a. Data

Evaluation of ARCSyM performance for 1990 and interpretation of the unusual climatic conditions during April–September requires a variety of datasets with relatively high spatial and temporal resolution. Sea ice data used include Special Sensor Microwave Imager (SSM/I)-derived daily and monthly mean sea ice fractions (area covered by sea ice per grid cell) obtained from the National Snow and Ice Data Center (Weaver et al. 1987), daily and monthly mean ice velocities derived from SSM/I 85-GHz data (Agnew et al. 1997; Emery et al. 1997; Kwok et al. 1998), trajectories of drifting buoys provided by the International Arctic Buoy Program (IABP; Thorndike and Colony 1981; Rigor and Colony 1994), and sea ice charts prepared by the National Ice Center. The SSM/I-derived ice fractions calculated using the National Aeronautics and Space Administration team algorithm (Cavalieri and Gloersen 1984) are typically accurate to within 5% during cold conditions. During peak melt in summer and when meltponds are present, the SSM/I algorithm underestimates ice fraction by approximately 10%–20% (Steffen et al. 1992). The velocity error in the daily SSM/I-derived ice motion is about 6 cm s⁻¹ due to the relatively low spatial resolution of the sensor. In most areas, however, the estimates are unbiased, so that errors are reduced when velocities are averaged over time or space. Surface melt and atmospheric water content limit the retrieval of the SSM/I-derived ice velocities, so that the remotely sensed ice motions are typically available only for autumn through spring.

Air temperature data were obtained from the Earth Observing System Polar Exchange at the Sea Surface (POLES) project (Martin and Munoz 1997). The POLES temperatures used here are interpolated from station data and drifting buoys for 1979–93. Approximate errors in the data are 0.1°C for coastal stations and 2.0°C for the drifting buoys. The resulting gridded fields of 2-m air temperatures show a positive bias of 0.3°–1.3°C compared to data from Soviet drifting stations (Overland et al. 1997). Shortwave and longwave radiative fluxes were obtained from the Arctic Ocean Radiative Fluxes (AORF) POLES dataset (Schweiger and Key 1994). AORF downwelling fluxes were estimated using the International Satellite Cloud Climatology Project (ISCCP)-C2 cloud data (Rossow and Schiffer 1991) and a radiative transfer model. Schweiger and Key (1994) and Key et al. (1997) describe the likely sources of errors in the ISCCP-C2 cloud amounts and derived fluxes, including the tendency for cloud amount to be underestimated over sea ice. Comparison of fluxes calculated using similar methods with surface measurements at Barrow, Alaska, and at the South Pole indicates errors of 10–25 W m⁻² in downwelling shortwave radiation (Rossow and Zhang 1995). As discussed later, these errors are similar in magnitude to the differences between the various ARCSyM simulations examined. However, the regional and seasonal patterns in the AORF fluxes are reasonable and are used here to assess the ability of the coupled model to reproduce these patterns. Other radiation data for 1990 were taken from the summary of Soviet drifting station observations compiled by Marshunova and Mishin (1994). Mean cloud fractions used for comparisons were obtained from ISCCP-D2 data (Rossow et al. 1996) as compiled by Serreze et al. (1997a). These cloud data include improvements in calibration and cloud detection algorithms, relative to the previous ISCCP-C2 version.

SLP data used include daily and monthly mean European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses obtained from the National Center for Atmospheric Research (NCAR; Trenberth 1992), National Center for Environmental Prediction (NCEP)–NCAR reanalysis fields (Kalnay et al. 1996), and monthly mean SLP interpolated from IABP buoys (Rigor and Colony 1994). Other datasets examined include satellite-derived ocean surface temperatures from the Advanced Very High Resolution Radiometer (AVHRR) Global Area Coverage Ocean Pathfinder Project [as presented in Parkinson (1997)], the cloud climatologies of Hahn et al. (1988) and Warren et al. (1986, 1988), and new radiation, cloud, and temperature climatologies summarized for the Arctic by Serreze et al. (1997a).

b. ARCSyM

ARCSyM is a limited-area mesoscale climate model that incorporates representations of the atmosphere, land surface, sea ice, and ocean. The atmospheric model is based upon the NCAR RegCM version 2 (Giorgi et al. 1993a,b), a hydrostatic, primitive equation model with a terrain-following vertical (sigma) coordinate. Coupled to this atmospheric model is the NCAR land surface model (Bonan 1996), and a dynamic–thermodynamic sea ice model. The sea ice model within ARCSyM consists of the cavitating fluid dynamics scheme of Flato and Hibler (1992) and the thermodynamics of Parkinson and Washington (1979) with surface flux modification following Ebert and Curry (1993). The cavitating fluid assumption is a parameterization of the internal ice stress, which limits the amount of compaction due to convergence. This simplification provided some advantages for climate model use (Flato and Hibler 1992), but tends to overestimate ice velocities in areas of extreme convergence, such as in the Arctic Ocean near the Canadian Archipelago. A Mohr–Coulomb shear stress modification is used here to provide resistance to shear, and thus better reflect conditions in such locations. The Mohr–Coulomb approximation provides substantial improvement in areas such as the central Arctic Ocean, but motion is still overestimated in highly convergent areas, with a tendency to steer the flow at an angle to locations with thick or compact ice (Flato and Hibler 1992).

