Nonhydrostatic nested climate modeling: A case study of the 2010 summer season over the western United States

Authors

  • Bereket Lebassi-Habtezion,

    Corresponding author
    1. Department of Environmental Earth System Science and Woods Institute for the Environment, Stanford University, Stanford, California, USA
    • Corresponding author: B. Lebassi-Habtezion, Department of Environmental Earth System Science, Stanford University, Stanford, CA 94305–4216, USA. (bereketl@yahoo.com)

    Search for more papers by this author
  • Noah S. Diffenbaugh

    1. Department of Environmental Earth System Science and Woods Institute for the Environment, Stanford University, Stanford, California, USA
    Search for more papers by this author

Abstract

[1] The potential importance of local-scale climate phenomena motivates development of approaches to enable computationally feasible nonhydrostatic climate simulations. To that end, we evaluate the potential viability of nested nonhydrostatic model approaches, using the summer climate of the western United States (WUSA) as a case study. We use the Weather Research and Forecast (WRF) model to carry out five simulations of summer 2010. This suite allows us to test differences between nonhydrostatic and hydrostatic resolutions, single and multiple nesting approaches, and high- and low-resolution reanalysis boundary conditions. WRF simulations were evaluated against station observations, gridded observations, and reanalysis data over domains that cover the 11 WUSA states at nonhydrostatic grid spacing of 4 km and hydrostatic grid spacing of 25 km and 50 km. Results show that the nonhydrostatic simulations more accurately resolve the heterogeneity of surface temperature, precipitation, and wind speed features associated with the topography and orography of the WUSA region. In addition, we find that the simulation in which the nonhydrostatic grid is nested directly within the regional reanalysis exhibits the greatest overall agreement with observational data. Results therefore indicate that further development of nonhydrostatic nesting approaches is likely to yield important insights into the response of local-scale climate phenomena to increases in global greenhouse gas concentrations. However, the biases in regional precipitation, atmospheric circulation, and moisture flux identified in a subset of the nonhydrostatic simulations suggest that alternative nonhydrostatic modeling approaches such as superparameterization and variable-resolution global nonhydrostatic modeling will provide important complements to the nested approaches tested here.

1 Introduction

[2] The prospect of “committed” climate change [e.g., MacCracken, 2008] raises the likelihood that natural and human systems will have to adapt to additional climate change, regardless of any climate change mitigation policies that may be implemented [Kelly and Adger, 2000; Smit and Wandel, 2006; Diffenbaugh et al., 2011b; etc.]. This emphasis on adaptation has increased the need to understand the processes that regulate the local-scale phenomena that most impact natural and human systems [Easterling et al., 2000; Diffenbaugh et al., 2005, 2008; White et al., 2006]. Those local-scale phenomena are influenced by processes that span a broad range of scales, including the global-scale general circulation of the atmosphere and ocean, synoptic-scale weather systems, and meso- and submesoscale phenomena such as fronts, drylines, land/sea breezes, mountain/valley circulations, coastal ocean upwelling, and urban heat island [UHI] effects. The fact that the current generation of global climate models is mostly run at grid spacing that are too coarse to explicitly resolve these important submesoscale processes provides a primary motivation for developing high-resolution climate modeling approaches.

[3] High-resolution climate models offer at least two important opportunities to improve the simulation of local-scale climate phenomena. First, high-resolution climate models improve the representation of boundary conditions such as topography, coastline, and land cover, thereby increasing the accuracy of the simulated climate in comparison with state-of-the-science global climate models [Bell et al., 2004; Zhang et al., 2009; Diffenbaugh and Ashfaq, 2010]. Improved resolution of these boundary conditions can be particularly important for capturing the role that fine-scale climate processes play in shaping the response of impacts-relevant climate phenomena to changes in radiative forcing [e.g., Diffenbaugh et al., 2005, 2006, 2011a; White et al., 2006; Rauscher et al., 2008]. Second, by allowing for nonhydrostatic model formulation at high resolution (hereafter nonhydrostatic resolution) of the atmosphere and ocean, high-resolution models offer the prospect of permitting—rather than parameterizing—convection, potentially improving the representation of phenomena such as the structural evolution of mesoscale convective systems [Skamarock et al., 1994], cold air pooling in mountainous terrain [McNider and Pielke, 1984], and coastal winds and the land-sea breeze [Lebassi et al., 2009]. Given the importance of these phenomena for natural and human systems, high-resolution nonhydrostatic climate model experiments can increase understanding of the processes that are likely to shape climate change impacts, thereby benefiting decision makers in a range of domains, including energy, agriculture, hydrology, air quality, and health.

[4] Nonhydrostatic atmospheric models use nonhydrostatic model formulations where hydrostatic balance in a concurrent hydrostatic model is replaced by a vertical momentum equation [Janjic et al., 2010], which creates additional terms in the nonhydrostatic formulation. Nonhydrostatic treatments are important especially in a complex topography and thermal heating of unstable convective boundary layers [Molemaker and Dijkstra, 2000] and cloud dynamics [Cotton and Anthes, 1989]. Previous research has conducted decomposition of nonhydrostatic versus hydrostatic [Song et al., 1985; Seman, 1994] and linear versus nonlinear terms [Dalu et al., 2003; Weidman and Pielke, 1983] of atmospheric model equations in order to study the importance of the contribution of the nonhydrostatic and hydrostatic terms in an atmospheric model. All but two of the 4 km grid-spacing simulations conducted in this study are nonhydrostatic.

[5] Janjic et al. [2001] compared hydrostatic and nonhydrostatic dynamical formulations and showed that at high-resolution nonhydrostatic model generally produced more robust and smooth solutions than their hydrostatic counterparts. At 8 km grid spacing even though the difference in the dynamics appeared to weak, a visible change was observed in the orographic precipitation.

