Modeling Antarctic Sea Ice Variability Using a Brittle Rheology

Abstract

Sea ice is a composite solid material that sustains large fracture features at scales from meters to kilometres. These fractures can play an important role in coupled atmosphere-ocean processes. To model these features, brittle sea ice physics, via the Brittle-Bingham-Maxwell (BBM) rheology, has been implemented in the Lagrangian neXt generation Sea Ice Model (neXtSIM). In Arctic-only simulations, the BBM rheology has shown a capacity to represent observationally consistent sea ice fracture patterns and breakup across a wide range of time and length scales. Still, it has not been tested whether this approach is suitable for the modeling of Antarctic sea ice, which is thinner and more seasonal compared to Arctic sea ice, and whether the ability to reproduce sea ice fractures has an impact on simulating Antarctic sea ice properties. Here, we introduce a new 50-km grid-spacing Antarctic configuration of neXtSIM, neXtSIM-Ant, using the BBM rheology. We evaluate this simulation against observations of sea ice extent, drift, and thickness and compare it with identically-forced neXtSIM simulations that use the standard modified Elastic-Visco-Plastic (mEVP) rheology. In general, using BBM results in thicker sea ice and an improved correlation of sea ice drift with observations than mEVP. We suggest that this is related to short-duration breakup events caused by Antarctic storms that are not well-simulated in the viscous-plastic model.

Key Points

• We introduce neXtSIM-Ant, the first Antarctic configuration with a sea ice model using a brittle rheology

• Sea ice properties in neXtSIM-Ant are consistent with observations of extent, thickness and drift

• Drift variability is better captured using the brittle rheology compared to a standard visco-plastic rheology at low resolution

Plain Language Summary

We studied how a new model configuration for Antarctic sea ice evolution, called neXtSIM-Ant can simulate sea ice fracturing in comparison to previous models. We tested this model with real-world observations and found that it does a better job of predicting where and how the ice will move and break apart, especially during storms. This new configuration could help improve our understanding of sea ice variability and future climate simulations.

1 Introduction

Sea ice in the Southern Ocean plays important roles in the carbon cycle evolution, Antarctic deep water formation, and in the Southern Hemisphere climate. However, contemporary global climate models (GCMs) struggle to represent Antarctic sea ice mean state and variability (Eayrs et al., 2021; Roach et al., 2018, 2020), whereas Arctic sea ice state simulations have improved in recent years (Horvat, 2021; Notz & SIMIP Community, 2020). These differences in the quality of sea ice representation may result from atmospheric forcing and/or differences in the sea ice properties. Focusing on ice properties, Southern Ocean sea ice is more seasonal, thinner, and comprised of smaller floes (Maksym, 2019; Toyota et al., 2016; Wadhams, 2000) than its Arctic counterpart. Antarctic ice pack has a land boundary in the poleward direction which strongly affects sea ice stresses. The Southern Ocean sea ice is also more saline, and more frequently includes formation from ocean water flooding snow layers and this affects ice strength (Eayrs et al., 2019). In addition, the Southern Ocean is more exposed to higher wind speeds from frequent low-pressure systems (Maksym, 2019; Vichi et al., 2019) and high ocean surface wave activity (Alberello et al., 2022; Horvat et al., 2020; Young et al., 2020). Those processes can lead to ice fracture and the generation of linear kinematic features (LKFs, Kwok et al., 1998), which affect the response of sea ice to stresses and atmosphere-ocean-ice interactions (Brenner et al., 2023; Kohout et al., 2014; Vichi et al., 2019). Moreover, these fractures/LKFs may contribute to the feedback between storms and sea ice which is critical for sea ice evolution in both hemispheres (Blanchard-Wrigglesworth et al., 2022; Boisvert et al., 2016; Horvat, 2022; Kohout et al., 2014).

The representation of brittle fracture and LKFs in Arctic sea ice simulations has been an active topic in recent years. Hutter et al. (2022) and Bouchat et al. (2022) have shown that models using the common visco-plastic rheology (used in all GCMs) fail to reproduce LKFs unless they run at small grid spacing ($${< } $$5 km) that is not feasible in the current era of global climate simulation. At coarser resolutions, the neXt generation Sea Ice Model (neXtSIM, Rampal et al., 2016) that uses a Brittle rheology (e.g., the Bingham-Maxwell (BBM, Ólason et al., 2022)) and a Lagrangian sea ice mesh was able to accurately capture Arctic sea ice fracture dynamics across multiple orders of magnitude in space and time (Bouchat et al., 2022; Hutter et al., 2022; Ólason et al., 2022; Rampal et al., 2019). This accuracy is achieved in this type of rheology by using a scalar parameter, called damage, that corresponds to the density of fractures at the subgrid-scale and represents the effect of fracture formation on sea ice mechanical strength. Combined with the Lagrangian framework to limit numerical diffusion and conserve the high gradients associated with LKFs, the model can represent the effects of sea ice breakup (i.e., LKFs) from grid to basin scale.

Improvements to brittle rheologies in neXtSIM through the Scale Aware Sea Ice Project have enabled decadal simulations and the reproduction of Arctic sea ice mass balance consistent with observations (Boutin et al., 2023; Ólason et al., 2022, 2025; Regan et al., 2023), and have illustrated the importance of small-scale dynamics on the winter ice production in the Arctic (Boutin et al., 2023). This has motivated the implementation of the BBM rheology in sea ice models other than neXtSIM (Brodeau et al., 2024). However, it remains unclear how using brittle rheologies, such as the BBM, would affect the representation of Antarctic sea ice given the differences between sea ice properties in the two hemispheres. In particular, the representation of sea ice deformation or drift in an environment dominated by seasonal, thin and mobile ice. It is also unclear how brittle rheology parameters tuned against Arctic observations (Boutin et al., 2023; Ólason et al., 2022) would perform in the Southern Ocean.

In this study, we present a new model configuration, neXtSIM-Ant, for simulating Antarctic sea ice using neXtSIM with the BBM rheology (Figure 1) over the period between 2015 and 2021. We describe this model in Section 2, and assess the state of sea ice simulated in neXtSIM-Ant in comparison to an identically-forced integration with the elastic-viscous-plastic rheology in Section 3. Finally, we discuss new avenues to study Antarctic sea ice and coupled processes using a brittle rheology in Section 4.