The oceanic component of ARCSyM comprises the one-dimensional mixed-layer ocean model of Kantha and Clayson (1994) based on a second-order turbulence closure boundary layer scheme, which explicitly simulates the whole vertical water column. Ocean bathymetry was obtained from the TerrainBase datasets compiled by the National Geophysical Data Center. The surface heat balance over ice or snow is computed as a net sum of the surface sensible and latent heat fluxes, the net downward longwave radiation, and the absorbed solar radiation. The mixed-layer model takes as input the net heat flux as above with the addition of a precipitation heat flux and the net salinity flux due to precipitation/evaporation and ice melt/growth. A conductive flux at the base of the ice is also computed along with the prescribed oceanic heat flux due to entrainment from below and lateral advection, and the latent heat due to ice growth or melt. These balances are computed in the ice model at each time step to find new surface, snow–ice interface, and ice bottom temperatures, which produce new fluxes to be passed back to the atmospheric model.

The ARCSyM model has been previously validated in several different contexts. At smaller scales, the model successfully simulated a wind-driven coastal polynya south of St. Lawrence Island in the Bering Sea (Lynch et al. 1997). The observed formation and extent of the polynya compared well to the model simulations and remotely sensed data provided by the SSM/I and AVHRR. Ice displacement over a 3-day period corresponded well to synthetic aperture radar derived vectors, and average ice drift speed was 0.35 m s⁻¹ compared to the observed 0.33 m s⁻¹. In addition, high-resolution simulations of the Alaskan North Slope (Lynch et al. 1999) show a strong correspondence between station data and land surface behavior, particularly with regard to circulation, surface temperatures, and turbulent fluxes of heat and moisture. At larger scales, the ARCSyM model has also been validated in simulations over the western Arctic (Lynch et al. 1995; Bailey et al. 1997). Lynch et al. (1995) found that the inclusion of both sea ice dynamics and the ice phase in large-scale cloud formation improved the simulation of surface temperature and precipitation, in particular. Bailey et al. (1997) found that an ocean mixed-layer model improved the overall ice distribution compared to the ocean “swamp” treatment (a constant heat flux applied at 30-m depth).

c. Stand-alone ice model

A two-dimensional dynamic–thermodynamic ice model (Maslanik and Dunn 1997) is used to provide initial ice thicknesses for ARCSyM that reflect 1990 climatic conditions. Results from the stand-alone model are also used to help assess the contributions of sea ice dynamical and thermodynamical processes toward generating the 1990 ice extent anomaly. This model includes three ice categories (first-year, second-year, and multiyear ice) adapted from Walsh and Zwally (1990), an approximation of multiple thickness levels used in the energy budget calculations (Walsh et al. 1985), a detailed surface albedo treatment (Ebert and Curry 1993), and a viscous-plastic ice rheology (Hibler 1979) with an improved solution scheme (Zhang and Hibler 1997). The model domain is a Cartesian grid with 80-km cells. Simulations using a 6-h time step were performed with and without climatological ocean currents (Hibler and Bryan 1987). The model includes a slab ocean treatment with prescribed climatological ocean heat fluxes from Hibler and Bryan (1987).

d. Model experiment design

The model experiments are summarized in Table 1. By varying the treatment of sea ice within ARCSyM and the stand-alone model, these experiments address the sensitivity of ice transport and melt to variations in atmospheric conditions, and the response of the atmosphere to changes in ice cover. In particular, the simulations focus on the key features associated with the record reduction in ice extent in the Siberian Arctic. ARCSyM experiments were carried out in fully coupled mode with dynamic and thermodynamic sea ice processes included (designated as “DT”), and with thermodynamics-only ice (e.g., without ice transport) (“TO”). To further isolate the effects of ice fraction and extent on atmospheric conditions, ARCSyM was then constrained using prescribed ice cover (“PR1” and “PR2”). PR1 includes the full dynamic–thermodynamic ice model but with ice fraction prescribed to the generally higher fraction generated by the TO simulation. For the PR2 experiment, ice fraction and ice extent are prescribed using daily SSM/I data. By constraining the ice conditions to match observations, PR2 addresses the degree to which the simulation of atmospheric conditions in ARCSyM is affected by inaccuracies in the simulated ice cover. Results from the stand-alone ice model with dynamic–thermodyamic ice (“SDT”) and thermodynamics only (“STO”) forced with NCEP-derived winds are included to investigate the importance of specific SLP patterns in producing the ice extent anomaly, and to further define the relative importance of ice transport versus ice melt.

For the test period, ARCSyM was initialized and driven at the boundaries using ECMWF reanalysis fields. The ARCSyM experiments described here use the Kuo cloud convection scheme and an implicit grid-scale moisture approach. The vegetation treatment was modified slightly based on measurements of Arctic plant characteristics made in Alaska [see Lynch et al. (1999) for more details]. Initial ice fraction for 15 April 1990 was prescribed from SSM/I data. Initial sea ice thickness was obtained from the stand-alone ice model. The stand-alone model results are taken from simulations performed using daily NCEP reanalysis fields for 1985–93 (Maslanik and Dunn 1997). The NCEP SLP used for the 1985–93 simulations are very similar to ECMWF SLP fields, and both SLP sets compare well with pressure fields estimated from IABP buoy observations. However, the NCEP downwelling shortwave radiation and spring–summer air temperatures over the Arctic Basin are overestimated compared to observations, which yielded excessive ice melt in the stand-alone model simulations when realistic surface albedos were used (Maslanik and Dunn 1997). To compensate for this, downwelling shortwave radiation was scaled by a factor of 0.9 to better agree with observations, and spring–summer air temperatures were adjusted based on climatology. Using these modified NCEP forcings, the resulting ice model simulations yielded realistic ice thicknesses (Maslanik and Dunn 1997) and reproduced the general characteristics of the summer 1990 ice extent, including the record expanse of open water (defined as area with ice fraction less than 15%) in the Siberian sector of the Arctic Ocean (Fig. 1).

