1. Introduction

Climate modeling inevitably requires a compromise between greater model sophistication and realism, on the one hand, and faster, more efficient throughput, on the other. For applications where the trade-off must necessarily emphasize the latter, it is essential to develop and evaluate low-resolution model versions. The low-resolution version based on Community Climate System Model, version 1 (CSM1; Boville and Gent 1998) was developed as an extension primarily for paleoclimate applications and so was referred to as PaleoCSM. It included improved ocean physics and its features included a more realistic El Niño–Southern Oscillation (ENSO) variability (Otto-Bliesner and Brady 2001) and a robust meridional overturning circulation (Otto-Bliesner et al. 2002).

More recently, the Community Climate System Model version 2 (CCSM2) was released (Kiehl and Gent 2004). The atmospheric component was the Community Atmosphere Model version 2.0 (CAM2.0), and the Parallel Ocean Program (POP) replaced the National Center for Atmospheric Research (NCAR) CSM Ocean Model (NCOM). The standard ocean model resolution was nominally 1° in the horizontal with 40 vertical levels. The attempt to replace PaleoCSM with a low-resolution version of CCSM2 was not successful. This deficient model is referred to as 2T31x3 to reflect its CCSM2 base and its horizontal resolution, T31 spectral truncation (3.75° by 3.75° transform grid) for the atmosphere and land, and nominally 3° in the ocean and sea ice components with 25 vertical levels in the ocean. It was found to be unsatisfactory in several respects. In particular, the meridional overturning circulation of the North Atlantic Ocean spins down in present day scenarios (T. Stocker 2003, personal communication), rendering the model unsuitable for studies of thermohaline collapse in past and future scenarios.

An overview of the latest coupled model (CCSM3) is provided by Collins et al. (2006a). The ocean component (Danabasoglu et al. 2006) has two possible resolutions; nominal 1° horizontal with 40 vertical levels and nominal 3° horizontal with 25 levels, which has roughly 17 times fewer grid points. Uncoupled, these POP configurations will be referred to as x1ocn and x3ocn, respectively. The sea ice component, which has the same horizontal resolution as the ocean, is version 1 of the Community Sea Ice Model (CSIM), as described by Briegleb et al. (2004). The CAM3 implementations and performance are described in Collins et al. (2006b). The standard atmospheric configuration has been T42 (2.8° by 2.8° transform grid) for more than a decade (Hack et al. 1994; Kiehl et al. 1998), and the current uncoupled version will be referred to as T42cam, while T85cam and T31cam refer to higher and lower uncoupled resolutions, respectively. The above models are combined in three standard CCSM3 coupled versions: T85x1 for the highest, T31x3 for the lowest and T42x1 for an intermediate. Comparisons of T85x1 and T42x1 physics and solutions can be found in Hack et al. (2006a) and Large and Danabasoglu (2006).

The purpose of this paper is to describe the novel aspects of the T31x3 configuration, contrast its coupled behavior in present-day conditions with the highly constrained forced solutions of T31cam and x3ocn, evaluate its performance relative mainly to T42x1 but also to T85x1, and present the relative computational costs, so that informed decisions can be made regarding the utility of this low-resolution version of CCSM3. In other words, do the benefits of drastically reduced wall clock time and CPU cost outweigh the disadvantages associated with some degradation in the climate simulation? At a minimum, the T31x3 configuration must satisfy the demand for an inexpensive, yet not unrealistic, climate system model suitable for routine multicentury or even longer integrations required for some paleoclimate and biogeochemistry work. All data from these simulations are freely available, and more extensive evaluation beyond what is presented here is encouraged.

It is essential that both high- and low-resolution model evolution follow the same development path so that major new model improvements can span all resolutions. However, the CCSM2 experience demonstrated that in order to qualify as a viable low-resolution climate model, a very different ocean model configuration would be needed. This development is detailed in section 2, together with a description of the resolution-specific modifications made to the atmosphere and sea ice components. The coupled spinup of T31x3 over nearly 900 yr is presented in section 3 and compared to higher-resolution spinups. Section 4 describes how the uncoupled low-resolution atmospheric simulation differs from T42cam and to what extent these differences transfer to the coupled solutions. The quality of the ocean and ice simulations is addressed in sections 5 and 6, respectively. The interannual variability of the T31x3 coupled control, in particular the ENSO simulation, is examined in section 7. The final section compares the computational requirements and efficiencies of the various CCSM3 configurations.

2. Development of low-resolution CCSM3

The strategy adopted for developing a low-resolution version of CCSM3 is based on the primary uses of CSM1 and the experience with CCSM2. It involves reconfiguring the ocean grid, modifying the ocean physics and retuning atmosphere and sea ice parameters. The first priority is to maintain a robust meridional overturning circulation (MOC) in the North Atlantic. Estimates from observations place the strength of this overturning at about 18 Sv with an error range of 3–5 Sv (Talley et al. 2003). The corresponding values from T85x1 and T42x1 are about 22 and 19 Sv, respectively (Bryan et al. 2006). The target minimum for T31x3 is 14 Sv, so that the North Atlantic MOC would not be farther from observed in T85x1.

