General circulation models (GCMs) designed for projecting climatic change have exhibited a wide range of sensitivity. Therefore, projected surface warming with increasing CO2 varies considerably depending on which model is used. Despite notable advances in computing power and modeling techniques that have occurred over the past decade, uncertainties of model sensitivity have not been reduced accordingly. The sensitivity issue is investigated by examining two GCMs of very different modeling techniques and sensitivity, with attention focused on how moisture processes are treated in these models, how moisture simulations are affected by these processes, and how well these simulations compare to the observed and analyzed moisture field. Both GCMs predict increases of atmospheric moisture with doubled CO2, but the increment predicted by one model is substantially higher (approximately twice) than that predicted by the other. This same difference is seen in responses of the boundary layer diffusive moistening rate. Calculations with a radiative–convective model indicate that the differences in predicted equilibrium atmospheric moisture, including both column amount and vertical distribution, have contributed to the largest differences in model sensitivity between the two models. We argue that in order for climate models to be credible for prediction purposes, they must possess credible skills of simulating surface and boundary layer processes, which likely holds the key to overall moisture performance, its response to external forcing, and in turn to model sensitivity.
General circulation models (GCMs) will likely remain primary tools for climate prediction in the foreseeable future. Model simulation of surface warming for a given forcing, or climate sensitivity, has, however, exhibited large uncertainties. For example, projected surface warming due to doubled CO2 can differ by a factor of 2, based on two versions of the National Center for Atmospheric Research (NCAR) Community Climate Model (CCM) released 10 yr apart (Kothavala et al. 1999; Hu et al. 2000). Large differences in model sensitivity are not uncommon to GCMs; for instance, the most recent Intergovernmental Panel on Climate Change (IPCC) Assessment Report (Houghton et al. 2001) gives a sensitivity range of 2.0–5.1 K with doubled CO2, using projections from 15 newer generation GCMs worldwide. And, in fact, this range of sensitivity has hardly changed since the previous report (Houghton et al. 1996), relying largely on older models.
Attempts have been made to explain the differences in model sensitivity by differences in simulated changes of atmospheric moisture, including both column amount and vertical distribution. Modeling of convection has been suggested as the primary responsible process for the differences in sensitivity inside the Tropics (Hu et al. 2000). Given the abundances of atmospheric moisture in different phases, its large spatial-temporal variability and fast response to a forcing, and the fact that water vapor is the most important greenhouse gas in the atmosphere, climate modeling has indeed evolved into a comprehensive test of our knowledge of the hydrological cycle (see Chahine 1992; Hack et al. 1998) and, in particular, its dependence on a wide variety of physical processes. The connection between moisture simulation and climate deserves close scrutiny as it likely holds the key to our understanding, and eventually resolving, issues surrounding model uncertainties. Toward this end, we examine systematically how atmospheric moisture processes are treated in climate models, how model moisture simulations are affected by these processes, and how well these simulations compare to the observed and analyzed moisture field.
Atmospheric moisture distribution is controlled by processes of extremely diverse scales, both in time (e.g., typical convective events last minutes to hours, while major climate events may take decades or longer) and space (from molecular, cloud microphysical, to circulations of global significance). Even with today’s most advanced numerical models, it is still impossible to resolve fully processes and motions of such diverse scales, let alone their potential interactions. This is not only because of the exceptional computational demand, but also because of our inadequate understanding and knowledge of how to represent these processes within a unified modeling system. So far, only large-scale advection is adequately resolved in current-generation GCMs, leaving all other processes, such as convection, cloud and radiation, and surface and boundary layer (BL) processes, to name a few, to rely on parameterizations of various degree in order to bring the system to a “closure,” or simply to make it “more manageable.” Consequentially, model sensitivity can differ depending on how, and how well, these processes are parameterized in individual models. A systematic examination of how well these processes are represented by climate models, and how these processes behave and respond to external forcing, both individually and collectively, is of great interest to climate research.
In this paper, we analyze model moisture processes and model sensitivity by examining closely two NCAR CCMs, the CCM1 and CCM3 (Williamson et al. 1987; Kiehl et al. 1996). The two models are chosen primarily for their known differences in modeling approaches and sensitivity, while also considering their large user pool supported by clear and complete documentation. Our uncertain knowledge of the actual atmospheric moisture field must be put into proper perspective while we evaluate model performances, hence we first compare the CCM performances with other available sources, including moisture observations from the National Aeronautics and Space Administration (NASA) Water Vapor Project (NVAP; see Randel et al. 1996), moisture simulation made by the newest version of the NCAR Community Atmospheric Model (CAM2), and the analyzed moisture field from the National Centers for Environmental Prediction (NCEP; see Kalnay et al. 1996) and the European Centre for Medium-Range Weather Forecasts (ECMWF; Gibson et al. 1997). Among them, NVAP is the only source derived from direct observations, although it has its own uncertainty (to be discussed later), whereas the reanalyses are merely “quasi-observations,” as they combine observations with models. Because of the frequent forced initialization using observations, however, the reanalysis models are expected to mimic more closely the working of the real atmosphere than does a “pure” model; moreover, the reanalysis models are able to relate the moisture field to specific moistening processes, making possible the evaluation of GCM performance down to the level of individual processes.
