3.1 A Robust and Sustained Pause in Arctic Sea Ice Loss

We first investigate the recent observed trends in Arctic sea ice cover, focusing on the annual minimum, September. The trend of September Arctic sea ice extent for the most recent two decades 2005–2024 is −0.35 and −0.29 million  per decade according to the NSIDC and OSISAF sea ice indices respectively (Figures 1a and 1b). The key point, we emphasize, is that these trends are not statistically significantly different from zero at a 95% confidence level. This is also seen in Figures 1c and 1d where 20-year trends are plotted versus the end year: note how trends ending in 2024 retreat inside the uncertainty envelope. According to the OSISAF record, the 2005–2024 trend is the slowest rate of sea ice area loss over any 20-year period since the start of the satellite record, and second slowest for the NSIDC record. For both data sets, this insignificant trend is four to five times smaller than the peak 20-year sea ice loss trend recorded (1993–2012). These results are robust to the choice of sea ice area or sea ice extent (Figure S1 in Supporting Information S1). The slowdown in September sea ice loss mainly occurs in the Pacific and Eurasian sector, from the Beaufort Sea westward to the Barents Sea (not shown).

While sea ice loss in September is of particular interest because that month is the annual minimum, the current pause in Arctic sea ice loss is seen in every single month throughout the year, as shown in Figures 1e and 1f. This suggests that the underlying mechanism(s) must explain not just the summer trends (R. Zhang, 2015; Francis & Wu, 2020) or winter trends (Yeager et al., 2015), but sea ice trends throughout the entire year. We speculate that a negative sea ice thickness feedback and changes in ocean heat transport convergence near the sea ice edge could play important roles (Bitz et al., 2005; Oldenburg et al., 2024; Polyakov et al., 2023; Sumata et al., 2023).

The same picture—indicating a severe slowdown in Arctic sea ice loss—also emerges when considering sea ice volume. The loss of Arctic sea ice volume has stalled for at least the past 15 years (Figure S2 in Supporting Information S1), consistent with previous work (J. Zhang, 2021). For the period 2010–2024, the simulated annual mean Arctic sea ice volume has an approximately flat trend, decreasing by only 0.4 million  per decade, a value that is seven-times smaller than the long-term simulated loss for the period 1979–2024 of 2.9 million  per decade, and again is not statistically significant. This result is most evident in the Barents Sea (Onarheim et al., 2024).

Various aspects of this pause in Arctic sea ice loss have been documented in recent studies (Polyakov et al., 2023; Stern, 2025; Sumata et al., 2023), building a consistent picture of a thinner ice cover which has remained basically unchanged since the late 2000s. Altogether, strong observational evidence reveals a sustained and pervasive pause in Arctic sea ice loss over the recent 15–20 years, highly robust to the choice of sea ice metric, observational product, and season, and we are led to ask: is such a pause unexpected? To answer that question we turn to analyzing comprehensive climate model simulations. We seek to determine if they are able to capture pauses such as the observed one and, if so, to establish if such pauses are exceedingly rare or relatively frequent events.

3.2 Comprehensive Climate Models Suggest 20-Year Pauses Are Not Rare

To understand whether comprehensive climate models can simulate a multi-decadal pause of Arctic sea ice loss, we search through the CMIP5 and CMIP6 large ensemble archive to identify members which exhibit ice loss pauses. Consistent with previous studies (Kay et al., 2011; Lee & Liu, 2023; Swart et al., 2015), we find that nearly all models can simulate reductions in September Arctic sea ice area smaller than observed during the period 2005–2024. The two models which do not feature any such trends, UKESM1-0-LL and CanESM5-1, are both models with large climate sensitivities (Meehl et al., 2020), for which overly strong anthropogenically-forced sea ice loss does not allow for pauses such as the observed one.

Figure 2a shows the percentage of members with less sea ice loss than observed. The main result here is that the multi-model average suggests a nearly 20% chance of this pause in Arctic sea ice loss (Figure 2a, column 1). However, we note a large spread across the CMIP5 and CMIP6 models, with the probability of a smaller-than-observed 2005–2024 trend varying from 0% and approximately 40%.

The central estimated value of nearly 20% doesn’t change substantially if models are selected following the criteria from Notz and SIMIP Community (2020), or according to their climate sensitivity (Sherwood et al., 2020), or using models according to their ability to simulate climatological sea ice conditions, or how sensitive they are to the biomass burning variability in the CMIP6 forcing (DeRepentigny et al., 2022) (Figure 2a, column 2–5). The central estimate is only slightly higher if we assess the probability of a non-statistically significant trend (Figure S3 in Supporting Information S1). Therefore from the multi-model perspective, what we have observed in the Arctic over the past two decades is not a rare event, but rather one that should be expected to happen with reasonable frequency. This result is insensitive to how models are sub-selected, to the metric of interest, or to the choice of sea ice extent or area (compare Figure S4 in Supporting Information S1 and Figure 2). This then raises the question of whether this pause in Arctic sea ice loss could continue and for how long.

