To evaluate the impact of ocean and sea ice coupling on sea ice and atmospheric forecasts in the medium range, three sets of 10 d forecast experiments (hereafter referred to as forecasts) were run with the ECMWF IFS for the period DJFM 2017/18. The experiments are initialised each day at 24:00 UTC during the period: one in which dynamic coupling with sea ice concentration and ocean is switched on (coup-SSTSIC), one atmosphere-only where sea ice concentration and SST anomalies are persisted from the initial time (pers-SSTSIC), and another atmosphere-only with updated observed sea ice concentration and SSTs (obs-SSTSIC). These were all run with Cycle 46r1 of the IFS at TCo1279 (∼ 9 km) resolution with 137 vertical levels, which is the current resolution of the operational high-resolution 10 d forecasts at ECMWF.

In the uncoupled forecasts (pers-SSTSIC and obs-SSTSIC), sea ice concentration and sea surface temperature (SST) from the Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA; Donlon et al., 2012) are prescribed at the surface. The sea ice fields in OSTIA are derived from the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) Ocean and Sea Ice Satellite Application Facility (OSI SAF) sea ice concentration product (Tonboe and Lavelle, 2016) but are interpolated onto the OSTIA grid and adjusted to make the ice concentration consistent with the OSTIA SSTs. In pers-SSTSIC sea ice concentration and the SST anomalies are persisted from the initial time to the end of the forecast, as was done in ECMWF high-resolution operational forecasts until the IFS Cycle 45r1 upgrade in June 2018 (the ECMWF Ensemble forecasts have included dynamic sea ice since the Cycle 43r1 upgrade in November 2016). In this experiment a persisted anomaly approach is used for the SSTs, where the anomaly at the initial time is added to the daily climatology appropriate for each forecast lead time. In the other uncoupled experiment, obs-SSTSIC, the SST and sea ice fields are updated daily throughout the forecast, again using OSTIA. This experiment allows one to assess the maximum potential benefit of correctly representing SSTs and sea ice for atmospheric forecast skill (assuming the evolution of the overlying atmosphere is consistent with the sea ice changes).

The OSTIA fields used in the ECMWF operational analysis (during the period used in this study) are updated daily at 12:00 UTC and fixed until the next day at the same time. The OSTIA analysis fields are based on observations collected during a 36 h window centred on 12:00 UTC the previous day (Donlon et al., 2012). As a result the sea ice concentration and SST fields used in the pers-SSTSIC and obs-SSTSIC runs are approximately 36 h old at the initial time.

For the coupled forecasts (coup-SSTSIC), the IFS atmosphere is coupled to NEMO (Madec, 2008) model version 3.4.1 and LIM2, using the ORCA025 horizontal grid (with a resolution of approximately ∼ 10 km in the Arctic) with 75 levels in the vertical. The IFS makes use of a single executable framework, using a sequential coupling procedure described in Mogensen et al. (2012). Operationally the only sea ice model variable that is coupled to the atmosphere is the sea ice cover. For the SSTs the atmosphere and ocean are fully coupled in the tropics at all lead times but only partially coupled in the extra-tropics to avoid SST biases that would degrade forecast skill. In these regions of partial coupling, during the first 4 d of the forecast, rather than the atmosphere seeing the actual SST field from the ocean model, SSTs from OSTIA are provided at the initial time. These are then updated by adding the SST tendencies from the ocean model onto the initial field.

In the coupled forecasts, the ocean and sea ice fields are initialised from the ECMWF OCEAN5 analysis (Zuo et al., 2019). OCEAN5 uses the same NEMO-LIM2 model used in the coupled forecasts, forced by atmospheric forecast fields at the surface, and assimilates observations using the NEMOVAR 3D-Var assimilation system. OSTIA sea ice concentration and SSTs are assimilated in its production in addition to other ocean data sources. The atmosphere of all three experiments is initialised from the ECMWF operational analysis.

Since the forecast experiments were performed for winter (DJFM), which is a period of ice expansion in the Northern Hemisphere, persisting sea ice (in the pers-SSTSIC runs) results in negative ice concentration (Fig. 1a) and extent (Fig. 2) biases which grow with lead time. The pattern of the bias is more complicated in the coup-SSTSIC (Fig. 1d–f). The bias is corrected, by construction, in the obs-SSTSIC runs (not shown) and improved relative to pers-SSTSIC in some regions in the coup-SSTSIC runs since the ice extent is free to increase (or decrease) in the coupled model.