While noting that models have been undergoing constant improvement over time—during which many modeling techniques have been improved or simply replaced—it is important to be able to identify and connect model performance with modeling processes and to view these changes through the history of model development. For models intended for climate prediction, it is essential for them to reproduce the observed climate statistics with the right physics; in particular, they must be able to produce the right moisture field with the right moistening processes in order to be useful for prediction. It should be noted, however, that even if a model performed reasonably when simulating the present climate, it is still unclear whether its response to a perturbation remains credible. Hence evaluating model performance through individual processes is only a necessary, rather than sufficient, condition for evaluating a model’s prediction skills (see also Houghton et al. 2001). In this paper, we limit our effort to linking modeling processes to model sensitivity, rather than to assessing general prediction skills of the models.
In section 2, we shall summarize the models used in this study and their baseline moisture performances. We shall also outline our analysis method. Comparisons of individual moisture processes and their moistening rates are presented in section 3, along with our analyses. In section 4, we attempt to connect modeling processes with model sensitivity by showing that modeling of the boundary layer and surface processes likely holds the key to overall moisture simulation and, in turn, to model climate sensitivity. Finally, in section 5, we shall summarize our findings and further discuss areas of potential importance to model sensitivity.
2. Models, performance, and approach
a. Models and performances
The models used for this study are the NCAR CCM1 and CCM3, although we shall often need to compare them with the reanalysis models in order to gauge the CCM performances. The CCM1 is a climate model built in the late 1980s, featured by the use of diurnally averaged solar forcing, a simple treatment of convection with convective adjustment, an Eulerian moisture transport with “numerical fixing” (to avoid negative values), a “bucket-type” land surface model, and an optional slab ocean model with a constant mixed layer depth, but without an explicit treatment of the atmospheric boundary layer. CCM1 represents a generation of climate models built at a more formative stage, with the intention of achieving tangible results while staying within the computational capability of the time. The CCM3, on the other hand, is a model that collectively combines major advances in modeling techniques developed up to the late 1990s, featuring the use of diurnal solar forcing; sequential executions of deep and shallow convection routines; the conservative semi-Lagrangian approach for moisture transport that eliminates numerical fixing; an explicit treatment of boundary layer process with prognostic layer height and nonlocal transport; a more sophisticated land surface model that accounts for different soil, vegetation types, and soil colors; and the coupling of it with an ocean model with variable mixed layer depths. The NCAR CAM2 inherited the CCM3 dynamical core, but with improved vertical resolution and a further improved land surface model. The CAM2 simulation used here was performed with prescribed sea surface temperature (SST). The NCEP and ECMWF reanalysis models (Kalnay et al. 1996; Gibson et al. 1997; also see http://www.ecmwf.int/research/ifsdocs), in contrast, are often precursors of next-generation GCMs, with fine-tuned processes and subsystems, although in some cases merely different implementations of the same modeling techniques used by other models. The reanalysis models are intended for short integration only, thus allowing for higher spatial resolutions. Both the NCEP and ECMWF models adopt the now-popular semi-Lagrangian approach for moisture transport, and both models use prescribed SST over the oceans. Each model, however, uses its own land surface parameterizations. In addition, NCEP adopts a simplified Arakawa–Schubert parameterization for convection, whereas the ECMWF uses an approach that combines deep, shallow, and midlevel convections in the same model, similar to the CCM3 approach. A comparative summary of the models is given in Table 1.
Besides the advantage of clear documentation that facilitates understanding, the two NCAR CCMs are chosen primarily for their reported differences in modeling approaches and model sensitivity, the primary focus of this paper. In spite of the similarities noted above, there exists one crucial difference between the CCMs and the reanalysis models in that the latter use comprehensive observations for frequent model initialization. This process allows the analysis models to produce regularly gridded, global atmospheric datasets that are not only consistent with the governing laws of physics, but also in close agreement with observations. For this reason, the analyzed fields are treated as quasi-observations and used to evaluate pure model performances such as those from the GCMs. This is particularly helpful when the roles of individual moisture processes, often not observable, are the subjects of investigation. The NVAP dataset, in comparison, is a composite dataset combining several observational methods, including microwave and infrared remote sensing, as well as the traditional radiosonde network (Randel et al. 1996); it makes possible a regularly gridded, global dataset of layer-integrated moisture amount for three vertical layers of great importance to climate, that is, surface–700, 700–500, and 500–300 mb. Although this vertical resolution is rather poor and further limited by an observational uncertainty of about 15%, NVAP still provides us with a much-needed benchmark dataset for model validation purposes.
