Results from an isotope-enabled general circulation model are presented in order to determine the isotopic signal of a warmer climate on Antarctica. The warming is forced using CO2 forecasts for the next century. For unforced interannual climate variability the temporal gradient and correlation between stable water isotopes and surface temperature is small. The relationship is much stronger for the CO2 forced event. There is little regional coherence between temporal gradients for the forced and unforced climates, implying that correlations between stable water isotopes and temperature from instrumental records of a couple of decades cannot be applied to larger warming events. Additionally, there are strong discrepancies between the forced warming temporal gradients and present-day spatial gradients of isotopes against temperature. We show that it is difficult to obtain a local spatial gradient since it is systematically affected by the geographical size of the spatial sample. For the forced warming, the temporal gradient derived for the warming event over Dome C is less than half the value generally applied. We determine, through means of a new frequency decomposition, that a large portion of this decrease from the expected value is due to changes in the seasonal precipitation temperature covariance. This low isotopic sensitivity to a CO2 driven warming implies that current and future warming trends may have rather small isotopic signals in Antarctica.
 The primary control on the spatial distribution of stable heavy water isotopes at high latitudes has been shown to be local temperature [Dansgaard, 1964]. For this reason, the stable water isotopic content of ice cores is used as an indicator of past temperatures. The fractional content of the stable water isotope oxygen-18 is usually expressed as the deviation from a standard water isotope sample, so for H218O: δ18O = [1000 × (18O/16O)V–SMOW] − 1, with V-SMOW = Vienna standard mean ocean water [e.g., Rozanski et al., 1992]. The method generally applied for central Antarctic ice cores is to substitute the local spatial gradient between surface temperature (TS) and isotopic content δ18O (or equivalently using deuterium δD) for the expected temporal gradient. This has been used to obtain past temperature records from the isotopic content of Antarctic ice cores including the Vostok [Lorius et al., 1985; Jouzel et al., 1987], EPICA Dome C [EPICA Community Members, 2004; Jouzel et al., 2007] and EDML (EPICA Dronning Maud Land) ice cores [EPICA Community Members, 2004].
 In the vicinity of Vostok, with temperatures between −20°C and −55°C, a spatial gradient of 9‰ δD per °C fits the data over a large geographical range [Jouzel et al., 1987; Petit et al., 1999]. Similar geographical spatial gradients apply to the EPICA [EPICA Community Members, 2004] Dronning Maud Land (EDML) and EPICA Dome C [Jouzel et al., 2007] sites. The more recent spatial gradients applied, after making corrections for sea-water isotopic content and altitude change through time, to obtain temperatures from these ice cores have been between 0.75 and 0.82 ‰ °C−1 in δ18O equivalent. Various authors have suggested that there is a 20% uncertainty on these gradients [e.g., Jouzel et al., 2003, 2007; Stenni et al., 2001]. The δ18O equivalent values are obtained by using the factor of 8 difference (due to the average relationship between hydrogen and oxygen isotope ratios) between TS against δD and TS against δ18O gradients. If these geographically derived gradients do not fit the paleotemperature against isotope record from Antarctic ice cores, this could have implications for our understanding of past temperature changes in Antarctica. Several authors have noted there is not necessarily any causal relationship between the spatial and temporal gradients [e.g., Noone and Simmonds, 2002].
 The temporal relationship between isotopes and temperature at Antarctic sites can only be directly observed using the short instrumental TS and δ18O records. Linear fits to interannual TS against δ18O time series show lower gradients and correlations than occur in geographical (spatial) data sets [e.g., Dansgaard et al., 1975; Robin, 1983; Schlosser and Oerter, 2002; Werner and Heimann, 2002; Schmidt et al., 2007]. It is not clear if these weak relationships from short interannual time series are analogous to the larger climatic shifts that the ice cores record [e.g., Jouzel et al., 1994]. Jouzel et al.  suggest that stronger “forced” temperature changes increase the δ18O against TS gradient.
 For Greenland, because of high accumulation rates, it is possible to use inverse models based on borehole thermometry as an independent constraint on the millennial temperature isotope relationship. These, alongside model results, show that precipitation seasonality changes between glacial and interglacial periods reduce the long-term temporal gradient between TS and isotopic content to a much lower value than the current spatial gradient [e.g., Cuffey et al., 1995; Hoffmann and Heimann, 1997; Fawcett et al., 1997; Krinner et al., 1997; Werner et al., 2000]. Inland Antarctic core sites borehole thermometry evidence is ambiguous because of rapid signal diffusion due to low precipitation rates [Salamatin et al., 1998]. Additionally, in contrast to Greenland, the Antarctic does not seem to have undergone the same degree of abrupt past temperature changes that Greenland experienced, thus gas isotopic contents (15N/14N and 40Ar/36Ar) have provided only very limited additional paleotemperature evidence [Caillon et al., 2001]. Thus there is no good independent proxy or direct observations of the long-term Antarctic relationship between the stable water isotopes and past temperature.
Jouzel et al.  suggested that a good alternative test of temporal gradients is to calculate gradients over a range of climates using an isotopically enabled general circulation model (GCM). For EPICA Dome C, Jouzel et al.  show this type of test supports the spatially derived conversion gradient of approximately 0.75 ± 0.15‰ °C−1 in δ18O (see Jouzel et al.  supplements). Modeling work by Delaygue et al.  and Krinner and Werner  also finds only very limited changes in the seasonality of precipitation between glacial and interglacial times, which supports the use of this gradient. However, Salamatin et al. , on the basis of perhaps ambiguous borehole evidence, suggest a temporal slope that might be 50% lower than the local Vostok spatial gradient. Likewise, using different assumptions and again somewhat ambiguous evidence, Caillon et al. , Parrenin et al.  and Schwander et al.  all suggest that the glacial interglacial Antarctic gradient might be 20% lower than the local spatial values.
 Here, we use this multiple experiment GCM approach to investigate the isotope-temperature relationship under forced warming conditions. This is attractive for several reasons. Globally warmer time periods are a topic of key current relevance. GCMs are more thoroughly validated for the present day, and the warmer climate conditions we are trying to model are closer to the present day than a glacial climate state. Boundary conditions over a CO2 driven warming are probably better defined than those for a cold event. The models have also been more thoroughly investigated for a CO2 forced warming event, and thus we can have more confidence that the climate processes that change during the warming are better documented, and perhaps understood.
