Fine Scale Modeling of Agricultural Air Quality over the Southeastern United States Using Two Air Quality Models. Part I. Application and Evaluation

Two air quality models, the U.S. EPA Community Multiscale Air Quality (CMAQ) model and ENVIRON’s Comprehensive Air Quality Model with extensions (CAMx), are evaluated for their applications in simulating ambient air quality, in particular, the fate and transport of agriculturally-emitted NH3 over an area in the southeastern U.S. in January and July 2002 using a fine-scale horizontal grid resolution of 4-km. Both models moderately overpredict maximum 1-hr and 8-hr ozone (O3) and fine particulate matter (PM2.5) in January, due likely to a weaker vertical mixing and insufficient dry and wet removal of PM2.5 species simulated by the models. They either slightly underpredict or overpredict O3 but significantly underpredict PM2.5 in July. The large underprediction in PM2.5 is due to an excess wet deposition removal of sulfate, an excess dry deposition removal of precursors, and an underestimation of emissions of primary PM and precursors of secondary PM and secondary organic aerosol concentrations. Both models show large biases in the simulated concentrations of several gases (e.g., CO in CAMx, NO in CMAQ, NO2 in both models in both months and NH3 by both models in July) and PM species (in particular, nitrate in both months and carbonaceous PM in July), visibility indices, and dry and wet deposition fluxes. They also show some inaccuracies in reproducing temporal variations of NH3, PM2.5, dry and wet deposition fluxes. Differences in model performance between the two models are attributed to different model treatments such as vertical mixing, wet and dry deposition, SOA formation, and PM size representations. These results indicate a need to improve accuracies of the emissions and measurements of NH3, the emissions of primary PM and precursors of secondary PM, as well as model treatments of vertical mixing and dry and wet removal processes.


INTRODUCTION
Agriculturally-emitted species such as ammonia (NH 3 ), hydrogen sulfide (H 2 S), methane (CH 4 ), nitrous oxide (N 2 O), and volatile organic compounds (VOCs) have important impacts ambient air quality, the eutrophication of the ecosystems, as well as global and regional warming.Among these species, NH 3 is most concerned, because it is the most abundant alkaline gas in the atmosphere and plays an important role in the nitrogen cycle in the ecosystem, neutralization of acids in the air, and the formation of particulate matter having an aerodynamic diameter of 2.5 μm or less (PM 2.5 ).The wet and dry deposition of NH 3 to the soil is a source of nitrogen, providing nutrients to plants; however, too much nitrogen runoff into coastal waterways and estuaries can lead to eutrophication, increasing harmful algal blooms (Kinzig and Socolow, 1994;Paerl, 1997).NH 3 in the atmosphere neutralizes acids, such as nitric acid (HNO 3 ) and sulfuric acid (H 2 SO 4 ), creating salts (i.e., ammonium sulfate ((NH 4 ) 2 SO 4 ) and ammonium nitrate (NH 4 NO 3 )), which are a major component of PM 2.5 .Some studies have also indicated that NH 3 may play an important role in the formation of new particles through nucleation (Napari et al., 2002;Zhang et al., 2010).The ammonium ion (NH 4 + ) has a longer lifetime (up to 15 days) (Aneja et al., 2001) than NH 3 (up to 10 days) (Seinfeld and Pandis, 2006); it is thus capable to be transported and deposited to regions further from the source.NH 3 emissions can also cause odors, impacting the lifestyles of people in the area.For these reasons, an accurate understanding of the emissions, fate, and transport of agricultural livestock NH 3 (AL-NH 3 ) emissions is important in improving the air quality and ecosystem of the surrounding areas.The role of NH 3 further increases as the emissions of sulfur dioxide (SO 2 ) and nitrogen oxides (NO x ) are being reduced in many regions in the world as a result of air pollution control policy.Agricultural sources such as livestock, including cattle, poultry, swine and sheep contribute to 81% of NH 3 -nitrogen emissions in the U.S. (Battye et al., 1994).High NH 3 emissions are of major concern in the southeastern United States (U.S.), in particular, in the eastern North Carolina (NC) and northeastern Georgia (GA) because of a high density of animal feeding operations.
While most air quality modeling for the State Implementation Plans (SIPs) is performed at a horizontal grid resolution of 12-km, the U.S. EPA has suggested that SIP modeling (U.S. EPA, 2007), particularly over areas with high gaseous precursor emissions and primary PM sources, may benefit from increased grid resolution from 12-km to 4-km (U.S. EPA, 2007).Given high emissions of SO 2 and NH 3 in the southeastern U.S., 3-D agricultural air quality modeling at a fine-scale (< 12-km) will be necessary from both scientific and regulation perspectives.In this work, two commonly-used air quality modeling systems are applied to simulate agricultural air quality at a horizontal grid resolution of 4-km over a portion of the southeastern U.S.They are the U.S. EPA Community Multiscale Air Quality (CMAQ) (Binkowski and Roselle, 2003;Byun and Schere, 2006) modeling system and the ENVIRON's Comprehensive Air Quality Model with extensions (CAMx) (ENVIRON, 2006), both are driven by meteorological predictions from the Pennsylvania State University (PSU)/National Center for Atmospheric Research (NCAR) 5 th generation Mesoscale Model (MM5) (Grell et al., 1995).A simulation with CMAQ at a horizontal grid resolution of 1.33-km is also conducted over a nested domain.Four additional CMAQ simulations are conducted at 4-km with 50% of reduction in the emissions of PM 2.5 precursors to study the responses of PM 2.5 predictions to these precursors.The objectives of this work are to perform a comprehensive model evaluation to assess the models' capability in simulating agricultural air quality, examine the sensitivity of model predictions to different air quality models (i.e., CMAQ vs. CAMx) and horizontal grid resolutions (i.e., 12-km vs. 4-km vs. 1.33 km), and assess the relative importance of the major PM precursors (i.e., SO 2 , NO x , and NH 3 ) in PM 2.5 formation.Part I describes the model setup, including episode description, model configuration, evaluation database and protocol, and performance evaluation of the baseline simulations using both models at 4-km.Part II describes the sensitivity evaluations conducted, including sensitivity to grid resolution and reductions of emissions of precursors of PM 2.5 .