Presented in Fig. 1 are intercomparisons of layer-integrated moisture amount in three vertical layers at two seasons [December–January–February (DJF) and June–July–August (JJA)]. Compared are simulations from the two CCMs, the CAM2 model, analyzed fields from the NCEP and ECMWF, and the NVAP observations.
All models produce credible moisture distributions that are in reasonable agreement with each other and with the observations. The latitudinal dependence of moisture amount and seasonal shifts of the tropical moisture maxima are clearly depicted by all the models involved, and in general, closer agreements are seen in the DJF than in the JJA seasons. We shall focus our attention on comparisons with the NVAP as it is the only source based on direct observations. When so compared, ECMWF exhibits a moderate moist bias in the lower and a larger dry bias in the upper tropospheres while showing somewhat better agreements in the middle. In contrast, NCEP shows improved moisture performance, in particular for the upper troposphere. Among the NCAR models, no significant differences are seen between the CCM3 and CAM2 (a reminder that the two share the same dynamical core), but systematic differences between the CCM1 and CCM3 are readily identifiable. In almost all cases, CCM1 is significantly drier than NVAP, and the differences are particularly large in the middle and upper tropospheres for JJA in the Northern Hemisphere. In comparison, CCM3 shows much improved agreement with NVAP in the lower and middle tropospheres, although a moist bias is seen in the upper troposphere during JJA—at a latitude surrounding the intertropical convergence zone (ITCZ). Finally, all models appear to exhibit significant dry biases in the Southern Hemispheric upper troposphere; however, poor network coverage, along with existing difficulty of measuring moisture in the cold and dry upper troposphere using radiosonde (Elliott and Gaffen 1991), may be partially responsible for the discrepancies.
The analyzed moisture field also shows significant departures from the NVAP as evident in the JJA months. This is likely due to the different implementations of the parameterized physics in individual models (see Trenberth and Guillemot 1995) since the same period (i.e., 1989–93) was used for both NVAP and the reanalyses. Even by accounting for the NVAP observational uncertainties, it still appears quite likely that the CCM1 has suffered a systematic, and occasionally rather severe, dry bias in the Northern Hemisphere during the JJA season.
To summarize, agreements among the CCM3, the reanalysis models, and the NVAP observations are significantly improved over that with CCM1, and, overall, the NCEP model shows closer agreement with NVAP than does the ECMWF, with more improvements seen in the upper troposphere. As pointed out earlier, models must be able to produce the right moisture characteristics with the right processes in order to possess prediction skills. Since direct observations (e.g., NVAP) are insufficient to allow this type of comparison, we must resort to the reanalyses to gauge CCM performances.
b. Our approach
We approach the problem of understanding the atmospheric moisture processes by representing the zonally averaged moistening rates due to large-scale motion with two distinct physical modes: the time-averaged mode (TAM) and the transient eddy mode (TEM). A simplified form of the zonal and time-averaged atmospheric moisture budget equation can then be written as
where q is specific humidity, υ and ω are, respectively, the meridional and vertical (pressure) velocities, with square brackets for zonal, and overbars for temporal averaging, and primes denote deviations from the time-averaged mode (i.e., the transient eddies). Note that the TAM can be further expressed by a zonally symmetrical component (the mean meridional circulation) and a zonally asymmetrical component (i.e., the standing eddies; Peixoto and Oort 1992). Starting from the right-hand side, the first term represents moisture changes due to time-averaged, large-scale circulations; the second term represents that due to transient eddies; and C and E denote moisture changes due to condensation and evaporation (or reevaporation of condensates), respectively. The next two terms are for convection (unresolved, thus parameterized) and vertical turbulent diffusion, with Kυ as the turbulent diffusivity. Parameterizations linking turbulent diffusivity to local shear and convective instabilities are used in the models, although implementation details may differ (to be discussed later). In statistical equilibrium, all the terms approximately balance out.
Following this procedure, the atmospheric moisture budget is controlled and balanced by (i) time-averaged circulations, (ii) transient eddies, (iii) convection, (iv) condensation and evaporation, and (v) surface processes and boundary layer diffusion. Each of these processes will be discussed in the next section, but because of the limited scope of this paper, only the mean states of the processes will be examined.
Moistening rate by a given process is a critical measure of the strength of that process. When multiplied by the latent heat of vaporization (L = 2.50 × 106 J kg−1) and divided by the specific heat at constant pressure (Cp = 1004 J K−1 kg−1), moistening rate can be expressed in equivalent heating units of K day−1. This is especially useful when energy balance is also of interest, in addition to moisture (see also Roads et al. 1998). We use this quantity to measure and compare the intensities of different processes in cycling moisture through the atmosphere.