The second priority is to have an equatorial circulation and ENSO variability that is comparable to that produced at higher resolution. The mean zonal velocity in the Equatorial Undercurrent (EUC) should approach the observed, >100 m s⁻¹ (Johnson et al. 2002), with 80–120 cm s⁻¹ the target range. Equatorial Pacific variability is affected by biases in the ocean mean state, including the simulation of the EUC core and equatorial thermocline. ENSO is not particularly well simulated in T85x1 (Deser et al. 2006), so the hope for T31x3 in this regard is only that its associated interannual SST variance not be significantly worse than at either of the two higher resolutions.

The overall experience with T31 truncation grids in the atmosphere has been positive since the simulation quality is comparable in many ways to T42 but at less than half the computational cost. Nevertheless, there are a number of systematic biases that are intrinsically associated with the lower-resolution grid. The major challenge in configuring a T31 atmosphere for CCSM3 is to maintain the quality of the top-of-atmosphere (TOA) radiation budget, which is strongly modulated by the simulated cloud field.

Finally, the distribution and thickness of sea ice in both hemispheres is considered. Climate sensitivity depends on the thickness, which observations in the central Arctic place in the 2–3-m range. Sea ice coverage depends on many factors including surface winds and ocean currents and heat transport. The tendency for ice area to be too extensive, especially at low resolution, needs to be minimized.

3. The T31x3 spinup

We now examine the first 880 yr of the T31x3 integration under the present-day (1990) atmospheric conditions given by Otto-Bliesner et al. (2006). The ocean component was initialized with World Ocean Atlas 1998 climatology (Levitus et al. 1998), merged with the Polar Science Center Hydrographic Climatology (PHC) Arctic data (Steele et al. 2001), hereafter WOA/P. The model physics remained constant for the final 850 yr of the run, following a change from upwinded to centered-differenced ocean advection at year 30. At year 133, the freshwater imbalance from the Baltic Sea was redistributed (section 2b). This was accompanied by a one-time, nonphysical correction of North Sea salinities back to the WOA/P values.

Figure 3 shows the globally averaged time series of surface temperature for both the T31x3 and T42x1 simulations. By this measure, T31x3 produces a remarkably stable climate. It becomes colder than the observed NCEP climatology by about 0.5°C and does not exhibit the cooling trend seen in T42x1, which by year 800 is about 0.2°C warmer than observed. The surface temperature trend in T42x1 is dominated by the Southern Hemisphere extratropics and is associated with a linear trend of increasing SH sea ice. In contrast, sea ice in T31x3 is stable in both the Northern and Southern Hemisphere, although the ice volume and area significantly exceed observational estimates (section 6).

In the ocean, neither T31x3 nor T42x1 reaches equilibrium by year 880 because of the long deep ocean time scales. However, the drift in global mean ocean temperature of T31x3 is small and nearly linear at approximately 0.01°C century⁻¹ (Fig. 4b). Corresponding rates for the T42x1 and T85x1 controls have the opposite sign and are −0.04° and −0.05°C century⁻¹, respectively. The T31x3 trend reflects a positive bias in total surface heat flux into the ocean of about 0.05 W m⁻² (Fig. 4a), compared to −0.2 W m⁻² for both higher-resolution models. At the same time about 0.05 W m⁻² passes in the bottom and out the top of the atmosphere, which is roughly a quarter as much as in the higher resolutions. By year 800, the global mean ocean temperature is only about 0.04°C warmer than the WOA/P climatology. Time series comparisons to WOA/P of zonally averaged ocean temperature as a function of depth (not shown) indicate that much of the trend in Fig. 4b is due to a warming in the Pacific between 400 and 2500 m. In contrast, the drift in the T42x1 ocean is primarily due to Pacific cooling everywhere below about 1000 m.

The freshening trend in global average ocean salinity (Fig. 4c) is small (−4 × 10⁻⁴ psu century⁻¹), but not as small as either T42x1 (−2 × 10⁻⁴ psu century⁻¹) or T85x1 (−0.5 × 10⁻⁴ psu century⁻¹). The North Sea salinity adjustment is evident at year 133. Around year 800, the global mean salinity is only 0.003 psu fresher than the WOA/P climatology but continues to exhibit a linear trend. This trend is mostly due to freshening above 500 m in the Pacific, Indian, and Southern Ocean regions.

Relatively short-lived transients associated with the spinup have largely disappeared by year 200. Time series plots show that model startup from a state of rest triggers large amplitude fluctuations in almost all global ocean measures. For example, Drake Passage transport (Fig. 4d) drops by more than 40 Sv during the first 100 yr. After year 500 it becomes relatively steady at 110–120 Sv. Despite a slow recovery, this transport is still below the observed range estimated by Whitworth (1983) and corrected by Whitworth and Peterson (1985). Figure 4e shows that the highest priority requirement for T31x3 is achieved. The strength of the North Atlantic MOC, as given by the maximum Atlantic overturning below 500 m and north of 28°N (Fig. 4e, thick line), maintains a steady value of around 16 Sv after initial fluctuations. The range estimated by Talley et al. (2003) is shown for comparison. These large-amplitude early transients underscore the importance of multicentury climate simulations that permit analysis of a climate system that has reached a quasi-equilibrated state.

