In this study, we use the medium-resolution NorESM version 1 NorESM1-ME NorESM1-ME,. NorESM is based on the Community Earth System Model CESM,, but with the ocean and atmosphere components changed. The original Community Atmosphere Model version 4 (CAM4, ), is replaced with the atmospheric model CAM4-OSLO with a prognostic aerosol life cycle formulation using emissions and new aerosol-cloud interaction schemes . The ocean component is replaced by the Bergen Layered Ocean Model (BLOM), a general circulation model which uses an isopycnal layered coordinate system. The sea ice component is the Los Alamos Sea Ice Model version 4 CICE,. The land component is the Community Land Model (CLM) version 4 . The model components are coupled using version 7 of the CESM coupler .

The atmosphere and land models have a horizontal resolution of 1.9° in latitude and 2.5° in longitude and 26 hybrid sigma-pressure vertical levels. The ocean and sea ice models have a horizontal resolution of 1° × 1°. BLOM includes 51 isopycnal layers and a bulk mixed layer, with time-evolving thicknesses and densities. NorESM runs with CMIP5 historical forcings (CMIP5 Historical +RCP8.5) and the Representative Concentration Pathway 8.5 (RCP8.5) beyond 2005. It is worth noting that using a more realistic external forcing would improve the model representation of the atmospheric trends and improve sea ice prediction skill of NorCPM.

The ice thickness distribution in CICE is represented using five thickness categories, which is optimal for representing the sea ice cover at reasonable computational cost and speed . The thermodynamic equations are solved by the one-dimensional vertical Bitz and Lipscomb model , including parametrisations for melt pond , aerosol and radiation transfer parameterisations . The incremental remapping scheme of is used to solve for the horizontal transport of sea ice. The sea ice stresses are formulated using the elastic-viscous-plastic rheology . The remapping scheme of transports the sea ice in thickness space.

NorCPM uses a sequential data assimilation method characterised by a Monte Carlo update followed by a linear analysis update . Specifically, we use a deterministic formulation of the EnKF , which solves the analysis without needing to perturb the observations. This method is multivariate, and all or any of the model state vector (depending on user choice) can be updated through their ensemble covariances with the observations. The deterministic EnKF is particularly useful for many applications as it is more computationally efficient and outperforms the standard EnKF for smaller ensemble sizes.

In this study, we use monthly assimilation, assimilating monthly averages of observations on the 15th day of each month. The model state is instantaneously updated based on all observations . This update is split into two steps: firstly, ocean model variables and the multi-category ice concentration per unit grid cell are updated based on the anomaly assimilation of the ocean and SIC variables. In the second step, the multi-category ice concentration per unit grid cell is updated with full-field assimilation of SIT observations.

In NorCPM, anomaly assimilation is employed for ocean and sea ice concentrations because the ocean and sea ice concentrations (that are tightly linked) are strongly attracted to their bias state, leading to forecast drift and a transfer of bias into non-observed quantities via ensemble covariance . Please note that the rationale for anomaly assimilation in NorESM has been described and verified in previous papers . We do not see that the assimilation causes artefact or instability, which we relate to the stabilizing properties of the Ensemble Kalman Filter. Unlike for sea ice concentration, the SIT bias can take up to a decade to re-emerge . Therefore, we use full-field assimilation in this study to correct the SIT bias, which, in turn, can influence the system's variability. This two-step method of anomaly and full-field assimilation has already been employed successfully and produces stable performance .

The assimilation of ocean temperature and salinity profiles, sea surface temperature (SST), SIC and SIT observations is performed as described in , though the observational datasets for SST, SIC and SIT differ. Strongly coupled ocean-sea ice data assimilation is performed, meaning that both ocean and sea ice observations are used to update both components . The full ocean model state vector is updated in isopycnal coordinates and the full multi-category ice concentration is updated. However, the sea ice volume (SIV) per unit grid cell (“vicen” in CICE) is estimated so that the thickness of each thickness category remains identical to that of the prior. This prevents the need to reshuffle ice to a different thickness category in the post-analysis, which counteracts the emergence of a drift and has been shown to be optimal in an idealised twin experiment . Similarly, the energy budget for each multi-category sea ice quantity is recomputed .

