a. Atmospheric models

The first atmospheric model used in our experimentation is the Canadian operational model GEM (Côté et al. 1998). The GEM model is used exactly in its regional configuration with a horizontal grid spacing of 24 km (Bélair et al. 2000). The primitive hydrostatic equations are integrated on a variable-resolution grid. The grid spacing in the central uniform domain is 0.22° (∼24 km, 270 × 353 points) and increases uniformly outside the central region with a stretching factor of about 10% (total: 353 × 415 points). Figures 1a and 1c present the topography and the land–water mask used in the model (24 km) for the region of interest.

The second model, used to downscale at a higher resolution (4 km), is the MC2 (Benoit et al. 1997; Thomas et al. 1998). This model is a limited-area model (LAM) running on a polar stereographic grid. Figure 1b shows the entire domain of this LAM and also its topography (400 × 400 points). The surface fields of the 4-km configuration clearly have more details and give a better representation of the land–sea interface (Figs. 1c and 1d). The land–sea mask field uses a land–water representation from 0% (water only) to 100% (land only) instead of a binary one. This more continuous representation of the land–sea interface allows smoother horizontal gradients of physical fluxes and permits generation of physical patterns that can be better resolved by numerical methods. A binary mask representation can cause sharp gradients that are difficult to resolve and are not necessarily desirable in numerical models. With a fourth-order advection scheme such as the one used in our atmospheric models (semi-Lagrangian with cubic interpolation), these types of sharp gradients can cause a loss of mass and energy. A value of 80% means that 80% of the surface fluxes used come from the land scheme and 20% from the ocean model. The land–sea interface of the 4-km configuration (Fig. 1d) fits better with the coastlines compared to the 24-km configuration (Fig. 1c). At low resolution, the numerical land–sea interface is stretched through the coastlines.

Both models use the same unified physics package developed at Recherche en Prévision Numérique (RPN). This package consists of a comprehensive multioption set of physical process parameterizations. A detailed description can be found in Mailhot et al. (1998). The main components used by both configurations are summarized in Table 1. The 4-km resolution configuration allows a better representation of the local winds and physical fluxes while the advanced microphysical equations (Kong and Yau 1997) are used to simulate cloud-scale processes.

The 24- and 4-km models are run with time steps of 600 and 50 s, respectively, and 28 and 35 vertical levels, respectively. Analyses at a 24-km horizontal grid spacing were generated via a 3D variational data assimilation scheme (Gauthier et al. 1996) using the GEM model. These analyses were used as initial conditions by both the 24- and 4-km models. The boundary conditions used by the LAM came directly from the 24-km simulation. The results presented here are from 48-h forecasts starting at 0000 UTC 13 March 1997.

b. Ocean-ice model

The ocean-ice model is the Gulf of St. Lawrence (Fig. 2) model (Saucier et al. 2003). The gulf is treated as a semienclosed sea opened to the Atlantic Ocean through Cabot Strait and the Strait of Belle Isle. The model domain extends between open boundaries near Cabot Strait and the Strait of Belle Isle, and the upper limits of tidal influence near Montréal and at the head of the Saguenay Fjord. It runs on a 5-km horizontal resolution grid for the region between the outer straits and Ile d'Orléans and uses a one-dimensional model for the momentum transfer from Québec City to Montréal. The surface and bottom layers adjust to the local water level and depth, respectively. The vertical resolution is uniform at 5 m from the surface down to 300-m depth and then becomes uniform at 10 m below. The time step used by the model is 300 s. The GSL ocean model is coupled to a Multi-category Particle-In-Cell (McPIC) sea ice model from Flato (1994). The initial ice conditions were obtained from charts produced each day by the Canadian Ice Service (CIS) incorporating Radar Satellite (RADARSAT) images, aerial reconnaissance, and ship-delivered data. This coupled model is developed for daily experimental sea ice and ocean forecasting at the Canadian Department of Fisheries and Oceans.

