Sensitivities and uncertainties in a coupled regional atmosphere-ocean-ice model with respect to the simulation of Arctic sea ice

Authors


Abstract

[1] A series of sensitivity experiments using a coupled regional atmosphere-ocean-ice model of the Arctic has been conducted in order to identify the requirements needed to reproduce observed sea-ice conditions and to address uncertainties in the description of Arctic processes. The ability of the coupled model to reproduce observed summer ice retreat depends largely on a quasi-realistic ice volume at the beginning of the melting period, determined by the relationship between winter growth and summer decay of ice. While summer ice decay is strongly affected by the parameterization of the sea-ice albedo, winter ice growth depends significantly on the parameterization of lateral freezing. Reciprocal model biases due to uncertainties in the atmospheric energy fluxes can be compensated to a certain extent. However, potential underlying weaknesses of the model cannot be eliminated that way. Since lateral freezing also determines the ice concentration during winter, and thus the heat loss of the ocean and the near-surface air temperature, the model tuning possibilities are limited. A large uncertainty in the model relates to the simulation of long-wave radiation most likely as a result of overestimated cloud cover. The results suggest that uncertainties in the descriptions for Arctic clouds, snow, and sea-ice albedo, and lateral freezing and melting of sea ice, including the treatment of snow, are responsible for large deviations in the simulation of Arctic sea ice in coupled models. Improved descriptions of these processes are needed to reduce model biases and to enhance the credibility of future climate change projections.

1. Introduction

[2] Sea ice plays a prominent role in the Arctic climate system because the presence of sea ice modifies the exchange of heat, moisture, and momentum between atmosphere and ocean, and therefore the atmospheric and oceanic processes and circulations which in turn have an impact on the existence and spatial distribution of sea ice. Furthermore, the sea-ice-albedo feedback effect is an important factor in the amplification of climate change in the Arctic [e.g., Curry et al., 1995], so that the changes in Arctic sea ice have the potential to impact Arctic and global climate significantly [Dethloff et al., 2006]. Hence a realistic simulation of Arctic sea ice is one of the major challenges in coupled climate modeling.

[3] Recent coupled-model intercomparison studies have shown that different global atmosphere-ocean-ice models (hereafter referred to as AOI models) produce quite different sea-ice thickness and extent already in their present-day climate [Walsh and Timlin, 2003; Holland and Bitz, 2003; Flato and Participating CMIP Modeling Groups, 2004]. Therefore it is not surprising that projections of the 21st-century ice extent by these models differ considerably from each other and are strongly dependent on the models’ simulations of present-day ice extent [Walsh and Timlin, 2003]. The outcome of this is a wide range in the projected polar amplification of climate change and thus in the magnitude and regional pattern of high-latitude warming and its potential consequences. The uncertainty in modeling present-day Arctic sea-ice conditions and in its implications for climate projections indicates the need for improved descriptions of physical processes involved in atmosphere-ice-ocean feedbacks.

[4] This paper addresses the ability of a pan-Arctic coupled regional AOI model to reproduce present-day Arctic sea-ice conditions. The regional model approach allows for realistic large-scale atmospheric conditions at the model’s lateral boundaries as well as the simulation of atmosphere-ice-ocean interactions with high resolution. Previous studies with pan-Arctic coupled and uncoupled regional models by Maslanik et al. [2000] and Rinke et al. [2003] have highlighted the importance of the atmospheric circulation in driving anomalous sea-ice retreat during summer 1990. However, the fully coupled versions of their models were not able to simulate both sea-ice anomaly and atmospheric circulation in a satisfactory manner. A potential shortcoming of the two studies could be that the associated coupled-model simulations were carried out only for a single year using initial ice thicknesses from stand-alone simulations.

[5] In the following sections, we first describe the model components in sections 2.1, 2.2, and 2.3, the coupling procedure in section 2.4, and the boundary conditions in section 2.5. An overview about the model experiments is given in section 3, while the results of these experiments are presented in section 4. There, we comment on the need for an adequate spin-up time in modeling Arctic sea ice with a coupled regional model. We further address sensitivities and uncertainties in the model, and we discuss some necessary improvements of Arctic process descriptions required to simulate Arctic sea ice in a more realistic fashion. The primary focus is set on the simulation of the Arctic sea-ice anomaly during summer 1998 when unusually strong reductions in ice cover occurred in the Beaufort and Chukchi seas and along the Canadian coast [e.g., Maslanik et al., 1999], but some of the results are transferable to other years as well. The year 1998 was chosen because a number of observations are available from the Surface Heat Budget of the Arctic Ocean (SHEBA) project, which included a yearlong field campaign on a drifting ice station in the Beaufort Sea from October 1997 through October 1998.

2. Model Description

[6] The coupled regional AOI model used in this study is a composite of the two stand-alone models HIRHAM and NAOSIM. Each of them has been applied for a wide range of Arctic climate studies [e.g., Dethloff et al., 2002; Rinke et al., 2004; Karcher et al., 2003; Kauker et al., 2003], and also an earlier version of the coupled HIRHAM-NAOSIM has already been used by Rinke et al. [2003] in a case study for anomalous Arctic sea-ice conditions. In accordance with the stand-alone models, domain choice and configuration of the coupled model’s components have been retained unchanged.

2.1. Atmosphere Model

[7] The atmospheric component HIRHAM [Christensen et al., 1996] was set up on an integration domain that covers the whole Arctic north of about 60°N (see Figure 1) at horizontal resolution of 0.5° (∼50 km) on a rotated latitude-longitude grid with the North Pole on the geographical equator at 0°E [Dethloff et al., 1996]. In the vertical, the model has 19 unevenly spaced levels in hybrid sigma-pressure coordinates from the surface up to 10 hPa with the highest resolution in the lower troposphere. HIRHAM includes prognostic equations for horizontal wind components, temperature, specific humidity, cloud water, and surface pressure and diagnostic equations for vertical wind, geopotential, cloud cover, and mixed clouds. The equations are solved using a time step of 240 s.

