The fifth-generation PSU–NCAR Mesoscale Model (MM5)-based regional climate model (CMM5) capability in simulating the U.S. precipitation annual cycle is evaluated with a 1982–2002 continuous baseline integration driven by the NCEP–DOE second Atmospheric Model Intercomparison Project (AMIP II) reanalysis. The causes for major model biases (differences from observations) are studied through supplementary seasonal sensitivity experiments with various driving lateral boundary conditions (LBCs) and physics representations. It is demonstrated that the CMM5 has a pronounced rainfall downscaling skill, producing more realistic regional details and overall smaller biases than the driving global reanalysis. The precipitation simulation is most skillful in the Northwest, where orographic forcing dominates throughout the year; in the Midwest, where mesoscale convective complexes prevail in summer; and in the central Great Plains, where nocturnal low-level jet and rainfall peaks occur in summer. The actual model skill, however, is masked by existing large LBC uncertainties over data-poor areas, especially over oceans. For example, winter dry biases in the Gulf States likely result from LBC errors in the south and east buffer zones. On the other hand, several important regional biases are identified with model physics deficiencies. In particular, summer dry biases in the North American monsoon region and along the east coast of the United States can be largely rectified by replacing the Grell with the Kain–Fritsch cumulus scheme. The latter scheme, however, yields excessive rainfall in the Atlantic Ocean but large deficits over the Midwest. The fall dry biases over the lower Mississippi River basin, common to all existing global and regional models, remain unexplained and the search for their responsible physical mechanisms will be challenging. In addition, the representation of cloud–radiation interaction is essential in determining the precipitation distribution and regional water recycling, for which the new scheme implemented in the CMM5 yields significant improvement.
The development of a nested regional climate model (RCM) was first proposed by Dickinson et al. (1989) and Giorgi (1990). Following pioneering work by Giorgi and Bates (1989), long-term continuous RCM simulations have been increasingly available for a wide range of applications, including model validation, sensitivity studies, and climate change assessments (see a recent review by Giorgi and Mearns 1999). It is now commonly accepted that an RCM downscaling integration is more skillful at resolving orographic climate effects than the driving coarse-grid general circulation model (GCM) simulation, especially for near-surface variables (Giorgi 1990; Jones et al. 1995; Giorgi et al. 1997, 1998; Christensen et al. 1998; Laprise et al. 1998; Leung and Ghan 1999; Hong and Leetmaa 1999; Fennessy and Shukla 2000; Pan et al. 2001; Roads et al. 2003). However, there exist systematic RCM biases that cannot be fully explained and large downscaling skill uncertainties over regions with relatively flat terrain (Takle et al. 1999; Leung et al. 1999; Pan et al. 2001; Anderson et al. 2003). In addition, the RCM results are sensitive to its dynamic configurations (e.g., domain and resolution) and physics representations (e.g., cloud–radiation and surface–atmosphere interactions). Therefore a typical RCM must be “customized” for each specific regional application (Giorgi and Mearns 1999). This includes careful sensitivity experiments and rigorous validation studies using observed initial and lateral boundary conditions (LBCs).
For the United States, several studies have provided important references for RCM customization. Giorgi et al. (1993b) first developed an improved LBC assimilation scheme to facilitate more realistic RCM climate simulations over large domains. Seth and Giorgi (1998) then found that small domains produce spurious LBC dynamical effects and thus cause the RCM to generate unrealistic responses to internal forcings. Liang et al. (2001) further demonstrated that the RCM performance depends significantly on the accuracy of the driving LBCs within the buffer zones. In contrast, Juang and Hong (2001) showed that the impacts of the domain size and LBC errors are small for a regional spectral model where the large-scale waves are preserved through the domain and spectral nesting. The difference in sensitivity arises because the large-scale waves in the former are generated by the dynamic relaxation of LBCs only within four narrow edge buffer zones, whereas for the latter they are prescribed by the continuous nudging toward the driving general circulations throughout the entire computational domain. If the driving circulations are perfect everywhere, both approaches shall produce a similar result. However, the existing global reanalyses, which are the best available proxies for observations, contain large inconsistencies, especially over oceans and near the Tropics where observational data are lacking and the RCM buffer zones are generally located (Liang et al. 2001). It is impossible for RCMs to produce a realistic regional climate via the dynamic relaxation of erroneous LBCs based on the reanalyses. In contrast, the nudging over the entire domain minimizes the LBC effect and enables the RCM to force a more accurate regional climate with the realistic regional circulation of the reanalyses that assimilate the most complete in situ observational data. On the other hand, if the RCM is driven by a GCM simulation that likely contains regional circulation biases over the inner domain due to the incomplete physics representation, the LBC relaxation approach may yield a more realistic regional climate because it gives more freedom for the more-complete RCM physics representation to be effective.
The preceding problems make it difficult in validation studies to separate the RCM downscaling skill and its sensitivity to the physics representation from the uncertainties in the driving reanalyses. Giorgi and Shields (1999) showed that the cumulus parameterization of Grell (1993) produces an overall more realistic regional climate over the continental United States than those of Kuo (Anthes 1977) and Zhang and McFarlane (1995). In contrast, Gochis et al. (2002) found that the Kain and Fritsch (1993) scheme produces superior vertical thermodynamic structures and hence more realistic convective precipitation associated with the North American monsoon (NAM) than the Betts–Miller–Janić (Betts and Miller 1986; Janić 1994) and Grell schemes. Leung et al. (2003) also indicated the superiority of the Kain– Fritsch to Kuo and Grell schemes for the NAM simulation. On the other hand, Xu and Small (2002) demonstrated that the Kain–Fritsch scheme yields too much NAM rainfall and fails to represent intraseasonal and interannual variations, while the Grell scheme is generally more realistic. It is essential to objectively determine whether the poor RCM performance is caused truly by the unrealistic model physics representation or purely by the LBC errors contained in the reanalyses. The findings are expected to strongly depend on climate regimes and locations.
The purpose of this study is to determine the capability of an RCM to reproduce the observed annual cycle of U.S. precipitation and to better understand the causes of the corresponding model climatology biases. This is facilitated by a continuous RCM baseline integration for the period 1982–2002 as driven by a global reanalysis and a suite of seasonal sensitivity experiments with various driving reanalyses and physics representations. As recommended by Giorgi and Mearns (1999), long-term continuous RCM integrations offer many advantages over ensembles of short simulations, including minimal effect from atmospheric spinup, improved equilibrium between the regional climate and surface hydrology cycle, accurate representation of the model internal climatology and better detection of systematic model physics deficiencies. Thus the 1982– 2002 mean climatology provides a robust statistical description to identify significant systematic model errors in the annual cycle. The sensitivity experiments are carefully designed and branched off from the restart conditions of the baseline integration to provide physical insights into the probable causes for specific RCM biases.