4. The atmospheric simulation

In most respects, the T31 uncoupled atmosphere (T31cam) bears a close resemblance to the T42cam simulation, and in general there is a similar correspondence between the atmospheric solutions in coupled T31x3 and T42x1. However, prior experience developing low-resolution configurations of the atmospheric model has revealed a few strong resolution-dependent model biases that can have important ramifications in the fully coupled system. The specific T42x1 to T31x3 differences discussed below are concerned with the precipitation, especially in the Tropics, the low-level dynamics (winds), radiation, and surface air temperature. In most cases, uncoupled resolution sensitivity is similar, and so differences between T31cam and T42cam are useful in understanding differences in the coupled solutions.

However, an important example of different coupled and uncoupled resolution sensitivity is seen in the annual average precipitation along the equatorial west Pacific. The region between 150°E and the date line is characterized by a strong west-to-east decrease, which the Large and Yeager (2004) balanced climatology gives as 7.7 → 3.9 mm day⁻¹. This rainfall is similar in T42cam (7.8 → 3.5 mm day⁻¹), but lower in T31cam (6.2 → 3.0 mm day⁻¹). This reduced T31cam equatorial Pacific precipitation gradient is seen in Fig. 5a. The increase due to coupling in T42x1 is only about 10% at 150°E and even less farther east. In contrast, precipitation in T31cam coupled to an x3ocn with a diurnal cycle nearly doubles throughout the region (11.5 → 6.0 mm day⁻¹). The serious consequences of this excessive precipitation are noted in section 2b, and prompted the removal of the diurnal cycle from the T31x3 ocean model. A clean comparison finds that this change alone reduces the precipitation by about 1.7 mm day⁻¹ in the west and by 3 mm day⁻¹ at the date line, so that the gradient in T31x3 (Fig. 5b) becomes a more feasible 10.0 → 4.2 mm day⁻¹. Even so, this precipitation remains larger at low resolution when coupled, as opposed to smaller when uncoupled.

The global distributions of annual average precipitation for T31x3 and T42x1 shown in Fig. 5, and for T85cam, T42cam, T85x1, and T42x1 in Hack et al. (2006b) share many of the same large-scale characteristics. Resolution sensitivity is relatively small compared to the large systematic precipitation differences from observations in both coupled (Large and Danabasoglu 2006) and uncoupled (Hack et al. 2006a) simulations, so comparisons with observations are not repeated in Fig. 5. The most significant bias is the zonal band of excess rainfall in the tropical South Pacific. This pattern is symptomatic of the double ITCZ problem, which is enhanced by coupling. Only the coupled simulations produce the overly extensive rainfall over the tropical Atlantic, because the source of the problem is the warm SST biases that develop off southwest Africa in all coupled configurations (Large and Danabasoglu 2006). Both coupled and uncoupled models produce excessive precipitation over the African continent, and too little in the South Atlantic off the coast of Brazil, but Fig. 5 shows that there is little change in the coupled biases with lower resolution.

The excessive meridional shift in tropical precipitation between December–February (DJF) and June–August (JJA), which occurs in the high-resolution CCSM3 (Hack et al. 2006b) is seen also in T42x1 and T31x3. Zonal mean precipitation curves for both lower resolution CCSM3 controls are nearly identical for boreal winter and summer (not shown), with slightly lower peak precipitation rates than in T85x1. Anemic interannual precipitation variability between 10°S and 10°N in the equatorial Pacific in the T31x3 is also quite similar to that seen in T85x1 (Hack et al. 2006b), a result that is related to deficient ENSO variability in each of the CCSM3 integrations. To first order, the T31x3 exhibits the same mean, seasonal, and interannual precipitation biases as the higher-resolution versions, and is not noticeably worse in terms of simulated hydrological cycle than the more expensive CCSM3 resolutions.

The low-level dynamical circulation in T31cam exhibits large-scale anomalies that have a mostly zonal character, so zonal averages of ocean wind stress components are used to display their coupled manifestation in Fig. 6. In general, model winds are too strong at all resolutions, as shown by the comparison of zonally averaged wind stress magnitude (Fig. 6c). This is especially true at storm track latitudes in both hemispheres and in the Northern Hemisphere trade wind zone, but there is slightly anemic wind stress over the Arctic in CCSM3 due to weaker than observed meridional stress. The observed mean stress is computed from coupling 2000–04 6-hourly blended Quick Scatterometer (QuikSCAT) winds (Milliff et al. 2004) to monthly observed SST. In both uncoupled and coupled atmospheric models there is an unrealistic migration of the Southern Hemisphere storm track toward the equator as resolution is lowered. However, the weaker T31x3 zonal stress and greater displacement conspire to give better agreement with observations at some latitudes, particularly ∼55°S where T31x3 wind stress magnitude coincides with the peak in observed Southern Ocean westerlies. Similarly, T42x1 is an improvement over T85x1 at some latitudes. An unrealistic weakening of westward wind stress in the equatorial Pacific of T31cam relative to T42cam is not a strong bias when the atmospheres are coupled, and the storm tracks are the only latitudes where significant change with resolution is evident in the coupled solutions. The excessive convergence of meridional wind associated with the double ITCZ in the Pacific is present for all resolutions in Fig. 6b. The effects of these dynamically related resolution sensitivities on the ocean and sea ice of the coupled system are discussed in sections 5 and 6, respectively.

