Journal of Geophysical Research: Atmospheres

Comparison of future climate change over North America simulated by two regional models

Authors


Abstract

[1] Climate versions of the Pennsylvania State University/National Center for Atmospheric Research (NCAR) MM5 and of the International Centre for Theoretical Physics/NCAR RegCM2 are intercompared with two 10-year simulations over North America for present (1990–1999) and close to doubled CO2 (2090–2099), driven by the same general circulation model forcing from a 100-year transient run of the NCAR Climate System Model. Some model options and modules are identical in MM5 and RegCM2, including radiation and the entire land surface model. MM5 and RegCM2 agree in their general patterns of surface temperatures. Details related to mesoscale effects are well represented in both models. Precipitation is more consistent in MM5 and RegCM2 in winter than in summer. Differences in temperature and precipitation between MM5 and RegCM2 are small in winter due to the strong large-scale forcing but are large in summer because of the increasingly important contribution of local processes. Despite the general agreement between MM5 and RegCM2, significant differences exist over some areas due to the regional models' internal atmospheric physics and dynamics. The magnitude of these differences is large enough in some areas to cause serious uncertainties in any potential assessment of societal impacts of regional climate change.

1. Introduction

[2] Regional model studies have undergone rapid development in the past decade. Representations of surface processes and mesoscale physics have become more detailed, and regional models have been widely used to simulate climate and climate change scenarios in specific regions. With high-resolution and detailed descriptions of surface characteristics, regional models are expected to provide more accurate representations of climate and climate change. Applications of the Pennsylvania State University (PSU)/National Center for Atmospheric Research (NCAR) MM5 and of the International Centre for Theoretical Physics/NCAR RegCM2 models to climate study in North America have provided encouraging results [Giorgi et al., 1994; Takle et al., 1999; Dutton and Barron, 2000].

[3] However, considerable uncertainties still exist in regional model applications for a number of diverse reasons:

[4] (1) The extent to which the results of different regional models agree is unclear. In seeking to explain anomalous weather and climate phenomena, sometimes conflicting conclusions are drawn depending on the internal processes and physical feedback in different regional models. For example, in studies of the 1993 Midwest flood, Beljaars et al. [1996] found that dry soil reduced the amount of precipitation, but Paegle et al. [1996] suggested that wet soil could lead to a reduction of precipitation by weaker diurnal heating of the planetary boundary layer. Henderson-Sellers et al. [1996] suggested that small differences among land surface schemes could lead to large differences in climate. Feedback between surface and atmosphere may introduce large biases into the coupled surface-atmosphere system. The relatively long-term memory of the land surface makes this especially important in climate studies.

[5] (2) Regional models are strongly influenced by large-scale circulation from driving general circulation models (GCMs) [Christensen et al., 1997], and biases in GCMs may be amplified in regional circulation models (RCMs). Therefore the same regional model can have different results if driven by different GCMs.

[6] (3) Giorgi et al. [1996] and Seth and Giorgi [1998] found that the choice of model domain can influence model sensitivity to various internal and external forcings.

[7] (4) In their recent study of internal variability in RegCM2, Giorgi and Bi [2000] suggested that internal variability in RCMs can be significant, especially in summer and on daily to seasonal timescales. Its effect can be as large as those due to plausible modifications in the model physics. A similar investigation by Christensen et al. [2001] found that internal variability is a consequence of nonlinear physical feedback in the RCM. However, although these studies found large effects on individual periods and on a relatively short timescale, the effects on multiyear climatological monthly means were smaller than the typical magnitudes we present in sections 3 and 4 of this climatological study. Hence we suggest that internal variability is not important for our results.

[8] (5) Previous sensitivity studies of RCM climatology have found that plausible changes to physical schemes such as boundary layer mixing, radiation, and convection can have large impacts on climatologic results [e.g., Giorgi et al., 1993a, 1993b]. For example, precipitation is sensitive to the values of the critical bulk Richardson number in the boundary layer scheme [Hong and Pan, 1996]. It is possible that other combinations of parameter values or physical schemes could significantly affect our model difference results shown in sections 3 and 4; however, we consider it unlikely that they could be substantially reduced in this way while still retaining realistic present-day results.