2. Regional climate model and simulations
The RCM used in this study is a climate extension of the fifth-generation Pennsylvania State University– National Center for Atmospheric Prediction (PSU– NCAR) Mesoscale Model (MM5) version 3.3 (Dudhia et al. 2004), hereafter referred to CMM5. The CMM5 is an improved version of Liang et al. (2001). Important modifications include incorporation of more realistic surface boundary conditions and cloud cover prediction as well as LBCs from an updated global reanalysis. A brief description of the CMM5 model configuration and climate simulations is given next.
a. Physics configuration
The CMM5 incorporates the same physics configuration as in Liang et al. (2001) except for an improved cloud–radiation interaction prediction. The land surface process is represented by the Oregon State University (OSU) model (Chen and Dudhia 2001). The planetary boundary layer is parameterized by the Medium-Range Forecast Model (MRF) countergradient (nonlocal) turbulence transport scheme (Hong and Pan 1996). Precipitation is determined by a combination of the Goddard Space Flight Center (GSFC) explicit cloud microphysical solution (Tao and Simpson 1989), the Grell (1993) cumulus parameterization, and shallow convection. Solar and infrared radiation are incorporated as in the NCAR Community Climate Model (CCM2; Hack et al. 1993) and are calculated every 30 min, where the radiative effects of both cumulus and nonconvective clouds are considered. Liang et al. (2001) showed that this physics configuration simulates the overall structure and magnitude of the observed rainfall pattern over the Midwest and northern Plains during the 1993 summer flood. This is associated with the accurate representation of both the westerly jet stream and the Great Plains low-level jet (LLJ). The RCM reproduces the observed different climate regimes, where rainfall was associated with the periodic (5 day) passage of midlatitude cyclones in June and persistent synoptic circulations in July. The model also correctly simulates the major flood area rainfall diurnal cycle (with the peak amount at 0900 UTC), which follows the LLJ cycle by approximately 3 h.
The CCM2 radiation requires distributions of cloud cover fraction CCF and cloud water path CWP (g m−2) to determine cloud radiative effects. These two fields were originally parameterized in terms of mainly relative humidity RH and air temperature T, respectively. For each vertical model layer, convective cloud cover CC was defined as
where subscripts l and s denote large and small grid-size bounds; N = kt − kb + 1 is the number of convective layers; and kt and kb are the cumulus top and base levels, which are predicted by the cumulus parameterization. The default values are Dl = 200 km, Ds = 10 km, CCl = 0.3, CCs = 1.0. For the current CMM5 grid spacing D = 30 km, the column total convective cloud cover CC0 = 0.86. Note that CC = 0 except within the cumulus column between layers kt, kb as in Eq. (1), where clouds are assumed to randomly overlap. The resolvable or large-scale cloud cover CS was computed by
where δσ is the layer thickness in the σ coordinate. The default values are Dl = 100 km, Ds = 10 km, RHl = 0.75, RHs = 0.9. For the current CMM5 grid spacing D = 30 km, the threshold relative humidity RHc = 0.867. The total layer cloud cover was then calculated as
When CCF > 0, the cloud water path was parameterized by
where δz is the layer thickness (m). For temperature bounds Tn, Tm, Tx = 220, 265, 295 (K), the default cloud water contents are specified as CWCn, CWCm, CWCx = 0.03, 0.2, 0.4 (g m−3).
The preceding scheme was found to overestimate both cloud cover fraction and cloud water path. This causes a deficit of up to 80 W m−2 in solar radiation reaching the surface as compared with station measurements in Illinois (see section 5). A similar deficit was found over the NAM region (Xu and Small 2002). As a result, the simulated surface air temperatures were colder than observations by 2°–5°C during summer over broad regions. Note that the scheme does not make use of the cloud hydrometeors (liquid, ice, rain, snow, graupel), which are predicted by the explicit cloud microphysical solution. Hence the cloud–radiation interactions are not consistently represented. To overcome this problem, the CMM5 incorporates a new cloud prediction scheme. Following Xu and Randall (1996), the resolvable cloud cover is parameterized in terms of relative humidity and cloud hydrometeor concentrations:
where qc, ql, qr, qs, qg, qt are liquid, ice, rain, snow, graupel, and total cloud mixing ratios (kg kg−1), q* is the saturated water vapor mixing ratio (kg kg−1). Equation (1) is retained for convective cloud cover, while Eq. (3) is changed to
Note that the new parameterization, Eq. (5), incorporates the principal physics for cloud formation and thus is independent of the climate model grid size, at least in the range of 64–512 km (Xu and Randall 1996). Its application for somewhat smaller sizes, such as 30 km in the CMM5, is still reasonable (K.-M. Xu 2004, personal communication). In addition, the cloud water path is now calculated as in Dudhia (1989) by
where ρ is the air density (kg m−3), δp is the layer pressure thickness (Pa), g is the gravitational acceleration (m s−2), αi = 0.510417, αr = 0.002292, αs = 0.016250 are the scaling factors to account for differences in absorption coefficients of ice (0.0735), rain (0.33 × 10−3) and snow (2.34 × 10−3) relative to liquid (0.144). The additional quarter-power factors are used to approximate the effects caused by the size spectrum dependence on precipitation (rainfall and snowfall) intensity.
Offline calculations based on the 1993 summer baseline integration indicate that, in absence of feedback, the old cloud–radiation interaction scheme, Eqs. (2)– (4), produces 0.14 (0.52 versus 0.38) more CS and 217 (297 versus 80) larger CWP than the new one, Eqs. (5)– (7), as integrated throughout the column and averaged over all domain grids containing observational precipitation data (over the land of the United States and Mexico; see Figs. 3a–d). The new formulation incorporates direct and consistent interactions between convection, cloud microphysics and radiation processes and, as discussed below, significantly improves the overall CMM5 performance in simulating the regional climate over the United States.
b. Computational domain and lateral boundary conditions
Figure 1 shows the CMM5 computational domain. The domain is centered at (37.5°N, 95.5°W) using the Lambert conformal map projection and a 30-km horizontal grid spacing, with total grid points of 197 (west– east) × 139 (south–north). The domain covers the whole continental United States and represents the regional climate that results from interactions between the planetary circulation (as forced by the LBCs) and North American surface processes, including orography, vegetation, soil, and coastal oceans. The buffer zones are located across 14 grid points along all four domain edges, where LBCs are specified throughout the entire integration period using a dynamic relaxation technique [see Eqs. (1)–(2) in Liang et al. 2001 with L = 15 − 1]. There are 23 vertical layers with the model top at 100 hPa (∼16.3 km above sea level) and 6 layers below 850 hPa (∼1.5 km). This configuration has produced the most skillful simulation of the 1993 summer flood (Liang et al. 2001) and 1986 fall heavy rainfall episodes over the central United States (Kunkel and Liang 2001).
