We analyze data from the first ensemble member of 26 CMIP6 models (see Table ) that were chosen based on the availability of the sea ice variable SIE, sea ice concentration (SIC), and SIV. The nominal horizontal resolution of the ocean or sea ice component from the different CMIP6 models varies between 25 and 250 km, with the majority using a resolution of 100 km (Table ). We use output from historical simulations, which cover the period 1850 to 2014, except for the EC-Earth3 large ensemble spanning the period from 1970 to 2014. We also employed two sets of climate projections following low- and high-warming scenarios, specifically SSP1-2.6 and SSP5-8.5, which correspond to top-of-atmosphere radiative forcings in 2100 of 2.6 and 8.5 Wm-2 with respect to preindustrial levels . Under SSP1-2.6, the Arctic SIE continues to decline in the earlier decades of the 21st century before stabilizing towards the latter part of the century (Fig. S1 in the Supplement). The climate projections cover the period 2015 to 2100, except for the CAMS-CSM1-0 model (model no. 4 in Table ), which ranges from 2015 to 2099. We only use the years covered by all models and, as such, our study focuses on the time period 1970–2099.

In addition to the CMIP6 multimodel ensemble that allows for a detailed investigation of model uncertainty, we analyze historical and SSP5-8.5 simulations from five large ensembles to better understand the role of internal climate variability in our results: ACCESS-ESM1.5 40 ensemble members;, CanESM5 25 ensemble members;, EC-Earth3 50 ensemble members;, MIROC6 50 ensemble members;, and MPI-ESM1.2-LR 30 ensemble members;. These large ensembles and their ensemble size were chosen based on the availability of sea ice variables. Using multiple large ensembles offers a robust comparison of forced responses and internal climate variability across models .

The primary sea ice output used in this study is the Arctic SIE, labeled siextentn in the CMIP6 output. In cases where siextentn was unavailable, we computed the SIE time series using the SIC data, labeled siconc in the CMIP6 output. SIE is calculated as the total area of all grid cells where SIC exceeds 15 %. SIE is a commonly used metric for model comparisons , and our choice of SIE metrics aligns with the existing definitions of RILEs to maintain consistency . However, it is important to note that a limitation of SIE compared to sea ice area (SIA), as highlighted by , is its strong dependency on grid resolution. Additionally, changes in SIA can occur with relatively little change in SIE, which suggests that RILEs defined in terms of SIA may represent fundamentally different processes than those defined using SIE. Nonetheless, we find that our conclusions using SIE are generally consistent with results using SIA (not shown). We also analyzed SIV, labeled sivoln in the CMIP6 output. If SIV was not available, we computed the total Arctic SIV from sea ice thickness (SIT), labeled sivol (grid-cell-averaged ice thickness) or sithick (sea ice thickness averaged over the ice-covered portion of a grid cell) in the CMIP6 output. When only sithick was provided, we calculated SIV by multiplying sithick by SIC and the grid cell area. By taking into account the vertical dimension, the SIV metric offers a more comprehensive representation of the condition of the Arctic sea ice cover as it relates more directly to the thermodynamic processes governing its evolution .

Climate models are powerful tools for analyzing the mean state, trends, and variability of the climate system and how these will evolve in the future. However, the reliability of the conclusions related to sea ice drawn from modeling studies is dependent on the accuracy of the representation of Arctic sea ice and the underlying physical processes embedded within models. To ensure the robustness of our results, we evaluate the performance of the sea ice simulations used in this study with the newly developed SITool . This tool is designed to assess the skill of CMIP6 simulations by comparing various sea ice metrics with observational references. To do so, we rely on observations of SIE and SIC obtained from the NASA Team (NSIDC-0051) dataset and reanalysis of SIT from PIOMAS . Observed SIC and SIT reanalysis data are available from 1979 and, as such, the evaluation of CMIP6 models focuses on the period 1979 to 2014.