 We aim therefore to use these experiments to help fill the gap that exists between our understanding of the short instrumental plus satellite record and the long ice-core record [Schneider et al., 2004]. We set the type of temperature observation that can be obtained from instrumental data (e.g., temperature) in the context of proxy water isotope data [e.g., Petit et al., 1999], and investigate how the proxy should behave under the conditions expected over the next century. This allows us to supplement predictions of temperature change with predictions of changes in precipitation δ18O over the next century and may allow use of the δ18O signal to determine when a significant temperature change has occurred.
 We run an isotopically enabled climate model over a CO2 forced climate warming event. The event is represented by a series of six experiments each with a progressively warmer climate. We use the present-day experiment to examine the spatial distribution of temperature and isotopes, in conjunction with a recently compiled Antarctic isotope database [Masson-Delmotte et al., 2008]. We carry out the analysis using equal area representation of the model output. Additionally we examine the interannual temporal gradients found in the present-day experiment, and compare them to the gradients found over the whole set of warming experiments. To analyze the CO2 forced warming results, we introduce a new frequency decomposition of the local influences of precipitation and temperature covariance on isotope record of temperature.
2. Model and Experiments
 The atmospheric model HadAM3 has a regular latitude longitude grid with a horizontal resolution of 2.5° × 3.75°, and 19 hybrid coordinate levels in the vertical [Pope et al., 2000]. The modeled Antarctic, and global, climatology is similar to that observed [Turner et al., 2006; Pope et al., 2000]. Details of the new stable water isotopic submodel incorporated into HadAM3 that we use here are presented by Tindall et al. . The accuracy of the HadAM3 climatology over Antarctica should enable the isotope output to match observations better than that from some other GCMs since it is generally biases in model climatology which produce the largest errors in modeled isotope output. For example, MUGCM and GENESIS both have significant Antarctic warm biases [Noone and Simmonds, 2002; Mathieu et al., 2002], while GISS has a cold bias [Jouzel et al., 2003]. The Antarctic climate of HadAM3 does seem to be quite close to the later versions of ECHAM [e.g., Hoffmann et al., 1998], which also has quite an accurate simulation of the Antarctic climate.
 Approaching the poles, the model grid cell size tends toward zero. This causes problems both with a constant time step (of 30 min) and with the parameterizations used (which are partly grid-size dependent). Therefore to prevent errors and instabilities, due to the Courant-Friedrichs-Lewy criterion, fields are Fourier-filtered near the poles [Pope et al., 2000, and references therein]. This means that cells close to the poles are less independent of each other, although in any GCM there will generally be correlation between nearby grid cells.
2.1. Experimental Setup
 We present results derived from six experiments: one for the present day (PD) based on 1990, and five between 2020 and 2100 (see Table 1). The PD boundary condition data are based on a monthly average of 1980–1999 HadISST sea surface temperature and sea-ice data. The HadISST data set compares well with other published sea surface temperature analyses [Rayner et al., 2003]. The five future experiments are based on sea surface temperature and sea-ice conditions obtained from the World Climate Research Programme's (WCRP's) Coupled Model Inter-comparison Project phase 3 (CMIP3) runs with HadCM3 [Gordon et al., 2000]. The full ocean and atmosphere Special Report on Emissions Scenarios (SRES) A1B HadCM3 experiment provide a good basis for the future experiment boundary conditions: the run is well-documented and consistent with our isotope enabled HadAM3 model.
 The A1B future scenario is not necessarily considered more likely than any of the other CMIP3 future scenarios. In terms of global temperature, it is about the middle of the range of changes projected by the different CMIP3 scenarios and suggests a centennial warming of ∼3°C [Bracegirdle et al., 2008]. The run scenario ends at the year 2100 which constrains the forcing period for the isotope enabled HadAM3 experiments. Note that as Table 1 shows, although the warming is driven by a mixture of greenhouse gases (GHG), here we use CO2 and GHG driven warming interchangeably; that is, where CO2 is written, we wish to imply CO2 equivalent GHG forcing.
 The HadCM3 SRES A1B sea surface temperature output has some regional biases compared with the PD HadISST sea surface temperature. These biases can affect the modeled Antarctic climatology. For example, tropical warm sea surface temperature biases in HadCM3 over Indonesia and the eastern tropical Atlantic cause mean sea level pressure errors around Antarctica [Lachlan-Cope and Connolley, 2006]. The effect of these sea surface temperature model biases is minimized by applying the HadCM3 SRES A1B sea surface temperature fields as anomalies to the HadISST sea surface temperature boundary conditions [e.g., Krinner et al., 2008]. Use of 10-year average HadCM3 SRES A1B sea surface temperatures avoids interannual run variability affecting the experiments. As an example, the 2080 experiment sea surface temperature field is formed by subtracting a 10-year monthly mean of 2000 to 2010 from a 10-year monthly mean of 2070 to 2080 and then adding the resultant anomaly to the present-day HadISST sea surface temperature.
 The use of the anomalies, while minimizing HadCM3 error bias, means that the 2020 is effectively separated by a smaller warming anomaly than would be expected (effectively 10 years) from the present-day run. Beyond 2020, the experiments are each separated by 20-year anomalies. As a result, the set of runs is indicative of anomalous SRES A1B warming over a 90-year rather than 100-year SRES A1B scenario HadCM3 run. Use of anomalies displaces any possible absolute dates of the future experiment. However, for the purposes of this investigation we only wish to model a plausible magnitude CO2 warming signal: “absolute” dates are not meaningful or of interest and have no effect on the findings presented.
 After the sea surface temperature anomalies are applied, any resulting sea surface temperatures of less than −1.8°C are set to −1.8°C to avoid improbably low values (as done by Paul and Schäfer-Neth ). We do not use sea-ice anomalies since negative sea ice occurs. Therefore sea-ice conditions from the HadISST data set are used for the present-day experiment then from the SRES A1B HadCM3 data set for the 2020 to 2100 experiments. Figure 1 shows the mean annual anomalous sea surface temperatures and the sea-ice conditions at 2100 for mean annual, summer (DJF) and winter (JJA) seasons. The change between experiments PD to 2100 in sea surface temperature is relatively linear, so the 2020 to 2080 experiments are similar to scaled versions of the 2100 anomalies shown.