The LBCs are constructed from the National Centers for Environmental Prediction–Department of Energy (NCEP–DOE) AMIP II reanalysis (R-2; Kanamitsu et al. 2002). The R-2 is regarded as an updated and human error–corrected version of the NCEP–NCAR reanalysis (NRA; Kalnay et al. 1996). Liang et al. (2001) showed that RCM simulations are less sensitive to domain choice when the LBCs are based on the NRA rather than the European Centre for Medium–Range Weather Forecasts (ECMWF) reanalysis (ERA; Gibson et al. 1997). This implies that large uncertainties exist in ERA when the domain size increases, especially if the south buffer zone is located in the Tropics. In addition, the ERA data is available only up to 1993 December, while the NRA and R-2 data continue to the present. Given these considerations, the R-2 is chosen to drive the baseline integration during 1982–2002, while the ERA is used in seasonal sensitivity experiments to depict the impacts of LBC uncertainties. Important improvements of the R-2 over NRA include soil moisture, that affects the CMM5 initialization, as well as land precipitation, which provide a better reference of the reanalysis model product for CMM5 downscaling skill assessment. All reanalysis data are the 2.5° latitude × 2.5° longitude grid and pressure-level product at a 6-hourly interval.
c. Surface boundary conditions
The CMM5 requires the specification of surface boundary conditions, including green vegetation fraction over land and sea surface temperature (SST) over oceans. Vegetation fraction was originally fixed in the MM5 at the initial condition. For climate applications, the CMM5 updates daily vegetation variations by linear temporal interpolation of the monthly mean climatology derived from the National Oceanic and Atmospheric Administration/Advanced Very High Resolution Radiometer (NOAA/AVHRR) data (Gutman and Ignatov 1998). The SST daily variations were first incorporated by Liang et al. (2001) using the NRA surface skin temperature, which is a model product over land. This was later found problematic in coastal oceans and the Great Lakes, where the coarse NRA grids cannot appropriately resolve land versus water points. For example, the Gulf of California is not resolved as a water surface, but treated as land points. Given the strong diurnal cycle over land, the surface skin temperature at 0000 UTC (near 1630 local time) is 5°–10°C warmer than the adjacent ocean. This results in overwhelming deep convection and excessive precipitation in the Gulf of California. To eliminate the problem, the CMM5 incorporates observed daily variations based on conservative spline fit from the weekly optimum interpolation SST (OISST) data (Reynolds et al. 2002). The OISST is available over the global oceans on a 1° × 1° grid mesh from 1981 November onward. A time series analysis showed that peak SSTs in September over the Gulf of California were about 5°C colder in 1982 than all other years. This unexplained bias did not occur elsewhere. Note that daily SSTs interpolated directly from weekly values (treated as if they were at the middle of the week) do not conserve weekly means nor preserve the extremes (Taylor et al. 2000). For this reason, an iterative spline-fit procedure was used to interpolate daily SST variations from the weekly data while conserving the weekly means. These daily SST variations were then incorporated into the CMM5 by daily updates. This procedure also effectively preserves the extremes revealed in the original weekly data.
Note that the CMM5 SST so constructed has small differences from those of the R-2, since the latter also used the same OISST with a similar time interpolation procedure (Kanamitsu et al. 2002; Fiorino 2000). Comparisons between the two SST data, including monthly mean rms errors and interannual correlations, show that they are essentially identical except over limited coastal oceans and the Great Lakes where differences of ±0.5°– 3°C occur. Hence the differences do not affect regional climate signals forced by the planetary SST anomalies nor explain the substantial precipitation contrasts between the CMM5 and the driving R-2 discussed later.
d. Model simulations and validation data
Given the limited availability of the OISST data, the CMM5 baseline integration was initialized at 0000 UTC 2 December 1981, and completed at the end of 2002. The initial month is considered as a spinup period, which is generally sufficient based on the examination of daily rainfall variations. One exception is near the Gulf of California, where the OISST contained apparent biases in 1982. Analysis showed that the removal of 1982 has little impact on the climatology. This study compares the CMM5 baseline and observed climatologies during 1982–2002 to determine the model biases in simulating the precipitation annual cycle over the United States. The climatologies are also compared with that produced by the reanalysis assimilation model to assess the skill enhancement in resolving regional climate variations as a result of the CMM5 downscaling approach. To understand these downscaling skills and biases, many sensitivity experiments were conducted that branch off from the restart conditions of the baseline integration. These experiments vary in the simulation periods (initial and length) depending on their specific application purposes. All experiments contain at least one month spinup time, where the model outputs are not used in the analysis. Note that some experiments cannot restart from the baseline conditions because a new physics scheme requires different or more initial states. In this case, a spinup of 2 months is assumed.
Several major daily datasets are utilized in the validation. For the circulation fields (200/850 hPa wind and sea level pressure), the R-2 data are taken as the best proxy of observations because they assimilated all available measurements. They are available on a global 2.5° × 2.5° grid mesh during the entire validation period. For a quantitative comparison, all these fields were mapped onto the CMM5 grids using bilinear spatial interpolation.
Daily precipitation data are a composite of three analysis sources, all based on gauge observations over land. The primary source was constructed by the Surface Water Modeling (SWM) group at the University of Washington, providing daily precipitation on a 1/8° latitude × 1/8° longitude grid mesh over the continental United States before 31 July 2000 (Maurer et al. 2002). It was produced from the daily observations of the National Weather Service Cooperative Observer Network using first the synergraphic mapping system algorithm (Shepard 1984) for geographic interpolation and then adjusted to match the 1961–90 monthly climatology of the Parameter-Elevation Regression on Independent Slopes Model (PRISM) developed by Daly et al. (1994, 1997). Given the strong precipitation dependence on elevation, the topographic adjustment is critical because the cooperative stations over mountainous regions are preferentially located at lower elevations and thus tend to underestimate the true spatial average. The SWM data were mapped on the CMM5 grids by mass-conservative spatial averaging.