3.3 The Observed Pause in Sea Ice Loss Could Foreseeably Continue for Another Decade

To investigate how much longer this current pause is likely to last into the future, we now examine those large ensemble members which do exhibit minimal sea ice loss in the period from 2005 to 2024 (Figure 2a), and quantify how long the simulated slowdowns persist in the future (Figure 2b). The multi-model average suggests that pauses in September sea ice loss for the period 2005–2024 have a nearly 1 in 2 chance of persisting for a further 5 years, and a 1 in 4 chance of persisting for a further 10 years (note, however, the considerable intermodel spread of 25%). Higher emissions scenarios typically tend to show slightly lower probabilities of sustaining the muted pace of Arctic sea ice loss in the future.

On average, the sea ice extent in ensemble members which simulate pauses in sea ice loss for the period 2005–2024 is 0.6 million  larger in 2025 than in ensemble members from the same models and scenarios in which there is no pause (Figure 2c). This source of predictability mostly decays within a decade in the multi-model mean, and after that the September sea ice area in ensemble members with and without pauses are indistinguishable from each other.

To ensure robust estimates, we included only models with at least five ensemble members simulating sea ice pauses during 2005–2024. This selection may bias our results toward models with larger ensembles or those more frequently simulating pauses. The probability of a sea ice loss pause in this subset, closer to 25%, is slightly higher than the estimate from all models. Thus, if the observed slowdown is in fact rare, our approach could overestimate its persistence. Nonetheless, over two-thirds of suitable models are included, and our findings align with previous research indicating that multi-decadal pauses in sea ice loss remain plausible in the late 21st century under medium emissions scenarios (Swart et al., 2015).

Whilst we have outlined the chance of the pause in sea ice loss persisting, all simulations suggest that sea ice conditions will inevitably revert back toward the longer term forced sea ice loss trend. In fact, there is substantial agreement among the model simulations that, for the five year period following the identified pause (2024–2028), members with pauses exhibit a more rapid loss of sea ice extent than members without pauses (comparing red and blue markers in Figure S5 in Supporting Information S1). The multi-model average indicates that sea ice loss could be 0.6 million  per decade faster than the broader long-term decline.

Taken together, the wealth of available CMIP5 and CMIP6 simulations suggest it is possible that the present pause in sea ice decline may continue for a further 5–10 years, which would be consistent with timescales of Pacific ocean variability (Screen & Deser, 2019). If that were the case it may then imply the occurrence of an early ice-free Arctic is less likely than raw model output would suggest (Arthun et al., 2020; England & Polvani, 2023; Jahn et al., 2016, 2024; Wang et al., 2021). However, we find that the existence of this slowdown also predisposes the sea ice cover for a more rapid decline in the near future.

3.4 Climate Models Suggest an Important Role for Climate Variability

Whether the present slowdown persists in the future or not, one final question remains to be answered: is the recent pause a response to anthropogenic forcings alone, or is there an important role for internal climate variability? When attempting to isolate the forced component of any observed trend from internal variability, it is important to keep in mind that—assuming the model simulations faithfully capture a plausible reality—the observations are expected to have the same broad features as individual ensemble members, that is that they are a combination of a forced trend plus one particular realization of internal variability (although one doesn’t expect the observations to precisely match any one member). Given the well-established importance of internal climate variability in Arctic sea ice trends (Dörr et al., 2023; England et al., 2019; Kay et al., 2011; Swart et al., 2015), we next assess whether a change in the forced response could also be substantially contributing to the observed pause in ice loss.

First, we show the forced September sea ice loss for the period 2005–2024 as estimated by the linear trend of the ensemble mean for each model and scenario (Figure 3, horizontal axis). We find there are only two model/scenario combinations (GFDL-ESM2M RCP8.5 and CESM2 SSP3-7.0) for which the forced trend is estimated to entirely explain observed trends in ice loss (shown in Figure 3, as gray vertical lines) with minimal role for internal variability. Over 85% of the models we analyze here have a larger forced sea ice loss for this period than observed (as they lie to the left of the observed trends), implying that internal variability has acted to reduce the pace of ice loss.

Second, we ask: is there evidence that the forced response itself is slowing down relative to the previous two decades? On the vertical axis of Figure 3, therefore, we plot the ratio of the 2005–2024 forced trend to that of the preceding 20 years, 1986–2005, for each model and scenario combination: a ratio of 1 indicates no change in the pace of ice loss,  indicates an acceleration and  indicates a deceleration. Again we find that only GFDL-ESM2M RCP8.5 and CESM2 SSP3-7.0 suggest that the reduction in the forced trends accounts entirely for the observed slowdown. While the results from all the other models agree that this observed pause is not entirely a forced response, the remaining models are relatively evenly split (Figure 3): roughly half the models suggest that the forced sea ice loss trend has modestly decelerated over the past two decades relative to the prior two decades, and roughly half suggest it has modestly accelerated. Examining the ensemble mean sea ice cover in each model (Figure S6 in Supporting Information S1) confirms that most models show no discernible evidence for a forced slowdown over the last two decades.

In summary then: while the modeling evidence is uncertain as to whether anthropogenic forcings account—even in part—for the recent pause in Arctic sea ice loss, it is very likely that internal climate variability is contributing to the slowdown in an important way.