Episode and Model Configurations
MM5/CMAQ v4.5.1 and MM5/CAMx v4.42 simulations are conducted at a 4-km horizontal grid resolution in January and July 2002.The January and July 2002 episodes are selected in order to compare the fine scale simulation results with the coarser scale CMAQ simulations at a 12-km horizontal grid resolution previously completed by the Visibility Improvement State and Tribal Association of the Southeast (VISTAS) (Morris et al., 2007).When possible, the 4-km CMAQ and CAMx simulations use the model configurations and physics that are consistent with those of CMAQ used in the Phase II of the VISTAS modeling study at 36-km and 12-km horizontal grid resolution.MM5 version 3.7 with Four Dimensional Data Assimilation (FDDA) is used to drive the meteorology fields supplied to the Sparse Matrix Operator Kernel Emissions (SMOKE) modeling system, CMAQ and CAMx.Table 1 summarizes the configuration of MM5.The cumulus scheme and shallow convection are turned off because all clouds are assumed to be resolved at a 4-km grid resolution.Analysis nudging is used aloft for temperature, moisture, and winds and at the surface for winds using reanalysis data from NCAR (ds464.0 and ds353.4)(Olerud and Sims, 2004).The initial and boundary conditions (ICON and BCON) are derived from the VISTAS 12-km MM5 and CMAQ simulations and prepared for CAMx using the cmaq2camx tool.Emissions, based on the 1999 National Emission Inventory version 2 and additional data for VISTAS states (MACTEC, 2008), are prepared for CMAQ using SMOKE version 2.1.
Table 2 summarizes the major model configurations used in model simulations.When available, both models use the same or similar options, e.g., the Carbon Bond IV gas-phase mechanism, the ISORROPIA inorganic aerosol thermodynamic module, and the Regional Acid Deposition Model (RADM) aqueous chemistry.One of the major differences between the models is the treatment of vertical advection.The Yamartino-Blackman Cubic advection option in CMAQ utilizes the Piecewise Parabolic Method (PPM) for horizontal advection, and then uses the density from MM5 to calculate the vertical velocity at each grid cell that satisfies the continuity equation.While CAMx also uses PPM for horizontal advection, the vertical diffusion and advection are calculated implicitly (ENVIRON, 2006).The difference in the treatment of vertical mixing between the models results in weaker vertical mixing of pollutants, and thus higher pollutant concentrations near the surface, by CAMx, as compared with CMAQ (Zhang et al., 2004).
The second difference between the two models is the representation of PM size distribution.CMAQ represents PM using a modal distribution (i.e., three log-normal modes: nuclei, accumulation, and coarse), whereas CAMx uses a sectional approach (e.g., 2 or more bins as specified by the user).Increasing the number of sections can improve the model representation of PM distribution; however, it also increases the computational time.Two size sections are used in the CAMx simulation in this study.The third difference between the models is the treatment of secondary organic aerosols (SOA).Both models simulate SOA formation from VOCs included in the CB-IV gas-phase mechanism.ENVIRON enhanced the SOA module in CMAQ by including sesquiterpenes, additional SOA formation from isoprene, and polymerization of SOA species (Morris et al., 2006(Morris et al., , 2007)).The models also treat dry and wet deposition differently.The dry deposition of gases in CAMx is based on Wesely's (1989) resistance model, which calculates the deposition velocity using the aerodynamic, boundary layer, and surface resistances.The dry deposition of gases in CMAQ v4.5.1 is the Models-3 dry (M3DRY) deposition model, which is coupled with the Pleim-Xiu land-surface model (LSM) to use the aerodynamic and boundary layer resistances simulated from this LSM in MM5 in order to improve the stomatal resistance over a non-coupled model (Otte and Pleim, 2010).A more advanced treatment for NH 3 such as the bi-directional air-surface exchange algorithm for NH 3 is not included in CMAQ v4.5.1 used in this work, but it has been included in CMAQ v. 4.7 or newer (Cooter et al., 2010).CAMx simulates PM dry deposition using the approach of Slinn and Slinn (1980).CMAQ uses the approach from the Regional Particulate Model developed by Binkowski and Shankar (1995).The wet deposition of gases in both models is based on Henry's law equilibrium; however, the absorption of particles into cloud water is treated differently between the models.CAMx assumes that all particles in a cloudy grid cell are contained within cloud water, while CMAQ assumes the accumulation and coarse mode particles are completely absorbed by cloud water and particles in the nuclei mode are slowly scavenged (Byun and Schere, 2006).These differences lead to differences in predictions using the two models.More details on model setup can be found in Olsen (2009).

