A three-dimensional chemical transport model (PMCAMx) is used to simulate particulate matter (PM) mass and composition in the eastern United States during the four seasons of the year (July 2001, October 2001, January 2002, and April 2002). The model predictions are evaluated against daily average PM2.5 measurements taken throughout the eastern United States by the IMPROVE and STN monitoring networks and the EPA Supersites program. During the spring and summer the model reproduces the measured daily average PM2.5 concentrations with an error of less than 50%, two thirds of the time. The PM2.5 error is less than 30% for 43% of the measurements during these seasons. For the fall and winter the PM2.5 predictions are within 50% of the measurements for 51% of the data points and within 30% for 34% of the time. The performance of the model in reproducing sulfate, organic mass, elemental carbon, and total PM2.5 concentrations varies from average to good depending on the season. Uncertainties in ammonia emissions during the fall cause errors in the corresponding ammonium predictions, while the ammonia emissions inventory appears to be satisfactory during the other seasons. The ability of the model to reproduce the aerosol nitrate concentrations in the spring and summer is limited by difficulties in simulating the heterogeneous nighttime formation rate of nitric acid. During the summer and fall the model performance in reproducing the organic PM concentrations and diurnal patterns is good. The predicted organic PM during the summer is on average 60% primary and 40% secondary. The secondary contribution to organic PM drops to around 20% during the winter. The used average EC emission rate of approximately 0.55 ktons d−1 (0.45 ktons d−1 during the weekends) is consistent with the observed EC concentrations. The nighttime chemistry of NOx determines the PM nitrate concentrations during most days in the winter, spring, and fall and its mathematical description needs improvement. The good agreement between the predicted and observed temporal profiles for most species suggests a reasonable understanding and depiction in the model of the corresponding processes. Additional strengths and limitations of current modeling approaches for this modeling domain and for the different seasons are further discussed.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Atmospheric particles have adverse effects on human health and contribute to visibility reduction, the formation of acid rain and acid fogs, and influence the energy balance of the planet. Air quality standards have been proposed that limit the daily and yearly average concentrations of PM2.5 and PM10 (particulate matter less than 2.5 μm and 10 μm in diameter, respectively). Three-dimensional chemical transport models that can accurately and efficiently describe the physical and chemical atmospheric transformations of these pollutants are crucial for the development of emission control strategies to achieve these air quality standards.
 The first multicomponent particulate matter chemical transport models (CTMs) were developed and applied in California. An Eulerian model was used to describe the transport and formation of pollutants in the south California air basin by Russell et al. . A three dimensional Eulerian CTM simulating the major inorganic and organic PM components and their size distribution was developed by Pilinis and Seinfeld . Both of these initial models were evaluated in Los Angeles for the 30–31 August 1982 smog episode. UAM-AERO [Lurmann et al., 1997] and the CIT photochemical airshed model [Harley et al., 1993] were applied in the South Coast Air Basin of California during the 1987 Southern California Air Quality Study (SCAQS). More recent model evaluation studies in southern California include those of Jacobson , Kleeman and Cass , Zhang et al. , and Morris et al. [2005a]. The October 1995 PM episode in the South Coast Air Basin of California was simulated by Gaydos et al.  using PMCAMx. Fewer studies have been conducted in areas such as the eastern United States, however, where PM mass is dominated by sulfate and organics, in contrast to the high ammonium nitrate concentrations seen in California. IMPROVE observations collected for the eastern United States during June 1995 [IMPROVE, 1995] were used for evaluation of the Community Multiscale Air Quality (CMAQ) model results [Mebust et al., 2003]. Other model evaluation studies in the eastern United States have been conducted by Boylan et al. , Morris et al. [2005b], and Marmur et al. . The overall performance of PMCAMx was evaluated in the eastern United States for the July 2001 episode by Gaydos et al. . The model predictions were evaluated against hourly measurements of PM and gas phase data collected during the Pittsburgh Air Quality Study (PAQS) [Wittig et al., 2004], as well as daily average PM measurements taken throughout the eastern United States by the IMPROVE and STN [U.S. Environmental Protection Agency (U.S. EPA), 2002] monitoring networks. More recent studies have applied CTMs to the entire continental United States. Eder and Yu  evaluated the performance of CMAQ for 2001 against monitoring data of four nationwide networks. Tesche et al.  evaluated and compared CMAQ and CAMx during the full calendar year 2002. CTMs have also been applied in Europe in areas such as the central and eastern Mediterranean [Sotiropoulou et al., 2004], western Mediterranean [Jimenez et al., 2005], and London [Sokhi et al., 2006].