Figure 1.

Geographical position of the model domains of the coupled model’s atmosphere component HIRHAM and its ocean-ice component NAOSIM. The domain of coupling (given by the overlap area) covers the whole Arctic Ocean, including all marginal seas, the Nordic seas, and parts of the northern North Atlantic.

[8] The physical parameterizations for subgrid-scale processes are adapted from the general circulation model ECHAM4 [Roeckner et al., 1996] and include comprehensive descriptions for radiative transfer, atmospheric boundary layer processes, gravity wave drag, cumulus convection, and large-scale condensation. In addition, balance equations for energy and hydrology at the land surface and a heat conductivity equation for the temperatures in five soil layers between the surface and 10 m depth are solved for all land grid cells. Over ocean grid cells, which are at least partly covered with sea ice, a heat balance equation is solved for the uppermost snow/ice layer to obtain the snow/ice surface temperature and the residual heat flux (Qai) available for freezing or melting of sea ice or melting of snow.

2.2. Ocean Model

[9] The ocean-ice component is the high-resolution regional version of NAOSIM [Karcher et al., 2003; Kauker et al., 2003]. It uses a horizontal resolution of 0.25° (∼25 km) on a rotated spherical grid, where the model equator corresponds to the geographical 30°W/150°E meridian, and 30 unevenly spaced z-coordinate levels in the vertical. NAOSIM’s ocean component is based on the Geophysical Fluid Dynamics Laboratory modular ocean model MOM-2 [Pacanowski, 1996] and includes prognostic equations for horizontal velocity components, potential temperature, and salinity and diagnostic equations for vertical velocity, density, and pressure. The model domain encloses the northern North Atlantic, the Nordic seas, and the Arctic Ocean (see Figure 1). The southern model boundary of NAOSIM is approximately located at 50°N in the Atlantic. Here an open-boundary condition has been implemented following the work of Stevens [1991], while all other boundaries, including the Bering Strait, are treated as closed walls. The model computations for both ocean and sea ice are carried out with a time step of 900 s. For more details on the model configuration, see the paper of Karcher et al. [2003] or Kauker et al. [2003].

2.3. Sea-Ice Model

[10] The NAOSIM version used here incorporates a dynamic-thermodynamic sea-ice model with elastic-viscous-plastic (EVP) dynamics according to the study of Hunke and Dukowicz [1997] and zero-layer thermodynamics following the work of Semtner [1976] with prognostic equations for ice thickness and ice concentration based on the widely used two-level sea-ice model by Hibler [1979] with two ice thickness categories (thick ice and thin ice/open water). In addition, NAOSIM allows for a prognostic snow layer on sea ice. The continuity equations for mean snow thickness (hs), mean ice thickness (hi), and ice concentration (ci) are given by

equation image
equation image
equation image

where equation imagei is the ice velocity; Sh, Ss, and Sc are the thermodynamic sources and sinks in terms of production rates; Ps is the snow production rate due to snowfall on sea ice; and Dc is the lead formation due to shear strain (dynamical ridging). Note that hi and hs are defined here as grid-cell mean ice thickness and snow thickness, respectively. This means that the ice (snow) mass per unit area is simply the product of ice (snow) density ρi (ρs) and ice (snow) thickness. Equations (1) and (2) thus reflect the conservation of snow and ice mass, while equation (3) is an empirical equation for the ice concentration. The thermodynamic production rates take the form

equation image
equation image
equation image

where Lf is the latent heat of fusion, Qa and Qo are the net atmospheric and oceanic heat fluxes, respectively, and h0 is a fixed reference thickness for lateral freezing, also referred to as lead closing parameter or demarcation thickness between thin and thick ice. h0 is a purely empirical parameter and will be discussed in section 3. The net atmospheric heat flux is given by the weighted average

equation image

where Qai and Qao are the net atmospheric heat fluxes over the sea-ice covered and the open-water fraction of the grid cell, respectively, which are calculated in the atmosphere model. Note that all fluxes are here defined positive upward and negative downward.

[11] The oceanic heat flux toward the ice-ocean interface is assumed to be directly proportional to the difference between the upper-ocean mixed-layer temperature (To) and the temperature at the ice-ocean interface, which in turn is assumed to be identical with the freezing temperature of seawater (Tfs):

equation image

Here ρw is the density and cpw is the specific heat capacity of seawater, Δz is the mixed layer depth, and τ0 is a fixed damping time constant for a delayed adaptation of the mixed-layer temperature, which is taken to be 3 days. Alternative formulations for the ratio Δz/τ0, which can be interpreted as a heat transfer rate, have been discussed by McPhee [1992] and Omstedt and Wettlaufer [1992] who relate the heat transfer rate to kinematic variables.

[12] The freezing temperature of seawater depends on the salinity and is calculated according to the formula

equation image

where Tf0 = 0°C is the freezing temperature of freshwater, So is the salinity at the ocean surface, which is assumed to be equal to the upper-ocean mixed-layer salinity, and βf = −0.0544 K/psu is a constant conversion coefficient. Note that the salinity in NAOSIM is actually given as deviation from the reference salinity Sref = 35 psu according to equation image = SoSref.

2.4. Model Coupling

[13] The coupling between HIRHAM and NAOSIM is carried out every hour, while the exchange of variables between ocean and ice model takes place at each time step. A full list of the exchange variables is given in Table 1.