[6] The “next frontier” in high-resolution climate modeling is to conduct global simulations at nonhydrostatic scales (less than 8 km). However, although proof-of-concept simulations are beginning to appear in the literature [Qian et al., 1998; Tomita et al., 2005; Satoh et al., 2008; Putman and Suarez, 2011], computational resources remain a limiting factor in conducting global nonhydrostatic climate change simulations. At least three primary alternatives to global nonhydrostatic simulations have emerged: (1) nested modeling, in which a limited-area nonhydrostatic model is nested within a global hydrostatic model [Trapp et al., 2011; Lebassi-Habtezion et al., 2011; Mahoney et al., 2012], (2) variable-resolution modeling, in which a single global model is run at nonhydrostatic resolution over a limited area and hydrostatic resolution over the remainder of the globe [Skamarock et al., 2012], and (3) “superparameterization,” in which a component (such as a cloud resolving model) is run at nonhydrostatic resolution at each grid point within the global hydrostatic model [Khairoutdinov et al., 2005].

[7] In this study, we explore prospects for nesting a nonhydrostatic model over a large regional domain, using the western United States (WUSA) as a case study. Although nonhydrostatic limited-area models are widely used for operational short-term weather forecasting and hydrostatic limited-area models are widely used for “regional climate downscaling,” nonhydrostatic simulations that cover large regional areas and are integrated over seasonal timescales remain rare in the literature [Leung et al., 2003; Trapp et al., 2011; Zhang et al., 2009; Lebassi-Habtezion et al., 2011; Comarazamy and González, 2011; Zhao et al., 2011; Mahoney et al., 2012; Oglesby et al., 2010; Lucas-Picher et al., 2012]. Further, although high-resolution models show improvements over their low-resolution counterparts [e.g., Räisänen et al., 2004; Bell et al., 2004; Diffenbaugh and Ashfaq, 2010], high resolution does not guarantee accurate simulation [Boyle, 1993; Senior, 1995; Walker and Diffenbaugh, 2009], and previous work highlights the importance of careful evaluation of the nested climate model configuration prior to conducting climate change simulations [Giorgi and Marinucci, 1996; Räisänen et al., 2004; Prömmel et al., 2010].

[8] Our goal here is not forecasting the regional climate but rather it is to explore the potential for nonhydrostatic nesting over a large region that exhibits complex climate dynamics for developing and validating baseline simulations for future climate change experiments. As stated above, even though the “next frontier” in high-resolution climate modeling is to conduct global simulations at high-resolution nonhydrostatic scales, nested nonhydrostatic approaches have less computational needs and will be vital in understanding local-scale climate phenomenon in climate change studies. In the present study, we use the so-called “perfect boundary condition” approach [e.g., Mearns et al., 2012], i.e., using reanalysis boundary conditions to drive our regional model in order to develop baseline simulations with which to evaluate the model simulation of the present climate. This perfect boundary condition approach is motivated from the perspective of developing the capability to eventually generate present and future decadal-scale simulations of impacts-relevant climate phenomena such as extreme temperature, heavy precipitation, drought, snow accumulation, and coastal winds.

[9] The WUSA's large population and economy [Leung et al., 2004], important agricultural industry [White et al., 2006], limited and snow-dependent water resources [Pavelsky et al., 2011], and intense coastal development [California Energy Commission Report, 2006] all create prominent concerns about climate change impacts. In addition, the complex, multiscale processes shaping the summer climate dynamics of the region present an interesting case study for nonhydrostatic climate model evaluation. For example, the North American monsoon (NAM) influences the summer climate over much of the WUSA and northwestern Mexico. The NAM is a result of intense summer heating that creates zonal movement of the subtropical ridge starting in early July, allowing low-level moisture transport from the eastern tropical Pacific, Gulf of California, and Gulf of Mexico [Adams and Comrie, 1997]. In addition, the Gulf of Mexico may also contribute to upper-level moisture advection into the region [Anderson et al., 2004]. This moisture that is transported northward into the WUSA is lifted by the major topographic features of the region (e.g., Sierra Madre Occidental), resulting in summer precipitation over many areas.

[10] The local-scale summer climate of the WUSA is also influenced by the orographic and topographic complexity that characterizes the region. The coastal mountain ranges effectively limit the moist marine air penetration from the Pacific Ocean, creating intense rain shadows, while the interior mountains are major accent areas for local convection [Adams and Comrie, 1997]. The steep ridges of the mountains also limit ventilation of the adjacent valleys, creating favorable conditions for cold air pools and inversions [Daly et al., 2010]. Although high-resolution hydrostatic modeling improves the climate simulation over complex orography and topography [e.g., Leung and Qian, 2003; Bell et al., 2004; Gao et al., 2006; Diffenbaugh and Ashfaq, 2010], nonhydrostatic resolutions are necessary to explicitly resolve the associated local-scale processes [Case et al., 2002; Colle et al., 2003; Cros et al., 2004; Wang et al., 2004].

[11] Further, on the west coast of the WUSA, the Pacific High, the coastal ocean, and the continental thermal low produce strong pressure, temperature, and moisture gradients between the ocean and the land. These gradients help to produce the along-shore winds that drive the California Current, which results in upwelling of cold, nutrient-rich water to the surface of the coastal ocean. These gradients also produce the daytime, on-shore, coastal marine airflow that transports cool, moist air over coastal land areas during the summer season. Subsidence from the Pacific High also produces an elevated inversion layer that caps the shallow (<1 km deep) marine boundary layer (MBL). The inversion is strongest (up to 20°C) through a layer of 250 m and its base is lowest (shallow) just off the coast, where upwelling water results in a cooling of the MBL [Seaman et al., 1995].

[12] The complex climate dynamics of the WUSA provide important controls on the response of impacts-relevant climate phenomena to global warming [e.g., Diffenbaugh et al., 2005, 2006; White et al., 2006; Rauscher et al., 2008]. However, while a number of high-resolution climate modeling studies have identified the importance of fine-scale climate processes for climate change in the WUSA, those studies have not applied the nonhydrostatic resolution that is likely necessary to capture the local-scale processes that will ultimately determine many climate change impacts. We therefore seek to begin closing the gap between long-term, large-domain hydrostatic simulations [e.g., Diffenbaugh et al., 2005, 2006; White et al., 2006; Rauscher et al., 2008] and short-term, small-domain, nonhydrostatic simulations [Lebassi-Habtezion et al., 2011; Zhao et al., 2011, Mahoney et al., 2012] by simulating a full summer season over the WUSA using hydrostatic and nonhydrostatic resolutions, single and multiple nesting approaches, and low- and high-resolution reanalysis boundary conditions.

