Numerical simulation of ice-ocean variability in the Barents Sea region
- First Online:
- Cite this article as:
- Budgell, W.P. Ocean Dynamics (2005) 55: 370. doi:10.1007/s10236-005-0008-3
- 235 Views
A dynamic–thermodynamic sea ice model has been coupled to a three-dimensional ocean general circulation model for the purpose of conducting ocean climate dynamical downscaling experiments for the Barents Sea region. To assess model performance and suitability for such an application, the coupled model has been used to conduct a hindcast for the period 1990–2002. A comparison with available observations shows that the model successfully tracks seasonal and inter-annual variability in the ocean temperature field and that the simulated horizontal and vertical distribution of temperature are in good agreement with observations. The model results follow the seasonal and inter-annual variability in sea ice cover in the region, with the exception that the model results show too much ice melting in the northern Barents Sea during summer. The spatial distribution of the winter simulated sea ice cover is in close agreement with observations. Modelled temperatures and ice concentrations in the central Barents Sea are biased too high and too low, respectively. The probable cause is too high inflow of Atlantic Water into the Barents. The seasonal and inter-annual fluctuations in temperature and sea ice cover in the central Barents are, however, in excellent agreement with observations. Salt release during the freezing process in the numerical simulation exhibits considerable inter-annual variability and tends to vary in an opposite manner to the net inflow volume flux at the western entrance of the Barents Sea. Overall, the model produces realistic ice-ocean seasonal and inter-annual variability and should prove to be a useful tool for dynamical downscaling applications.
KeywordsIce-ocean modelComparison model-dataBarents Sea
A regional coupled ice-ocean model will be used to study future climate change scenarios in the Barents Sea based on results from the global Bergen Climate Model (Furevik et al. 2003). However, before future regional climate regimes can be simulated, it is necessary to demonstrate that the downscaling model can reproduce current climate conditions with realistic seasonal and inter-annual variability, which is the main goal of this paper.
Section 2 provides a description of the coupled ice-ocean model together with the configuration of the large-area (basin-scale) and nested (regional-scale) models. The boundary and initial conditions are defined and the surface forcing fields used in the models are discussed. Section 3 contains the results of the validation of the regional model against observations. Comparisons of model results with satellite imagery, hydrographic sections and spatially integrated time series are provided. Section 4 offers examples of the results of ice–ocean interaction in the Barents Sea. Brine formation and meltwater production are illustrated and discussed. Finally, the paper concludes in Sect. 5 with a summary and discussion of results.
2 The ice-ocean model
Several model studies of circulation in the Barents Sea region have been carried out over the past decade or so, including those of (Ådlandsvik and Loeng 1991; Stolehansen and Slagstad 1991; Harms 1992; Loeng et al. 1997; Harms 1997; Ådlandsvik and Hansen 1998; Asplin et al. 1998; Ingvaldsen et al. 2004a; Maslowski et al. 2004). The current study is based on a new ice-ocean model system run at high spatial resolution for a multi-year simulation validated against available observations.
2.1 The ocean model component
The ocean model component is based on Regional Ocean Modelling System (ROMS) version 2.1. ROMS is a three-dimensional baroclinic general ocean model, the development of which is described in a series of papers (Song and Haidvogel 1994; Haidvogel and Beckmann 1999; Haidvogel et al. 2000; Shchepetkin and McWilliams 2003). ROMS uses a topography-following coordinate system in the vertical that permits enhanced resolution near the surface and bottom (Song and Haidvogel 1994). Orthogonal curvilinear coordinates are used in the horizontal. A spline expansion can be used for vertical discretization which provides for an improved representation of the baroclinic pressure gradient (Shchepetkin and McWilliams 2003), vertical advection and vertical diffusion of momentum and tracers. ROMS has been designed from the ground up to run efficiently in both distributed (MPI) and shared (OpenMP) memory parallel computing environments, thus enabling computationally intensive dynamical downscaling experiments to be conducted.
2.2 The ice model component
Large portions of the Barents Sea are ice-covered for much of the year. Thus, it is important to include the effects of ice drift, melting and freezing upon the ocean fields. To accomplish this, a dynamic–thermodynamic sea ice module has been developed and coupled to the ocean model. The ice dynamics are based upon an elastic-viscous-plastic (EVP) rheology after Hunke and Dukowicz (1997) and Hunke (2001). The EVP scheme is based on a time-splitting approach whereby short elastic time steps are used to regularize the solution when the ice exhibits nearly rigid behaviour. Because the time discretization uses explicit time-stepping, the ice dynamics are readily parallelizable and thus computationally efficient. Employing linearization of viscosities about ice velocities at every elastic (short) time step, as recommended by Hunke (2001), has the desirable property of maintaining the ice internal stress state on or in the plastic yield curve. That is, the ice deforms as a plastic material unless it is in a rigid state. Another desirable property of the Hunke (2001) linearization is that the EVP ice dynamics are found to provide a good transient response to rapidly varying winds as well as to inertial and tidal dynamics, particularly in the marginal ice zone.