 For the 2100 experiment the annual average sea surface temperature anomalies at 50°S are 1.3 to 1.5°C warmer in a nearly zonally constant pattern with a limited seasonal signature. South of 50°S the sea surface temperature anomaly become more strongly seasonal and zonally dependent. At 70°S the zonally averaged annual mean anomaly is 2.3°C with a seasonal spread of 3°C (zonal mean anomaly in summer of 0.5°C and winter of 3.5°C). Sea-ice loss is at a maximum at 70°S with approximately 25% losses in fractional cover for the annual average. The maximum losses in summer are farther south at 72.5°S and slightly northward at 67.5°S in winter.
 The surface ocean δ18O value used for all experiments is uniformly zero. Over polar regions, initial test experiments (not shown here) using the current global ocean (and lake) surface distribution of δ18O produced δ18O in precipitation results that were almost identical to those from uniform source condition experiments. Using an atmosphere only model means that we cannot simulate the oceanic and lake warmer experiment distribution of δ18O. The limited polar differences in δ18O in precipitation simulated by using the observed oceanic and lake distribution of δ18O leads us to believe that the δ18O in precipitation (for the warming event simulated here) will not be sensitive to realistic changes in this boundary condition. Thus our constant and uniform surface ocean δ18O seems a reasonable approach.
2.2. Present-Day Climate
 The match between HadCM3 modeled and observed Antarctic climatology is closer than for most other GCMs [Bracegirdle et al., 2008]. Even using HadCM3 (which is HadAM3 coupled to an ocean GCM), which is known to have incorrect sea level pressure around Antarctica due to tropical sea surface temperature errors, the atmospheric model rates amongst the top few models (fourth of 19 of the IPCC AR4 models) for its atmospheric representation of the Antarctic [Connolley and Bracegirdle, 2007]. Additionally we note that without the tropical sea surface temperature errors, eliminated here by the use of the HadISST sea surface temperature data, the simulation is more accurate [e.g., Krinner et al., 2008]. The geographical distribution of the modeled precipitation (Figure 2a) minus evaporation and temperature compares quite favorably both with observations, and with ECMWF output for the majority of the continent [Turner et al., 2006].
 Mean Antarctic precipitation for the present-day run is 182.8 kg m−2 a−1 (Figure 2a); sublimation is 13.1 kg m−2 a−1; giving a total mean accumulation rate of 169.7 kg m−2 a−1. This compares to mean accumulation estimates of 143 ± 4 kg m−2 a−1 [Arthern et al., 2006]; 151 kg m−2 a−1 [Turner et al., 1999]; and 182 kg m−2 a−1 [Monaghan et al., 2006]. We note that the modeled values could be slightly inflated owing to the model land masking. High-accumulation regions such as the Ross Sea Embayment and the Peninsula region are represented by regions which are too large (e.g., see difference between true coast and modeled coastline in Figure 2). However, the modeled accumulation values are still within the range of observationally derived estimates.
 The present-day experiment has a low-amplitude seasonal precipitation signal. There is only rather limited observational information on the inland Antarctic seasonal cycle [e.g., Bromwich, 1988; Ekaykin et al., 2002] with which we can compare the modeled precipitation seasonality. Bromwich  suggested that most precipitation in East Antarctic may fall in winter, despite the low average moisture content of the air, owing to more intense winter cyclones activity. However, Marshall  suggests in total the evidence is somewhat equivocal with some of the records referenced by Bromwich  indicating a slight winter maximum (for Vostok and a suggestion of a semiannual pattern for South Pole). The modeled present-day run has an East Antarctic (see Figure 3 for regional definitions) summer precipitation maximum peaking in December at 11% above the mean annual precipitation value. This is supported by the recent Fujita and Abe  East Antarctic record for Dome F which also indicates a summer maximum. By 2100, the modeled East Antarctic summer maximum has reduced to +7% in December and the cycle has become flatter with a semiannual pattern featuring a second +7% peak in May (spatial pattern of increase between PD and 2100 shown by contouring on Figure 2a). For the rest of the continent, there is a winter maximum in modeled present day precipitation: +17% from May till August for the West Antarctic; +22% for the Peninsula, with a hint of a semiannual cycle; and the average for the whole continent is +15% between April and June. For regions with a winter maxima in precipitation the 2100 cycle amplitude is larger than that for the present day with the regions all increasing by about 10%. We also note a substantial increase in the annual mean precipitation total for all regions over the warming event of between 15 and 20%.
2.2.2. Temperature and δ18O in Precipitation
 The isotopic results from HadAM3 enable us to examine in detail the relationship between stable water isotope values and temperature. In this case we do this specifically within the context of a warming world. Although we note that the deuterium excess can provide useful isotope derived climate insight [e.g., Masson-Delmotte et al., 2008], herein, we confine our analysis to the oxygen-18 isotope results.
Figures 2b and 2c show the modeled present-day pattern of surface temperature and δ18O. The largest absolute errors in the mean annual PD experiment surface temperature TS are in the Dronning Maud land close to 80°S, 20°W and near 45°E, where the orography is too high and too low, respectively. Additionally the highest points of East Antarctic do not reach more than 4000 m, and the coastal and Peninsula regions have topographic slopes which are too low (not shown). Thus the extremes of both coastal slopes and high plateaus are not fully represented by the model. This is due to the relatively coarse resolution of the model grid. Errors in orography also produce the largest temperature errors in comparable isotopically enable GCMs [Hoffmann et al., 1998; Mathieu et al., 2002]. Apart from the orographically induced errors, the mean annual PD experiment TS compares quite well with in situ estimates. Turner et al.  note that the low-temperature inversion is up to 5°C too strong in the central interior. Apart from this region, the basic inversion pattern and strength generally matches the best observations within a few percent. But comparative inversion observations are rather sparse and may themselves not be representative [Connolley, 1996].
 In terms of absolute Antarctic TS (surface temperature) and δ18O values the PD run compares well with the Masson-Delmotte et al.  database (see Figure 2c). (See Tindall et al.  for the global modeled distribution of TS and δ18O values compared to observations.) The lack of the coldest temperatures and most depleted δ18O values in the observations seems to reflect the distribution of the observations with none in the Masson-Delmotte et al.  database for the coldest Dome A region around 85°S 70°E (see Figure 2b). The absence of the warmest TS and least depleted δ18O values on the model grid is partly due to the relatively coarse grid resolution at the coast; the highest TS model sites represent averaged areas farther inland than the observations. The maximum TS in the regridded data is sensitive to whether we interpolate using marine temperature values or use only continental grid cells. If we regrid including marine temperatures, the maximum TS is higher. Using marine grid cells can also bring the modeled δ18O range into closer agreement with the Masson-Delmotte et al.  database. However, here we use a more conservative approach and do not interpolate any values beyond the natural grid continental edge midpoint of the cells. We do the same for the most southerly cells, meaning we lack values for the most poleward and most coastal parts of the continent.