SITool reveals noticeable differences between models of the multimodel ensemble in their representation of SIE (Fig. S2 in the Supplement); however, the multimodel mean demonstrates good performance relative to observational data for both March and September (Fig. a). Additionally, we find that the majority of CMIP6 models effectively replicate the mean and variability of SIC, SIE, and SIT as well as the spatial distribution of the ice edge (Figs. S2 and S3 in the Supplement). Note that some models exhibit larger disparities in one or more metrics when compared to observed references: BCC-CSM2-MR, CAMS-CSM1-0, NESM3, EC-Earth3, and MIROC-E2SL. However, we find that screening out these models does not affect our conclusions (not shown) and, therefore, we have retained this subset of models for the analysis. We also conducted the same evaluation of all members of the five large ensembles for SIE and found that ACCESS-ESM1.5 and MPI-ESM1.2-LR demonstrate better performance in reproducing the mean state, standard deviation, and trend of Arctic sea ice extent, while CanESM5, EC-Earth3, and MIROC6 show slightly lower performance (Figs. b and S4 in the Supplement). Specifically, CanESM5 exhibits a particularly negative trend during 1979 to 2014 compared to observations, and MIROC6 shows a less negative trend compared to observations (Fig. S5 in the Supplement) and greatly underestimates SIE from November to June .

Several definitions exist in the scientific literature for RILEs in the Arctic, each emphasizing distinct criteria and temporal characteristics . used the rate of change exceeding a specific threshold, determined through the derivative of the 5-year mean smoothed time series of SIE. Based on this definition, a RILE is identified when sea ice loss surpasses 0.5×106 km2yr-1, with the event's duration based on the period during which SIE decreases by more than 0.15×106 km2yr-1. In contrast, defined rapid sea ice declines based on a period lasting at least 4 years, with the trend in the 5-year running mean minimum SIE consistently lower than or equal to -0.3×106 km2yr-1. characterized a RILE as a drop in summer SIE exceeding 1.2×106 km2. According to their definition, a RILE can manifest itself as a single large drop (“one-step event”) or a series of up to three consecutive steps involving smaller year-to-year drops (“multiyear event”). Finally, assessed rapid ice change events in the Barents Sea using 5-year linear trends of winter (November–April) SIA and the criteria of trends exceeding 2 standard deviations of the distribution of 5-year trends in the Community Earth System Model Large Ensemble (CESM-LE) between 2007 and 2025.

For this study, we use the definition of , for which a RILE is a period lasting at least 4 years, during which the trend in the 5-year running mean minimum SIE is lower than or equal to -0.3×106 km2yr-1. We chose this definition as it emphasizes the total amount of loss during a RILE, with the 5-year running mean filtering out interannual variability, and it limits RILEs to events that last several years rather than single-year events, thus focusing on events with a higher impact on climate, ecosystems, and society. We apply the definition of to all months of the year, maintaining the same threshold. According to this definition, a RILE is even more extreme than the most rapid observed sea ice loss to date. Indeed, over the period 1979–2024, the observed SIE in the Arctic decreased by 0.037×106 km2yr-1 in March and by 0.078×106 km2yr-1 in September , with the most rapid sea ice decline in September over the period 2001–2008 reaching -0.28×106 km2yr-1.

When examining the occurrence of RILEs throughout the year from 1970 to 2099, we find a distinct regime difference between the first and last 6 months of the year (Figs.  and ). Indeed, the characteristics of RILEs (e.g., the total number of RILEs simulated, duration over multiple years, and intraseasonal consistency over several months during 1 year) are noticeably different between winter–spring and summer–fall RILEs under both the high- and low-warming scenarios (Fig. ). From January to June, very few RILEs are simulated by the CMIP6 multimodel ensemble between 1970 and 2050, with an increasing frequency toward the end of the 21st century under the high-warming scenario (SSP5-8.5) in about one-third of the models (Fig. a and b). These winter and spring RILEs also exhibit intraseasonal consistency, meaning that they extend over multiple months of the same year (see the darker colors in Fig. a and b). For the low-warming scenario (SSP1-2.6), only a few RILEs are simulated over the 130 years of our study period, indicating a large contribution from scenario uncertainty to the probability of occurrence of future winter and spring RILEs (Figs. e and f and a). In contrast, between July and December, RILEs are more abundant, though more short-lived, and appear to be randomly distributed throughout the time period when sea ice is present 1970 to consistently ice-free conditions; for both warming scenarios (Fig. c, d, g, and h). We also see a smaller impact of the choice of future scenario on RILEs occurring in the last 6 months of the year, especially for summer RILEs (July–August–September; Fig. c and g). This suggests that forcing factors predominantly influence winter and spring conditions, with little to no role in summer–fall conditions.