In some regions the coup-SSTSIC forecasts exhibit a positive bias. This is particularly the case in the region north of Svalbard, along the Labrador Sea coast of North America and in the Sea of Okhotsk (Fig. 1), as can also be seen in the time series of Northern Hemisphere sea ice extent (Fig. 2a). The spatial pattern of the bias in the coupled forecasts is consistent across lead times, but the magnitude increases. This suggests that the pattern of the bias, with respect to OSI SAF, is present in the OCEAN5 analysis, i.e. already at the initial time of the forecast, and is inherited from biases in the background forecasts from the ice–ocean model, which advances the ice edge too rapidly, consistent with the findings of Tietsche et al. (2015) and further discussed in Sect. 3.2.

For objective scoring of the forecasts, we calculate the IIEE with respect to OSI SAF, which provides an integrated measure of forecast performance for the position of the sea ice edge in a given region. It provides a detailed picture of how accurate the forecasts are in predicting which grid boxes are covered with sea ice by taking misplacement of the ice as well as the total area of ice into account (Goessling et al., 2016). It was calculated for the Northern Hemisphere as a whole and also for the Nordic Seas (65–83∘ N, 20∘ W–60∘ E), western Atlantic/Labrador Sea region (55–70∘ N, 62–30∘ W) and the Sea of Okhotsk (45–62∘ N, 135–157∘ E). These regions are shown in the coloured boxes in Fig. 1c and f.

IIEE increases with lead time for both pers-SSTSIC and coup-SSTSIC, with pers-SSTSIC generally increasing more rapidly. The lead time at which coup-SSTSIC becomes more skilful than pers-SSTSIC, based on the IIEE, varies from region to region. In the Northern Hemisphere and Nordic Seas region, coup-SSTSIC shows an improvement early in the forecast. However, the coup-SSTSIC forecasts have a larger IIEE in the Labrador and North-East Atlantic region at all lead times and in the Sea of Okhotsk until day 6. If, and at what lead time, the coupled forecast becomes more skilful depends on both the size of the initial error in coup-SSTSIC and the rates of error growth of both the coupled model and a persistence forecast. All these factors are regionally dependent: for example, in the Sea of Okhotsk there is a large initial bias in ice extent which grows with lead time, leading to worse performance than persistence during the first 6 d. In contrast the coup-SSTSIC initial error in the Nordic Seas is smaller than pers-SSTSIC and biases are more modest, and so performance is better than persistence from day 0. The skill of a persistence forecast will also vary from region to region based on the local ice dynamics.

It is interesting to note the large initial (day 0) IIEE in both pers-SSTSIC and coup-SSTSIC experiments. Errors in the pers-SSTSIC forecast at day 0 are partly due to the lag in the time at which sea ice fields are available to the operational analysis (see Sect. 2.1 for explanation) and partly due to small differences between the OSI SAF fields, used for evaluation, and OSTIA fields, used as boundary conditions for the model, due to interpolation and adjustment for consistency with the SSTs used in the OSTIA product.

Initial errors in coup-SSTSIC are inherited from the OCEAN5 analysis (also noted by Zampieri et al., 2018) as indicated by the similarity between the coup-SSTSIC IIEE at day 0 and the IIEE of OCEAN5 with respect to OSI SAF (i.e. comparing the orange and red curves in Fig. 3). It is worth noting that the coup-SSTSIC day-0 error for the Northern Hemisphere, with respect to OSI SAF, is more than 50 % of the IIEE of the coupled forecast at day 9. It is also striking that the error in the analysis, expressed as the IIEE of the OCEAN5 fields, is larger than the difference in IIEE between the coup-SSTSIC and pers-SSTSIC forecasts (Fig. 3), which suggests that using OCEAN5 instead of OSI SAF for verification would dramatically change the outcome of the evaluation. The next section will focus on evaluation of OCEAN5 sea ice concentration fields.

Subjective evaluation of the sea ice edge position in the OSI SAF and ECMWF OCEAN5 analyses against True Color and Corrected Reflectance images from the Moderate Resolution Imaging Spectroradiometer (MODIS) for selected forecast dates confirms the day-0 differences seen in Figs. 1–3. In the MODIS Corrective Reflectance product (shown in Fig. 4b, d and f) snow and ice appear bright red. Thick ice and snow appear vivid red (or red-orange), while ice crystals in high-level clouds appear reddish-orange or peach. Open ocean appears dark, but small liquid water drops in clouds appear white. This strong contrast between the red (sea ice) and the dark (open ocean) shades in the images can be used to infer the position of the sea ice edge.