To handle sampling errors that can cause significant problems in ensemble-based data assimilation systems we use a series of ad-hoc techniques that are available and widely tested in NorCPM . We use the R-factor inflation method, in which the observation error is inflated by a factor of 4 (i.e. R-factor = 4) when updating the ensemble anomaly . This prevents spurious deflation of the ensemble spread. We also use the K-factor formulation, in which the observational error is artificially inflated if the assimilation pushes the update beyond twice the ensemble spread (K-factor = 2) . This reduces the risk associated with overly strong updates (e.g., from erroneous observations, or a model not able to cope with real variability). Lastly, we use localisation , whereby only local observations are used in the analysis update to limit spurious long-range correlations. We define a localisation radius of 800 km for the sea ice model and use a localisation radius dependent on latitude for the ocean – varying between 1500 and 2300 km. A smooth distance-weighted Gaspari and Cohn function is used to ensure continuity in the analysis update.

In this study, we assume that the observation errors are independent of each other. The observation errors of the SST and SIC products may, however, be correlated due to the post-processing in the production of the final gridded product. We mitigate this in NorCPM by only using the nearest observation of SST, SIC and SIT respectively in the local analysis . All hydrographic profile data within each local window are used .

Hydrographic data profiles of temperature and salinity are taken from the EN4.2.2 dataset . The data in this dataset are split into different categories depending on their quality . We use only the highest quality data (category 1) in this study. The observation errors used are as in , accounting for instrumental and representation error.

SST and SIC monthly average data are from the UK Met Office operational sea surface temperature and sea ice analysis (OSTIA) dataset . All the satellite data used in this dataset is sourced from the EUMETSAT’s Ocean and Sea Ice Satellite Application Facility (OSISAF) . SST is sourced from the ESA SST CCI dataset , while SIC is sourced from the OSI-450, OSI-430 and OSI-430-b datasets . In-situ SST data from drifting and moored buoys sourced from the HadIOD version 1.2.0.0 dataset are additionally used to produce the OSTIA dataset. The SST and SIC data are synchronised for the product, and no SST data is assimilated in grid cells where sea ice is present. The OSTIA dataset is available on a 0.05° × 0.05° horizontal grid, a much higher resolution than our model and likely higher than its effective resolution. We have used superobing , so that the observation estimate is the average of all observations within that grid cell, and the error is the harmonic mean of the daily error estimate from the provider. However, it has been found necessary to inflate the observation error standard deviation by a factor of three, to sustain good reliability in our ensemble and have an observation estimate within the same ballpark as other products (e.g. HadISST2 used in NorCPM previously, ). Furthermore, we impose that the observation error standard deviation be no lower than 0.1 °C. SIC observation errors are not provided with the dataset, so we use a 20 % constant value , which is the generally agreed upon value for the highest uncertainties in the summer melt season .

SIT data are from the LEGOS dataset . We assimilate both the Arctic and the Antarctic data, though we do not present any analysis of Arctic results in this study. In the Arctic, where melt ponds obfuscate radar returns, the dataset is only available in the winter months (October–April). This problem is reduced in Antarctica, because snow over sea ice is much thicker there , which prevents melt pond formation and its impact on SIT retrieval. As such, the data is available year-round from April 1994 to June 2023. LEGOS SIT is the first observation dataset to provide a consistent spatiotemporal resolution and representation of Antarctic SIT covering such a long period. As such, it represents an ideal dataset for assessing the impact of SIT on reanalysis and prediction over a sufficiently long period to provide robust conclusions. The LEGOS SIT dataset is derived from along-track radar freeboard from ERS-1, ERS-2, ENVISAT and CryoSat-2. ERS-1, ERS-2 and ENVISAT used a different, less well-understood mode to retrieve sea ice freeboard compared with CryoSat-2, which uses synthetic aperture radar. A calibration to correct for biases introduced by the different retrieval methods is done using a neural network . The observational uncertainty is estimated in using an ensemble-based approach. The principal sources of uncertainty in the SIT retrieval (including snow depth, snow density, and ice density) are identified, and multiple in-situ, satellite-derived, and model estimates of these quantities are combined to generate an ensemble estimate of the retrieval uncertainty. However, notable differences remain between independent SIT products, highlighting the substantial uncertainty associated with basin-scale Antarctic SIT retrievals. The deterministic EnKF framework used in NorCPM does not treat the observations as the exact representation of the true sea ice state but uses the prescribed observational uncertainties from LEGOS SIT (and the additional sampling error handling techniques described above in Sect. 2.2) to determine the strength of the observational constraint on the model state. In September, typical uncertainty values for LEGOS SIT range between 0 and 1 m across most of the sea ice cover, reaching up to 2.5 m for the thickest ice in the Weddell Sea. In February, the uncertainty values show slightly greater spatial variability over the remaining sea ice at the end of the melt season but within a similar range.