[9] As a result of these and other uncertainties, projections of future climate change differ not only between GCMs and RCMs but also between RCMs themselves [Giorgi et al., 1994, 1998; Pan et al., 2001]. For example, the Commonwealth Scientific and Industrial Research Organisation (CSIRO) GCM simulated changes in precipitation under doubled CO2 over the central Great Plains were −1% in summer and 37% in spring, but changes predicted by RegCM2 were 6% in summer and 24% in spring when driven by the CSIRO GCM [Giorgi et al., 1998]. Discrepancies between GCM-RCM simulations are also found in East Asia [Hirakuchi and Giorgi, 1995]. In their earlier RegCM2 simulations under doubled CO2 over North America, Giorgi et al. [1994] found that precipitation increased 18.7% in the cold season and 24% in the warm season over the central Great Plains when driven by Global Environmental and Ecological Simulation of Interactive Systems (GENESIS). These values varied from 6% in summer to 24% in spring when RegCM2 was driven by the CSIRO GCM. The later simulation was for a smaller domain and a longer period. Leung et al. [1999] used the Pacific Northwest National Laboratory (PNNL) RCM to study climate sensitivity in the Pacific Northwest (PNW), and their results are quite different from those of Giorgi et al. [1994]. The PNNL RCM predicted less precipitation in winter in southern California and the southwestern U.S., but RegCM2 predicted more precipitation in this area. In the Pacific Northwest, RegCM2 predicted up to 2 mm day−1 precipitation increases over both ocean and land, whereas the PNNL RCM predicted precipitation decreases in coastal areas and over the adjacent ocean. However, some of these differences are probably due to different GCM driving fields: Community Climate Model, Version 3 (CCM3) for Leung and Ghan [1999] and GENESIS for Giorgi et al. [1994]. In a recent study of uncertainties in regional climate change simulations, Pan et al. [2001] examined precipitation climatology from two regional models, RegCM2 and HIRHAM, and found large differences between these RCMs. However, they only analyzed precipitation and focused on biases due to various factors such as model performance and boundary forcing. Their integration period for future change corresponds to 2040–2049, i.e., CO2 ∼ 480 ppm. Some important physical schemes such as cumulus parameterization and explicit precipitation are also significantly different in RegCM2 and HIRHAM. A similar comparison study has been undertaken by Takle et al. [1999] as part of the Project to Intercompare Regional Climate Simulations (PIRCS). Their results indicate that limited-area models can generally reproduce bulk temporal- and spatial-scale characteristics of meteorological fields. Episodes of mesoscale and convective precipitation are represented in a more stochastic sense, with less precise agreement between models in temporal and spatial patterns. Although their integration period covers only 2 months, which is too short for climate studies, their results still suggest some common tendencies of model results despite their differences in convective and surface parameterizations and in their methods of assimilating lateral boundary conditions.

[10] This study is a continuation of the work of Pan et al. [2001] and Takle et al. [1999]. Climate versions of MM5 and RegCM2 are applied over North America, both for the present climate and for hypothesized CO2 levels near the end of the next century, i.e., CO2 ∼ 660 ppm. To avoid inconsistency, both models are driven by the same NCAR Climate System Model (CSM) GCM and cover exactly the same model domain with the same horizontal and vertical resolution. The model domain extends into both the Pacific and Atlantic Oceans, which is large enough for regional models to freely develop their own synoptic-scale and mesoscale dynamics. This also helps to avoid possible errors caused by spatial interpolation over complex terrain areas in lateral boundary zones. The choice of physical schemes in each model is made carefully to maintain consistency in model physics. Ten-year integrations have been performed to provide sufficient information for the seasonal variation and spatial structure of the model climate.

[11] This paper is organized as follows. Brief descriptions of climate versions of MM5 and RegCM2 and of the experimental design are given in section 2. In section 3, present climate simulations by MM5 and RegCM2 are compared; the extent to which the two sets of regional models agree is examined, and reasons are suggested for their discrepancies. In section 4, common features of the future change scenarios predicted by both regional models are summarized; consistencies and discrepancies between MM5, RegCM2, and the driving GCM CSM are addressed, focusing on mesoscale characteristics. A concluding discussion is contained in section 5.

2. Model Description and Experiment Design

2.1. Model Description

2.1.1. Climate Version of MM5

[12] The climate version of MM5 used here is adapted from the standard version [Grell et al., 1994] for long-term climate simulation. MM5 is a nonhydrostatic primitive equation model with terrain-following σ coordinates, developed originally for medium-range forecasting. To apply to climate simulations, some important longer-term processes have been incorporated. Surface temperatures over water areas have been made time dependent, either by specification from the GCM sea surface temperature field (over ocean) or by the prediction of a slab mixed-layer model (over the Great Lakes). The simple ice scheme [Dudhia, 1989] is used for large-scale precipitation. The planetary boundary layer is parameterized by the medium-range forecast scheme [Hong and Pan, 1996]. The dynamic relaxation technique (Newtonian nudging and horizontal diffusion) [Davies and Turner, 1977] is applied in the MM5 buffer zone, with linear decay relaxation coefficients [Grell et al., 1994].

2.1.2. RegCM2

[13] RegCM2 [Giorgi et al., 1993a, 1993b] is the climate version of the NCAR/PSU mesoscale model MM4. It is a primitive equation, terrain-following, limited-area model with noncompressible hydrostatic balance. The explicit planetary boundary layer formulation [Holtslag et al., 1990] is used here. Large-scale precipitation is described by a simplified explicit cloud water scheme [Giorgi et al., 1993a, 1993b], based on the work of Hsie et al. [1984]. Surface temperature over water is set up the same way as in MM5. A similar dynamic relaxation technique is applied in the regional model buffer zone, with exponential decay relaxation coefficients [Giorgi et al., 1993b].