The remaining period, from 1 August 2000 onward, was supplemented by a Cressman (1959) objective analysis using daily measurements from 7235 cooperative stations over the United States. The station density is generally high and compatible with the CMM5 30-km grid mesh except for mountainous regions in the Rockies. To best resolve observational details, the objective analysis first uses an influence radius of 40 km for each station. For the total of 8471 model grids over the continental United States, approximately 200–400 grids (depending on dates) have no data after the first analysis. A second analysis using an 80-km radius fills most of these missing values. The remaining 1–10 points are defined by a third analysis with a 100-km radius. Note that this supplementary analysis, hereafter referred to as STM (station to model), lacks the PRISM adjustment for the statistical topography–precipitation relationships.
The third source is the NCEP historical daily precipitation analysis over the United States and Mexico using a modified Cressman scheme (Higgins et al. 2000). This is to supplement the data over Mexico, where about 300 (600) gauge stations before (after) 1990 were incorporated in the analysis. This analysis has a coarse resolution, with a 1° × 1° grid mesh and available till the end of 2001. This is continued with the NCEP real time daily precipitation analysis following the same approach but a finer resolution on a 1/2° × 1/2° grid mesh. Given the relatively coarse resolution, these NCEP analyses were mapped onto the CMM5 grids using bilinear spatial interpolation.
The composite analysis thus contains daily precipitation data on the CMM5 grids over the United States and Mexico during the entire baseline integration period. No data is available over Canada. Certain inconsistencies exist due to the subtle differences outlined earlier. Figure 2 compares the annual mean climatologies processed on the CMM5 grids from the SWM and STM analyses during the overlap period 1982–99. As expected, apparent differences (strongest in winter and weakest in summer) occur over major mountainous regions with sparse gauge station densities. In the western United States where westerly flow prevails during October–May, the SWM yields persistently heavier precipitation (by 1–3 mm day−1) than the STM on the west sides (upstream) of the Coast Range, the Cascade–Sierra Nevada ranges, and the peak ranges of the Rocky Mountains in Montana (Big Belt Mountains), Wyoming (Absaroka and Wyoming ranges) and Colorado (Colorado Front and Sangre De Cristo ranges). The PRISM seems to realistically capture the mesoscale patterns of the orographic precipitation enhancement by the forced uplifting. In the eastern United States, the orographic effect of the Appalachian Mountains is smaller and the density of observing stations is sufficient to capture the main features. The small SWM and STM difference there may also reflect the increased density of gauge stations. The lack of the PRISM effect in the composite during 2000–02 accounts for an underestimation of less than 0.5 mm day−1 in the 1982–2002 climatology over the aforementioned mountain ranges. In addition, the NCEP and STM differences are generally small except for up to ±1 mm day−1 in northern California and the western half of both Washington and Oregon as well as near the coastal areas of the southeast states. These differences resulted mainly from the analysis resolution changes.
3. Annual cycle of precipitation
Figure 3 illustrates the 1982–2002 averaged seasonal mean geographic distributions of the observed and CMM5 simulated precipitations. Observations exhibit generally wetness east of 100°W and west of the Rockies, with a dry zone in between along the downstream slopes of the Rockies. Over the central United States, precipitation decreases rapidly from southeast to northwest in winter while remaining rather uniform in summer, with transitions in spring and fall. The CMM5 well reproduces these characteristics. The model downscaling skill, however, shows distinct regional dependencies.
The CMM5 realistically simulates mesoscale orographic precipitation patterns and their seasonal variations west of the Rockies. When compared to the composite data, the model produces generally more winter precipitation west of the Rockies, especially over the west slopes of all major peak mountain ranges except for a deficit west of the Coast Range. The differences decrease in spring and fall, and are not evident in the summer dry season. These regions overlap with the areas of significant orographic precipitation effects (Fig. 2) generated by upslope motions within the eastward-moving Pacific storm systems. Given the lack of direct measurements at high elevations over steep mountainous regions, the uncertainties of the PRISM orographic adjustments, and hence the composite analysis, may likely be large. The negative (positive) CMM5 differences from the composite analysis along the west (east) side of the Coast Range, indicating a phase error, are likely caused by the difference in terrain elevation at a grid spacing between 30 km for the model and 1/8° (∼10 km) for the PRISM.
The CMM5 accurately simulates the transition dry zone on the east downstream slopes of the Rockies due to the precipitation shadowing effect of mountain ranges. The model also correctly reproduces seasonal variations of precipitation over the Midwest. The biases over these regions are mostly within ±0.5 mm day−1, except for the Great Lakes region in winter and spring where overestimations (by <2 mm day−1) occur. Data quality analysis indicated that rain gauge measurements in this region generally underestimate the actual snowfall water equivalent (Groisman et al. 1994, 1996). This may also apply for the western mountainous regions. Hence the earlier CMM5 differences from the composite analysis may partially be attributed to observational uncertainties. The CMM5's success in the central United States, especially in summer, is encouraging since it is a common difficulty for other RCMs to realistically reproduce the local precipitation variations (Takle et al. 1999) that are dominated by regional mesoscale circulations, including the nocturnal LLJ (Stensrud 1996; Higgins et al. 1997; Liang et al. 2001) and mesoscale convective complexes (Maddox 1980; Fritsch et al. 1986).
On the other hand, the CMM5 essentially fails to simulate the summer NAM rainfall pattern. The NAM results from a combination of large-scale circulation forcing and local diabatic heating (Barlow et al. 1998). The former is incorporated into the CMM5 through the dynamic relaxation of the LBCs, while the latter is determined by the model physics, especially surface–atmosphere and cumulus-cloud–radiation interactions. The CMM5 failure is caused by substantial uncertainties in the LBCs over the west and south buffer zones due to the lack of observations over oceans assimilated into the reanalyses (Liang et al. 2001), large errors in surface forcing by the use of coarse-resolution SST data (Gao et al. 2003), and strong diabatic heating sensitivities to physics representations. The NAM simulation remains a challenging issue in both RCM and GCM communities. Even the global reanalyses cannot faithfully produce a consistent NAM system (Barlow et al. 1998). In addition, the CMM5 underestimates rainfall (by <3 mm day−1) in the Gulf States throughout the year and eastern states in summer and fall. Similar underestimations were identified in other RCM simulations (Giorgi et al. 1994; Takle et al. 1999; Pan et al. 2001). This deficiency may partially be explained by large LBC errors over the south and east buffer zones. The correct mechanisms responsible for these biases are not known and will be explored in the subsequent sections through circulation analyses and sensitivity experiments.