Figure 1

Maps of sea ice concentration from the mEVP (left) and BBM (right) runs at 21hr 05/07/2016 UTC. The different segments of the Southern Ocean are highlighted in magenta on the left plot where AB represents the Amundsen and Bellingshausen Seas. On the right, the black contours represent surface atmospheric pressure shown at every 10 hPa and the magenta circle highlights the center of a low-pressure system on the Weddel Sea. The mEVP and BBM runs are forced by the same atmospheric reanalysis (ERA5).

2 Methods

2.1 The neXtSIM Antarctic Configuration

We use the stand-alone version of the neXtSIMv2 model based on Ólason et al. (2022), meaning that our simulations are forced by atmospheric and oceanic reanalyzes (see below). neXtSIM is a Lagrangian sea ice model that allows for the generation and propagation of multi-scale ice fracture features through the ice pack, from the model grid scale to basin scale (Arctic Ocean) (Bouchat et al., 2022; Hutter et al., 2022; Ólason et al., 2022; Rampal et al., 2019). It also provides different options for the sea ice rheology to test the impact of including brittle physics.

In this study, we run Antarctic sea ice simulations using neXtSIM with the Brittle Bingham Maxwell rheology (BBM run) and the modified Elasto-Visco-Plastic rheology (mEVP run) following the same approach conducted by Ólason et al. (2022). The mEVP rheology (Bouillon et al., 2009; Kimmritz et al., 2015, 2016) is based on the same physical principles as the visco-plastic rheology suggested by Hibler (1979) for simulating the aggregate behavior of large floes on climate length scales (Toyota et al., 2022) but with improved numerical convergence. The mEVP rheology, or related EVP-style physics, is used by most current-generation modeling centers (Notz & SIMIP Community, 2020), including in regional configurations (e.g., Nazarenko & Tausnev, 2001). Thus, we use the mEVP simulations as a reference here, to identify the potential advantages and drawbacks of using BBM to simulate Antarctic sea ice.

The model bathymetry is ETOPO1 with 1 arc-minute of horizontal resolution. We applied additional corrections to the coastline to account for the Antarctic ice shelves. neXtSIM uses a moving mesh with an average horizontal resolution of 50 km. This average resolution is maintained during the simulation by remeshing in a physically consistent manner (Samaké et al., 2017), when any angle of a triangular element of the mesh is smaller than $${10}^{o}$$ (Rampal et al., 2016). This remeshing generally occurs a few times per simulated day, and its cost remains low at 50 km but increases with the resolution. The results we present in this study are remapped to a fixed polar stereographic with a grid spacing of 25 km for visualization and comparison against observations. Atmospheric forcing was taken from the hourly ERA5 reanalysis at 1/$${4}^{\mathrm{o}}$$ horizontal resolution (Hersbach et al., 2020).

Oceanic forcing (sea surface currents, height, temperature and salinity) are taken from the Biogeochemical Southern Ocean State Estimate (BSOSE, Mazloff et al., 2010, 2023). BSOSE is an ocean–sea ice reanalysis also forced by ERA5 atmospheric boundary conditions. We use the latest version of this reanalysis (Iteration 139, http://sose.ucsd.edu/SO6/ITER139/, last visited on 5 July 2024), available as a gridded data set at 1/$${6}^{o}$$ horizontal resolution, and covering the period between 2013 and 2021 as 5-day averages. We initialize sea ice thickness and concentration using BSOSE while setting the initial snow thickness to zero. All simulations are initialized on the 6th of January 2013 and ended on the 29th of December 2021 to match the period of BSOSE data availability. To give a rough idea of the computational efficiency of the model, it took 14 hr (1.5 hr/year) using 40 Intel(R) Xeon(R) Gold 6230 CPUs at 2.10 GHz without hyperthreading to run the 9 years of simulations. The first 2 years of all simulations are discarded to account for model spin-up period. BSOSE is based on MITgcm and it is a dynamically-consistent ocean-sea ice reanalysis based on the 4-dimensional variational (4D-Var) data assimilation method. We include BSOSE sea ice properties in our analysis as a reference of how a standard, Eulerian sea ice model evolves under the same atmospheric and ocean forcing. The thermodynamics between BSOSE and neXtSIM also differ. BSOSE's thermodynamic ice model is a variation of the 0-layer thermodynamics of Semtner (1976) (Fenty & Heimbach, 2013). neXtSIM includes two ice classes with different thermodynamic models. The first, young ice, corresponds to ice growing in open water and thinner than $${\sim} $$30 cm which is similar to BSOSE and uses the 0-layer thermodynamics. The second class represents consolidated ice after it has grown thicker than $${\sim} $$30 cm. Its thermodynamics are modeled using the 3-layer model described in Winton (2000).

2.2 Experimental Setup

In this study, we conduct three simulations using neXtSIM. First, we run two neXtSIM simulations using the BBM and mEVP rheologies with nearly the same parameters described in Ólason et al. (2022) (their Table 1). We refer to these two simulations using the acronym of their rheology (BBM and mEVP). The BBM simulation is the main focus of this study, while the mEVP run serves as a reference to estimate what the results would be using a “standard” rheology with the same thermodynamics such as in Ólason et al. (2022). The main differences to the Ólason et al. (2022) study are the use of a larger grid spacing (50 vs. 10 km), longer time steps (1,800 and 60 s), and reduced quadratic wind drag coefficient to $$1.6\times 1{0}^{-3}$$ such as in Boutin et al. (2023). For the BBM run, we also increased the exponent $$cf$$ (compression factor) from 3/2 to 2 (please see Equation 8 in Ólason et al., 2022). A larger $$cf$$ produces thinner coastal ice and increased offshore ice thickness and slightly improves modeled deformation and thickness in the Arctic (Figure 11 in Ólason et al., 2022).

