On the sensitivity of Antarctic sea ice model biases to atmospheric forcing uncertainties

Abstract

Although atmospheric reanalyses are an extremely valuable tool to study the climate of polar regions, they suffer from large uncertainties in these data-poor areas. In this work, we examine how Antarctic sea ice biases in an ocean-sea ice model are related to these forcing uncertainties. Three experiments are conducted in which the NEMO-LIM model is driven by different atmospheric forcing sets. The minimum ice extent, the ice motion and the ice thickness are sensitive to the reanalysis chosen to drive the model, while the wintertime ice extent and inner pack concentrations are barely affected. The analysis of sea ice concentration budgets allows identifying the processes leading to differences between the experiments, and also indicates that large and similar errors compared to observations are present in all three cases. Our assessment of the influence of forcing inaccuracies on the simulated Antarctic sea ice allows disentangling two types of model biases: the ones that can be reduced thanks to better atmospheric forcings, and those that would require improvements of the physics of the ice or ocean model.

Appendix: Ocean vertical resolution

As noted in Sect. 2.3, the vertical resolution of the ocean in the present model configuration is higher than in most previous studies. Usually, the ORCA1 grid has indeed been used with 46 layers, whose thicknesses range from 6 m at the surface to 20  at 100 m depth, and to 250 m for the bottommost layer. For comparison, at the same depths, the layer thicknesses in the eORCA1 grid with 75 levels are 1, 10 and 200 m. In this appendix, we examine the effects of this change. To this end, an additional experiment is conducted, named LVR. It is the exact equivalent of DFS except for its lower vertical resolution.

Fig. 13

Differences in September sea ice concentration (left) and thickness (right) between experiments LVR and DFS (LVR-DFS), averaged over 1985–2014

As seen in Fig. 1, changing the ocean model vertical resolution has a negligible effect on the summer Antarctic sea ice extent. Regional differences in ice concentration between experiments LVR and DFS actually exist during the melting season, but they are small and they compensate each other in the total sea ice extent. By contrast, the latter is clearly reduced during the expansion period, by up to 2 $\times$ 10$^6$ km$^2$, even though the ocean-sea ice model is used in forced mode. The strongest decrease in ice concentration takes place close to 130$^\circ$W (Fig. 13), a region where DFS already underestimated the sea ice area compared to observations. Furthermore, concentrations are lower inside the ice pack in LVR, but they remain nevertheless higher than satellite estimates.

In spite of a smaller extent, the sea ice volume in experiment LVR is larger than in DFS (Fig. 1). The difference peaks at 3 $\times$ 10$^3$ km$^3$ in October, and a fraction of it persists throughout the year. It is related to an ice thickening of up to a few tens of centimeters in LVR, in most regions but the southern and southwestern parts of the Weddell Sea, where a slight thinning occurs (Fig. 13).

These changes between the simulations with different vertical resolutions are explained by how the model handles new ice formation in open waters. In conditions of ice growth, the open water heat loss to the atmosphere is split into two parts (Barthélemy et al. 2016a). The first one is used to lower the temperature of the top oceanic grid cell to the freezing point. The remaining heat loss must then be compensated by latent heat released by seawater freezing, which is associated with expansion of ice in the open water fraction of the grid cell. The thickness of the first ocean layer being 1 m in DFS and 6 m in LVR, at equivalent ocean temperature, more energy is needed in LVR to cool the surface down to the freezing point, and less energy is left for sea ice formation in open waters. This is the reason why ice concentrations are lower in LVR than in DFS. The lesser contribution of thermodynamic processes to the expansion of sea ice in LVR is indeed visible in the winter ice concentration budget based on online model diagnostics (not shown).

Fig. 14

August ocean temperature profiles, averaged over 1985–2014 and south of 65$^\circ$S, in the WOA13 climatology (Locarnini et al. 2013) and in experiments DFS and LVR. The horizontal lines show the simulated mean mixed layer depths

Subsequently, convection due to surface cooling and brine rejection from ice growth will homogenize the winter mixed layer. Since a thicker ocean layer needed to be cooled in LVR to allow for ice formation in open waters, the temperature of the mixed layer will become lower in this simulation. This is visible in the mean August temperature profiles displayed in Fig. 14. Close to the surface and on average south of 65$^\circ$S, LVR is about 0.05 $^\circ$C colder than DFS, with a larger difference at greater depths. A lower sea surface temperature implies a smaller oceanic heat flux to the ice base. In August, just before the sea ice maximum, the mean heat flux is reduced by half in the low resolution case (not shown). Differences of several tens of W m$^{-2}$ occur along the ice edge, where the flux is the highest in DFS due to relatively warm waters. In turn, the decreased oceanic heat supply to the ice in LVR explains the increase in ice thickness in this experiment.

Finally, the simulated mixed layers are deeper in LVR, as a consequence of the surface cooling and of enhanced brine rejection linked to larger sea ice production. On average south of 65$^\circ$S, the difference between the two simulations amounts to 15 m. While LVR appears to be in better agreement with WOA13 temperature profiles, the increase in mixed layer depth enhances the deep bias already present in DFS with respect to an observation-based climatology (Pellichero et al. 2017).

Based on this understanding of the first order effects of a reduced vertical resolution, we can re-examine the spatial patterns of changes in Fig. 13. The fact that sea ice is not thicker in LVR in the southern parts of the Weddell Sea can be explained by ocean temperatures close to the freezing point in that area, implying that the process described above does not play an active role there. By contrast, subsurface temperatures are the highest in the Amundsen Sea. In this region, the vertical mixing induced by new ice production at low resolution causes a upward heat transport which offsets the initial cooling. Higher surface temperatures decrease strongly the sea ice concentration and prevent the ice thickening visible elsewhere.

Depending on the variable considered, the best agreement with observations is provided by either DFS or LVR. More importantly, these results demonstrate that changing the vertical resolution of a model could require adjusting the treatment of some physical processes as well. The issue is similar to that of the increase in the horizontal resolution of ocean models, in which the eddy parameterizations must be adapted when the simulations become eddy-resolving (e.g., Iovino et al. 2016). In our case, it is not realistic that only the 1 m surface layer of the ocean is associated with new sea ice formation in DFS. In reality, turbulent mixing continuously mixes the upper water layer, over a thickness larger than than 1 m, thereby coupling it with the freezing taking place at the surface. Even if vertical mixing is realistically simulated, connections between the surface and the second ocean layer (and the deeper ones) can only occur at the model time step frequency. The results described above suggest that a more advanced representation of the formation of new ice in open waters (e.g., Wilchinsky et al. 2015; Barthélemy et al. 2016b) might help reduce the modeled winter sea ice concentrations.