Table 1. List of the Exchange Variables Between Atmosphere (A), Ocean (O), and Ice (I) Components of HIRHAM-NAOSIM
VariableStatusaFromToRequired for Computing of
  • a

    Status of “i” means instantaneous, “a” means averaged over coupling interval, while “r” means running mean over 12 coupling intervals. The running mean is applied to avoid sudden changes in the forcing due to strongly varying wind fields.

Wind stressrAO + IOcean current, ice drift
Heat fluxesaAO + ISea surface temperature, ice growth
Moisture fluxesaAOSea surface salinity
Precipitation rateaAO + ISea surface salinity, snow accumulation
Air temperatureaAIDifferentiation of snow/rain
Sea surface temperatureiOA + IHeat and moisture fluxes, ice growth
Sea surface salinityiOA + IFreezing point of seawater
Ocean currentiOIIce drift
Ice concentrationiIAGrid cell averaged fluxes
Ice thicknessiIAIce surface temperature
Snow thicknessiIAIce surface temperature
Momentum fluxiIOOcean current
Heat fluxiIOSea surface temperature
Freshwater fluxiIOSea surface salinity

[14] Since HIRHAM and NAOSIM use grids with different orientation and resolution, the variables must be interpolated from one grid to the other. For this purpose, an auxiliary grid is constructed in advance by extending the coarser grid, here the atmosphere grid, in such a way that it encloses the whole area of the finer grid. Each grid cell of the finer grid is then subdivided into 3 × 3 subgrid cells, and each of these subgrid cells is then assigned to a grid cell of the auxiliary grid.

[15] The interpolation between the grids is carried out in three steps. First, all land grid cells are filled up with respective values of the surrounding ocean grid cells by successive extrapolation. This is necessary to avoid inclusion of improper values into the interpolation due to different land-sea masks. The next step represents the actual interpolation, which is performed by averaging the values of all subgrid cells that have been assigned to the current grid cell of the auxiliary grid (ocean/ice to atmosphere) or by averaging the respective nine values of the auxiliary grid cells that have been assigned to the nine subgrid cells of the current ocean grid cell (atmosphere to ocean/ice). The wind stress vectors are transformed into the coordinate system of the ocean model by a simple rotation before their components undergo the interpolation procedure. As a last step, all ocean grid cells that lie outside of the overlap area of the two model domains are marked as uncoupled domain and afterward treated as in uncoupled mode, i.e., the models’ surface forcing in those regions is taken from observational data as described in the following section.

2.5. Boundary Conditions

[16] As the models are regional, forcing data are required at HIRHAM’s lateral boundaries and also at HIRHAM’s lower- and NAOSIM’s upper-boundary points that lie outside of the overlap area of the two model domains. These data are taken from the most recent European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses (ERA-40) and are updated every 6 hours at the HIRHAM’s lateral boundaries and every 24 hours at the surface boundaries outside of the overlap area.

[17] NAOSIM’s lateral boundaries, which are given as open boundaries (see section 2.2), are treated in a different way because reanalysis data are not available for the ocean. The model determines outflow and inflow points and allows for the outflow of tracers and the radiation of waves. At inflow points, temperature and salinity are restored with a time constant of 180 days toward a yearly mean climatology [Levitus and Boyer, 1994; Levitus et al., 1994]. The baroclinic part of the horizontal velocity is calculated from a simplified momentum balance, while the barotropic velocities normal to the boundary are taken from a lower-resolution version of the model that covers the whole Atlantic Ocean.

[18] At NAOSIM’s upper-boundary points outside of the overlap area, the input data comprise daily means of 2-m air and dew-point temperature, cloud cover, precipitation, wind speed, and wind stress. The atmospheric fluxes are calculated using standard bulk formulas, which are also used in the stand-alone version of NAOSIM. Since runoff from the land is not explicitly included, a salinity-restoring flux with an adjustment timescale of 180 days is added to the surface freshwater flux.

[19] Although there are only few ocean grid cells within the HIRHAM domain, which are not covered by NAOSIM (primarily over the northernmost part of the Pacific), a simple extrapolation from NAOSIM values onto these grid cells is not reasonable because they are too far away from the NAOSIM domain that similar sea surface conditions can be expected. For this reason, sea surface temperature and sea-ice cover fraction are taken from ERA-40, whereas ice thickness, snow thickness, and sea surface salinity are prescribed by default values (hi = 1 m, hs = 0 m, So = 35 psu).

3. Experimental Design

[20] To isolate the importance of uncertain process descriptions and initial conditions with respect to the thermodynamic evolution of sea ice in the model, a series of sensitivity experiments was conducted with HIRHAM-NAOSIM (Table 2). Each of these experiments covers the period from May 1989 to December 1999, and in each case, the atmosphere was initialized with ERA-40 data, while initial ocean and sea-ice fields were taken from a stand-alone run of NAOSIM described by Karcher et al. [2003] except for an initial ice experiment (h1.2-uni) where ice thickness and ice concentration were prescribed in the following way: If the stand-alone run yielded an ice concentration greater than 0.5, the initial ice concentration was set to 1.0 and the ice thickness was initialized with 1.0 m. In the other case, initial ice concentration and thickness were set to zero.

