The Weather Research and Forecasting (WRF) model is used to explore the sensitivity of the large-scale atmospheric energy and moisture budgets to prescribed changes in Arctic sea ice and sea surface temperatures (SSTs). Observed sea ice fractions and SSTs from 1996 and 2007, representing years of high and low sea ice extent, are used as lower boundary conditions. A pan-Arctic domain extending into the North Pacific and Atlantic Oceans is used. ERA-Interim reanalysis data from 1994 to 2008 are employed as initial and lateral forcing data for each high and low sea ice simulation. The addition of a third ensemble, with a mixed SST field between years 1996 and 2007 (using 2007 SSTs above 66°N and 1996 values below), results in a total of three 15-member ensembles. Results of the simulations show both local and remote responses to reduced sea ice. The local polar cap averaged response is largest in October and November, dominated by increased turbulent heat fluxes resulting in vertically deep heating and moistening of the Arctic atmosphere. This warmer and moister atmosphere is associated with an increase in cloud cover, affecting the surface and atmospheric energy budgets. There is an enhancement of the hydrologic cycle, with increased evaporation in areas of sea ice loss paired with increased precipitation. Most of the Arctic climate response results from within-Arctic changes, although some changes in the hydrologic cycle reflect circulation responses to midlatitude SST forcing, highlighting the general sensitivity of the Arctic climate.
 Analysis of the satellite data record reveals downward linear trends in Arctic sea ice extent for every month, largest (−12.4% per decade) at the end of the melt season in September [Stroeve et al., 2011]. September sea ice extents for the last five years are the five lowest in the record, with the record minimum occurring in 2007 [Stroeve et al., 2008]. While natural variability has played a role in the observed decline, anthropogenic forcing appears to be the major driver [e.g., Kay et al., 2011b].
 A range of global climate model simulations project that the Arctic Ocean will become seasonally ice-free somewhere between the middle of the present century to well beyond the year 2100. Local effects of reduced sea ice cover include an increase in absorbed solar radiation by the ice-free ocean, raising sea surface temperatures [Steele et al., 2008], and increasing the vertical heat fluxes to the atmosphere. Based on reanalysis data, Serreze et al.  and Screen and Simmonds identified enhanced ocean-to-atmosphere heat fluxes due to ice loss as a primary contributor to observed strong increases in surface and lower-tropospheric air temperature over the Arctic Ocean in autumn and winter. On climatic time scales, through changes in ocean stratification and the thermohaline circulation, loss of the Arctic sea ice cover could potentially have significant global impacts.
 Numerous model sensitivity studies have been conducted to isolate the effects of Arctic sea ice loss on the atmosphere. Most have focused on the autumn and winter seasons. Budikova  provides a comprehensive review. As for some examples, Deser et al.  and Magnusdottir et al.  used the National Center for Atmospheric Research (NCAR) Community Climate Model (CCM) version 3 (an atmospheric general circulation model) to investigate the atmospheric response to observed trends in both North Atlantic sea surface temperature (SST) and sea ice. Positive SST anomalies reduced the atmospheric static stability directly over the regions with changed lower boundary conditions, while the opposite was true for the negative SST anomaly experiments. They also noted a stronger atmospheric response to warm rather than cold SST anomalies. Along with Strong et al. , they also found that sea ice patterns associated with a positive NAO phase tend to generate atmospheric patterns that resemble the negative phase of the NAO, suggesting a negative circulation-sea ice feedback. Similarly,Alexander et al. , using the community atmospheric model (CAM) version 3, found a significant shallow atmospheric response to reduced sea ice cover characterized by stronger upward surface heat fluxes, atmospheric warming, and an enhanced hydrologic cycle.
 Several recent studies have focused on potential atmospheric responses to the record low ice extent observed in 2007. Strey et al. ran the Weather Research and Forecasting (WRF) model (a mesoscale atmospheric model) with sea ice extent for 2007, the record low September sea ice extent in the satellite record, and 1984, a year close to the median observed extent in the satellite record. The direct response to reduced ice cover includes a persistent increase in near-surface temperature and a local decrease in sea level pressure (SLP), with a nearly barotropic ridge/trough response over North America in autumn. Using the ECHAM5 atmospheric general circulation model forced with the sea ice and SST conditions in 2007,Blüthgen et al.  find a large surface warming (compared to climatology) over the eastern Arctic and increased ocean heat uptake. In autumn, an additional heat flux of 60 W m−2 was released to the atmosphere from the warmed ocean. They also find a significant reduction of up to −2 hPa in SLP in the eastern Arctic in July, August, and September; a pattern that acts to export additional ice from the Arctic, but find no significant SLP response in the autumn. Kay et al. [2011a] used CAM version 4 and summer surface conditions in 2007 and find modest increases in near surface temperature and humidity, but much larger decreases in stability and increased relative humidity aloft.
 There is growing observational evidence for circulation responses to reduced sea ice. Overland and Wang  argued that the observed surface warming associated with sea ice loss has led to an increase in the 1000–500 hPa thicknesses over the Arctic Ocean and a stronger Beaufort Sea high at 850 hPa. They also speculate that reduced ice extent helps to explain the diminished occurrence of the positive mode of the Arctic Oscillation in recent years and its replacement by a more meridional pattern, termed the Arctic Dipole. Francis et al. used satellite and conventional meteorological observations to show that reduced autumn sea ice extent affects large-scale atmospheric features in the following autumn/winter through a slackening of the poleward thickness gradient and associated weakening of the polar jet stream, suggesting a long-lived extra-polar response to reduced Arctic sea ice. The subsequent study ofFrancis and Vavrus  lends further weight to this view; in particular, they argue that with a reduced poleward gradient in the 1000–500 hPa thickness, easterly wave propagation slows and tends to follow a higher amplitude trajectory, resulting in slower moving weather systems. Strong et al.  used a vector autoregressive model to find a negative feedback relationship between the North Atlantic oscillation (NAO) and sea ice anomalies in the Greenland, Barents, and Labrador Seas.