The T31cam simulation also exhibits important large-scale differences from T42cam in the radiation budget that are associated with the behavior of parameterized cloud processes. These biases are seen both at the TOA and at the surface and are strongly correlated with similar anomalous structures in the precipitation (e.g., Fig. 5) and precipitable water fields. They are especially apparent in the Indian Ocean extending into the tropical western Pacific, and along the South Pacific convergence zone. Spatially coherent signals exceeding 10 W m⁻² are seen at the TOA in both the longwave and shortwave radiation budgets. The corresponding surface signals are evident in the net surface heat flux difference of Fig. 7, which is dominated by changes in the radiative components. The contributions from the longwave component appear to be associated with biases in clear-sky radiative transfer, which are largely explained by a systematic drying of the atmosphere in regions of deep convection. The net absorbed solar radiation in these regions is also significantly increased, with large regions exhibiting increases of 20 W m⁻² or greater. Wittenberg et al. (2006) show that the range of available estimates of tropical surface heat flux across the Pacific, averaged from 5°S to 5°N, is between 40 and 100 W m⁻². Ranges at least as large are expected at other longitudes, so such estimates are not able to discriminate between T31cam and T42cam fluxes, even though the differences in Fig. 7 are significant. In the coupled models these radiation differences are similar, but noticeably weaker, particularly over the Indian Ocean and tropical West Pacific. More significant energy budget differences are associated with relatively minor shifts in circulation features, and in the distribution of sea ice at high latitudes (section 6).

The change from mid- to low-resolution coupled CCSM3 results in significantly lower surface temperatures throughout the Eurasian Arctic, especially in the Barents Sea region where temperatures drop more than 12°C below the T42x1 mean. This is by far the largest surface temperature difference between the two coupled solutions anywhere. The warm bias relative to observations that exists in T42x1 in this region becomes a cold bias in T31x3, of nearly equal magnitude. The coupled feedbacks related to ice growth in the Barents Sea region complicate the attribution of this bias, which arises as a result of the resolution-related sensitivities of both CAM3 and CSIM and their complex coupled interactions. Although the ice coverage in this region becomes too extensive, the colder T31x3 Arctic has the advantageous effect of reducing the higher than observed DJF land surface temperatures which exist over the Eurasian continent in both T42x1 and T85x1 (Collins et al. 2006a) by up to 4°–6°C.

5. The ocean simulation

Figure 4e shows that after year 400, the strength of the NAMOC, as given by the maximum Atlantic overturning below 500 m and north of 28°N, becomes relatively steady between about 14 and 18 Sv. Multidecadal averages are roughly 16 Sv, which is well within the target for low-resolution CCSM3 (section 2) as well as the error ascribed by Talley et al. (2003) to their observational estimates (18 ± 3 − 5 Sv). For comparison, a typical value for the strength of the Atlantic MOC in low-resolution CCSM2 after 200 yr is only 6 Sv, while for PaleoCSM, the maximum Atlantic overturning was too strong at around 30 Sv. The global overturning in T31x3 generally tracks that of the North Atlantic with a positive offset of about 6 Sv.

The latitude–depth distribution of the mean MOC in T31x3 is shown in Fig. 8, for both the globe and the North Atlantic. The 6-Sv offset is not uniform, but confined to the vicinity of the maximum around 40°N and 700-m depth, in accord with observationally based estimates of ∼8 Sv for the amplitude of the North Pacific deep water cell (Talley et al. 2003). The maximum NAMOC is lower, but not worse, than in both the T42x1 (∼20 Sv; Bryan et al. 2006) and the forced x3ocn (>20 Sv), and the maximum is found at a similar latitude and depth in all three ocean solutions. The T42x1 and x3ocn have very comparable Atlantic overturning streamfunctions with more concentrated flow near 60°N (∼10 Sv reaching ∼1500 m) associated with the deep western boundary current downstream of the Denmark Strait and Faroe Bank overflows. Weaker deep water formation at high latitudes in the Atlantic appears to be the primary cause of the reduced overturning circulation when the low-resolution ocean is coupled to the T31 atmosphere.

The less vigorous overturning in T31x3 is consistent with a much reduced northward heat transport in the Atlantic relative to all other model configurations (Fig. 9, lower panel). The peak value of about 0.8 PW at approximately 25°N is smaller than either inferred from ocean observations, ∼1.27 PW (Ganachaud and Wunsch 2003), or implied by surface heat flux climatologies, ∼1.1 PW (Large and Yeager 2004). There can be little doubt that the T31x3 underperforms in this regard. But it appears that the ocean model is not wholly to blame, since x3ocn forced with observed atmospheric boundary conditions generates a much more reasonable transport. Boning et al. (1996) find a direct linear relationship between North Atlantic heat transport and NAMOC strength, with variations between similar physical models caused by different wind and thermohaline forcing in the north. It follows that the forcing differences between T31x3 and x3ocn are the likely cause of the reduced North Atlantic MOC and heat transport in the former. This weakness of coupled Atlantic heat transport relative to uncoupled is also seen in the high-resolution ocean configurations. However, the coupled configurations generate more global total heat transport because of increased Pacific transport when coupled to an atmospheric model. Apart from uniformly high transports near 50°N, all curves in the global panel of Fig. 9 appear to fall within the error bars of global meridional heat transport obtained from inverse methods applied to World Ocean Circulation Experiment (WOCE) hydrographic data (Ganachaud and Wunsch 2003).