The initial conditions for the ocean have been generated by running the coupled system from 10 to 12 March 1997. The initial conditions for temperature, salinity, and currents used to start the ocean model on 10 March came from a combination of observations and climatological data (Petrie et al. 1996). The spinup is produced only for the motion, which is enough to obtain a quasi-stationary regime. In March the water is near the freezing point from the surface until approximately 100 m. The ice analysis produced by the CIS is used as the initial ice condition. The GSL ocean model is run in a continuous mode from 10 to 12 March. The 24-km configuration of the atmospheric model is coupled for this initialization period. The atmospheric and ice conditions (for the sea ice model) have been reinitialized from the analysis after 24 and 48 h of ocean simulation, that is, for 11 and 12 March. Thus, this initialization period is a combination of three atmospheric and ice forecasts of 24 h each and of one ocean simulation of 72 h (10–13 March). The purpose is to generate the best possible ocean conditions for 13 March. All the simulations presented in this paper are started using ocean conditions obtained this way.

a. The coupler

The RPN coupler has been designed to transfer data between models without modifying their model codes too much. Since this tool is a connector modulus between many environmental models that are developed by different research centers and universities, it is important that the coupler is a general application, is easy to implement on different computer systems, and is computationally efficient.

Figure 3 schematically represents the two-way coupled atmosphere–ocean-ice model system where two processes work in opposite directions to allow simple and efficient linking between the models. The coupler is divided into two main components. The first is the communication package that allows the physical linkage between the models and the coupler. This Globally Organized System for Simulation Information Passing (GOSSIP) package (Lee and Vallin 2000) is a relatively short C code using the pipefile technique. Many options are included to ease the use of the coupler. The code is well documented and is now open source. The second component of the coupler is the package that transfers physical information from one grid to another by method of interpolation or aggregation. Many popular types of grids used in atmospheric science are already supported by the package. The main idea is to minimize the related modifications or dependencies in the codes of the numerical models that are being coupled together. In this way it becomes easier to support, in parallel, the evolution of many numerical models by many organizations that are collaborating in the coupled modeling activity. Only a couple of additional subroutines have to be included in the numerical models to allow direct connection with the coupler. Each model passes and receives the physical information from the coupler directly onto its numerical grid. Once the coupling mechanism is installed in a model, it is then easy to migrate it into future versions and use it with other various external models. For example, the same coupler allows our atmospheric models to communicate with land surface schemes, hydrological models, and wave models (e.g., Desjardins et al. 2000). The RPN coupler is very similar to the Ocean Atmosphere Sea Ice and Soil (OASIS) coupler (Valcke et al. 2003). The main differences are located in the techniques used to exchange the data. OASIS uses message passing interface (MPI) while the RPN coupler uses GOSSIP.

b. Relationship between the time–space strategy and physical aspects

The two-way interactive coupling between two different numerical models requires a certain level of approximation in the representation of the physical connections. To maximize the effects of the coupling between both models, an attempt was made to respect these three principles: 1) to conserve energy and mass, 2) to represent as much as possible all the physical details produced by the models, and 3) to reduce the possible numerical discrepancies. Different techniques can be used to couple many models together, but they do not necessarily satisfy these goals. Here are two different coupling scenarios that strive to meet them.

