Impact on modeled cloud characteristics due to simplified treatment of uniform cloud condensation nuclei during NEAQS 2004



[1] Subgrid-scale cloud condensation nuclei (CCN) heterogeneity is not represented in global climate models (GCM) and potentially contributes systematic errors to simulated cloud effects. High-resolution WRF-Chem model simulations were performed to investigate the impact of assuming a uniform CCN distribution on cloud properties and surface radiation over a region the size of a GCM grid column. Results indicate that a prescribed CCN distribution allowing for vertical and temporal fluctuations does substantially better in simulating cloud properties and radiative effects than does a prescribed uniform and constant CCN distribution. Spatially and temporally averaged net effects on downwelling shortwave radiation are between −3 and −11 W m−2 for the fluctuating and uniform distributions, respectively, versus a control simulation with fully interactive aerosols. Both prescribed CCN distributions produce optically thicker clouds more often than the control, with the mean cloud optical depth increasing by over 25% when using the uniform and constant CCN distribution.

1. Introduction

[2] Clouds are one of the most difficult physical phenomena for atmospheric models to reproduce because many of the processes that uniquely define cloud characteristics are smaller than typical model grid spacings. Subgrid factors include small-scale differences in the number and composition of aerosols serving as cloud condensation nuclei (CCN). For example, near power plants where SO2 stack emissions produce narrow sulfate plumes, the cloud albedo [Twomey, 1991] and lifetime [Albrecht, 1989] of along-plume clouds can differ from those in the surrounding region due to aerosol indirect effects, even though all the clouds experience similar meteorological conditions.

[3] Modelers continually balance increased resolution and more detailed representations of atmospheric physicochemical processes with computational costs. Simplified or temporally-/spatially-constant treatments that reduce costs without sacrificing accuracy are increasingly important when refining grid resolution. For example, it would be advantageous to treat aerosols in global climate models (GCMs) on a coarser grid than clouds since the cost of treating aerosol physics within global cloud resolving models [e.g., Miura et al., 2005; Tomita et al., 2005] is prohibitive. Another potential savings is via a simplified handling of CCN. However, to date, the importance of subgrid-scale CCN variability has not been explored.

[4] This paper investigates the effect of CCN heterogeneity on cloud properties and surface radiation by simulating a region approximately the size of a GCM grid column with realistic and idealized CCN size distributions. By simulating regional aerosol properties at high resolution and then using these values to construct two different “ideal” CCN scenarios with horizontally uniform aerosols, we show that including temporal and vertical heterogeneity of CCN (and aerosol) is important. However, subgrid-scale horizontal variations contribute much less error and could be considered adequate for some applications in light of the typical error in simulated clouds.

2. Model Description and Experimental Setup

[5] The chemistry version of the Weather Research and Forecasting model (WRF-Chem) [Grell et al., 2005] as modified by Fast et al. [2006] was used in this study. The modified model employs the CMB-Z gas-phase chemistry mechanism and the MOSAIC aerosol module, with eight size sections. Additional features added to WRF-Chem include parameterizations for aerosol nucleation [Napari et al., 2002], coagulation [Jacobson et al., 1994], activation/resuspension and wet scavenging [Easter et al., 2004; Ghan et al., 2001], and aqueous chemistry [Fahey and Pandis, 2001]. Aerosol activation (and droplet nucleation) is based on a maximum supersaturation determined from a Gaussian spectrum of updraft velocities and the internally mixed aerosol properties within each size bin, similar to the methodology used in the MIRAGE GCM [Ghan et al., 2001]. As droplets evaporate, particles return to the interstitial phase.

[6] The first and second aerosol indirect effects are implemented in WRF-Chem by linking simulated droplet number to the Goddard Space Flight Center Shortwave and the Lin et al. microphysics schemes, respectively [Skamarock et al., 2005]. The Lin scheme now predicts droplet number, and the autoconversion is dependent on droplet number, following Liu et al. [2005]. Aerosol particles acting as CCN are now tightly coupled with the cloud physics portion of the model. This coupling allows for fully interactive feedbacks: aerosols affect cloud droplet number and cloud radiative properties, and clouds alter aerosol size and composition via aqueous processes and wet scavenging.