2.3. Forced Warming
 The total mean Antarctic warming of 2.6°C over the experiment set lies quite centrally in the range of the Intergovernmental Panel on Climate Change  CMIP3 SRES A1B output [Bracegirdle et al., 2008]. While different models disagree in detail on the Antarctic climatological changes throughout the A1B scenario [Bracegirdle et al., 2008], this set of experiments provides a basis for assessing the isotope-temperature relationship over a meaningful CO2 forced warming. The warming rate is very similar to the total global warming of 2.7°C although total Southern Hemisphere warming is 1.9°C (Northern Hemisphere is 3.5°C), owing to the slow rate of warming in response to greenhouse gas increase in the Southern Ocean [Manabe et al., 1991] (see Figure 1). Sea-ice changes are important, particularly in the coastal regions [Bracegirdle et al., 2008].
 In Antarctica, recent change has been small over most of the continent [Schneider et al., 2006; Monaghan et al., 2006], but very significant warming has been observed in the Antarctic Peninsula region [Vaughan et al., 2003], and this has been associated with the loss of a number of ice shelves. We attempt to elucidate how the different regions will respond by defining four different Antarctic regions, shown in Figure 3 (the fourth region not labeled on Figure 3 is the whole of the Antarctic), and examine the temperature changes and isotope changes within them individually. Additionally, we look at the isotope response to temperature change at five core sites.
 The Peninsula shows the strongest warming until 2080, probably owing to the greater influence of sea-ice losses on the warming rate [e.g., Bracegirdle et al., 2008]. The East Antarctic shows linear warming after a slight cooling in 2020. By 2100 the warming is of similar magnitude to that in the Peninsula region. The West Antarctic shows quite linear warming, i.e., quicker initially than in the East Antarctic, then a rate slower than the East Antarctic over 2060 to 2100. By 2100 the West Antarctic is slightly less warmed compared to the East Antarctic. See Figure 1 for overall warming. These trends are amplified in the winter mean TS. Summer trends are all more linear. By 2100 the annual mean TS anomaly of the Peninsula is +2.9°C; the East Antarctic +3.0°C; the West Antarctic is +2.4°C. Most of the warming takes place in winter: the whole Antarctic is +3.4°C; the Peninsula is +3.7°C; the East Antarctic +3.9°C; the West Antarctic is +2.9°C whereas in summer the whole Antarctic is +1.8°C; the Peninsula is +1.4°C; the East Antarctic +2.7°C; the West Antarctic is +1.9°C. The autumn and spring pattern of warming south of 50°S is strongly zonally and seasonally dependent. After 2020 all the experiments show relatively linear warming in all seasons and in the mean average. But we note that the 2020 results are slightly anomalous in their seasonality, compared with the PD and the later experiments. This appears to reflect changes in the sea ice since PD uses observed sea ice while 2020 and beyond use HadCM3 sea ice. But generally, all experiments from 2040 onward and all regions show a weak summer warming and a stronger winter warming.
3. Using δ18O as a Proxy for Temperature
 To use δ18O as a proxy for spatial or temporal temperature, it is the general convention to define a gradient as the linear relationship δ18O = aSTS + b so that aS = Δδ18O/ΔTS, where TS is the surface temperature. This relationship can be applied to sets of spatial or temporal δ18O and TS observations to obtain either spatial aSSPACE or temporal aSTIME gradients. We use the least squares fit method since it provides a straightforward means to obtain linear regression fit statistics and gradients for the δ18O against TS relationship.
 Since δ18O is recorded in precipitation, and is therefore already “precipitation-weighted” we also compare δ18O with a precipitation-weighted version of the surface temperature for each longitudinal and latitudinal position (x, y) calculated,
where t is increments of time (herein t is daily, so that t = 1.365) and P is precipitation.
 Warmer air tends to be associated with higher absolute humidity so that warm air masses can contribute more precipitation. Additionally cloud cover will tend to reduce radiative cooling [e.g., Kohn and Welker, 2005]. So changes in TSPt may tend to be smaller than changes in TS. Werner and Heimann  have shown that in some instances TSPt can deviate significantly from TS. Understanding these differences is essential to fully comprehending how changes in TSPt relate to changes in TS and thus how δ18O relates to TS. We call this difference “biasing” B so
 Variations in B (with space and/or climate) causes differences between the spatial and temporal gradients calculated using TS and those that would be obtained using an equivalent precipitation derived ice-core record. It is useful therefore to define (as above) gradients between δ18O and TSPt. We term the new spatial and temporal gradients of δ18O against TSPt: aSPtSPACE and aSPtTIME, respectively. Therefore aSPtSPACE and aSPtTIME are versions of aSSPACE and aSTIME using precipitation-weighted temperature.
3.1. Spatial Gradients of δ18O Against TS and TSPt
 Spatial gradients aSSPACE and aSPtSPACE are presented in Table 2 for the regions shown in Figure 3, and for the whole of the Antarctic. Using standard model latitude longitude gridded output it is not simple, near the poles, to calculate the uncertainty associated with spatial relationships between δ18O and temperature. Adjacent spatial data points cannot be fully independent within a GCM output, or an observation set, because there is correlation between nearby points. But this problem is exacerbated approaching the poles, because the degree of independence decreases with the grid cell size (which is partly why the model uses Fourier filtering near the poles to decrease the model degrees of freedom approximately in proportion with an equal area representation). To counteract these problems of latitudinally dependent grid autocorrelation affecting results, particularly in the calculation of uncertainties, and to allow a better comparison with available data, we regrid the model output to an equal area grid. Equal area regridding onto a 50-km grid gives 20 points to represent a natural model grid point at 65°. This linearly decreases to 12 points at 75°S (at Dome C) and 0 at the poles. We regrid using a linear two-dimensional interpolation model. Note that if an appropriately rotated grid was used, to provide a polar grid which was approximately equal area, this regridding would be unnecessary.
Table 2. Spatial Gradients aSSPACE and aSPtSPACE, Using Mean From 20-Year Experimentsa
Region or Core Site
Using Unweighted TS
Using P-Weighted TSP
Notation: a, gradient; CI95, mean 95% confidence interval for a, assuming model grid point independent at 65°S (see text); EV, explained variance (R2).
