Inherent sea ice predictability in the rapidly changing Arctic environment of the Community Climate System Model, version 3
- First Online:
- Cite this article as:
- Holland, M.M., Bailey, D.A. & Vavrus, S. Clim Dyn (2011) 36: 1239. doi:10.1007/s00382-010-0792-4
- 626 Downloads
Seasonal predictions of Arctic sea ice have typically been based on statistical regression models or on results from ensemble ice model forecasts driven by historical atmospheric forcing. However, in the rapidly changing Arctic environment, the predictability characteristics of summer ice cover could undergo important transformations. Here global coupled climate model simulations are used to assess the inherent predictability of Arctic sea ice conditions on seasonal to interannual timescales within the Community Climate System Model, version 3. The role of preconditioning of the ice cover versus intrinsic variations in determining sea ice conditions is examined using ensemble experiments initialized in January with identical ice–ocean–terrestrial conditions. Assessing the divergence among the ensemble members reveals that sea ice area exhibits potential predictability during the first summer and for winter conditions after a year. The ice area exhibits little potential predictability during the spring transition season. Comparing experiments initialized with different mean ice conditions indicates that ice area in a thicker sea ice regime generally exhibits higher potential predictability for a longer period of time. In a thinner sea ice regime, winter ice conditions provide little ice area predictive capability after approximately 1 year. In all regimes, ice thickness has high potential predictability for at least 2 years.
The Arctic environment is undergoing rapid change across the marine, terrestrial, and atmospheric systems (e.g. Serreze et al. 2007; Overland et al. 2004; Stroeve et al. 2007; Francis and Hunter 2007). Perhaps the most dramatic aspect of this change is manifested in the reduction and thinning of the Arctic floating ice cap (Serreze et al. 2007). Near-term forecasts of sea ice conditions provide important information on the marine accessibility of Arctic seas. These serve a number of uses and a number of institutions provide operational forecasts in order to meet these needs.
Many of these forecasting systems use multiple regression models that rely on statistical relationships present in the historical record (e.g. Walsh 1980; Drobot and Maslanik 2002; Drobot et al. 2006; Lindsay et al. 2008). The predictors used in these systems include a mix of information on previous ice (e.g. concentration, albedo, age), ocean (e.g. temperature) and atmospheric (e.g. incoming longwave radiation) conditions. Forecasting techniques that use ensemble simulations from ice–ocean coupled models with prescribed atmospheric forcing from the historical record have also been explored (Zhang et al. 2008a). Additional work has examined the influence of different factors for an extreme September sea ice anomaly using an adjoint of an ice–ocean coupled model (Kauker et al. 2009), and found that the winter/spring ice thickness and summer wind and air temperature variations played a particularly important role.
Recently, the Study of Environmental Arctic Change (SEARCH) program has requested end-of-summer sea ice “outlooks” from the scientific community in an effort to better assess the methods used to provide short-term forecasts of sea ice conditions. These are focused on the expected September sea ice extent minimum at lead times of from 1 to 4 months (although investigators often base their outlooks on the previous spring, winter or summer conditions). Information from the official website (http://www.arcus.org/search/seaiceoutlook/index.php) indicates that there are many methods being explored in these outlooks and that these provide considerably different quantitative forecasts.
The various methods documented in these studies do show some skill in forecasting the summer ice cover at a variety of lead-times. However, in the presence of the rapidly changing Arctic environment, statistical relationships used in these forecasting methods (or diagnosed from an ocean–ice coupled model adjoint; Kauker et al. 2009) may not remain valid. Additionally, using prescribed historical atmospheric forcing in ensemble ice–ocean numerical model forecasts neglects feedbacks to the atmosphere. This has important implications for the forecasts, which may be dependent on the mean climate state. To date, limited work has been done to explore the inherent predictability of Arctic sea ice cover. Regional climate modeling experiments focused on the 1980s and 1990s have elucidated aspects of Arctic ice and near-surface climate predictability and the role of regional versus large-scale atmospheric circulation for interannual Arctic climate variability (Döscher et al. 2009). Additionally, Koenigk and Mikolajewicz (2008) have assessed high latitude climate predictability using ensemble experiments with a global coupled climate model and found that central Arctic ice thickness is highly predictable for up to 2 years due to persistence. In contrast, the inherent predictability of sea ice concentration was very low. These simulations used control conditions without rising greenhouse gases and hence did not assess changing aspects of predictability in a changing Arctic environment.