Table 1. Wind and Ocean Drag Coefficient Values in the Different Simulations Used in This Study

Parameter (symbol)	BBM	mEVP	mEVP+	BSOSE
Wind drag coefficient $$\left({C}_{a}\right)$$	$$1.6\times 1{0}^{-3}$$	$$1.6\times 1{0}^{-3}$$	$$2.0\times 1{0}^{-3}$$	$$1.0\times 1{0}^{-3}$$
Ocean drag coefficient $$\left({C}_{w}\right)$$	$$5.5\times 1{0}^{-3}$$	$$5.5\times 1{0}^{-3}$$	$$5.5\times 1{0}^{-3}$$	$$8.5\times 1{0}^{-3}$$

• Note. The last column gives the values used in the sea ice model part of BSOSE, as an indication.

The third simulation (mEVP+) is also run with the mEVP rheology, but with a larger quadratic wind drag coefficient $$\left(2.0\times 1{0}^{-3}\right)$$ than in the mEVP run. This is because using default parameters taken from Ólason et al. (2022), the mEVP and BBM runs generate different average drift speeds which impact divergence/convergence, sea ice thickness and other properties. The larger drag coefficient generates average sea ice drift speeds similar to the BBM run, so the mEVP+ run allows for further investigation of differences in simulated sea ice properties that can be more attributed to the rheology rather than the resulting sea ice drift speed. The drag coefficient values used in each simulation are presented in Table 1. We also provide the standard parameters for the sea ice model implemented in BSOSE for reference.

2.3 Observations and Model Evaluation Metrics

Model results from all experiments are compared against satellite observations of sea ice extent, drift and thickness. Sea ice extent is computed using the sea ice concentration product generated by the climate data record of the EUMETSAT Ocean and Sea Ice Satellite Application Facility (OSI-SAF, Lavergne et al., 2019). Grid cells are considered ice-covered if their sea ice concentration exceeds 15%. We use the most updated product version (OSI-430-b) which covers from 2016 to present. To evaluate the simulation, we compute the integrated ice extent error (IIEE, Goessling et al., 2016) for summer and winter seasons. In addition, we calculate what we refer to as the ice extent accuracy (IEA %), that is the fraction of all grid points where both model and observation agree that is, the sum of all “true positives” over the model grid sea ice extent. This is compared with an ice extent inaccuracy (IEI), the area of all sea ice extent points that are “false negatives” divided by the model total sea ice extent. IEA (percentage of true positives) added to false positives (overestimation) gives 100%.

For sea ice drift velocities, we use the low-resolution ($${\sim} $$60 km) daily OSI-SAF sea ice drift product, which provides 2-day integrated sea ice displacement (Lavergne et al., 2010). We used observations covering the period from 2015 to 2020. In our evaluation, we compute the complex vector correlation of ice drift vectors (Kundu, 1976) from these daily vectors. This metric provides information about both magnitude and direction correlation between the simulations and the observations.

Sea ice thickness comparisons are made using observations from Cryosat-2 (Fons et al., 2023), covering the period 2010–2021. This data set provides monthly averages for the period analyzed and a level of uncertainty associated with each month. These uncertainty levels are due to the difficulties of estimating the snow depth over Antarctic sea ice and the lack of in-situ data available for the calibration/validation of remote sensing estimates. Fons et al. (2023) estimated that month average uncertainty levels can reach 40% in the Austral winter. Additional sea ice thickness/volume observations from Kacimi and Kwok (2020) are included in the model validation. These authors computed sea ice volume using LiDAR (ICESat-2) and radar (CryoSat-2) altimetry, with retrievals of freeboard, snow depth, and ice thickness from April to November 2019. Kacimi and Kwok (2020) also provide estimates of minimum and maximum Antarctic sea ice volume for each month. Model SIT are objectively validated using root mean square error (RMSE) given by:

$$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}{{\Sigma }}_{i=1}^{n}{\left({x}_{i}-{y}_{i}\right)}^{2}};$$$ (1)

between observed (x) and modeled (y) results, where i = 1,2,…,n are the observation months. SIT RMSE is computed using monthly averaged model results interpolated to observation locations inside the domain.

We compared ice growth rates, ridging ratio and divergence between the neXtSIM runs. New ice growth is computed by neXtSIM as the amount of ice formed at the surface by freezing and is expressed in meters per day. Ridging ratio is the volume fraction of ridged ice over total ice in a grid cell and is computed within neXtSIM. More details about ridging calculations are provided in Appendix A. Divergence (div) is defined as $$\mathrm{d}\mathrm{i}\mathrm{v}$$ = $$\partial $$u/$$\partial $$x + $$\partial $$v/$$\partial $$y where x and y (u and v) are the zonal and meridional components of space (drift). Convergence is negative divergence. We define summer here as January-February-March (JFM) because it symmetrically surrounds the minimum observed sea ice extent around Antarctica. Similarly, we define winter as August-September-October (ASO) as it is centered around the maximum observed Antarctic sea ice extent.

The evaluation metrics used in this study capture critical aspects of Antarctic sea ice characteristics, focusing on extent, drift, and thickness, each with specific sensitivities and limitations. The Integrated Ice Extent Error (IIEE) quantifies seasonal discrepancies between modeled and observed ice extent, while IEA and IEI measure agreement on ice coverage, though both are sensitive to the choice of concentration threshold (e.g., 15%) and may vary along the ice edge. For sea ice drift, the complex vector correlation assesses alignment in drift patterns, capturing magnitude and direction but potentially overlooking finer-scale variations that are not resolved by the available observational data. Sea ice thickness (SIT) RMSE provides a reliable measure of model performance in thickness representation, yet is influenced by uncertainties in observational data, especially during winter. Additional metrics, including new ice growth, ridging ratio, and divergence, further inform ice formation and deformation processes within the model. Together, these metrics offer a robust framework for evaluating model performance, though they are sensitive to spatial resolution, observational uncertainties, and regional variability, particularly near the ice edge. For that reason, we performed analyses in regional sub-divisions, to enhance the understanding of model performance in areas of high spatial and temporal variability.

