The performance of a set of 15 global climate models used in the Coupled Model Intercomparison Project is evaluated for Alaska and Greenland, and compared with the performance over broader pan-Arctic and Northern Hemisphere extratropical domains. Root-mean-square errors relative to the 1958–2000 climatology of the 40-yr ECMWF Re-Analysis (ERA-40) are summed over the seasonal cycles of three variables: surface air temperature, precipitation, and sea level pressure. The specific models that perform best over the larger domains tend to be the ones that perform best over Alaska and Greenland. The rankings of the models are largely unchanged when the bias of each model’s climatological annual mean is removed prior to the error calculation for the individual models. The annual mean biases typically account for about half of the models’ root-mean-square errors. However, the root-mean-square errors of the models are generally much larger than the biases of the composite output, indicating that the systematic errors differ considerably among the models. There is a tendency for the models with smaller errors to simulate a larger greenhouse warming over the Arctic, as well as larger increases of Arctic precipitation and decreases of Arctic sea level pressure, when greenhouse gas concentrations are increased. Because several models have substantially smaller systematic errors than the other models, the differences in greenhouse projections imply that the choice of a subset of models may offer a viable approach to narrowing the uncertainty and obtaining more robust estimates of future climate change in regions such as Alaska, Greenland, and the broader Arctic.
Global climate models (GCMs) are the most widely used tools for projections of climate change over the time scale of a century. The periodic assessments by the Intergovernmental Panel on Climate Change (IPCC) have relied heavily on global model simulations of future climate driven by various emission scenarios. The global model simulations show a polar amplification of the greenhouse-driven warming and of other variations in climate (Serreze and Francis 2006; Wang et al. 2007), although the ratio of the models’ projected changes to the natural variability is not necessarily greater in the Arctic than in lower latitudes (Kattsov and Sporyshev 2006).
Given the likelihood that the Arctic will experience greater climate changes than most other regions over the next century, the credibility of the model simulations of Arctic climate becomes a key issue. The absence of databases for the validation of future climate simulations increases the importance of evaluations of models’ ability to simulate recent climate, for which syntheses of observational data are available.
Greenhouse-driven climate change represents a response to the radiative forcing associated with increases of carbon dioxide, methane, water vapor, and other radiatively active gases, as well as associated changes in cloudiness. The response varies widely among models because it is strongly modified by feedbacks involving clouds, the cryosphere, water vapor, and other processes whose effects are not well understood. While changes in the radiative forcing associated with increasing greenhouse gases have thus far been relatively small (only a few watts per square meter; see Solomon et al. 2007), a much more potent change in forcing occurs each year through the seasonal cycle of solar radiation. Herein, we place the models’ ability to capture the seasonal cycle of present-day climate at the core of a strategy for evaluating the models’ simulations of Arctic climate. Our evaluation is motivated by regional applications of the climate model output in the Arctic, specifically the climatically sensitive regions of Alaska and Greenland. Both of these regions have surface states that can be fundamentally altered by relatively small climate changes: much of Alaska is underlain by permafrost (thick and continuous in the north, discontinuous and thin in much of the interior), while Greenland is dominated by an ice sheet that may already be responding to climate change (Rignot and Kanagaratnam 2006; Dowdeswell 2006). Greenland also has permafrost in much of the narrow strip of land between the ice sheet and the surrounding seas.
The analysis of the model results described here is directed at the following questions:
How does model performance over Alaska and Greenland compare with performance over the broader pan-Arctic and the Northern Hemisphere domains? Specifically, are the models with the smallest errors in the broader Northern Hemisphere also the models with the smallest errors in the Arctic and, particularly, in Alaska and Greenland?
Are the models’ errors attributable primarily to relatively uniform biases, that is, to the fact that the models are consistently too warm, too cold, too wet, too dry, etc?
Do twenty-first-century projections of Arctic change show any systematic dependence on the validity of the different models’ simulations of present-day climate?