Climate model integrations of the twentieth and twenty-first centuries have been performed with a number of different modeling systems (IPCC 2007). While these models vary in the quality of their Arctic simulations (Zhang and Walsh 2006; Arzel et al. 2006; Gerdes and Koberle 2007; Holland et al. 2008a), they do provide a useful tool to assess changing aspects of Arctic sea ice predictability. Through the use of “perfect initialization” experiments in which multiple ensemble integrations are initialized with identical ice, ocean and terrestrial conditions, we can assess the inherent predictability in the sea ice system within the climate model context. Obtaining initial conditions from different time periods of standard twentieth to twenty-first century integrations allows us to test how predictability characteristics change with the changing climate state.
Here we assess a set of “perfect initialization” ensemble integrations from the Community Climate System Model, version 3 (CCSM3; Collins et al. 2006a) to address a number of interrelated questions regarding sea ice predictability on seasonal to interannual timescales. In particular, within the climate model system, we assess: (1) What is the inherent predictability of Arctic sea ice? (2) How important is “preconditioning” versus intrinsic variability for subsequent ice conditions? and (3) Do predictability characteristics change with a changing ice state? We assess the predictive capability for both Arctic sea ice thickness and area, with a particular focus on end-of-summer (September) sea ice cover since this has been the subject of a number of recent studies (Drobot et al. 2006; SEARCH outlook activity).
2 Climate model integrations
Climate model integrations from the fully coupled Community Climate System Model, version 3 (CCSM3) are examined. This model includes atmosphere, ocean, land and sea ice components (Collins et al. 2006a). For the integrations considered here, the atmosphere model (CAM3) (Collins et al. 2006b) is run at T85 resolution (approximately 1.4 degrees) with 26 vertical levels. The ocean model (Smith and Gent 2004) includes an isopycnal transport parameterization (Gent and McWilliams 1990) and a surface boundary layer formulation following Large et al. (1994). The dynamic-thermodynamic sea ice model (Briegleb et al. 2004; Holland et al. 2006b) uses the elastic-viscous-plastic rheology (Hunke and Dukowicz 1997), a sub-gridscale ice thickness distribution (Thorndike et al. 1975; Lipscomb 2001) and the thermodynamics of Bitz and Lipscomb (1999). Both the ice and ocean models use a nominally 1-degree resolution grid in which the north pole is displaced into Greenland. The grid spacing over the Arctic Ocean varies from about 20 km · 60 km near Greenland to 50 km · 70 km in the East Siberian Seas. As such, the Arctic Ocean is reasonably well resolved and there is an open channel within the Canadian Arctic Archipelago. The land component (Bonan et al. 2002) includes a subgrid mosaic of plant functional types and land cover types based on satellite observations. It uses the same spatial grid as the atmospheric model.
A description of the three sets of perfect initialization ensemble integrations
Number of members
The control integrations from which the initial conditions are obtained can be considered another member of the predictability ensembles as they have identical initial ice–ocean–terrestrial conditions for the years in question and they are treated as such for our results in Sect. 3. Different standard CCSM3 ensemble runs were used for the selected twentieth century (1970) initial conditions and twenty-first century (2016 and 2017) initial conditions. This was due to the initial condition availability from these runs. However, the standard CCSM3 twentieth to twenty-first century ensemble members are subject to the same timeseries of external forcing (e.g. changing atmospheric greenhouse gas concentrations, volcanic forcing, and solar variability at the top-of-atmosphere) and only differ slightly in their initial preindustrial (year = 1870) state. All of these integrations exhibit long-term declines in Arctic sea ice thickness and area (Holland et al. 2006). Differences across the various ensemble members reflect the contribution of simulated natural variability in the system and as such the initial states chosen for the predictability ensemble experiments represent valid CCSM3 conditions for the respective time periods.