Using 20-year means of the PD or 2100 experiment as marked; see Table 1.
The degrees of freedom calculated for this sampling radius are about 15. Any geographical sampling radii less than 300 km gives an insufficient amount of independent locations (too few degrees of freedom in the model output) to obtain a meaningful linear fit.
 The equal area regridded model results are used to obtain aSSPACE and aSPtSPACE. Even using equal area output, it is not clear how to objectively estimate the degrees of freedom in each spatial data set. Here as a simple approximation to the true degrees of freedom we assume the natural model grid points at 65°S are independent. This means that for the 50 km2 equal area grid version each set of 20 adjacent data points represent one degree of freedom. We use this consistently throughout the spatial gradient analysis. While this is still not correct, because the cells at 65°S will be partially spatially correlated, 65°S is generally out with the region of model Fourier filtering and the approach at least gives a consistent approach to estimating the uncertainties.
 The values of aSSPACE shown in Table 2 are similar to the published spatial gradient from observations. Masson-Delmotte et al.  collate the most complete database of surface Antarctic snow isotopic composition and use the data set to calculate spatial gradients. Their δ18O gradient for the entire new Antarctic spatial database is aSSPACE = 0.80 ± 0.01‰ °C−1. This is 0.09 ± 0.05‰ °C−1 lower than the modeled present-day value in Table 2. We note that the PD model value of aSSPACE = 0.89 ± 0.04‰ °C−1 is also higher than previous observation-based estimates of δ18O aSSPACE that were around 0.75‰ °C−1 [e.g., Lorius and Merlivat, 1977; Jouzel et al., 2003, and references therein].
 The Masson-Delmotte et al.  database is not evenly geographically distributed across Antarctica. More than 82% of the observations are located in the East Antarctic region, and there mostly in the sector from 90°E to 180°E. However, our East Antarctic gradient is 0.73 ± 0.08 which matches the observed gradient for the best comparable data in the Masson-Delmotte et al.  observation set. Additionally, we note that our two modeled values (Antarctic and East Antarctic) of aSSPACE enclose the observational based estimate of aSSPACE, and the observation geographical sampling seems to fall somewhere between our definition of an evenly sampled Antarctic and our East Antarctic. Thus the discrepancy between the modeled and observational estimate of aSSPACE appears to be at least as likely due to the uneven geographical sampling of the observations as to inaccuracy, or resolution-dependent imprecision, in the model output. The difference between the 0.75‰ ° C −1Lorius and Merlivat  estimate and the 0.80 ± 0.01‰ °C−1Masson-Delmotte et al.  gradients are also likely to be due to the differing geographic samples.
Masson-Delmotte et al.  examine the geographic pattern of local spatial gradients by calculating aSSPACE using multiple local subsets of δ18O and TS observations. Each subset of values is selected according to whether it lies within a set distance from the location of interest. We follow the same methodology here to examine the variation of aSSPACE across Antarctica for the experiments. Figure 4 shows the results of aSSPACE fits for the PD experiment using fits based on sampling radii as noted. We also calculate values of aSPtSPACE. These are generally slightly higher than aSSPACE, but the spatial pattern is very similar so is not shown.
Figure 4 shows clear geographical structure in aSSPACE (repeated in aSPtSPACE, not shown). The general geographical pattern is similar to that shown by Helsen et al.  obtained using a backward trajectory isotope model, suggesting that the pattern obtained is robust between different types of model. Likewise using sampling radii similar to those used by Masson-Delmotte et al.  the geographical structure of aSSPACE seems to compare both quantitatively and qualitatively quite well to that from the observations. For example both observation and model results show low gradients in central regions; locally around Vostok aSSPACE can be less than 0.2‰ °C−1. Additionally, both show a strong southward decrease in aSSPACE along the Peninsula and high aSSPACE values in the Weddell Sea embayment. The sparse nature of the observations make a more detailed comparison difficult.
 Precipitation-weighted gradients aSPtSPACE tend to be a little higher than aSSPACE (on average about +0.02‰ °C−1). The largest difference between aSSPACE and aSPtSPACE occurs in the Peninsula region. This is due to the large meridional gradient in B over this area, with small B at the northern tip and larger B values inland (not shown here). This B correlates with TS so that aSPtSPACE is larger than aSSPACE since, for the colder southerly Peninsula portions, the reduction in TSPt is lower than the reduction in TS. Thus the spatial reduction in δ18O is applied over a lesser reduction in TSPt than TS, leading to a higher aSPtSPACE value. In general, any regions with a spatial correlation between B and TS are liable to produce differences between aSSPACE and aSPtSPACE. The other regions show relatively small changes between aSSPACE and aSPtSPACE because there is a less systematic relationship between B and TS (and δ18O).
 Interestingly, we find a strong dependence of aSSPACE (and aSPtSPACE) on the applied subset sampling radius. This is illustrated by the change between Figures 4a and 4b, and is clarified in Figure 5a. The mean Antarctic values of aSSPACE and aSPtSPACE increase strongly with the spatial sampling radius used to define the local data subset. Likewise the mean correlations obtained for the linear fits are also related to the spatial radius (Figure 5b). This implies that it is difficult to obtain a local spatial gradient, since it is clearly strongly systematically dependent on the subset sampling radius used. This indicates that a serious problem may occur if we try to define a “representative” local aSSPACE value.
Table 2 also shows core site results for aSSPACE for 2100. None of the gradients calculated are significantly different between the experiments, although there does seem to be slight tendency for the 2100 aSSPACE values to be higher than those for PD. However, in general the model results suggest the observed stability of the spatial slope in the Dome C region over 20 years [Masson-Delmotte et al., 2008] remains relatively unchanged (over the majority of the continent) over this CO2 forced warming event.
3.2. Unforced Interannual Temporal Gradients
 One of the interesting questions in the literature is whether there really is a difference between the values of aSTIME calculated using the short records of temperatures and isotopes over a decade or two where no forced change is applied [e.g., Schlosser and Oerter, 2002; Werner and Heimann, 2002], and those from situations where forced centennial or millennial climate change occurs [e.g., Jouzel et al., 1997]. To examine this we present (unforced, UF) UFaSTIME calculated from PD experiment mean annual values over a 20 year run (Table 3), and compare these to (forced, F) FaSTIME calculated over the 2.6°C forced warming event charted by the 6 experiments (see Table 1) and shown in Table 4.