3 Results

3.1 Sea Ice Extent and Presence

Figure 2 shows the mean seasonal cycle of sea ice extent computed for the three model runs, BSOSE's sea ice outputs, and the OSI-SAF observational products. The observed daily SIE seasonal cycle shows a minimum between late February and March ($${\sim} 2.5\times 1{0}^{6}$$ $${\text{km}}^{2}$$) which is well-captured by all models. After that, Antarctic sea ice expands at the rate of 2.5$${\times} $$ $${10}^{6}$$ $${\text{km}}^{2}$$/month. This expansion is captured by all neXtSIM simulations, but sea ice grows more rapidly in the BSOSE reanalysis ($$3.5\times 1{0}^{6}$$ $${\text{km}}^{2}$$/month). Observed Antarctic SIE reaches a peak ($${\sim} 18\times 1{0}^{6}$$ $${\text{km}}^{2}$$) around late August, September and October that is overestimated by all model runs, but that is most inaccurate for the BSOSE reanalysis ($${\sim} 22\times 1{0}^{6}$$ $${\text{km}}^{2}$$). This positive bias in all model simulations can be a result of a colder lower atmosphere in ERA5 which increases SIE during winter. Otherwise, this positive bias could be attributed to colder waters in BSOSE which results in large winter SIE in BSOSE and neXtSIM runs. The origin of the difference between the ice extent predicted by neXtSIM and the one from BSOSE is uncertain, as these models use different grid frameworks (Lagrangian vs. Eulerian), thermodynamics and dynamics (see Section 2).

Figure 2

Mean Antarctic sea ice extent (SIE) from OSISAF-ease2 observations (green) of sea ice concentration Climate Data Record (CDR, release 3), Biogeochemical Southern Ocean State Estimate reanalysis (red), and mEVP (orange), mEVP+ (blue) and BBM (black) runs. The solid lines show daily averages and the shades represent the daily standard deviation between 1st of Jan 2016 and 31st of Dec 2021.

To understand the origin of these biases, we investigate the spatial distribution of sea ice extent (SIE) discrepancies between neXtSIM simulations and observations (Figure 3). In summer (JFM), the observed SIE has an average of 3.97 $${10}^{6}$$ $${\text{km}}^{2}$$ which is overestimated by all simulations, especially BSOSE (8.21 $${10}^{6}$$ $${\text{km}}^{2}$$) (top left panel in Figure 3) due to its large SIE values in January (Figure 2). This overestimation mostly comes from the Ross Sea area, where no simulation can reproduce the tongue of open water in the Ross Sea (southernmost red shades in the left column of Figure 3). The mEVP and BBM runs improved the SIE results compared to BSOSE, reducing JFM SIE by more than 20%. Winter (ASO) SIE is an average of 18.24 $${10}^{6}$$ $${\text{km}}^{2}$$ in observations, and is overestimated by 5.51 $${10}^{6}$$ $${\text{km}}^{2}$$ by BSOSE (top right panel in Figure 3). This overestimate is distributed uniformly along the ice edge.

Figure 3

January-February-March (JFM–left-hand column) and August-September-October (ASO–right-hand column) averages of Antarctic SIE from Ocean and Sea Ice Satellite Application Facility observations (white and blue shades within the black contour), Biogeochemical Southern Ocean State Estimate reanalysis (top row), mEVP (middle row), and BBM (bottom row) runs. Red shaded areas represent regions where the model overestimates SIE, and blue-shaded area are regions of underestimate. The text in each section of the Southern Ocean shows integrated ice extent error (IIEE, Goessling et al., 2016) which corresponds to the sum between over- and underestimated areas. The numbers over Antarctica represent the total sea ice extent. Averages are computed using data from year 2016–2021.

In Figure 4 we show the ice extent accuracy (IEA–percentage of true positives, the white area within the black contour in Figure 3) and inaccuracy (IEI– percentage of false negatives, blue area in Figure 3) for all models and BSOSE. Sea ice extent predicted by neXtSIM is very consistent with observations from April to November (IEA higher than 80% and IEI lower than 10%), and all neXtSIM simulations outperform BSOSE during this period. The smallest IEA generated by BSOSE is explained by the large overestimation of SIE (false positives) which reduces its accuracy. In all models, there is a significant reduction in IEA in summer months when IEI reach a maximum in March. BBM and mEVP simulations behave very similarly overall. BBM has the highest IEA in winter but the lowest in summer. The IEA (IEI) tends to reduce (increase) in summer because the model total sea ice extent reduces, and those indexes are calculated over it. When the total sea ice extent is reduced (summer), smaller errors create larger inaccuracies.

Figure 4

Average Antarctic sea ice extent (SIE) accuracy index from Biogeochemical Southern Ocean State Estimate reanalysis (red), mEVP (orange), mEVP+ (cyan) and BBM (black) runs compared against Ocean and Sea Ice Satellite Application Facility observations of sea ice extent. The solid lines show the percentage of true positive (or accurate) points—white area in Figure 3. The percentage of true positives added to false positives (overestimation) gives 100%. Dashed lines represent the percentage of false negative (or underestimated) points—blue area in Figure 3. The percentages are computed over the model total sea ice extent (red and white shades in Figure 3).

3.2 Sea Ice Drift

Observed sea ice drift speed reaches a maximum around September and a minimum in February–this seasonality was matched by the numerical experiments (Figure 5), although this seasonal cycle has a lower amplitude. All model runs underestimate sea ice drift between June and November and overestimate it from January to March. Summer ice drift provided by OSI-SAF, however, is highly uncertain. It is mostly based on a free-drift model and is known to be negatively biased (Lavergne et al., 2010). This is true for other satellite retrievals of summer ice drift and most products and analyses tend to be focused on the (winter) months April to October (e.g., Holland & Kwok, 2012). In that context, the model comparison with OSI-SAF for the summer months shown in Figure 5 should be interpreted carefully. Sea ice drift speed in the BBM simulation is generally larger than in mEVP, for all months, and therefore agrees better with OSI-SAF in the winter. This difference is not necessarily due to an advantage of the BBM rheology over mEVP because sea ice drift depends on both rheological parameters and constant drag coefficients that are poorly constrained (e.g., Ólason et al., 2022). Here, default parameters for the drag coefficients in neXtSIM may favor the BBM rheology. As a demonstration, a simple increase of the wind drag coefficient from $$1.6\times 1{0}^{-3}$$ to $$2.0\times 1{0}^{-3}$$ in mEVP+ significantly improved the drift speed seasonality amplitude with a better agreement with OSI-SAF than the BBM simulation.