3. Moisture processes and moistening rates
We have shown that NCEP has an overall moisture simulation somewhat superior to the ECMWF reanalysis, and we therefore select the NCEP results as “proxy” observations for process comparisons. The processes of interest include: time-averaged circulation, large-scale eddies, convection, condensation, evaporation, and surface processes and boundary layer diffusion. Moistening rates by these processes, based on simulations by CCM1, CCM3, and the NCEP model, are here compared. (Note: we use CCM3 rather than CAM2 for our analyses since CAM2 has since only recently been released, and few runs exist, in particular no 2 × CO2 runs are yet available. Since CAM2 is essentially an updated version of CCM3, this should have little bearing on our analyses.)
a. Time-averaged circulation
Moistening rates due to large-scale, time-averaged circulations are calculated according to Eq. (1). Contributions due to vertical and horizontal motions are computed separately and added together to obtain the total moistening rates, as shown in Fig. 2.
Net moistening associated with large-scale ascending and drying with large-scale descending motions of the time-averaged circulations are easily recognizable. In addition, they appear to be consistent with observational analysis (Peixoto and Oort 1992) that shows general moistening along the ITCZ and drying at subtropical latitudes, and the patterns move with season following the ITCZ. Contributions associated with large-scale convergence or divergence are also recognizable, notably inside the boundary layer. Comparison among the models indicates the CCM1 moistening or drying patterns are substantially narrower, and the amplitudes much larger, than shown by the CCM3 and the NCEP reanalysis. Differences in the moistening behaviors are more evident in the JJA months as the patterns shift northward and the effects of Northern Hemispheric summer monsoon and standing topographic waves become more important for the Northern Hemispheric moisture budget. CCM1’s moistening rates are generally stronger but with relatively narrow coverage, whereas the NCEP and CCM3 suggest more widespread moistening patterns.
The amplitudes and character of the differences between the CCM1 and the other two models are in line with our finding (results not shown) that the CCM1 generally has difficulty reproducing the correct intensity of the mean meridional circulation at the right latitudes. It is not clear at this point, however, how much of the differences can be attributed to its lower spatial resolution.
b. Transient eddies
Large-scale transient eddies represent eddy motions resolvable by model resolution. It is well known that eddies transport a significant amount of moisture through the atmosphere and play an important role balancing the global moisture budget (the same is true for heat and momentum as well, but we limit our discussion to moisture only here). Eddy moistening rates are calculated for horizontal and vertical components, respectively, and are added together to obtain the total eddy moistening rates.
Figure 3 shows the total eddy moistening rates from the two CCMs and the NCEP reanalysis. Some distinctive moistening patterns are readily seen, such as the general drying at lower latitudes and lower altitudes and the general moistening at high latitudes and high altitudes. These distinctive patterns reflect the distinctive roles played by the vertical and horizontal components of the eddies: the vertical component dries the boundary layer and moistens the regions above, while the horizontal component dries the low and moistens the high latitudes. The models appear to have captured the essence of this moistening mechanism when compared to limited observational evidence seen in Peixoto and Oort (1992). The lack of detailed observation forces us to rely on the quality of the reanalysis to fill in the gaps.
Analysis of the NCEP results indicates that eddy drying is essentially limited to the triangular region covering ±50° at the surface and 400 mb over the equator (see Fig. 3), with moistening elsewhere. This is qualitatively consistent with observational analysis. The CCM3 has largely captured this moistening behavior whereas the CCM1 has not. Discrepancies between the CCM1 and CCM3 are in fact rather striking: the CCM1 produces a large moistening rate outside the equatorial boundary layer where the CCM3 (and NCEP) suggests very little moistening, and the CCM1 boundary layer drying is also far too strong compared to the other models. Elsewhere in midlatitudes, all models produce moistening maxima that are consistent with observational analysis, although the CCM1 moistening maxima appear somewhat too close to the equator compared to others.
CCM3 generally shows much-improved agreement with the NCEP model than does the CCM1, in both moistening patterns and rates. Further analysis of this moistening mechanism indicates that the discrepancies are primarily caused by errors in the CCM1 vertical eddy component. The CCM1 vertical eddy moistening mechanism is likely too strong in the inner tropical belt of ±10°; even with partial cancellation from the horizontal component (which dries it), strong moistening is still seen in this region. These apparent CCM1 deficiencies are likely more problematic when coupled with deficiencies seen in its time-averaged circulations (last section) and may lead to even larger errors in tropical moisture simulation.
In summary, large-scale eddies transport great amounts of moisture to high latitudes and high altitudes. Model simulations of this mechanism produced mixed results, with the CCM3 showing reasonable skills of reproducing many of the “observed” features whereas the CCM1 was in general less satisfactory and, on occasion, misrepresented the moistening patterns and rates. The problem with CCM1 is particularly troublesome in the inner Tropics (±10°), where the errors tend to add up, instead of canceling out, with those identified earlier (i.e., errors associated with time-averaged circulation). This will seriously implicate tropical moisture simulation and model sensitivity.