2. Model output and validation data
Our evaluation is based on the twentieth-century simulations by the models used in the Fourth Assessment Report (AR4) of the IPCC (Solomon et al. 2007). These models are being used in the third phase of the Climate Model Intercomparison Project (CMIP; information available online at http://www-pcmdi.llnl.gov/projects/cmip/index.php) and hereafter are referred to as the CMIP3 models. The output used here consists of the monthly grids of surface air temperature, precipitation, and sea level pressure (SLP) for 1958–2000, which is a subperiod of the twentieth-century simulations of these models and is also the period spanned by the validation fields (see below). Most of the model simulations were begun in the 1800s and continued through 2000 with prescribed greenhouse gas concentrations and, in some cases, estimated sulfate aerosols and variable solar forcing (Table 1; see discussion in Wang et al. 2007). Simulations were continued through the twenty-first century, with forcing prescribed from the Special Report on Emissions Scenarios (SRES; A2, A1B, B1, etc.) of the IPCC (Nakicenovic and Swart 2000). For the evaluation performed in this study, we use only the output from the twentieth-century simulations (20C3M) to evaluate the models’ performance.
The CMIP3 model output is compared here against the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40), which directly assimilates observed air temperature and SLP observations into a reanalysis product spanning 1958–2000. Precipitation is computed by the model used in the data assimilation, that is, observed precipitation data are not assimilated into the reanalysis. ERA-40 is one of the most consistent and accurate gridded representations of these variables available, and it compares favorably with other reanalyses of the Arctic (Bromwich et al. 2007). It is therefore a logical choice for observational analyses from which we determine the model biases of late-twentieth-century surface air temperatures, precipitation, and SLP. (Data and documentation for ERA-40 can be found online at http://www.ecmwf.int/research/era/Products.) While ERA-40 was performed at T106 (∼125 km) resolution with 60 levels, we use the version of the output archived on a 2.5° latitude × 2.5° longitude grid for compatibility with the climate model output. Because our evaluation is limited to surface air temperature, precipitation, and sea level pressure, no upper-air levels of the reanalysis output are used here.
To facilitate GCM intercomparison and validation against the reanalysis data, all monthly fields of GCM temperature, precipitation, and SLP are interpolated to the common 2.5° latitude × 2.5° longitude ERA-40 grid. Our evaluation of the models’ simulated fields uses monthly, seasonal, and annual climatological means for the late-twentieth-century period of 1958–2000.
3. Validation method
The core statistic of our validation is the root-mean-square error (RMSE) evaluated from the differences between ERA-40 and each model for each grid point and calendar month. For those models for which ensembles of simulations were archived, we use only the first ensemble member for consistency with models that have only one simulation in the CMIP3 archive. In all cases, the differences are between climatological means for the 1958–2000 period. The RMSE calculations are performed for each of the 15 models for each calendar month, and are area weighted for each of the following four domains: Alaska (shown in Fig. 1), Greenland (shown in Fig. 1), the “pan-Arctic” polar cap (60°–90°N), and a middle–high-latitude “Northern Hemisphere” domain (20°–90°N). The Alaskan and Greenland domains are approximately equal in area and are contained within the Arctic and Northern Hemisphere domains (except for a small portion of southeastern Alaska), so the results for the various domains are not independent. The reason for our choice of the four overlapping domains is that, although our primary interest is in the models’ performance for Alaska and Greenland, the simulation of both present-day and future climate in these regions will depend on the simulation of regions from which weather systems move toward Greenland and Alaska. In particular, the larger-scale circulation over much of the Northern Hemisphere influences Alaska and Greenland via advection and teleconnections, so credible simulations of future changes will depend on the models’ ability to capture the large-scale circulation of the pan-Arctic and Northern Hemisphere domains.