For this study, we have used the NorCPM setup described in Sect. to produce two new reanalyses and hindcast datasets. One assimilates ocean and SIC, while the other additionally assimilates SIT. The two reanalyses cover the period from January 1994 to December 2023 and have 30 ensemble members. Both reanalyses are branched on 15 January 1994 from the pre-existing NorCPM reanalysis of , which assimilates temperature and salinity profiles from the EN4.2.2 dataset, SST and SIC from the Hadley Center Sea Ice and SST dataset , and is initialised in 1982 from the NorESM ensemble of free runs. In this study, we do not assimilate data in the atmospheric component of NorCPM. A NorESM ensemble of free runs is used in addition, when informative, to identify part of the variability related to forcings (e.g., anthropogenic, solar, and volcanic). Two seasonal hindcast datasets are initialised from the reanalyses every year from 1995 to 2022, in January, April, July and October of each year, each run for a year. Each hindcast uses 10 ensemble members, chosen from the first 10 ensemble members of the respective reanalyses (all members are equally likely with the EnKF). This choice represents a compromise between computational cost and ensemble size, as the seasonal prediction experiments require a large number of integrations from 1995–2022 across multiple initialisation dates and lead times. While a larger ensemble could further reduce sampling uncertainty, the skill metrics presented in this study are computed from the ensemble mean over multiple years and 10 ensemble members, which helps mitigate the impact of ensemble sampling variability. We also see no large discontinuity in the validation metric, suggesting that this ensemble size is sufficient to extract the main patterns in the ensemble mean. Note that while we pick the first 10 members, this choice does not matter because, with the Ensemble Kalman Filter, all members are equally likely to yield the best results. The only difference between the two reanalyses – and as a consequence the two hindcast datasets – is in the SIT observations assimilated in EXP-OCT:

NorESM: A free run of the NorESM model. 30 ensemble members are randomly selected from a stable pre-industrial simulation in 1850 and are then integrated until 2023.

We subdivide the Antarctic into five regions (as in , Fig. ), based on longitude. These are Weddell Sea (60° W–30° W), Indian (-30° W–70° E), West Pacific (70° E–150° E), Ross Sea (150° E–150° W) and the Amundsen–Bellingshausen Seas (150° W–70° W). These five regions are chosen due to their distinct geography (particularly differences in the sea ice cover) and model performance .

The reanalysis and hindcasts of SIC and SIE are evaluated against the monthly averaged OSTIA dataset (Sect. ). The reanalysis and hindcast SIT are validated against LEGOS SIT, SMOS Antarctic SIT dataset v3.3 and the ICESat deep-learning derived SIT . Only SMOS SIT below 1 m is used for validation. The ICESat data from 2003–2009 are excluded because missing days in each coverage period result in incomplete sampling for monthly mean calculations. When comparing and validating against the ICESat-2, SMOS, and LEGOS datasets, we restrict the analysis to grid cells and time periods for which both the model output and the observational data are available. Monthly values are therefore computed only when a sufficient number of daily observations are available to form a representative monthly mean, ensuring that all metrics are evaluated over consistent spatiotemporal subsets of the data.

Various metrics are used to analyse the reanalyses and hindcasts in this study. We define the bias and RMSE as follows: 1bias=∑i=1NWi(xi-yi),2RMSE=∑i=1NWi(xi-yi)2, where N is the total number of data points, Wi is a normalising area-based weight constant – i.e. when the metric is computed considering grid cells that do not have the same area. For point-wise bias and RMSE calculations, Wi is 1N. xi is the model ensemble mean values, and yi is the observed values, which are both anomalies from their respective climatology (model or observation). These are usually averaged spatially or temporally (where each grid cell is then weighted by its area).

