Dynamical modeling is widely utilized for Antarctic sea ice prediction. However, the relative impact of initializing different model components remains unclear. We compare three sets of hindcasts of the Norwegian Climate Prediction Model (NorCPM), which are initialized by ocean, ocean/sea‐ice, or atmosphere data and referred to as the OCN, OCNICE, and ATM hindcasts hereafter. The seasonal cycle of sea ice extent (SIE) in the ATM reanalysis shows a slightly better agreement with observations than the OCN and OCNICE reanalyzes. The trends of sea ice concentration (SIC) in the OCN and OCNICE reanalyzes compare well to observations, but the ATM reanalysis is poor over the western Antarctic. The OCNICE reanalysis yields the most accurate estimation of sea ice variability, while the OCN and ATM reanalyzes are comparable. Evaluation of the hindcasts reveals the predictive skill varies with region and season. Austral winter SIE of the western Antarctic can be skillfully predicted 12 months ahead, while the predictive skill in the eastern Antarctic is low. Austral winter SIE predictability can be largely attributed to high sea surface temperature predictability, thanks to skillful initialization of ocean heat content. The ATM hindcast from July or October performs best due to the effective initialization of sea‐ice thickness, which enhances prediction skills until early austral summer via its long memory. Meanwhile, the stratosphere‐troposphere coupling contributes to the prediction of springtime. The comparable skill between the OCN and OCNICE hindcasts implies limited benefits from SIC data on prediction when using ocean data.

In this study, the subseasonal Antarctic sea ice edge prediction skill of the Copernicus Climate Change Service (C3S) and Subseasonal to Seasonal (S2S) projects was evaluated by a probabilistic metric, the spatial probability score (SPS). Both projects provide subseasonal to seasonal scale forecasts of multiple coupled dynamical systems. We found that predictions by individual dynamical systems remain skillful for up to 38 days (i.e., the ECMWF system). Regionally, dynamical systems are better at predicting the sea ice edge in the West Antarctic than in the East Antarctic. However, the seasonal variations of the prediction skill are partly system‐dependent as some systems have a freezing‐season bias, some had a melting‐season bias, and some had a season‐independent bias. Further analysis reveals that the model initialization is the crucial prerequisite for skillful subseasonal sea ice prediction. For those systems with the most realistic initialization, the model physics dictates the propagation of initialization errors and, consequently, the temporal length of predictive skill. Additionally, we found that the SPS‐characterized prediction skill could be improved by increasing the ensemble size to gain a more realistic ensemble spread. Based on the C3S systems, we constructed a multi‐model forecast from the above principles. This forecast consistently demonstrated a superior prediction skill compared to individual dynamical systems or statistical observation‐based benchmarks. In summary, our results elucidate the most important factors (i.e., the model initialization and the model physics) affecting the currently available subseasonal Antarctic sea ice prediction systems and highlighting the opportunities to improve them significantly.

Sea ice thickness (SIT) presents comprehensive information on Arctic sea ice changes and their role in the climate system. However, our understanding of SIT is limited by a scarcity of observations and inaccurate model simulations. Based on simultaneous parameter optimization with a micro genetic algorithm, the North Atlantic/Arctic Ocean–Sea Ice Model (NAOSIM) has already demonstrated advantages in Arctic sea ice simulations. However, its performance in simulating pan-Arctic SITs remains unclear. In this study, a further evaluation of Arctic SITs from NAOSIM was conducted based on a comparison with satellite and in situ observations. Generally, NAOSIM can reproduce the annual cycle and downward trend in the sea ice volume. However, deficiencies can still be found in the simulation of SIT spatial patterns. NAOSIM overestimates the SIT of thinner ice (1.5 m) in the central Arctic and is unable to capture the upward trend in the SIT in the north of the Canadian Archipelago as well as to reproduce the intensity of the observed SIT variability. In terms of SIT simulation, NAOSIM performs better as the time approaches the optimization window (2000–2012). Therefore, in the context of rapid changes in Arctic sea ice, how to optimize this model based on limited observations still remains a challenge.

Antarctic sea ice is an important part of the Earth's system and provides key habitats for wild animals. This study assesses the variability and predictability of regional Antarctic sea ice, particularly focusing on the impact of the initialization of different components on its seasonal predictions. We run three hindcasts (retrospective forecast) experiments within the Norwegian Climate Prediction Model (NorCPM), whose atmosphere, ocean, or sea ice is initialized, respectively. These hindcasts are conducted four times per year over 1985-2010 and last for 13 months. We first evaluate three reanalyses that provide initial conditions for the hindcasts. It shows that solely constraining the ocean or atmosphere cannot reproduce the overall observed sea ice concentration (SIC) trend, but has some skill in capturing the variability of the SIC and Antarctic sea ice extent (SIE). Sea ice assimilation further improves the estimate of the SIC/SIE trend and variability, but the performance in the Pacific section is degraded. According to the hindcast experiments, the predictive skill varies with region and season. For example, winter SIE in the Weddell Sea and Amundsen-Bellingshausen Sea can be skillfully predicted 12 months in advance and the predictive skill in the Pacific Section is lower. Among the three hindcast experiments, atmosphere initialization generally yields comparable or even more prolonged prediction skills when compared to ocean or ocean/sea-ice initialization. Compared to ocean initialization, additional sea ice initialization improves prediction in the Indian Section, Pacific Section, and the Ross Sea, but degrades in the Amundsen-Bellingshausen Sea. Further analysis demonstrates that a large part of regional SIE predictability can be explained by high SST predictability on seasonal timescale. In addition, sea ice thickness plays a key role in prolonging the prediction skill in the Ross Sea until the summer season due to the memory of thick ice.