Evaluation Database and Protocol
The meteorology and air quality model results are evaluated against available observations to assess model performance.Table 3 summarizes data from the surface networks for model evaluation.These include national networks (e.g., the Clean Air Status and Trends Network (CASTNET), the Speciation Trends Network (STN), the Interagency Monitoring of Protected Visual Environments (IMPROVE), the Aerometric Information Retrieval System -Air Quality Subsystem (AIRS-AQS), and the National Atmospheric Deposition Program (NADP), as well as state, local, and private agencies (e.g., the Southeastern Aerosol Research and Characterization (SEARCH) study, the North Carolina Department of Environmental and Natural Resources (NCDENR), and the North Carolina State Climate Office (NC SCO).The meteorological variables evaluated include hourly temperature at 1.5-m (T1.5), relative humidity at 1.5-m (RH1.5),wind speed (WS10) and direction at 10-m (WD10), and weekly total precipitation (Precip).Chemical variables evaluated include 1-hr and 8-hr maximum O 3 , carbon monoxide (CO), nitrogen oxide (NO), nitrogen Table 3.The observational networks used in model evaluation along with the variables evaluated, the sampling frequency, and the number of sites within the 4-km and 1.33-km modeling domains.The evaluation of model performance is conducted using statistics, spatial distributions, and temporal analysis.The statistics, including mean observation, mean simulation, correlation, and normalized mean bias and error (NMB and NME, respectively), are separated by networks because of their varying characteristics in terms of sampling frequency and resolution, monitoring approach, and type of area (e.g., urban vs. rural) following several studies (e.g., Eder and Yu, 2006;Zhang et al., 2009).The simulated dry deposition velocities and fluxes of some species are compared with the results of the Multilayer Model (MLM), which uses species concentrations and leaf area index recorded at the CASTNET sites.More details on the MLM are provided in Section 3.4, as well as Meyers et al. (1998) and Finkelstein et al. (2000).The statistics for evaluation of simulated dry deposition velocities and fluxes is calculated against values from MLM. Nine locations are selected for temporal analysis; a coastal site (Beaufort (BFT), NC), a mountain site (Great Smoky Mountain (GRS), TN), two urban sites (Raleigh (RAL), NC, and the Jefferson Street (JST), downtown Atlanta, GA), two rural sites (Yorkville (YRK), GA, and Candor (CND), NC), and three sites in the eastern NC (Kinston-Lenoir (LCC), Clinton (CLT), and Jamesville (JMV)) where NH 3 emissions are high and the measurements of NH 3 mixing ratios are available.Among them, JST and YRK are SEARCH sites, BFT, GRS, and CND are CASTNET sites, and RAL, LCC, CLT, and JMV are NCDENR sites.