The ice thermodynamics are based on those of Mellor and Kantha (1989) and Häkkinen and Mellor (1992). Two ice layers and a single snow layer are used in solving the heat conduction equation. The snow layer possesses no heat content, but is, in effect, an insulating layer. Surface melt ponds are included in the ice thermodynamics. A molecular sub-layer (Mellor et al. 1989) separates the bottom of the ice cover from the upper ocean. The inclusion of the molecular sub-layer was found to produce much more realistic freezing and melting rates than if the ice-ocean heat flux is based purely on the ice bottom–ocean upper layer temperature difference.
2.3 The large-area model
The large-area model (Fig. 2) is used to supply boundary and initial conditions to the regional Barents model. A stretched spherical coordinate grid (Bentsen et al. 1999) is used in the horizontal, with the North Pole situated in central Asia and the South Pole situated in the Pacific Ocean west of North America. In the Barents Sea region, the horizontal resolution is approximately 50 km. There were 30 generalized σ-coordinate (s) levels, stretched to increase vertical resolution near the surface and bottom.
A time step of 1,800 s was used for both the ocean internal mode and ice thermodynamic time step. A ratio of 40 was used between the ocean internal and external mode time steps. A ratio of 60 was used between ice thermodynamic and dynamic time steps.
No tides were included in the simulation. The vertical mixing scheme employed was the LMD (Large et al. 1994) parameterization. The LMD scheme was used for the large-area model because it has been found to produce good agreement with observed mixed-layer behaviour in the deep ocean (Large and Gent 1999). The lateral boundaries were closed with the exception that a constant volume flux of 1 Sv was input across Bering Strait and the same quantity was removed along the southern boundary.
The atmospheric forcing was obtained from the NCEP/NCAR reanalysis data (Kalnay et al. 1996). Daily mean wind stress, and latent, sensible, downward shortwave radiative and net longwave radiative heat fluxes were applied as surface forcing after correcting for differences in model and NCEP surface conditions, such as in surface temperature and ice concentration. The flux corrections applied were developed by Bentsen and Drange (2000) and provide a feedback between the model surface temperature and applied heat fluxes, thus minimizing problems with drift in model surface temperatures. Precipitation was taken from the daily mean NCEP values. Snowfall was taken to be precipitation, corrected for snow density, when air temperature was less than 0°C. Evaporation was computed from the latent heat flux.
River runoff was computed using the NCEP/NCAR Reanalysis daily accumulated surface runoff values over land that were routed to ocean discharge points using the Total Runoff Integrated Pathways (TRIP) approach of Oki and Sud (1998). The hydrographs were modified for areas north of 60°N to account for permafrost hydrology and storage in snow cover.
The model was started from rest at January 1, 1948 with the January mean temperature and salinity from the Polar Hydrographic Climatology (Steele et al. 2001). Initial sea ice concentration was obtained from the daily mean NCEP/NCAR Reanalysis sea ice concentration for January 1, 1948. Initial average ice thickness (ice volume per unit area averaged over each grid cell) was specified by multiplying the initial ice concentration by 2 m.
A simulation was then conducted to spin up the model over the period 1948 to the end of 1987. The model fields at 0000Z January 1, 1988 were then used as initial conditions for a hindcast conducted from January 1, 1948 to the end of 2003.
2.4 The Barents regional model
The regional model domain is shown in Fig. 1. The regional model uses the same horizontal coordinate system as the large-area model, but at higher spatial resolution. The horizontal grid size varies between 7.8 and 10.5 km, with an average of 9.3 km. In the vertical, 32 s-coordinate levels are used, with enhanced resolution near the surface and bottom.
The generic length scale (GLS) scheme (Warner et al. 2005) was used for subgrid-scale mixing of mass and momentum, with the two-equation k-kl model parameters. The k-kl model is a modified form of the Mellor-Yamada 2.5 closure (Mellor and Yamada 1982). The GLS k-kl scheme was selected for use in the Barents regional model since it was found to produce good results in coastal applications where tidal mixing is important (Warner and Geyer 2005).