Convection occurs on subgrid scales; it is therefore parameterized in climate models although the complexity of parameterization may vary considerably. Convection is often parameterized with dependence on large-scale variables, so it is not strictly independent from large-scale fields. In CCM1, convection is treated by moist convective adjustment of Manabe et al. (1965), in which heat and moisture are mixed and transported upward while proper conservation laws are followed. In a single adjustment event, moisture and heat are lost from a lower level and gained at higher levels (if condensation occurs during the event, instantaneous rainout is assumed). CCM3 introduces a more sophisticated “two-stage” moist convection, beginning with a penetrative, deep convection of Zhang et al. (1998) and followed by a series of shallow convections moving upward from the surface level (Hack 1994). Sequential execution of the two is believed to have contributed to the marked improvements in simulated climate (Kiehl et al. 1998; Hack et al. 1998; see also Hu et al. 2000). The NCEP model uses a simplified Arakawa–Schubert parameterization for convection (see Kalnay et al. 1996). Although it is generally believed that convection plays a crucial role in redistributing moisture in the free atmosphere, it is unclear to what extent it may affect moisture distribution globally. Because of the lack of direct observations, our understanding of this process will largely rely on modeling efforts.
Figure 4 shows the convective moistening rates from all three models. Models suggest convection results in predominant drying, especially inside the boundary layer, with very limited moistening tendency (such as seen in the subtropics, atop the boundary layer). The models, however, exhibit large differences in their behaviors: the NCEP model suggests that the strongest drying occurs in the lower–mid part of the boundary layer, where it can reach up to 10 K day−1 inside the Tropics—results largely supported by the CCM3—whereas the CCM1 produces a drying that is substantially weaker inside the boundary layer, in addition to a distinct, jet-like moistening core that reaches the equatorial upper troposphere. Moreover, the CCM1 shows a rather peculiar convective drying at higher vertical levels outside the Tropics while showing no evidence of moistening atop the subtropical boundary layer, as found in the other two models.
It is not clear which model is more realistic in representing the convective moistening rates as a result of the lack of observations, although it is widely believed that convective adjustment is probably too simplistic to represent these complex processes. What does become clear from the analysis here is, however, that convection can indeed alter vertical moisture distribution, and depending on the convective parameterizations used, the vertical moisture distributions can differ rather widely. The ability of altering vertical moisture distribution is expected to influence model climate sensitivity.
d. Condensation and evaporation
Condensation is the ultimate moisture sink of all processes. Model-simulated total condensational heating rates by the two CCMs are presented in Fig. 5, along with a proxy result from the NCEP reanalysis (identical field from the latter is not readily available). The NCEP proxy contains the total condensational heating but also contains an unwanted sensible heating contribution. However, numerical calculations suggest that condensational heating dominates under most conditions, making such a proxy comparison nonetheless worthwhile.
As expected, the general pattern of condensational heating follows closely the large-scale cellular motion, with the most active condensation coinciding with the ascending Hadley branch and the least active with the subsidence. Significant condensation is also seen associated with the midlatitude storm tracks. Generally speaking, the CCM1 depicts a narrowly defined, strong condensation core with weak seasonal shifts (same as seen in time-averaged circulation). In contrast, the CCM3 suggests a condensation pattern that is more widespread in space and weaker in intensity, while maintaining a more visible seasonal swing—features supported also by the NCEP model. The condensation patterns are largely consistent with the moistening mechanisms discussed in previous sections. Despite the notable differences between the two (e.g., the NCEP model suggests that the strongest condensation occurs in the tropical upper troposphere, while CCM3 points to lower levels), the CCM3 resembles the NCEP more closely in simulating the broad features of this mechanism. Further calculation of reevaporation of precipitation, which was made possible only by the CCM3 using the Sundqvist (1988) parameterization, indicates that this factor alone is rather insignificant throughout the atmosphere and thus unlikely to make a significant atmospheric moisture source. Finally, readers should be reminded of the speculative nature of the comparisons shown in this section due to the use of proxy fields and limitations of the models (not all of them were able to calculate the reevaporation).
e. Surface and boundary layer diffusion
Surface and boundary layer processes are intimately related. The fundamental atmospheric moisture source is undoubtedly from surface evaporation, a process with critical dependence on surface type, moisture availability, temperature, roughness, and wind. In both the CCMs and the reanalysis models, the atmospheric boundary layer is further divided into a surface layer, where exchange between the earth’s surface and atmosphere takes place, and the planetary boundary layer, where unresolvable turbulent eddies dominate the vertical moisture (as well as heat and momentum) transfer. Parameterizations using a bulk aerodynamic formula, in which moisture fluxes are assumed proportional to the differences between moisture values at the surface and the lowest atmospheric level, are used in all models.