Figure 10 shows how the mean current structure of the equatorial Pacific in T31x3 compares both to observations (Johnson et al. 2002) and the standalone ocean solution (x3ocn). The maximum zonal speed of the EUC in T31x3 is less than 90 cm s⁻¹ (bottom panel), but still within the target range (section 2). Westward surface currents extend too deep in the eastern half of the Pacific in both coupled and uncoupled ocean solutions compared to observations, but this is a bias seen in the high-resolution ocean solutions as well (Fig. 10 of Large and Danabasoglu 2006). In the west, there is too much vertical shear near the surface of T31x3 because low wind variability fails to generate the westerly wind bursts seen in observations (and present in the observed forcing of the x3ocn), but again, the low-resolution model would appear to be no worse in this regard than T42x1 or T85x1 (Large and Danabasoglu 2006). The most significant degradation of T31x3 relative to x3ocn (and T42x1) is that the core of the EUC west of about 230°E is constant at about 100-m depth. In contrast, the observations show that the EUC core deepens westward of 230°E, reaching ∼200 m at 160°E (top panel). As a result of this bias, the EUC source waters are too warm in T31x3.

A series of sensitivity experiments have shown that this flattening of the EUC core is related to excess precipitation in the western Pacific warm pool. In T31x3, this problem would be catastrophic if the model were configured with an ocean diurnal cycle because the resulting warmer equatorial SST would increase the precipitation and stabilize the ocean, thereby increasing the SST even more. Therefore, the cold SST bias relative to the Reynolds and Smith SST climatology (Reynolds and Smith 1994) in the central equatorial Pacific (Fig. 11) is essential in order to avoid such a runaway situation. Thus, removing the ocean diurnal cycle in the low-resolution CCSM3 improves the subsurface equatorial solution, but at the cost of physical realism. Another consequence of excess coupled model rainfall, in particular south of the equator, is that this more symmetric forcing produces zonal flow that is also much too symmetric about the equator. For example, both T31x3 and T42x1 generate both northern and southern branches of the westward-flowing South Equatorial Current (SEC), but in T31x3 (as well as T85x1), the SEC is nearly as strong south of the equator as to the north, instead of being much weaker as in observations (see Fig. 11 of Large and Danabasoglu 2006).

The most serious deficiencies of the SST simulation in T31x3 are the same as those seen in the higher resolution CCSM3 controls: large errors in the vicinity of poorly represented western boundary currents as well as in the eastern boundary upwelling regions of the major basins (Large and Danabasoglu 2006). While the mean equatorial Pacific SST has a more negative bias in T31x3, the seasonal cycle along the equator is not obviously worse than in the highest resolution simulation. It has the same erroneous double peak east of 200°E seen in the T85x1 (Large and Danabasoglu 2006). In T31x3, the amplitude of the seasonal variation is too large as opposed to too small in T85x1, but the same phase biases are present in both configurations.

The existence of warm mean SST biases in the stratocumulus regions off the subtropical continental west coasts of South America (Peru/Ecuador/Chile), North America (Baja/Southern California), and southwest Africa is a problem in all CCSM3 configurations, and demonstrated for T85x1 by Large and Danabasoglu (2006). In these eastern subtropical ocean regions, the two most significant differences between T42cam and T31cam are the representation of stratus clouds and the overall wind stress forcing of the ocean. Potentially problematic is the tendency for both to amplify the warm SST bias. T31cam exhibits a reduced stratocumulus cloud cover in the oceanic regions one to two grid points off the coast, resulting in significantly increased absorbed solar radiation which can easily exceed 50 W m⁻² seasonally. Also, the upwelling favorable longshore surface wind stress is too weak in T42cam compared to observations and even weaker in T31x3 (not shown). Such weakening of the subtropical dynamical circulation would be expected to produce less coastal upwelling, and contribute to even warmer surface temperatures.

The severity of eastern boundary SST anomalies at all coupled resolutions is quantified in Table 1, which lists the climatological difference of model SST from observed, averaged over strips within 15° longitude of the west coasts. Unexpectedly, T31x3 has biases lower than T42x1 along all subtropical eastern boundaries and lower than T85x1 everywhere but along the coast of South America. This result likely follows from x3ocn exhibiting generally less of a bias than x1ocn. At all resolutions, coupling exacerbates these ocean biases. Lower SST anomalies in the low-resolution ocean are related to colder subsurface temperatures, not enhanced coastal upwelling. This suggests that there are differences in large-scale ocean circulation between the models that account for the differences in severity of the problem, and which more than compensate for the inherent warming tendencies of T31cam. However, the T31x3 bias off Africa appears to be still too large to improve the tropical Atlantic precipitation (Fig. 5), as was achieved with prescribed coastal temperatures and salinities in Large and Danabasoglu (2006).