Overall, the OSI SAF sea ice edge is more consistent with the MODIS images than the OCEAN5 ones, which further justifies the use of OSI SAF as the reference data for the sea ice verification in this study. The largest difference between OSI SAF and OCEAN5 in the Nordic Seas is in the region north and west of Svalbard (Fig. 1d). A large area of open water is present in this region, and OCEAN5 overestimates the extent of the sea ice (see the region annotated “NS” in Fig. 4a), with the OSI SAF ice edge being a closer match to the MODIS image (i.e. comparing the line contours to the edge of the red ice-covered area in Fig. 4b). In the western Atlantic the differences between OSI SAF and OCEAN5 are largest in the regions adjacent to the east and west coast of Greenland, where OSI SAF is more extensive, stretching further south along both coastlines than in OCEAN5 (Fig. 4c and d). On 3 March 2018 there is a clear discrepancy along the eastern coast of Greenland for example (in the region annotated “EG” in Fig. 4c), where ice is completely missing in the OCEAN5 but present in MODIS and OSI SAF. There is a similar situation in the northern part of the Sea of Okhotsk (in the region indicated with the initials SO in Fig. 4e), where OCEAN5 is more extensive than the OSI SAF fields. The truth is harder to see for the bias west of Greenland, indicated by WG in Fig. 4c, due to the high level of cloud cover in the image shown and in others that were inspected on different dates that are not shown.

In Fig. 4 the OSI SAF fields from the previous day are plotted in orange to provide an indication of how large the difference one would expect from the OCEAN5 field simply due to a lag in the sea ice concentration (SIC) field used in the assimilation. The similarity between the OSI SAF fields from one day to the next and the difference between these and the OCEAN5 fields suggest that errors in the OCEAN5 are systematically biased in these regions, rather than due to a simple lag in the availability of sea ice concentration data for use in the production of the ocean analysis mentioned in Sect. 2.2.1. The fact that biases in these regions show pronounced growth with lead time (see Figs. 1 and 2) further supports the idea that these biases are inherited from the ice–ocean model via the background fields used in the assimilation process. This can be seen most clearly in the time series of sea ice extent for the Sea of Okhotsk where there is a systematic positive bias in the OCEAN5 analysis (red dots) compared to OSI SAF (black dots), which grows with lead time during the forecasts (orange lines; see Fig. 2d).

Another factor is the assimilation methodology: in the NEMOVAR 3D-Var used in OCEAN5, the minimisation of the cost function for sea ice is separate from temperature and salinity (Zuo et al., 2019), so increments in sea ice concentration will not necessarily be consistent with other fields. This may be a particular issue around the sea ice edge, which is quite poorly sampled in terms of in situ observations of temperature and salinity. Another potential avenue for development is the use of observational error estimates for sea ice concentration. These are provided by OSI SAF and OSTIA but are not currently utilised in the data assimilation process.

In this section we investigate whether the difference in skill between pers-SSTSIC and coup-SSTSIC is consistent across the period of study or is larger during specific episodes. Time series of the IIEE at day 3 (Fig. 5) and day 9 (Fig. S1 in the Supplement) show that in general the pers-SSTSIC forecasts are more variable in all regions at both lead times, except for the Sea of Okhotsk at day 3. However, the magnitude of the IIEE is highly variable from day to day, with the values of the IIEE varying by as much as an order of magnitude between the most and least skilful forecasts, depending on the region. This high variability in the error in pers-SSTSIC shows that persistence is at times a very good forecast but that at times it is very poor. Decomposing the IIEE of the pers-SSTSIC forecast into the absolute extent error (AEE) and misplacement error (ME) (following Goessling et al., 2016) allows us to investigate this further.

On average the AEE and ME contribute roughly 50 % each (Fig. S2 in the Supplement) during the period investigated (DJFM), although this ratio may depend on the season. However, further inspection of the time series reveals that situations where persistence is a poor forecast tend to be where the AEE, rather than ME, is large for the persistence forecast (Fig. S2), with the AEE explaining ca. two-thirds of the variance in the IIEE in each region. This shows that the reason that persistence forecasts are particularly poor in these episodes is because rapid expansion and contraction of the area covered with ice is missed.