Meteorology
Table 4 summarizes the performance statistics for meteorological variables.For WS10, a cut-off value of 1.5 knots (i.e., 0.771 m/s) is used because of instrumentation limitations in reporting calm wind speeds following Olerud and Sims (2004).When the observed WS is less than this value, the data pair is not included in the statistical calculations.The NMBs of T1.5 and RH1.5 are generally within ± 10%, with a few exceptions in January.The large cold bias in T1.5 in January is likely due to too cold soil   Zhang et al., Aerosol and Air Quality Research, 13: 1231-1252, 20131236 initial temperatures and inappropriate snow treatments (Liu et al., 2010).Larger cold biases at the SEARCH and SCO sites may be due to an additional model limitation in capturing an urban heat island effect.For a similar reason, the overprediction in RH1.5 at the SEARCH sites is larger than that at the CASTNET sites.WS10 is overpredicted in both months with a better performance in January.The MBs for WD10 are within ± 12°, indicating an overall small deviation from observations.Precip is underpredicted with an NMB of -10.9% in January but significantly overpredicted with an NMB of 148.5% in July due to the limitation of MM5 in capturing convective rainfall in terms of its frequency and intensity (Olerud and Sims, 2004).This overprediction indicates an inability of MM5 in reproducing a drier than historical average season in the region.It affects the removal of pollutants through wet deposition, as shown in Section 3.4.