This regime difference between winter and summer RILEs is also present in the large ensembles (Fig. ). Additionally, we see a large contribution of model uncertainty to the probability of occurrence of future winter RILEs (Fig. a–e), something that is also apparent in the CMIP6 multimodel ensemble under a high-warming scenario (Fig. a and b). Model uncertainty is reflected in the timing when winter RILEs first occur as well as their intraseasonal consistency for multiple months of the year (Fig. ). In Sect. , we take a closer look at some important characteristics of RILEs that will shed light on the physical processes leading to this model uncertainty.

While the overall pattern reveals an increase in RILE occurrence from late spring through winter, differences emerge across the models (Fig. a). The CanESM5 large ensemble displays a relatively uniform distribution of RILEs throughout the year, with an average number of RILEs per simulation ranging from 2 in March/April to 2.7 in October. In contrast, the EC-Earth3, ACCESS-ESM1.5, MIROC6, and MPI-ESM1.2-LR large ensembles exhibit more pronounced seasonal variability, with a higher occurrence of RILEs from late spring to early winter. The RILE seasonal patterns for the EC-Earth3 and ACCESS-ESM1.5 large ensembles resemble that of the CMIP6 multimodel ensemble for the SSP5-8.5 scenario (Fig. a). On the other hand, the MIROC6 and MPI-ESM1.2-LR large ensembles exhibit a seasonality pattern in RILE similar to the CMIP6 SSP1-2.6 multimodel distribution, even though all large-ensemble analyses here are based on the SSP5-8.5 scenario. The occurrence of RILEs in MIROC6, being similar to the RILE occurrence in the multimodel ensemble under the SSP1-2.6 scenario despite the stronger forcing of the SSP5-8.5 scenario, can be attributed to the relatively weak long-term SIE trend in MIROC6, as shown in Fig. S5. However, the comparison between ACCESS-ESM1.5 and MPI-ESM1.2-LR further underscores the complexity: while the SIE values in both models show similarly weak SIE trends, they differ in their RILE seasonality. This suggests that, while the long-term SIE trend plays a role in determining the seasonality of RILE occurrence, other factors – such as the mean state and internal variability – are also important. For instance, SIE in ACCESS-ESM1.5 has a higher internal variability than MPI-ESM1.2-LR but a similar mean state (Fig. ), which likely contributes to the differences in their seasonal distributions.

Because of the expected increase in sea ice variability as the thickness of the ice cover decreases as well as the extreme sea ice loss associated with RILEs, one could expect an early transition toward consistently ice-free conditions in models that simulate many RILEs. However, we find no clear relationship between RILE occurrence and the timing of consistently ice-free conditions in September, with instances of September ice-free conditions occurring at the end of multiple, few, or no RILEs at all. Indeed, some models from the five large ensembles simulate many RILEs before reaching consistently ice-free conditions in September (e.g., CanESM5, EC-Earth3, and ACCESS-ESM1.5), while others simulate only a few if any (e.g., MIROC6 and MPI-ESM1.2-LR; Fig. ). We also find that some models start simulating winter RILEs immediately after the occurrence of consistently ice-free conditions in September (e.g., CanESM5), while for other models (e.g., ACCESS-ESM1.5, MIROC6, and EC-Earth3) there is a lag of around 20–30 years between the timing of consistently ice-free conditions in September and the onset of winter RILEs. Additionally, some models simulate a large number of winter RILEs across all ensemble members (CanESM5, ACCESS-ESM1.5, and EC-Earth3), whereas only a few winter RILEs are simulated for MPI-ESM1.2-LR and no RILE is simulated in March for MIROC6 (Fig. a). This confirms that uncertainty regarding the initiation of winter RILEs is large and strongly model-dependent in addition to the choice of future scenario, as discussed above.

According to the CMIP6 multimodel ensemble, around 50 % of September RILEs (SRILEs) end at a SIE value under 2×106 km2 and 30 % between 0 and 1×106 km2 (i.e., consistently ice-free conditions) for both warming scenarios (Fig. d). This is also the case for the large ensembles: 18 %–37 % of the SRILEs end below the 1×106 km2 threshold (Fig. d). For MIROC6, the majority of SRILEs have a SIE at around 2.5×106 km2 at their onset, which is the lowest value compared to other large ensembles (Fig. b). Accordingly, 37 % of MIROC6 SRILEs end at a SIE below 1×106 km2 (Fig. d). While the timing of consistently ice-free conditions in September shows little correlation with the occurrence of RILEs, long-lasting SRILEs can directly lead to ice-free conditions, although such events are relatively rare. Specifically, RILEs lasting more than 10 years frequently result in ice-free conditions, but these extended events account for less than 15 % of all RILEs across both the multimodel ensembles and large ensembles. Indeed, most SRILEs typically persist for 4 to 6 years (Figs. b and b). Moreover, this pattern appears consistent across models, suggesting that SRILE duration is not strongly model-dependent.