In the first scenario, the atmospheric model has a spatial grid resolution (24 km) that is coarser than the resolution of the ocean-ice model (5 km). The time step of the atmospheric model (600 s) is larger than the time step of the ocean model (300 s). Even if the average motion is faster in air than in water (∼100 times), the efficiency of the semi-Lagrangian advection scheme, together with the lower grid resolution, allows the atmospheric model (Courant number ∼ 0.8) to have a time step that is 2 times larger than the one required by the Eulerian scheme of the ocean model (Courant number ∼ 0.3). Figure 4 schematically presents this first case for heat flux exchanges only. The large arrows represent a time step for the atmospheric (up) and the ocean-ice model (down). The thin black arrows represent the different exchanges, and the numbers represent the order in which the different processes (steps) are executed. Given that the numerical grid of the ocean-ice model is at a finer resolution, all the information coming from the atmospheric model is interpolated by the coupler to the ocean-ice model grid. Conversely, the heat flux computed on the ocean grid is aggregated by the coupler onto the atmospheric model grid. For more information, see section 3a. To simplify the following description, the models are considered to communicate directly together instead of through the coupler. The atmospheric model passes the field of air temperature near the surface to the ocean model (step 1 in Fig. 4, thin black arrow), and the latter computes the heat flux on its own grid using the air temperature and the surface temperature. To keep the maximum amount of physical information, the heat fluxes are computed on the grid having the higher resolution (principles 1 and 2). It is more natural to aggregate energy than to aggregate ice. For example, the latent heat flux computed directly on a high-resolution grid of an ocean-ice model and aggregated onto a low-resolution atmospheric grid will not necessarily be equivalent to the flux computed on the low-resolution grid using aggregated ice and ocean fields. In an attempt to conserve energy (principle 1), the models need to exchange energy rather than variables. The atmospheric model receives the heat flux [Fc(0), step 2 in Fig. 4] and then makes its first time step. The ocean model receives the updated air temperature and can then make its first time step (steps 4, 5 in Fig. 4). After these time steps, the ocean model computes the heat flux field that will be used for the next atmospheric time step. The energy amount passed to the atmospheric model [Fc(1) in Fig. 4] is the amount used to complete the two oceanic time steps. With this technique, the coupling is accomplished between both models by a time and process splitting that reduces as much as possible the temporal and spatial approximations (principle 3).

In the second scenario, the atmospheric model has a spatial grid resolution (∼4 km) approximately equivalent to the ocean-ice model (∼5 km). The increased spatial resolution is associated with a decrease of the atmospheric model time step to 50 s. Thus the atmospheric model executes six time steps (50 s) for one oceanic time step (300 s). The coupling is at each oceanic time step. To simplify Fig. 5 and the following description, consider that the atmospheric model produces two time steps during one oceanic time step. The description of this second scenario is similar to the previous one (Fig. 4). The difference is that for the same real time the atmospheric model executes more time steps than the ocean model. The heat fluxes used by the intermediate time steps (between two coupling; steps 3 and 6 in Fig. 5) are always calculated on the ocean grid (principles 1 and 3). This will maintain complete consistency with the ocean model and the coupling time steps as well as consider the maximum physical information produced by the models. Steps 4 and 5 in Fig. 5 represent the connections, between both models, used to calculate the heat flux for the atmospheric model. In the beginning, the atmospheric model communicates with the ocean-ice model (steps 1 and 2 in Fig. 5) to pass the air temperature and to obtain the corresponding heat flux calculated from the initial air and surface temperatures: F_c(0) = σ[T_s(0) − T(0)], where σ is the transfer coefficient. Afterward, the atmospheric model makes its first time step (step 3 in Fig. 5). Even if it is not a coupling time step, the atmospheric model communicates with the ocean model to obtain an updated heat flux [F_c(1/2)] computed from the updated air temperature [T(1/2)]. The surface temperature used for the intermediate time step is unchanged: F_c(1/2) = σ[T_s(0) − T(1/2)]. With this technique, the fluxes are definitely calculated, not only with the maximum information available, but also exactly in the same manner for all the atmospheric time steps. Keeping the fluxes unchanged between two coupling time steps could induce numerical discrepancies or instabilities. In this case, the gridpoint spacings are approximately the same in both models; thus, calculation of the fluxes is permitted directly in the atmospheric model. The goal is to keep as close as possible the same setup as in the previous configuration (24 km) and thus facilitate the physical comparisons. The similarities in resolutions of both atmospheric and ocean-ice models require using the interpolation technique for all exchanges of physical fields.

In both coupling scenarios, the fields passed to the ocean-ice model are the surface temperature, humidity, precipitation, short- and longwave radiation coming from the atmosphere, and surface winds. The ocean model transfers the heat and humidity fluxes, the longwaves radiation coming from the surface, and the averaged ice coverage to the atmospheric model. For this experiment the surface wind stress transferred to the atmosphere is calculated within the atmospheric model using the ice coverage. In future experiments, the wind stress used by the atmospheric model will come directly from the ocean-ice model. A more consistent treatment will allow conservation of the momentum fluxes but could also improve the winds forecasted by the 24-km atmospheric configuration.