2 Methods

2.1 Model Simulations

[13] We employ the Weather Research and Forecasting (WRF; version 3.2.1) atmospheric model [Skamarock et al., 2008] to simulate the climate of the WUSA at nonhydrostatic and hydrostatic resolutions. The numerics in WRF use full compressible equations with contributions divided into hydrostatic and nonhydrostatic formulations, which facilitates comparisons of hydrostatic and nonhydrostatic solutions by selecting the desired option (nonhydrostatic or hydrostatic) in the namelist file. Then WRF solves the Reynolds-averaged primitive equations on different projections [Skamarock et al., 2008]. It uses terrain-influenced sigma coordinates and an Arakawa-C staggered grid on which thermodynamic and moisture variables are defined at grid-volume centers, with velocity components on grid-face centers perpendicular to each component [Mesinger and Arakawa, 1976].

[14] We generate seven WRF simulations using various combinations of horizontal grid spacing, grid nests, and large-scale boundary conditions. Each of the seven WRF simulations are initialized on 28 April 2010, and run without reinitialization through 1 September 2010.

[15] The simulation matrix for this experiment is summarized in Table 1:

  1. [16] WRF forced by the National Centers for Environmental Prediction (NCEP) (WRF.4.1.NCEP) uses one grid covering the WUSA at 4 km (nonhydrostatic) horizontal grid spacing, and initial and boundary conditions (ICBCs) from NCEP global reanalysis with 2.5° × 2.5° grid spacing (NCEP Reanalysis data provided by the National Oceanic Atmospheric Administration (NOAA), Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/).

  2. [17] WRF forced by North American Regional Reanalysis (NARR) (WRF.4.1.NARR) uses one grid covering the WUSA at 4 km (nonhydrostatic) horizontal grid spacing (Figure 1), and ICBCs from the NARR regional reanalysis with 32 km grid spacing (NARR data provided by the NOAA, Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/).

  3. [18] WRF.4.1C.NARR uses one grid covering the WUSA at 4 km (but hydrostatic with cumulus parameterization turned on) horizontal grid spacing, and NARR.

  4. [19] WRF.4.1D.NARR uses one grid covering the WUSA at 4 km (but nonhydrostatic with cumulus parameterization turned on) horizontal grid spacing, and NARR.

  5. [20] WRF.25.2.NCEP uses an outer grid with 100 km (hydrostatic) horizontal grid spacing and an inner grid covering the WUSA at 25 km (hydrostatic) horizontal grid spacing (Figure 1), and ICBCs from NCEP.

  6. [21] WRF.50.2.NCEP uses an outer grid with 150 km (hydrostatic) horizontal grid spacing and an inner grid covering the WUSA at 50 km (hydrostatic) horizontal grid spacing (Figure 1), and ICBCs from the NCEP global reanalysis.

  7. [22] WRF.4.3.NCEP uses three nested grids, including outer grids of 100 km (hydrostatic) and 20 km (hydrostatic) horizontal grid spacing and an inner grid covering the WUSA at 4 km (nonhydrostatic) horizontal grid spacing (Figure 1), and ICBCs from the NCEP global reanalysis.

Table 1. Matrix of the Study
No.Model NameModel FormulationCumulus ParameterizationBCGridsLevelsΔx, Δy (km)
1WRF.4.1.NCEPNonhydrostaticOffNCEP1504, 4
2WRF.4.1.NARRNonhydrostaticOffNARR1504, 4
3WRF.4.1C.NARRHydrostaticOnNARR1504, 4
4WRF.4.1D.NARRNonhydrostaticOnNARR1504, 4
5WRF.4.3.NCEPNonhydrostatic NCEP3504, 4
6WRF.25.2.NCEPHydrostaticOnNCEP25025, 25
7WRF.50.2.NCEPHydrostaticOnNCEP25050, 50
8NARRHydrostaticOnNCEP 2932, 32
9PRISM  Obs.  4, 4
10NCEPHydrostaticOn  29250, 250
Figure 1.

(Top left) WRF.4.1.NARR, WRF.4.1.NCEP, and WRF.4.1C.NARR domains with topography of the region (color) and METAR stations (black dots) used for evaluation; (top right) WRF.4.3.NARR domains; (bottom left) WRF.25.2.NCEP domains; and (bottom right) WRF.50.2.NCEP domains.

[23] Together, these six simulations allow us to test differences between nonhydrostatic and hydrostatic resolutions, single and multiple nesting approaches, and high- and low-resolution reanalysis boundary conditions.

2.2 Model Configuration

[24] We employ WRF grids with horizontal grid spacing of 4, 20, 25, 50, 100, and 150 km (Figure 1). We use the WRF Single-Moment 3-class (WSM3) microphysics scheme of Hong et al. [2004]—which is a simple efficient scheme with ice and snow processes suitable for mesoscale grids sizes—for the grids with 4 km horizontal grid spacing. We use the cumulus parameterization of Kain and Fritsch [Kain, 2004] for the grids with horizontal grid spacing of 20, 25, 50,100, and 150 km. Kain-Fritsch is a deep and shallow convection subgrid scheme, which uses mass flux approach and with moist updrafts and downdrafts that include the effects of detrainment and entrainment.

[25] We use the Noah Land Surface Model (Noah LSM) [Pan and Mahrt, 1987; Chen et al., 1997; Chen and Dudhia, 2001; Ek et al., 2003] in all of the WRF simulations. The Noah LSM is a unified NCEP/NCAR/AFWA land surface model with four soil layers and canopy moisture and snow cover prediction [Chen and Dudhia, 2001]. We use the Mellor-Yamada-Nakanishi-Niino (MYNN) level 2.5 [Nakanishi and Niino, 2004] planetary boundary layer (PBL) scheme, which is a Turbulent Kinetic Energy (TKE)-based local mixing scheme. We also use the NCAR Community Atmospheric Model (CAM) radiation package [Collins et al., 2004], which considers water vapor, cloud fraction, and aerosol and tracer gas concentrations. The radiation package calculates atmospheric heating due to radiative flux divergence and surface radiations for the ground heat budget.