The probability of having at least one SRILE over the period 1970–2099 in the CMIP6 multimodel ensemble is 92 % for both scenarios, with a maximum of five SRILEs projected during this period for one single simulation (Fig. c). When looking at results from the large ensembles, we find disparities across models in terms of the probability of occurrence of SRILEs. There is a 78 % probability of having at least one SRILE per simulation with the MIROC6 large ensemble, with 46 % of the simulations having only one RILE over the period 1970–2099. In contrast, the EC-Earth3 large ensemble shows a 100 % probability of having at least one SRILE over the same period, with 90 % of the simulations projected to have more than one SRILE (Fig. c). This range of results may be related to differences in SIE mean state, variability, and trends among models (Figs.  and ). This again highlights the important role of model uncertainty and the models' mean state in the probability of occurrence of RILEs.

The percentage of simulations exhibiting a RILE before 2030 was analyzed with the multimodel ensemble under both scenarios (SSP1-2.6 and SSP5-8.5) and large ensembles for the SSP5-8.5 scenario (Figs. d and d). Approximately 60 % of the simulations show a SRILE before 2030, with intermodel differences in the probability. MIROC6 shows a minimum of 26 %, while CanESM5 reaches 92 %, highlighting strong intermodel variability. This large range of probabilities across models shows that a large sea ice model spread remains a concern for CMIP6 models and that analyzing multiple models is crucial for best characterizing the uncertainty inherent in current sea ice projections. While systematic biases in CMIP6 models remain a concern – models can reproduce current sea ice trends for incorrect levels of global warming, as shown by for CMIP5 models – our results provide insights by relying on a multimodel ensemble of 26 models and five large ensembles. As the forcings for the SSP5-8.5 and SSP1-2.6 scenarios remain comparable until 2030, the probability of RILEs occurring before 2030 is similar across multiple models under both the SSP5-8.5 and SSP1-2.6 scenarios. The analysis of large ensembles reveals that models with high SRILE occurrences before 2030 (80 %–92 %; i.e., CanESM5, ACCESS-ESM1.5, and EC-Earth3) also exhibit increased variability in sea ice extent starting in the late 2010s (Fig. a). This enhanced variability increases the likelihood of RILEs before 2030. In contrast, models with lower variability (MIROC6 and MPI-ESM1-2) and an underestimated mean sea ice extent in March (Fig. ) show a lower (26 %–30 %) probability of SRILE occurrence before 2030. While the different SIE interannual variability in models influences the probability of RILEs, their occurrences remain more frequent in summer than in winter, especially from August to October, stabilizing around 60 % in the multimodel ensemble (Fig. d). Outside the summer season, this probability decreases sharply but does not drop to 0 % for the multimodel ensemble, indicating that RILEs, although less frequent, could still occur before 2030 during other months of the year as well. However, there is a clear model dependence in the seasonal distribution of RILEs. For instance, MIROC6 does not project any RILEs before 2030 outside the summer months, suggesting a strong seasonal confinement in this model.

Interestingly, we find an increased probability of SRILE occurrence after a period of no sea ice loss (i.e., a 10-year period with a neutral or positive SIE trend; Fig. ). Indeed, while the overall probability of a RILE occurring in the multimodel ensemble from 2015 until consistently ice-free conditions under SSP5-8.5 is 7 %, the likelihood increases to around 20 % following a period of no sea ice loss. This increase in probability after a decade of neutral or positive trends is consistent across the CMIP6 multimodel ensemble (Fig. ) and the five large ensembles (Fig. S7 in the Supplement), and those differences are all highly significant (z score>5). In addition, the SIV trend during the 10-year period of no trend or a positive trend does not increase the probability of a RILE occurring in the subsequent years (not shown).

SRILEs start occurring in the late 20th century and early 2000s for the CMIP6 multimodel ensemble (Fig. a). By 2025, 50 % of SRILEs have already occurred, and by 2070 all events have taken place for both scenarios. For the large ensembles, the initiation of SRILEs is similar to the CMIP6 multimodel ensemble: 50 % of the total number of SRILEs have already occurred at the earliest by year 2020, as in CanESM5, and at the latest by 2040 for models such as MIROC6 and MPI-ESM1.2-LR (Fig. a). In these models, the timing of SRILEs is consistent with an increase in September SIE variability : sea ice variability in CanESM5 starts to increase around 2010 and peaks around 2025, whereas the peak in SIE variability for MIROC6 and MPI-ESM1.2-LR occurs about 20 years later (Fig. a).