We also use the mean squared skill score (MSSS), which quantifies the reduction in error relative to the free run. MSSS is computed using the mean squared error (MSE), and defined as 3MSSS=1-MSEforecastMSEreference, where MSE is defined as 4MSE=1n∑i=1n(x-y)2, where x and y are the model and observation values respectively. One benefit of MSSS is that it normalises errors by their pointwise variability and, as such, can help visualise improvements that would otherwise be masked when the amplitude of the error is smaller than in other regions.

The anomaly correlation coefficient (ACC) is also used to test the skill in reproducing the observed variability in the reanalyses and hindcasts. The ACC is defined as 5ACC=∑i=1Nxi′yi′∑i=1Nxi′2∑i=1Nyi′2, where xi′ and yi′ are model and observation values. In this study, we always analyse detrended ACC values, i.e., detrending the time series of the experiments and the observations before calculating the ACC. The trend is usually removed because it is considered to be easy to predict . We test the statistical significance of the ACC using a Student's t-test at a 5 % significance level (95 % confidence level), with degrees of freedom calculated as in .

The final metric used is the integrated ice edge error (IIEE) . This metric is a stronger indicator of sea ice edge location skill than SIE. IIEE is defined as follows: 6IIEE=∫Amax(cx-cy,0)dA+∫Amax(cy-cx,0)dA, where A is the area, c=1 where the SIC is above 15 % and 0 elsewhere, with subscripts x and y denoting the model and observations respectively. IIEE then, is a sum of all areas where SIE is overestimated (first term), and underestimated (second term). IIEE is useful for ice-edge location skill because SIE only accounts for the overall values.

We first present results from the two reanalyses, starting with an analysis of the seasonal cycle of SIE and SIV, followed by a detailed analysis of SIT with the assimilated and independent datasets, and followed by an analysis of the SIC trend and then the interannual variability.

The differences in the seasonal cycle of SIE between the two reanalyses are negligible. As we use anomaly assimilation for ocean and SIC, they both show a large bias compared to observations. On the contrary, the seasonal cycle of SIV is strongly different between EXP-OC and EXP-OCT (Fig. b). Antarctic SIV is increased in EXP-OCT in austral winter, and decreased during the rest of the year in comparison with EXP-OC and NorESM, which brings closer agreement with the derived LEGOS SIV. The timing of the peak volume season also differs in EXP-OCT, as ice volume almost reaches its peak already in August.

The sea ice thickness (SIT) bias is substantially reduced in EXP-OCT (Fig. ) when compared to the assimilated LEGOS product as expected. In both February and September, EXP-OCT exhibits smaller overall biases than EXP-OC, particularly in regions of thick ice in the Weddell and Ross Seas, where the observational data is most reliable. The weaker improvements in thinner-ice regions are primarily attributed to larger observational uncertainties in the assimilated sea ice thickness (SIT) products. In contrast, thicker-ice regions exhibit stronger improvements partly because the model tends to overestimate ice thickness there, likely due to underestimated melting processes, leading to larger innovations (i.e., model and observation mismatches) and thus stronger corrections during assimilation. While alternative datasets such as SMOS SIT may offer potential improvements near the ice edge, their shorter temporal coverage limits their suitability for robust assessment at seasonal timescales. In February, EXP-OCT better reproduces the thin, patchy late-summer ice cover, whereas EXP-OC tends to show thicker SIT in the Weddell and Ross Seas than in observations. In September, when sea ice reaches its maximum extent, both simulations underestimate SIT in the interior pack but display positive biases along the ice edge. These positive edge biases coincide with areas where the model simulates a higher SIE than observed (Fig. ), suggesting that part of the SIT bias arises from extent mismatches rather than from intrinsic thickness errors. The extent mismatches are a result of positive SST biases present in NorCPM .