[7] WRF-Chem was configured with three two-way interactive nested domains, using grid spacings of 18, 6, and 2 km, and 57 vertical levels. Twenty-seven model levels reside within the lowest 2 km to fully resolve boundary layer transport. Domain 1 covers the eastern United States. Domain 3 is slightly less than 200 km square (90 by 84 grid nodes) and encompasses most of western Pennsylvania, corresponding to the location of a suite of ground and aircraft measurements obtained during the New England Air Quality Study 2004 (NEAQS2004). The extent of Domain 3 corresponds to the typical size of a single GCM column. The period 5 August 2004 12Z through 11 August 2004 21Z was selected for simulation with the times before 9 August 6Z used as spin up and the remaining times used for analysis. The first six hours of the analysis period were almost cloud free due to a ridge over the area and sufficient cloudiness occurred during the remaining time to permit investigations of cloud-aerosol interactions. The latter portion of the simulation includes the passage of a warm and cold front pair. The simulation period is also concurrent with field campaign data to permit comparisons of observed and simulated values.

[8] WRF-Chem meteorological initial and boundary conditions were taken from the North American Regional Reanalysis [Mesinger et al., 2006]. Initial and inflow boundary conditions for trace gases and aerosols were derived from averaged August values for northeastern North America in MIRAGE GCM simulations [Easter et al., 2004]. Hourly aerosol and trace gas emissions were based on the U.S. EPA's 1999 National Emissions Inventory (NEI99) [EPA, 2001]. To better reflect actual emissions occurring during the simulation period, NEI99 SO2 and NOX emission estimates for SO2 point sources emitting ≥24 tons SO2 day−1 in Domains 2 and 3 were replaced with continuous emissions monitoring (CEM) stack emissions data reported by the U.S. EPA. Additionally, all NOX and SO2 emissions from stacks greater than 100 m in height that were not replaced with CEMs data were adjusted by recommended factors of 0.51 and 0.87, respectively, to reflect 1999–2004 point source emission trends [Frost et al., 2006].

[9] Three simulations were conducted to investigate the impacts of aerosol spatial heterogeneity. The “interactive aerosol” simulation (IA) uses the full suite of WRF-Chem modules and trace gas and particulate emissions described above in fully interactive aerosol-cloud feedback mode. The simulated CCN size distribution is a product of MOSAIC's eight sectional-size bins, resulting in a CCN distribution available for cloud formation that varies in space and time. Of the three simulations, IA is the most realistic and acts as our control. The two “prescribed aerosol” simulations (PA) use idealized, prescribed CCN distributions that are horizontally uniform, which mimics the grid cell mean that would be available in a GCM. The prescribed distributions in both PA runs are based on hourly output from IA during the analysis period and replace the aerosol module in the model. To assure an overall size distribution similar to that in IA, the hourly aerosol number, radius, and hygroscopicity for each MOSAIC size bin are averaged from the IA Domain 3 values and used to generate the prescribed CCN. A five point ring around the edge of Domain 3 is excluded from the averaging process to eliminate the transition zone between the nested domains. The two PA simulations differ in how the averaging is performed. In PAXYZT, the averaging is performed horizontally, vertically within the lowest 20 model levels (approximately the depth of the daytime boundary layer), and temporally, yielding a CCN size distribution that is spatially and temporally uniform. In PAXY, the averaging is performed only horizontally, yielding CCN profiles that vary in height and time. The PAXYZT approach is comparable to GCMs using a fully prescribed aerosol distribution or, given the short length of this analysis, a GCM that has slowly varying CCN such as those derived from monthly means [e.g., Ming et al., 2007]. The PAXY approach is comparable to GCMs that include the production and advection of aerosols [e.g., Storelvmo et al., 2006]. The direct and semi-direct effects in both PA simulations are replicated using the spatially and temporally varying aerosol optical properties that are output hourly from IA; this ensures that PA-IA simulation differences are solely due to indirect effects.

[10] Differences between IA and PAXYZT results demonstrate the impact of assuming uniform CCN characteristics within a GCM column. Differences between IA and PAXY results, relative to the magnitude of IA and PAXYZT differences, demonstrate whether the overall differences in the IA-PAXYZT comparison result mainly from horizontal heterogeneity or from the vertical and temporal variability of the aerosols.

3. Results

[11] Confidence in the validity of the model interpretation is dependant upon realistic conditions in the IA simulation. While space precludes presenting detailed validation results, the model's accuracy in predicting cloud and aerosol properties can be gauged qualitatively via comparison with NEAQS2004 ground observations from Indiana, PA, where a multi-filter rotating shadowband radiometer (MFRSR) operated. Aerosol and cloud optical depths (AOD and COD) were derived from MFRSR data following Michalsky et al. [2001] and Barnard and Long [2004], respectively, while downwelling shortwave radiation (SW) was measured directly. Figure 1 shows simulated versus observed AOD, SW, and COD for IA plus several days of the spin up period. The model captures both the increasing AOD trend from 7 August through 9 August and the general diurnal cycle. A slight model bias exists with simulated AOD overestimated on some days. Based on comparing simulated and observed SW and COD, the model captures the timing of most cloudy periods well. This is encouraging, as much of the simulation period experienced partly cloudy skies, a particularly difficult scenario for models to reproduce. The IA SW and COD ranges of values for the nine cells surrounding Indiana, PA indicate that the model is also producing highly variable clouds, e.g. on 7, 10 and 11 August.