Table 3. Temporal Gradients UFaSTIME and UFaSPTIME, Using Interannual PD Experiment (Unforced)a
Using Unweighted TS
Using P-Weighted TSP
Notation: a, gradient; CI95, mean 95percnt; confidence interval for a; EV, explained variance (R2). Unforced fits are from 20 years of mean annuals, using annual means averaged across each region for each of the 20 years of the PD experiment.
Table 4. Forced Multiannual Multirun Results for aSTIME and aSPTIMEa
Using Unweighted TS
Using P-Weighted TSP
Notation: a, gradient; EV, explained variance (R2); SIG, significance.
Forced Fits Using 20 Years of Each Run
Mean Summer Results (DJF)
Mean Winter Results (JJA)
Averaging Radius 200 km (51 × 50 km2Grid Points)
 The results for the UFaSTIME and UFaSPtTIME are calculated using the unforced mean annual values for the 20-year present-day run. As in section 3.1 we use the equal area gridded results for the linear least square fitting. In most cases we do some spatial averaging prior to the fitting. This is indicated in each case, as in the previous section, by a spatial sampling radius. Using the same grid and equivalent sampling procedures helps to make the temporal gradient calculated directly comparable to the spatial gradients.
 The linear fit gradients are shown in top of Table 3 and by Figure 6. As observed by Schlosser and Oerter  and modeled by Hoffmann et al.  and Werner and Heimann  the regionally averaged unforced PD temporal fits have low explained variances. The linear fits show that TS generally accounts for less than 10% of the variance in δ18O except in the Peninsula region where the explained variance rises to 30%. Figures 6a and 6b show that the spatial pattern of UFaSTIME is variable with the possibility of better correlations and higher gradients in local, mainly coastal, locations. We have not ascertained here what determines these locations. Note that these short-term (20-year) time series are from simulations run with climatological mean SST and sea ice. This reduces the model degrees of freedom owing to remote large-scale interannual variability (described by, e.g., Schneider and Noone  which is liable to reduce our modeled interannual correlations and gradients between δ18O and TS compared to the observations and the coupled interannual relationships modeled by Schmidt et al. .
 As Schlosser and Oerter  note from observations at Neumayer, the correlation coefficients for UFaSPtTIME using the precipitation-weighted temperature are considerably higher than those found with UFaSTIME (differences between top and bottom panels on Figure 6), as are the gradients themselves. The Antarctic average for UFaSTIME is 0.20‰ °C−1, and for UFaSPtTIME it increases to 0.42‰ °C−1. Thus it seems that as Schlosser and Oerter  suggest, a large degree of δ18O variance unexplained by UFaSTIME may relate to interannual variability in seasonality, or other local parameters. This shows that TS is not very well correlated with TSPt for these interannual temperature values. The coefficients of correlation contoured on Figures 6c and 6d also show that even using UFaSPtTIME, more than 50% of the δ18O variance remains unexplained by a linear fit between δ18O and TSPt. Again we note that this may be an underestimate due to the use of climatological mean boundary conditions.
3.3. Forced Warming Temporal Gradients
 Using the same linear least square fitting procedure we calculate gradients FaSTIME and FaSPtTIME for the forced (F) warming results. These are calculated by fitting a line through the six experiment averages (20-year means). The results in Figure 7 and in Table 4 show that the FaSTIME gradients and correlation coefficients are higher for a forced warming event than for the unforced interannual results. For the forced run results, the regionally averaged results in Table 4 indicate that the forced gradients are very consistent between the different regions: FaSTIME = 0.36 ± 0.04‰ °C−1 and FaSPtTIME = 0.53 ± 0.03‰ °C−1 for each of the four regions examined. In general, the ice core site gradients calculated seem to reflect the regional East Antarctic gradients, although Figure 7b shows that over the whole continent FaSTIME can vary between about 0.1 to 0.6‰ °C−1. EDML and Dome F FaSTIME values are generally a little larger than the regional values. Dyer tends to reflect the regional Peninsula gradient values quite closely. Byrd has a rather low explained variance (EV) value, which suggests that it is difficult to interpret. Generally however, the EV associated with the forced warming linear fits is high, with less than 20% of the δ18O variance left unexplained by the local temperature. This does not leave much room for the fitted gradients to be significantly in error.
 Preliminary Last Glacial Maximum (LGM) model runs using HadAM3 (unpublished) indicate that the gradient FaSTIME for the whole continent is for the LGM to PD climate change 0.92 ± 0.06 ‰ °C−1 and for the East Antarctic FaSTIME = 0.74 ± 0.15‰ °C−1. These results require further experiments and analysis to confirm them, however they agree with previously published LGM to present day model run gradients. Werner et al.  calculate FaSTIME = 0.73 ± 0.06‰ °C−1 for the East Antarctic and 0.85 ± 0.07‰ °C−1 for the West Antarctic. Jouzel et al.  (see their supplements) present a gradient in δD of 6.2‰ °C−1, which is approximately equivalent to FaSTIME = 0.75‰ °C−1 for δ18O. This agreement with previous temporal modeling studies of δ18O and temperature implies that it is unlikely that the low FaSTIME values we calculate over the modeled warming event are specific only to this model.
 We note that the gradient for the Dyer site (typical for the Antarctic Peninsula) is 0.38‰ °C−1. This implies that, for regions that have experienced the warming observed on the western Antarctic Peninsula, represented by the station Faraday/Vernadsky [Vaughan et al., 2003], of 2.8°C in the period 1951–2001, then the forced warming gradient would lead us to expect perhaps an increase in δ18O of approximately 1‰ to have occurred (if this can be considered a forced warming).
Figure 7 indicates that the central East Antarctic Plateau region tends to have low forced warming FaSTIME values. Note, in particular that the Dome C value of the critical FaSTIME gradient is around 0.34‰ °C−1, far below the value equivalent to 0.75‰ °C−1 that was used in recent temperature reconstructions [Jouzel et al., 2007]. This suggests that, near present-day climate conditions, small isotopic changes in the EPICA Dome C region due to GHG change could indicate much larger than expected increases in surface temperature.