Given the same large-scale forcing, the CMM5 and R-2 may produce different regional precipitation patterns over the United States, especially when local surface and convective processes dominate. Given no direct incorporation of precipitation measurements, the R-2 precipitation was generated by the global assimilation model in balance with the constraint of observed atmospheric circulation fields. Thus the R-2 result provides a reference for the CMM5 downscaling skill enhancement due to its refined spatial resolution and physics representation. Figure 4 illustrates the 1982–2002 averaged seasonal mean geographic distributions of the R-2 precipitation. Because of the coarse-resolution T62 (∼210 km), the R-2 cannot resolve the mesoscale orographic precipitation patterns west of the Rockies, showing a single broad center with a gradual inland reduction from the coastal Pacific Ocean. Precipitation is underestimated on the west slopes of all major mountain ranges, while overestimated on the east slope of the Cascade Range. The R-2 produces substantially heavier summer rainfall over Mexico as compared to observations (by >4 mm day−1). Similar and weaker overestimations occur in other seasons. These are opposite to the CMM5. Another opposite bias appears in summer over the Gulf and eastern states, where rainfall is excessive in the R-2 while insufficient in the CMM5. The NRA yields even greater wet biases over these regions, which are identified with excessively large diurnal cycles in the assimilation model (Higgins et al. 1997). The R-2 dry biases over the Gulf States are greater (smaller) than those of the CMM5 during winter (fall), while comparable in spring. The R-2 poorly simulates the transition dry zone on the east slopes of the Rockies with more precipitation in spring and summer.
Figure 5 compares the observed, the R-2, and CMM5 simulated monthly precipitation variations averaged over eight key regions (see specification in Fig. 1). These regions have distinct climate regimes and/or model biases, for which extended discussions are given later. The CMM5 has the most realistic simulation over the Cascades, with a substantial improvement in cold rainy seasons compared to the R-2. The seasonal cycle amplitudes for the northern Rockies, Midwest, and Northeast are enhanced in the CMM5, which produces more (less) precipitation in winter (fall) than observations. The amplitudes are also increased in the R-2, though more (less) precipitation than observations occurs in summer (winter). For the Southeast, central Great Plains, and NAM region, the CMM5 simulations are generally realistic in winter and spring, while substantial underestimations are identified in summer and partially in fall (except the Southeast). As discussed in section 5, the summer dry biases over the Northeast, Southeast, central Great Plains, and the NAM region are largely removed when the CMM5 uses the Kain–Fritsch instead of Grell cumulus scheme. The Gulf States seem to be the most problematic, where the CMM5 has overall large dry biases except in August while the R-2 generates substantially larger (smaller) rainfall in summer (winter) than observations. The use of the Kain–Fritsch scheme provides little improvement (see below).
Table 1 compares the observed, CMM5, and R-2 simulated seasonal mean precipitation statistics over four broad sectors (see Fig. 1 for the boundary specification) and the entire domain. The statistics includes spatial averages and deviations, pattern correlations (see Giorgi et al. 1993a), as well as percentages of grid points where the CMM5 produces smaller precipitation absolute errors than the R-2 and vice versa. The calculation is based on the 1982–2002 climatologies and accounts for all domain grids containing observational precipitation data (over the land of the United States and Mexico; see Figs. 3a–d). The CMM5 realistically simulates the observed precipitation spatial variability over all sectors in each season, whereas the R-2 produces substantially smaller (greater) deviations in fall, winter, and spring (summer). The CMM5 generates smaller absolute errors over more extensive areas than the R-2 for the fall and winter; the opposite occurs in summer, for which a significant improvement can be made by an ensemble simulation using the Grell and Kain–Fritsch cumulus schemes (see section 5); the spring values are compatible. The pattern correlations with and the average differences from observations show no systematic tendency for the superiority between the CMM5 and R-2. These two measures seem unable to represent mesoscale structure contrasts. Note that the R-2 assimilated all local observations of circulation fields, including wind, temperature, humidity, and surface pressure, to produce a balanced precipitation. As such the R-2 biases indicate important assimilation model deficiencies in physics representation, especially the boundary layer treatment, explicit cloud microphysics, and cumulus parameterization. On the other hand, the CMM5 biases are overall smaller than the R-2, indicating a high credibility of the downscaling skill. The next two sections will explain the existing CMM5 biases.
4. Associated regional climate circulation
A serious question is whether the CMM5 precipitation biases result from model formulation deficiencies or LBC forcing errors. Since the LBCs were constructed from the R-2 that assimilated all available observations of circulation fields, the CMM5 simulations of these fields are expected to have the least differences from the R-2, especially within the buffer zones. These fields are first examined in this section to see if they are correctly represented through dynamic relaxation of the LBCs. Figure 6 shows seasonal mean distributions of sea level pressure and 850-hPa wind differences between the CMM5 and R-2.
At sea level, the Gulf and eastern states are under the influence of the Atlantic subtropical high throughout the year, while the western states are affected by the Pacific subtropical high. During the winter, a stationary high persists over the Great Basin and traveling anticyclones prevail east of the Rockies. Cyclones and hurricanes penetrate through or overwhelm these high pressure systems to produce local precipitation. The CMM5 closely reproduces these large-scale features (not shown). Subtle but important biases exist. The sea level pressure biases are especially large in fall and summer when significantly less precipitation is simulated over the southern and eastern states. When the domain means are removed, the pressure biases are generally opposite to the precipitation biases. In particular, both Atlantic and Pacific subtropical highs are too strong and farther inland, causing less precipitation in the Gulf and eastern states while the NAM is totally missed. On the other hand, the pressures are underestimated over the Great Lakes region, where more precipitation is simulated.
The reasons for the CMM5 sea level pressure biases are two-fold. The first explanation may arise from the relative LBC errors. The errors are especially large in the south, west, and east buffer zones, mostly over oceans where the R-2 assimilated few observations. In contrast, many more observations were used in the north buffer zone, mostly over land, which forces a balanced atmosphere circulation that is consistent with observations. Without this observational constraint in other buffer zones, the R-2 model formulation deficiencies introduce large inconsistencies between the LBCs of major circulation fields (e.g., wind, temperature, humidity, pressure). Together with incompatible model representations at the coarse R-2 versus fine CMM5 resolutions, these inconsistencies could result in large CMM5 biases, starting in the unrealistic buffer zones and ultimately spreading into the inner domain. On the other hand, the pressure biases can also be internally generated by model deficiencies (see section 5).