a. Overview of ARCSyM results for May–September 1990

Air temperatures simulated by the ARCSyM DT experiment are generally higher than the POLES observations. The errors are greatest for June and July (positive biases of 5.3° and 5.4°C, respectively) and decrease through August and September (positive bias of 3.2°C). The temperatures are overestimated over the entire domain (Fig. 2), but with the largest errors confined to the Canadian Archipelago and the North Atlantic. Positive biases over the central Arctic are substantially smaller, averaging about 3°C. The bias decreases to approximately 1° in the eastern Arctic Ocean, which is within the level of uncertainty of the POLES temperatures. In general, lowest observed mean temperatures occur in the western Arctic, where the greatest positive bias in the ARCSyM temperatures are present. This bias may be linked to the tendency for ARCSyM to underestimate anticyclonic conditions (discussed later). ARCSyM also underestimates the strength of the boundary layer surface inversion.

The overestimate of air temperatures during summer can be attributed in part to insufficient cloud cover, which consequently influences downwelling shortwave radiation at the surface (F_R; Pinto and Curry 1997). Over an annual cycle, clouds over sea ice have a net warming effect at the surface due to enhanced downwelling longwave radiation (F_L), except for a relatively short period during midsummer (Curry and Ebert 1992). However, the largest overestimates in ARCSyM air temperature occur in lower-latitude, open-ocean areas where the summertime cloud forcing will yield a net cooling. Cloud amount averaged over May–September (Fig. 3) indicates that the underestimate in cloud cover indeed occurs mainly over open ocean rather than over the Arctic as a whole, as compared to the cloud summaries in Serreze et al. (1997a). Amounts are underestimated in the North Atlantic by about 40% and by about 20% over the Arctic Ocean relative to the ISCCP-D2 data. Mean cloud fractional coverage averaged over the model domain for the DT experiment decreases during midsummer, ranging from 0.43 in May to 0.38 in June and July, and increasing to 0.59 in September. Climatologies based on observations (Warren et al. 1986, 1988; Hahn et al. 1988) indicate typical summer cloud fraction of about 0.7 in May, increasing to nearly 0.8 in September. The mean cloud amounts produced by ARCSyM are closer to, but still less than, the ISCCP-D2 estimates (Serreze et al. 1997a). This apparent underestimate of cloud amount in the ARCSyM experiments appears typical of most current GCMs (Chen et al. 1995). Unlike these GCMs though, ARCSyM reproduces the observed tendency for cloud amount to increase in early autumn.

The difference in F_R between the AORF data and the ARCSyM DT results as averaged over the model domain is 6% over the May–September period (Fig. 4a). The DT experiment yields greater F_R in all months except August. The greatest difference of 10% (29 W m⁻²) occurs in June. Consistent with the spatial patterns of cloud amount, the differences between the AORF and ARCSyM fluxes are smallest over the Arctic Basin and largest in the North Atlantic and along the coastal ice margins in the Arctic Basin. The large differences over open ocean may be due to ARCSyM’s underestimate of cloud cover, while the differences over the ice pack may reflect the tendency of the Community Climate Model 2 radiation scheme used in ARCSyM to overestimate F_R in clear-sky conditions (Pinto et al. 1997). Pinto et al. (1997) also demonstrate the sensitivity of F_R to cloud optical depth, surface albedo, and the vertical moisture profile. In their example, a decrease in albedo from 0.80 to 0.40, or an increase in cloud optical depth from 2 to 4, reduces F_R by 50 W m⁻² over summertime ice. Since this sensitivity of F_R to optical depth increases as albedo decreases, and since the DT experiment tends to underestimate ice fraction as described below, the fully coupled simulation may be particularly sensitive to cloud conditions. The ISCCP-derived fluxes themselves may be overestimated (Curry et al. 1996; Serreze et al. 1998), so the actual error in the ARCSyM fluxes may be larger than that inferred here. Comparisons to monthly mean fluxes for 1990 measured at the North Pole drifting station NP 30 (Marshunova and Mishin 1994) suggest this, but such differences for a single location may be due as much to variations in synoptic conditions as to inherent biases in the retrieval methods. Clearly, given these sensitivities in the context of the magnitude of error in F_R over sea ice in the DT experiment, slight changes in cloud or surface properties could account for the observed differences.

Monthly averaged AORF and ARCSyM F_L for the model domain are fairly similar in all months. The mean difference over the ocean area covered by the POLES datasets is −10 W m⁻² for May–September (Fig. 4b). The maximum differences occur in May, when F_L from the DT experiment is 11% (24 W m⁻²) greater than the AORF estimate. The mean DT overestimate for June through September is 3%. The smaller differences seen in F_L compared to F_R over open ocean may be due in part to the greater net radiative cloud forcing for F_R expected during midsummer over open ocean, as well as to ARCSyM’s overestimate of air temperatures seen in Fig. 2, which yields increased F_L to offset decreases due to the underestimate of cloud cover.