Table 2. HIRHAM-NAOSIM Sensitivity Experiments With Respect to Initial Ice Thickness, Lead Closing Parameter (h0), and Ice Albedo Scheme
ExperimentDescription
h1.2-stdControl run with h0 = 1.2 m, standard albedo scheme, and standard ice initialization
h1.2-uniAs h1.2-std but with uniform initial ice thicknesses of 1.0 m
h0.5-stdAs h1.2-std but with h0 = 0.5 m
h1.2-albAs h1.2-std but with new snow and ice albedo scheme
h2.0-albAs h1.2-alb but with h0 = 2.0 m
nospinupShort-term run for the year 1998 without model spin-up

[21] In addition, a coupled model experiment without an adequate own spin-up time was carried out (nospinup). This experiment was started on 1 December 1997 using initial fields from the stand-alone run of NAOSIM mentioned above and was exceptionally running only for the year 1998. In this short-term run, h0 was set to 1.0 m since the same value was also used in the stand-alone run (see comments in the paragraph after next).

[22] The thermodynamic evolution of sea ice depends largely on the atmospheric heat fluxes. During the cold season, the atmospheric heat flux over open water is generally much larger than over an insulating ice layer (i.e., QaoQai > 0). The area of open water is therefore important for the total heat loss of the ocean-ice system and determines the rate of ice growth. In equation (6), the empirical parameter h0 directly controls the rate of increase in ice concentration under freezing conditions (Qao > 0). This indirectly affects Sh by modifying the weighting of the respective contributions to Qa in equation (5). Consequently, a faster/slower increase in ice concentration leads to a decelerated/accelerated ice thickness growth. In this manner, h0 determines the relationship between basal and lateral sea-ice growth during the cold season and might indirectly influence summer sea-ice and atmospheric processes.

[23] As a result of the empirically based conservation equation for ice concentration, it is impossible to deduce a default value for h0 from physical principles. Hibler [1979] suggested a value of h0 = 0.5 m, derived from a simple thermodynamic experiment; on the other hand, Bjornsson et al. [2001] argued that a value of h0 = 1.0 m is suitable when modeling the large-scale Arctic sea-ice cover, while a value of h0 = 0.3 m leads to more reasonable ice concentrations within polynyas. Here two experiments were performed using h0 = 0.5 m (h0.5-std) and h0 = 1.2 m (h1.2-std), respectively.

[24] In contrast to winter conditions, the atmospheric heat fluxes during the melting period are dominated by the contribution of solar radiation. Because of the importance of the ice-albedo feedback for summer ice retreat [e.g., Lynch et al., 2001], a new snow and ice albedo scheme following suggestion 2 of Køltzow et al. [2003] has been tested in the coupled model. This scheme was derived from measurements during the SHEBA project and differs from the standard scheme in many respects. It includes, for instance, a quite different temperature dependency of the ice albedo due to an explicit temperature-dependent parameterization of melt ponds (Figure 2) and allows for areal fractions of snow, melt ponds, and bare ice. The major difference is that the new scheme decreases the ice albedo in most instances, particularly for melting conditions when the overlying snow cover has already disappeared.

Figure 2.

Temperature dependency of the sea-ice albedo in the standard scheme (dark gray lines) and the new sea-ice albedo scheme (light gray lines). The solid/dashed lines represent lower/upper limits of the albedo arising in case of no/complete snow cover.

4. Simulation Results

4.1. Ice Volume and Extent

[25] The top panel of Figure 3 shows the simulated ice volume within the NAOSIM domain (see Figure 1) from the experiments h0.5-std, h1.2-std, and h1.2-uni. Although at least the experiments h0.5-std and h1.2-std were initialized with sea-ice fields from a stable run of the stand-alone ocean-ice model, the modeled ice volume is far from a steady state at the beginning in all coupled model experiments. This is in contrast to the results of the stand-alone run in which the ice volume nearly persists at values like in the early stages of the coupled simulations (see Figure 4). However, for the coupled model, the initial ice volume appears to be too high in h0.5-std and h1.2-std, while it is obviously much too low in h1.2-uni where the model started from uniformly 1-m-thick ice.

Figure 3.

Simulated monthly means of sea-ice volume (top) and sea-ice extent (bottom) within the model domain from May 1989 (month 5) to December 1999 (month 132). The sea-ice extent is here defined as the area of all grid cells with at least 15% sea-ice concentration. For comparison, the SSM/I satellite-derived sea-ice extent (solid line) was calculated for the same domain. The model simulations were carried out with ho = 0.5 m (h0.5-std, dotted lines), with ho = 1.2 m and standard ice initialization (h1.2-std, dashed lines), and with ho = 1.2 m and initialization with uniform 1-m ice thickness (h1.2-uni, dot and dash line).

Figure 4.

As Figure 3 but for the uncoupled NAOSIM simulation (dashed line) and coupled model simulations with the new snow and ice albedo scheme using h0 = 1.2 m (h1.2-alb, dotted lines) and h0 = 2.0 m (h2.0-alb, dot and dash lines), respectively.

[26] Nevertheless, all coupled simulations arrive at a quasi-stationary cyclic state of equilibrium after about 6–10 years, and this equilibrium is only little affected by the initial sea-ice state even though the initial ice volume differs by a factor of 4. On the other hand, the coupled model’s state of equilibrium depends significantly on the rate of increase in ice concentration given by h0. The simulation with h0 = 0.5 m results in a mean ice volume that is just about half as large as using h0 = 1.2 m. This dependency on h0 is in qualitative agreement with the findings of Holland et al. [1993] who found an increase in mean ice thickness of about 1 m when decreasing the growth rate of ice concentration to 50% (equivalent to a doubling of h0).

[27] The corresponding sea-ice extent of the two h0 experiments is shown in the bottom panel of Figure 3 in comparison with SSM/I satellite-derived data using the NASA Team algorithm [Cavalieri et al., 1990, updated 2004]. The model generally overestimates the sea-ice extent during winter, and none of the experiments has been able to reduce this shortcoming substantially. In contrast, the simulation with h0 = 1.2 m agrees quite well with the observed summer ice extent after some years, while the simulation with h0 = 0.5 m tends to underestimate the summer ice extent considerably.