As expected, the wind differences are generally small within all buffer zones except over the Sierra Madres where steep mountains cause large spatial interpolation errors in the LBCs. Systematic wind differences at 850 hPa in the Rockies depict terrain elevation contrasts at the CMM5 fine versus R-2 coarse resolutions. These differences are often below the ground and thus not meaningful. There are, however, important regional biases that are consistent with those in precipitation and sea level pressure discussed earlier. First, the CMM5 produces 850-hPa anticyclonic flow departures (reversal aloft with the center tilted inland) over the Pacific Ocean coast near Mexico in summer, indicating a much weaker monsoon circulation and hence higher sea level pressure and less NAM rainfall. Second, the summer CMM5 LLJ is stronger than the R-2 (by ∼35% in the core). Given the distinct LLJ diurnal cycle and inability for the twice-daily soundings (thus the R-2) to resolve it (Liang et al. 2001), this difference may not be a CMM5 deficiency but an improvement over the R-2 due to resolution enhancement. On the other hand, the CMM5 generates enhanced lower (upper)-level southerly/southeasterly (southwesterly) flows over the Gulf of Mexico and its coastal states, which are identified with higher sea level pressure and insufficient precipitation. Third, in fall, the easterly flow on the south flank of the Atlantic subtropical high is increased, indicating a systematic northward shift of the high that leads to less precipitation over the Gulf and eastern states. Fourth, over the Great Lakes region in winter and spring, the CMM5 simulates stronger westerly (easterly) flow to the north (south) at 200 hPa, suggesting a northward shift of the upper-level jet stream. This is accompanied by stronger lower-level westerly flow in winter and southwesterly flow in spring. These differences are associated with upper-level divergence, sea level pressure cyclonic perturbations and increased regional precipitation.
5. Possible reasons for model biases
Figure 3 reveals that CMM5 dry biases are particularly large over the NAM region and the East Coast in summer, over the lower Mississippi River basin in fall, and over the Gulf States in winter. In addition to the LBC uncertainties discussed earlier, these biases can also be generated internally by deficiencies in the CMM5 physics representation. To address this issue, sensitivity experiments are conducted for the 1984 fall and winter as well as the 1988 and 1993 summers. The CMM5 precipitation biases in these seasons are typical of the long-term means.
Figures 7a–d illustrate the 1988 and 1993 summer rainfall differences from observations for the CMM5 simulations using the Grell versus Kain–Fritsch cumulus schemes with all other model dynamic and physic configurations (including the new CCF and CWP) being identical. The 1988 and 1993 summers were characterized by extreme drought and flood conditions, respectively, over the Midwest (Trenberth and Branstator 1992; Kunkel et al. 1994), accompanied by heavy and light NAM precipitation. They are identified with physical mechanisms at both the planetary and local scales (Trenberth and Guillemont 1996) and thus are widely used for evaluation of RCM performance (Giorgi et al. 1996; Bosilovich and Sun 1999; Hong and Leetmaa 1999; Liang et al. 2001). For both summers, the baseline integration with the Grell scheme underestimates rainfall by 1–4 mm day−1 over the NAM region and the East Coast. By contrast, the Kain–Fritsch scheme generally overestimates these regional rainfall amounts by 1–4 mm day−1, although a weaker dry bias still exists in the northern NAM area. As such, the two schemes produce substantial precipitation differences while their composite contains smaller biases compared to observations, especially along the East Coast. Note that the Kain–Fritsch scheme generates systematically excessive rainfall over the Atlantic Ocean as compared with the analysis of Xie and Arkin (1996). On the other hand, it fails to simulate the 1993 flood over the Midwest. The significant CMM5 sensitivity to cumulus parameterizations arises because of the convective predominance in summer precipitation (Heideman and Fritsch 1988) and its strong dependence on subtle ambient forcing differences (Crook and Moncrieff 1988). In particular, over the central United States, mesoscale convective systems prevail and account for much of the regional precipitation (Fritsch et al. 1986).
Diagnoses of the 1993 summer simulations show that the Kain–Fritsch scheme consumes convective available potential energy (CAPE) much more effectively than the Grell scheme to produce heavier rainfall, especially over oceans. The partition between the parameterized and resolved precipitation is also very sensitive to the cumulus schemes. The CMM5 with the Grell scheme generates relatively greater resolved precipitation (by the explicit cloud microphysics), while producing smaller convective rainfall. Given the complexity of nonlinear interactions between various physics processes, it is extremely difficult to identify which formulation contrasts that likely explain the CMM5 result differences when using the Grell versus Kain–Fritsch schemes (see further discussion in section 6). Nonetheless, the opposite responses to the two schemes provide an excellent opportunity to utilize the CMM5 to better understand the predictability of the U.S. warm season precipitation.
The impact due to the LBC uncertainties is not negligible. Figures 7e–f compare the CMM5 simulations with the Grell scheme as driven by the R-2 versus ERA LBCs. For the 1988 summer, the ERA forces the CMM5 to produce more rainfall (by 1–2 mm day−1) over the southern Great Plains, which approximately removes the biases resulted from the R-2 forcing. On the other hand, the ERA forcing generates excessive rainfall over the coastal Atlantic Ocean, much resembling the biases caused by the Kain–Fritsch cumulus scheme. During the 1993 summer, the ERA LBCs also somewhat reduce the dry biases, with heavier precipitation over the southern Great Plains. In contrast, the responses over the Atlantic Ocean and the southwestern Gulf Coast are opposite to those in 1988 and differ from the result of the Kain– Fritsch cumulus scheme.
The representation of cloud–radiation interaction plays an essential role in determining regional precipitation. For the 1993 summer, the old scheme substantially underestimates rainfall (by 1–4 mm day−1) over the Midwest flood area, while generating excessive precipitation in the coastal Atlantic Oceans (Fig. 7h). A similar but weaker sensitivity occurs in 1988 summer (Fig. 7g). The result is attributed to the fact that the old scheme largely overestimates cloud cover and water path, and thus significantly reduces surface solar insolation and consequently evapotranspiration to produce much weaker regional water recycling over land. For example, the 1993 (1988) summer mean solar insolation measured at Illinois stations was 246.8 (288.6) W m−2. The corresponding values are 223.9 (285.6) and 164.3 (223.6) W m−2 for the CMM5 incorporating the new and old schemes, respectively. These differences are approximately balanced by opposite changes in surface latent heat flux and thus affect regional water recycling.
