IPCC AR5 WG1 ch12 on Long-term Arctic sea ice decline:
http://www.climatechange2013.org/images/report/WG1AR5_Chapter12_FINAL.pdf"In the NH, in accordance with CMIP3 results, the absolute rate of decrease of the CMIP5 multi-model mean sea ice areal coverage is greatest in September.
The reduction in sea ice extent between the time periods 1986–2005 and 2081–2100 for the CMIP5 multi-model average ranges from 8% for RCP2.6 to 34% for RCP8.5 in February and from 43% for RCP2.6 to 94% for RCP8.5 in September. Medium confidence is attached to these values as projections of sea ice extent decline in the real world due to errors in the simulation of present-day sea ice extent (mean and trends—see Section 9.4.3) and because of the large spread of model responses. About 90% of the available CMIP5 models reach nearly ice-free conditions (sea ice extent less than 1 × 10^6 km2 for at least 5 consecutive years) during September in the Arctic before 2100 under RCP8.5 (about 45% under RCP4.5). By the end of the 21st century, the decrease in multi-model mean sea ice volume ranges from 29% for RCP2.6 to 73% for RCP8.5 in February and from 54% for RCP2.6 to 96% for RCP8.5 in September. Medium confidence is attached to these values as projections of the real world sea ice volume. In February, these percentages are much higher than the corresponding ones for sea ice extent, which is indicative of a substantial sea ice thinning.
A frequent criticism of the CMIP3 models is that, as a group, they strongly underestimate the rapid decline in summer Arctic sea ice extent observed during the past few decades (e.g., Stroeve et al., 2007; Winton, 2011), which suggests that the CMIP3 projections of summer Arctic sea ice areal coverage might be too conservative. As shown in Section 9.4.3 and
Figure 12.28b, the magnitude of the CMIP5 multi-model mean trend in September Arctic sea ice extent over the satellite era is more consistent with, but still underestimates, the observed one (see also Massonnet et al., 2012; Stroeve et al., 2012; Wang and Overland, 2012; Overland and Wang, 2013). Owing to the shortness of the observational record, it is difficult to ascertain the relative influence of natural variability on this trend. This hinders the comparison between modelled and observed trends, and hence the estimate of the sensitivity of the September Arctic sea ice extent to global surface temperature change (i.e., the decrease in sea ice extent per degree global warming) (Kay et al., 2011; Winton, 2011; Mahlstein and Knutti, 2012). This sensitivity may be crucial for determining future sea ice losses. Indeed, a clear relationship exists at longer than decadal time scales in climate change simulations between the annual mean or September mean Arctic sea ice extent and the annual mean global surface temperature change for ice extents larger than ~1 × 10^6 km2 (e.g., Ridley et al., 2007; Zhang, 2010b; NRC, 2011; Winton, 2011; Mahlstein and Knutti, 2012). This relationship is illustrated in Figure 12.30 for both CMIP3 and CMIP5 models. From this figure, it can be seen that the sea ice sensitivity varies significantly from model to model and is generally larger and in better agreement among models in CMIP5...
Today, the optimal approach for constraining sea ice projections from climate models is unclear, although one notes that these methods should have a credible underlying physical basis in order to increase confidence in their results (see Section 12.2). In addition, they should account for the potentially large imprint of natural variability on both observations and model simulations when these two sources of information are to be compared (see Section 9.8.3). This latter point is particularly critical if the past sea ice trend or sensitivity is used in performance metrics given the relatively short observational period (Kay et al., 2011; Overland et al., 2011; Mahlstein and Knutti, 2012; Massonnet et al., 2012; Stroeve et al., 2012)....
[V]arious methods all suggest a faster rate of summer Arctic sea ice decline than the multi-model mean. Although they individually provide a reduced range for the year of near disappearance of the September Arctic sea ice compared to the original CMIP3/CMIP5 multi-model ensemble, they lead to different timings (Overland and Wang, 2013). Consequently, the time interval obtained when combining all these studies remains wide: 2020–2100+ (2100+ = not before 2100) for the SRES A1B scenario and RCP4.5 (Stroeve et al., 2007, 2012; Boé et al., 2009b; Wang and Overland, 2009, 2012; Zhang, 2010b; Massonnet et al., 2012) and 2020–2060 for RCP8.5 (Massonnet et al., 2012; Wang and Overland, 2012). The method proposed by Massonnet et al. (2012) is applied here to the full set of models that provided the CMIP5 database with sea ice output. The natural variability of each of the four diagnostics shown in Figure 12.31a–d is first estimated by averaging over all available models with more than one ensemble member the diagnostic standard deviations derived from the model ensemble members. Then, for each model, a ±2 standard deviation interval is constructed around the ensemble mean or single realization of the diagnostic considered. A model is retained if, for each diagnostic, either this interval overlaps a ±20% interval around the observed/reanalysed value of the diagnostic or at least one ensemble member from that model gives a value for the diagnostic that falls within ±20% of the observational/reanalysed data. The outcome is displayed in
Figure 12.31e for RCP8.5.
Among the five selected models (ACCESS1.0, ACCESS1.3, GFDL-CM3, IPSL-CM5A-MR, MPI-ESM-MR), four project a nearly ice-free Arctic Ocean in September before 2050 (2080) for RCP8.5 (RCP4.5), the earliest and latest years of near disappearance of the sea ice pack being about 2040 and about 2060 (about 2040 and 2100+), respectively. It should be mentioned that Maslowski et al. (2012) projected that it would take only until about 2016 to reach a nearly ice-free Arctic Ocean in summer, based on a linear extrapolation into the future of the recent sea ice volume trend from a hindcast simulation conducted with a regional model of the Arctic sea ice–ocean system. However, such an extrapolation approach is problematic as it ignores the negative feedbacks that can occur when the sea ice cover becomes thin (e.g., Bitz and Roe, 2004; Notz, 2009) and neglects the effect of year-to-year or longer-term variability (Overland and Wang, 2013). Mahlstein and Knutti (2012) encompassed the dependence of sea ice projections on the forcing scenario by determining the annual mean global surface warming threshold for nearly ice-free conditions in September. Their best estimate of ~2°C above the present derived from both CMIP3 models and observations is consistent with the 1.6 to 2.1°C range (mean value: 1.9°C) obtained from the CMIP5 model subset shown in Figure 12.31e (see also Figure 12.30b).
The reduction in September Arctic sea ice extent by the end of the 21st century, averaged over this subset of models, ranges from 56% for RCP2.6 to 100% for RCP8.5. In light of all these results, it is very likely that the Arctic sea ice cover will continue to shrink and thin all year round during the 21st century as the annual mean global surface temperature rises. It is also likely that the Arctic Ocean will become nearly ice-free in September before the middle of the century for high GHG emissions such as those corresponding to RCP8.5 (medium confidence)."