 Numerous questions remain regarding the atmospheric responses to reduced sea ice and Arctic Ocean warming. What is the vertical structure of expected atmospheric warming and moistening? What changes will occur in cloud type and height? Will increased evaporation linked to open water and a warmer ocean result in increased precipitation, and if so, will changes be limited to the Arctic Ocean or extend to other areas? Finally, how do extra-polar SSTs affect the Arctic atmosphere and how do within-Arctic changes impact lower latitudes? The present paper tries to address some of these questions on the basis or regional modeling simulations. Similar to the recent studies ofStrey et al. , Kay et al. [2011a] and Blüthgen et al. , we focus on isolating atmospheric responses to the record low sea ice extent observed in 2007. We make comparisons with the findings from these studies in section 5.
2. Model Experiments
 We carried out three experiments using WRF ARW (Advanced Research WRF) Version 3.2.0. This model version has been used for both very short forecasts [Benjamin et al., 2004] and dynamical downscaling of regional climates [Bromwich et al., 2010]. Similar to Porter et al. , the experiments were run on a domain of 205 × 275 grid points (Figure 1) on a polar stereographic projection, using a horizontal resolution of 50 km on an Arakawa C-grid, and 40 levels in the vertical. The WRF ARW (hereafter referred to as “WRF”) is a non-hydrostatic, fully compressible model that uses terrain- following vertical levels. The model top is set to a constant-pressure of 50 hPa.Skamarock et al.  discuss the specifics of the ARW model dynamics and physical parameterizations. Guided by Cassano et al. , we use the physics options showing the best performance at high latitudes. Some of these options include Goddard microphysics and Grell-Devenyi cumulus scheme for large-scale and sub-grid scale cloud representation, the Noah land surface model (LSM) which has the ability to represent fractional sea ice cover [Bromwich et al., 2009], the Monin-Obhukov (Janjic Eta) surface layer, and the Mellor-Yamada-Janjic planetary boundary layer (PBL).
 The initial atmospheric state, lateral forcing, and lower boundary conditions over land were specified using data from the European Centre for Medium-Range Weather Forecasts (ECMWF) Interim Reanalysis (ERA-Interim hereafter). Full pressure-level ERA-Interim data were obtained from the ECMWF data server (http://data-portal.ecmwf.int/data/d/interim_daily/levtype=pl/). ERA-Interim is one of the most modern global reanalyses, employing new physics parameterizations, a 12-h four-dimensional variational data assimilation (4DVAR) system, and variational bias correction of observed satellite-derived radiances [Dee et al., 2011]. Additionally, we include observed sea ice concentrations from the National Snow and Ice Data Center (NSIDC) daily bootstrap product [Comiso, 1990], available at http://nsidc.org/data/nsidc-0002.html.
 For each year from 1994 to 2008, the WRF simulations were started on 16 June and run continuously through 1 December using the corresponding ERA-Interim data for initial and lateral boundary conditions. For each of the 15 model years, WRF was run first with daily sea ice concentration and SSTs from 1996 (“high1996”) and then again with daily 2007 (“low2007”) observations. Hence, only two of the 30 simulations, 1996 with 1996 lower boundary conditions (LBCs) and 2007 with 2007 LBCs, have consistent initial, lateral, and lower forcing data. The years 1996 and 2007 have the highest and lowest September mean sea ice extent in the satellite record, respectively. This experiment design yields two parallel streams of model simulations, where each represents an Arctic atmosphere forced by imposed sea ice concentrations and SSTs from either high or low sea ice years. If we consider each model year as an individual member of an ensemble, then we have created two 15-member ensembles. By using observed lateral boundary conditions we introduce not only “noise” but also observed variability in large-scale modes of the climate system, such as ENSO, that will force our simulation domain. We feel that this ability to capture varying large-scale modes of climate variability in the lateral forcing is an important distinction compared to other methods of generating an ensemble of simulations. However, specifying sea ice in WRF eliminates many atmosphere-ice-ocean feedbacks.
 A third ensemble was then created that helps resolve the fraction of the total response, seen in the low2007 and high1996 experiments, that is due to changes in extra-polar SSTs. This third ensemble of 15 members uses 1) 2007 sea ice, 2) SSTs from 2007 north of 66°N, and 3) SSTs from 1996 south of 66°N. In this way, the third ensemble, referred to as the “mixedSST” experiment, is identical to the low2007 ensemble except that SSTs below 66°N latitude are set to the values from 1996.
 While the model experiments are similar to those of Strey et al.  who also used WRF, and Blüthgen et al.  who used ECHAM5, to explore the atmospheric response to 2007 sea ice conditions, there are many important differences. In comparison with Strey et al. , we use lateral forcing data with a realistic range of natural variability, specify Arctic SSTs that match the observed sea ice cover, include the summer months to better observe the transient response to seasonal sea ice loss and increased SSTs, and used a newer version of the WRF model (3.2.0 versus 3.0.1). Some of the improvements in WRF 3.2.0 important for high latitude domains are the inclusion of fractional sea ice, snow-specific thermal diffusivity, proper use of the latent heat of sublimation and vaporization, and more realistic soil moisture and snow density values over glacial ice. Further description of polar improvements included in WRF version 3.2.0 can be found inBromwich et al. . While much more similar to our study, there are a few important differences with Blüthgen et al. , such as their use of a global model at much lower horizontal (T63) and vertical resolution (31 levels) and monthly mean sea ice and SST fields interpolated to daily values. Additionally, their model setup also allowed them to discern the importance of extra-polar SST changes on the evolution of the within-Arctic response. However, by using climatological SSTs, they fail to capture the varying modes of atmospheric circulation that we include in the current paper by using observed lateral forcing data for 15 different years. Taken as a whole, the many similarities between these three studies act to increase the robustness of any direct comparisons.