Even though the timing of SRILEs varies by up to 2 decades across models, the peak probability of RILE onset as a function of SIE is quite consistent for both the CMIP6 multimodel ensemble and large ensembles at slightly less than 4×106 km2 (Figs. b and b), which suggests that the dependence of RILE onset on the mean state is similar across models. This is also the SIE at which the large ensembles simulate the largest values of September sea ice variability (Fig. ). EC-Earth3 and CanESM5 show a double peak distribution for SIE at the end of SRILEs (Fig. b and d), which is consistent with early and late RILEs due to high sea ice variability for EC-Earth3 and the early increase in variability (2010–2015) for CanESM5 (Fig. ).

In contrast to September, March RILE (MRILE hereafter) occurrences are mostly simulated after 2050. The peak MRILE occurrence is around 2075 and for SIE values around 11×106 km2 in the CMIP6 multimodel ensemble following the high-warming scenario (Fig. a and b). Of the large ensembles simulating MRILEs (i.e., EC-Earth3, CanESM5, and ACCESS-ESM1.5), the mean SIE at the onset of MRILEs is similar to the CMIP6 multimodel ensemble, except for ACCESS-ESM1.5, for which the peak in MRILE occurrences is around 15×106 km2 (Fig. b). Uncertainty in the timing of MRILEs is more pronounced across these large ensembles (Fig. a), highlighting once more the large contribution of model uncertainty to winter RILEs. This uncertainty can be explained by differences in sea ice mean state and/or variability. For example, although the onset years for winter RILEs differ between CanESM5 and EC-Earth3 (Fig. a), both large ensembles show an increase in sea ice variability as the March SIE falls below 10×106 km2 (Fig. ). In contrast, MPI-ESM1.2-LR exhibits much lower March sea ice variability as a function of both time and SIE (Fig. ), resulting in few winter RILEs (Fig. a). Both ACCESS-ESM1.5 and MIROC6 show an increase in sea ice variability over the last few decades of the 21st century as they reach a lower SIE (Fig. ), resulting in an increase in the occurrence of winter RILEs at that time (Fig. ). The probability density functions of SIE at the end of MRILEs are flatter and wider than the ones for SRILEs (Figs. d and d), which is explained by the large contribution of model uncertainty to the initial SIE as well as the duration and intraseasonal consistency of MRILEs.

A peak of RILE occurrences between 5000 and 7500 km3 of SIV is evident for both September and March (Fig. c) in the CMIP6 multimodel ensemble. This parity in the probability of the initial SIV between September and March suggests that the average ice volume state may serve as preconditioning for RILE occurrences. Indeed, we find that more than 20 % of SRILEs begin at a SIV ranging from 4000 to 6000 km3. The declining mean state of SIV likely drives the similarity in SIV at the onset of SRILEs and MRILEs. By 2060–2080, reduced winter SIV (mean: 9.75×103 km3 in March) approaches early-21st-century summer values (mean: 8.23×103 km3 in September 2000–2020), indicating that future winters will resemble today's summers and contribute to RILEs in all seasons (Fig. S8 in the Supplement). Additionally, SIV differs between these periods: March SIV in 2060–2080 shows lower interannual variability (mean SD ≈1.5×103 km3) than September SIV in 2000–2020 (mean SD ≈2.0×103 km3, range 0.5–4.5×103 km3). This greater summer variability suggests that occasional high summer SIV values in the early 21st century can match low winter SIV levels in the middle to late 21st century.

It is important to note that reaching a critical sea ice state is not a sufficient condition for winter RILEs to occur. MIROC6 simulates no winter RILEs before the last 2 decades of the 21st century (Fig. d) despite showing a less extensive and thinner winter ice cover (Figs. b and S6b in the Supplement). showed that, in addition to a forcing perturbation, an adequately thin ice cover is necessary to initiate RILEs in September, and our results suggest that this finding is also applicable to winter RILEs, except for EC-Earth3. Indeed, EC-Earth3 simulates winter RILEs at the end of the 20th century (Fig. e) without meeting the thin sea ice condition.