Chemical Concentrations of Gaseous and PM Species
The statistics for several gaseous (i.e., O 3 , CO, NO, NO 2 , HNO 3 , SO 2 , and NH 3 ) and PM species (i.e., PM 2.5 , NH 4 + , SO 4 2-, NO 3 -, EC, OC, and TC) are provided in Tables 5 and 6.Despite underpredicted T1.5 by MM5 at most sites, the maximum 1-hr and 8-hr O 3 mixing ratios are slightly overpredicted by both models in January (except at the SEARCH sites) and by CAMx in July, due likely to a weaker than actual vertical mixing that prevents the dispersion of precursor species.The maximum 1-hr and 8-hr O 3 mixing ratios are slightly underpredicted by CMAQ in July, due partly to underpredicted peak T1.5 by MM5 on most days at most sites and partly to the underestimate in emissions of precursors, particularly at the AIRS-AQS sites, that dominates over the effects caused by a weaker than actual vertical mixing simulated by CMAQ.This weaker vertical mixing can be reflected in the overprediction of CO mixing ratio by both models.The CO statistics also indicate that CAMx has a much weaker vertical mixing than CMAQ, resulting in a much larger overprediction of CO in both months.In January, both models underpredict the mixing ratios of NO and overpredict those of NO 2 .In July, the mixing ratios of NO are underpredicted by CMAQ but overpredicted by CAMx; both overpredict those of NO 2 .Despite overpredicted NO 2 mixing ratios in both months, the underprediction and overprediction in Precip in January and July, respectively, lead to an overprediction of HNO 3 mixing ratios (which is highly soluble) in January at all sites and an underprediction of HNO 3 mixing ratios at the SEARCH sites in July by CMAQ.CAMx underpredicts the mixing ratio of HNO 3 at all sites in both months, due partly to different dry deposition treatments as compared  -7), resulting in lower mixing ratios of HNO 3 .Both models overpredict SO 2 mixing ratios in both months, with higher values by CAMx than by CMAQ, due mainly to a weaker vertical mixing and a lower (by 69% on average) dry deposition velocity of SO 2 calculated by CAMx.Both models significantly underpredict NH 3 mixing ratios, particularly in July.One possible reason is underestimation of NH 3 emissions in the eastern NC, despite significant efforts by the VISTAS program on improving NH 3 emissions (e.g., Morris et al., 2007).Another possible reason is the large uncertainties in the NH 3 measurements.The difficulties of measuring NH 3 with high temporal resolution, due to its "sticky" nature, are well documented (e.g., von Bobrutzki et al., 2010).The concentrations of NH 3 measured at LCC in July using a Thermo Environmental Instruments (TEI) Model 17C Ammonia analyzer (Shendrikar, 2006) are much higher (e.g., by a factor of 8.7 for monthly mean values, as compared with the mean values of Walker et al. (2004) than measurements using other methods such as the annular denuder system during summers 1999-2000 at the same site and at Clinton, an agricultural site with the highest emission density that is located 52 miles southwest of Kinston (e.g., Robarge et al., 2002;Walker et al., 2004).Furthermore, the concentrations during late July are higher than concentrations typically observed in high emission density areas of eastern NC (e.g., using the ALPHA passive sampler and the Tropospheric Emission Spectrometer (TES), Pinder et al., 2011).Shendrikar (2006) clearly showed that the extremely high NH 3 concentrations in 2002 does not reflect the typical annual cycle observed at LCC and CLT sites.According to the historic climate records at the NOAA's National Climatic Data Center, June through August 2002 was much warmer and drier than average summers in the U.S. and North Carolina.The emissions of NH 3 would likely high or the sampling artifact by the TEI Model 17C Ammonia analyzer may be high under extremely high temperature conditions.At present it is uncertain whether the high summertime concentrations at LCC are the result of a temporary local source or a sampling artifact.The weaker vertical mixing simulated by both models can also impact PM 2.5, which is also overpredicted by both models in January.Despite a weaker vertical mixing in July, other factors, such as the underestimation in the emissions of primary EC and OC and SOA concentrations, and the overprediction in the removal of SO 4 2-through wet deposition, are dominant, resulting in the underprediction of PM 2.5 in July.The performance of individual species vary, e.g., the overprediction of SO 4 2-by CAMx but underprediction by CMAQ at the CASTNET and STN sites in January and much larger underprediction of SO 4 2by CMAQ than CAMx in July.This indicates several other factors and/or differences between the models are more influential to PM species in both months than vertical mixing alone.First, the oxidation of SO 2 to SO 4 2-may be 28.4 -0.04 -34.9 75.9 CAMx 6.9 -0.06 -98.4 98.4 1 Dry deposition fluxes and dry deposition velocities are not observed measurements, but calculated from measurements using the Multilayer Model (MLM) (Meyers et al., 1998;Finkelstein et al., 2000)., and NO 3 -, which can partly explain differences in their predicted concentrations.For example, compared with CMAQ, CAMx gives lower dry deposition fluxes for all these species in both months and lower wet deposition fluxes for all species in January but higher wet deposition fluxes for NH 4 + and SO 4 2-in July (see Table 7).In January, the lower dry and wet deposition fluxes of SO 4 2-explain higher concentrations of SO 4 2-by CAMx than by CMAQ.In July, despite a higher wet deposition flux of SO 4 2-, the dry deposition flux of SO 2 is lower and the vertical mixing is weaker by CAMx, both factors lead to higher mixing ratios of H 2 SO 4 and thus higher concentrations of SO 4 2-by CAMx than by CMAQ.
The statistics for EC, OC, and TC in both months indicate a generally higher prediction by CAMx than CMAQ, which is likely affected by a weaker vertical mixing and additional SOA formation from VOCs in CAMx, as well as the differences in removal processes between the two models.Both models, however, underpredict OC, EC, and TC, due likely to the underestimation in the emissions of primary EC and OC and SOA formation.Among all PM species, both models generally perform the worst for NO 3 -.For example, NO 3 -concentrations are significantly overpredicted by both models at the STN sites in January, which is associated with the overpredictions in NH 4 + concentrations.In July, although large underprediction occurs for the concentrations of NO 3 -, it has little impact on the PM underprediction because it is the smallest component of PM.Underprediction in the concentrations of SO 4 2-and NO 3 -can explain underprediction in the concentration of NH 4 + at all sites in July by both models.
The simulated monthly-mean PM 2.5 concentrations by CMAQ and CAMx are overlaid with observations in January and July in Fig. 1.In both months, CAMx predicts higher values than CMAQ throughout the domain, likely due to a weaker vertical mixing and lower dry and wet deposition fluxes of secondary PM 2.5 (except for the wet deposition flux of SO 4 2-in July) by CAMx than CMAQ.In January, the overprediction of PM 2.5 by both models is significant near Atlanta, GA, which may be partly due to overestimate of SO 2 emissions from the Bowen power plant in Bartow, GA.The Bowen plant has been ranked among the top 50 emitters of SO 2 in the U.S. In July, the underprediction occurs throughout the domain.In the eastern portion of the domain, PM 2.5 predictions are higher in January than July, contradictory to observations, which may be attributed to a weaker vertical mixing in January than in July.
Fig. 2 shows observed and simulated hourly O 3 mixing ratios at six sites.In January, the predicted O 3 mixing ratios by both models are similar at non-urban sites such as YRK, GRS, and BFT but are apparently different at urban sites such as JST and RAL and a site with high NH 3 mixing ratio, i.e., LCC.CAMx gives higher O 3 mixing ratios during the daytime at these sites due to higher precursor mixing ratios as a result of a weaker daytime mixing.CMAQ tends to give higher O 3 mixing ratios at nights than CAMx due to less titration of O 3 by NO x as a result of a stronger nocturnal vertical mixing.The minimum O 3 mixing ratios are generally captured at JST, YRK, and GRS by CAMx and overpredicted by both models at BFT.The maximum O 3 mixing ratios are underpredicted throughout the month at JST and during the first half of the months at other sites by both models.In July, CMAQ also gives higher O 3 mixing ratios at night and lower daytime O 3 for the same reasons.Similar to January, the minimum O 3 mixing ratios are generally captured at JST, YRK, RAL, and GRS by CAMx and overpredicted by both models at BFT and LCC.The maximum O 3 mixing ratios are underpredicted by both models in the beginning of the month and overpredicted at the end of the month.The diurnal cycle is better captured by both models at the urban, rural, and mountain sites (i.e., JST, RAL, YRK, and GRS) but not well represented at the costal site (BFT) and the site with high NH 3 (LCC).Fig. 3 shows observed and simulated mixing ratios of NH 3 at three sites in the eastern NC.Large differences exist in simulated mixing ratios of NH 3 , particularly at LCC and JMV, due to different treatments in vertical mixing, dry and wet deposition, and gas-to-particle conversion processes (as a result of different size representations and different versions of ISORROPIA).While both models overpredict mixing ratios of NH 3 at CLT in January, they significantly underpredict those at LCC in January and CLT in July on most days, and at LCC in July and JMV in both months throughout the months.In particular, the model fails to capture the extremely high mixing ratios of NH 3 (up to 183 ppb) at LCC in July.The large discrepancies indicate large uncertainties in NH 3 emissions in terms of both magnitudes and spatial distributions (e.g., a possible underestimate in winter and an overestimate in summer, Zhang et al. (2006)) and in NH 3 measurements as mentioned previously.
Fig. 4 shows the observed and simulated hourly concentrations of PM 2.5 at five sites.In January, the models show an overall good agreement with the observations at YRK and GRS in terms of both magnitude and temporal variations, but a large overprediction exists at other sites, particularly at JST by CAMx, because of overprediction in the concentrations of SO 4 2-, NH 4 + , and EC.In July, both models significantly underpredict the concentrations of PM 2.5 at most sites during most time periods, particularly at RAL and GRS, although they do capture some of the long term trends (i.e., the increase and decrease of PM 2.5 at JST, YRK, and RAL during some periods and the three peaks at LCC).The large underprediction at all sites in the beginning of July is likely due to the long range transport of PM 2.5 mass from a forest fire in Canada (DeBell et al., 2004) that is not accurately represented in the emission inventories used by both models.While forest fire emissions are included in the inventory, many assumptions are made (MACTEC, 2008) in estimating the emissions, resulting in possible errors in the inventory.Other factors contributing to underpredictions may include underestimate of primary PM emissions and the formation of secondary inorganic aerosol and SOA.