Figure 8 depicts the CMM5 differences of sea level pressure and 850-hPa wind between the Kain–Fritsch versus Grell cumulus schemes in the 1988 and 1993 summers. For both cases, the Kain–Fritsch scheme simulates systematically lower sea-level pressures, especially over the East Coast, the Atlantic Ocean, and the Gulf of Mexico. These largely reverse, and most often overwhelm, the positive biases in the Grell scheme (see Fig. 6). Meanwhile, stronger southerlies over the Gulf of Mexico enhance moisture convergence and consequently rainfall over the Gulf States. Stronger easterlies over Texas and northeasterlies from the Midwest increase mass convergence toward Colorado and New Mexico, causing more rainfall over the northern NAM region while a deficit over the Midwest. Stronger southwesterlies lead to convergence and heavier rainfall over the Atlantic coast. These circulation changes are the consequence of differential heating differences due to the latent heat release between the two schemes (Gochis et al. 2002). The result suggests that the cumulus parameterization alone could explain a large portion of the CMM5 sea level pressure biases (Fig. 6).
Figure 9 shows 1984 fall precipitation differences between the CMM5 simulations using the Grell versus Kain–Fritsch cumulus schemes as well as the replacement of ERA LBC forcing and the removal of all precipitating processes over the Gulf of Mexico. The last experiment was motivated by the speculation that moisture from the Gulf of Mexico could rain out prematurely before moving inland to produce sufficient precipitation over the Gulf States (Pan et al. 2001). Again, the Kain– Fritsch scheme produces more rainfall over the NAM region, which counteracts the dry biases in the baseline integration. On the other hand, the scheme increases the major dry biases over the lower Mississippi River basin. It also continues to generate unrealistically excessive precipitation over the Atlantic Ocean. When the precipitating processes, including the explicit cloud microphysics and cumulus parameterization, are completely turned off over the Gulf of Mexico, the CMM5 simulates modest precipitation increases (by 0.5–2 mm day−1) over the lower Mississippi River basin. This may support the notion that more precipitation over the basin could occur if the Gulf storms do not prematurely rain out moisture. The basin precipitation gain, however, is much less and does not compensate for the dry biases, even though the storms produce not a single drop of rain over the Gulf. In addition, the dry biases are somewhat enhanced when only the cumulus parameterization is turned off over the Gulf (not shown). Pan et al. (2001) also argued that the omission of rainwater as a predictive variable may cause cloud water to rain out immediately, and thus reduce downstream moisture transport. In the CMM5, the rainwater together with other hydrometeors is predicted by two comprehensive explicit cloud microphysics solutions (Tao and Simpson 1989; Reisner et al. 1998). Both solutions, however, produce similar dry biases.
The impact from LBC uncertainties on 1984 fall precipitation is particularly large over the southeast portion of the domain. The ERA LBCs force the CMM5 to generate much more rainfall over the southeast coast, where the baseline integration with the R-2 forcing is generally realistic. On the other hand, the sensitivity over the southern Great Plains is relatively small. Pan et al. (2001) also found small regional precipitation influences from a 1500-km equatorward shift of the south buffer zone. None of these results, however, can justify concluding that LBC errors are not the cause of the problem for the major dry biases over the lower Mississippi River basin. Given any of the existing reanalyses, LBC uncertainties are especially large over the south, east, and west buffer zones, mostly over oceans where observations are lacking for data assimilation. It is not known, however, which reanalysis, and perhaps none of them, best represents the reality. Since the south buffer zone in the Pan et al. model and the current CMM5 is located in an already problematic area, its further southward expansion will not improve the lateral forcing but instead may worsen it (Liang et al. 2001).
The fall dry biases over the lower Mississippi River basin are common not only to RCMs (Giorgi et al. 1994; Takle et al. 1999; Pan et al. 2001), but also to GCMs (Duffy et al. 2003). Figure 10 illustrates the basin-mean fall precipitation climatology during 1979–95 for 21 AMIP (Atmospheric Model Intercomparison Project; Gates et al. 1998; see online at http://www-pcmdi.llnl.gov/projects/amip/index.php; for updates) and the 30-yr control period for 9 CMIP (Coupled Model Intercomparison Project; Meehl et al. 2000) GCMs. The observed, R-2, and CMM5 simulated precipitation are 3.6, 3.2, 2.4 mm day−1, while AMIP model values range from 0.7 to 2.8 mm day−1 and CMIP model values from 1.3 to 3.0 mm day−1, with the composite means of 1.8 and 2.2 mm day−1, respectively. Hence, similar to CMM5, all AMIP and CMIP models have large dry biases in fall. The AMIP and CMIP composites (averages of all GCMs) indicate that, in fall, the minimum 850-hPa wind speed line (along the east–west axis of the Atlantic subtropical high) in the Gulf of Mexico is shifted to the north and extended to the west. This shift causes westerly to northwesterly biases over the Gulf States and coastal oceans, which lead to an overall weaker inland moisture advection from the Gulf of Mexico and consequently precipitation deficit over the basin. Similar 850-hPa flow biases exist in the CMM5 (see Fig. 6). Most recently, Gutowski et al. (2004) 1 showed that the primary contributor to their RCM dry biases is a deficit in evapotranspiration. They speculated that this deficit may result from the underrepresentation of terrestrial water reservoirs and/or unrealistic moisture convergences due to mesoscale circulation errors. The exact physical mechanisms responsible for these biases remain unknown and will continue to be a challenging issue for RCM and GCM studies.
The winter dry precipitation biases over the Gulf States may not be caused by deficiencies in the CMM5 physics representation. CMM5 experiments with the cumulus parameterization and explicit cloud microphysics solution show little winter response in general. Nor are these biases common to GCMs as diagnosed from the AMIP and CMIP simulations. The most likely reason seems to be associated with the LBC errors. Note that the R-2 itself produces broader and greater dry biases over the region (cf. Figs. 3 and 4). Given that winter precipitation episodes over the Gulf States are produced primarily by storms originating over the Gulf of Mexico (Businger et al. 1990), the R-2 seems to inadequately resolve these storms and thus the resultant LBCs in the south buffer zones are erroneous, causing the CMM5 dry biases. More comprehensive diagnoses are needed to better understand the model capability in resolving the Gulf storms, their frequencies, intensities, trajectories, and landfalls.
Figure 11 summarizes seasonal mean precipitation amounts averaged over the eight key regions (see specification in Fig. 1) as observed and simulated by the CMM5 using the Grell and Kain–Fritsch cumulus schemes for all sensitivity experiments discussed above. The CMM5 responses are generally small for all regions in the 1984 winter. This is also true for the 1984 fall except for the NAM (Gulf States) where the Kain– Fritsch scheme produces more (less) rainfall, an improvement (deterioration) over the Grell scheme. The responses are particularly large in summer for both 1988 and 1993, where the Kain–Fritsch scheme yields overall heavier and more realistic precipitation over the continental United States. There are two important exceptions. Over the Midwest, the Kain–Fritsch scheme underestimates the 1993 flood by approximately 25% while the Grell scheme is very close to observations. In contrast, over the Southeast for both 1988 and 1993, the Kain–Fritsch scheme results in excessive rainfall while the Grell scheme gives large deficits.