Figure 5

Average Antarctic sea ice drift speed observations (green), Biogeochemical Southern Ocean State Estimate reanalysis (red), mEVP (orange), mEVP+ (cyan) and BBM (black) runs. The solid lines show averages of filtered timeseries using a 10-day moving mean for better visualization. The green shade represents the seasonal standard deviation of the 10-day moving average filtered observations.

We investigate the ability of the model to capture the daily magnitude and direction of sea ice drift using the complex vector correlation between modeled and observed drift. The BBM run has the largest complex vector correlation (mean = 0.76) (Figure 6), and it remains relatively stable throughout the year, with a small drop at the end of summer (February to April). Increasing the wind drag coefficient in the mEVP+ run (mean = 0.60) slightly improved the complex correlation compared to the mEVP run (mean = 0.58). Still, it is not enough to reach the levels of correlation simulated by the BBM run, and its lower correlation levels occur in the summer months (December to April). We explore the potential reasons for these discrepancies in Section 3.4.1.

Figure 6

Mean Antarctic sea ice drift vector complex correlation between observations and mEVP (orange), mEVP+ (cyan) and BBM (black) runs. The lines show averages of filtered timeseries using a 10-day moving average for better visualization.

Maps of complex vector correlation between model runs and observations also show a more consistent representation of Southern Ocean daily sea ice drift in the BBM run compared to the mEVP run (Figure 7). The BBM run shows correlation coefficients larger than 0.7 in most regions of the Southern Ocean with an average of 0.72 in summer and 0.74 in winter (Figure 7). The mEVP simulation has correlation coefficients below 0.7 for most regions in the two seasons, except for a small area in the Ross Sea and a few scattered spots in the Southern Ocean during summer. Increasing the ice-atmosphere drag coefficient in the mEVP+ significantly reduces the drift speed bias but results in a small increase (0.02) in the average correlation coefficient in winter (0.03 in summer). This suggests that the difference in average complex correlation (Figure 6) between the BBM and mEVP/mEVP+ simulations may be attributed to the ability of the BBM simulation to better capture the daily drift variability and its spatial patterns. We discussed this further in Section 3.4.1. We also note that sea ice drift in waters deeper than 300 m is better simulated than coastal sea ice for the mEVP, mEVP+ and BBM runs, in both winter and summer.

Figure 7

January-February-March (JFM–left-hand column) and August-September-October (ASO–right-hand column) averages of Antarctic sea ice drift complex vector correlation between observations and the mEVP (top row), and BBM (bottom row) runs. The black contour line represents a correlation coefficient of 0.7. Averages are computed using data from year 2015–2020. The average complex correlation is shown for each segment of the Southern Ocean, as well as total, coastal (depth $${< } $$ 300 m), and deep (depth $${\ge} $$ 300 m) averages.

3.3 Sea Ice Thickness and Volume

Sea ice thickness (SIT) differences between models and observations show that the mEVP and BBM runs generate thicker sea ice near the coast and thinner in the deep ocean compared to CS2WFA data (Figure 8). The positive bias is larger in the BBM run with thicker ice in the Weddell Sea and on the west side of the Ross Sea region. However, Southern Ocean total average SIT is still underestimated by the two model runs with larger negative biases in summer (mEVP = −0.43 m; BBM = −0.33 m) compared to winter (mEVP = −0.22 m; BBM = −0.09 m). The mEVP+ shows similar results to the mEVP run with a slightly smaller SIT average bias (−0.40 m) in summer and the same average difference in winter. This negative bias mostly originates from sea ice offshore, while coastal areas tend to be positively biased, particularly in the BBM simulation. The positive bias in the Weddell Sea corresponds to the location of its perennial sea ice.

Figure 8

January-February-March (JFM–left-hand column) and August-September-October (ASO–right-hand column) averages of Antarctic sea ice thickness (SIT) difference to observations (mod-obs) from the mEVP (top row), and BBM (bottom row) runs. Averages are computed using data from Jan. 2015 to Aug. 2021. The average SIT difference is shown for each segment of the Southern Ocean, as well as observed total, coastal (depth $${< } $$ 300 m), and deep (depth $${\ge} $$ 300 m) averages.

The root mean square error (RMSE) analysis shows slightly larger values in the BBM run (0.72 m) compared to the mEVP run (0.69 m) in summer (JFM) (left-hand column in Figure 9). A similar pattern occurs in winter (ASO), albeit with smaller errors, when the BBM run has an average RMSE of 0.48 m and the mEVP RMSE is 0.47 m (left-hand column in Figure 9). This larger RMSE in the BBM run is mostly attributed to the SIT overestimation in the Weddell Sea and on the west side of the Ross Sea region. The BBM run, however, has slightly smaller RMSE in the Amundsen and Bellingshausen Seas in summer (0.04 m) and winter (0.01 m) which are associated with larger SIT underestimation in that region by the mEVP run (Figure 8). In any case, these differences in model RMSE are small when compared to the uncertainties in the observations. SIT uncertainty has an average of 0.38 m and can be as large as 7 m in certain coastal locations in the Fons et al. (2023) data set.

Figure 9

January-February-March (JFM–left-hand column) and August-September-October (ASO–right-hand column) root mean square error (RMSE) of Antarctic sea ice thickness (SIT, in meters) compared to observations (CS2WFA) from the mEVP (top row), and BBM (bottom row) runs. RMSEs are computed using data from Jan. 2015 to Aug. 2021. The average SIT RMSE is shown for each segment of the Southern Ocean, as well as total area, coastal (depth $${< } $$ 300 m), and deep (depth $${\ge} $$ 300 m) regions.

Timeseries analysis of observed monthly SIT shows thicker ice in February, at the end of summer/melting season (Figure 10), when there is only thick sea ice ($${\sim} $$1.50 m) left near the coast (Fons et al., 2023; Parkinson, 2019). During autumn and winter, new and thinner sea ice forms ($${\sim} $$1.15 m) which dominates the Pan-Antarctic average until the beginning of the next melting season in October/November. All numerical simulations show lower averages for the full period compared to the observations. However, they tend to remain within the uncertainty range of the observations most years during winter (green shade in Figure 10).

Figure 10

Monthly averages of Antarctic sea ice thickness (SIT, in meters) from observations (green line) and their uncertainty (green shade), Biogeochemical Southern Ocean State Estimate reanalysis (red), mEVP (orange), mEVP+ (blue) and BBM (black) runs. The solid lines show month averages from model output interpolated to observations and masked where observations are not available.