[28] A common result of the experiments is that summer ice extent is significantly correlated with the ice volume at the beginning of the melting period (ensemble correlation coefficient of 0.92 between ice volume in April and ice extent in September). However, the modeled ice volume is generally affected by ice growth during winter as well as ice decay during summer, and thus not only by h0 but also by the parameterizations relating to Qai and Qao in the atmosphere model.

[29] The experiments conducted using the new snow and ice albedo scheme show that due to a decrease of the ice albedo (h1.2-alb versus h1.2-std; see annotation in section 3), the energy input into the ocean-ice system increases, leading to quicker decay of sea ice during summer and accordingly to reduced ice volume at the end of the summer (Figure 4). However, the impact on ice volume is not as strong as due to the change of h0, and the resulting ice volume of experiment h1.2-alb lies finally between that of h1.2-std and h0.5-std. The same is true for the summer ice extent.

[30] As the new albedo scheme may be considered as more sophisticated than the standard scheme, leading in principle to a more realistic simulation of the ice-albedo feedback and the associated ice decay during summer, one may conclude that underestimated ice growth during winter is the major reason for the apparent underestimation of the ice volume in experiment h1.2-alb. This conclusion suggests that simulations with the new albedo scheme may yield more reasonable summer ice extent and concentration when using a higher value of h0 at the same time. Then, quicker ice loss during summer due to lower ice albedo is balanced by increased ice growth during winter, and the model’s state of equilibrium in ice volume is situated higher than before. This presumption is supported by the experiment h2.0-alb where the new albedo scheme is applied in conjunction with h0 = 2.0 m. The summer ice volume is here almost identical to that of h1.2-std, and also the summer ice extent agrees here much better with the SSM/I data than in experiment h1.2-alb. The impact on winter ice extent is marginal again.

[31] A rough comparison with available ice thickness observations, for instance with the climatologies of Bourke and Garrett [1987] and Laxon et al. [2003], shows that the simulated ice thicknesses in h0.5-std and h1.2-alb are definitely too thin in the model’s steady state (roughly from month 70 onward), while they are much closer to, but still somewhat thinner than, the observations in h1.2-std and h2.0-alb. On the other hand, the initial ice thicknesses (for May 1989) are clearly too thick in these four experiments. The latter holds for the whole stand-alone run and consequently for the run without spin-up as well. However, a detailed validation of the modeled ice volume is currently not possible since all available observational data include regional gaps.

[32] A comparison with ice thickness estimates derived from the extensive observations during SHEBA [Lindsay, 2003] is presented in Figure 5. The ice thickness data are available via URL http://www.joss.ucar.edu/cgi-bin/codiac/dss/id=13.122. Although this data set represents only regional ice conditions in the vicinity of the SHEBA ice camp, it indicates that the modeled ice thicknesses in h1.2-std and h2.0-alb can be considered as more realistic than in h0.5-std and h1.2-alb. This conclusion can also be drawn by comparing the modeled ice thicknesses with the ice thickness curve at the SHEBA Pittsburgh site produced by Huwald et al. [2005]. However, Huwald et al. [2005] have also shown that there is high spatial variability of ice thickness gauge measurements at small spatial scales, indicating the difficulties in validating modeled ice thicknesses with individual measurements.

Figure 5.

Mean ice thickness estimated from observations near the SHEBA drifting ice camp in the Beaufort Sea during 1997–1998 [Lindsay, 2003] and from simulations of HIRHAM-NAOSIM. Simulated ice thicknesses were interpolated from the model grid onto the respective position of the ice camp. The mean ice thickness includes the open-water areas. The term “SHEBA day” on the x axis corresponds to the day from the start of 1997. The time series were smoothed using a 7-day running mean.

4.2. Winter Ice Concentration

[33] Figure 6 shows the satellite-derived and modeled sea-ice concentration in March 1998. At first view, there is a rather good agreement between the simulations and the observation with respect to the ice edge, except for the Labrador Sea, where the model clearly overestimates the formation and persistence of ice. This model bias appears in all long-term experiments of the coupled model and all winters after the first melting season and is responsible for the overestimation of the total ice extent. In contrast, the uncoupled NAOSIM simulation and the run without spin-up agree better with the satellite data over the Labrador Sea, but on the other hand, they overestimate sea ice over the Greenland Sea more than the coupled model, so that the total ice extent is quite similar.

Figure 6.

Sea-ice concentration in March 1998 from SSM/I satellite-derived data (top left), an uncoupled NAOSIM run (top right), and the HIRHAM-NAOSIM experiments described in Table 2.

[34] At closer inspection, it is also visible that there is a well-defined correlation between h0 and the modeled open-water fraction within the ice covered area: With increasing h0, the open-water fraction increases too. In particular, in experiment h2.0-alb, there are large areas where the ice concentration is well below 95%. This is contrary to the observation. As a more quantitative measure, the root mean square error (RMSE) of sea-ice concentrations greater than 50% has been calculated and confirms that, with respect to winter ice concentration, h2.0-alb (RMSE = 8.4%) can be regarded as the most unrealistic and h0.5-std (RMSE = 6.5%) as the most realistic long-term experiment. The RMSE of the other experiments lies in-between (h1.2-std: 7.0%; h1.2-uni: 7.2%; h1.2-alb: 6.9%).