 With respect to some known model biases, Porter et al.  found that the albedo of sea ice in WRF, set at 0.8, is too high in the summer. In agreement with previous studies looking at surface albedo's effect on mesoscale weather models [Cassano et al., 2001], the increase in reflected shortwave in WRF nevertheless has few implications for evolution of the model's atmosphere when considering a melting snow or ice surface. WRF has several large biases when compared to the forcing data [Cassano et al., 2011] and atmospheric reanalyses [Porter et al., 2011], including overly high SLP in the North Pacific and surface temperatures that are too low over the central Arctic Ocean. One example of a model bias that could affect the response is deficiencies in the representation of low-level temperature inversions, which could inhibit or enhance the response if inversions are too strong or weak, respectively. While spectral or grid nudging reduces such biases [Porter et al., 2011], nudging can overly constrain the model for experiments such as those presented here by reducing the model sensitivity to changes in lower boundary conditions. Because we are analyzing the difference between two sets of simulations with high and low sea ice states, such biases will likely have only small impacts on our conclusions.
 Unlike experiments with coupled models, these experiments with WRF do not allow for feedbacks from the atmosphere back to the sea ice cover and hence isolate the one-way response of the atmosphere to imposed lower boundary changes. By isolating the ocean and ice to atmosphere response, model results are not necessarily realistic. For example, elimination of the albedo-temperature feedback likely leads to an underestimation of the response. However,Deser et al. find that the response of the net surface heat flux to reduced sea ice is similar in both atmosphere-only and coupled climate models.
 The main analyses employed are comparisons of ensemble means and variances of the low2007, high1996, and mixedSST sets of simulations. Differences between these three ensembles are considered the “response” or the sensitivity of the Arctic atmosphere, at least in WRF, to imposed changes in both sea ice cover and SSTs. By taking the difference (response) between different pairs of ensemble means and comparing their variances, we are able to isolate WRF's response to different lower boundary forcings and their significance. By differencing the low2007 and high 1996 simulations, we can examine the model's response to all LBC differences (both polar and lower-latitude SSTs) between the high and low sea ice years. To isolate just the model response to within-Arctic sea ice and SST changes, we difference the mixedSST and high1996 ensemble means, which isolates the low sea ice concentration and warm polar SSTs observed in 2007. Comparing the low2007 and mixedSST scenarios, where the only differences are lower-latitude SSTs, will reveal any within-Arctic climate responses to extra-polar SST forcing.
 Similar to previous Arctic energy budget studies [Porter et al., 2010; Serreze et al., 2007], we include an analysis of areal averages of the region north of 70°N latitude (Figure 1), termed the polar cap. The polar cap domain, which has a total surface area of 15 × 106 km2, consists of 72% ocean and 28% land and covers 6% of the Northern Hemisphere. By comparison, the model domain used for these simulations, with an area of 162 × 106 km2, is 10.8 times larger than the polar cap domain. The use of such a large domain provides the model with more freedom to respond to the imposed LBC changes in these experiments.
 For many of the comparisons presented below we take differences in monthly mean model fields. In addition to monthly averages, we also calculate 30-day moving averages of daily mean quantities to better understand the temporal evolutions of different responses. The running means are calculated using the ensemble means of daily averages and are centered between 1 July and 15 November. Polar cap averages for the moving average data are calculated similarly to the monthly means.
4.1. Prescribed Sea Ice Concentration and Sea Surface Temperatures
Figure 2a shows monthly mean sea ice concentrations differences used in the high1996, low2007, and mixedSST experiments. Note that the monthly mean ice concentrations for mixedSST and low2007 scenarios are identical. The largest negative differences are for August, September, and October, in the Beaufort, Chukchi, East Siberian and Kara seas. Positive differences (more ice in 2007 than in 1996) are found in the Fram Strait region and around the Greenland Sea, mostly associated with increased ice export out of the Arctic Ocean [Kwok et al., 2009]. By November, the largest differences are in the Chukchi, northern Barents, and Labrador seas.
 Specified daily SSTs from ERA-Interim, also repeated for each simulation from either 1996 or 2007, match the specified sea ice. By repeating SSTs in the same manner as for sea ice concentration, this method reduces sharp, unphysical temperature gradients near the ice edge. As expected, the region with the largest negative ice concentration anomaly in 2007 shows large positive SST differences (Figure 2b). In the North Atlantic the differences are more variable, but lower SSTs do occur in the region of increased ice cover. Importantly, in the low2007 and high1996 experiments, large differences in SST outside of the Arctic basin are also present, resulting in a potential response in WRF due to lower latitude SST differences that are not associated with changed polar sea ice cover and SSTs. Figures 2c and 2d show the difference in SSTs between the low2007 and high1996 scenarios and the mixedSST ensemble.
4.2. Direct Effects: Sensible and Latent Heat Fluxes
 A major pathway for the imposed changes in lower boundary conditions to impact the atmosphere is through surface turbulent latent and sensible heat fluxes. Accordingly, areas of increased SST are associated with a positive response (Figure 3a) in the turbulent sensible heat flux (QH). Because open water and surface temperatures govern potential evaporation, and therefore the vertical gradient in specific humidity, there is also an increase in the turbulent latent heat flux (QE, Figure 4a). The sign convention used here is that a flux is positive when directed into the atmosphere: upwards at the surface and downward at the top of the atmosphere (TOA). As the high latitudes come out of their summer insolation maximum, both QE and QH begin to turn from positive to negative over sea ice (not shown).