Since the ensemble simulations apply an identical initial ice–ocean–terrestrial state (perfect initial conditions) and an identical model (perfect forecast model), the results give a measure of the maximum possible predictability in the system on seasonal-interannual timescales. The three different sets of ensemble members (Table 1), with different initial conditions (Fig. 2), allow us to explore the importance of the changing mean sea ice state for the seasonal-interannual predictability characteristics. Set 1 has relatively thick sea ice, whereas Sets 2 and 3 have thinner conditions more similar to the recently observed state (e.g. Stroeve et al. 2008).
The timeseries of northern hemisphere September ice area and January Arctic basin averaged ice thickness for 1950–2060 from the control integrations that are used to obtain initial conditions in the ensemble sets are shown in Fig. 3. These standard twentieth to twenty-first century simulations, including the Arctic sea ice conditions, have been assessed previously (e.g. Holland et al. 2006a; Meehl et al. 2006) and were discussed in the IPCC-AR4 (IPCC 2007). Similar to observations and hindcast simulations forced with atmospheric data (e.g. Maslanik et al. 2007; Rothrock et al. 2003), the simulated ice volume shows a considerable decline over the late twentieth to early twenty-first century in all of the CCSM3 standard runs with a corresponding reduction in the summer sea ice area. As discussed by Stroeve et al. (2007), the reduction in September ice extent in these simulations is consistent with observed sea ice loss over the satellite record.
3.1 Ensemble integrations
Statistics for the September ice characteristics from the different ensemble sets, including the standard deviation of ice area across the ensemble members, the autocorrelation of the change in September area from one year to the next for year 1 and for year 2, and the fraction of ensemble members that exhibit an increase in September ice area in year 1 compared to the September conditions obtained just prior to the ensemble set initialization
Fraction that increase in Year 1
From the 1979–2008 satellite observations, the year-to-year changes in September ice area have a significant negative 1-year lagged autocorrelation (R = −0.59) meaning that an increase in September ice cover is typically followed by a decrease in ice cover the next year. This occurs even with the large downward trend in September sea ice over this time period because the one-year reductions in ice are considerably larger than the recovery that often occurs the next year. A similar lagged autocorrelation behavior is seen in the members from Ensemble Sets 1 and 2 (Table 2), where the September ice area change in Year 1 and Year 2 are significantly correlated at −0.50 and −0.46, respectively. For Ensemble Set 3, the correlation is not significant at −0.05. The observed and simulated negative correlation is likely related to negative feedbacks, such as the fact that thinner ice cover grows more rapidly subject to the same forcing. This fundamental aspect of sea ice thermodynamics gives rise to a negative feedback with a stabilizing influence on sea ice conditions (Bitz and Roe 2004). The low correlation in Ensemble Set 3 suggests that the influence of these stabilizing feedbacks is overwhelmed by other factors such as the inherent variability of the system and positive feedbacks such as that due to surface albedo changes. The standard deviation in September ice area increases for the second year of integration with the increase being largest for Set 3. This suggests that there is some “knowledge” of the initial conditions during the first September of integration, which decreases the following year. The stabilizing feedback implicit in the ice area change correlation should counteract this but is not strong enough to overcome it.
Over time, the perfect initialization ensemble members diverge due to the chaotic nature of the system. The spread in sea ice conditions across the ensemble members results from variations in dynamic and thermodynamic sources and sinks of sea ice. The resulting changes in these terms (i.e. changes in ice melt, growth and divergence) directly influence the ice thickness and indirectly modify ice area when, for example, entire regions of the ice cover melt out in summer. During model run time, the monthly change in ice volume at each model grid cell that results from thermodynamic processes and from dynamic processes is separately diagnosed. Similar diagnostics are computed for the change in ice area that results from thermodynamic and dynamic processes. An analysis of these terms allows us to separate the influence of thermodynamic change versus dynamic change for the evolving sea ice conditions. The thermodynamic processes include all growth and melt terms, including frazil and basal ice formation, snow-to-ice conversion, and surface, basal, and lateral melting. The dynamic processes include ice divergence due to transport and, in the case of ice area change, ridging and rafting processes that reduce ice area while conserving ice volume.