2.1.3. Common Features in MM5 and RegCM2

[14] Some important physical processes in shaping model climate are set identical in our versions of MM5 and RegCM2. To improve the representation of surface physics, the Land-Surface-Transfer scheme (LSX) is coupled into both model systems. LSX is a relatively detailed physical column model of soil, snow, and the vegetation canopy [Pollard and Thompson, 1995]. It computes exchanges of momentum, thermal energy, and water mass between the atmosphere and the land surface. The main LSX equations predict vegetation temperature and canopy air temperature given current RCM conditions above and current soil and snow conditions below. Intercepted liquid and snow and the cascade of precipitation through the canopy are included, and fluxes of radiation, sensible heat, and water vapor above the upper canopy are passed to the atmospheric model. Similarly, the net downward fluxes of heat and water at ground level are calculated and passed to the soil and snow models. Grell et al.'s [1994] cumulus parameterization scheme is applied in both models. Since most summer precipitation is cumulative, this is important for warm season precipitation simulation.

2.1.4. Differences Between MM5 and RegCM2

[15] The main differences between our versions of MM5 and RegCM2 are their descriptions of planetary boundary turbulence transport, explicit precipitation, and cloud cover, which will influence the radiative transfer calculation through cloud-radiative feedback. The same NCAR CCM2 radiation transfer package [Briegleb, 1992] is used in both models, but stratiform cloud amounts are parameterized differently. In MM5, cloud fraction is a function of large-scale relative humidity, while in RegCM2 it is equal to 0.75 when resolvable-scale clouds are calculated by the model's explicit moisture scheme. Radiation is also affected by cloud water content, which is predicted prognostically. Nevertheless, present-day cloud amounts are quite similar in both models and are reasonably realistic on large scales (not shown).

[16] Grell et al.'s [1994] cumulus parameterization scheme is used in both MM5 and RegCM2, but some parameter values, such as the convective timescale and the maximum depth of the stable layer between the updraft origination level and the level of free convection for a parcel to undergo deep convection, are set differently in the two models. The default values in RegCM2 are 1800 s and 150 mb, respectively, while in MM5 they are 900 s and 50 mb, respectively. Previous studies [Giorgi et al., 1993b; Kain and Fritsch, 1992] have shown that Grell et al.'s [1994] scheme is particularly sensitive to the maximum stable layer thickness, which determines whether deep convection occurs. A deeper stable layer allows convection to occur more easily, while a shorter convective timescale allows the model to approach instantaneous convective adjustment. Differences in these parameter values may lead to differences in model results; however, the complex interdependence between them makes it difficult to interpret their effect and to choose the best values for a specific model and experiment [Yang and Arritt, 2002]. For this study we chose to use the parameter values from standard versions of both models that have been shown to yield reasonably realistic present-day North American climatology [Giorgi and Mearns, 1999]. Moreover, when we used the RegCM2 values in MM5 in short present-day tests, summer precipitation became unrealistically strong and results were degraded compared to the standard MM5.

[17] At 50-km resolution the dynamical cores of MM5 and RegCM2 are essentially the same, despite the fact that RegCM2 and MM5 use hydrostatic and nonhydrostatic dynamics, respectively. Note, however, that the buffer zone for lateral boundaries is 250 km in MM5 and 400 km in RegCM2, as in their standard versions. The main differences between the two models are listed in Table 1.

Table 1. Main Differences Between the Physics in Our Climate Versions of RegCM2 and MM5
 RegCM2MM5
Explicit schemesimplified explicit cloud water scheme [Giorgi and Shields, 1999] based on explicit moisture scheme of Hsie et al. [1984]explicit moisture scheme with resolved-scale ice phase parameterization [Dudhia, 1989]
Planetary boundaryHoltslag et al.'s [1990] nonlocal formulation of vertical transportHong and Pan [1996] medium-range forecast nonlocal-k approach
Cloud formationGiorgi et al. [1993a]Grell et al. [1994]
Dynamicshydrostaticnonhydrostatic
Lateral boundarydynamic relaxation technique [Davies and Turner, 1977] that applies Newtonian nudging and horizontal diffusion within buffer zone; exponential decay relaxation coefficients [Giorgi et al., 1993a, 1993b] adopteddynamic relaxation technique [Davies and Turner, 1977] that applies Newtonian nudging and horizontal diffusion within buffer zone; linear decay relaxation coefficients [Grell et al., 1994] adopted

2.2. Experiment Design

[18] The model domain is centered at 99°W, 38.5°N, with a grid size of 50 km and 135 × 92 grid points (Figure 1). There are 15 vertical levels, with higher resolution in the boundary layer below 1500 m. The model top is at 100 mb.

Figure 1.

Model domain, topography (m), and six subregions (see text): PNW, Pacific Northwest; NC, north–central; NE, northeastern; BJC, Baja California; SE, southeastern U.S.; GPW, the Great Plains and adjacent western U.S.

[19] Ten-year simulations have been performed by both models for present and for close to doubled CO2. Both models are driven by the same large-scale forcing from an archived 100-year transient experiment of the NCAR CSM [Boville and Gent, 1998]. The CSM was run at T42 horizontal spectral resolution, with 18 vertical hybrid-σ levels. It was first spun up at present conditions for 15 years, after which CO2 was held constant at its present level for 10 years (years 15–24). Beginning with year 25, CO2 was increased at 1% per year, with doubling achieved in year 95. The output of the CSM in years 15–24 (termed “present-day”) and years 95–104 (termed “doubled CO2”) are used as driving fields for the integration of both regional models. Lateral boundary conditions are provided at 6-hourly intervals, and linear time interpolation is done at intervening time steps.