[35] The impact of the ice concentration on 2-m air temperature during winter is shown in Figure 7. HIRHAM-NAOSIM generally tends to overestimate the winter temperatures over the ice-covered ocean. The magnitude of this model bias depends on the ice thickness (compare nospinup and h0.5-std, which show similar ice concentrations but quite different ice thicknesses) but particularly on the open-water fraction due to enhanced heat transfer from the ocean to the atmosphere in the absence of an insulating ice cover. This finding agrees with the analyses of uncoupled HIRHAM simulations by Rinke et al. [2006]. The temperature overestimation increases considerably if the open-water fraction increases too. While the temperatures are about 8–12 K too warm in h0.5-std, they are more than 12 K too warm in h2.0-alb over most of the Arctic Ocean. Even if a value of h0 >1.0 m yields more reasonable ice thicknesses in the model, it prevents total freezing of the Arctic Ocean in winter and results not only in unrealistic sea-ice concentrations but also in too-warm temperatures.

Figure 7.

Difference of 2-m air temperature in winter 1997/1998 (December to March) between the same simulations of HIRHAM-NAOSIM as in Figure 6 and ERA-40 reanalysis data.

4.3. Summer Ice Concentration

[36] The satellite-derived and modeled sea-ice concentration in September 1998 is presented in Figure 8. The experiments demonstrate the effects of an unrealistic ice volume on summer ice extent and concentration: If the sea ice is too thin at the beginning of the melting period (as most notably in h0.5-std), the ice cover is quicker to open with the result of stronger ice retreat and underestimation of sea-ice concentration throughout the Arctic. In contrast, too-thick sea ice (as in the short-term experiment without spin-up) results in effects exactly the opposite to the above. The experiments in which the ice thicknesses are likely to be closest to reality (h1.2-std and h2.0-alb) also show the best agreement in ice extent and concentration.

Figure 8.

As in Figure 6 but for September 1998.

[37] This finding does not hold for the uncoupled model. Here a rather good agreement with the satellite data has been obtained with an ice thickness distribution which is similarly too thick as in the coupled model experiment “nospinup”. At least the observed opening of the Beaufort Sea is present in the uncoupled model, while it is completely missing in “nospinup” where the ocean-ice model is coupled to an atmosphere model. The different behavior of the coupled and the uncoupled model appears also in the ice concentration itself. The coupled model tends to overestimate the open-water areas within the summer sea-ice cover.

[38] Although experiment h1.2-alb shows quasi-realistic sea-ice retreat in the Beaufort Sea and also in the Barents and Kara Seas, there are considerably larger areas of open water in the Laptev and East Siberian seas. This underestimation of sea ice is associated with differences in the atmospheric circulation during the previous summer months (Figure 9). In contrast to observations and all other experiments, this experiment shows a pronounced cyclonic flow over the Laptev Sea which provides an atmospheric wind stress for drifting ice away from the East Siberian Sea toward the central Arctic Ocean and Kara Sea. The redistribution of ice mass within the Arctic leads to a situation in which thermodynamic loss of ice is regionally either intensified by dynamic ice loss or partly compensated by increased influx of ice.

Figure 9.

Difference of mean sea-level pressure in summer 1998 (June to September) between the same simulations of HIRHAM-NAOSIM as in Figure 8 and ERA-40 reanalysis data.

[39] As the effect of the atmospheric circulation on the sea-ice distribution is quite evident, the atmospheric response to incorrect sea-ice concentrations or thicknesses is not that definite. In particular, there is no linear relationship between over- or underestimated ice concentrations or thicknesses and deviations in the atmospheric circulation. The underlying nonlinear feedback process is not yet understood. However, it is remarkable that significant pressure deviations occur predominantly over the Laptev and Kara seas, a region where also strong deviations in the simulated ice edge appear. Such differences in the positions of the ice edge are able to trigger model deviation in mean sea-level pressure [Rinke et al., 2003, 2006]. On the other hand, the position of the ice edge in the Beaufort Sea, which varies considerably among the simulations, has obviously no or only marginal impact on the simulation of the atmospheric circulation. Differences in the simulated ice drift can therefore not be regarded as the main reason for the model variations in the strength of the simulated ice retreat in the Beaufort Sea.

4.4. Atmospheric Heat Fluxes

[40] Some potential shortcomings of the coupled model become apparent when looking at the mean seasonal cycle of atmospheric heat and radiative fluxes averaged over all sea areas north of 70°N (Figure 10). As noted before, a higher value of h0 leads to lower ice concentrations (sea-ice cover) during winter and to an increased heat transfer from the ocean to the atmosphere. This effect is visible in both the sensible and latent heat fluxes. However, the differences among the model simulations as well as between simulation and ERA-40 data are on average clearly lower than 10 W/m2. In accordance with the largest deviations in ice concentration, h2.0-alb also shows the largest deviations in the heat fluxes.

Figure 10.

Mean seasonal cycle (1996–1999) of selected variables averaged over all sea areas north of 70°N from ERA-40 reanalysis data and simulations of HIRHAM-NAOSIM. Observed mean sea-ice cover is based on SSM/I data instead of ERA-40. Fluxes are positive toward the ocean-ice surface.

[41] On the other hand, all simulations show an overestimation of net long-wave radiation in winter of about 20 W/m2 compared to ERA-40. This means that the net surface heat flux during winter is approximately 10–20 W/m2 too high in the coupled model. The outcomes of this are too-warm ice surface temperatures (and consequently too-warm 2-m air temperatures) in winter and an underestimated conductive heat flux through the ice with the result of too-low ice growth. The long-wave radiation bias is partly compensated by an opposed bias in sensible and latent heat flux, most notably in h2.0-alb, but with the consequence of the largest temperature bias.