Owing to the large vertical gradients of moisture concentration, only turbulent vertical diffusion is of importance to climate modeling and is often parameterized in a classic harmonic form, with diffusivity a dependent variable of local shear and convective instabilities so that stronger down-gradient transport and mixing can be realized during unstable conditions. The CCM1 has adopted this basic “local” (related only to local instabilities) approach (Williamson et al. 1987). The CCM3, in contrast, takes into account also the “nonlocal” transport due to dry convection and where diagnostic calculation of the boundary layer height becomes an integral part of the boundary layer parameterization (see Holtslag and Boville 1993). We shall only examine the diffusive moistening rate here for its role in redistributing moisture in the boundary layer (surface evaporation is important but is highly correlated with vertical diffusion, and therefore will not be separately discussed).
Model simulations of diffusive moistening rates are shown in Fig. 6. It is noted that this process is almost exclusively limited to the atmospheric boundary layer, which moistens. The resulting moistening rate can be as large as 10 K day−1 in the middle boundary layer. Direct comparison of moistening rates due to this and other processes discussed so far suggests that diffusive moistening dominates the boundary layer moisture sources. Although all the models show somewhat similar moistening habits in the boundary layer, quantitative differences are easily seen. The NCEP reanalysis shows that the moistening is highly concentrated inside the atmospheric boundary layer; it first increases with altitude and then decreases rapidly after peaking at around 950 mb—by the level of 850 mb, its intensity has dropped by about a factor of 10. The CCM3 result is similar in amplitude and structure, but it decreases less rapidly with altitude after peaking at 950 mb while it reaches somewhat higher altitudes. The CCM1, in contrast, does not peak in the intermediate levels; instead, its moistening rate decreases monotonically with altitude and reaches even higher. Given that CCM1 is the only model that has not used a boundary layer parameterization and that it has only two vertical levels below 850 mb (see Table 1), the above noted differences are not surprising. It is obvious now that in order to resolve the moistening structures in the boundary layer, there must be sufficient vertical levels within it. CCM1 does not have enough vertical resolution in this region; therefore, its moisture simulations are more likely to err when compared to the other models.
The role of vertical diffusion may also be seen through its influences on other processes, though more indirectly. For instance, by depositing and distributing moisture inside the boundary layer, diffusion provides the necessary moisture source needed for convection. And by the same token, it will modulate moistening rates by other processes, such as the time-averaged circulation and transient eddies. Given that vertical diffusion is the only process that directly links surface moisture source to the atmospheric sink above and indirectly influences other moisture processes through feedbacks, we argue that it is vertical diffusion that controls the overall intensity of the hydrological cycle; its response to an external forcing, then, will be indicative of how the atmospheric moisture field may respond to such a forcing. We will get back to this point in the next section when moisture simulation and climate sensitivity are discussed.
4. Moisture simulation and climate sensitivity
We have examined and compared major processes that control moisture distribution in the atmosphere. These processes jointly maintain the presently observed or simulated atmospheric moisture state and may respond, in their own ways, to any forcing imposed upon the system. The ability to simulate the observed moisture distribution, with substantial skills linking moisture distribution to moistening process, will undoubtedly enhance the prediction capability of a model (in contrast, a model without the right processes may still be able to produce many features of the present-day moisture field but will lack prediction skills). Besides playing a role in maintaining a new, and likely different, equilibrium state with imposed forcing, it is not clear what specific influence each individual process, with the exception of boundary layer diffusion, may have on model sensitivity. Boundary layer diffusion stands out as the only process that directly connects surface evaporation to the free atmosphere where other processes take place. In this unique role, it determines the overall intensity of the hydrological cycle by controlling the amount of moisture entering the atmosphere. We now examine how the amount of moisture entering the atmosphere through diffusion may respond to increasing CO2, and how it will relate to the moisture amount staying in the atmosphere in equilibrium. The latter has direct bearings on climate sensitivity.
The changes in diffusive moistening rates with doubled CO2, based on CCM1 and CCM3 climate change simulations, are shown in Fig. 7. Differences between the two are quite evident: the increments predicted by the CCM1 are approximately twice as large as predicted by the CCM3. This means that there is about twice as much additional moisture entering the CCM1 free atmosphere as entering the CCM3 with doubled CO2. We need to know how much of the increased moisture will remain in the atmosphere after a new equilibrium state is achieved.
Figure 8 shows changes of moisture amount in the three vertical layers when equilibrium states from before and after the doubling of CO2 are compared. The changes are presented in both absolute and relative amounts. The largest absolute changes are seen in the lower tropical levels as expected, but the largest relative changes are generally seen at higher latitudes and higher levels. The absolute changes in atmospheric moisture amount show strong correlations with the changes of diffusive moistening rates, for both seasons and for all the vertical levels. In all these cases, the amount of moisture increase is approximately twice as much in the CCM1 as in the CCM3 atmospheres.