A time series of daily radiative and turbulent fluxes for an individual grid cell in the Beaufort Sea (Fig. 5) illustrates the general situation noted above for the monthly means. Compared to other values typical of this location in the Arctic Basin (e.g., Curry and Ebert 1992; Fig. 9a), the net fluxes are reasonable. Spatial patterns of ocean salinity and temperatures (Fig. 6) are also reasonable. Simulated temperatures in most open ocean areas agree well with AVHRR-derived sea surface temperatures in May–August, although the model overestimates temperatures in the Greendland Sea, Baffin Bay, and Arctic Basin coastal waters by about 5°C in September. These areas correspond to locations where cloud amount is underestimated and where F_R is too large.

Monthly mean SLP from the ARCSyM DT experiment (Fig. 7) show many of the basic features seen in the ECMWF fields but with some significant differences in spatial patterns, as discussed in detail below. Differences in SLP averaged over the model domain are less than 2 hPa. The anomalous low pressure in the eastern Arctic in May is reproduced reasonably well. The center of the low is located correctly in the DT experiment, but is weaker than in the ECMWF data. The DT simulation lacks the observed high pressure cell over the eastern Beaufort Sea. In June, the model SLP patterns and magnitudes are again similar to the ECMWF fields. ARCSyM produces a ridge of high pressure extending across the Arctic Basin, although the DT simulation continues to underestimate pressures in the Beaufort and also in the eastern Siberian Sea. In July, ARCSyM captures the general SLP patterns, but fails to produce the closed low pressure region north of Greenland and the Canadian Archipelago. The largest differences between the ARCSyM and ECMWF SLP fields occur in August. Here, the tendency of the model to underestimate pressures in the Beaufort Sea is most apparent. In September, the ARCSyM SLP are again significantly in error, with high pressure located in the Arctic Basin rather than centered over the central Siberian coast.

Statistics of cyclone frequency and position summarize the model’s ability to reproduce the short-period variability in SLP patterns that affect ice deformation and transport. To this end, the 6-h SLP output from the ARCSyM experiments and the 12-h ECMWF output were passed through a modified version of the cyclone detection and tracking algorithm described by Serreze (1995) and Serreze et al. (1997b). This algorithm detects cyclones by inspecting whether gridpoint SLP values are surrounded by higher pressure values based on a user-specified detection threshold, and then tracks systems using a “nearest neighbor” analysis of cyclone positions on subsequent SLP charts. Modifications for application to the ARCSyM results involve setting the cyclone detection threshold to 0.5 hPa as opposed to 1.0 hPa to account for the model’s higher grid resolution, and decreasing the maximum distance search parameters for system tracking to be appropriate to 6-h as opposed to 12-h output. For the 12-h ECMWF fields, larger search distances were used but we retained a 0.5-hPa detection threshold.

As summarized for the entire model period, ARCSyM correctly depicts the ECMWF results in showing frequent cyclone activity over Alaska and the Yukon and relatively high counts over central and eastern Siberia. These are known to be regions of frequent summertime cyclogenesis (Serreze 1995). The DT model experiment and the ECMWF analyses also show frequent cyclone activity in eastern Canada and Baffin Bay but only limited activity over the Atlantic side of the Arctic, the latter illustrating the typically weak North Atlantic storm track in summer. As expected, total counts are much lower in the ECMWF data due to the use of 12-h SLP instead of the 6-h fields available from ARCSyM. The most salient feature in the model output is relatively frequent activity over the Arctic Ocean in the Laptev Sea arising from the persistent storminess during May and reflected in the SLP fields for this month.

Figure 8 shows frequency histograms of mean cyclone central pressure calculated from the ARCSyM DT and ECMWF sea level pressure fields. The TO experiment results included in Fig. 8 are discussed later in section 3c. So as to be most compatible with the ECMWF data, the results shown here are based on the 0000 and 1200 UTC ARCSyM output only. There is no evidence of significant differences in central pressures or in other cyclone statistics between the 0000 and 1200 and the 6000 and 1200 UTC ARCSyM output. There is little difference in cyclone central pressures between the DT experiment and the analyses—both show approximately normal distributions with a mean of about 1000 hPa. Figure 9 summarizes temporal changes in cyclone activity as bimonthly means of central pressure. The ARCSyM SLP agree with ECMWF in showing a summertime increase in central pressures. The tendency for summer systems to be weaker than those in winter, spring, and autumn is well documented in earlier studies (e.g., Serreze 1995). In contrast with central pressure, intensity falls off throughout the model period, indicative of an overall decline in the background pressure field. While lower intensities characterize the ECMWF SLP for the reasons noted above, the ECMWF-derived cyclone statistics show a general decline in intensity over the summer.

As is the case for the other variables discussed above, the DT simulation reproduces some of the general characteristics of the sea ice cover, but with significant biases and regional differences compared to observations. DT ice fractions in May and June are similar to the SSM/I values within the interior pack, and underestimated by about 20% along the Alaskan and Canadian coast. Ice fraction is reduced along the Siberian coast, but with no actual open water area generated. In June, the DT ice fraction in the Beaufort Sea remains greater than 0.95, but with a gradual decrease eastward toward the Siberian Sea and still with no open ocean along the coast. In contrast, the SSM/I data show substantial open-water area in the eastern Arctic, and a much more consolidated and abrupt ice edge, with ice fractions of about 0.9 close to the ice margin in the Siberian Sea. The SSM/I data continue to indicate large areas of greater than 0.85 ice fraction over most of the Arctic in July, but with an increasingly extensive ice-free area in the Siberian Arctic. The DT experiment produces some regions of 0.85 ice fraction in the central Arctic, but consistent with its overestimation of surface temperature and downwelling fluxes in summer, ARCSyM also generates large areas with less than 0.7 ice cover. However, only small areas of open ocean are present along the coasts in the DT experiment, and the simulated ice edge in the Siberian Arctic is less compact than observed. By September, the DT experiment still produces relatively little open water along the Siberian coast, in contrast to the large area of open water apparent in observations, although the reduction in ice extent in the Beaufort Sea is reasonable (see Fig. 1). As noted earlier, the stand-alone ice model forced with NCEP fields (the SDT experiment) was able to reproduce the large reduction in ice extent, but the stand-alone model’s results are similar to ARCSyM in that the simulated ice edge was less compact than observed.