[42] The overestimation of net long-wave radiation in winter is associated with an overestimation of cloud cover of about 10% compared to ERA-40 (Figure 10) but more than 20% compared to some satellite products [e.g., Schweiger et al., 1999]. Despite general difficulties in validating modeled wintertime Arctic cloud cover as pointed out by Wyser and Jones [2005], the overestimation of low-level clouds over the Arctic Ocean during winter is a well-known problem in HIRHAM and was already discussed by Rinke et al. [1997]. In the coupled model system, this bias reduces the thermodynamic growth of sea ice and leads to an increased winter temperature bias as a result of a positive feedback. While one might expect that the bias in cloud cover could also increase due to increased moisture flux from the surface in case of larger open-water areas, no such positive feedback is apparent in the simulations.

[43] During summer, the differences between the model simulations are rather low with respect to the atmospheric fluxes. Even the two simulations with the new albedo scheme show only minor deviations from the standard simulations. In particular, all simulations show a similar underestimation of net short-wave radiation of about 20 W/m2 and also an underestimation of 2-m air temperatures of 2–3 K during summer.

[44] While the reason for the underestimation of net short-wave radiation is still unclear (lower sea-ice cover and lower cloud cover in summer should actually be accompanied by an overestimation of net short-wave radiation), the underestimation of the 2-m air temperature can be explained by the fact that the model limits the ice surface temperature to values below the salinity-dependent freezing point of seawater as long as ice is present in the grid cell. This is indeed a rather rough assumption. On the one hand, the snow and ice surface temperature, at least, may be expected to rise up to the freezing point of freshwater, which is up to 2 K higher, since the ice crystal structure accommodates only negligibly small concentrations of salt. On the other hand, incoming atmospheric energy over open water can be used either to melt ice laterally or to warm up the mixed layer, even above the salinity-dependent freezing temperature, assuming the matter of fact that horizontal mixing does not occur instantaneously.

[45] Maykut and Perovich [1987] have argued that the possible elevation of the upper-ocean temperature above the freezing point depends on the lead width and can be up to 5 K for nearshore conditions. They have further noted that the lateral melt rate depends on lead width and ice thickness in a very complex manner. For central Arctic conditions, just about 20% of the total melt rate can be attributed to lateral melt. In the current setup of the coupled model, all atmospheric energy over open water is used to melt snow and ice, in fact independent from the present lead width and ice thickness. This oversimplification may lead to an overestimated ice and, in particular, snow melt in the model with the result of an overestimated magnitude of the ice-albedo feedback effect.

4.5. Snow Thickness

[46] Validation of the modeled snow thickness is pretty difficult because area-wide snow thickness measurements are not available yet. Anyway, the SHEBA project has also provided snow measurements [Perovich et al., 1999] which can be used for a rough validation of the modeled snow thickness. However, measurements of snow depth at SHEBA show high variability at small spatial scales as pointed out by Huwald et al. [2005] and are not necessarily representative for a larger area. Compared to the measurements at the “snow mainline” (Figure 11), all model simulations overestimate snow thickness in winter and spring, but the onset of snow melt coincides quite well with the observation (around the middle of May, corresponding to SHEBA days 490–500). While the observed snow melt continued till early July, the model clearly overestimates snow melt and does not show any snow already in the middle of June (approximately SHEBA day 530) when just half of the observed snow had disappeared.

Figure 11.

Mean snow thickness from measurements at the “snow mainline” of the SHEBA drifting ice camp in the Beaufort Sea during 1997–1998 [Perovich et al., 1999] and from simulations of HIRHAM-NAOSIM. Simulated snow thicknesses were interpolated from the model grid onto the respective position of the ice camp. Effective snow thickness refers to the ice covered part of the grid cell (excluding open-water areas). The term “SHEBA day” on the x axis corresponds to the day from the start of 1997. The time series were smoothed using a 7-day running mean.

[47] The model simulations differ about 2 weeks at most in the beginning and the end of the snow-melt period, but the length is almost equal and there is no clear indication that the new albedo scheme has significant impact on the snow-melt rate. However, the end of the snow-melt period coincides with the beginning of the ice-melt period (see Figure 5). The total amount of thermodynamic ice loss depends mainly on the date of disappearance of the snow cover. Too-early disappearance of the snow cover due to overestimated snow melt leads to increased decay of sea ice and vice versa. It may be supposed that the underestimation of ice concentration over the central Arctic Ocean during summer originates from an overestimated snow-melt rate in the coupled model and, consequently, from the simplicity of the thermodynamic ice scheme, which is reflected in equations (4) and (5).

5. Summary and Conclusions

[48] A pan-Arctic coupled regional AOI model was applied to gain insight into uncertain and sensitive Arctic process descriptions in a coupled model, which need to be improved in order to reproduce observed sea-ice conditions in a more realistic fashion. Particular attention was paid to the ability of the model to reproduce the Arctic sea-ice anomaly during summer 1998 and the associated atmospheric conditions.

[49] Because the boundary conditions at the interface of atmosphere and ocean-ice model are normally not perfect in a coupled model system, the ice volume arising in a simulation of the stand-alone ocean-ice model may differ considerably from that of a corresponding coupled model simulation. It has turned out that a spin-up time of about 6–10 years is needed to reach a quasi-stationary cyclic state of equilibrium in the coupled model if the initial ice conditions are far from this state. The coupled model’s steady state level is rather independent from the initial ice conditions but depends significantly on uncertain process descriptions affecting thermodynamic growth and decay of ice in the model.