 Significant (marked by hatching in the spatial plots) Arctic monthly mean turbulent heat flux responses are as large as 26 W m−2 in September, the month with largest sea ice anomalies (Figures 3a and 4a). Note that by comparing each pair of ensembles, it can be determined that most of the high latitude responses are due to within-Arctic surface forcing (e.g.,Figures 3a and 3c) and not extra-polar SST changes (e.g.,Figure 3b). At this time there are also significant decreases in both QH and QEin areas of increased sea ice cover (e.g., Fram Strait). Reflecting strong air-sea temperature differences, the largest contrasts in QE and QH, both in spatial extent and magnitude, occur in October (Figures 3, 4, and 5). This is especially true for those areas where the atmosphere is usually largely decoupled from the ocean due to the presence of sea ice cover, but was ice free in the autumn of 2007. The monthly mean difference in QH for October is as large as 40 W m−2 in the eastern Arctic Ocean and the Beaufort Sea (in both Figures 3a and 3c). The response in QE is smaller, but still up to 24 W m−2 in the Chukchi and East Siberian Seas (in both Figures 4a and 4c).
 When comparing the QE response in the low2007 and high1996 runs to those using the mixedSST simulation (Figure 4b and 4c), the most obvious feature is that the large, extra-polar responses in the North Pacific and North Atlantic oceans are linked to SST changes in these locations. Also,Figures 3c and 4c highlight that although there are no midlatitude SST differences when comparing the mixedSST and high1996 ensembles (Figure 2d), there are significant extra-polar responses in the latent heat flux for all months. Such large monthly mean responses, over 30 W m−2 at 33°N in the Pacific in September (Figure 4c), reinforce how within-Arctic sea ice and SST changes can have significant extra-polar effects.
Figures 5a and 5bshow 30-day moving averages of polar cap averaged QH and QE, respectively. Vertical bars indicate the one standard deviation spread among the ensemble members. Differences in QE lag those in QH. Responses in both QH and QE become substantial in September, the month with largest difference in sea ice concentration. However, the largest flux response is in October and November, with polar cap differences of 6 and 4 W m−2 for QH and QE, respectively.
 The increase in sensible and latent heat fluxes for 2007 ice conditions relative to 1996 is reflected in the 2 m temperature (T) and specific humidity (q). Polar cap 30-day moving averages ofT and q at 2 m (Figures 5c and 5d) show a delayed autumnal decline due to the increased upward sensible and latent heat fluxes, respectively. A significant warming and moistening response at 2 m occurs in September, with a maximum response in October. Specifically, the low2007 responses in early autumn are as large as +6°C and 0.8 g kg−1(56% increase) compared to the high1996 scenario. There remains a significant response at the end of the simulation, almost 3 months after largest lower boundary condition forcing. A significant increase of over 2°C in the low2007 simulations remains for the 30-day polar cap moving average that is centered on 15 November.
 In the mixedSST ensemble (green line in figures), the responses of the polar cap surface averages shown in Figure 5are very similar to those in the low2007 scenario. Hence, most of this model response is due to within-Arctic changes. However, they are not identical, meaning that lower-latitude SST changes are having some effect on the Arctic atmosphere, although this is of secondary importance.
4.3. Vertical Extent of Warming and Moistening
Figure 6shows ensemble mean polar cap 30-day moving average vertical profiles of temperature for high1996 (Figure 6a), low2007 (Figure 6b), and difference plots based on the three differences between the low2007, high1996, and mixedSST ensembles. As seen in the low2007 minus high1996 difference plot (Figure 6c), despite only modest increases in QH in August, there is a substantial response in the temperature below 3 km. By September, the lowest kilometer shows a strong positive response, resulting in weaker inversion (Figure 6b). The largest response in both QH and T at 2 m occurs in October, so it is not surprising that this month also shows the strongest tropospheric warming (4°C in the lowest kilometer) (Figure 6c). The response decreases in November. However, the difference in upper level temperatures is largest at this time (over 1.4°C warming above 9 km). The monthly progression of the atmospheric temperature response shows that the warming is first confined to the lower levels, but becomes vertically deep in the autumn months.
 Comparing the difference plots Figures 6d and 6e illustrate which responses seen in Figure 6care due to within- and extra-polar LBC changes. Given the similar responses inT and q at 2 m for the low2007 and mixedSST simulations (Figures 5c and 5d), it is not surprising that the largest response is due to within-Arctic sea ice and SST changes (mixedSST minus high1996,Figure 6e), while the model is generally less responsive to changes in lower-latitude SSTs (low2007 minus mixed SST,Figure 6d) than it is to within-Arctic sea ice and SST changes (mixedSST minus high1996,Figure 6e). The exception to this is the warming from 2 to 9 km in late October in the total response (low2007 minus high 1996, Figure 6c) which is due to the extra-polar SST changes (low2007 minus mixed SST,Figure 6d), while the strong low-level warming of over 5°C up to 2 km is a result of the reduced sea ice (mixedSST minus high1996,Figure 6e). The elevated warming seen in the low2007 minus mixedSST response, especially when there is little surface response (Figure 6d), is likely a result of changes in atmospheric circulation patterns driven by extra-polar SST differences. However, the elevated tropospheric warming in the total response (low2007 minus high1996,Figure 6c) in early October and the stratospheric warming from mid-October to mid-November appears to be due primarily to the surface warming (low2007 minus mixedSST,Figure 6e) propagating vertically because of reduced static stability in the low2007 simulation. This extra heat in upper atmospheric levels at the end of the simulation suggests that the surface forcing maximum in September could have effects lasting into winter, potentially influencing the large-scale wintertime circumpolar vortex [Overland and Wang, 2010].
 Turning to specific humidity (Figure 7), the low2007 minus high1996 difference plot (Figure 7c) reveals modest moistening of the lowest 1 km in July and August that is consistent with the modest enhancement of QE. The humidity response becomes stronger in magnitude in September and is maximized in October with moistening of the lowest 4 km. At this time, the peak moisture response is found at the surface. Remarkably, there is a 33% increase in specific humidity below 1 km for low2007 compared to the high1996 scenario. By isolating the effects of lower-latitude SST changes, it is clear that most of the total response is due to within-Arctic forcing (mixedSST minus high1996,Figure 7e).