Table 2 compares various aspects of ocean circulation in T31x3 to other model configurations and to a set of observed ocean benchmarks: NAMOC (Talley et al. 2003), peak northward Atlantic heat transport (NAHT; Bryden and Imawaki 2001), volume transport between Florida and Cuba (FCT; Hamilton et al. 2005), Drake Passage transport (ACC; Whitworth 1983; Whitworth and Peterson 1985), the Indonesian Throughflow (ITF; Gordon 2001), and the Bering Strait Throughflow (BST; Roach et al. 1995). Both the ACC transport through Drake Passage and the Gulf Stream transport between Florida and Cuba are too small, but probably for different reasons. The ACC compares quite well to observations in both x1ocn and x3ocn, so it is likely the coupled forcing that is to blame; T31 storm-track migration toward the equator (section 4) implicates the zonal winds. The Southern Hemisphere westerlies that drive the ACC are too strong in all coupled configurations, but the latitude of the peak in zonal mean winds shifts systematically northward with decreasing atmospheric resolution. In the case of T31x3, this shift is nearly 10° at Drake Passage, which results in a zonal stress that is weaker than observed at ACC latitudes by as much as 0.07 N m². As a consequence, the T31x3 ACC is low, but the T42x1 and T85x1 transports are higher than observed because these configurations generate generally stronger than observed stress over latitudes between 50° and 60°S (Fig. 6).

The Florida–Cuba transport (FCT) is too small in both uncoupled and coupled low-resolution configurations, but too high in both x1ocn and T42x1. This suggests that the larger lateral viscosity required by the lower resolution numerics, and the poorer representation of ocean topography and coastlines retard the transport in both the forced x3ocn and coupled T31x3. Other factors such as sea ice extent may be contributing to the smaller than observed transport from the Pacific to Arctic through the Bering Strait, because the x3ocn value is more reasonable. Finally, the ITF in T31x3 falls in the estimated range, while this transport appears too strong in T42x1.

6. The sea ice distribution

The equilibrium ice model solution in T31x3 is characterized by excessive Northern Hemisphere (NH) ice. Figure 12 shows mean aggregate ice area and ice thickness from the final 5 yr of the integration. The thick line in the ice area plots (top panels) shows the observed climatological location of 10% ice coverage derived from 1979–99 Special Sensor Microwave Imager (SSM/I) satellite data (Comiso 1999). Apart from a small region in the Greenland Sea, the NH ice edge is too extensive throughout the Arctic. In contrast, both higher-resolution configurations of CCSM3 show deficient ice coverage in the Barents Sea (Holland et al. 2006). The T31x3 ice model bias is related to an atmospheric surface temperature cold bias in the Barents Sea of more than 12°C relative to T42x1 (section 4). Both configurations generate surface temperature biases in this region relative to observations, but of opposite sign. The NH ice thickness distribution is qualitatively quite similar to that of T42x1, but thicker; the mean ice thickness in the central Arctic is 3.5–4 m, compared to the observed value of 2–3 m and the T42x1 value of 2.5–3 m. As in T42x1, there is an unrealistic accumulation along the coast of eastern Siberia and deficient ice buildup along the Canadian coastline, the latter of which DeWeaver and Bitz (2006) have linked to poor Arctic summer surface wind forcing at low atmospheric resolution.

In the Southern Hemisphere (SH), ice concentration in the T31x3 is reduced relative to the T42x1 in the eastern Atlantic and Indian Ocean sectors, resulting in somewhat better agreement with the line of observed 10% ice coverage (cf. to Fig. 5 of Holland et al. 2006). There is too much ice coverage in the quadrant centered at Cape Horn. This is probably related to biased wind forcing in this region (Fig. 6), among other factors. There is also excessively thick ice on the eastern side of the Antarctic Peninsula, although it would not appear to be significantly worse than in T42x1 (Holland et al. 2006).

In general, there is increasingly excessive NH ice coverage as CCSM resolution is lowered. Figure 13 shows that the T42x1 ice area bias in the NH is roughly doubled in the T31x3 throughout the year, with the largest deviation from observed in the wintertime. However, T31x3 aggregate ice coverage in the SH is less extensive than T42x1 from summer through winter, and hence more in line with observations (Fig. 13, lower panel). As in the NH, the largest deviations from observed sea ice area occur in the wintertime.

7. Interannual variability

ENSO-like variability in the T31x3 is qualitatively quite comparable to the observed record over one particular 50-yr period near the end of the simulation. The minimum and maximum Niño-3.4 region anomalies over these years (830–880) are −1.7° and 2.9°C compared to −1.9° and 2.7°C for observations between 1950 and 2000. The frequency of large-amplitude anomaly events is also very comparable to the observed record. The number of large, positive Niño-3.4 anomaly events (>1°C) over the time period above is eight for T31x3 and seven for observed; the number of large, negative events (<−1°C) is five for T31x3 and eight for observed.