[20] To evaluate the model results, Climate Research Unit (CRU) global observations of decadal means during the period of 1981–1990 [New et al., 1999] are used. CRU data are at a resolution of 0.5° × 0.5° latitude and longitude, which is close to our regional model resolution (∼50 km). These are comparable to our present-day results only as climatological averages since the GCM forcing does not correspond to actual weather sequences. Some results are presented in section 3.5 in terms of subregions with distinctive climates, as shown in Figure 1. Our focus is on intermodel comparisons, though some evaluation of individual model aspects is included. Surface temperature and precipitation are emphasized in this study since they strongly influence socioeconomic consequences of future climate change. Synoptic circulation patterns are also analyzed to seek physical explanations for model climate and climate change.

3. Present Climate Simulation

3.1. Spatial Characteristics of Surface Temperature

[21] Monthly average temperatures for January (Figure 2) and July (Figure 3) from both regional models, from the CSM, and from CRU observations show that the general pattern of surface temperatures is reasonably simulated. Spatial patterns of surface temperature are quite similar for the two models and are consistent with CRU observations. Compared with CSM results, regional models provide finer-scale details and mesoscale effects. The observed low-temperature centers over high topography in the Rocky Mountains and the high-temperature centers in summer in the southern and southwestern U.S. are well represented. As observed, the Great Lakes region is warm in winter and cold in summer in both regional models, which is not the case in the CSM, for which the Great Lakes are not distinguished from land. The north-south temperature gradient in winter, and the warm tongue extending northward along the Great Plains in summer, are well simulated in both regional models. The largest difference between MM5 and RegCM2 is that MM5 predicts higher temperatures in winter over most of the model domain and in summer over the northern Great Plains (Figure 4). Conversely, temperatures for MM5 are lower than those for RegCM2 in summer in the southern U.S. and Mexico.

Figure 2.

Monthly average temperatures (°C) for January as determined by (a) MM5, (b) RegCM2, (c) the CSM, and (d) CRU observations [New et al., 1999].

Figure 3.

Monthly average temperatures (°C) for July as determined by (a) MM5, (b) RegCM2, (c) the CSM, and (d) CRU observations [New et al., 1999].

Figure 4.

Differences in monthly mean temperatures (°C) between MM5 and RegCM2 (under the present conditions) for (a) January and (b) July.

[22] Seasonal spatial correlation coefficients for surface temperature between MM5, RegCM2, and CRU observations are given in Table 2. The correlations are generally high, with the lowest values in summer for both models. One possible reason is that summer experiences more convective activity, which may influence surface temperature simulations via water and energy redistribution. Another possible reason is that with strong solar radiation in summer, mesoscale thermodynamic effects of local topography and surface characteristics may become stronger, which can modify local weather and climate details.

Table 2. Spatial Correlation Coefficients of Temperature Between MM5, RegCM2, and CRU Observations For 3-Month Averages Over the Whole Model (Land) Domain
 MM5 Versus CRURegCM2 Versus CRU
Spring0.960.95
Summer0.860.86
Autumn0.960.94
Winter0.980.97

3.2. Seasonal Cycle of Surface Temperature

[23] Monthly mean surface temperatures for MM5, RegCM2, and CRU observations areally averaged over six subregions are shown in Figure 5. Both regional models simulate reasonable annual temperature cycles over land. In the PNW and northern interior regions, annual cycles of temperatures for MM5 and RegCM2 are almost identical. Biases from CRU cycles are within 3°C in most months over all six subregions except for winter in Baja California, where both models underestimate temperatures by >3°C. Temperatures in summer for MM5 and RegCM2 agree closely to within 1°C. In the other three seasons, MM5 produces higher temperatures than RegCM2, with the largest differences of up to 5°C occurring in winter in the southeastern U.S. (SE). Generally, the MM5 simulation is more consistent with CRU observations than with RegCM2, especially in winter, when temperatures in all subregions are underestimated by both models but more so by RegCM2.

Figure 5.

Area average monthly mean surface temperatures (°C) over the six subregions shown in Figure 1 (only calculated over land): solid lines are for MM5, dotted lines are for RegCM2, and dash-dotted lines are for CRU observations.

3.3. Precipitation

[24] Monthly average precipitation for January is shown in Figure 6. Concentrations of precipitation on the Pacific Northwest and Atlantic seaboards are well captured in both models, but precipitation is overestimated in the southern Great Plains area, and a fictitious precipitation center occurs in the vicinity of Baja California. These errors may be due partly to biases in the large-scale forcing. The CSM 500-mb synoptic pattern (not shown) shows an unrealistic deep trough and an intensified cyclonic circulation over western Mexico, which causes overestimation of convergence and precipitation there. This error in the CSM circulation is consistent with its unrealistic precipitation maximum in western Mexico and Baja California. The same error is also introduced into both MM5 and RegCM2, resulting in overestimation of precipitation in the southern Great Plains. MM5 and RegCM2 produce similar spatial patterns of precipitation, and their results are closer to those of the CSM than to the CRU observations.