Figure 1.

Time series of concatenated spin up and IA model results versus observations for aerosol optical depth (AOD), downwelling shortwave radiation (W m−2), and cloud optical depth (COD) at Indiana, Pennsylvania. The symbols represent observations and the orange shading represents the range of values in the nine model cells surrounding the observation location for the IA simulation. Note that observed AOD is only available for clear-sky times and observed COD is only available for fully overcast times. The vertical line at 9 August 6Z indicates the transition from the spin up simulation to IA.

[12] Multiple aircraft flights throughout the model domain also occurred during the simulation period. Overall, comparisons between aircraft data and IA (not shown) indicate that the simulated gas and particulate quantities are in agreement with the measurements except that the model SO2 is sometimes too high near stack exits and occasional errors in simulated wind direction lead to plume shifts. Shifting plume locations will not affect our investigation of the effects of prescribed versus interactive aerosols on spatially averaged cloud properties and radiation. Over-estimating SO2 potentially could influence IA sulfate aerosol number, radius, hygroscopicity, and mass and the corresponding derived PA values. However, the model does quite well in capturing the range of observed sulfate/SO2 ratios (suggesting part of the discrepancy may be due to uncertainties in input SO2 emissions), and both observed and simulated sulfate concentrations are within the range of values expected for a polluted continental region.

[13] Based on the reasonable agreement between observed and simulated quantities in IA, we proceed to examine the effects of CCN heterogeneity on cloud and radiative properties within Domain 3. As shown in Table 1, the impact of CCN heterogeneity on COD can be large. Overall, there are more cloudy columns appearing in both PA simulations with the mean COD between PAXYZT-IA and PAXY-IA increasing by 26.9% and 5.8%, respectively. Additionally, there is a shift in cloud thickness with formation of fewer thin clouds (i.e., fewer PA columns with CODs between 1 and 20) and more thick clouds (i.e., more PA columns with COD between 20 and 200). This point is emphasized in Figure 2, where the hourly 25th, 50th (median) and 75th COD percentiles are shown for IA and PAXYZT. The 25th percentile values exhibit only small differences between the simulations, but the 75th percentile values show substantial hourly variation. IA-PAXY results (not shown) exhibit similar temporal trends, but the 50th and 75th COD percentile values for each hour of each simulation agree more closely than in the IA-PAXYZT comparison.

Figure 2.

Time series of fine domain averaged cloud optical depth (COD) for IA and PAXYZT for columns with COD > 1. The bars indicate the range of COD values between the 25th and 75th percentiles within the domain, and the lines indicate the median values.

Table 1. Statistics Comparing Cloud Optical Depth (COD) for the Three Simulationsa
RunCount for Indicated COD RangeCount Change Versus IA, %Mean 0 to Max.Mean Change Versus IA, %
1 to 2020 to 2001 to Max.1 to 2020 to 2001 to Max.
  • a

    The “count” represents the number of occurrences of a column within the model domain having a COD within the given range of values during the simulation. The mean includes all model cells including those with COD < 1 to facilitate comparison to GCM results. Differences between PA and IA means for COD > 1 are similar, with differences of 24.1% and 5.5% for PAXYZT and PAXY vs. IA, respectively.


[14] Changes in cloud characteristics also lead to changes in the radiation budget, with less SW reaching the surface in both PA relative to IA (Table 2). As shown in the time series of Figure 3, instantaneous spatially-averaged PA-IA SW differences approach 60 W m−2, with sequential hours having differences of more than 15–20 W m−2. The sign of the difference is almost always negative due to the generally greater optical thicknesses of these clouds and the higher fraction of the domain containing clouds in both PA simulations.

Figure 3.

Time series of cloud optical depth, cloud fraction (as percent of fine domain columns with COD > 1), downwelling shortwave radiation, and cloud water path. The black line is the difference between PAXYZT and IA, the orange line is the difference between PAXY and IA, and the blue line is the value of the variable for IA. The blue lines correspond to the secondary y-axes on the right side.