The modeled SIT seasonal cycle tends to peak in November or December, earlier than observed values, which generally peak in January or February (Figure 10), regardless of the simulation. This phase difference from observations is similar to what is seen in models that use a 0-layer thermodynamic approach, where sea ice does not have a heat capacity. However, neXtSIM applies the 3-layer model introduced by Winton (2000), which should reduce this lag. This is illustrated by mean SIT in BSOSE (0-layer model) which peaks earlier than neXtSIM (e.g., 2015, 2016, and 2019). Therefore, we suggest this error could originate from uncertainties in the forcings or in the simplistic treatment of the snow layer in the Winton (2000) 3-layer model that may not be adapted to Antarctic conditions. The modeled mean SIT decreases from January until it reaches its minimum in March. Sea ice thickness retrieved using Cryosat-2 does not show such a minimum in March. Instead, it shows a sharp decrease in SIT from February to March followed by a plateau that lasts until October/November. The discrepancy between the model and observations is associated with a lack of thick ice adjacent to the coast in all regions but the Weddel Sea (Figure 8). Modeled sea ice thickness in the Weddel sea is generally higher than in observations, especially in the BBM run.

Sea ice volume results from multiplying grid cell area, thickness and concentration. The latter variable (concentration/area) has a larger influence on it and makes total sea ice volume (Figure 11) resemble sea ice extent variability (Figure 2) which differs from the mean SIT timeseries (Figure 10).

Figure 11

Time series of Antarctic sea ice volume (SIV) from Biogeochemical Southern Ocean State Estimate (BSOSE) reanalysis (red), mEVP (orange), mEVP+ (blue) and BBM (black) runs. The solid lines show 6-hourly (5-day for BSOSE) averages and the dashed lines represent annual means. Observed monthly average estimates from Fons et al. (2023) are shown as purple squares and estimates of sea ice volume from Kacimi and Kwok (2020) as green stars assuming different values of CroySat-2 snow depth adjustment.

Annual averages of the sea ice volume timeseries show clear differences between the numerical simulations (dashed lines in Figure 11). The BBM run has the largest volume followed by the mEVP and mEVP+ runs which had similar results. Increasing the wind drag coefficient in the mEVP+ run does not make much difference in sea ice volume compared to the mEVP run.

All model data sets show a similar seasonal cycle for sea ice volume. The peak values reach between $${\sim} $$15,500 $${\text{km}}^{3}$$ (BSOSE) and $${\sim} $$19,000 $${\text{km}}^{3}$$ (BBM) in the winter of 2019. September 2019 average estimates from Kacimi and Kwok (2020) show a larger sea ice volume range which oscillated from $${\sim} $$10,500 $${\text{km}}^{3}$$ and $${\sim} $$21,500 $${\text{km}}^{3}$$ considering no ice freeboard (minimum estimate) and ice freeboards from CryoSat-2 (largest value) respectively. Intermediate values were computed using adjustments of 6 cm (third largest) and 3 cm (second largest) applied to snow depth estimates from CryoSat-2 (Kacimi & Kwok, 2020). The latter estimate is similar to the peak in volume in the mEVP run ($${\sim} $$17,500 $${\text{km}}^{3}$$), but this peak occurs later, in October, like in the other runs and in BSOSE. This delayed peak in sea ice volume is related to the late peak in sea ice extent which can be attributed to a colder atmosphere (ERA5), ocean (BSOSE) or both. The September 2019 average sea ice volume from Fons et al. (2023) falls between the BBM and mEVP simulations winter maxima in 2019, suggesting winter sea ice volume in these simulations is acceptable.

In summer 2019, sea ice volume minimum in neXtSIM runs and BSOSE are close to estimates from Fons et al. (2023), but underestimate the volume in April 2019 compared to estimates by Kacimi and Kwok (2020). This corresponds to the refreezing period, and this underestimate could be due to a delay in the refreezing in the model. The agreement between the model and BSOSE suggest this underestimate could also be impacted by our choice of ocean forcings.

Model interannual variability showed similar patterns to observations. Winter peaks of sea ice volume are larger in 2015 and decrease until a local minimum in 2017 similar to observations (Fons et al., 2023). The local minimum sea ice volume in 2017 also corresponded to the small Antarctic SIE found by Parkinson (2019). The model runs also reached another local maximum in 2020 which is similar to what was shown by Fons et al. (2023) in their Figure 8.

3.4 Analysis of the Differences Between the Simulations

3.4.1 Difference in Model Drift Skill

Model sea ice drift from the mEVP and BBM runs showed similar mean sea ice drift magnitude for large and long scales (e.g., Pan-Antarctic and seasonal). Moreover, sea ice drift can be tuned with unconstrained parameters such as drag coefficients (see Figure 5). However, complex correlation applied to daily ice drift vectors suggests that the BBM rheology better represents sea ice drift than the mEVP rheology at a coarse resolution (50 km) (Figures 6 and 7). Previous results from the Arctic Ocean suggest that this discrepancy originates from the BBM rheology's ability to capture the variability of the drift associated with extreme events (Rheinlaender et al., 2022) which are common around Antarctica.

To support this hypothesis, we show sea ice drift speed and direction from observations and neXtSIM simulations on 05/07/2016 when a low-pressure system (Figure 1) generated a cyclonic sea ice circulation in the Weddell Sea (magenta circles in Figure 12). In the mEVP and mEVP+ runs, we observe a cyclonic sea ice circulation with a center displaced eastward and closer to the coast than in the observations. The BBM run shows a strong westward coastal drift and a cyclonic shape similar to the observations. The offshore drift in the Ross Sea is another feature worth noting which has larger negative differences in the mEVP and mEVP+ runs compared to the BBM run.

Figure 12

Maps of average sea ice drift (km/day, arrows and green shade) from observations (a), and drift (arrows) and drift difference (red and blue shades in km/day) from mEVP (b), mEVP+ (c), and BBM (d) runs on 05/07/2016 UTC. The magenta circle highlights the position of the atmospheric cyclone (center of the low-pressure system) shown in Figure 1.