Figure 6.

Monthly average precipitation (mm day−1) for January as determined by (a) MM5, (b) RegCM2, and (c) the CSM and (d) as observed [New et al., 1999].

[25] Precipitation for July (Figure 7) shows large differences between regional models as well as between regional and global models. A common feature in MM5, RegCM2, and the CSM is the concentration of precipitation in the northeastern part of the domain. In most interior areas, precipitation is greater in summer than in winter. The CSM overestimates precipitation along the Rocky Mountains, but this overestimation doesn't occur in RegcM2 and MM5. Generally, RegCM2 produces more summer precipitation than MM5 over most land areas. Note that summer precipitation is seriously overestimated over the Gulf coast and Mexico in MM5, which might be due to unrealistically strong moisture transport and vertical motion forced by topography. The subtropical high simulated by MM5 in July is much stronger than in RegCM2 (Figures 9d and 9e), causing large moisture transport from the subtropical ocean to the Gulf coast and Mexico. The warm moist air is forced to rise along the upwind side of the Mexican central mountain range (Figure 1) and generates heavy precipitation there. The released latent heat helps intensify the vertical motion and produces more precipitation. With enough supply of warm moist air from the subtropical ocean, the positive feedback of topographically forced vertical motion to precipitation to latent heating to stronger vertical motion and more precipitation explains why precipitation is seriously overestimated.

Figure 7.

Monthly average precipitation (mm day−1) for July as determined by (a) MM5, (b) RegCM2, and (c) the CSM and (d) as observed [New et al., 1999].

[26] Differences between precipitation simulated by MM5 and RegCM2 are shown in Figure 8. The differences are small in winter but large in summer. In winter the largest difference between the regional models occurs in the southern U.S., where MM5 has higher precipitation than RegCM2. Over most land areas, differences in winter precipitation between MM5 and RegCM2 are within 0.6 mm day−1. In summer, however, in the northwestern interior, MM5 precipitation is up to 1 mm day−1 less than RegCM2 and is up to 2.5 mm day−1 less than observed. The largest difference occurs along the west Gulf coast, where MM5 produces more precipitation than RegCM2, but in parts of the southeastern U.S., MM5 gets much less precipitation. These results suggest that the influence of large-scale forcing is strong in winter but weak in summer, when local processes contribute more to model results. Differences in the physics and thermodynamics of individual models may have a strong influence on model results during summer.

Figure 8.

Differences in monthly mean precipitation (mm day−1) between MM5 and RegCM2 (under the present conditions) for (a) January (b) July.

[27] Since both regional models are driven by the same large-scale forcing, their spatial patterns of precipitation fields tend to be similar in most regions, especially in winter. However, precipitation differs significantly in some regions and especially in summer. It is clear that MM5 and RegCM2 are under a stronger influence of synoptic-scale forcing in winter than in summer, and in summer the regional models are relatively independent of lateral forcing. In a study of internal variability in RegCM2, Giorgi and Bi [2000] have found similar results. Since the two regional models have different dynamical structures and their explicit precipitation and boundary transfer schemes are also different, it is difficult to identify the main reasons for the differences in precipitation simulation in the two regional models. However, some insight can be gained from the analysis of their synoptic circulation patterns, presented in section 3.4.

3.4. Synoptic Circulation and Moisture Transport

[28] Monthly average circulation patterns at 500 hPa for MM5, RegCM2, and the NCAR-NCEP Reanalysis Project (NNRP) of 1958–1981 [Kalnay et al., 1996] are shown in Figure 9. In January the polar front in both models is positioned over the southeastern U.S., corresponding to the position of the largest temperature gradient. High 500-mb wind speeds extend in a SW-NE direction in this area. Concentrations of winter depressions and related precipitation (shown in Figures 6a and 6b) coincide with the position of this jet stream aloft. Since winter temperatures over the southeastern U.S. are higher for MM5 than for RegCM2 (Figures 2a, 2b, and 5), the temperature gradient between ocean and land is larger in MM5, which causes stronger meridional circulation in MM5 than in RegCM2. The stream axis in MM5 tends to be oriented more SSW-NNE, with stronger cyclonic disturbances (Figures 9a and 9b). Corresponding to these circulation patterns, MM5 has more precipitation in the southeastern U.S. than does RegCM2.

Figure 9.

Geopotential height (m) at 500 mb for January as determined by (a) MM5, (b) RegCM2, and (c) the NNRP and for July as determined by (d) MM5, (e) RegCM2, and (f) the NNRP.