4.4. Response of Cloud Cover
 Comparison of the low2007 simulation with the high1996 simulation indicates that the warmer and moister Arctic atmosphere in autumn (Figures 6c and 7c), in response to a decrease in sea ice cover, generally also results in an increase in cloud cover (low2007 minus high1996, Figure 8c). The 30-day moving average liquid cloud content (qcloud) for the polar cap indicates decreased low-level cloud cover (below 1 km) and increased in mid-level cloud cover (from roughly 1 to 3 km) through the middle of September (Figure 8c). The increase in lower-tropospheric temperature outpaces the increase in moisture, resulting in a reduced relative humidity at the surface and an increase in boundary layer height. The pattern of decreased low-level cloud paired with increased mid-level cloud in the low2007 minus high1996 comparison (Figure 8c) weakens through September. From October through the end of the simulation, the dominant signal is more low-level cloud. Indeed, after about 25 September, there is a 31% increase in the low-level liquid cloud content in the low2007 ensemble compared to the high1996 ensemble. The response in low-level clouds in early autumn is the result of large increases in bothT and qat a time when overall Arctic cloud cover is decreasing. This reflects a delayed transition from the summer cloud regime, characterized by low-level liquid stratus clouds, to the elevated maximum in WRF-simulated liquid clouds in the autumn. A delayed seasonal transition acts as a memory mechanism by prolonging the effects of sea ice loss through changes in the surface energy budget. For the polar cap domain specifically, the availability of late autumn moisture is limited because of the growing sea ice cover. In the low2007 scenario, this moisture limitation in October and November is diminished, resulting in an increase in the lowest clouds.
Figures 8d and 8eindicate that the total response in polar cap liquid clouds is mostly due to within-Arctic forcing. Interestingly, the cloud response to lower-latitude SST changes in late summer is an opposite pattern to the total response. These changes in summertime cloud cover are associated with altered circulation patterns induced by the extra-polar SST changes.
 Ice clouds, as simulated by WRF, are generally found between 5 and 9 km in summer, with maximum ice content around 6 km in September. Ice clouds in the autumn and early winter occur between the surface and 7 km, with the height of the ice content maximum decreasing to 4 km by mid-November (Figures 9a and 9b). The change in ice content due to changes in lower boundary conditions is shown in Figures 9c to 9e. The higher temperatures in the lowest 2 km of the atmosphere in late October and early November, due to reduced Arctic ice cover (Figures 6c and 6e), leads to a reduction in low ice clouds at this time (Figures 9c and 9e). This manifests a delayed transition to the winter pattern, when ice clouds can be found at low atmospheric levels. The modeled increase in ice cloud content above 6 km is likely due to the increase in moisture at these levels, where it is still too cold to form liquid cloud drops. It is evident from a comparison of Figures 9d and 9e with the total change (low2007 minus high1996, Figure 9c) that a large portion of the total response of ice clouds, particularly between 3 and 9 km, results from changes in extra-polar SST, likely linked to changes in circulation. The low-level decrease in ice clouds in autumn, however, is a reflection of the increased temperatures due to sea ice loss and Arctic SST increases (Figures 9c and 9e).
4.5. Response of Radiation Fluxes
 The response in surface downwelling radiation is similar in both the low2007 and mixedSST simulations (Figure 10). In July and August, there is little difference in the downwelling longwave radiation between the low2007, high1996, and mixedSST ensembles (Figure 10b) because the decrease in low-level liquid clouds in the low2007 and mixedSST ensembles relative to the high1996 ensemble is mostly counteracted by the increase in mid-level liquid clouds (Figure 8c). The increase in ice cloud amount in July (all responses, Figure 9) can probably account for the observed decrease in downwelling shortwave radiation at the surface in the low2007 and mixedSST ensembles (Figure 10a). The response in the net surface shortwave radiation is strongly negative in the low2007 and mixedSST experiments (Figure 10c) since the reduced surface albedo of the polar cap increases the amount of absorbed solar radiation. Starting in September, the increase in total cloud cover (mainly low-level features at this time), results in a substantial increase in downwelling longwave radiation for these two experiments relative to the high1996 experiment (Figure 10b). There is little downwelling longwave radiation response to sea ice changes in summer, but cloud increases over newly opened water in autumn explain the increased downwelling longwave radiation in the low2007 and mixedSST experiments at this time of year. Importantly, the positive feedback of the increased downwelling longwave radiation, implied by but prohibited by our experimental setup, suggests that the real response in autumn could be larger.
 When considering the response in longwave radiation at the surface, we focus on the difference between the low2007 minus high1996 results since the Arctic SSTs are identical in the mixedSST and low2007 simulations. In addition to increased turbulent heat transfer from the ocean to atmosphere (Figures 5a and 5b), the open water also leads to an increase in the upwelling longwave radiation relative to ice-covered grid points (Figure 11a). The largest longwave response occurs in September and October in areas where there are large differences in sea ice concentration between the low2007 and high1996 ensembles. For areas with a positive response in upwelling longwave radiation, these are partly balanced by increases in downwelling longwave (Figure 11b) resulting in net increased longwave flux to the atmosphere (Figure 11c). Although the net surface longwave response is positive (Figure 10d), the response is much smaller than the response in the other surface heat flux components in early autumn. While warmer than the high1996 ensemble mean, the polar cap averaged net surface longwave responses for both low2007 and mixedSST comparisons are not statistically significant (Figure 10d). Interestingly, there are local significant responses in the net longwave term in areas without changes in SSTs or sea ice cover. A negative response in net surface longwave radiation, a heat gain by the surface, is found over high latitude land areas, as well as the North Pacific and locations with increased sea ice cover (Greenland and Barents seas) (Figure 11c).