By employing a moving 50-yr window over several hundred years of model integration, the mean and range of the standard deviation for each Niño region was computed for the three different CCSM3 configurations. The results (Fig. 14) show that there are significant variations in modeled ENSO variability over the course of the control simulations. Whereas comparison of individual observation-length segments usually highlights differences in Niño variability between the CCSM3 resolutions, the overlap of the standard deviation ranges suggests a basic similarity. As is common in coupled climate models (e.g., see Wittenberg et al. 2006), equatorial SST variability is relatively high in the western Pacific (Niño-4, Niño-3.4), and low compared to observed in the eastern equatorial Pacific (Niño-3, Niño1 + 2). The natural rise in SST variability from the west to the east in the Pacific does not occur in CCSM, at any resolution. Although T31x3 has the lowest mean variance in the three easternmost Niño regions, it is highest in the Niño-4 region. As hoped, its range includes the mean values of the higher resolutions in all four measures.

Despite the qualitative realism of the T31x3 Niño-3.4 time series mentioned above, comparing the power spectra of 50-yr segments of Niño-3.4 from the model with that of the data record over 1950–2000 reveals a general shift in the peak of power toward higher frequencies than is observed, a result seen in both higher resolution configurations (Deser et al. 2006). For example, over model years 830–880, both T31x3 and T42x1 show broad peaks in power centered near a period of 2 yr instead of near 4 yr as in nature, with less overall variance in T31x3 than in T42x1.

However, wavelet analysis reveals that over the course of the T31x3 simulation, time periods can be found during which there is a much more realistic peak of Niño-3.4 spectral power than in T42x1. Figures 15a,b show the wavelet power spectra of the Niño-3.4 index for T31x3 and T42x1, respectively, over 400 yr of integration near the end of the runs. To the right, time-averaged wavelet power of model Niño-3.4 anomalies are compared to observations (1950–2000, in red). There is a clear focus of wavelet power near a period of 2 yr in T42x1 throughout the 400 yr, but the period of peak power is much less well-defined in T31x3 and occasionally shifts to longer periods. During the time interval 650–700, in particular, Niño-3.4 wavelet power in T31x3 peaks between a period of 4 and 6 yr, generating a time-average spectrum whose shape closely resembles observed, but with lower maximum power (Fig. 15a, green curve). The long-term mean (480–880) for T31x3 does peak near a period of 2 yr, but of course no observed record of equivalent length is available for comparison.

In contrast, no 50-yr interval can be found when the T42x1 wavelet power shows a similar shift to longer periods. Although the time period 810–860, for example, does show a relative increase in power at longer periods, the peak remains at 2 yr (Fig. 15b, green curve). For T42x1, the long-term mean power curve (480–880) faithfully represents the frequency distribution of power for observation-length segments of the control integration.

The time history of scale-averaged wavelet power in the period band of 3–8 yr [equivalent to average variance in this band, see Torrence and Compo (1998)] is shown for T31x3, T42x1, and observations in Fig. 15c. This frequency band is where observations of the last half-century show maximum power for Niño-3.4 SST. The T31x3 integration goes through several multidecadal segments when variance in this band increases dramatically, to levels comparable to observations. The intervals 660–690 and 700–730 are particularly notable. Niño-3.4 variance in the T42x1 and T85x1 (not shown) control integrations does not reach the same levels in the 3–8-yr band.

The T31x3 simulation of other major modes of climate variability shows the same basic level of skill as in T42x1. with significantly greater discrepancy between CCSM3 and observation than between different versions of the CCSM3 model. Figure 16 shows the first empirical orthogonal function of mean December–March (DJFM) sea level pressure north of 20°N (top panels) and monthly nonseasonal sea level pressure south of 20°S (bottom panels) for T31x3 (years 700–879), T42x1 (years 700–879), and National Centers for Environmental Prediction (NCEP) observations (1948–2002). The observed patterns of pressure variation are known as the Arctic Oscillation (AO) and Antarctic Oscillation (AAO), respectively. We have used the full NCEP–NCAR reanalysis back to 1948 for both hemispheres, despite indications that data quality over Antarctica is lower prior to 1979 (Marshall 2002). Both model resolutions generate an AO that is much more tripolar than observed, with a strong North Pacific signal that is barely seen in nature. This mode explains more variance in both models than in observations, and it appears to be more strongly exhibited in T31x3, with larger amplitudes and even greater variance explained.

In the Southern Hemisphere, both T31x3 and T42x1 generate an AAO that is too weak over the continent of Antarctica and too strong in the band between ∼30°–50°S. Extensions of the polar maximum into the Atlantic, eastern Indian, and eastern Pacific sectors are not as pronounced as observed, at either resolution. The T42x1 does seem to do a somewhat better job than T31x3 of reproducing the enhanced variability which is observed in the Southern Ocean near 120°W. Still, the resolution-related differences between T31x3 and T42x1 are slight compared to the inherent biases seen in the CCSM3 family of model solutions.

A 1% yr⁻¹ increasing CO₂ experiment branched off of the T31x3 control at year 400 indicates that the transient climate response of the T31x3 (change in global average surface air temperature at the point of doubling of CO₂) is 1.4°C. This value is a 20-yr average centered about the point of doubling. The equivalent numbers for the T85x1 and T42x1 resolutions are 1.5° and 1.4°C, respectively. The transient response to greenhouse gas forcing in fully coupled CCSM3 does not show an unambiguous increase with increasing resolution, as is found to be the case for CAM3 equilibrium sensitivity (Kiehl et al. 2006). At the point of quadrupling of CO2 in the 1% increase experiments, the response is 3.5°, 3.3°, and 3.4°C for T85x1, T42x1, and T31x3, respectively. Thus, the climate sensitivity of the fully coupled low-resolution CCSM3 is not significantly different from that of the higher-resolution configurations.