[29] In summer a midcontinental pressure ridge and related dry air control large inland areas west of 95°W in both models (Figures 9d and 9e). The high-pressure ridge for MM5 is stronger and extends farther northward than it does for RegCM2; thus underestimations of precipitation are more severe for MM5. The summer trough above the central-eastern U.S. is shifted north of its true position for both models and to a greater extent in MM5. The midatmospheric trough provides favorable conditions for precipitation in the northeastern domain in both models. In the southeastern U.S. and Gulf coast, however, circulation patterns for MM5 and RegCM2 differ. Figures 9d and 9e show that the midtroposphere trough along the Gulf coast is not well developed in both models, but the subtropical high is overly developed in MM5 and extends too far northward. Warm moist air from the subtropical ocean flows along the southern edge of high pressure, reaches the western Gulf coast, is forced to rise by topography, and results in heavy precipitation in MM5. The subtropical high is weaker in RegCM2 than it is in MM5. Thus summertime precipitation in MM5 in the southeastern U.S. is less than it is in RegCM2. The above discussion about differences in synoptic patterns partly explains their differences in precipitation, although different physical schemes (convection and explicit precipitation, cloud, boundary transfer) in these two models may also contribute.

3.5. Seasonal Cycle of Surface Water Budget in MM5 and RegCM2

[30] Monthly mean precipitation, evapotranspiration, and surface water convergence are shown in Figure 10 for the selected subregions. The seasonal variation of precipitation for RegCM2 is stronger than for MM5, mainly because RegCM2 precipitation in warm months (May to October) is higher but precipitation in cold months is nearly the same as in MM5 (except for SE and the Great Plains and adjacent U.S. regions, where RegCM2 precipitation is less). Evaporation is controlled by surface energy and water availability, and the largest evaporation occurs in summer for both models and over all subregions. Similarly to precipitation, summer evapotranspiration for MM5 is less than for RegCM2. A common feature in both models is that precipitation has almost the same magnitude as evapotranspiration during summer, which suggests that most precipitation is reevaporated into the atmosphere. Thus, despite the difference of summer precipitation for MM5 and RegCM2, the total soil moisture convergence is similar for both models. They share the same characteristics in that water content is recharged in cold months but is lost in warm months, which is a typical phenomena in midlatitude continental inland areas.

Figure 10.

Area average monthly mean precipitation, evaporation, and soil moisture convergence (precipitation minus evaporation and runoff) over the six subregions shown in Figure 1. All units are in mm day−1. Solid lines are for MM5, and dashed lines are for RegCM2.

[31] These results indicate that regional models are capable of self-adjustment in their surface water budget. Precipitation and evapotranspiration are positively correlated, though surface energy availability is also important in determining evaporation. Seasonal soil moisture variation does not exactly follow moisture convergence but lags it by ∼2–3 months (not shown). With its higher temperature in winter (Figure 4) and less precipitation in summer (Figure 8), MM5 has a warmer and wetter winter and drier summer than does RegCM2.

4. Climate Change Scenario Under Doubled CO2 Conditions

4.1. Temperature Change

[32] Monthly average temperature changes (doubled CO2 minus present) are shown for January and July in Figure 11. The large-scale changes in both models are similar in that higher-latitude areas in North America experience stronger warming than lower-latitude areas in winter; lower-latitude regions experience stronger warming in summer than in winter; and summertime warming of 1.5°–4°C is relatively homogeneous over almost the whole model domain. However, there exist some differences between the two model results. In winter, RegCM2 produces stronger warming in higher latitudes. Cooling of 0.5°–2°C occurs in MM5 over large areas of lowland Mississippi and the Ohio River valley, but in RegCM2 the same area experiences only slight changes of <1°C. A similar cooling is also found in the CSM GCM, though less severe than in MM5. Student's t-tests show that this cooling in both the CSM and in the regional models is not significant at the 95% confidence level, indicating that it could well be due to interannual variability over these particular decades. In summer the largest difference between the regional models occurs in a small area between the Great Lakes and the northeastern coast, where MM5 produces <1.0°C warming but RegCM2 produces up to 3°C. MM5 also has some scattered small areas with <1.0°C temperature warming, while RegCM2 has temperature increases of 1.5°C or more almost everywhere.

Figure 11.

Monthly average temperature change (°C) under doubled CO2 for January as determined by (a) MM5, (b) RegCM2, and (c) the CSM and for July as determined by (d) MM5, (e) RegCM2 , and (f) the CSM.

[33] Comparing the temperature changes with those in the CSM GCM, the spatial patterns in MM5 are generally more consistent with the CSM, except that the surface cooling in winter over the southeastern U.S. extends farther southward and is greater in MM5. Summertime warming is larger in both regional models than in the CSM over most areas. Figure 12 shows the changes in cloud cover due to doubled CO2 simulated by MM5 and RegCM2. Most of the large-scale change patterns are similar in both models for both winter and summer. In summer, there is a clear correlation between cloud changes and temperature changes (Figures 12c and 12d versus Figures 11d and 11e), with larger warming coinciding with greater cloud reductions and vice versa. However, cloud radiative fields were not saved, so it is hard to tell if clouds are the primary forcing. In winter, no such correlation is apparent between cloud and temperature changes (Figures 12a and 12b versus Figures 11a and 11b).

Figure 12.

Monthly mean cloud cover change under doubled CO2 (%) for January as determined by (a) MM5 and (b) RegCM2 and for July as determined by (c) MM5 and (d) RegCM2.