 Both the net surface heat flux (FSFC, positive upwards) and the net TOA radiation (FRAD) are sensitive to changes in sea ice cover. Figures 10e and 10f show the polar cap moving average FRAD and FSFC in WRF for the high1996, low2007, and mixedSST experiments. As for other variables discussed above, the low2007 and mixedSST FRAD and FSFC are nearly identical. At the TOA in July and August (Figure 10e), the decrease in the reflected shortwave flux more than compensates for the increased loss of longwave radiation, leading to reduced net radiative loss by the atmosphere until mid-September. In autumn, as incoming shortwave radiation approaches zero, the increased longwave emission to space results in a decrease in FRAD (Figure 10e). The surface shows a similar response (Figure 10f). In summer, the increase in absorbed shortwave radiation due to sea ice loss overwhelms other responses, resulting in a negative (downward) response in FSFC. By October, the increased upward turbulent heat fluxes accompanies an increase in the net longwave flux to the atmosphere (Figure 10f). The similarity of the low2007 and mixedSST results (Figures 5 and 10) show that most of the total response in radiation and large-scale energy budget terms is a consequence of the changes in Arctic lower boundary condition, with lower-latitude changes playing a much smaller role. Under the energy budget framework, sea ice loss acts to amplify the seasonal cycle of both FRAD and FSFC, which will in turn impact the horizontal energy flux divergence and atmospheric energy storage tendencies [Serreze et al., 2007].
4.6. Hydrologic Cycle Response
 In addition to increased cloud height in summer and altered cloud amount in autumn, the WRF experiments point to impacts of decreased sea ice cover on the hydrologic cycle. Figure 12shows 30-day moving averages of evaporation (E), precipitation (P), and precipitation minus evaporation (P-E). In autumn (September, October, and November), when the response in the turbulent fluxes is largest, there is very little difference in P-E between the low2007 and high1996 ensembles. This is due to compensation by similar magnitude increases in both E and P. During this time, most of the additional water vapor in the Arctic, due to the reduced sea ice cover, is recycled to the surface as precipitation. The mixedSST ensemble mean P is smaller than in the low2007 scenario (Figure 12, green line), meaning that extra-polar SST changes have impacts on the within-Arctic hydrologic cycle. This suggests that the polar cap would become drier (reducedP-E) in October for the low2007 minus high1996 comparison if not for the contribution from lower-latitude SST changes. While there are only small changes inP-E in the low2007 scenario, compensating responses in P and E can have impacts on the atmospheric energy budget by taking energy from the surface and depositing it in the atmosphere.
Figure 13 shows the spatial pattern of the changes in E, P, and P-E between the low2007 and high1996 ensembles. For areas with open water in 2007, there is a large increase in both evaporation and precipitation. In these areas, the increase in atmospheric moisture is not entirely balanced by precipitation, leading to a negative response in P-E. This excess moisture is carried away from these regions of increased E, a divergence of latent energy, to other areas where it then precipitates. The opposite is true for areas with more sea ice (such as in the Greenland Sea) in the low2007 scenario.
 An interesting result from Figure 13 is a large hydrologic response in areas without a difference in sea ice or SSTs between the low2007 and high1996 scenarios. This includes areas like the Gulf of Alaska and many land areas, such as East and West Siberia, Alaska, and Greenland (Figures 2a and 2b). These areas show only a small response in evaporation while at the same time a large response in precipitation. These responses are therefore related to changes in moisture gradient or atmospheric circulation, which is supported by a significant response in the latent energy flux divergence (Figure 13d). Areas that experience a wetter surface (positive P-E difference) as a result of increased latent energy convergence include the north slope of Alaska and eastern Siberia as well as the southwest coast of Greenland. The southeast coast of Greenland, western Siberia, and other land areas, are influenced by circulation changes that do not favor latent energy convergence, leading to drier land surfaces. In addition, changes in atmospheric circulation results in some places having large changes in evaporation but no change in precipitation, such as the central Barents Sea and Baffin Bay.
 In July and August, when there is little response in evaporation, there is an increase in P (July) and then a decrease (August) in P in the low2007 ensemble relative to the high1996 ensemble, respectively (Figure 12). These changes are likely associated with circulation differences and associated patterns of moisture convergence. Looking at the patterns of the monthly mean SLP response (an indicator for circulation changes), the periods of higher and lower P in July and August, respectively, correspond to negative and positive responses in Arctic mean sea level pressure (Figure 14c). This is consistent within the view that negative monthly mean pressure anomalies manifest stronger or more frequent storms entering the Arctic. While the positive high pressure response in the central Arctic in August appears to be attributable to within-Arctic changes in lower boundary conditions, (Figures 14c and 14e), such a large response in SLP (Figure 14c) in July is surprising given the relatively small sea ice anomalies at this time. The results in Figure 14d suggest that the summertime differences in midlatitude SSTs between high1996 and low2007 have significant impacts on the early part of the simulations, although these midlatitude SST differences are not necessarily implicated in the sea ice differences between those years [e.g., Graversen et al., 2011]. In autumn, SLP differences are dominated by the within-Arctic lower boundary condition forcing. The lower Arctic SLP in October and November (Figures 14c and 14e) is consistent with the increased precipitation at this time in both comparisons. This suggests that the increased precipitation is not only a function of increased moisture (Figure 7) but also increased storminess. This helps explain why the mixedSST ensemble does not show as large an increase in autumn precipitation as the low2007 ensemble (Figure 12).
 Several of the results presented in this paper are qualitatively similar to other recent model sensitivity experiments [Blüthgen et al., 2012; Kay et al., 2011a; Strey et al., 2010], which increases our confidence in the modeled atmospheric response to the sea ice and SST anomalies observed in the summer and autumn of 2007. In addition, our results also identify additional atmospheric responses not seen in these previous studies.