The fractional changes of atmospheric moisture as a function of pressure are shown in Fig. 9 in which the averages are calculated for the Tropics and the entire globe (area weighted). The amount of moisture has increased in all vertical levels with higher fractional changes experienced by higher levels. Since water vapor is a strong greenhouse gas, larger changes in moisture amount will lead to stronger warming due to the water vapor greenhouse effect. Its influences are now evaluated.
We calculated the changes of moisture amount predicted by the CCM1 and CCM3, for the three vertical layers, and averaged them over the Tropics and the globe. The results are summarized in Table 2. The total surface warming predicted by the two CCMs is shown as “ΔTs.” The would-be surface warming, due to the change of moisture in each layer, denoted as “water vapor feedback (WVFB),” is calculated using a radiative–convective model (Hu 1996) and shown in the last columns. This factor alone is able to explain most of the differences between the CCM1 and CCM3 model sensitivity, with an explained fraction of 54% (i.e., 1.3 out of 2.4 K) for global and 74% (i.e., 1.4 out of 1.9 K) for tropical averages. It should be noted that the direct warming of doubling CO2 (i.e., the forcing) amounts to about 1 K of the total surface warming according to our calculation, with all the rest explained by various feedback mechanisms in which water vapor feedback is the most prominent one. For the three layers we have considered, both models suggest higher sensitivity from the top and bottom than from the intermediate layers. But overall, the vertical dependence is not particularly strong.
The results indicate that for all the cases considered, the warmings are proportional to the moisture increases, as would be expected from an energy balance point of view. And in all these cases, the predicated moisture increase is roughly twice as much in the CCM1 as in the CCM3, and so is the warming. We thus conclude that the differences in model sensitivity are mostly explained by the differences in predicted moisture, in which the modeling of surface and boundary layer processes is most likely to be responsible.
5. Discussion and conclusions
We have compared atmospheric moisture simulations by two NCAR CCMs with several other sources, including the NCAR CAM2 simulation, the NVAP observations, and the NCEP and ECMWF reanalyses. We have also compared the essential moistening processes by the two models with the NCEP reanalysis model. The CCMs represent different stages of climate model development, with the CCM1 for a generation of climate models developed prior to the late 1980s, designed to achieve tangible results while staying within existing computational capacity, whereas the CCM3 contains many model advances made up to the late 1990s. These changes have resulted in many improvements in simulated climate as reported elsewhere (e.g., Kiehl et al. 1998), but it is unclear how these improvements would translate into better moisture simulation, and perhaps more accurately, modeled climate sensitivity.
It is no surprise to see that significant differences exist among the different sources themselves, especially for moisture in the cold and dry upper troposphere. This reflects the level of uncertainty of the moisture field itself to the best of our knowledge. While we have seen differences between modeling results and the observations that appear rather systematic in character (such as seen in the Southern Hemispheric upper troposphere), suggesting that the quality of observations may be partially responsible, we have also seen larger differences involving the CCM1, in particular in the Northern Hemisphere summer season, that are notably beyond the uncertainties of the observations involved, indicating probable model deficiency. For all the cases of interest, with the exception of the JJA 500–300-mb layer (where a limited moist bias was found around the equator), the CCM3 shows clear signs of overall improvements in moisture simulation.
Beyond the seemingly reasonable agreements in reproducing many observed features of the moisture field, we were especially concerned about the roles of individual moistening mechanisms and their influences on the simulated moisture field. The NCEP was chosen for process comparisons for its overall quality of moisture simulation and for the lack of direct observations available. The time-averaged, large-scale circulation is found to moisten the inner tropical troposphere and to dry the subtropics, while the large-scale eddies generally transport moisture from low to high altitudes and from low to high latitudes. Convection and condensation both dry the atmosphere by removing moisture from it, whereas the reevaporation of stable precipitation only plays an insignificant role in the moisture budget. Boundary layer diffusion, which connects surface moisture source to the free atmospheric moisture sink above, is found to be in a very unique position controlling the amount of moisture entering the atmosphere and thus exerting, though indirectly, influences on other processes. Moreover, the amount of moisture increase in the equilibrium atmosphere is found proportional to the increase in diffusive moistening rate in the event of an external forcing. We conclude that regarding overall moisture simulation, it is boundary layer diffusion that controls the amount of moisture entering the atmosphere and regulates its responses to external forcing.
Both CCM1 and CCM3 predicted a more active hydrological cycle (as measured by moistening rates) and significant increases in atmospheric moisture with the doubling of CO2. The changes of diffusive moistening rate with doubled CO2 are in general twice as large in the CCM1 as in the CCM3, and so are the changes in equilibrium moisture amount. On global and annual averages, CCM1 predicted a 33% moisture increase, and CCM3 predicted 16%. Corresponding increases in the Tropics are at 29% and 14%, respectively. Calculations with a radiative–convective model indicate that the differences in predicted equilibrium moisture amount, along with vertical distribution, have contributed to the largest differences in model sensitivity as exhibited by the two CCMs. The explained differences are, respectively, at 54% (1.3 out of 2.4 K) for global and 74% (1.4 out of 1.9 K) for tropical averages.