3.4. Relationship Between Unforced and Forced Temporal Gradients
 Owing to the low gradients, correlations, and the confidence intervals shown for UFaTIME the regional gradients are not significantly different between the unforced and forced temporal gradients. From Figures 6 and 7 it is difficult to find any spatial relationship between the forced warming gradients and the unforced gradients (which are approximately equivalent to those derived using short instrumental records). Figure 8 shows how they directly relate to each other. UFaSTIME and UFaSPtTIME are plotted against FaSTIME and FaSPtTIME for locations where the minimum correlation coefficient of the two fits is greater than 0.6 (i.e., TS (Figures 8a and 8b) or TSPt (Figures 8c and 8d) explains at least 36% of the local temporal δ18O variance). The bunching of the data points by color shows that there is spatial structure in the relationship. Note that some auto correlation is present in all the model output, regardless of spatial averaging. Therefore trails of autocorrelated points appear in all of the Figure 8 panels. But generally, Figure 8 confirms that spatially there is almost no relationship between the short term instrumental type temporal gradients, and forced warming temporal gradients. This reinforces the observational findings, and suggests that it may not be possible to use short-term (2-decade-long) δ18O against temperature records to interpret warming events in longer-term ice-core records.
3.5. Synoptic Versus Seasonal Precipitation-Weighted Biasing
Table 4 and Figure 7 show that the FaSPtTIME gradients are higher than FaSTIME for each region and most core sites (except for Dome F). Figure 9 shows the difference between the gradients. This strong difference between the temporally precipitation-weighted FaSPtTIME compared to FaSTIME indicates that B is changing over the warming event. For brevity we write the overall change, between the beginning and end of the warming, in B as
where B2100 is the precipitation-weighted biasing (equation (2)) at each location (x,y) for the 2100 experiment and BPD is for the present-day experiment.
 In the PD experiment the inland B varies between +8 and +10°C, see Figure 10a (mean of +9.2°C for East Antarctic; +6.7°C for West Antarctic; and +6.5°C for the more coastal Peninsula). By 2100 this has changed by −0.79°C, −0.79°C and −1.27°C, respectively (see the colored shading on Figure 10a for ΔB). The reduction in B is almost linear over the simulated warming for each of the regions (not shown). This means that for the regional results, ΔB accounts for almost all the difference in the linear fits between FaSTIME and FaSPtTIME; this is why Figures 9 and 10d are similar. (The ΔB changes come out as 26%, 34%, and 41% of ΔTS (where we define ΔTS similarly to equation (3)) for East Antarctica, West Antarctica, and the Peninsula, respectively. This compares with equivalent changes of 25%, 32%, and 45%, in (FaSPtTIME − FaSTIME)/FaSPtTIME for each of the regions.) Lower aSTIME compared to aSPtTIME indicates that precipitation-weighted TSPt tends to attenuate temperature changes through time in TS. This implies that the changes in δ18O related to temperature tend to be smaller than would otherwise be expected.
 We can gain insight into why B changes during the warming event by decomposing B using bandpass filtering. We obtain frequency filtered TS and precipitation P signals. Here we use three frequency bands: high-pass (e.g., high-pass temperature is termed THPS) synoptic (subseasonal) frequencies of <60 days; midpass seasonal frequencies of 61–375 days; and low-pass interannual >375 day frequencies. We then use these band-pass filtered signals of TS and P to generate a frequency decomposed version of B,
T′HP St is the high-pass filtered synoptic (subseasonal) frequency signal of T′S. We define T′S as TS with the time mean TS removed. T′HP St is calculated using a standard inverse discrete fast Fourier transform of T′S. P′HP t is the equivalent high-pass filtered precipitation signal. Frequency decomposed BSEAS (seasonal) and BINTER (interannual) are defined in the same way, using the appropriately frequency filtered signals. Although equation (4) requires BINTER to fully reconstruct B, for the model output here, BINTER is always less than ∣0.5∣ °C, and ΔBINTER (changes between PD and 2100) are generally less than ∣0.1∣ °C. Mean Antarctic ΔBINTER is −0.07°C, σ = 0.12°C. Therefore since BINTER is very small we omit BINTER and ΔBINTER from the following results. Note however, that for δ18O and T data time series with more low-frequency climate variability, or example, with El Niño changes [Tindall et al., 2008], BINTER and ΔBINTER may become important.
 The results (Figures 9 and 10) indicate that the majority of the present-day biasing B (Figure 10a) is due to BSYNOP (Figure 10c). The synoptic biasing of ∼+6°C for the East Antarctic (Figure 10c) is determined by the precipitation regime. It is not clear from observations if a regime that is dominated by clear skies precipitation [Bromwich, 1988] would be liable to substantial subseasonal P against TS covariance. GCMs of the type used here do not specifically parameterize clear-skies precipitation. However, in possible favor of BSYNOP ≈ +6°C, winds from lower warmer regions can correlate with P, so higher-frequency covariance of P and TS seems likely. Bromwich  indicates that for inland East Antarctic precipitation from clouds and clear skies are both generated by the same mechanism of orographic lifting of moist air, so the lack of clear-skies parameterization may not be critical. But, in general, the frequency-magnitude distribution of inland Antarctic precipitation events is not well known, thus it is difficult to further test whether the modeled subseasonal B and Δ B are really representative of the present-day Antarctic or of warmer climate conditions. Further observations of inland P against TS covariance would be useful for testing these BSYNOP results.
 Interestingly, it is changes in both synoptic covariance ΔBSYNOP (Figure 10f) and the lower-frequency seasonal covariance ΔBSEAS (Figure 10e) over warming periods that reduce FaSTIME compared to FaSPtTIME, i.e., reducing overall biasing B over the period of the warming. Where the synoptic biasing difference ΔBSYNOP is negative, this indicates a lesser difference under warmer climate between the temperature under weather systems which carry precipitation and the mean local temperature. The BSYNOP changes could be accounted for by changes in the frequency, magnitude, or duration of the synoptic weather systems. We note that synoptic changes seem to be more important around the coastal region, but it is the seasonal changes which are critical and explain most of ΔB for the forced warming for inland Antarctica. Negative ΔBSEAS is due to the annual precipitation cycle becoming seasonally flatter, or biased toward colder/winter values.
 The ΔBSEAS (changes in the precipitation seasonality over a warming event) reduce the temporal gradients around the Dome C region (and apparently also near Vostok) from about 0.53‰ °C−1 to about 0.34‰ °C−1. Although there are also smaller areas where ΔBSEAS (and ΔBSYNOP in the more coastal areas) is positive, increasing B over the warming event.