We also evaluate the experiments against the SMOS Antarctic thin SIT dataset and the deep-learning derived ICESat SIT dataset, which are independent SIT products. Both datasets are converted into monthly means and re-gridded onto the NorCPM grid for comparison. SMOS retrievals are limited to austral winter months (April–October) and are generally considered reliable primarily for thin sea ice (0–1 m), with the highest accuracy for ice thicknesses up to approximately 0.5 m; uncertainty increases progressively beyond this range . In contrast, ICESat-2 employs a lidar instrument, which exhibits greater uncertainty at lower ice thicknesses, with accuracy improving as thickness increases. However, its estimates are additionally affected by uncertainties in snow depth, which remain poorly constrained over Antarctic sea ice. While ICESat-2 appears to outperform LEGOS for thicker ice conditions, its relatively short temporal coverage limits the ability to robustly assess its added value for seasonal predictions. Comparisons with ICESat-2, which provides accurate estimates of thicker ice, but covers a more restricted time period, show an interesting complementary analysis.

Although the standalone LEGOS product can exhibit larger errors than EXP-OC when evaluated against independent datasets, its assimilation in EXP-OCT nevertheless leads to a reduction in RMSE overall (Fig. ). This reflects that the data assimilation framework estimates the most likely sea ice state by optimally combining model states with multiple observational constraints, rather than directly reproducing any individual dataset. Differences between LEGOS, ICESat-2, and SMOS are also expected due to their distinct retrieval approaches and differing sensitivities to sea ice and snow properties. EXP-OCT achieves the lowest RMSEs in austral spring and austral autumn, while errors are higher in February, at the end of the melt season, and during winter, when sea ice is thickest and most spatially variable. The largest RMSE reductions relative to ICESat-2 occur outside the winter period. This seasonal dependence likely reflects variations in both model and observational uncertainty. During transition seasons, larger model biases in sea ice thickness provide greater scope for improvement through assimilation. In contrast, during winter, increased variability in snow depth and density introduces larger uncertainties in the conversion from freeboard to thickness, reducing the effectiveness of assimilation increments due to higher observational noise.

SMOS uncertainty increases with ice thickness – becoming substantial above 0.5 m and too large beyond 1 m . Therefore, we further examined model skill by SIT ranges (Fig. 5). EXP-OCT shows improved performance across all thickness bins, with the largest reductions in RMSE for ice between 0.2 and 0.8 m. Improvements are therefore not limited to the ranges where SMOS uncertainty is greatest, indicating that the assimilation enhances SIT representation even within the more reliable SMOS measurements below 0.5 m. However, it should be noted that the remaining RMSE values are still relatively large compared with the observed SMOS thickness, indicating that the system continues to exhibit substantial errors in representing thin sea ice. The main improvement shown here is a relative reduction in error compared with EXP-OC, rather than a strong agreement with the observed thickness values.

Spatial bias analyses further illustrate how SIT assimilation impacts regional patterns. Comparisons with SMOS (Fig. ), are particularly relevant for thin ice. Both reanalyses overestimate SIT along the sea ice edge and near the continental shelf. EXP-OCT substantially reduces this overestimation in May – when thin ice dominates – in the Indian Sector and the Ross Seas, but yields a slight degradation in the Weddell Sea and parts of the West Pacific Sector. In September, at a time when SIT is thicker and despite the overall bias being reduced, the bias reduction is only evident in the West Pacific Sector, and a rather strong degradation is found in the Weddell and Amundsen–Bellingshausen Seas. As with SMOS, EXP-OCT reduces widespread positive biases most in February relative to EXP-OC, with the largest improvements occurring in the Weddell and Ross Seas near the continental shelf. Improvement remains near the continental shelf in September. As with the SMOS comparison, the bias increases near the ice edge in the Weddell and Ross Seas. This corresponds to the region where we observed a clear improvement over LEGOS (Fig. ), suggesting that SIT in LEGOS dataset may be too thick in that area.

Overall, the assimilation of LEGOS SIT yields a more balanced and spatially consistent ice thickness distribution, improving both thin ice and thick ice regimes. The cross-dataset agreement with SMOS and ICESat-2 confirms that the assimilation benefits propagate dynamically, producing a more realistic and physically coherent representation of Antarctic SIT. The assimilation of SIT substantially reduces NorCPM’s SIT biases and is expected to improve SIE prediction skill through the known influence of SIT memory on seasonal prediction skill, particularly in the Arctic and the Weddell Sea . The consistent RMSE reductions across both independent datasets indicate that assimilating LEGOS SIT improves NorCPM’s representation of sea ice conditions rather than merely enforcing the LEGOS SIT variability in our model.