For all of the ensemble sets, the scatter across the ensemble members in the summer thermodynamic (i.e. melt) driven ice volume change (Fig. 6a) is significantly correlated to the following September ice area. This indicates that differences in summer ice melt, and the consequent ice thickness, are largely responsible for the spread in September ice area in the different ensemble members. In contrast, scatter in the dynamic driven anomalies in ice volume (i.e. ice export; Fig. 6b) generally show smaller correlations with the subsequent September ice area. Indeed, significant correlations only occur for ensemble Set 3 during the summer months and ensemble Set 1 during September; these are considerably smaller than the correlations with thermodynamically driven ice volume changes.
Differences in the ice volume simulation across the ensemble members affect ice area by influencing the region that can potentially melt out during summer or by modifying the strength of the ice pack and resulting ice motion and convergence. An examination of the relationship between the ice area change terms and subsequent September ice area (Fig. 7) indicates that different mechanisms are important in the different ensemble sets. During winter, dynamic ice area loss in the Arctic resulting from export or convergence is balanced by thermodynamic ice area gain as newly opened water rapidly refreezes. This leads to an almost perfect anti-correlation between the Arctic thermodynamic and dynamic sources of open water during winter (not shown) and causes comparable (but opposite sign) correlations in Fig. 7a,b for January through May. It also allows the Arctic Ocean to remain almost completely ice covered through April in all of the ensemble sets.
The dynamic and thermodynamic ice area change no longer balance over the melt season, resulting in a reduction in ice area. During this time period, it is illustrative to consider the ice area changes during each month (Fig. 7c, d), as opposed to the time-integrated changes since initialization (Fig. 7a, b). For all ensemble sets, ice area loss resulting from net melt during the summer plays an important role in the ultimate scatter of September ice cover across the ensemble members (Fig. 7c). The timing of these important melt anomalies differs among the ensemble sets and generally occurs a month earlier in the thinner ice regime (Set 2 and Set 3) where summer melt more readily leads to open water formation. Additionally, the processes that enhance thermodynamic ice area loss (melt out) in some ensemble members are different across the ensemble sets. All ensemble sets exhibit enhanced summer open water formation for integrations with larger summer ice volume melt. However, ensemble Set 3 also has enhanced melt out (thermodynamic ice area loss) in simulations which have higher dynamic ice area loss during the previous winter. This is reflected in the significant positive correlations (Fig. 7d) found in January and February for Ensemble Set 3. Indeed, in Set 3, the June thermodynamic ice area loss (melt out) is significantly correlated to dynamic ice area loss during winter (at R = 0.6; not shown). While any winter dynamic ice area loss quickly freezes over leading to little change in ice area, it does redistribute the ice and modify the ice thickness distribution. In particular, enhanced ridging leads to a broader distribution with a larger area of thin ice, which is then more able to melt out the following summer. This appears to play an important role in Set 3, but not in the other ensemble sets. This is likely related to the thin conditions used for initialization of the Set 3 ensemble members. Since thinner ice is weaker and more able to converge and shear, larger variations in ridging are possible across the Set 3 ensemble members. That different factors determine the resulting scatter in simulated September sea ice conditions for the various ensemble sets, strongly suggests that different ice thickness regimes and extreme events have different predictability characteristics.
3.2 Prognostic potential predictability of sea ice conditions
The potential predictability of the sea ice system can be quantified by diagnosing how quickly the ensemble simulations diverge and comparing this to the natural variability in the system (e.g. Koenigk and Mikolajewicz 2008; Pohlmann et al. 2004). We assess the expected spread due to natural variability using a multi-century present-day integration and a multi-century preindustrial integration of the model. The present-day and preindustrial simulations have fixed external forcings (e.g. greenhouse gas levels, top-of-the-atmosphere solar forcing, etc.) that are consistent with observed conditions for 1990 and 1870 time periods, respectively.