[50] A quasi-realistic ice thickness distribution at the beginning of the melting period has been found to be a decisive precondition for the ability of the coupled model to reproduce observed summer sea-ice extent. In contrast to the stand-alone ocean-ice model, where the atmospheric conditions are prescribed, the coupled model has to calculate the atmospheric fluxes in response to the given ice conditions and vice versa. Owing to positive feedbacks arising from this interaction, the coupled model may be expected to be more sensitive to the ice thickness distribution than the stand-alone ocean-ice model.

[51] Some indications exist that those experiments of the coupled model, which show the best agreement in simulated ice extent with observations, also offer the most realistic ice thickness distribution. Nevertheless, it is not well defined what a really realistic ice thickness distribution is because ice thickness observations are only sparsely available. This uncertainty remains a fundamental issue in Arctic climate modeling in general and sea-ice modeling in particular and will make the development of improved parameterizations difficult.

[52] It has turned out that the widely used, but rather intuitively reasoned parameterization of lateral ice growth, which is especially reflected in an arbitrary reference thickness for lateral freezing (h0), provides an opportunity to tune a coupled model toward a quasi-realistic state of equilibrium in terms of the ice volume. Even though summer sea-ice retreat develops more realistically in this case, the simulation of winter ice concentrations and near-surface temperatures may be getting worse. The right choice of h0 is likely to depend on the model used, particularly on the parameterizations of atmospheric processes that determine the surface fluxes. Because of the prime importance of the areas of open water for the oceanic heat loss during the cold season, the parameterization of lateral freezing can be regarded as a key process for the ice growth at large. This conclusion might be universally valid even if using other ice models that include empirical parameters with different interpretations [e.g., Mellor and Kantha, 1989; Häkkinen and Mellor, 1992].

[53] In the particular model used in this study, a higher value of h0 has to be used in order to give reasonable ice growth in winter. The associated higher heat loss of the ocean due to larger open-water areas has to compensate for a systematic overestimation of net surface long-wave radiation of about 20 W/m2. The reason for this model bias might be an overestimation of low-level clouds over the Arctic Ocean. Uncertainties in the simulation of Arctic clouds are among the major problems in most regional [e.g., Maslanik et al., 2000; Mikolajewicz et al., 2005] and global models [Chen et al., 1995]. It has become apparent that they cannot only account for unrealistic summer ice decay but particularly also for unrealistic winter ice growth in a coupled AOI model.

[54] A change in a surface flux-related parameterization is able to affect both atmospheric circulation and sea-ice conditions, as demonstrated by the integration of a more sophisticated snow and ice albedo scheme into the coupled model. The ice albedo has direct impact on the ice-albedo feedback, which definitely represents one of the most dominant processes for thermodynamic loss of ice during summer. Even small changes in a coupled model’s ice albedo scheme may lead to significant changes in the simulation of summer sea ice due to this positive feedback. Hence any basic change of the ice albedo scheme requires readjusting the relationship between growth and decay of sea ice.

[55] Furthermore, there are indications that the magnitude of the ice-albedo feedback effect is overestimated in the coupled model associated with too-early disappearance of the snow cover. It is supposed that an elaborated subdivision of the incoming atmospheric energy into snow and ice melt from above, lateral ice melt, and mixed layer warming will be able to overcome this shortcoming. Such schemes have already been partly realized in ice models [e.g., Tremblay and Mysak, 1997] with specific intent to allow for more realistic sea-ice retreat during summer, but they are rather seldom in fully coupled dynamical AOI models.

[56] In order to achieve a realistic regional distribution of sea ice in late summer, it also requires that the coupled model reproduces the observed atmospheric circulation during the preceding summer months. Nevertheless, in contrast to the clear response of the sea-ice cover to the atmospheric circulation, the atmospheric response to incorrect sea-ice cover is not that definite. Unrealistic sea-ice cover, as a result of incorrect thermodynamic ice loss, may favor model deviations in atmospheric circulation, but these deviations can clearly differ in their strength, probably in consequence of regional feedbacks. Owing to the variety of processes involved in such regional feedbacks, it is hard to distinguish between cause and effect of model deviations in a coupled model system without systematic sensitivity experiments. Some of such experiments have been presented in this paper, but several further experiments, especially with respect to the cloud scheme and the treatment of snow and ice melt, are required to assess the importance of individual processes for the simulation of Arctic sea ice and to develop improved parameterizations for these processes.

[57] The results of this paper suggest that uncertain process descriptions for Arctic clouds, snow, and sea-ice albedo, and lateral freezing and melting of sea ice, including the treatment of snow, might also be the reason for the large deviations in the simulation of Arctic sea ice with global AOI models. A coupled AOI model responds definitely more sensitively to such uncertainties than a pure atmosphere or ocean-ice model due to the feedbacks arising between the components of the Arctic climate system. Given that the magnitude of the long-wave radiation bias over the Arctic Ocean even exceeds the estimated change in the radiative forcing up to the year 2100 from all emission scenarios [Cubasch et al., 2001], it becomes apparent that model biases in polar regions must be further reduced in order to enhance the credibility of future climate change projections.

Acknowledgments

[58] This work was funded by the German Federal Ministry for Education and Research (BMBF) project ACSYS II (BMBF grant 03PL034C) and the European Union project GLIMPSE (EU grant EVK2-CT-2002-00164). The model simulations were carried out on the parallel IBM-p690 computer system at the North German Supercomputing Center (HLRN) under project ID hbk00014. SSM/I sea-ice concentrations were obtained from the National Snow and Ice Data Center (NSIDC), Boulder, CO, SHEBA data were provided by the SHEBA Project Office, University of Washington, and ERA-40 data were provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). Finally, we would like to thank Jan Sedlacek and two anonymous reviewers for their helpful comments to improve the manuscript.

Ancillary