a. Observed sea ice patterns

Given that the ice coverage changes and moves very rapidly in only 24 h (13–14 March), it is very important for the validation to compare the real and the simulated ice coverage for a specific and common time. However, the different observation tools presently used to generate an analysis on the St. Lawrence produce only one analysis per day (Canadian Ice Service). These analyses combine observations acquired in different areas at different times of the day, making it difficult to examine the subdaily variability. As mentioned earlier, it takes approximately 3 days for RADARSAT to cover the simulation domain entirely, and the flight and ship observations are not necessarily valid at the same time. It is possible to obtain a complete picture of the sea ice cover on the St. Lawrence valid for a specific time by using a visible satellite image. The challenge with the visible image is to be able to distinguish between clouds and sea ice or snow. Fortunately, a satellite picture without clouds over the major part of the ocean grid (valid at 2000 UTC 14 March 1997) corresponded to 42 h into the forecast of the simulation. It is, however, difficult to distinguish between snow/sea ice or open water using the infrared picture given approximately the same surface temperature for both water phases (melting period, ∼0°C). It is easier, given the large difference of temperature between clouds and snow/sea ice, to use the infrared picture to produce a cloud mask for the visible picture. This deduced mask has been superimposed on the visible picture (red regions in Fig. 7a). Only the mid- to high-level clouds have been masked. The infrared picture allows us to clearly detect the regions free of mid- to high-level clouds. Most of the gulf and estuary of the St. Lawrence are free of low-level clouds except for the Magdalen Islands and Cape Breton, where thin layers of low-level clouds are noticed.

With this combined infrared/visible picture, it is now easier to see the regions of the St. Lawrence estuary where a direct comparison with the observed and the simulated ice is possible. Efforts were made to examine the position of the sea ice margin. Nevertheless, this comparison is a very good indicator of the quality of the forecast, given the rapid change in the ice coverage during these 42 h of simulation. The red lines delineate, in Fig. 7a, eight segments (A–H) of the sea ice margin identified in the satellite image. These will be used for validation.

b. Differences between the one-way and two-way coupling

As discussed in the previous section, Fig. 7a is the observed ice field valid at 2000 UTC 14 March 1997. Figures 7b and 7c show the 42-h forecasts of simulated ice cover valid at the same time. Figure 7b shows the results from one-way coupling compared to those from two-way coupling in Fig. 7c. The integrations for the one-way version have exactly the same configuration as those for the two-way coupling except for the physical information received by the atmospheric model. On the first time step, the atmospheric model receives the physical information related with the initial ocean-ice conditions. With the one-way coupling these related ocean physical fields (water and ice temperature, ice thickness, and ice coverage) remain invariant for the rest of the simulation, while the ocean model continues to receive the updated atmospheric conditions at each time step. In the one-way coupling the ocean model does not influence the atmospheric model. This is very similar to the operational configuration of the Canadian atmospheric forecast model. The experiments presented here have been produced with the higher-resolution configuration of the atmospheric model. In Fig. 7, the dark blue areas represent water (0%) and the other colors indicate ice coverage (light blue: 20%–40%; green: 41%–60%; gray: 61%–80%; white: 81%–100%). The model has reproduced the segment labeled A (Fig. 7c), which is generated by the stronger wind regime due to the large-scale wind channeling and funneling effect. As the melting is not so efficient in the one-way coupling case (segment A, Fig. 7b), the ice edge does not seem to propagate southward rapidly enough compared with the observations (Fig. 7a) and the two-way experiment (Fig. 7c). The “V” shape (segment A) is also better reproduced by the interactive experiment. But, in general, both configurations have relatively well reproduced the observed segment A. The right position of the ice edge after 42 h of forecast indicates a very good approximation of the real surface wind produced by the atmospheric model. The segments C, D, and E clearly demonstrate the advantages of the two-way simulation. The larger ice-free areas generated between Québec and New Brunswick (D and E) have clearly been underestimated by the one-way coupling (Fig. 7b). The ice edges from the two-way simulation (red lines D, E in Fig. 7c) correspond better with the observations (red lines D, E in Fig. 7a), although the eastward advection of the two ice margin segments D and E is still underestimated. The same statement can be made about the ice segment C located along the western Newfoundland coast (red line C, Fig 7a). The ice coverage produced by the one-way coupling between Anticosti Island and Newfoundland is clearly more extensive than in the two-way simulation. The supplementary energy input in the two-way simulation gives a better reproduction of the shape and the location of the segment C (wave segment, Fig. 7c).