 From the precipitation observations available it is difficult to ascertain if the modeled present-day precipitation seasonality is accurate (see section 2.2.1) but we note first that the geographical distribution of precipitation seasonality in Figure 10b is very similar to the summer precipitation pattern shown by the Helsen et al.  modeling study, secondly we do not seem to be in disagreement with the available observations described in section 2.2.1. Together this provides limited support for the general modeled reduction of FaSPtTIME about 0.53‰ °C−1 to FaSTIME about 0.34‰ °C−1 over the GHG induced warming.
 A gap exists between our understanding of Antarctica's short instrumental and satellite records and isotopic ice core data from Antarctica [Schneider et al., 2004; Monaghan et al., 2006]. It is therefore of considerable interest to investigate how isotopes behave under the conditions expected over the next century.
 By introducing an equal area regridding of δ18O and TS we show that the East Antarctic spatial gradient aSSPACE = 0.73 ± 0.08‰ °C−1 we model is not statistically different from the observed gradient for the best comparable data from the Masson-Delmotte et al.  observation set. Additionally, our spatial variation aSSPACE (using our equal area sampled model output) is similar to that observed both by Masson-Delmotte et al.  and modeled by Helsen et al.  in a present-day backward trajectory model. This gives confidence that the general geographical structure in locally subsampled model aSSPACE is robust. We find that the spatial gradients calculated are strongly dependent on the applied subset sampling radius and tend to increase with increasing sampling radii. Likewise the mean correlations obtained for the linear fits are also related to the spatial radius (Figure 5b). While we would expect the correlations, and the gradient to be related (owing to the dependence of the degrees of freedom on radius) this still requires further investigation since it implies that it is difficult to calculate a “representative” local spatial gradient. These systematic sample definition problems may occur, on top of the spatial variation shown by Helsen et al.  and Masson-Delmotte et al.  and here, whenever trying to define a representative local aSSPACE value.
 The temporal δ18O against TS gradients we have calculated here over a 2.6°C global warming simulation are consistent between different Antarctic regions: FaSTIME = 0.36 ± 0.04‰ °C−1 for each of the four regions examined (F stands for forced warming), although local EDML and Dome F ice-core site gradients are a little higher. The Dome C and general East Antarctic FaSTIME is about 0.36 ± 0.04‰ °C−1, about half the value equivalent to 0.75‰ °C−1 that was used in recent temperature reconstructions [Jouzel et al., 2007]. This suggests that, GHG warmings may leave a smaller than expected imprint on the stable water isotope record across Antarctica. Agreement with previous LGM modeling studies of the δ18O and TS relationship suggests that the low FaSTIME values we calculate over the modeled warming event are not specific to HadAM3.
 For the forced warming event, decadally averaged δ18O can explain around 75% of the centennial warming variability in TS. This compares to unforced climate variation from short-term (20-year) time series of annually averaged δ18O which typically explain less than 10% of the interannual variability in TS. We note that these short-term (20-year) time series are from simulations run with climatological mean SST and sea ice. Therefore they are likely to underestimate the true correlations and gradients between δ18O and TS. However, they do provided limited support for the observational findings of low explained variance at Neumayer by Schlosser and Oerter  and modeled results by Werner and Heimann . There is little agreement between δ18O against TS gradients for the forced warming and those from short unforced (climatological mean driven) 20-year interannual time series for any particular Antarctic regions or locations. This would agree with these previous authors who have highlighted the difficulties in trying to use short-term isotope-temperature records to aid in the interpretation of longer-term ice-core records.
 The precipitation-weighted gradient, for the warming event, is larger (at FaSPtTIME = 0.53 ± 0.03‰ °C−1) than the standard gradient (FaSTIME) using surface temperature for the regions examined. For the CO2 warming event modeled here FaSPtTIME is 40–50% larger than FaSTIME for the East Antarctic and Dome C. This is in contrast to modeling study results by Krinner et al. ; Delaygue et al. ; Werner et al.  who note that the glacial-interglacial changes in precipitation seasonality are small for the Antarctic. To investigate this, we have introduced a frequency decomposition of the local influence on δ18O. This decomposition, using bandpass filtering, shows that the majority of the present-day biasing B is due to synoptic (between 1 and 60 day) frequency variations in the covariance in TS and P, but a significant proportion in the East Antarctic plateau region is due to seasonal covariance. Over the warming event simulated this biasing tends to reduce across most of Antarctica. The reduction in FaSTIME (at 0.36 ± 0.04‰ °C−1) compared to the precipitation-weighted temperature FaSPtTIME (at 0.50 ± 0.03‰ °C−1) is mainly due to changes in the seasonal covariance part of B; especially inland in the East Antarctic region. However, changes in the synoptic frequencies have some effect on the spatial pattern of B, particularly in the more coastal regions. The zonal changes in precipitation may be related to changes in the circumpolar low-pressure trough influencing the mean cyclone track [Bromwich, 1988], further investigation of which is merited.
 This study has not addressed other important processes which might be affecting the forced warming gradients. In particular, the model is not currently set up to enable the tracking of water vapor sources, thus we do not know if changes in the temperature of source regions may also be impacting the temporal δ18O − TS gradients, and/or changes in atmospheric overturning circulation [Noone, 2008]. We note that experiment boundary conditions include some sea surface temperature warming at all latitudes, so if the source region remained unchanged than the difference in source and site temperature would be significantly less than the change in site temperature. Additionally, we have not looked for systematic changes in condensation (inversion) temperatures (although the model inversion structure for the present day is reasonable). Source temperature and condensation temperature changes could both contribute to the value of FaSPtTIME = 0.53 ± 0.03‰ °C−1 being lower than equivalent LGM to present-day gradients. Without source tracking, it is difficult to determine source temperature, and space constraints preclude a detailed investigation of condensation temperature changes here.
 There could be implications from these results for detection of warming trends across all of Antarctica. Our results raise the possibility that the commonly used gradients may not be appropriate for all the climate changes observed in the ice core record of past Antarctic climate. Until now, the isotopic experiments have been carried out for the coldest periods in the record, but further study of the different warm periods is clearly warranted.
 We acknowledge NERC RAPID ISOMAP for funding the model development; Masson-Delmotte et al.  for their unpublished manuscript; the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI), and the WCRP's Working Group on Coupled Modelling (WGCM) for their roles in making available the WCRP CMIP3 multimodel data set; and three thorough and helpful anonymous reviewers.