b. Details of the 1990 ice extent anomaly

This inability of ARCSyM to reproduce the observed reduction in ice extent in 1990 sheds additional light both on the mechanisms involved in driving the ice anomaly, as well as the level of model performance needed to correctly simulate interannual variability in ice cover. Here, we consider further the details of the observed ice and atmospheric conditions associated with the 1990 ice reductions, and the ability of ARCSyM to reproduce these conditions.

Key to controlling the 1990 ice extent minimum appears to be the above-normal cyclonic activity in April–May, and in late July–early August, combined with southerly winds resulting from the more typical midsummer pressure patterns present in June–July. Atmospheric circulation in most months generated southerly or easterly winds that transported ice away from the Siberian coast, advected warmer air into the region, and affected downwelling radiation as noted below. In keeping with the strong low pressure system over the Laptev Sea in May, ice transport observed in the SSM/I data and buoy motions and simulated by the stand-alone model SDT experiment (Fig. 10) show strong northward ice drift near the New Siberian Islands. The SSM/I-derived ice velocities and ice fraction in Fig. 10 also illustrate how closely the motion patterns associated with the cyclonic pressure regime in May appear linked to the production of the ice-free zone adjacent to the Siberian coast. This coastal polynya began developing in late April and extended from about 140° to 170°E in May. On average in May, open water (ice fraction less than 0.15) covered about 22 000 km², with an additional area of 130 000 km² with ice fraction between 0.15 and 0.50 in the SSM/I data. Examination of the daily SSM/I ice concentrations, ice motion, and daily forecast-model pressure fields suggests that this coastal polynya began as reductions in ice fraction in April in conjunction with strong low pressure systems in the eastern Arctic. Frequent and persistent cyclonic activity continued through May, generating the open water along the coast earlier than seen in any year from 1979 to 1997.

Typical summer mean circulation patterns through July continued to favor southerly winds in the Siberian region. Closed high-pressure systems over the Beaufort and Barents Seas in July, combined with low pressure north of Greenland, combined to continue the general pattern of ice advection from the eastern Arctic to the central Arctic. In late July–early August, atmospheric circulation was again unusual and somewhat similar to May, with frequent low pressure systems in the Laptev and Barents regions and high pressure in the western Arctic. SSM/I data show below-normal ice extent through July, followed by a large removal of ice-covered area at the end of July through the first half of August. This reduction in ice extent coincided with southerly and southeasterly winds driven by high pressure over the Beaufort Sea and low pressure in the Laptev–Barents region, with particularly strong low pressure systems developing in early August (below 980 hPa on 8 August). From 25 July through 15 August, these southerly winds contributed to a 13% decrease in Arctic Basin ice extent. This reduction in ice cover was maintained through September, yielding the minimum observed ice extent for 1979–97. Due at least in part to northward ice advection, the SSM/I-derived ice fraction was greater in the northern Beaufort Sea in summer 1990 by about 10% than at any time from 1979–96 (e.g., mean of 0.95 vs a mean of 0.85 for all Augusts from 1979–96).

In comparison to these observed conditions, the DT results suggest that relatively slight errors in SLP patterns in spring, combined with larger errors later in the summer, significantly affected the simulation of the ice extent reduction. Ice fraction in the DT experiment is reduced adjacent to the Siberian coast in May, but no grid cells consist entirely of open water. This is in part due to the model resolution, but even if the averaging effect of the 100-km grid cells is considered, ice fractions remain overestimated. The changes in the simulated ice cover instead appear mainly as a reduction in ice thickness in May–June. As shown earlier in Fig. 10, the region of maximum simulated ice transport is shifted to the west into the Chukchi Sea instead of near the New Siberian Islands, and the strength of the ice transport is less than observed due to the weaker Laptev Sea cyclone and the underestimate of SLP over the Beaufort Sea.

In subsequent months, the errors in SLP in the DT experiment tend to favor a net gain of ice in the Siberian Sea due to ice transport, whereas the stand-alone experiment SDT shows a net decrease due to transport through midsummer. Bearing in mind that ARCSyM was initialized with the ice thicknesses from the stand-alone model, the maximum reduction in ice cover due to dynamics in the region of greatest northward transport and divergence in the Siberian Sea is 0.3 m in May in the DT simulation, compared to 0.6 m in the SDT experiment. By the end of June, a total of 0.3 m of ice is lost due to melt over most of the Siberian Sea in the DT experiment, while the net reduction due to transport decreases to 0.05 m, indicating advection of ice into the region rather than the small amount of export seen in the SDT simulation during June. In July, the DT experiment did not reproduce the closed low-pressure system north of Greenland. The net effect of this low pressure region as determined from the SDT results is to advect ice northward within the central Arctic. By failing to generate this low, the DT simulation continues to transport ice southward from the central Arctic into the Siberian Sea. This low pressure region also produces an area of ice divergence north of the New Siberian Islands in the SDT simulation, rather than the uniform southward motion simulated by ARCSyM. Later in summer, the DT pressure fields continue to yield a slight but generally southward ice advection into the Siberian Sea from the central Arctic. In contrast, the strong high pressure system in the western Arctic shown in the August ECMWF, NCEP, and IABP sea level pressure fields generates a well-defined Beaufort Gyre in the SDT experiment, with advection into the Siberian Sea from the regions of low ice concentration in the southern Beaufort and Chukchi Seas, and with rapid drift northward along the Transpolar Drift Stream originating north of the New Siberian Islands.