[34] Differences in temperature under doubled CO2 between MM5 and RegCM2 are shown in Figure 13. The general patterns are quite similar to those of the present conditions (Figure 4). In the eastern U.S., temperature is higher in winter but lower in summer for MM5. Big differences also occur in summer in the northern Great Plains, where temperatures for MM5 are ∼1°–4°C higher, and in northern Mexico, where temperatures are almost 4°C lower for MM5.

Figure 13.

Differences in monthly mean temperature (°C) between MM5 and RegCM2.

4.2. Precipitation Change

[35] In winter, precipitation changes due to doubled CO2 are similar for MM5 and RegCM2 (Figures 14a and 14b). In both models the southeastern U.S. and the Pacific coast from Oregon to California experience large increases in precipitation. A zonal belt between ∼27°–37°N and areas around Hudson Bay in North America also experience more precipitation. Between this belt and the Hudson Bay, and in Baja California, precipitation tends to decrease. These precipitation changes are similar to changes in the CSM. The precipitation increase in MM5 is stronger in the southeastern U.S. but weaker on the Pacific coast than in RegCM2. In summer, precipitation increases in both models in the eastern U.S. and in most areas of the northern domain (Figure 14d and 14e). These areas with precipitation increases are not positioned exactly the same for both models, but their general positions are similar. Precipitation decreases mainly occur in parts of the southwestern U.S. and southeastern Canada, but these decreases are less severe for RegCM2 than for MM5. Compared with winter, these changes in summer are less dependent on CSM results, indicating a stronger influence of regional internal dynamics and physics.

Figure 14.

Monthly average precipitation change (mm day−1) under doubled CO2 for January as determined by (a) MM5, (b) RegCM2, and (c) the CSM and for July as determined by (d) MM5, (e) RegCM2, and (f) the CSM.

[36] Compared with changes in the CSM, both regional models predict small-scale centers with more violent changes. Spatial variations in precipitation changes and the magnitude of these changes at smaller scale are much larger in the regional models than in the CSM. More importantly, in areas downwind of the Great Lakes, precipitation increases in both MM5 and RegCM2 but decreases in the CSM. These features in the regional models are due to their higher resolution and are not resolved by the CSM GCM. The area in the southwestern U.S. with less summer precipitation in the CSM is enlarged in MM5 but decreased in RegCM2.

[37] Differences in precipitation under doubled CO2 (Figure 15) for MM5 and RegCM2 are similar to those under the present conditions (Figure 8). Small differences of <0.6 mm day−1 occur in winter, but up to 4 mm day−1 more precipitation occurs along the western Gulf coast and up to 1.5 mm day−1 less precipitation occurs in the southeastern U.S in summer for MM5 as compared with RegCM2. The precipitation changes discussed above are closely related to changes in synoptic scale circulation, which is discussed in section 4.3.

Figure 15.

Differences in monthly mean precipitation (mm day−1) between MM5 and RegCM2 (under doubled CO2 conditions) for (a) January and (b) July.

4.3. Circulation Change

[38] Figure 16 shows the circulation change at 500 mb for MM5, RegCM2, and the CSM. The synoptic-scale circulation in regional models is strongly influenced by large-scale forcing. This influence is most pronounced in winter (Figures 16a, 16b, and 16c). Changes in circulation patterns induced by doubled CO2 for both MM5 and RegCM2 are quite similar to those for the CSM. In winter the polar front is positioned over the southeastern United States, and the jet stream above is oriented SW-NE. Winter depressions and related rainfall coincide with the position of the jet stream aloft. Under doubled CO2 the axis of this jet stream becomes more SSW-NNE and cyclonic circulation is intensified, indicating stronger and more frequent disturbances. This explains why both models have more precipitation in the southeastern U.S. in winter. Although the overall circulation is similar between MM5 and RegCM2, there are some significant differences related to their different temperature responses to doubled CO2. Surface cooling in the wintertime southeastern U.S. is stronger and covers a larger area in MM5 than in RegCM2 (Figures 11a and 11b), and the ocean-land temperature gradient becomes larger in MM5. The meridional circulation in the southeastern U.S. is thus more intense for MM5, and consequently, its cyclonic circulation extends farther north than does that of RegCM2. In the northern part of the domain, however, because of the strong warming in the north and the weak warming or even cooling in the south under doubled CO2, the midlatitude jet stream is weakened in both models, which causes less storm activity and less precipitation northward of 40°N. Westerly winds strengthen on the Pacific coast of Oregon, Washington, and northern California under doubled CO2, causing increases in precipitation over these limited regions. MM5 and RegCM2 show similar patterns of circulation change in winter. These changes in circulation pattern are closely related to temperature and precipitation changes and are under the strong influence of the large-scale CSM forcing.

Figure 16.

Monthly average changes in the 500-hPa wind field (m s−1) under doubled CO2 for January as determined by (a) MM5, (b) RegCM2, and (c) the CSM and for July as determined by (d) MM5, (e) RegCM2, and (f) the CSM.