We analyse the capability of the SIT assimilation to improve the SIC trends (Fig. ). EXP-OC does show some agreement with the observed trend, but misses some important details, such as in the Ross Sea where the observation shows a strong positive trend, as well as in other regions where the trend is negative. EXP-OCT is a better match to the observed trends, increasing the pattern correlation from 0.37 to 0.5. In particular, EXP-OCT captures the negative trends observed in the Indian and West Pacific regions, while also exhibiting the strong positive trend seen in the Ross Sea.

The detrended interannual variability of pan-Antarctic SIE and SIV for EXP-OC and EXP-OCT is shown in Fig. at the seasonal minimum and maximum in February and September. MSSS and ACC values for SIE over both the pan-Antarctic and individual regions are given in Table . Results vary substantially between regions and months, with a modest overall increase in pan-Antarctic MSSS and ACC for both February and September. The largest improvements occur in the Weddell, Ross, and Amundsen–Bellingshausen Seas, where sea ice is thickest and LEGOS data are most reliable. In contrast, performance decreases in the West Pacific and Indian sectors, where sea ice is thinnest and uncertainty in LEGOS SIT is greatest. EXP-OC and EXP-OCT show some skill in reproducing the observed variability in the OSTIA SIC observations, though at a lower amplitude. EXP-OC and NorESM generally show minimal interannual SIV variability, whereas EXP-OCT exhibits larger variability, including two notable events: a pronounced increase in 2013–2014, followed by a sharp decrease in 2023. NorESM demonstrates limited skill in reproducing observed variability compared with EXP-OC and EXP-OCT for both SIE and SIV. As expected, the spread in both SIE and SIV is reduced in EXP-OC and EXP-OCT relative to NorESM due to data assimilation reducing uncertainty.

Errors in SIE may compensate for over- and underestimations of the true SIC. The IIEE provides a more robust measure by revealing the net displacement of the predicted ice edge relative to observations. The monthly climatology of pan-Antarctic IIEE (Fig. ) shows that SIT assimilation overall benefits ice edge skill from June to December, with notable improvement towards the end of the year, while from January-March IIEE is increased in EXP-OCT. Surprisingly, the West Pacific where ice is thin and SIT assimilation has a weaker impact, contributes most to the reduced IIEE over the pan-Antarctic. The increased IIEE in EXP-OCT from January–March in the pan-Antarctic is caused by the Indian sector and the Weddell Sea. This increase in IIEE at this time is likely from thicker ice created by the SIT assimilation in the Weddell Sea, some of which advects into the Indian sector during the summer when sea ice velocity is at its highest. The reduction in IIEE during the last six months of the year is promising for the predictions, including in the West Pacific, where NorCPM has previously had poor prediction skill .

We assess Antarctic SIE, SIT and SIC prediction skill using detrended anomaly correlation coefficients (ACC) in Figs. , and .

In the West Pacific sector, SIE prediction skill in Exp-OC is generally low, while SIT skill remains skillful longer – i.e., SIE skill is limited during the first half of the year, while SIT skill can extend until the end of the year from August to December. SIT assimilation greatly improves the prediction skill of SIT, which now extends to a 12-month lead time for all seasons except the austral winter (July–August). This leads to a notable improvement in SIE that now extends up to 9 lead months in the austral summer. A reduction in IIEE during the austral spring (Fig. ) likely contributes to enhanced skill during the subsequent growth season.

In the Ross Sea, the primary benefits of SIT initialisation are not as clear, and there seems to be a degradation for SIT prediction, particularly in austral winter, when a degradation was found with other SIT products (e.g., Fig. ). This region is also challenging, being highly dynamic and strongly seasonal . In the austral summer and autumn (January–May), SIT was showing some improvement in the reanalysis, and the SIT prediction skill is more consistent up to five lead months, which also translates into improved SIE prediction skill during that time. SIE prediction skill from the October initialisation is significantly improved, with skill extending up to 10 months.