The discussion so far has been focused on the dynamical effects on precipitation. It is generally understood that precipitation is strongly affected by thermodynamics (e.g., through changes in buoyancy production due to temperature differences). Diagnoses however indicated complex relationships, if any, between precipitation and surface air temperature biases (not shown). As compared with the STM data during 1982–2002, the CMM5 produces winter warm biases (1°–3°C) west of the Rockies, with the spatial pattern closely resembling Fig. 2 where precipitation enhancement results from the orographic effect. A similar bias pattern is generated in summer without obvious rainfall correspondence. In spring, the CMM5 has (1°–3°C) warm biases over the Great Plains where rainfall biases are small (Fig. 1). The fall temperature biases are small everywhere, within ±1°C of observations over the entire domain. The reason for the lack of coherent precipitation–temperature bias relationship warrants further investigation.
A continuous CMM5 baseline integration for the period 1982–2002 driven by the R-2 reanalysis and supplementary seasonal sensitivity experiments with various LBCs and physics representations were conducted to determine the RCM capability to reproduce the observed annual cycle of U.S. precipitation and to better understand the causes for model biases. The CMM5, incorporating an improved cloud–radiation interaction parameterization, well reproduces the major characteristics of geographic distributions and seasonal variations of observed precipitation. It simulates more realistic regional details with less overall biases than the R-2, indicating a high credibility of its downscaling skill.
Several important regional biases were identified. Sensitivity experiments using the ERA versus R-2 data indicate that the LBC uncertainties, especially in the south, east, and west buffer zones, account for certain portions of these biases. The winter dry biases in the Gulf States may likely result from the LBC errors. The summer dry biases in the NAM region and East Coast can be largely removed by replacing the Grell with the Kain–Fritsch cumulus scheme, although the latter often yields heavier rainfall than observations, especially in the Atlantic Ocean. Over the Midwest, the Grell scheme produces realistic summer rainfall, while the Kain– Fritsch scheme generates large dry biases. The CMM5 simulations using the two cumulus schemes contain opposite biases in summer and their composite tends to be the best match for observations. The fall dry biases over the lower Mississippi River basin are common to all existing RCMs and GCMs, but the responsible physical mechanisms are not known. The winter and spring wet biases over the Great Lake region and northern Rockies may partially be attributed to the rain gauge underestimation of actual snowfall water equivalent and the inaccurate representation of orographic precipitation enhancement by the PRISM with sparsely distributed, mainly low-elevation stations over the major mountains. We speculate that these wet biases could also result from unrealistic model physics representations, though the available cumulus parameterization and explicit cloud microphysics schemes reveal little sensitivity.
In the coming series of the companion papers, we will elaborate on the CMM5 capability to downscale U.S. precipitation variability at various time scales, including the diurnal cycle, daily fluctuations, extreme events, intraseasonal evolutions, and interannual variations. In particular, we will further evaluate the CMM5 responses to the Grell versus Kain–Fritsch cumulus schemes at these scales. Although many studies have demonstrated the superiority of the Kain–Fritsch scheme (Kuo et al. 1996; Wang and Seaman 1997; Cohen 2002; Gochis et al. 2002), the Grell scheme has its own compelling advantages. As discussed in this and upcoming papers, the Kain–Fritsch scheme generates excessive rainfall over the Atlantic Ocean, underestimates summer precipitation in the Midwest, and fails to reproduce the diurnal cycle in the Great Plains, while the Grell scheme does all of these reasonably well. Furthermore, these regional responses may change when the CMM5 is coupled with different representations of cloud–radiation interactions (Xu and Small 2002) and land surface processes (Pan et al. 1996) or driven by different LBCs containing large uncertainties. Unfortunately, no general theory of cumulus parameterization is available, nor has a single scheme been proven to work perfectly across a wide range of weather systems and grid sizes, especially at 10–50-km resolutions where mesoscale processes are partly resolved and parameterized (Molinari and Dudek 1992; Emanuel and Raymond 1993; Arakawa 1993).
The Grell scheme is based on Arakawa and Schubert (1974) with a simplified single-cloud model, while the Kain–Fritsch scheme is developed from Fritsch and Chappell (1980) with a sophisticated cloud model. Although both represent updrafts and downdrafts, the Kain–Fritsch scheme incorporates detailed cloud microphysics as well as entrainment and detrainment between the cloud and environment, all of which are absent in the Grell scheme. When convection is triggered, the Kain–Fritsch scheme removes all CAPE within the relaxation time, whereas the Grell scheme adjusts the buoyancy toward an equilibrium state depending on the strength of cloud-base vertical motion. The rainfall is then parameterized as the product of precipitation efficiency with integrated water vapor and liquid flux about 150 hPa above the lifting condensation level in the Kain–Fritsch scheme, but with total condensate and cloud-base mass flux in the updraft for the Grell scheme. The precipitation efficiency is a function of mean wind shear of the convective column in both schemes, with additional dependence on the cloud-base height for the Kain–Fritsch scheme. It is beyond the scope of this study to identify which, and perhaps all, of the previously mentioned formulation contrasts that are responsible for the CMM5 result differences.
We thank Jimy Dudia and George Grell for numerous discussions on the model results, and Steven Hollinger, David Kristovich, and two anonymous reviewers for constructive comments on the manuscript. We appreciate Wei Wang and John Michalakes for their assistance during the MM5 implementation. We are grateful to Edwin Maurer for providing the SWM rainfall data. We acknowledge NCAR for access to the MM5 system and the NCEP–DOE and ECMWF global reanalysis data, and FSL/NOAA and NCSA/UIUC for the supercomputing support. Station data were provided by the Midwestern Regional Climate Center under NOAA Cooperative Agreement NA67RJ0146. The research was partially supported by NOAA/GAPP Grant NA06GP0393. The views expressed are those of the authors and do not necessarily reflect those of the sponsoring agencies or the Illinois State Water Survey.
Corresponding author address: Dr. Xin-Zhong Liang, Illinois State Water Survey, University of Illinois, Urbana–Champaign, 2204 Griffith Dr., Champaign, IL 61820-7495. Email: firstname.lastname@example.org
We noticed this work at the final revision process.