These periods of rapid sea ice change are exactly the situations where one would expect the dynamic coupling to add the most value. This can be seen most intuitively from Fig. 2, which shows that pers-SSTSIC is a particularly poor forecast during periods of rapid extent change. During periods where the observed extent increases (decreases), the persistence forecast is biased low (high). As the sea ice in the coup-SSTSIC forecasts is dynamic, the model is able to follow these variations seen in the observations. As a result, the reduction in IIEE (IIEEcoup-SSTSIC - IIEEpers-SSTSIC) is largest during periods of rapid ice extent change. There is a significant correlation between IIEEcoup-SSTSIC - IIEEpers-SSTSIC and the magnitude of the observed change in ice extent between the initial and verification times in the Nordic, Labrador and Okhotsk seas at day 3 (r is -0.48, -0.47 and -0.35, respectively, all highly significant) and day 9 (r is -0.54, -0.62 and -0.61, respectively, all highly significant) (see Fig. 6). Indeed, even for lead times where IIEEcoup-SSTSIC is larger than IIEEpers-SSTSIC on average (e.g. day 3 in the Sea of Okhotsk – Fig. 4d), events with exceptionally large changes in extent are better forecasted by coup-SSTSIC (see e.g. Fig. 6e). This confirms quantitatively what one can see intuitively in Fig. 2.

Periods of rapid ice extent change involve either an advancing or a retreating sea ice edge. Splitting the forecast dates into terciles, based on the observed change in sea ice extent over the first 3 d of the forecast, and compositing the initial conditions demonstrate that mean north-easterly (south-westerly) flow is associated with advances (retreat) in sea ice cover in the Nordic Seas region (see Fig. 7a and b). During north-easterly (south-westerly) flow situations southward (northward) advection of the sea ice due to anomalous wind stress and antecedent conditions for thermodynamic growth (melt) leads to a positive 3 d change in sea ice concentration (Fig. 7c and d). Comparison of these composites with the forecast 3 d change in concentration shows that the forecast does not increase as rapidly as seen in observations (Fig. 7e and f). These behaviours can also be seen in the time series of sea ice extent in Fig. 2b, where it can be seen that sea ice extent in coup-SSTSIC does not increase as rapidly as OSI SAF during a period of rapid ice growth in late January 2018 or the subsequent period of rapid decline in early February 2018 in the Nordic Seas. A similar picture can be seen in the other sectors (not shown). Note that the larger IIEE improvement in ice advance cases compared to ice retreat cases likely reflects a positive bias in the rate of ice growth, which tends to favour performance during ice advance.

As air is advected across a boundary between distinct surface types, such as between sea ice and open water (as in the situation shown in Fig. 6), its properties are modified as it adjusts to the new set of boundary conditions (Oke, 1987). This modification begins at the surface and is propagated upwards through turbulent diffusion. The layer of air whose properties have been affected by the new surface is called an internal boundary layer, and its depth grows with distance downstream of the boundary between the media, known as the leading edge, which in our case is the edge of the sea ice. Here we will focus on variations in the sea ice edge due to their large spatial scale (and therefore importance to global NWP); however air-mass advection between ice-covered and open-water regions in the form of leads and polynyas within the ice is also known to strongly affect air-mass properties (Moore et al., 2002; Pinto and Curry, 1995).

Two such idealised situations are represented in schematics in Fig. 8a and b. Figure 8a shows a situation with off-ice flow, typically seen during marine cold-air outbreaks, where a cold polar air mass is advected across the sea ice edge and over the open ocean. Since this air is much colder and drier than conditions at the surface of the open ocean, upward sensible and latent heat fluxes are induced which act to warm and moisten the internal boundary layer (Renfrew and King, 2000; Spensberger and Spengler, 2021). Conversely, Fig. 8b shows an idealised on-ice-flow situation where a marine air mass is advected over the sea ice. Since this air mass is much warmer and more humid than conditions at the surface, anomalous downward sensible and latent heat fluxes are induced which act to cool and dry the internal boundary layer (e.g. Pithan et al., 2016, 2018).