In keeping with the southerly winds associated with the low atmospheric pressure in the eastern Arctic in April and May, and possible additional heating provided by the polynya along the Siberian coast, observations of surface air temperatures in May show that temperatures were above normal north of the Siberian coast. This is consistent with the tendency for higher temperatures over Siberia during years with above-normal cyclonic activity in the eastern Arctic (Rogers and Mosley-Thompson 1995), as was the case in 1990. Surface air temperatures for May 1990 averaged −0.5° at approximately 75°N, 160°E versus −8.5° in 1988, a year with spring SLP patterns typical of most previous years. In 1990, melting of the ice pack (defined as days with air temperatures of 0°C or greater) in the Siberian Sea started earlier and occurred for more days than in any other year from 1979–96. The average date for start of melt and the number of melting days for a 1200 km × 1200 km region encompassing the Siberian Sea for 1979–96 is 11 May and 47 days, respectively. In 1990, melt started on 25 April, with 61 days of melt. ISCCP-derived cloud fractions in May were below normal over the eastern Siberian Sea, resulting in above-normal downwelling shortwave fluxes. The spatial patterns of higher air temperatures (Fig. 11a), decreased cloud fraction, and increased downwelling flux were well defined and collocated with the region of subsequent ice reduction, and thus likely contributed to the observed decrease in ice extent (e.g. Maslanik et al.).

Overall, the ARCSyM DT experiment produced a muted version of these conditions affecting melt. Surface air temperatures are higher over the eastern Arctic than elsewhere in the surrounding area (Fig. 11b) but are slightly lower than the POLES temperatures. Temperatures are overestimated elsewhere in the Arctic Basin, such that the temperature contrast from the central Arctic to the Siberian and Chukchi Seas is considerably less than observed. The higher temperatures are shifted toward the Chukchi Sea, in keeping with the slight misplacement of the center of the simulated mean low pressure area in May. Similar to the AORF data, downwelling radiation is greater over the eastern Siberian Sea than over the rest of the Arctic Basin, but again with less well-defined spatial differences compared to the satellite-derived mean radiation data for May 1990.

In addition to the effects of discrepancies in the simulated atmospheric conditions on ice transport and melt, characteristics of the ice model itself could have affected the ability of ARCSyM to reproduce the observed ice extent reductions. For example, the albedo treatment in the stand-alone ice model is dependent on snow and ice thickness. A thinner ice and snow cover results in a lower albedo, thus providing a positive feedback to increase melt. In contrast, ARCSyM’s current albedo scheme is a simple function of surface temperature, which may have underestimated the amount of melt in the Siberian Sea where the above-normal F_R should have favored greater melt. However, the lack of open water production along the Siberian coast in April through May appears to have depended more on dynamics than melt, and perhaps reflects aspects of the ice rheology as well as the grid resolution of the model. Convergence of the pack in response to compressive forces might have been particularly important for a case such as spring 1990, where southerly winds attempted to transport ice northward into the consolidated and predominantly multiyear ice pack. With the cavitating fluid rheology, ice flow tends to deflect around such areas of thick and highly convergent ice cover as noted earlier. This is mitigated by increasing resistance to shear through the use of the Mohr–Coulomb modification, but the DT ice velocities suggest some tendency for increased lateral flow of the ice in resistance to the northward advection in 1990. Removal of ice along the coast in the model will also be affected by ice strength, parameterized in ARCSyM as a function of ice thickness and ice fraction. As noted above, in a case such as 1990 when melt was likely to have been more extensive than usual, the strength parameterization may have overestimated the actual resistance of the ice pack to convergence. Also, the width of the polynya as reflected in the original, 25-km resolution SSM/I data was about 200 km in mid-May. The dynamics involved in producing such a relatively narrow open-water feature between the land margin to the south and a relatively compact ice pack to the north likely require finer model resolution than the 100-km grid cells used here for ARCSyM.

Other factors that may be significant for simulating the 1990 ice extent reduction include the effects of heat absorption in leads on lateral versus basal ice melt (Maykut and Perovich 1987) and the parameterization of the rate of change in ice fraction during melt (Holland et al. 1993). In the DT experiment, ice fraction appears to decrease at the expense of ice thickness. As an illustration, Fig. 12 shows mean ice thickness for September. In contrast to the relatively smooth spatial distribution of ice fraction simulated by ARCSyM, the thickness contours indicate thick ice in the central Arctic, but a fairly extensive area of thinner ice in the Siberian Arctic. The location of fairly abrupt increase in thickness along the northern portions of the Siberian and Laptev Seas corresponds well with the observed ice edge location in Fig. 1. Lower albedo and additional allocation of heat to basal melt combined with southerly winds to advect the remaining ice northward might substantially improve the simulation of 1990 ice conditions.