[39] Circulation pattern changes in summer for MM5 and RegCM2 are more different than in winter (Figures 16d, 16e, and 16f). MM5 seems more dependent on large-scale forcing since its circulation is more similar to the CSM, while RegCM2 develops its own pattern. For MM5 the high-pressure ridge over the inland continent and the cyclonic circulation above the Atlantic coast in summer are both intensified. These changes are favorable for more precipitation over the eastern U.S. and less precipitation over the southwestern U.S. Compared to the present conditions, southerly winds are intensified in the northwestern areas of the model domain for MM5 and RegCM2, indicating more moisture transport there. These circulation pattern changes in summer explain the precipitation changes discussed in section 4.2 to some extent, although some other physical processes may contribute to the changes.

5. Discussion and Conclusions

[40] In this study, climate versions of MM5 and RegCM2 are used for long-term simulations of present and close to doubled CO2 climates over North America, driven by an archived 100-year transient experiment of the NCAR CSM. Model results are compared with each other as well as with the large-scale fields of the NCAR CSM. The extent to which regional models are in agreement and the reasons for their discrepancies are discussed.

[41] Even though the same large-scale forcing and the same land surface model (LSX) are used in MM5 and RegCM2, significant differences are found in their monthly mean results. Our principal result is that even with the same large-scale forcing, various internal physical and dynamical processes in different regional models can lead to large uncertainties in model results. MM5 and RegCM2 generally agree better in winter than in summer. As found by Dutton and Barron [2000] in their ensemble study of intra-annual and interannual regional climate variability, this is due mainly to the stronger constraint of synoptic-scale forcing in winter. They concluded that in winter, interannual variability of RegCM2 is strongly controlled by the interannual variability of synoptic-scale forcing, whereas in summer the regional model variability is less constrained by the lateral boundary and stems from internal fluctuation. This is consistent with our results: In summer, regional-scale differences between MM5 and RegCM2 are much larger than in winter due to relatively weak synoptic-scale forcing and the increased influence of interior dynamics and local processes. In their study of internal variability in RegCM2, Giorgi and Bi [2000] conclude that the model response to small perturbations is much stronger in summer than in winter, which also helps to explain why our differences between MM5 and RegCM2 are larger in summer than in winter. The amplitudes of these regional differences in temperature and precipitation can be up to ∼5°C and 4 mm day−1, respectively.

[42] Therefore, in order to improve confidence in regional climate predictions, both synoptic-scale forcing and internal processes in regional models need to be considered carefully. Model differences can be amplified by the complex feedback between various physical and thermodynamic processes in the regional model itself and between the regional model and synoptic-scale forcing. In an earlier regional model intercomparison study, Pan et al. [2001] analyzed the uncertainties in precipitation from RegCM2 and HIRHAM [Pan et al., 2001]. Driven by three sets of synoptic forcing, RegCM2 and HIRHAM were integrated for 10 successive years with CO2 levels of ∼330 ppm and ∼480 ppm, respectively. They found that intermodel bias is largest in summer, probably due to different subgrid scale processes in each model. They also found significant biases in both models' precipitation patterns as compared with observed. Our study extends their intermodel comparison to more climate variables, including temperature, precipitation, and mesoscale and synoptic-scale circulation simulated by regional models. Furthermore, in the present study, some important physical component schemes such as the land-surface model and the radiativeon transfer scheme are identical in MM5 and RegCM2. It should be noted that although both models use Grell et al.'s [1994] cumulus parameterization scheme, some parameter values are set differently, as discussed in section 2. Thus the differences between these two models can be attributed to model dynamics and physical schemes (explicit moisture scheme, boundary scheme, etc.) and to the feedback between the various processes.

[43] Previous sensitivity studies of physical schemes in RCMs have shown that acceptable present-day climatologies can be achieved in different ways, possibly for the wrong reasons, with canceling errors. For instance, although boundary layer models in RegCM2 [Holtslag et al., 1990] and MM5 [Hong and Pan, 1996] both involve K-theory and countergradient transport terms, Giorgi et al. [1993a, 1993b] showed that the boundary scheme in RegCM2 in itself induces an overall increase in total precipitation. However, this is offset by a decrease in convective precipitation due to drier low levels caused by rapid upward moisture transport. Further sensitivity experiments conducted by Hong and Pan [1996] indicated that parameters such as critical bulk Richardson number, countergradient terms, and the diffusivity coefficient profile in the MM5 boundary scheme all have significant and sometimes canceling impacts on modeled precipitation. Although tuning these parameters can achieve satisfactory results for specific case studies or regions, it is harder to do so over continental areas and for long-term climate. Therefore it is perhaps not surprising that significant discrepancies can occur between different RCMs in climate change experiments, even though present-day control results may be similar.

[44] One major objective for the national assessment of potential consequences of future climate change is to assess regional vulnerabilities and impacts of climate change. As a tool capable of enhancing GCM information in specific regions through the description of mesoscale processes, regional models are increasingly being applied to future climate change studies. At this stage it is important to understand how reliable the future predictions of RCMs are. We have shown that significant differences exist between two RCMs due only to their internal atmospheric physics and dynamics. The magnitude of these differences is large enough in some regions to cause significant uncertainties in any potential assessment of societal impacts. Further studies will focus on determining the contribution to uncertainties made by individual factors, such as dynamics and various physical schemes, in order to improve our confidence in regional model simulations.

Ancillary