Visibility
CMAQ uses two methods to calculate these optical properties: one calculates β ext based on aerosol size distribution and the other is based on the species mass concentration (Binkowski and Roselle, 2003;Mebust et al., 2003).The latter method is selected here for the evaluation because it is similar to the calculation used by IMPROVE and can be readily adapted for CAMx.The relative humidity factor (i.e., f(RH)) is needed to calculate β ext .While f(RH) is obtained from a lookup table based on Malm et al. (1994) in CMAQ, it is calculated for each site by CAMx based on U.S. EPA (2003).In a pristine environment, β ext = 0.01 km -1 (Mebust et al., 2003).The HI, reported in deciviews (dcv), is calculated based on β ext .In a pristine environment, the HI is equal to 0. CAMx does not internally calculate any optical properties but the mass concentrations  of PM species at each IMPROVE site can be used to calculate β ext and HI.Table 6 shows performance statistics of β ext and HI at the IMPROVE sites from both models.Both models simulate optical properties worse than observed, i.e., higher β ext , due to the overpredicted concentrations of PM 2.5 in January, with a cleaner environment simulated by CAMx than CMAQ.This may be a result of using different f(RH) values in CAMx than those used in CMAQ.The HI was slightly overpredicted by CMAQ and underpredicted by CAMx.In July, the underpredictions in the simulated β ext and HI values are consistent with the underpredictions of PM 2.5 concentrations, with lower predictions by CMAQ than by CAMx.