While the ensemble sets have these general characteristics in common, there are some considerable differences as well. Ensemble Set 1, with relatively thick sea ice conditions, retains significant potential predictability throughout most of the first year of integration and has particularly high values over the summer. It also retains significant potential predictability during all or most of the second summer (depending on the characterization of natural variability used for comparison). Ensemble Set 2 also exhibits significant potential predictability over the spring and summer of Year 1. After April of the second year, Set 2 has essentially no significant potential predictability except during September of the second year. Ensemble Set 3 generally has the lowest PPP values. It has little significant potential predictability from June through December of Year 1, although September does exhibit marginally significant predictability in the first year. This set has no significant PPP values after March of Year 2 except during the final month of integration.
Taken together, these results suggest that regardless of the sea ice regime, prognostic potential predictability is generally significant for the first and second winters, ice area during the spring transition season shows less predictability, and summer ice area has potential predictability with a 9-month lead time. Much of the significant winter predictability is associated with conditions in the Labrador Sea (not shown), especially in the thick sea ice regime, which is consistent with results from Koenigk and Mikolajewicz (2008). Although it varies somewhat for different ensemble sets, the significant summer predictability during Year 1 is influenced by high predictability within the Barents Sea region (not shown). This differs from Koenigk and Mikolajewicz (2008) who found higher, but not significant, PPP values there. Contrasting the results from the various ensemble sets suggests that a thicker sea ice regime generally exhibits higher prognostic potential predictability in ice area for a longer period of time and summer ice area is potentially predictable for up to 2 years. In the thinner sea ice regime, winter ice conditions generally give less predictive capability for summer ice conditions, especially after a year.
The winter ice area exhibits potential predictability for a year in all of the Ensemble Sets. As discussed by Bitz et al. (2005), the location of the mean winter ice edge is strongly related to ocean heat flux convergence. It follows, that variations in ocean heat transport are likely important for the winter ice edge variability. Because of relatively long ocean advective timescales this allows for considerable predictability in the winter ice edge after a single year (and potentially much longer, although we can not assess this with our current experiments). This is true across all of the ensemble sets and is not strongly related to the mean Arctic climate state.
4 Conclusions and discussion
The inherent predictability of Arctic sea ice in the CCSM3 model on seasonal to interannual timescales has been investigated through the use of perfect initialization ensemble experiments. These simulations are initialized with identical ice–ocean–terrestrial conditions on January 1 and integrated for 2 years. We perform three different such ensemble sets, with different initial conditions, to assess the role of the mean sea ice state on predictability characteristics. The initial conditions are obtained from a standard twentieth to twenty-first century integration and represent the kind of changes in sea ice state that have occurred during the last 30 years. As such, the results from this study provide useful information on the potential forecasting of Arctic sea ice conditions in the rapidly changing Arctic environment that we are currently experiencing.
Our results suggest that, regardless of the initial ice state, the Arctic ice area exhibits potential predictability during the first few months of integration, which then drops during the spring transition season, and rises again during summer. All of the ensemble sets simulate significant prognostic potential predictability of ice area during the first September of integration (although it varies considerably across the ensemble sets and is quite low in Ensemble Set 3). This indicates that winter sea ice preconditioning (i.e. the January sea ice initial state in our experiments) provides some summer ice area predictive capability, which is consistent with ice–ocean coupled model studies (Kauker et al. 2009). As such, efforts should be made to better observe the Arctic ice thickness distribution, a conclusion also arrived at by the SEARCH Outlook effort.
A correlation analysis reveals that limits on the predictability of September sea ice area are primarily related to intrinsic variations in the thermodynamic-driven ice volume (i.e. melt/growth) anomalies that then translate into different amounts of open water formation over the melt season. Summer (JJAS) melt variations are particularly effective at driving scatter across the ensemble members, although for some ensemble sets, winter ice growth anomalies also play a role. Dynamic driven ice volume and ice area anomalies have generally a smaller influence on reducing the September ice extent predictability, except in Ensemble Set 3. In Ensemble Set 3 integrations, which are initialized with thin sea ice following a large summer ice-loss event, there is evidence that variations in the amount of ridging among the ensemble members in spring are important for redistributing the ice cover, allowing larger variations in summer melt-out and affecting the end-of-summer scatter in ice area.