Timeseries analysis of wind speed and complex vector correlation between observed and simulated sea ice drift shows improved complex correlation during large and strong wind events in the Pan-Antarctic outlook (top panel in Figure 13). The BBM simulation shows larger complex correlation ($${\sim} $$0.8) associated with slightly higher linear correlation with wind speed compared to the mEVP runs. Weddell Sea timeseries analysis highlights the higher magnitude of relationship between wind speeds and the ability of the BBM simulation to reproduce observed sea ice drift patterns (complex vector correlation) (bottom panel in Figure 13). During large wind speed events (5th, 18th, 22nd and 25th of July 2016), BBM's drift complex vector correlation with observations largely increases. For the same cases, sea ice in mEVP and mEVP+ simulations tend to respond (increased complex correlation) with later/delayed peaks in wind speed (e.g., 5th and 25th of July 2016) and does not correlate well with the wind forcing in the Weddel Sea. This suggests sea ice responds faster to the wind in our BBM simulation compared to mEVP/+ runs.

Figure 13

Time series of daily spatial drift complex correlation (left axis) from the mEVP (orange), mEVP+ (blue) and BBM (black) runs computed for the Southern Ocean (top row) and Weddel Sea (bottom row). Hourly wind speed from ERA5 (green line) for each region is also shown on the right-hand axis. The values shown in the figure legend represent the linear correlation between wind speed and drift complex correlation with observations. Please refer to Figure 1 to identify the Weddell Sea study area.

This analysis corroborates results from Ólason et al. (2022) and Rheinlaender et al. (2022) for the Arctic Ocean. These authors showed that 10-km resolution models using the BBM rheology captured the timing and spatial patterns of local sea ice deformations (i.e., derivatives of the drift) associated with a breakup event in a way that the mEVP rheology, as implemented in neXtSIM, could not. Viscous-plastic rheologies can generate such fractures and may improve drift simulations in high resolution models ($${< } $$5 km) (Bouchat et al., 2022; Hutter et al., 2022). However, these resolutions make their application unfeasible for climate modeling studies.

The difference in sea ice drift response to storms between the BBM and mEVP simulations likely originates from the BBM model's ability to account for mechanical damage and its rapid impact on reducing sea ice strength. In the brittle framework, strong winds quickly cause shear stress that increases damage within just a few timesteps, leading to a rapid drop in ice strength. This weakened ice then undergoes significant localized deformation, which promptly affects the drift field over short (e.g., daily) timescales. In contrast, the mEVP rheology, and VP rheologies in general, lack this quick feedback mechanism. Here, ice strength only decreases when sea ice diverges, as it thins and reduces in concentration. These processes occur more slowly, over hours rather than minutes. This delay may slow the drift and deformation response during storms, and higher spatial resolution is needed to capture large local changes in thickness and concentration. Our results, like previous findings from Ólason et al. (2022), Rheinlaender et al. (2022), emphasize the importance of simulating the rapid drop in sea ice strength due to breakup events to capture an accurate sea ice drift response, particularly around Antarctica.

3.4.2 Models' Ability to Represent Antarctic Sea Ice Extent and Thickness

In Section 3, we have shown, for the first time, that using neXtSIM with the BBM rheology generates reasonable results for large-scale sea ice properties of the Antarctic sea ice, without requiring any particular retuning. For sea ice extent, the results depend little on the choice of the rheology. This is likely generated by the constraints imposed by the reanalyzes (BSOSE and ERA5) used to force the sea ice models. These reanalyzes have generated/experienced larger sea ice extent, and, likely, neXtSIM's sea ice extent cannot significantly differ from that. Nevertheless, there is a small improvement in our simulated sea ice extent compared to BSOSE's sea ice extent, the cause of which is not clear.

Sea ice thickness from the mEVP simulation is slightly more consistent with the observations than the BBM run. However, regional differences remain relatively small and they could be affected by re-tuning both thermodynamics and dynamics in one or another simulation. In addition, observations are associated with high uncertainties. We interpret that there is no major difference in skill at a large scale depending on the rheology since the differences between models' SIT averages (0.12 m) and mean RMSE (0.02 m) are smaller than the observation average uncertainty (0.38 m) (Fons et al., 2023).

Sea ice is generally thicker with the BBM rheology than with mEVP which is similar to the Arctic Ocean (Ólason et al., 2022). The thicker ice simulated in the BBM run is attributed to feedback from the damage represented in the model, and convergent failure, that affect ice strength or elasticity during strong wind events. For instance, we evaluate sea ice response to the cyclonic event displayed in Figure 12 to illustrate how ice breakup could explain this difference in the resulting sea ice thickness and volume. In Figure 14, we show sea ice divergence, ice growth, and ridged sea ice percentage for mEVP and BBM runs. Results from the BBM run show significant sea ice strain in response to the cyclone with radiating divergence patterns (red stripes in top-right panel in Figure 14) indicating the formation of leads or less concentrated sea ice in the Weddell Sea (blue stripes in right panel in Figure 1). These divergent regions provide conditions for the growth of new ice in the openings, while ridging happens when convergence dominates the drifting patterns. This process explains the generation of thicker ice in the BBM run compared to mEVP/+ runs.

Figure 14

Maps of sea ice divergence (1/day—top row), ice growth via supercooling (new ice—middle row), and percentage of ridged ice over total ice (ridge ratio—bottom row) from the mEVP (left) and BBM (right) runs at 21hr 05/07/2016 UTC (same time as Figure 1). The magenta circles highlight the atmospheric low-pressure center shown in Figure 1.

Histogram analyses of divergence, convergence and ice growth show that the differences between the mEVP and BBM occur in the lower and intermediate part of those variables' spectra (Figure 15). Increasing the wind drag coefficient in the mEVP+ run increases the lower levels of divergence but is not enough to increase new ice growth at the magnitude of the BBM run. Convergence levels are higher in the BBM simulation compared to the mEVP/+ runs. mEVP and mEVP+ have similar results for convergence in all categories. The ridging ratio is generally higher in the BBM simulation, with significantly larger occurrences of grid cells with values between 30% and 80% compared to the mEVP/+ simulations. The BBM run, however, has fewer occurrences for ridging ratio below 20% than the mEVP/+ runs. The mEVP+ run has increased levels of ridging from 20% compared to the mEVP run and is similar to the BBM run in the range of values of 80% and above. This result shows that sea ice tends to pile up more frequently (subgrid parameterization) and mechanically generate thicker ice in the BBM run. Differences in the sea ice volume are only attributed to divergence and ice growth since convergence and riding do not increase the total ice volume directly but only thickness.