2.3 Boundary Conditions

[26] The simulations use variable-field model initialization and update processes, in which gridded three-dimensional temperature, horizontal wind, relative humidity, geopotential height, and soil temperature fields and two-dimensional surface pressure and sea-level pressure fields are provided at the lateral boundaries of WRF from the NCEP or NARR reanalysis product. The boundary conditions data are assimilated at 6 h (NCEP) or 3 h (NARR) intervals using Newtonian relaxation. WRF has one-way and two-way nesting capability and various options for physics, initialization, and postprocessing produce an end-to-end mesoscale simulation. For simulations that use multiple nested grids, we selected two-way interactive nesting, in which the inner higher-resolution grid updates the outer lower-resolution grid. This option keeps the solutions in the coarse and nested grids in phase, allowing well-resolved disturbances to exit the nested domains with only minor reflections [Harris and Durran, 2010].

[27] Regional Climate Model solutions can also be substantially affected by the domain size and placement. To minimize this effect, the domain sizes of our regional model simulations were set to be large enough [Jones et al., 1995], and lateral boundary conditions were also set far enough from the area of interest [Seth and Giorgi, 1998], in order to prevent feedbacks from the inner part of the domain.

[28] In addition to the atmospheric ICBC, the bottom boundary conditions of topography, land cover, and sea surface temperature (SST) are also needed for model initialization and integration. We use the 30 arc-sec (approximately 1 km) US Geological Survey (USGS) topographic heights [GTOPO30, 1996] and 30 arc-sec USGS Global Land Cover Characteristics (GLCC) [Loveland et al., 2000] land use data, which were developed from 1 km Advanced Very High Resolution Radiometer (AVHRR) satellite images obtained in 1992–1993. We also use the daily Real-Time Global Sea Surface Temperature (RTG_SST) SST data set, which is analyzed from in situ and remote infrared measurements using 1° AVHRR [Gemmill et al., 2007]. SST data are obtained from NCEP at a grid spacing of 0.5° for all simulations.

2.4 Comparison With Observational Data

[29] We compare the spatial distribution of WRF-simulated 2 m-surface maximum temperature and precipitation fields for the months of June, July, and August 2010 with the respective observational gridded fields from the Parameter-elevation Regressions on Independent Slopes Model (PRISM), station-based gridded data set. These comparisons show the simulated spatial features of the 2 m-surface maximum temperature and precipitation fields, along with the biases in each simulation as compared to the PRISM gridded observational data.

[30] We also added a high-resolution hydrostatic model (WRF.4.1C.NARR) to compare nonhydrostatic versus hydrostatic dynamical options at 4 km grid spacing. This simulation provides a case study illustrating the sensitivity to model formulation. It should be noted that this simulation could not be executed with exactly the same model options (cumulus parameterization was on because turning microphysics on made the model unstable even at small time steps) as the nonhydrostatic simulations, which was run with cumulus parameterization turned off. However, we added another nonhydrostatic simulation (WRF.4.1D.NARR) with cumulus parameterization turned on so that we have exactly the same configurations as WRF.4.1C.NARR except for the cumulus parameterization.

[31] To test a hypothesis that adiabatic temperature adjustment of the low-resolution runs (with grid spacing of 25 and 50 km) will produce comparable or better results as compared to the high-resolution runs, we adiabatically adjusted (9.8°C km−1) the temperature of the July low-resolution model runs (with grid spacing of 25 and 50 km) to include high-resolution topographic effects. If this hypothesis is true, running low-resolution model is computationally inexpensive and making those runs and adiabatic adjustment for topography would save a lot of computational time. However, there is also a challenge of how to correct other fields (e.g., precipitation) using such adjustments.

[32] We also compare the WRF-simulated July 2010 hourly surface 2 m maximum temperature, relative humidity, and 10 m wind speeds with observed values from 292 Meteorological Terminal Aviation Routine (METAR) weather stations (Figure 1). We first interpolate the WRF temperature and relative humidity values to the 2 m level and the wind speeds to the 10 m level, at the WRF grid point closest to each site. We then calculate the mean value of each hour of the day at each station for the month of July and then calculate the hourly mean across all of the stations in each State in the WUSA, creating a statewide mean daily cycle for July 2010 only. These comparisons show the simulated climatological daily cycle in different areas of the region, along with the biases in each simulation as compared to the METAR station data.

[33] Finally, we compare the WRF-simulated July 2010 atmospheric fields of mixing ratio, moisture fluxes, and vertically integrated moisture flux with the NCEP global reanalysis and NARR regional reanalysis. These comparisons show the simulated atmospheric circulation, along with the biases in each simulation as compared to the reanalysis data.

3 Results and Discussion

3.1 High-Resolution Features of Simulated and PRISM Temperatures and Precipitation

[34] To illustrate the spatial details in the nonhydrostatic WRF simulations, first we focus on the comparison of July 2010 mean daily maximum temperature and total monthly precipitation between one of the 4 km WRF configuration (WRF.4.1.NARR) and the 4 km observational PRISM data (Figure 2). The nonhydrostatic WRF simulation captures the spatial details of daily maximum temperature reflected in the PRISM data set, including the complex patterns of high and low maximum temperatures associated with the complex topography and orography of the region. For example, in addition to the broad patterns associated with the high elevations of the Rocky Mountains, Sierra Nevada, and Cascade Mountains and the low elevations of the Central Valley, Columbia Valley, and Snake River Valley, the WRF.4.1.NARR simulation also captures the localized temperature variations in the narrow mountain ranges and valleys such as the Grand Canyon, Death Valley, and the Basin and Range of Nevada. The model also captures the cool marine influences over the coastal zones (e.g., San Francisco Bay, Monterey Bay, Los Angeles Basin, etc.), which create substantially cooler daily maximum temperatures over the coast than over the adjacent inland areas.

Figure 2.

(Top) July 2010 total precipitation and (bottom) mean-monthly daily maximum temperature from the 4 km WRF.4.1.NARR simulation and the 4 km gridded observational PRISM data set. Precipitation is in mm, where red is relatively dry and blue is relatively wet. Temperature is in °C, where red is relatively warm and blue is relatively cool.

[35] The most prominent bias in the WRF.4.1.NARR simulation of July mean daily maximum temperature is a tendency towards excessively high temperatures in low elevation areas (Figure 2). For example, the highest daily maximum temperatures over the Central Valley are approximately 5°C greater in the WRF.4.1.NARR simulation than in the PRISM data set. Warm biases of similar magnitude also occur over southeastern California, southern Nevada, and southwestern Arizona.