 The present study adds to these previous efforts to evaluate our current understanding of the atmospheric processes responsible for the spatial, temporal, and seasonal variability of fine PM over the eastern United States. The evaluation focuses on the ability of PMCAMx, to reproduce not only daily average concentrations but also the diurnal variation of the major aerosol components during the various seasons.
 The current study begins with a brief description of PMCAMx and its application to four periods during the year (July 2001, October 2001, January 2002, and April 2002) in the eastern United States. The model predictions are compared against daily average PM measurements taken throughout the eastern United States by the IMPROVE [IMPROVE, 1995] and STN [U.S. EPA, 2002] monitoring networks and also the Pittsburgh Supersite semicontinuous measurements [Wittig et al., 2004]. The strengths and limitations of current modeling approaches for this modeling domain and for the different seasons are discussed in the conclusions section.
2. Model Description
 PMCAMx uses the framework of CAMx [Environ, 2003], which models the processes of horizontal and vertical advection, horizontal and vertical dispersion, wet and dry deposition, and gas phase chemistry. In addition, three detailed aerosol modules are used: inorganic aerosol growth [Gaydos et al., 2003; Koo et al., 2003], aqueous phase chemistry [Fahey and Pandis, 2001], and secondary organic aerosol formation and growth [Koo et al., 2003]. PMCAMx is the research version of the publicly available CAMx model.
 The general equation solved in PMCAMx to describe the change in concentration over time for a given species, ci, is:
The aerosol size and composition distribution is simulated using a sectional representation. Thirteen aerosol species are modeled (four secondary organic aerosols, primary organic aerosol, primary elemental carbon, crustal material, water, chloride ion, sodium ion, ammonium ion, nitrate ion, and sulfate ion) and ten size sections are used varying in size from 40 nm to 40 μm (Table 1). A two-section approach in CAMx is three times faster than the 10-section algorithm used in PMCAMx and the present study. However, the assumption that all secondary PM is in the first section (corresponding roughly to the accumulation mode) introduces significant uncertainty to the model predictions [Morris et al., 2005a].
Table 1. Aerosol Size Sections Used in PMCAMx
Diameter Range, μm
 As is typical in CTMs, an operator-splitting approach is used, with each process simulated separately for each time step. The time step changes to ensure numerical stability for horizontal advection, and typically ranges from five to fifteen minutes. A smaller time step is used within individual processes, such as condensation to obtain an accurate solution. The simulation order of the processes is: emissions, horizontal advection, vertical advection, vertical diffusion, horizontal diffusion, wet deposition, gas phase chemistry, and finally the aerosol processes: nucleation, coagulation, condensation, secondary organic aerosol growth, and aqueous phase chemistry. For nucleation, the ternary H2SO4-H2O-NH3 nucleation rate parameterization of Napari et al.  is used in the model. The nucleated mass is assigned to the first size section. The coagulation equation [Seinfeld and Pandis, 2006] is solved following the Tambour and Seinfeld  and Gelbard et al.  approach using a high-resolution distribution obtained by subdividing each section of the original distribution into three sections. The coagulation algorithm conserves volume and calculates the integrals in the coagulation equation iteratively. The equilibrium model employed in this work for solving inorganic aerosol condensation has been described by Capaldo et al. . The amount of each species transferred between gas and aerosol phases is determined by bulk aerosol thermodynamics using ISORROPIA [Nenes et al., 1998], and is distributed over the aerosol size distribution by using a weighting factor for each size section based on the effective surface area (condensational sink) of each size section [Pandis et al., 1993; Lurmann et al., 1997]. Equilibrium is assumed between the gas and organic aerosol phase on the basis of the Secondary Organic Aerosol Model (SOAM) II of Strader et al.  as implemented by Koo et al. . The condensable organic vapors are partitioned between the gas and aerosol phases on the basis of their volatility, the aerosol size distribution, and composition assuming a pseudo-ideal solution [Koo et al., 2003]. Four lumped condensable organic gas species (CG1–CG4) are used as secondary organic aerosol (SOA1–SOA4) precursors. CG1 and CG2 correspond to the low and high yield products, respectively, of the photo-oxidation of toluene and the other aromatics. The oxidation products of paraffins, anthropogenic olefins, and cresol are assigned to CG3. Finally, CG4 is formed from the oxidation of biogenic olefins. The temperature dependence of the saturation concentrations is described by the Clausius-Clapeyron equation. Additional details are given by Gaydos et al. .