This link between the position of the sea ice edge and the development of the downstream boundary layer is a potential source of error in forecasts where the sea ice is persisted from the initial analysis, as in pers-SSTSIC. Because the sea ice concentration is itself modified by anomalous surface wind stress and surface energy fluxes, during the situations described above, the position of the leading edge changes over time, influencing downstream boundary layer development. In particular, during persistent off-ice flows, the sea ice edge advances to cover more of the open ocean (see Fig. 8c). As a result, the polar air meets the open ocean further to the south, and the maximum in the turbulent fluxes is shifted further to the south. During persistent on-ice flows the opposite occurs, with the sea ice edge retreating to expose more open water (Fig. 8d). As a result the sea ice dynamics modify the point at which the air-mass transformation begins to occur. The impact that the ice dynamics can have on the internal boundary layer development and on the turbulent fluxes can be seen by comparing Fig. 8a and b, which show an idealised situation where the sea ice has been fixed, with Fig. 8c and d, respectively, which show the same situation but where the sea ice has been allowed to evolve dynamically. The difference in the position of the internal boundary layer is shown in faded hues.

The features described above can be clearly seen in the composites of mean temperature and specific humidity forecast errors during periods of ice advance (Figs. 8–10) and retreat (Figs. S3–S5 in the Supplement) in each of the three regions (shown in Fig. 1c) from the pers-SSTSIC forecast set during DJFM. The dates used in these composites are those for which the 3 d change in sea ice extent is in the top or bottom tercile, as used in Fig. 6.

The negative bias in ice concentration in pers-SSTSIC, during periods of ice advance, goes hand in hand with positive temperature and specific humidity bias around and to the south of the sea ice edge (Figs. 9a, d, g; 10a, d, g; and 11a, d, g). The coup-SSTSIC and obs-SSTSIC forecasts have higher sea ice concentrations than the pers-SSTSIC forecasts (Figs. 9b, c; 10b, c; and 11b, c), and, as a result, the turbulent heat flux (THF) is reduced in those regions. Because the area south of the sea ice edge is a local maximum of THF field and the ice edge is further south in these simulations compared to the uncoupled runs, the THF is increased somewhat to the south, resulting in a dipole in the heat flux field (i.e. dashed red contours to the south of the blue). This effect was also observed in studies with more dramatic changes in sea ice (Day et al., 2012; Deser et al., 2010).

During periods of ice advance obs-SSTSIC and coup-SSTSIC are cooler and generally drier than pers-SSTSIC in the region downstream of the sea ice edge (Figs. 8e, f, h, i; 9e, f, h, i; and 10e, f, h, i); this difference is larger where the difference in sea ice concentration is greatest but extends some ∼ 800 km downstream to the south of the region where the sea ice has changed. To put this into context, the maximum 925 hPa specific humidity difference relative to pers-SSTSIC is perhaps 0.2 gkg-1 northeast of Svalbard. At day 3 in that region the climatological humidity is about 2.5 gkg-1 and the RMSE is about 0.3 gkg-1 (not shown). So dynamic ice is important in terms of the 925 hPa specific humidity error (i.e. around 10 % of the total humidity).

This is consistent with previous modelling studies which have indicated sensitivity to the nature and position of the sea ice edge and marginal ice zone at similar distances downstream using large eddy simulation (LES) experiments (e.g. Gryschka et al., 2008; Liu et al., 2006), although clearly cooling of SST will also play a role. Note that near and downstream of the sea ice the coup-SSTSIC forecast is cooler and drier than obs-SSTSIC run due to having higher sea ice concentrations. These are at least partly associated with a positive bias in sea ice concentration in the Nordic sector (Fig. 1). Further, biases in sea ice in the coupled model SE of Greenland and in the Sea of Okhotsk already discussed go hand in hand with temperature and humidity biases (Figs. 8, 9, S3 and S4). This shows the potential for errors in the sea ice to influence the atmospheric fields.

Conversely, during periods of ice retreat in the Nordic Seas (which corresponds to on-ice flow), a positive bias in the uncoupled runs goes hand in hand with negative temperature and specific humidity biases over the sea ice (Fig. S3a, d and g). The coup-SSTSIC and obs-SSTSIC forecasts have lower sea ice concentrations than the uncoupled one, and as a result the THF is higher in those regions (Fig. S3b and c). Compared to the uncoupled forecast, the prescribed and coupled forecasts are warmer and more humid over the sea ice downstream of the changes in the sea ice edge (Fig. S3e, f, h and i). However, the response of the atmosphere and turbulent exchange to including the evolution of the sea ice is much more modest in these situations. This is consistent with the findings of Blackport et al. (2019), who argue that due to the orientation of the turbulent fluxes, the influence of the sea ice position on the atmosphere is much more modest during such situations. A similar picture can be seen in the Labrador Sea and Sea of Okhotsk (Figs. S4 and S5).