Figure 15

Histograms of divergence, ice growth, convergence and ridging ratio from the mEVP (orange), mEVP+ (blue) and BBM (black) runs computed using 6-hourly outputs from 1st of Jan to 31st of Dec 2016 for the whole model domain.

4 Discussion

Our results show marked differences in ice volume and drift between models using the BBM or mEVP rheology at 50 km resolution. This highlights the importance of properly simulating pack ice deformation for Antarctic sea ice and stresses the importance of accurately modeling sea ice dynamics in both hemispheres. The brittle rheology, initially designed to describe multiyear and pack ice behavior—characteristic of the Arctic region—reasonably captures pan-Antarctic sea ice properties despite its more seasonal nature. Differences with the simulations using mEVP arise mostly in areas where sea ice is more compact and less seasonal, like in the Weddell Sea. This is not unexpected, as this is where sea ice breakup in pack ice is more likely to occur. Here, we suggest a few research questions where the skills of the BBM rheology demonstrated in this study and previous work on Arctic sea ice could be relevant.

The BBM rheology could be a relevant tool to investigate the Antarctic sea ice thickness response to potential changes in sea ice deformations, in a way similar to what has been done in the Arctic (e.g., Boutin et al., 2023; Rheinlaender et al., 2024). Our results suggest that the BBM simulation has some skill in reproducing these deformations, given the ice drift can be interpreted as the $${0}^{\text{th}}$$ order of deformations. However, to provide relevant insights, such studies first require a direct assessment of the model's ability to reproduce sea ice deformations in Antarctic sea ice. This evaluation may not be as straightforward as in the Arctic, mostly because, to our knowledge, there is no reference data set such as the RGPS (Kwok et al., 1998) to serve this purpose.

Another possible use is to investigate the impact of atmosphere-ice interactions in the Southern Ocean on controlling the variability of sea ice extent and thickness (e.g., Hobbs et al., 2016). The ability of the BBM rheology to capture the imprint of storms demonstrated in these and previous results suggests it could capture their impact even in coarse-resolution climate models, offering new opportunities to investigate the sensitivity of sea ice to these extreme events. For instance, the current setup, with forced atmospheric and oceanic conditions, could be used to investigate potential changes in Antarctic sea ice deformations in the context of climate change, or assess their impact on atmosphere-ice-ocean interactions, that are more important in the Southern Ocean than in the Arctic Ocean (Maksym, 2019).

5 Conclusions

We introduced a new configuration of neXtSIM for the Southern Hemisphere in which we use a brittle rheology to represent Antarctic sea ice for the first time. neXtSIM has been originally built to simulate Arctic pack ice, and previous to this study, its applicability to Antarctic sea ice was unknown. Southern Ocean sea ice properties relevant for climate studies like extent, thickness, volume, and drift are generally well-represented. We investigated and discussed the differences in modeled properties that could be attributed to the change in rheology. We found that the brittle rheology we used led to a thicker ice cover resulting from small-scale deformation events (ice growth in leads and ridging) and a better representation of the drift variability on short temporal timescales. We suggested both these results were due to the ability of the brittle rheology to parameterize the subgrid effect of sea ice breakup and its impact on sea ice drift and deformation, as has been shown previously in the Arctic (Rheinlaender et al., 2022).

Our results have highlighted the potential importance of accounting for the effects of sea ice deformations even at a coarse resolution (50 km) and in the presence of mostly seasonal sea ice. Further evaluation is required to ensure these deformations are well represented. Still, our results support the idea that brittle rheologies, for example, BBM, could be useful in investigating the impact of atmosphere-ice-ocean interactions during extreme events, like strong cyclones, especially at coarse resolutions relevant for climate simulations. This is now feasible since BBM has recently been implemented in a sea ice model using a more widely adopted Eulerian framework (Brodeau et al., 2024), rather than the Lagrangian framework used in neXtSIM. Understanding the importance of these processes may help correct potential biases in current models, thus reducing the spread in the projections of Antarctic sea ice properties.

Acknowledgments

This work was supported by Schmidt Futures, a philanthropic initiative that seeks to improve societal outcomes by developing emerging science and technologies. R.S. thanks New Zealand's Antarctic Science Platform (ASP) for supporting this work. We thank everyone involved in the support teams from the New Zealand eScience Infrastructure (NeSI) and the Institute of Computing for Climate Science (Cambridge, UK). We also thank the anonymous reviewers for their thoughtful comments.

Appendix A: Ridging Calculations

neXtSIM tracks the volume ratio of ridged ice throughout the simulation. In the Lagrangian framework, when convergence occurs the element area is reduced, but the volume of ice in the element is conserved. If the element undergoing convergence is fully ice-covered, volume conservation implies that the area reduction results in an increase of the mean thickness H. Assuming that the thickness of level ice is unchanged by ridging, we can show that the ridge ratio R at time (n + 1) should be computed as:

$$${R}_{n+1}=1-\left(1-{R}_{n}\right){H}_{n}/{H}_{n+1}$$$ (A1)

where H is the mean ice thickness over the element.

Assumptions must be made for the evolution of R when ice grows or melts. We assume that when sea ice grows, the volume of ridged ice is conserved according to $${R}_{n+1}={R}_{n}{H}_{n}/{H}_{n+1}$$, and when the ice melts, ridged and level ice melt at the same rate $$\left({R}_{n+1}={R}_{n}\right)$$. More details about the ridging calculations can be found in (Ólason et al., 2025; Williams et al., 2021).

Open Research

Data Availability Statement

The version of code used to run the model can be found here: https://doi.org/10.5281/zenodo.14183764, Nansen Environmental and Remote Sensing Center (2024). The outputs from the model runs used in this paper are available for download here: https://doi.org/10.5281/zenodo.12770130, Santana (2024). BSOSE reanalysis is available at http://sose.ucsd.edu/SO6/ITER139/, Mazloff et al. (2023) (last access on the 5th of July 2024).