[36] The WRF.4.1.NARR simulation also captures the spatial details of monthly precipitation reflected in the PRISM data set, with the highest monthly precipitation (up to 400 mm) occurring over the southeast of the domain, widespread areas with monthly precipitation of greater than 25 mm occurring over the eastern half of the domain, and areas with monthly precipitation of less than 25 mm occurring over most of the western half of the domain (Figure 2). Although the effect of topography is less prominent for July monthly precipitation than July daily maximum temperature, a number of topographic effects that are seen in the PRISM data set are also seen in the WRF.4.1.NARR simulation. For example, along the Pacific coast, areas of relatively high precipitation are seen over the high elevations of the Cascade Mountains and Sierra Nevada in both the WRF.4.1.NARR simulation and PRISM data set. Similar areas of relatively high precipitation are also seen over the high elevations of Colorado, Utah, Idaho, and Montana.

[37] The most prominent biases in the WRF.4.1.NARR simulation of July monthly precipitation are lower-than-observed values over areas of Arizona, Utah, and Colorado. Many of these dry biases occur in areas where southerly advection associated with the NAM transports moisture over topographically complex terrain. The PRISM data set—which is generated by interpolating between relatively sparse station data [Daly et al., 2002] —tends to exhibit uniformly high precipitation values in these areas, while the WRF.4.1.NARR simulation tends to exhibit high precipitation values over the areas of peak topography, with lower precipitation values over lower elevation areas.

3.2 Effect of Grid Spacing, Model Formulation, and Boundary Conditions on Simulated Temperature

[38] In order to better understand the high-resolution WRF simulations, we compare our six simulations of June, July, and August 2010 with the statistically interpolated PRISM data set. In general, both the high and low-resolution WRF simulations resolve the pattern of monthly-mean daily maximum temperatures over the WUSA (Figures 3, 4, and 5). However, the high-resolution simulations (WRF.4.1.NARR, WRF.4.1D.NARR, WRF.4.1.NCEP, and WRF.4.3.NCEP) exhibit considerably greater spatial detail—and greater agreement with PRISM—throughout the domain, while WRF.4.1C.NARR also exhibits spatial detail over most part of the domain but larger warm biases over the high elevation areas. The improved agreement of the high-resolution models with PRISM is particularly clear over the areas of greatest topographic complexity. In addition, compared with the low-resolution simulations, the high-resolution simulations also more accurately reflect the cool coastal temperatures that are clearly seen in the PRISM data set. This improved representation of the coastal temperatures is hypothesized to be attributable to the resolution of the SSTs just offshore, as well as with the improved resolution of the topography and coastline in the high-resolution WRF grid. This improved resolution of the coastal boundary conditions and of the atmospheric circulation also likely improves the simulation of penetration of marine air through the coastal gaps.

Figure 3.

June average monthly surface 2 m temperature for: (i) PRISM gridded observations, (ii) NARR reanalysis, (iii) the WRF.4.1.NCEP simulation, (iv) the WRF.4.1.NARR simulation, (v) the WRF.4.3.NCEP simulation, the WRF.25.2.NCEP simulation, and (vi) the WRF.50.2.NCEP simulation. Temperature is in °C, where red is relatively warm and blue is relatively cool.

Figure 4.

Same as Figure 3 but for July.

Figure 5.

Same as Figure 3 but for August.

[39] As with the WRF.4.1.NARR simulation of July 2010 (Figure 2), all of the simulations show warm biases (1–7°C) over the low elevation areas (Figures 3, 4, and 5). These warm biases are generally more pronounced in the nonhydrostatic simulations than in the hydrostatic simulations (e.g., over the Central Valley, Columbia Valley, and Snake River Valley). Low elevations could be expected to be cooler in the lower-resolution simulations as the grid point elevations are generally higher at lower resolution than at higher resolution. In addition, irrigation has been shown to have a local cooling effect over areas of the WUSA [Sorooshian et al., 2011; Kueppers et al., 2007]. For example, a recent study by Sorooshian et al. [2011] using a 4 km inner domain of the MM5 model found temperature decreases of approximately 3–7°C and humidity increases of approximately 9–20% in all the irrigated grid cells over the Central Valley. These temperature decreases are of similar magnitude of the warm biases in our results, suggesting that the warm biases over the irrigated valleys in our results (Figures 3, 4, and 5) could be caused by the model parameterization of the irrigated land surface. However, the four nonhydrostatic simulations do not all exhibit the same biases over the irrigated valleys, with the triple-nested WRF.4.3.NCEP simulation generally exhibiting less severe warm biases than the single-nestedWRF.4.1.NARR, WRF.4.1D.NARR, and WRF.4.1.NCEP simulations. Indeed, the triple-nested WRF.4.1.NCEP simulation appears to be systematically cooler across the domain, with daily maximum temperatures that are cooler than PRISM over many high-elevation areas (Figures 3, 4, and 5).

[40] To test the hypothesis that adiabatic temperature adjustment of the 25 and 50 km grid spacing runs will produce comparable or better results as compared to the high-resolution simulations, we adiabatically adjusted (9.8°C km−1) the temperature of the July low-resolution model runs (grid spacing of 25 and 50 km) to include topographic effects (Figure 6). Running a low-resolution model is computationally inexpensive, meaning that using adiabatic adjustment for topography would save substantial computational time. However, results show very large temperature biases in most of the valleys and mountain areas of the entire domain. Even though the two-dimensional temperature features are better resolved, and topographic effects are present, the accuracy of the temperature fields is highly biased. The other downside of this method is that there is no way of correcting the precipitation or other fields using the adiabatic adjustment.

Figure 6.

July average monthly surface 2 m temperature with adiabatic adjustment (top right) WRF.50.2.NCEP and (top left) WRF.25.2.NCEP as compared to (bottom) PRISM gridded data set.