While the RMSEs were evaluated for all three variables (temperature, precipitation, and sea level pressure), the calculation of the RMSE for SLP differed from the calculation of the other RMSEs by the removal of the domain averages from all SLP grids (ERA-40 and CMIP3 models). This was done because the spatial gradients are the key features of the SLP fields, because it is the gradients of pressure that determine the winds.
As a seasonally inclusive measure of the models’ success in simulating the regional and larger-scale climates, we sum the RMSE values over the 12 calendar months. In this respect, we are evaluating the models’ ability to simulate the seasonal cycle and hence the models’ sensitivities to the cycle of solar forcing. The use of the seasonal cycle as a “surrogate” measure of climate sensitivity is based on the premise that the response to seasonal changes in solar radiation should depend on at least some of the same processes that determine a model’s response to other changes of forcing (Tsushima et al. 2005; Hall and Qu 2006). This reasoning has motivated a variety of past studies, and the results are somewhat mixed. Lindzen at al. (1995), for example, found that the seasonal cycle of globally averaged temperatures did not correspond well with the signature of CO2-induced climate change in one particular model, primarily because of differences in the seasonal cycles of globally averaged temperature at the surface and in the middle troposphere. On the other hand, Knutti et al. (2006) evaluated relationships between the CMIP3 models’ climate sensitivities (defined as the equilibrium global temperature response to a global change in forcing, e.g., from increasing greenhouse gas concentrations) and the amplitudes of their seasonal cycles. They found that the models with the largest climate sensitivities tended to overestimate the seasonal cycle of surface air temperature when compared with observations, and concluded that “the amplitude of the seasonal cycle in temperature provides a strong constraint on climate sensitivity” (Knutti et al. 2006, p. 4232). Hall and Qu (2006) took this approach a step further and focused on the snow albedo feedback as a determinant of both the amplitude of the seasonal cycle and of climate sensitivity. They found that the surface albedo–temperature sensitivity in the seasonal cycle was an excellent predictor of this sensitivity in greenhouse-driven climate change simulations. The snow albedo feedback is especially pertinent to the present study, because snowmelt in Alaska occurs primarily during April and May when insolation is relatively strong. Earlier work by Hall (2004) showed that about half of the high-latitude response to greenhouse forcing was attributable to the albedo feedback associated with retreating snow and ice.
After summation of the regional mean RMSEs over the 12 calendar months, the sums are used to rank the models. The model with the smallest 12-month sum of RMSEs is ranked 1 for that variable and region, while the model with the largest 12-month sum of RMSEs is ranked 15. The ranks can then be summed over different variables and/or different regions, depending upon a user’s priorities for variables and regional emphasis. The raw RMSE values for individual months also enable users to assess the utility of a particular model for a particular month or season.
One disadvantage to the above procedure is that it does not distinguish between two contributions to the RMSE—the portion resulting from the bias in a model’s annual mean and the errors in the seasonal cycle that are superimposed on the model’s annual mean. To address this distinction, we calculate a second set of RMSEs based on the models’ errors after removal of the annual mean error (bias). The differences between these two sets of RMSEs, both of which are presented in the following section, indicate the contributions of the annual mean biases to the total RMSEs.
The RMSEs vary widely among models, across seasons, and across regions. To illustrate the seasonal and regional dependencies of the RMSEs, the median model RMSE values of each variable for each season are shown in Figs. 2 –4 for surface air temperature, precipitation, and sea level pressure, respectively. In all cases, the RMSEs computed after the removal of the annual mean bias are shown as gray bars adjacent to the total RMSEs (black bars). It is apparent from Fig. 2 that approximately half of the total RMSE of temperature is attributable to the bias in the annual mean temperature, although the removal of the annual mean has a smaller impact on the RMSE in summer than in winter. Largely because the mean biases are smaller in summer than in winter, the total RMSEs are smaller in summer than in winter in all domains. The model median temperature RMSE is generally smaller for Alaska and Greenland than for the entire Arctic polar cap (60°–90°N) except in summer, although it is comparable to the broader hemispheric (20°–90°N) values. It should be noted that because of uncertainties in the prescribed sea ice concentrations the quality of the ERA-40 surface air temperatures may be worse over the Arctic Ocean (roughly 75°–90°N) and its peripheral seas than over the northern land areas.