Dry and Wet Deposition Fluxes
The dry deposition velocity and flux are calculated using MLM, which is based on the concentrations and vegetation data collected by CASTNET.Compared with the M3Dry or Wesely modules which treat the canopy as one layer, MLM is a more detailed model that uses 20 layers within the canopy layer and calculates individual boundary and stomatal layer resistances for each canopy layer (Meyers et al., 1998).MLM has been evaluated against some limited observations and found to have varying performance, ranging from an underestimation of SO 2 dry deposition velocity by 35% (Finkelstein et al., 2000) to an overestimation by 18.3% (Meyers et al., 1998).The MLM predictions are used as a benchmark to evaluate dry deposition fluxes and velocities simulated by CMAQ and CAMx in this work, as shown in Table 7 and Figs.5-8.Compared to MLM in both months, M3Dry in CMAQ shows some improvement in dry deposition velocities and fluxes of HNO 3 over the Wesely's deposition module in CAMx, although they are significantly overpredicted by both models.While the NMBs for SO 2 dry deposition velocity in CMAQ indicate significant overpredictions (159% and 91% in January and July, respectively), those by CAMx are within the biases of the MLM estimations against observations (-21% and -1%, respectively).MLM contains dry deposition velocities of SO 2 and HNO 3 only, which are shown along with those of CMAQ and CAMx in Fig. 5 at GRS, CND, and BFT.CMAQ gives much higher dry deposition velocities of SO 2 than those of MLM and CAMx on some days at GRS and most days at CND and BFT in January and some days at all sites in July but to a lesser extent.CAMx gives lower values than MLM on some days at CND and most days at GRS and BFT in January and agree quite well at all sites in July.For dry deposition velocities of HNO 3 , CAMx gives slightly higher values than CMAQ at all sites, both give higher values than MLM on most days in both months, with better agreement between MLM calculations and both model predictions in July than in January.These discrepancies in dry deposition velocities will propagate into dry deposition fluxes calculations.
The NMBs of SO 2 dry deposition fluxes simulated by CMAQ and CAMx are 58% and -21% in January and 56% and 16% in July, respectively.This is due partly to the overpredictions in the concentrations of SO 2 (see Table 5) and partly to the overestimation of the dry deposition velocity of SO 2 in both months.The latter can limit the amount of the gas available for gas-to-particle conversion.Since the degree of overpredictions of the concentrations    of SO 2 by both models is similar, the larger overprediction in its dry deposition fluxes by CMAQ is likely dominated by the larger overprediction in the dry deposition velocities of SO 2 than CAMx.The NMBs of HNO 3 dry deposition fluxes simulated by CMAQ and CAMx are 151.4% and 194.1% in January and 62.1% and 164.3% in July, respectively.While CMAQ overpredicts the concentrations of HNO 3 in both months, CAMx gives zero bias in January and moderately underpredicts them.The overprediction in dry deposition fluxes of HNO 3 is therefore caused primarily by the overpredicted dry deposition velocity in CAMx but mostly by overpredicted dry deposition velocity and to a lesser degree by the overpredicted concentrations of HNO 3 in CMAQ.The dry deposition fluxes of all three PM species (i.e., NH 4 + , SO 4 2-, and NO 3 -) are significantly underpredicted by both models in both months, which may be influenced by several factors including their underpredicted mass concentrations in July (see Table 6(b)), the underestimates in their estimated dry deposition velocities, and the overestimation of dry deposition fluxes of their gaseous precursors.
The hourly dry deposition fluxes of SO 2 , HNO 3 , NH 4 + , SO 4 2-, and NO 3 -calculated by MLM, CMAQ, and CAMx are shown in Figs.6-8 at GRS, CND, and BFT, respectively.Significant discrepancies exist in calculated dry deposition fluxes of all these species at all sites.CMAQ tends to give the highest dry deposition fluxes of SO 2 and CAMx tends to give the highest dry deposition fluxes of HNO 3 at most sites during most time periods in both months.Because NH 4 + is often associated with SO 4 2-, their dry deposition fluxes are generally similar in terms of magnitudes and variation patterns.However, both models fail to capture the maximum fluxes for NH 4 + , SO 4 2- , and NO 3 -determined by MLM at all three sites in both months.CMAQ tends to predict a much larger daily variation in NH 4 + and SO 4 2-dry deposition fluxes than CAMx and MLM.Among the three PM species, the dry deposition fluxes of NO 3 -exhibit the largest discrepancy against the MLM values at all sites in both months, due partly to the worst performance in the predicted concentrations of NO 3 -.The overall dry deposition results indicate large uncertainty in the simulated dry deposition using different dry deposition models with different levels of details in their treatments (e.g., the single-layer, bulk, and 20 layers).
Table 7 also shows the statistical evaluation of wet deposition fluxes of NH 4 + , SO 4 2-, and NO 3 -against the NADP measurements.Despite underpredicted precipitation, CMAQ overpredicts the wet deposition fluxes of all species in January, indicating that wet deposition is a more efficient removal process than dry deposition for those species.CAMx, however, underpredicts their wet deposition fluxes in January.The underpredictions of dry and wet deposition fluxes by CAMx also contribute to the higher PM 2.5 concentrations than CMAQ in January.Both models ovepredict the wet deposition fluxes of SO 4 2-in July because there is sufficient ambient SO 4 2-to be removed from the atmosphere by the excess precipitation simulated by MM5.Despite the overpredicted precipitation, NO 3 wet deposition fluxes are underpredicted by both models because there is limited NO 3 -available to be removed from the atmosphere in the southeast in July.There is a mixed performance for the wet deposition fluxes of NH 4 + , with an overprediction by CMAQ and an underprediction by CAMx.These differences may be attributed to the differences in PM size representations and influential model treatments such as wet and dry deposition and vertical mixing.Comparing with dry deposition flux predictions, the correlation between the biases in the predicted concentrations and wet deposition fluxes is not strong, and in many cases their biases indicate an opposite trend.For example, in July, both models underpredict the concentrations of NH 4 + , SO 4 2- , and NO 3 -, however, the wet deposition of SO 4 2-is overpredicted by both models.This indicates the high non-linearity in the wet deposition flux prediction, which depends on the rate limiting impacts of chemical concentrations and/or wet scavenging (i.e., the wet deposition fluxes are limited by insufficient ambient concentrations or less precipitation or both).