Table 2. Statistics Comparing Downwelling Shortwave Radiation and Cloud Water Path for the Three Simulations
RunShortwave Radiation, W m−2Cloud Water Path, g m−2
Daylight MeanChange Versus IA (Diff./%)MeanChange Versus IA (Diff./%)

[15] The second indirect effect is related to greater numbers of droplets in polluted clouds slowing droplet growth, and therefore increasing cloud lifetime. Table 2 shows PA-IA differences for cloud water path (CWP), where CWP is the vertically integrated mass of all condensed water phases. Overall, the net impact of the spatially uniform CCN is a small positive bias; PAXYZT and PAXY exhibit 5% and 3% increases in CWP, respectively, versus IA. The CWP time series in Figure 3 reveals that most of this increase is due to a large difference during a rainy period on 10 August.

4. Discussion

[16] Intuitively one expects clouds to have different characteristics when CCN are uniformly distributed throughout a region (PA simulations) relative to when CCN are concentrated into plumes (IA simulation). The simulation results confirm this expectation; there is a shift towards optically thicker clouds and the fraction of Domain 3 containing clouds increases slightly in the PA simulations. This may be due to a longer duration for optically thick clouds (second indirect effect), since the percent changes in thick-cloud occurrence account for much of the percent changes to mean COD, and the 10–11 August COD and CWP peaks are temporally broader in the PA simulations.

[17] Figure 2 clearly indicates the impact of assuming a homogeneous CCN distribution varies in time, presumably due to changing meteorological conditions and the instantaneous regional aerosol mix. Since PAXYZT-IA COD differences are larger than PAXY-IA, the temporal variability of CCN plays an important role; the temporal variability of the PAXYZT-IA COD difference indicates that the impact of meteorology is also important. The meteorological influence is in agreement with Barker and Räisänen [2004], who found that that for a given day, geographic location plays a substantial role in determining the effect of assumed simplistic droplet distributions.

[18] The model domain used in our investigation is a fairly polluted region, and it is reasonable to expect that COD and SW differences between interactive and prescribed CCN simulations will tend toward the high end. However, the importance of meteorological effects on temporal variability suggests that fully investigating the impact of using simplistic CCN distributions in GCMs will require studies over large regions (to capture areas with differing climatologies) and over long periods (seasons to years, to capture cyclic meteorological conditions).

[19] Accurately reproducing the impact of clouds in atmospheric models requires a reasonable representation of CCN. PAXY-IA differences shown here presumably are smaller than the overall error in the cloud fields generated by most cloud parameterizations. Alternatively, our results suggest that, under the meteorological conditions encountered during the simulation period, using a constant, fully prescribed CCN distribution, as in PAXYZT, can lead to large biases that would impact long-term simulations used for climate change assessment. It is potentially possible to tune a fully prescribed CCN distribution to yield accurate long-term means, but day-to-day results will have errors that are likely to impact regional energy budgets and whose frequency could change in altered climate scenarios. This suggests that GCMs which simulate vertical and temporal fluctuations in CCN distributions are likely to be much more accurate and better able to capture regional cloud variations.

[20] The cost of simulating fully interactive aerosols is substantial. For the 8-bin sectional aerosol model used here, the time necessary to run IA is roughly 13 times greater than the simpler PAXYZT where no chemical reactions occur and several hundred fewer variables are advected. Obviously, simpler aerosol models, such as those employing a modal size distribution, would yield a lower run time differential, but one must still justify the added computational cost when designing accurate cloud parameterizations.

[21] Since feedbacks between clouds and aerosols are important in both directions, one must consider which processes are important to capture. For the impact of clouds on aerosols, one possibility would be to simulate clouds on a fine grid and aerosols on a coarse grid, similar to how meteorology variables are split within the multiscale modeling framework [Randall et al., 2003]. For the impact of aerosols on clouds where errors from neglecting subgrid-scale CCN horizontal variability are too large, one potential simplification to simulating CCN at high resolution is to design a statistical representation of the spatial CCN heterogeneity within a region and then link this representation to the cloud parameterization. Recently, similar techniques for handling meteorological variability have shown promise for improving cloud parameterizations [e.g., Berg and Stull, 2005; Larson et al., 2005]. While both suggestions would add substantial cost to a GCM simulation, the expense would be substantially less than simulating both high-resolution clouds and aerosols.


[22] We thank Jim Barnard of PNNL for providing the processed MFRSR data. This research was supported by the U.S. DOE Atmospheric Sciences program of the Office of Biological and Environmental Research under contract DE-AC06-76RLO 1830 at PNNL. PNNL is operated for the U.S. DOE by Battelle Memorial Institute.