 Along with Blüthgen et al. , our maximum response in near surface temperature and surface heat fluxes occurs in newly ice-free areas. We both find increases of near surface air temperature of up to +3 K, increases in ocean heat uptake in summer (although we see a larger response of about 60 W m−2 compared to their value of 40 W m−2), and an increase in heat flux from the ocean to the atmosphere in autumn (again, our modeled response is about 110 W m−2 compared to their value of 60 W m−2). Estimates of the response in the net surface heat flux from the 10-yr freely evolving sensitivity experiments inKay et al. [2011a] are smaller in July, at only 6.2 W m−2.
 The duration of atmospheric responses to sea ice seen in this paper, lasting through mid-November, are similar to those ofStrey et al. , who find a response time of polar cap averaged 2 m temperature of about 2 months. Interestingly, although we have a more similar experimental setup to Blüthgen et al. , they find that the atmospheric response is strongest in July, August, and September, a result that is less consistent with observational work that shows a significant autumn response in surface heat fluxes and air temperature to reduced sea ice [Francis et al., 2009; Screen and Simmonds, 2010; Serreze et al., 2009].
 The response in cloud cover and the atmospheric boundary layer to changes in surface conditions is seasonally dependent. In summer, Kay et al. [2011a]show that melting sea ice and a small air-sea temperature gradient result in a rather weak response to reduced sea ice and increased SSTs. In contrast, along withBlüthgen et al. , we find an increase in cloud height and thicker boundary layer in summer, along with significant increases in upper-level ice clouds. The 30-day polar cap moving average liquid cloud water content for the polar cap (qcloud) indicates decreased low-level cloud cover (below 1 km) and increased mid-level cloud cover (from roughly 1 to 3 km) through the middle of September (Figure 8c). This counter-intuitive response has also been reported bySchweiger et al. , using data from the ECMWF 40 year reanalysis (ERA40) and infrared satellite observations, and is argued to result from decreased static stability. The response in our experiments, while similar, is not exactly comparable since we use polar cap averages while Schweiger et al.  analyze only areas with significant sea ice changes.
 In autumn, when the air-sea temperature gradient is large and atmospheric stability is reduced compared to other times of the year, increased turbulent heat fluxes to the atmosphere can lead to a pronounced cloud cover response [Kay and Gettelman, 2009], particularly at low levels over newly ice-free areas. The autumn cloud response in our study is characterized by an increase in liquid clouds and a decrease in ice cloud content, which in turn has large effects on the amplitude and seasonality of the net surface and TOA radiation budgets, as was also shown byKay et al. [2011a]. Comparison of the low2007 simulation with the high1996 simulation indicates that the warmer and moister Arctic atmosphere in autumn (Figures 6c and 7c), in response to a decrease in sea ice cover, generally also results in an increase in cloud cover (low2007 minus high1996, Figure 8c). Vavrus et al.  also see increased cloud cover attending reduced sea ice concentrations in autumn in global climate models, though they are unable to determine that changes in either variable is leading changes in the other when using monthly mean data. This response in autumn cloud cover to sea ice loss is consistent with recent observations [Kay and Gettelman, 2009; Palm et al., 2010]. These different responses in summer and autumn point to an intensification of the seasonal transition, shown also by Blüthgen et al. .
 The higher temperatures in the lowest 2 km of the atmosphere in late October and early November (Figures 6c and 6e), lead to a reduction in low ice clouds at this time (Figures 9c and 9e). This manifests a delayed transition to the winter pattern, when ice clouds can be found at low atmospheric levels. This result is consistent with de Boer et al. , who used lidar and radar data to show that that liquid cloud fraction of Arctic clouds decreases with decreasing temperature. Similar to Vavrus et al. , it is evident from a comparison of Figures 9d and 9e (low2007 minus high1996, Figure 9c) that a large portion of the total response of ice clouds, particularly between 3 and 9 km, results from changes in extra-polar SSTs, which are likely linked to changes in circulation.
 In agreement with findings by Kay and Gettelman , there is a small longwave radiation response to sea ice changes in summer due to clouds, but cloud increases over newly opened water in autumn helps to trap heat by increasing the downwelling longwave radiation. This effect of sea ice cover on downwelling longwave radiation in October and November counters the conclusion of Schweiger et al. , that sea ice has no significant net effect in autumn months. Vavrus et al.  find that cloud cover changes in 21st century projections in CCSM3 can accelerate rapid ice loss events in autumn, resulting in increased cloud radiative forcing through increases in downwelling longwave radiation and liquid cloud condensate. Serreze et al.  find that the observed strong warming of the lower troposphere over the Arctic Ocean, at least in atmospheric reanalyses, is maintained through increased upwelling longwave radiation associated with reduced sea ice cover. For this study, although the net surface longwave response is positive (a heat gain by the atmosphere), it is much smaller than the response in the other surface heat flux components in early autumn, a feature also found by Deser et al.  in CAM3 simulations.
 The intensified differential atmospheric heating due to reduced sea ice can result in changes in baroclinicity, leading to changes in atmospheric circulation [Jaiser et al., 2012]. Blüthgen et al.  find a consistent low SLP anomaly of −2 hPa over the whole eastern Arctic between July and September. Francis et al. find increased 500 hPa heights in winter in response to low September sea ice extent and warming surface temperatures. We find larger, though less consistent, responses in SLP and 500 hPa height fields to within-Arctic surface forcing. As pointed out byOverland and Wang , SLP fields contain more information than just the first-order thermodynamic response to increasing low-level temperatures, clouding the interpretation of these response patterns.
 Several results from this paper are either new or more thoroughly developed than previous efforts, including the following.