Precipitation is unique among the three variables in the sense that the removal of the annual mean bias generally increases the RMSE (Fig. 3), although the increase over the Alaskan domain is small (and slightly negative in winter, spring, and autumn). The interpretation is that the removal of the annual mean bias actually increases the bias in those seasons with the largest precipitation amounts (summer and autumn). By either measure of RMSE, the precipitation RMSEs for Alaska and Greenland are larger than those for the entire Arctic polar cap (but smaller than for the larger 20°–90°N domain. The precipitation RMSEs for Alaska and Greenland are smallest in spring, which is generally a dry season in both regions.
The RMSEs of sea level pressure (Fig. 4) show a similar seasonal cycle, with the largest in winter, to those of temperature, although it should be noted that the actual magnitudes of the temperature and SLP RMSEs are not directly comparable because of their different units. The annual mean bias accounts for a larger portion of the total RMSEs in the Alaskan and Greenland domains than in the larger domains, especially for 20°–90°N.
A substantial portion of the RMSE arises from biases in the models’ annual means, upon which the seasonal cycles are superimposed. The gray bars of Figs. 2 –4 show the RMSEs after removal of the model’s annual mean biases. It is apparent that the impact of the removal of the annual mean biases is greater for the smaller domains (Alaska and Greenland), and that the impact is also generally greater in the warm season (especially when the reduction of the RMSE error is viewed as a percentage of the total RMSE; see the black bars in Figs. 2 –4). However, the relative areal coverage of land fraction versus ocean in the smaller domains is greater than the land fraction of the larger domains, and the variability of temperature and precipitation are generally greater over land than ocean.
The main objective herein is to identify the models that are most successful at simulating the seasonal cycle of the climates of Alaska and Greenland. The relevant information is contained in Figs. 5 –7, which show the 12-month mean RMSEs of the different models, arranged in order of increasing RMSE, for the three variables. Each figure contains a separate display for Alaska, Greenland, the pan-Arctic (60°–90°N), and the extratropical Northern Hemisphere (20°–90°N). It is apparent from these figures, especially Fig. 5 for temperature, that the models vary widely in their ability to capture the seasonal climates of Alaska and Greenland. For example, the yearly averaged RMSE of temperature over Alaska varies from 2.9°C in the Max Planck Institute’s (MPI’s) ECHAM5 to 11.0°C in the Institute of Atmospheric Physics (IAP) Flexible Global Ocean–Atmosphere–Land System Model (FGOALS). The range for the pan-Arctic domain (60°–90°N) is even greater, from 2.9° to 13.6°C. While the ranges are smaller for the other variables, the RMSEs still vary across the models by nearly a factor of 2 for precipitation and by more than a factor of 2 for sea level pressure. Similar ranges are found for the larger domains, that is, the pan-Arctic and hemispheric polar caps.
Nevertheless, a noteworthy feature of Figs. 5 –7 is the tendency for some models to rank highest no matter which variable is evaluated. Moreover, the models that rank highest for Alaska tend to rank higher for Greenland. There is also a tendency for these same models to have the smallest RMSEs over the larger domains, although there are exceptions. Table 2 provides a synthesis of the model performance based on the two RMSE metrics: annual mean bias included (Table 2, top section) and annual mean bias removed (Table 2, bottom section). These tables rank the models from 1 (smallest RMSE) to 15 (largest RMSE) for each variable and domain. As in Figs. 5 –7, these ranks are based on RMSEs summed over all 12 calendar months, so they incorporate the models’ successes or failures in capturing the seasonal cycle. The rightmost columns of Table 2 are the sums of all 12 ranks (four domains × three variables) of the models. We refer to these columns as our “integrated ranks.” Because the domains are nested, the integrated rank effectively double weights the model performance over the Arctic polar cap (60°–90°N) and triple weights the model performance over the Alaskan and Greenland regions (i.e., Alaska and Greenland are included in both larger domains). While this weighting is admittedly ad hoc, it is consistent with our focus on Alaska and Greenland.