During the second year of integration, all of the ensemble members exhibit potential predictability for the winter sea ice conditions. We speculate that this is related to the role of the ocean (with intrinsically long-timescales) in determining the winter ice edge. For the remainder of Year 2, Ensemble Sets 2 and 3 exhibit little potential ice area predictability. In contrast, Ensemble Set 1 exhibits significant prognostic potential predictability during the second summer of integration. Indeed, the potential predictability is generally higher in this ensemble set throughout the 2 years of integration. This set is initialized with thick ice conditions obtained from 1970 of the CCSM3 twentieth century integration; whereas the other sets are initialized with sea ice conditions more consistent with the current observed Arctic conditions. Thus, while one might expect that the summer Arctic sea ice is becoming more predictable due to the strong downward trend, our results suggest that on seasonal to interannual timescales the opposite is true. Ice area is more predictable in a thick sea ice regime and future summer ice area will be harder to forecast with continued thinning of the pack ice. Additionally from a comparison of Ensemble Set 2 and Set 3, which are initialized just preceding and just following a large 2007-like summer ice loss event, we find that even one year changes in the January ice thickness used for initialization (albeit large changes in this case) modify the predictability characteristics, with thinner initial ice conditions (Ensemble Set 3) generally resulting in less predictive capability.
The difference in summer predictability in different sea ice regimes appears in many respects to be a simple matter of geography. The summer ice concentration variability is concentrated along the ice edge with little variability present in the thicker interior ice pack. For a thick ice regime, such as Ensemble Set 1, this variability occurs along the shelf regions and is confined by the coast. In the thinner ice regimes prescribed as the initial conditions in Ensemble Sets 2 and 3, there is a region along the shelf that melts out in the first summer of all the ensemble members with consequently little variability there. In the following year, these regions are able to retain sea ice in some integrations leading to higher variability across the ensemble members and less inherent predictability.
Our results have implications for the design of sea ice forecasting systems. In particular, they suggest that historically based observational relationships used in statistical regression models may have less relevance in future (or possible present) sea ice regimes. This indicates that physically based models may have greater utility in future sea ice forecasting systems. While several ice–ocean coupled modeling systems have been used in the SEARCH Outlook effort, the lack of feedbacks to the atmospheric state is a concern that requires further investigation. Regional coupled climate modeling systems that include these feedbacks (e.g. Döscher et al. 2009) may provide an alternative forecasting tool although issues still arise on whether they can be adequately initialized. Regardless, our results do suggest that for all sea ice regimes, the intrinsic atmospheric variability during summer months places a strong limit on the predictability resulting from winter sea ice/ocean conditions.
We have only examined simulations initialized on January 1 for this study, which was necessitated by the data availability. However, a number of existing seasonal prediction systems forecast end of summer sea ice cover using springtime information. It is unclear what the potential predictability of sea ice conditions is given a smaller lead time. Since our integrations indicate that summertime thermodynamic forcing was an important factor in reducing the inherent predictability of end-of-summer ice cover, it is not clear that a smaller lead time (e.g. initializing with March conditions) will provide a greatly enhanced potential predictability. Additionally, this study only assessed integrations from a single coupled model and issues such as model biases and model resolution may influence the simulated predictability characteristics. These issues will be explored further in future work.
We thank three anonymous reviewers for constructive comments that led to improvements in the manuscript. We acknowledge the numerous scientists involved in the building and testing of the Community Climate System Model. S Vavrus’s collaboration in this effort was supported under a visiting faculty fellowship program at the National Center for Atmospheric Research and by NSF OPP-0327664, ARC-0628910, and ARC-0652838. DA Bailey was supported by NSF OPP-0612388. The National Center for Atmospheric Research is sponsored by the National Science Foundation.