To some extent then, the inability of the DT simulation to generate the reduced ice extent seen in 1990 reflects a combination of relatively slight errors in SLP in spring and early summer compounded by larger errors in August–September, perhaps augmented by limitations in the ice model’s rheology and energy budget treatments. The fact that ARCSyM reproduces some aspects of the beginnings of the ice anomaly pattern in April and May, but failed to produce the large reduction in ice extent seen in August and September suggests that the unusual 1990 ice conditions required the frequent southerly winds and northward ice advection observed in late July and early August, which ARCSyM did not reproduce well. An ice extent anomaly such as that seen in 1990 thus appears quite sensitive to the timing and duration of weather systems, in addition to the strength and location of these systems.

It is important to reiterate that the mixed layer ocean model used here in ARCSyM does not include dynamics, and no ocean currents are prescribed. Based on the correspondence of observed ice reductions to SLP patterns, we assume that ocean currents were not as critical as atmospheric circulation for the development of the observed ice extent anomalies in 1990. Consistent with this is the fact that the stand-alone model with no ocean currents simulated the reduction in ice extent quite well, and the SDT experiment using climatological ocean currents produced only a slightly more extensive ice cover in 1990 (see Fig. 1). However, ocean currents in general clearly play a role in sea ice variability, either through direct transport of the ice or through effects on ice growth due to modification of ocean heat fluxes or salinity. The inclusion of an effective dynamical ocean model is a goal of the ARCSyM development effort.

c. The role of ice transport versus ice melt

The above discussion of atmospheric circulation patterns, air temperatures, and radiative fluxes suggests that melt and ice transport both contributed to the 1990 ice extent anomaly. Limiting the model simulations to thermodynamical processes so that ice transport is excluded allows us to further explore the degree to which the reductions in ice extent depended on a combination of melt and ice advection. Excluding dynamics in the TO experiment yields an increase in ice extent and substantially higher ice fractions compared to the DT results. Whereas some open ocean area appears adjacent to the Siberian coast in the DT simulation, the TO experiment produces no open water in the location of the observed reduction in ice extent. The companion stand-alone model experiment (STO) further suggests that a combination of dynamical and thermodynamical processes contributed to the ice extent anomaly in 1990. As with the TO experiment, the STO simulation failed to produce open water along the Siberian coast. As noted earlier, the anomaly was simulated well in the STD experiment (Fig. 1).

The greater reduction in ice extent seen in the DT and SDT experiments compared to the thermodynamics-only simulations suggests a dominant role for ice transport. However, a summation of net loss of ice in the Siberian sector due to dynamical and thermodynamical processes in the SDT experiment over the entire period showed that most of the decrease in ice cover in the Siberian Arctic occurred through melt rather than through ice advection from the region (Maslanik et al. 1997). This illustrates the effects of a potential feedback between ice melt and local and regional ice dynamics. For example, the presence of above-normal air temperatures and radiation observed in May 1990 may have combined with the likely disruption of the ice cover in response to the frequent and strong low-pressure systems to yield additional heat absorption within leads and the coastal polynya. This would result in increased lateral and basal melting of the sea ice, producing more open water and decreased ice albedo. Consistent with this, the analysis of net accumulation due to dynamical and thermodynamical processes in the SDT experiment suggests that the reduction in ice cover in the Siberian sector in May due to dynamics is critical for fostering enhanced melt in the later months. In this sector, depletion of ice due to dynamics is slight (but still a net loss) from June through July, but the area of greatest loss due to melt occurs in the region where ice fraction had been reduced by dynamical processes in May. In addition, if the ice pack was thinner than normal as the simulations and energy budget observations suggest, the thinner ice cover would have been more susceptible to ridging and rafting. This would have favored increased northward advection in response to the southerly winds that occurred in late July and August, and would have further contributed to the positive feedback between ice transport and increased melt.

d. Significance of ice–atmosphere coupling

The TO results also provide some insight into the degree of coupling between the atmosphere and sea ice cover. The most obvious differences in atmospheric conditions when ice dynamics are excluded are slightly lower mean air temperatures, a 3% decrease in cloud amount, a 2% increase in F_R, 2% decrease in F_L, and a slight change in mean SLP as averaged over the model domain. Although the mean effects are relatively small, some notable regional differences in SLP patterns emerge. In general, the TO experiment produces higher pressure in regions such as the Beaufort Sea, where the DT experiment underestimates SLP. For example, SLP in August (Fig. 13) is considerably improved compared to the DT experiment results shown earlier in Fig. 7. Similar SLP differences are found in other months, although like the DT experiment, the TO simulation fails to produce the low pressure area north of Greenland in July.

In terms of cyclone statistics, the difference field between ARCSyM experiments (TO-DT, Fig. 14) shows few coherent patterns, the primary exception being that the TO experiment depicts generally more cyclone activity in the Laptev Sea as well as an eastward shift in this activity, and fewer cyclones in the Chukchi and Beaufort Seas. This change is manifested in the model differences in SLP distributions and cyclone counts (given earlier in Figs. 8 and 9). There is little difference in cyclone central pressures between the three models. Corresponding histograms of cyclone intensity for the two ARCSyM experiments, based on the Laplacian of SLP at the cyclone centers, are also very similar (not shown). The Laplacian is proportional to the geostrophic relative vorticity and unlike central pressure gives a measure of intensity independent of changes in background central pressure. Temporal changes in cyclone activity in the TO experiment are similar to those in the DT experiment and the ECMWF fields, but the decrease in cyclone central pressures in June and August (Fig. 9) suggests some deepening of cyclones in addition to the more frequent Laptev Sea cyclones shown in Fig. 14. Mean SLP for these months reflect this, with lower SLP in the eastern Arctic and a strengthening of high pressure over the Beaufort Sea. As a result, overall differences in SLP between the TO and DT experiments are slight when averaged over the model domain. However, as discussed below, such changes can have a substantial effect on ice transport.