If the model ranks are based on individual variables or domains, the order shifts somewhat. The top-ranking models for the different domains, based on the computational strategies used for Table 2, are given in the top half of Table 3.
Based on the aggregate of the two sets of rankings, the top-performing models for Alaska are Geophysical Fluid Dynamics Laboratory Climate Model, version 2.1 (GFDL CM2.1), MPI ECHAM5, Centre National de Recherches Météorologiques Coupled Global Climate Model, version 3 (CNRM-CM3), Met Office (UKMO) third climate configuration of the Unified Model (HADCM3), and Model for Interdisciplinary Research on Climate 3.2 (MIROC3). The top-performing models for Greenland are GFDL CM2.1, MIROC3, CNRM-CM3 and MPI ECHAM5. For each of the two larger domains (60°–90°N and 20°–90°N), the two highest-ranking models are MPI ECHAM5 and GFDL CM2.1, while the MIROC3 model ranks third and fourth, respectively, for these two domains. Thus, there is strong overlap among the top performers for the larger domains (pan-Arctic and Northern Hemisphere) and the subregional domains (Alaska and Greenland).
When the ranks for a particular variable are summed over the four domains, the ranks for the three primary variables are shown in the bottom half of Table 3.
The summaries given in Table 3, together with Table 2, confirm a conclusion of many other climate model evaluations, that is, no single model outperforms all of the others for either all regions or all variables. On the other hand, several of the CMIP3 models consistently rank close to the top, demonstrating that their ability to reproduce the seasonal cycle of high-latitude climate of recent decades is superior to that of the other CMIP3 models. MPI ECHAM5 and GFDL CM2.1 are clearly the highest-ranking models overall. MIROC3 and UKMO HADCM3 models also rank highly. These higher-ranking models are the logical candidates for driving offline simulations of high-latitude variables such as permafrost, glaciers/ice sheets, and terrestrial or marine ecosystems.
With regard to the reasons for the different levels of skill over Alaska and Greenland (as well as the larger domains), no systematic relationship to model resolution emerged. The models with the smallest RMSEs, MPI ECHAM5 and GFDL CM 2.1 (Table 2), have resolutions that are neither the highest nor lowest of the 15 models. There is also no obvious relationship between model performance and the type of sea ice formulation in the models. Other candidates for explanations of the differences in model performance include the cloud and radiative formulations, which are now being investigated elsewhere; the planetary boundary layer parameterization; and the land surface schemes of the various models. Biases in the large-scale atmospheric circulation, perhaps driven by processes outside the Arctic, are also candidates to explain the across-model differences in temperature, precipitation, and sea level pressure.
The two highest-ranking models (MPI ECHAM5 and GFDL CM2.1) both included aerosol effects and ice-phase clouds in model versions that were released in 2005, which is more recent than the release date of most of the other CMIP3 models. Systematic experiments with these and other parameterizations are needed if the reasons for the relative success of these two models are to be established unambiguously. Nevertheless, there is a general tendency for the more recently released model versions to be the better performers in the Arctic.
To provide some perspective for the RMSEs of the different variables, we show the models’ bias fields for January in Figs. 8 –10. The composite fields included in these figures are based on the 14-model subset that excludes the IAP FGOALS model, which is such an outlier (Figs. 5 –7) that it skews the composite bias fields. Figure 8 shows that the 14-model composite biases of January temperature are generally less than about 3°C over most of the Arctic, except over the Barents Sea and far eastern Russian regions. The Barents Sea bias arises from the models’ tendency to oversimulate the extent of sea ice in this area during winter. There are more areas of negative (cold) bias than of positive (warm) bias, although the biases are generally no larger in the Arctic than in middle latitudes. The 14-model composite field averages out many biases of the opposite sign in the various models. The January bias fields for the individual models show that larger biases occur in some models, and that there are both positive and negative biases across the models at any location. Particularly noteworthy is the large negative bias over southern Alaska in MPI ECHAM5; this wintertime bias degrades the Alaskan temperature rank of this model, which otherwise ranks very highly in the Arctic.