CONCLUSIONS
The performance of two commonly-used air quality modeling systems, MM5/CMAQ and MM5/CAMx, is evaluated against observations from a number of networks for their applications over an area in the southeastern U.S.at a horizontal grid resolution of 4-km in January and July, 2002.Simulated 1.5-m temperature and relative humidity using MM5 are generally within ± 10% of the observations.Wind speed at 10-m is moderately overpredicted in both months.Precipitation in July is significantly overpredicted, which affects the removal of pollutants through wet deposition.Differences in simulated major air pollutants between CMAQ and CAMx simulation results can be attributed to several differences in model treatments, such as the treatment of vertical mixing and SOA formation, PM size representation, and wet and dry deposition modules.Both models overpredict surface concentrations of PM 2.5 and maximum O 3 mixing ratios in January, with higher values by CAMx.This is likely due to a weaker vertical mixing simulated by the models, particularly CAMx, as indicated by the overprediction in CO.Despite a weaker vertical mixing simulated in July by both models, the concentrations of PM 2.5 and O 3 are underpredicted.This is due to an excess removal of pollutants through wet deposition (i.e., SO 4 2-), an excess removal of precursors through dry deposition (e.g., SO 2 ), and an underestimation of emissions of primary PM and precursors of O 3 and secondary PM and SOA concentrations.While both models reasonably capture the diurnal variations of O 3 , they show inaccuracies in simulating those of NH 3 and PM 2.5 .Uncertainty in the emission estimates and the spatial distribution of the emissions and measurements may be responsible for large biases in simulated NH 3 mixing ratios, indicating a need to improve the accuracy of NH 3 emissions and measurements.Difficulties in simulating PM 2.5 arise from the volatile species (i.e., NO 3 -), lack of information in the formation mechanisms (i.e., SOA), and large uncertainties in the emissions of primary PM (e.g., EC and OC).Large discrepancies exist in the dry and wet deposition predictions from CMAQ and CAMx and from MLM and NADP.The discrepancies in dry deposition fluxes between the two models and between model predictions and MLM/NADP values are likely dominated by discrepancies in the dry deposition velocities of species.Those in wet deposition fluxes are likely dominated by differences in the wet scavenging of species, in particularly in July.These differences may be attributed to the differences in PM size representations and treatments for wet and dry deposition and vertical mixing.
Since both models have been used for regulatory applications, the results from this work provide an assessment of their capability in reproducing observed concentrations and deposition fluxes as well as visibility and identify several areas of improvement, which would be useful for their continuous development and applications in support of the SIP modeling efforts and the development of other emission control policies as well as in agricultural air quality modeling.The areas of improvement identified through this work should be given more research attentions to enhance the models' capability and fidelity.

a
There are a total of 73 NCDENR sites in the 4-km domain, some with collocated observations (46 sites with hourly O 3 observations, 37 sites with 24-hour average PM observations, 8 sites with hourly PM observations, and 8 sites with hourly meteorology observations).dioxide (NO 2 ), HNO 3 , SO 2 , NH 3 , PM 2.5 , ammonium (carbon (EC), organic carbon (OC), total carbon (TC = EC + OC), the dry deposition (DD) of NH 4 +, parameters evaluated are the extinction coefficient (β ext ) and haziness index (HI).
DV: Deposition Velocity; DD: Dry Deposition; WD: Wet Deposition.underestimated because SO 2 mixing ratio is overpredicted and SO 4 2-concentration is underpredicted by both models.Second, CMAQ and CAMx contain different treatments for dry and wet removals of NH 4 + , SO 4 2-

Table 1 .
MM5 model configurations used in this work.

Table 2 .
CMAQ and CAMx model configurations used in this work.

Table 4 .
Performance statistics for meteorological variables at a 4-km horizontal grid resolution for January and July 2002.

Table 5 .
Performance statistics for trace gases a 4-km horizontal grid resolution in January and July 2002.

Table 6 .
Performance statistics for PM species a 4-km horizontal grid resolution in January and July 2002.

Table 7 .
Performance statistics for dry and wet deposition of species a 4-km horizontal grid resolution in January and July 2002.