The equilibrium states of the atmospheric moisture may have complicated relationships to the dynamical and thermodynamical processes involved, but regardless of how the equilibrium atmospheric moisture distribution is reached and maintained by these processes, the fact that more moisture enters the atmosphere through diffusion and remains there in equilibrium means higher climate sensitivity—since water vapor is a strong greenhouse gas and it causes surface warming when its amount increases. The overall dry biases seen in the CCM1 present-day simulation, being rather large and systematic, are also likely to boost the model sensitivity due to the enhanced potential of gaining moisture with increasing CO2.
With regard to boundary layer processes and diffusion, it is important to note that sufficient vertical resolution is necessary to fully resolve the boundary layer. As noted earlier, one of the most significant improvements of the CCM3 over CCM1 is due to the introduction of an explicit atmospheric boundary layer parameterization with nonlocal transport and iteratively determined boundary layer height (Holtslag and Boville 1993; Kiehl et al. 1996). The combined effect of vertical resolution and boundary layer parameterization is clearly seen in the model-calculated moistening rates and their responses to doubling CO2. The NCEP model has six vertical levels in the boundary layer and is seen resolving some of the sharpest moistening features in this region. The CCM3 has four levels in this region; it resolved, marginally, some of the structures inside the layer, but the vertical gradients of the moistening rates have been significantly “smoothed.” The CCM1, with only two levels available below 850 mb and without an explicit boundary layer parameterization, has simulated a much weaker vertical gradient with very little structure. Sufficient vertical resolution is therefore a prerequisite to any boundary layer parameterizations.
Agreements among the models are mostly qualitative, a reflection that the models have used rather diverse approaches in process modeling, especially regarding processes of unresolved scales where different parameterizations are used. For example, the two types of convective modeling used by the CCM1 and CCM3 (considering NCEP as close to CCM3) have led to very different moistening patterns and rates (the same as seen for boundary layer diffusion). Even for large-scale simulations—where the differences are least expected—there are still, however, noticeable discrepancies among the models. These are most evident in the simulated moistening rates for both the time-averaged circulations and transient eddies in that the CCM1 was shown to have mischaracterized the intensity as well as the locations. Therefore, it is possible for models of marginal spatial resolution (CCM1 likely being one of them) to err even on the very basic level, such as simulating the large-scale circulations. There is little question that process modeling determines model behavior as well as response to external forcing and therefore affects model sensitivity.
Among the modeling processes that may affect model sensitivity is convection. Convection, however, affects model sensitivity by mostly changing the vertical distribution of atmospheric moisture. By assuming that the effects of large-scale circulation on vertical moisture distribution can be averaged out with sufficient spatial averaging, the use of different convective parameterizations may be singled out to explain the differences in simulated moisture distributions, and subsequently the differences in model sensitivity. This is the approach taken by Hu et al. (2000) in studying the tropical climate and its responses to increasing CO2. Results from here, however, suggest that vertical diffusion (and related boundary layer parameterizations as well) plays a more essential role in lifting moisture into the atmosphere in the first place, providing the necessary “fuel” for other processes—including convection. In other words, these processes hardly work independently, they instead coexist dynamically in a system with multiple interactions and feedbacks. In all likelihood, convection is primarily responsible for changing the vertical distribution of moisture in the atmosphere, rather than controlling its amount—which is made through surface process and boundary layer diffusion. Our analysis clearly points to the connections between boundary layer modeling to atmospheric moisture, and to model climate sensitivity.
It is not the focus of this study to assess the predication skills of the models, but based on our findings that the CCM3 has, in most respects, shown more realistic representation of major moisture processes important to climate—in addition to its overall superior performances in simulating the presently observed state of atmospheric moisture distribution—there is reason to believe that the CCM3 may be in a more advantageous position over the CCM1 for climate predictions, and with the likelihood of a more realistic representation of climate sensitivity (which is different from model sensitivity). But there is no guarantee, as it is still unclear whether the response of such a model to a perturbation remains credible. What does become clear from this case study is, however, that the sensitivity of a model depends critically on its moisture simulation, in which modeling of the moisture processes, in particular surface and boundary layer processes, is likely to hold the key.
The authors thank the reviewers for helpful comments and suggestions, Professor J. G. Anderson for encouragement and support, and Atmospheric and Environmental Research, Inc., for support. This work was supported by NSF Grant ATM-9530914, NASA Grant NAG5-8779, and NCAR computation Grants 36211018 and 35851000. The reanalysis results were obtained from NCAR. This paper is dedicated to the memory of Professor Barry Saltzman for his lifelong support and encouragement until his untimely death in February 2001.
Corresponding author address: Dr. Haijun Hu, Atmospheric and Environmental Research, Inc., 131 Hartwell Avenue, Lexington, MA 02421. Email: firstname.lastname@example.org