The precipitation biases (Fig. 9) are more spatially complex than the biases of temperature and pressure, and they show a distinct orographic signature in the 14-model composite and in most of the individual models. The negative (dry) biases of the models over the coastal mountains of southern Alaska, western Canada, southeastern Greenland, and western Norway are indicative of a smoothing of the mountains that are subject to upslope flow in the major storm-track regions. The corresponding wet biases immediately inland of the dry upslope biases in some areas also point to the models’ oversmoothing of the topography, which is unable to create the full “precipitation shadow” effect that occurs downstream of major mountain ranges. The raw output from ERA-40 has a much finer resolution than the CMIP3 models (Table 1), and hence it is better able to capture the pattern of heavy upslope flow and leeside shadowing. This pattern is strikingly apparent along the northwest coast of North America (Fig. 9), where the models are too dry on the windward side and too wet on the lee side of the coastal mountains. This bias is much less apparent in the summer, when the bias fields of most models show no orographic signature. These error patterns point to the importance of downscaling GCM output fields when constructing site-specific climate change scenarios, especially for temperature and precipitation at locations in or near highly varying terrain.
The composite sea level pressure biases, shown in Fig. 10, are also generally small, with magnitudes in the 14-model composite being less than 3–4 hPa almost everywhere in the middle and high latitudes. Greenland and the eastern Arctic Ocean show positive biases, while the biases over Alaska and the North Pacific are negative. The biases over the Bering Sea and Alaska imply cold advection over eastern Asia, consistent with the cold bias in that area during January (Fig. 8, upper left). The positive bias of 3–6 hPa over the eastern Arctic Ocean is similar to the errors found by Bitz et al. (2002) in the previous generation of global atmospheric models. Individual models show much larger biases of both signs (Fig. 10).
The final question raised in the introduction pertained to possible relationships between the models’ projected greenhouse changes and the relative accuracy of the simulations of present-day climate. Figure 11a shows the models’ projected warming for 60°–90°N plotted as a function of the integrated rank of the models. The integrated rank used here is based on the top section of Table 2 and includes only the 20°–90°N and 60°–90°N domains. (While the Alaska- and Greenland-specific domains are excluded because the projected warming is for the entire pan-Arctic domain, this makes little difference in Fig. 11 because the ranks across the various domains are highly correlated.) The warming is defined as the area-weighted linear change in surface air temperature from 2001 to 2099 under the IPCC A1B scenario. While there is considerable scatter in the warming, there is a tendency for the highest-ranking (better performing) models to simulate the greatest warming over the 60°–90°N domain. The projected warming versus model performance relationship is statistically significant at the 5% level. The projected changes of precipitation (Fig. 11b) show a similar dependence on the models’ integrated ranks, with better-performing models projecting larger increases of Arctic precipitation than the lower-ranking models. While there is more scatter in the corresponding SLP results (Fig. 11c), there is a tendency for larger twenty-first-century decreases in Arctic sea level pressure to be projected by higher-ranking models than by lower-ranking models.
The projections summarized above are based on an intermediate scenario (A1B) of greenhouse forcing. Because the models’ different climate sensitivities do not depend on the precise rates of increase of forcing, the models that produce the stronger (weaker) warming under A1B forcing tend to produce the stronger (weaker) warming under the A2 and B1 forcing scenarios (Solomon et al. 2007, p. 763, their Fig. 10.5). Hence, the results in Fig. 11 do not depend strongly on the choice of the A1B scenario.
