacific from 502 the ten CMIP5 models is associated with the tropical-midlatitude interaction related to the SST 503 uncertainty over the Pacific. 504 5.1.2 Dipole-like pressure response over the Euro-Atlantic region 505 In the ten CMIP5 models the dipole-like SLP response in the North Atlantic appears to result 506 from both local response to SST and remote influences from the Pacific. In CAM4, the SLP 507 responses in the SST perturbation runs and the SST+SIC perturbation runs are similar and again 508 show some correspondence to the ten CMIP5 SLP uncertainty pattern. In these CAM4 509 simulations, the Rossby wavetrain simulated in the SST perturbation runs is triggered by SST 510 over the Pacific. This is consistent with the results of Delcambre et al. The SLP responses of the two AGCMs to the SIC perturbation are consistent in the Arctic 526 and show lower SLP associated with less sea ice, as in the ten CMIP5 models. However, the 527 midlatitude SLP response of the two AGCMs diverges. Because the difference in the 528 midlatitude SLP response between the two AGCMs is statistically significant (Fig. S8) of PAISLP against the forced response in the DJF 10-hPa height is not statistically significant 539 (Fig. 12a). That means the uncertainty in the forced response of the stratospheric polar vortex 540 is not considerably correlated to PASLP. Consistently, the stratospheric response to the SIC 541 perturbation in the two AGCMs has small signal-to-noise ratios (Fig. 12f,g). Compared to 542 CAM4, IFS has a much higher model top and its stratospheric dynamics might be better 543 resolved, but the stratospheric responses of IFS to the SST perturbation and the SST+SIC 544 perturbation are weaker than CAM4 (Fig. 12b-e). Therefore, the impact of stratosphere on the 545 tropospheric circulation is not apparent in this study. Note that the results in this study are 546 related to the uncertainty in the forced SLP responses over the northeastern Pacific and the 547 northeastern Atlantic (i.e., PASLP). The weak linkage between PASLP and the stratospheric 548 circulation does not imply the uncertainty in the forced response of the tropospheric circulation 549 being unrelated to the stratospheric circulation in other models. Our results could also be 550 affected by the small number of high-top models (three out of ten models) (Charlton-Perez et 551 al. 2013) and the perturbation pattern of the Arctic SIC in this study. 552

Conclusions 553
The future projection of the winter SLP in the northeastern Pacific and the northeastern 554 Atlantic has a large inter-model spread, which covaries with the large-scale SST gradients and 555 the total Arctic sea ice extent. In this study, the sensitivity experiments of CAM4 and IFS 556 revealed that atmospheric responses to the same SST and SIC boundary conditions are generally 557 large, with more coherent responses over the oceans (in terms of the sign of response) and less 558 coherent remote responses to these boundary conditions. Specifically, we have learnt the 559 following points: 560 • The uncertainty in the forced SLP response over the northeastern Pacific is about twice as 561 large as the uncertainty related to internal climate variability. This uncertainty is better 562 explained by the SST perturbation, and it is associated with tropical-midlatitude 563 interactions and the propagation of a Rossby wavetrain towards North America. The 564 relative contribution from the tropical and extratropical Pacific should be investigated in 565 future; 566 • The uncertainty in the forced SLP response over the northeastern Atlantic is of similar 567 strength as internal climate variability and is even weaker than it at high latitudes. This 568 uncertainty is better explained by the joint effect of SST and SIC perturbations. It appears 569 to be related to a Rossby wavetrain from the North Pacific and with local air-sea 570 interactions, with the first more important in CAM4 and the second more important in IFS; 571 • The uncertainties over the northern hemisphere continents and at high latitudes appear to 572 depend sensitively on the atmospheric model. Furthermore, the response to SST and SIC 573 perturbations can be non-linear in some models (e.g., IFS), while quite linear in others (e.g., 574 CAM4). Thus, we should be cautious when using single climate model to understand the 575 physical mechanism responsible for the uncertainty in future climate projections from 576 multiple models; 577 • Future projections of the winter SLP might not be improved by constraining only the SST 578 and SIC projections. We should investigate other factors contributing to the inter-model 579 spread in the winter SLP (e.g., the atmosphere-SST-SIC coupled dynamics) in order to 580 provide more accurate climate projections. 581 One limitation of our study is that the uncertainty in future climate projections is computed 582 from only ten CMIP5 models with only three ensemble members, where these models have 583 more substantial weakening in the Icelandic low than the whole CMIP5 models. The SLP 584 pattern from the inter-model EOF analysis, as well as the corresponding SST and SIC patterns, 585 might also be sensitive to the number of models. To reduce uncertainties related to internal 586 variability and to make the analysis more representative, more models with more ensemble 587 members are required to separate the forced response from the internal climate variability, Centre at Linköping University (NSC). 609 Data availability statement: The CAM4 and IFS data will be available upon request and 610 will be located on the National Infrastructure for Research Data (NIRD) in Norway. 611 772   Table 1  dots denotes the grid points exceeding the 95% confidence interval of the zonal-mean mass 819 streamfunction and the zonal-mean precipitation, respectively. 820 Student's t-test. 834 Fig. 9. As in Fig. 8, but for the response in IFS. 835 Fig. 10. As in Fig. 5, but for the response of CAM4 and IFS in the SST perturbation runs. 836 Fig. 11. As in Fig. 6, but for the response in the SST perturbation runs: (left) CAM4 and (right) 837