The time step used in the simulation was 450 s for both the ocean internal mode and the ice thermodynamics. The ocean model is split mode explicit. The ratio of ocean internal to external mode time step is 45. The ratio of ice thermodynamic to dynamic time step is 60.
Flather (1976) and Chapman (1985) radiation open boundary conditions were prescribed for barotropic normal velocity components and the free surface, respectively. Flow relaxation scheme (Engedahl 1995) open boundary conditions were employed for three-dimensional velocity components and tracers. The regional model was forced at the boundaries with interpolated 5-day mean fields from the large-area model and with tidal velocities and free surface heights from eight constituents of the Arctic Ocean Tidal Inverse Model (AOTIM) by Padman and Erofeeva (2004).
Surface forcing for the regional model was the same as that applied in the large-area model simulation, but with the exception that the NCEP/NCAR Reanalysis cloud cover fraction was modified to provide the same monthly mean cloud cover climatology over the period 1983–2002 as the International Satellite Cloud Climatology Project (ISCCP) cloud cover data (Schiffer and Rossow 1985). This necessitated the modification of the downward shortwave radiation and net longwave radiation fluxes to be consistent with the new cloud data. In the first simulations of Barents Sea ice cover, we found excessive melting of sea ice in summer by shortwave radiation due to the NCEP/NCAR Reanalysis cloud cover fraction being too low by a factor of approximately 0.75. This problem was largely, but not entirely, remedied by the use of ISCCP cloud cover data.
Initial conditions for the Barents regional model were obtained from the archived 5-day mean large-area model fields interpolated to January 1, 1990. The Barents simulation was conducted for the period 1990–2002.
3 Model-data comparison
Model results are compared with data from a variety of sources, including satellite SST and passive microwave, hydrographic sections and time series of integral quantities.
3.1 Satellite SST
It can be seen that, in March, the model produces a realistic transport of warm water northward west of Spitzbergen in the West Spitzbergen Current and that the region encompassed by the 2° isotherm matches the satellite SST distribution, but the model results are ∼2° too cold in the middle of the Norwegian Sea near the left-hand (southern) boundary. The model results for September show good agreement with the satellite SST field.
3.2 Hydrographic sections
Given that the modelled fields are only averaged over a day and are compared to a (∼2 day) nearly synoptic hydrographic section, the agreement is remarkably good. The model results show good placement of the Polar Front along the Bjørnøya (northern) slope. The GLS mixing scheme produces mixed layers of the right thickness during the stratified late summer conditions. The modelled surface temperatures are too warm in the immediate vicinity of Bjørnøya at the northern end of the section.
3.3 Section mean time series
It can be seen that the model results track the seasonal and inter-annual variations in the Kola section temperatures very well.
Mean error (bias) and root-mean-square error with the bias removed for the Fugløya-Bjørnøya, Vardø North and Kola section mean temperatures based on the period 1991–2002
Mean error (°C)
RMSEu (°C) bias removed
3.4 Barents inflow
Over the period 1991–2002, the mean net volume flux of inflow to the Barents across the Fugløya-Bjørnøya section is computed to be 3.2 Sv (1 Sv=106 m3 s−1). Across the full Barents entrance between northern Norway and the southern tip of Spitzbergen, the net inflow over the period 1991–2002 is computed to be 3.6 Sv in this study. Maslowski et al. (2004) computed a value of 3.3 Sv net inflow across the same section in their study. From fixed current meter mooring arrays, Ingvaldsen et al. (2004b) estimated net Atlantic Water inflow to the Barents through the Fugløya-Bjørnøya section to be 1.5 Sv, averaged over the period August 1, 1997 through July 31, 2001. If Atlantic Water is taken to be water warmer than 3°C with salinity greater than 34.9, then on the same section, over the same period, a value of 2.5 Sv is obtained in this modelling study. The advective heat flux computed from current meter mooring temperature and currents for the same period is 35.4 TW, whereas the heat flux computed from model results using the same ‘sampling’ scheme as the observations is 80.4 TW.
Since the modelled temperatures at the Barents Opening (Fugløya-Bjørnøya) are unbiased, the excessive modelled heat flux must be due to too high inflow velocities. It is possible that the NCEP/NCAR Reanalysis daily mean wind stresses are too high. Renfrew et al. (2002) have noted that the roughness length formula employed in the computation of surface momentum flux in the NCEP reanalysis project is inappropriate for moderate to high wind speeds and that errors can be substantial in regions with large air–sea temperature differences and high wind speeds (such as the western Barents Sea and the Norwegian Sea). Such a formulation could produce an over-estimate of wind stress during storm events, precisely the conditions favourable to Barents inflow (R. Ingvaldsen, personal communication). This hypothesis is currently being tested through sensitivity analyses with a variety of atmospheric forcing fields.