 1. By analyzing the temporal evolution of vertical profiles of temperature, humidity, and cloud cover in response to changes in sea ice and SSTs, we can uniquely diagnose both the vertical structure and seasonality of the atmospheric response. The weak temperature response to within-Arctic surface forcing is confined near the surface until August when circulation changes result in a 1 K warming up to 9 km, with a peak warming response at about 2 km (Figure 6e). Starting in September and lasting through 15 November (the end of our analysis), the warming increased to over 2 K and is generally limited to below 2 km when forced with sea ice loss and Arctic-only SST increases. The liquid cloud response is characterized by an increase in cloud height in summer and an increase in low-level liquid clouds in autumn due to a delayed seasonal transition to ice clouds because of increased surface temperatures. The response in ice clouds above 3 km exhibits large month-to-month variability and appears to be more correlated with changes in circulation than temperature.
 2. The hydrologic cycle response to reduced sea ice can be large because of the resulting increase in evaporation, resulting in a moister Arctic atmosphere. Most of this positive response in evaporation is counteracted by an equally large response in precipitation when considering averages over the polar cap. We find significant responses over high-latitude land areas that are downwind of newly ice-free water, including the north slope of Alaska, the eastern Siberian coast, the Canadian archipelago, and western Greenland. Other land areas show a significant drying (negativeP-E response) due to sea ice loss, including southeast Greenland. Although not directly comparable, this agrees with Schuenemann and Cassano  who on the basis of CMIP3 model output showed that when changes in circulation patterns are isolated in a warmer and reduced sea ice climate, precipitation decreases in southeast Greenland. The response of precipitation in the North Atlantic in this paper is opposite that of Strey et al.  for autumn months, who find an increase in precipitation. That our results are different is not surprising given that Strey et al. did not alter the SST state, which will have large impacts on evaporation and therefore the precipitation response.
 3. Through our unique use of three ensembles, each of which is forced by observed atmospheres from 15 different years, we are able to determine how much of the total response to changed surface conditions is due to within-Arctic and extra-polar forcing.Blüthgen et al. find that the main response patterns are retained whether or not imposed SST changes are confined to the Arctic or not. However, we find that while most of the total atmospheric response to changed lower boundary conditions (low2007 minus high1996) results from within-Arctic forcing, extra-polar SST changes are shown to have some significant impacts on the Arctic atmosphere. We feel that part of this discrepancy is because we are forcing each ensemble member with observed atmospheres and, therefore, capturing true modes of natural variability (e.g., ENSO) not represented in the experimental design ofBlüthgen et al. . Also, we have shown exactly which parts of the overall responses are due to lower-latitude SST changes. Specifically, the effects of SST changes below 66°N include deep atmospheric cooling in August and warming in October and an increase in surface liquid clouds in summer over the polar cap and low SLP anomalies in the eastern Arctic. That lower-latitude SSTs alone can have significant impacts on the Arctic atmosphere in just a few months of mesoscale model simulation highlights the general sensitivity of the Arctic climate. Also, within-Arctic sea ice and SST changes can have significant extra-polar effects, such as large monthly mean turbulent heat flux responses of over 30 W m−2 at 33°N in the Pacific in September (Figure 4c).
6. Summary and Conclusions
 As the Arctic continues to warm, we can expect continued reduction in its sea ice cover with attendant increases in sea surface temperatures. Consistent with previous modeling studies, our WRF sensitivity experiments suggest that changes in surface fluxes due to reduced sea ice extent can significantly affect the overlying atmosphere with effects extending beyond the Arctic. The results from this study, along with comparisons to recent studies also looking at impacts of the 2007 sea ice minimum, support the following conclusions.
 1. Impacts of anomalous open water and positive SST anomalies are most pronounced in autumn when air-sea temperature differences are strongest, promoting increased turbulent sensible and latent heat fluxes. Warming and moistening of the Arctic boundary layer is maximized in early autumn and this signal subsequently spreads vertically in November, where there is a temperature response to reduced sea ice of 1.4°C above 9 km.
 2. Similar to several recent studies [Jaiser et al., 2012; Kay et al., 2011a; Schweiger et al., 2008], we also find a summertime increase in low-level temperatures and humidity acting to reduce static stability, raising the top of the boundary layer, and consequently the cloud height. The response of downwelling radiation to reduced ice cover is small in summer, but there is an increase in upwelling longwave radiation due to increased surface temperatures.
 3. In autumn, when the increase in turbulent heat fluxes is largest, so too is the response in low-level clouds and radiation. Enhanced liquid cloud content persists into autumn in the reduced sea ice scenario, along with a simultaneous decrease in ice cloud content, consistent with the increase in temperatures at low heights. A consequence of the response in cloud cover and surface temperature is increased amplitude of the seasonal cycle of both the net surface heat flux and TOA radiation budgets, which will have impacts on the energy convergence from lower latitudes.
 4. The hydrologic response to reduced sea ice cover is characterized by an autumnal increase in both evaporation and precipitation across the polar cap domain. Even as polar cap averaged P and E both increase resulting in little change in P-E, some areas experience large changes in P-E. Additionally, changes in precipitation or evaporation mostly drive changes in latent energy flux divergence, not the other way around. We identify responses in the hydrologic cycle that can in turn affect the Arctic energy budget by taking heat from the subsurface column and depositing it in the atmosphere. The increase in precipitation, which is mostly snowfall during these autumn months, will impact the surface albedo and thermal conductivity over sea ice.
 5. While most of the response seen in our simulations is due to changes in Arctic surface conditions, changes in lower latitude sea surface temperatures can also significantly impact the Arctic through altering atmospheric circulation patterns.
 6. Finally, we also find significant midlatitude responses in the surface heat fluxes and air temperature when changing only within-Arctic surface conditions.
 The authors would like to thank the three anonymous reviewers and the editor for their significant help in improving this manuscript. This work was supported in part by a grant of HPC resources from the Arctic Region Supercomputing Center at the University of Alaska Fairbanks as part of the Department of Defense High Performance Computing Modernization Program. ERA-Interim forcing data was kindly supplied by ECMWF. Support for this research comes from the National Science Foundation grants NSF ARC 0805821, ARC 0901962, ARC 0732986 and Department of Energy grant DOE DE-FG02-07ER64462.