The dependence of the Arctic’s climate sensitivity on the RMSE used here may be considered surprising, because it is just as likely that “bad” models are biased warm as cold. This finding raises the question about whether the sensitivity is correlated with the amplitude of the seasonal cycle or with the bias. The correlation between the projected warming and the amplitude of the simulated present-day seasonal cycle was found to be small and insignificant for the 60°–90°N domain on which the results in Fig. 11 were based.
This study represents an early step in the evaluation of global climate model performance over Arctic subregions, with an eye toward the narrowing of the uncertainty in regional Arctic climate projections by the selection of an optimal subset of models. It must be emphasized that the ranking process used here is targeted at particular applications (i.e., forcing of permafrost models, planning for climate change in Alaska and Greenland) and hence contains arbitrary elements. For example, the assignment of weights based on an ad hoc–postulated equal importance of different variables and regions will not be appropriate for other regions and applications. We present our approach and results in the spirit that they may stimulate enhancements or better alternative approaches to regional applications of climate model scenarios.
Several questions about model performance were posed in the introduction; based on the results presented on the preceding section, the answers to those questions are as follows:
Model performance over Alaska and Greenland is generally neither better nor worse than the performance over the pan-Arctic and Northern Hemisphere domains. During winter and autumn, the temperature errors over Alaska and Greenland are somewhat smaller, but the precipitation errors are generally larger, than over the larger domains. The specific models that perform best over the larger domains tend to be the ones that perform best over Alaska and Greenland, although a notable exception is the relative lack of success of the overall top-ranked model (MPI ECHAM5) in capturing the winter temperatures over Alaska.
The rankings of the models are largely unchanged when the bias of each model’s climatological annual mean is removed prior to the error calculation for the individual models. The annual mean biases typically account for about half of the models’ root-mean-square errors. Thus, the root-mean-square errors are not simply the manifestations of general tendencies for the models to be colder, warmer, wetter, drier, etc., than the corresponding observationally derived fields. However, the root-mean-square errors of the models are generally much larger than the biases of the composite output, indicating that the systematic errors differ considerably among the models.
There is a tendency for the models with the smaller errors to simulate a larger greenhouse warming over the Arctic. Because several models have substantially smaller systematic errors than the other models, the differences in warming imply that the choice of a subset of models may offer a viable approach to narrowing the uncertainty and obtaining more robust estimates of future climate change in regions such as Alaska, Greenland, and the broader Arctic. The results in Fig. 11 suggest that the uncertainty might be narrowed by eliminating the models with the larger biases. Because the models with the larger biases tend to have the weaker sensitivities to anthropogenic forcing, the aggregate of the remaining (retained) models would have a stronger sensitivity and would imply more confidence about that sensitivity. Such an approach has already been suggested by Overland and Wang (2007) and Kattsov and Sporyshev (2006), and will be pursued in future assessments of Arctic change.
This work was supported by the National Science Foundation, Office of Polar Programs, through Grant OPP-0612533 to the University of Alaska, Fairbanks, and by Grant OPP-0520112 to the University of Illinois at Urbana–Champaign. We acknowledge the following international modeling groups for providing their data for analysis: the Program for Climate Model Diagnosis and Intercomparison (PCMDI) for collecting and archiving the data, the JSC/CLIVAR Working Group on Coupled Modeling (WGCM) and their Coupled Model Intercomparison Project (CMIP) and the Climate Simulation Panel for organizing the model data analysis activity, and the IPCC Working Group I’s Technical Support Unit (TSU) for technical support. The IPCC Data Archive at Lawrence Livermore National Laboratory is supported by the Office of Science, U.S. Department of Energy. Finally, we thank three anonymous reviewers for comments and suggestions that improved the manuscript.
Corresponding author address: Dr. John E. Walsh, International Arctic Research Center, University of Alaska, 930 Koyukuk Drive, Fairbanks, AK 99775-7340. Email: firstname.lastname@example.org