8. Comparative computational efficiency

The significant economies associated with the low-resolution CCSM3 are quantified in Fig. 17. Performance data compiled from load balancing tests run on a variety of platforms have been plotted for each model configuration. The number of years of coupled model integration achievable per wall clock day is related to the total number of CPUs applied. The points plotted generally represent the best of a series of performance tests, and all ordinate values should be understood as approximate. Some of the platforms included are experimental at this stage. Also, the load balancing work has not been completed, and further refinement is likely to result in increased performance on the machines at Oak Ridge as well as on the Linux clusters at NCAR (Intel Xeon).

Direct comparisons between resolutions are only possible for select configurations. On the NCAR IBM Power 4, with 128 CPUs, going from T85x1 to T42x1 results in an increase in simulated years per day (syd) by more than a factor of 2.5. On the Cray X1 at Oak Ridge, T31x3 is more than 3 times faster than T42x1 when both models are run on 76 (multistream) processors. This results in a model throughput of 35 syd, the highest yet achieved for any coupled CCSM3 configuration. Running T31x3 on 16 processors of a Linux server (NCAR, Intel Xeon) generates as many simulated climate years per day as running the T85x1 on 192 processors of an IBM power 4 supercomputer.

The slope of the line through the data point and the origin in Fig. 17 gives the simulated years per day per CPU, a measure of efficiency. Higher slopes are more desirable, indicating that more climate simulation can be completed with fewer resources. The three rays drawn show the maximum efficiency achieved at each resolution of CCSM3. All of the T31x3 test cases have higher efficiency than the most efficient T42x1 case. As expected, T85x1 is the least efficient configuration, and all of these cases fall in the lower right-hand sector of the plot where large increases in CPU power are needed to achieve even modest gains in model throughput. Comparing performance numbers on either of the two IBM supercomputers at NCAR shows that there are much higher efficiency gains going from T42x1 to T31x3 than from T85x1 to T42x1. This is related to the simultaneous reduction of both atmosphere and ocean resolutions in the low-resolution CCSM3. Changing from T42x1 to T31x3 reduces the number of atmosphere grid points by almost a factor of 2, but it reduces the number of ocean grid points by a factor of almost 17. This drastic reduction in resolution puts T31x3 in a performance class by itself.

9. Discussion and conclusions

The results of the previous sections show that several features of the coupled climate at T31x3 are notably worse than in T42x1: the ice in the Northern Hemisphere is even more excessive, the Atlantic heat transport is relatively anemic, and SH storm tracks are shifted farther toward the equator. Many of these biases can be traced to inherent deficiencies of the individual component models at low resolution. Although the T31cam solution is similar in most respects to T42cam, there are low-level dynamical circulation differences as well as systematic biases related to parameterized cloud processes. Weaker deep water formation in the x3ocn contributes to a less vigorous thermohaline circulation and an anomalously low heat transport in T31x3. In some instances, however, the uncoupled biases do not leave strong signatures in the coupled solution. For example, large-scale radiation budget biases in T31cam are not as large in T31x3, and reduced stratocumulus and coastal wind forcing off subtropical west coast regions do not exacerbate the positive SST biases in the coupled context.

In fact, many aspects of the low-resolution coupled solution compare quite favorably with the higher-resolution configurations. The T31x3 generates a more stable climate than T42x1, with less ocean temperature drift, which increases the utility of T31x3 as a tool for climate studies. There does not appear to be a systematic degradation in modeled ENSO-like variability as CCSM3 resolution is lowered. On the contrary, T31x3 switched at least once into a regime where ENSO variability has a quite realistic spectral power distribution, unlike higher-resolution configurations in which ENSO power consistently peaks at a period of near 2 yr. A related result is that the T31x3 maintains a passable Pacific equatorial undercurrent despite having less than one-third of the longitudinal resolution of the T42x1. The eastern boundary ocean SST bias, which is present in all configurations of CCSM3 is least severe in the T31x3. The ACC transport in T31x3, while too weak, is closer to the observed value than in either high-resolution coupled configuration. The Indonesian Throughflow is within the observed range, although probably not for the correct reasons. Ice coverage and thickness in the Southern Hemisphere appear to be at least as good as in T42x1, and the seasonal cycle of total SH ice area is slightly closer to observed in the T31x3 configuration. The simulation of modes of atmospheric variability such as the AO and AAO and the transient climate response to anthropogenic forcings are not significantly degraded in the T31x3 compared to the more standard CCSM3 configurations. Finally, the magnitude of the meridional overturning in T31x3 is within the error bars of observation and maintains its strength over many hundreds of years of integration. This represents a significant improvement in CCSM low-resolution modeling.

Whether or not the shortcomings of the T31x3 climate are acceptable in light of the very large gains in efficiency described above (Fig. 17) is clearly a question that must be answered by the individual researcher. This evaluation will necessarily depend upon the nature of the phenomena under investigation. But efficiency is not the only benefit of T31x3; its unexpected skill in several measures relevant to climate studies will also recommend its use.