3.3 Hourly Climatology

[41] Given the effect of grid spacing on the WRF simulation of the monthly mean daily maximum 2 m temperature (Figures 3, 4, and 5), we compare the model time series of hourly average surface temperatures, relative humidity, wind speeds, and wind directions from the WRF simulations that are nested in the global NCEP reanalysis at 4, 25, and 50 km horizontal grid spacing (Figures 7-9) with the time series from the 292 observed METAR station data sets (Figure 1). Figure 7 shows the monthly average hourly surface temperature data for the month of July 2010 in the WRF.4.1.NCEP, WRF.25.2.NCEP, and WRF.50.2.NCEP simulations. The model generally agrees with observations at night for all cases, with the average WUSA bias mostly limited to 1.75°C or less during the nighttime hours. However, there is a daytime warm bias in WRF.4.1.NCEP over California (up to 5.77°C), Colorado (up to 4.79°C), New Mexico (up to 6.79°C), Arizona (up to 4.84°C), and Utah (up to 2.75°C). WRF.4.1.NCEP shows less agreement with observations than the lower-resolution counterparts for the southern states of Arizona, Colorado, New Mexico, and Utah, as well as California, with the WRF.4.1.NCEP simulation showing biases of up to 2.65°C and the WRF.25.2.NCEP and WRF.50.2.NCEP simulations showing biases of 0.55°C and 1.66°C, respectively.

Figure 7.

Observed and simulated hourly monthly average July 2010 temperature (°C). The simulations are from the WRF.4.1.NCEP, WRF.25.2.NCEP, and WRF50.2.NCEP simulations. The observations are from the METAR stations. Values are calculated at the WRF grid point nearest to each METAR station, and then averaged over each of the 11 Western US states.

Figure 8.

Same as Figure 7, but for humidity at 2 m.

Figure 9.

Same as Figure 7, but for wind speed at 10 m.

[42] In conjunction with this warm bias, the WRF model also shows a consistent dry bias in relative humidity of up to 33% over the southern states (Figure 8). In contrast, simulated relative humidity over Washington, Wyoming, Oregon, Idaho, Montana, and Nevada shows greater agreement with the station observations, with biases of up to −11.83%, 4.35%, −13.89%, −6.48%, 4.67%, and 6.11%, respectively (Figure 8). The hydrostatic WRF.25.2.NCEP and WRF.50.2.NCEP simulations also exhibit positive wind speed biases of up to 2 m s−1 over the WUSA (Figure 9), while the nonhydrostatic WRF.4.1.NCEP simulation exhibits high agreement with the observed values for most periods of the diurnal cycle, with biases mostly limited to 0.25 m s−1 or less. This discrepancy in wind speed between the hydrostatic and nonhydrostatic configurations could be associated with differences in the grid point elevations in different areas of the topographically complex WUSA, thereby more accurately representing roughness and topographic effects in the model. Differences between the simulated and observed wind vectors also suggest that the higher-resolution WRF.4.1.NCEP best captures the summer time surface winds (not shown).

3.4 Effect of Grid Spacing, Model Formulation, and Boundary Conditions on Simulated Precipitation

[43] As with surface air temperature, we compare the precipitation from our six simulations of June, July, and August 2010 with the statistically interpolated PRISM data set (Figures 10, 11, and 12). All seven simulations capture the basic structure of monthly precipitation seen in the PRISM data set, with dry conditions in the southwest of the domain and wet conditions in the east of the domain in June, July, and August. However, the nonhydrostatic simulations much more closely resolve the spatial heterogeneity of summer precipitation associated with the complex topography of the WUSA. These topographic effects can be most clearly seen over the Cascade Mountains and Sierra Nevada in June 2010, where the nonhydrostatic WRF.4.1.NARR, WRF.4.1.NCEP, WRF.4.1D.NARR, WRF.4.3.NCEP, and the hydrostatic WRF.4.1C.NARR simulations resolve the observed precipitation heterogeneity, but the hydrostatic WRF.25.2.NCEP and WRF.50.2.NCEP simulations do not.

Figure 10.

June total monthly surface precipitation for: (i) PRISM gridded observations, (ii) NARR reanalysis, (iii) the WRF.4.1.NCEP simulation, (iv) the WRF.4.1.NARR simulation, (v) the WRF.4.3.NCEP simulation, the WRF.25.2.NCEP simulation, and (vi) the WRF.50.2.NCEP simulation. Temperature is in °C, where red is relatively dry and blue is relatively warm.

Figure 11.

Same as Figure 10 but for July.

Figure 12.

Same as Figure 10 but for August.

[44] The WRF.4.1.NARR exhibits the closest agreement with the PRISM data set over the WUSA domain in July and August 2010, with the nonhydrostatic WRF.4.1.NCEP and WRF.4.3.NCEP simulations exhibiting dry biases of up to 200 mm over the southeastern part of the domain, the hydrostatic WRF.25.2.NCEP and WRF.50.2.NCEP simulations exhibiting wet biases of up to 200 mm over the southeastern part of the domain and the hydrostatic WRF.4.1C.NARR showed extremely wet biases of up to 400 mm over the high elevation areas of the north, west, and south west US, while nonhydrostatic counterpart (WRF.4.1D.NARR) showed dry bias over the same regions but is closer to the WRF.4.1.NARR. In addition, the clearest discrepancy with the PRISM data set is in the WRF.4.1.NCEP simulation. The contrast in precipitation over the southern states between the WRF.4.1.NARR and WRF.4.1.NCEP simulations in July and August could be caused by the fact that the difference in grid spacing between the reanalysis boundary conditions and the WRF grid is much greater for the NCEP global reanalysis (2.5 × 2.5° grid spacing) than for the NARR regional reanalysis (32 km). However, the WRF.4.3.NCEP simulation also exhibits dry biases over the southern states, suggesting that the dry bias in the WRF.4.1.NCEP simulation is not caused by the grid-spacing difference in nesting the 4 km WRF grid directly in the global NCEP grid. Further, both of the hydrostatic WRF simulations —which are nested in the NCEP global reanalysis—exhibit precipitation peaks over the southern states, suggesting that the lack of precipitation in the WRF.4.1.NCEP and WRF.4.3.NCEP simulations is not caused by deficiencies in the NCEP global reanalysis that would prevent conditions necessary for precipitation.