EXP-OCT shows good improvement over EXP-OC for both SIE and SIT skill in the Amundsen and Bellingshausen Seas, with higher ACC values sustained over longer lead times, particularly for winter and spring targets. SIT initialisation improves the representation of initial ice volume and melt timing, leading to improved prediction performance. EXP-OCT exhibits stronger SIC skill improvement in January from October initialisation (Fig. ). However, the increase in Ross Sea SIE skill does not correspond with higher SIT skill, which may suggest that improved sea ice conditions in the adjacent Amundsen Sea then later advect into the Ross Sea , which improves SIC predictions without improving local SIT skill. Transported or newly grown ice may melt, deform, or interact with existing ice in ways not accurately represented by the model.

The Weddell Sea (Fig. ) exhibits the most pronounced increase in SIT prediction skill thanks to SIT assimilation. SIE improvements coincide well with the seasonal timing of the skill but are slightly lower than for SIT. EXP-OCT extends significant SIE prediction skill by approximately three months from the January start time. Figure shows ACC increases in both SIC and SIT in the Weddell Sea for up to nine lead months, though SIC ACC weakens near the ice edge, consistent with smaller basin-wide SIE skill gains. identified atmospheric initialisation as an important mechanism during austral spring in the Weddell Sea, whereas we demonstrate that SIT initialisation is the key mechanism governing SIT predictability during austral summer and autumn.

In the Indian Ocean sector, the impact from SIT is marginal. SIT is only slightly improved on the October forecast, and the prediction skill is modest. SIE skill is degraded from the January start time in EXP-OCT. The thinner, more variable sea ice cover and stronger oceanic influence in this region likely limit the impact of SIT initialisation, especially in summer when there is very little ice cover in this region.

Basin-wide, EXP-OCT produces higher and more persistent ACC values than EXP-OC for SIT, although many differences are not statistically significant. Predictive skill increases markedly from the October initialisation, with both SIT and SIE skill remaining significant at lead times approaching one year. In January, there is an additional SIE skill for up to three months, though SIT improvement is less clear.

Regions showing improved SIT skill generally exhibit corresponding SIE improvements, indicating that assimilating SIT strengthens the model’s initialisation of ice volume. This enhances the persistence of sea ice anomalies and improves both SIE and SIT forecasts. At short lead times, both experiments perform similarly; however, at longer leads, predictive skill in EXP-OC decays more rapidly, while EXP-OCT retains skill in both SIC and SIT. As shown in Fig. , EXP-OCT exhibits strong initial SIT skill across the Antarctic, with SIC skill increases that persist at longer lead times from October. SIT and SIC prediction skills in EXP-OCT decline slightly toward the end of the melt season, but typically strengthen again in autumn and persist through winter.

In both EXP-OC and EXP-OCT, a substantial fraction of statistically significant ACC at the shortest lead times (0–2 months) does not exceed persistence, indicating that forecast skill can be largely attributed to the inherent memory of the sea ice system. However at longer lead times EXP-OCT shows more regionally and temporally localised improvements in SIE and SIT prediction skill, which are significant and exceed persistence. This demonstrates that the assimilation of SIT enhances NorCPM’s sea ice prediction skill beyond persistence at longer lead times.

Overall, the results demonstrate a robust link between SIE and SIT predictive skill. Parallel improvements are evident – for example, during austral spring in the Amundsen–Bellingshausen Seas, austral autumn in the West Pacific, and in pan-Antarctic forecasts from the October initialisation. Possible evidence for a prediction barrier for SIT emerges in August–September, preceding a corresponding barrier for SIE in the austral spring. This suggests that inadequate SIT estimates at the end of winter are associated with poor SIC and SIE estimates at the end of summer. However validation of the reanalysis for SIT with independent SIT products indicates that LEGOS SIT may have increased SIT too much during the peak sea ice volume period which may cause this barrier in our model. Figure further shows great improvements in SIT ACC at initialisation that persist through the forecast period, though they weaken slightly at three months but re-emerge at six lead months. SIC improvements are weaker but remain evident, particularly in the Ross and Weddell Seas.

The West Pacific sector is the only region that departs from this pattern, possibly because LEGOS observations reveal persistent pockets of thick coastal ice that are better represented in EXP-OCT, as indicated by the reduced bias in Fig. . Nevertheless, the model underestimates SIT growth relative to observations, which limits SIT predictive skill at the end of the growth season and subsequently degrades SIE prediction skill through spring and summer in experiments without SIT initialisation.