3.5 Satellite ice concentration
4 Ice–ocean interaction
The distribution of total ice-melt for 1993, shown in Fig. 17, is very different from that of ice formation. While most ice-melting occurs in a region encompassing the summer ice edge in the northern Barents, discontinuous lines of high total ice-melt are distributed throughout the portion of the model domain which encounters ice cover. A sample time series of ice-melt rates at one location situation on a high ice-melt line is shown in Fig. 18. The melting is concentrated in a 5–8 day period in early June, with no other appreciable melt activity occurring over the rest of the year. The rapid melting event corresponds to the transport of the ice edge into surface waters previously warmed by shortwave radiation. The melting occurs largely through lateral ablation.
5 Concluding remarks
The main goals of this study are to assess model performance in the light of available observations and determine the model’s utility for dynamical downscaling applications. The agreement between simulated and observed temperatures is, in general, excellent. The notable exceptions are the bias in simulated temperatures at the Vardø North and Kola sections. The likely cause of this bias error is excessive inflow of warm Atlantic Water at the Barents western entrance.
The agreement between modelled and observed salinities is not as close as with temperature, particularly at the Fugløya-Bjørnøya section. There, the model produces a weaker, more saline coastal current. This error is largely attributable to the relatively low salinity Norwegian Coastal Current not being adequately spatially resolved at the southern open boundary with forcing from the 50 km resolution Large Area model. The more saline, less stratified model results near Bjørnøya may be caused by the simulated ice edge lying a few kilometres too far to the north.
The agreement between total simulated and observed sea ice cover during 10 months of the year is excellent. During summer, however, the model results show excessive melting of sea ice occurring in the northern Barents. While it was found that, as noted by Serreze et al. (1998), NCEP/NCAR Reanalysis cloud fraction is too low and that ISCCP cloud cover data is more realistic at high latitudes, it is clear from the Makshtas and Korsnes (2001) study that the ISCCP cloud fraction in the northern Barents marginal ice zone in summer is still too low. An improved parameterization for the effect of low cloud/fog over the Barents summer marginal ice zone on the radiation budget is required if the summer ice melting is to be reduced. The model results underestimate the ice cover in the central Barents region with a nearly constant offset. The negative bias is consistent with too much inflow of Atlantic Water into the Barents. The modelled ice cover in the central Barents does, however, closely follow the observed inter-annual fluctuations.
The computed net inflow volume flux across the Barents Opening is within 10% of that of the Maslowski et al. (2004) study, in which very different forcing fields were used. However, the computed Atlantic Water net inflow in this study is 1.67 times that estimated by Ingvaldsen et al. (2004b) from current meter moorings. This discrepancy could be largely attributable to uncertainty in the wind forcing fields in the western Barents/eastern Norwegian Sea. The sensitivity of the modelled Barents inflow to different forcing data sets is being examined in a study currently being conducted.
The coupled model can be used to examine the role of ice freezing and melting processes in water mass formation and modification. The spatial distribution of ice freezing and melting zones is quite different with freezing taking place largely over banks and in shallow regions and melting occurring in the summer marginal ice zone and at the ice edge. Lateral ablation at the ice edge is very episodic over 5–8 day period and along discontinuous lines. Brines released during the freezing process are found to drain from the banks in plumes, forming a brine-enhanced high-salinity layer in troughs and depressions. The salt released during freezing exhibits considerable inter-annual variability and tends to follow the opposite trend to the net inflow volume flux at the Barents western entrance.
Overall, the coupled model produces realistic seasonal and inter-annual ice-ocean variability. The discrepancies with observations can largely be accounted for by uncertainties in the prescribed forcing fields. Thus, the model is demonstrated to be a viable tool for ocean climate dynamical downscaling purposes.
This work was supported by the Research Council of Norway Regional Climate development under global warming (RegClim) programme. This work has received support through the Programme for Supercomputing of the Research Council of Norway through a grant of computing time. I wish to thank Randi Ingvaldsen for making available her processed hydrographic section data and sharing her insight into circulation processes in the western Barents. I also wish to thank Jens Debernard and Øyvind Sætra for making available their EVP ice dynamics code and to Sirpa Häkkinen for making available her ice thermodynamics code.