The discussion above and in section 3b illustrates that accurate simulation of the reduction in ice cover seen in 1990 depends on the ability of the model to reproduce regional atmospheric circulation patterns that affect ice melt and transport. As indicated by the differences in SLP between the TO and DT simulations, changes in sea ice conditions affect atmospheric circulation in a way that can then feed back into changes in ice extent and ice fraction. Errors in model components therefore can combine and accumulate over time to further reduce the simulation performance. Prescribing the ice cover in ARCSyM allows us to investigate this sensitivity further. The PR1 experiment sets ice fraction to match the substantially greater fractions produced in the TO simulation, but includes full ice dynamics. Comparing the DT, TO, and PR1 results thus indicates the degree to which an increase in ice fraction affects atmospheric conditions.

SLP patterns for the PR1 experiment lie between those of the DT and TO results. For example, the anticyclone in the Beaufort Sea in May is stronger than in the DT experiment, but weaker than in the TO simulation. As shown in Fig. 15 though, even relatively subtle changes in the centers and strength of the high and low pressure systems such as seen in the PR1 experiment and the TO results discussed above are significant, given the contribution of wind-driven ice transport to the ice extent anomaly. In this case, the strengthening of the Beaufort high-pressure cell in May increases ice velocities and helps shift the region of maximum ice transport westward toward the correct location in the Siberian Sea. SLP patterns in August are similar to the TO results in that the location of high pressure over the Beaufort Sea and low pressure over the Laptev Sea agree better with the ECMWF SLP than was the case for the DT experiment. However, the pressure differential between the centers of the high- and low-pressure cells is weaker than in the ECMWF and TO experiment SLP. As a result, northward ice advection remains underestimated. Overall though, these changes in atmospheric conditions in the PR1 experiment yield a slightly greater retreat in ice extent by September than was produced by the DT simulation.

Given these sensitivities of atmospheric circulation to variations in ice cover, it is useful to consider the degree to which a more accurate representation of ice extent and ice fraction could further improve ARCSyM’s simulation of atmospheric conditions. The second prescribed-ice experiment (PR2) addresses this by setting the ice cover within ARCSyM to the ice fraction and extent estimated from SSM/I data. As outlined earlier, the observed ice conditions differ in three main ways compared to the ice cover simulated in the ARCSyM DT, TO, and PR1 experiments—higher ice concentrations throughout the Arctic, particularly in midsummer; less extensive ice cover, particularly along the Siberian coast; and more compact ice-pack margins.

The changes in monthly mean SLP in the PR2 experiment are consistent with the results of the TO and PR1 experiments, yielding relatively subtle changes in SLP with similar spatial patterns in May–August, but with fairly large differences in September in this case. The unusual presence of open water along the Siberian coast in April and May could have affected cyclone development by providing a potential source of turbulent heat flux (LeDrew et al. 1991). Perhaps reflecting this reduced ice fraction as prescribed by the SSM/I data, the low pressure system in May in the Siberian Arctic deepens by 2 hPa to 996 hPa in the PR2 experiment compared to the DT simulation. In keeping with the associated strengthening of the southerly winds over the Siberian Sea, the mean surface temperature over the ice pack north of the Siberian coast in May is approximately 0.75° greater than in the DT experiment, and 0.5° greater than the TO results. As noted earlier though, the typical width of the coastal polynya as reflected in SSM/I data was about 250 km or less. When the SSM/I data are averaged from 25-km to the 100-km resolution of ARCSyM, the polynya loses much of its character, with no ice-free grid cells remaining. As a result, the simulated surface temperatures and heat fluxes fail to reflect the typical 5°–10° contrast in temperatures expected between ice and open water in late April–May. The model resolution of 100 km is thus not sufficient to resolve the full effects of such a feature.

SLP is improved slightly in June, but the greatest difference between the DT and PR2 experiments is seen in September (Fig. 16). While the pressure patterns still differ significantly from the ECMWF SLP, the patterns are considerably improved compared to the DT results. The PR2 simulation correctly yields low pressure in the western Arctic and high pressure in the eastern Arctic, whereas the opposite was the case in the DT results shown earlier in Fig. 7. As is typical of the other experiments though, SLP within anticyclones tends to be underestimated. The high pressure area in the eastern Arctic is too weak and is misplaced westward toward the Barents Sea instead of over the Laptev Sea, while the strength of the low pressure region over the Beaufort Sea is overestimated. However, winds in the Siberian Arctic now have a westerly component, with a pressure gradient more similar to the ECMWF SLP. As in most other months, further strengthening of the high over the eastern Arctic would substantially improve the ice transport patterns.

Thus, the PR2 results illustrate that improving the accuracy of the model’s simulation of ice fraction and extent can improve the simulated SLP patterns and surface temperature. Remaining errors can be attributed to other model components such as the treatments of cloud–radiation interactions and the surface energy balance. Overall, the fact that SLP patterns are least changed in midsummer and more affected, and improved, in the colder months suggests that the effects of differences in ice cover are most apparent as changes in turbulent fluxes. As a result, errors in the sea ice simulation are likely to cause greatest errors in a coupled model during winter.