[45] Since the source of moisture in the southern states dry bias region is associated with the NAM, we compare the simulation of the tropospheric dynamics associated with the NAM between the five WRF configurations in order to understand the discrepancies in simulated precipitation over the southern states. Figure 13 shows July 2010 500 hPa mixing ratio and wind vectors for the NARR and the WRF model simulations. The NARR shows a well-defined northeast-southwest (NE-SW) trough (an elongated upper-level region of relatively low atmospheric pressure), with peak mixing ratio (up to 30 × 10−4 kg kg−1) localized over the southern states and northern Mexico. The nonhydrostatic WRF.4.1.NARR simulation exhibits very similar features at 500 hPa, with peak mixing ratio exceeding 32 × 10−4 kg kg−1 over the southern states. The 500 hPa circulation and mixing ratio features in the hydrostatic WRF.25.2.NCEP and WRF.50.2.NCEP simulations are also similar to NARR, with the WRF.25.2.NCEP simulation exhibiting the most widespread areas of mixing ratio above 20 × 10−4 kg kg−1. However, the NE-S oriented trough is dissipated in the WRF.4.1.NCEP and WRF.4.3.NCEP simulations, with the center of the subtropical ridge displaced southward, leading to more westerly onshore flow and reduced mixing ratio.

Figure 13.

Mean July 2010 500 hPa wind vectors and mixing ratio [shaded, (kg kg−1)] for the NARR reanalysis, the WRF.4.1.NARR simulation, the WRF.4.1.NCEP simulation, the WRF.4.3.NCEP simulation, the WRF.25.2.NCEP simulation, and the WRF.50.2.NCEP simulation. The reference vector is 10 m s−1. Mixing ratio is in (kg kg−1), where blue is relatively high mixing ratio and brown is relatively low mixing ratio.

[46] Previous research suggests that the dominant moisture source for the NAM moisture advection comes from combination of the eastern tropical Pacific, Gulf of California, and the Gulf of Mexico [Harrington et al., 1992; Schmitz and Mullen, 1996]. The discrepancies in July precipitation in the WRF simulations (Figures 10, 11, and 12) can therefore be explained in part by the 500 hPa circulation and mixing ratio, with the WRF.4.1.NARR simulation exhibiting very similar features to the NARR regional reanalysis, and areas of dry (wet) bias in the other simulations exhibiting reduced (enhanced) mixing ratio and/or reduced convergent flow. Analysis of the July vertically integrated horizontal moisture fluxes from 650 hPa to the top of the atmosphere (650 hPa-TOA) supports this assessment (Figure 14). This layer-integrated horizontal moisture flux is balanced by the vertical moisture flux into the layer and precipitation from the bottom. The WRF.4.1.NARR simulation again shows the greatest agreement with the NARR regional reanalysis, with midlevel moisture flux vectors indicating onshore flow over the Gulf of California and subtropical Pacific (with some indication of onshore flow over Gulf of Mexico in the southeast corner of the domain). However, the WRF.4.1.NCEP and WRF.4.3.NCEP simulations exhibit more westerly midlevel moisture flux vectors over the southern half of the domain, particularly the WRF.4.1.NCEP simulation, which exhibits the most severe dry bias over the southern states (Figures 10, 11, and 12). The July layer-integrated horizontal moisture flux likewise indicates greater onshore flux over the Gulf of California and Gulf of Mexico in the WRF.25.2.NCEP simulation than in the WRF.50.2.NCEP simulation, helping to explain the greater July precipitation values over New Mexico and Colorado in the WRF.25.2.NCEP simulation (Figures 10, 11, and 12).

Figure 14.

As in Figure 13, but for integrated 650 hPa-TOA moisture flux vector and its magnitude. The reference vector is 25 kg (m s−1). Moisture flux magnitude is in kg (m s−1), where green is relatively high moisture flux and brown is relatively low moisture flux.

[47] The simulated dynamics of the NAM can therefore help to explain the differences in precipitation exhibited by the WRF simulations. The NAM sources low-level moisture from the Gulf of California and upper-level moisture from the Gulf of Mexico. The moisture transported inland from the Gulf of California up to the western side of the Sierra Madre Occidental due to topographic heating creates convection and redistributes the moisture aloft, which finally could mix with the upper-level Gulf of Mexico moisture from the eastern side of the Sierra Madres Occidental. The moisture is subsequently transported into the southwestern United States via flow around the subtropical ridge. The moisture flux sources and sinks for the simulated NAM therefore depend on resolution of topography, coastline, and the mesoscale atmospheric circulation. Resolving the Gulf of California and other key topographic features for local convective activities shows greater improvement in the nonhydrostatic simulations than the hydrostatic simulations. However, the simulation of the atmospheric circulation and moisture is sensitive to the combination of reanalysis boundary conditions and grid spacing of the nested simulations. The comparison of WRF.4.1.NARR (nonhydrostatic with cumulus parameterization off) with WRF.4.1C.NARR (hydrostatic with cumulus parameterization on) and WRF.4.1D.NARR (nonhydrostatic with cumulus parameterization on) suggests that the differences between WRF.4.1.NARR and WRF.4.1C.NARR are most likely due to the dynamical core rather than the cumulus parameterization, and that the nonhydrostatic treatment without cumulus parameterization results in a closer match to the PRISM summer precipitation than either nonhydrostatic or hydrostatic with cumulus parameterization. In particular, as with July and August precipitation, the nonhydrostatic realization nested in the NARR regional reanalysis (WRF.4.1.NARR) exhibits the closest agreement with observational data, and the hydrostatic WRF.25.2.NCEP and WRF.50.2.NCEP simulations exhibit greater agreement than the nonhydrostatic WRF.4.1D.NARR, WRF.4.1.NCEP, WRF.4.3.NCEP, and the hydrostatic WRF.4.1C.NARR simulations.

[48] In addition to greater accuracy in orographic representation by the high-resolution nonhydrostatic models, to resolve local-scale phenomenon, the high-resolution input atmospheric BCs could also be hypothesized to have impacts in the higher accuracy of the NARR forced nonhydrostatic simulations. However, the model formulation was also important as our results in WRF.4.1C.NARR showed high precipitation biases with the NARR BCs but different model formulation. Further, the biases in 500 hPa circulation and moisture flux exhibited by the WRF.4.1.NCEP and WRF.4.3.NCEP simulations are consistent with the dry biases in precipitation seen over the southeastern quadrant of the domain in those simulations.