Another very interesting pattern allows validation of the two-way simulation. Good prediction of pattern C is essential to simulate well the trajectory represented by pattern B. In Fig. 7a, the red arrow (pattern B) represents the trajectory followed by sea ice floes along the north shore of Anticosti Island and exiting toward Cabot Strait. An animation of the ice field shows that the ice floes coming from the coast of Québec are accumulated on the north coast of the island and follow the coast in a southward direction until they move into the Laurentian Channel. The observed dimension, location, and amount of ice floes along this trajectory are very similar to those produced by the two-way simulation (Fig. 7c). It is more difficult to find the correspondence in the one-way coupling (Fig. 7b). As the ice located between the island and Newfoundland does not melt rapidly enough (without any two-way interactions, pattern C), the floes accumulated on the north coast and at the end of the island are larger than observed. In the one-way coupling, the ice coverage is too extensive (pattern C) and the floes cannot follow exactly the same trajectory with the same speeds as simulated by the two-way coupling. The floes cannot easily pass to the south of the island. The ice trajectory in the one-way coupling is thus shorter (pattern B) than in the two-way simulation, which seems to adopt an ice-edge shape more similar to the observed ones.

All the other patterns (F to H) have been relatively well reproduced by both coupled simulations. The fact that these patterns are relatively small has reduced the thermodynamic effects of the two-way coupling. These small openings in the ice field have essentially been produced by advection. It is, however, somewhat difficult to compare the pattern G produced near the Magdalen Islands, given the presence of low-level clouds. These clouds are noticed in the infrared picture (section 5a). By studying the ice thickness field (not shown here), it can be seen that thicker ice particles have been moved farther south in the two-way coupled run than in the one-way coupled run. The supplementary energy input obtained by the two-way coupling seems to generate a more realistic reduction in the simulated ice coverage and have an important impact near the ice edge.

Figure 8 shows the ice volume difference between the two-way and the one-way coupled runs as a function of ice coverage categories. Positive difference means that ice production is also positive while negative values indicate a loss in the ice field mass. The solid line represents this difference due to the coupling by expressing it as a percentage of the volume of each category. Note that the effect of the two-way coupling is to reduce the ice volume of the full coverage region (−10% for category 90%–100%) and to increase the other categories (+20% for category 50%–60%). In allowing a reduction of the ice thickness, the simulation with two-way coupling increases the regions with ice coverage between 0% and 75% and decreases the regions between 76% and 100%. This effect results in a thinner ice pack near the coasts with the two-way coupling.

In summary, the two-way coupling seems to have a positive impact on the prediction of the ocean-ice simulations. It is warmer than the one-way simulation that causes a reduction of 8% in the total ice volume when compared to the resulting ice from the one-way simulation. The following sections explore the impact of the coupling on the atmospheric aspect.

c. Observed and simulated clouds patterns

As seen in the previous section, the two-way coupling seems to improve the sea ice distribution forecast. This result was essentially associated with higher temperature due to a reduced sea ice cover seen by the atmospheric model during the simulation. The details of the impact on the atmospheric conditions are presented using a dynamical sea ice cover. The first atmospheric evaluation is rather objective and tries to indirectly evaluate the physical fluxes at the base of the atmospheric model. The low-level clouds generated directly over ice and water have been used to produce this evaluation. Clouds directly affect the energy reaching the surface and thus have an important impact on the predicted temperature, precipitation, and ice. Figure 9a shows a low-level cloud cover obtained from a satellite image (AVHRR). It is presented as a satellite view angle of the middle southern area of Fig. 7 rotated counterclockwise by 90°. The image is valid at 1230 UTC 14 March 1997. Dark blue areas represent ice-free water, gray ones correspond to ice, and yellow ones to land, while blue and white colors represent the low-level clouds. The geographical contours (white lines) are shown to better distinguish the coastlines of New Brunswick, Prince Edward Island, and the Magdalen Islands, which are covered by scattered low-level clouds. The red line represents the location of the cloud edge. The arrows (A, B, and C) indicate the three main trajectories for the surface winds. The gravity waves produced by Cape Breton Island and the Magdalen Islands are clearly visible in the clouds.