5 Summary and Conclusions

[49] Computational limitations create critical barriers to conducting global nonhydrostatic climate model simulations. Although high-resolution hydrostatic global climate model experiments are becoming feasible [e.g., Taylor et al., 2012], the current generation of global models remains far from resolving the mesoscale features that could determine the response of local-scale climate phenomena to global warming. In particular, impacts on hydrology, air quality, energy supply and demand, human health, and agriculture will likely depend on changes in local-scale temperature, precipitation, humidity, and atmospheric circulation, the simulation of which could be improved by nonhydrostatic resolution of the atmosphere. Although the current generation of global and regional climate models can resolve many key features of the climate system, the processes that shape phenomena such as marine boundary layer bases, sea/land breezes, mountain/valley circulations, super-cell thunderstorms, urban heat islands, and coastal upwelling are not resolved in the current generation of climate models.

[50] The potential importance of those local-scale phenomena motivates development of approaches to enable nonhydrostatic climate model simulations that are less computationally expensive than running the entire globe at nonhydrostatic resolution. In the current study, we test various approaches to nesting a nonhydrostatic grid over a subcontinental region, using the 2010 summer climate of the WUSA as a test case. Comparing nonhydrostatic 4 km grids with hydrostatic 4 km, 25 km, and 50 km grids allows us to test the impact of nonhydrostatic resolution on the simulation of meso- and local-scale climate features. In addition, comparing simulations forced by global and regional reanalysis allows us to test the impact of the grid spacing of the large-scale boundary conditions in which the nonhydrostatic grid is nested. Further, comparing simulations with single- and multiple-nesting approaches allows us to test the impact of differences in grid spacing between the large-scale boundary conditions and the nested grid.

[51] The experiments show that, compared with the hydrostatic simulations, the high-resolution nonhydrostatic simulations more accurately resolve the heterogeneity of surface temperature, precipitation, and wind speed that are associated with the topography and orography of the WUSA region.

[52] High-resolution atmospheric BCs could also be hypothesized to have impacts in the higher accuracy of the NARR forced nonhydrostatic simulations; however, the model formulation was also important as our results in WRF.4.1C.NARR showed high precipitation biases with the same BCs but different model formulation. The comparison of WRF.4.1.NARR (nonhydrostatic with cumulus parameterization off) with WRF.4.1C.NARR (hydrostatic with cumulus parameterization on) and WRF.4.1D.NARR (nonhydrostatic with cumulus parameterization on) suggests that the differences between WRF.4.1.NARR and WRF.4.1C.NARR are most likely due to the dynamical core rather than the cumulus parameterization, and that the nonhydrostatic treatment without cumulus parameterization results in a closer match to the PRISM summer precipitation than either nonhydrostatic or hydrostatic with cumulus parameterization. In addition, starting with grid spacing of 25 or 50 km model output and doing either a statistical or thermodynamic (i.e., adiabatic) downscaling on it did not produce accurate results. To clarify that we corrected the low-resolution model runs (with grid spacing of 25 and 50 km) to include topographic effects using adiabatic temperature adjustment and results showed very large temperature biases in most of the valleys and mountains. Even though the features were better resolved, and topographic effects are present, the accuracy is reduced. Another disadvantage is the need to accurately simulate the response of precipitation, winds, and other fields in addition to temperature in areas of complex terrain.

[53] In addition, the experiments show that the simulation in which the nonhydrostatic grid is nested directly within the regional reanalysis exhibits the greatest overall agreement with observational data. However, the experiments also show that nonhydrostatic resolution does not improve the agreement with observations for all features of the regional climate, with the nonhydrostatic simulations that were nested within the global reanalysis showing severe dry biases over the southern states in June and July 2010, along with biases in the atmospheric circulation and moisture flux. The fact that the simulation with multiple nests shows similar dry biases as the simulation with a single nest suggests that the biases are not caused primarily by the difference in grid spacing between the global reanalysis and the nonhydrostatic grid. Similarly, the fact that the hydrostatic simulations that are nested directly within the global reanalysis do not show the same biases in precipitation, atmospheric circulation, and moisture flux suggests that the dry biases are not due to systematic errors in the global reanalysis.

[54] The dry biases exhibited by two of the three nonhydrostatic simulations therefore highlight important challenges for moving forward with nested nonhydrostatic climate model experiments, as the optimal configuration will likely depend on the nonhydrostatic model and the large-scale boundary conditions in which the nonhydrostatic model is nested. Given the agreement between observational data and our nonhydrostatic simulation nested in the regional reanalysis, our experiments suggest that these issues can potentially be addressed by nesting the nonhydrostatic model within a hydrostatic high-resolution model (analogous to the regional reanalysis). However, we note that previous work over the eastern U.S. suggests that such an intermediary hydrostatic climate model can degrade critical atmospheric features that are passed to the nonhydrostatic model [Trapp et al., 2007].

[55] The improvement of local-scale climate features in all of our high-resolution nonhydrostatic simulations, and the accuracy of the nonhydrostatic simulation nested in regional reanalysis, both indicate that further development of nonhydrostatic nesting approaches is likely to yield important insights into the response of local-scale climate phenomena to increases in global greenhouse gas concentrations. Further development of such nested approaches is therefore an important activity for understanding the potential impacts of climate change, particularly given the large computational relative to global nonhydrostatic simulations. However, the biases in regional precipitation, atmospheric circulation, and moisture flux identified in a subset of the nonhydrostatic simulations suggest that alternative nonhydrostatic modeling approaches such as superparameterization and variable-resolution global nonhydrostatic modeling will provide important complements to the nested approaches tested here.

Acknowledgments

[56] We thank the National Centers for Environmental Prediction (NCEP) for providing access to the NCEP global reanalysis data set and the North American Regional Reanalysis (NARR) data set. We thank the PRISM Climate Group and Oregon State University for providing access to the PRISM observational temperature data set, and the University of Wyoming, Department of Atmospheric Science for providing station data at http://weather.uwyo.edu/. The WRF simulations were performed using computing resources provided by the Center for Computational Earth and Environmental Science (CEES) at Stanford University. The research reported here was supported by NSF award 0955283, NIH award 1R01AI090159-01, and DOE award DE-SC005171.

Ancillary

Advertisement