The physical patterns presented in Fig. 10 allow a good understanding of the mechanisms producing these low-level cloud patterns. This surface flow has been produced by the high-resolution (4 km) two-way system and has the same valid time as the satellite image. The black arrows represent the surface winds (knots), and the shaded (white contours) patterns represent the surface temperatures (degrees Celsius). The three dashed arrows represent the main trajectories followed by the surface flow and correspond to the arrows presented on the satellite image (Fig. 9a). The air parcels following the trajectory A came from the Québec lower north shore and are deviated by the Chic Chocs Mountains (∼1100 m; see Fig. 1b). They have an important interaction with the large ice-free area extending from the area west of Anticosti Island to the northwest of the Magdalen Islands clearly visible in Fig. 7a. Note that the cold air (−6°, −4°C contours, Fig. 10) moved over this large ice-free water area and gained heat and moisture (by heat and latent flux). It then reached the condensation threshold to produce the clouds clearly visible on the satellite image (Fig. 9a). The air parcels followed the trajectory B pass over the Chic Chocs Mountains before reaching the domain of interest. They are affected by a cooling and a drying (precipitation) during the ascending portion and are heated (adiabatic) on the downslope portion of their trajectory. They arrived at the ocean warmer and dryer than those following the previous trajectory (A). The pattern of temperature corresponding with trajectory B is clearly warmer (−4°, −2°C, Fig. 10). This dryer (Chic Chocs chinook) effect explains the region without cloud observed around trajectory B of the satellite image (Fig. 9a). The air parcels of the trajectory C originated directly from land (red portion of the arrow C, New Brunswick) with conditions relatively cold (−6°C, Fig. 10) and humid (∼91% RH). By passing over the ice-free waters (blue portion of the arrow C), the air parcels increased their temperature and specific humidity by sensible and latent heat flux. These same parcels continued their trajectory over the ice (green portion of the arrow, Fig. 9a), and on contact with this colder surface, they transferred a part of this new energy to the ice. This transfer changed the energy budget of the ice coverage that allowed a reduction of the ice thickness. The air parcels near the surface rapidly reached saturation (∼109%) and then generated the low-level cloud coverage located between the Magdalen Islands and Prince Edward Island (Fig. 9a). Figures 9b and 9c show the clouds and ice simulated for one-way and two-way coupling (4-km configuration) valid at the same time as the satellite image (Fig. 9a). The contours of simulated ice coverage shown are 0% (water: dark blue),>20% (light gray), and <20% (gray). The observed low-level clouds over the ice and on Prince Edward Island (red line in Fig. 9a) are relatively well reproduced by the simulation with two-way coupling (red line in Fig. 9c), while the simulation with one-way coupling is unable to generate them since it used an invariant ice field with no real ice-free area (see Fig. 6b). The shape of the simulated pattern is very similar to the observed pattern. These low-level clouds are caused by the presence of open waters during the simulation. The cloud and no-cloud patterns described by the three trajectories on the satellite image are relatively well reproduced by the atmospheric model (Fig. 9b). The same positive impact of the two-way coupling is also visible in comparing the clouds generated by the 24-km simulations (not shown here).

It is also very interesting to compare smaller features such as the observed limit of cloud formation captured by the satellite image with the model outputs. The satellite image (Fig. 9a) shows clearly that cloud forms very rapidly over the ice. The location of the cloud edge (red line, near the green arrow C) corresponds approximately to the location of the ice edge. This edge is shifted downstream in the two-way simulation (Fig. 9c). The numerical physical exchanges (fluxes) seem to be less efficient compared to reality. A second interesting feature to compare is the small line of ice produced near the coast of New Brunswick (letter G in Fig. 9c). This line can be relatively well distinguished on the satellite image (letter G in Fig. 9a). These small features bring complementary information that will be very useful in future studies.

This validation of low-level clouds demonstrates that the two-way coupled simulation has produced physical exchanges (fluxes) that correspond better with reality.

d. Observed and simulated temperature and precipitation