2.1. Gas Phase Chemistry
 The chemical mechanism used is based on the Carbon Bond mechanism 4 (CBM-IV) [Gery et al., 1989]. The mechanism includes 100 reactions for 60 species (35 gas species, 12 radicals, and 13 aerosol species). The gas phase chemistry reactions are numerically integrated using smaller time steps than the transport processes. The chemistry solver used is the CMC fast solver developed by Environ . The CMC solver uses the steady state approximation for fast reacting species (radicals). The slower reacting (state) species are separated into two groups. The differential equations for fast state species (with chemical lifetimes of seconds to a few minutes) are solved using a second-order implicit Runge-Kutta method, while those for slow state species (with chemical lifetimes longer than a few minutes) are solved explicitly. The chemical equation solver has been evaluated against benchmark numerical solutions [Environ, 2003] and the agreement for all species was excellent in all tests.
 Two modifications were implemented in the gas phase chemistry in order to avoid the unrealistic high amounts of HNO3 that the model used to produce [Gaydos et al., 2007]. First, the rate coefficient of the reaction of N2O5 and H2O was changed from a constant to a function of temperature [Atkinson et al., 1986]. The value used is:
based on the work of Wahner et al.  and Dimitroulopoulou and Marsh . In addition, the numerical accuracy of the solution of the system of algebraic equations used for the calculation of the pseudo-steady state species (N2O5 and NO3) was increased. Small errors in the numerical treatment of this system of reactions can sometimes result on unrealistically high HNO3 nighttime production rates. The heterogeneous formation of nitric acid is not simulated explicitly in the current version of the model. The effects of this simplification will be discussed in the model evaluation section.
 The change in concentration within or below a cloud due to precipitation is parameterized using a scavenging coefficient, Λ: = −Λci. The scavenging coefficient is calculated separately for gases and particles, on the basis of Seinfeld and Pandis . The mass transfer coefficient below the clouds depends on the droplet diameter and falling speed, which are calculated on the basis of the empirical estimates of Scott  and modified to better agree with the data provided by Seinfeld and Pandis . Within a cloud, Henry's Law is used to partition the total gas concentration of a species between the aqueous and gas phases. For the case of aerosols, all aerosol particles are assumed to be within the cloud water. Below the cloud layer, the collection efficiency depends on the particle diameter, dp [Seinfeld and Pandis, 2006].
3. Model Application
 PMCAMx is applied in the eastern United States for approximately a month during the four seasons of the year (July 2001, October 2001, January 2002, and April 2002). These four months have been selected to be representative of each season allowing us to reduce the computational cost of the overall effort and to concentrate on the scientific issues resulting from the model evaluation. The simulations start in relatively clean periods and the first two days of each simulation have been excluded from the analysis to limit the effect of the initial conditions on the results. The values of the organics, elemental carbon, nitrate, sulfate and ammonium concentrations at the boundaries of the domain were chosen on the basis of measurements taken in sites close to the boundaries of the model domain by the IMPROVE and STN monitoring networks during the periods of interest. The constant boundary conditions values for the above species for each season represent the background concentrations for the eastern United States and are shown in Table 2.
Table 2. Concentrations of Aerosol Species in Each Cell in the Boundaries of the Model Domain
Concentration, μg m−3
 The modeling domain covers a 3492 × 3240 × 6 km region in the eastern United States with 36 × 36 km grid resolution. Fourteen vertical layers are used for July and sixteen vertical layers for the other three months. The vertical extent of the model domain remains constant at 6 km for all seasons, but additional resolution for the nonsummer months was needed for the meteorological simulations. Inputs to the model include horizontal wind components, temperature, pressure, water vapor, vertical diffusivity, clouds, and rainfall, all created using the meteorological model MM5 [Grell et al., 1995] by the Lake Michigan Air Directors Consortium (LADCO). MM5 was exercised by Olerud and Sims  over the same domain using a science configuration derived from a substantial number of sensitivity and diagnostic experiments. Rigorous performance testing of the MM5 fields [Abraczinskas et al., 2004; Olerud and Sims, 2004] showed that the dynamic and thermodynamic fields generated by MM5 were sufficient for use in CTM simulations.
 The emission inventory used is Midwest Regional Planning Organization's Base E inventory [LADCO, 2003]. This inventory (including primary carbonaceous material) is based on the U.S. EPA's 1999 National Emissions Inventory (version 2.0) [U.S. EPA, 2001]. Temporal emission profiles for electric generation units were developed on the basis of continuous emission monitor data. The emission rate of organic PM in the inventory for the whole modeling domain is 1.28 ktons d−1 during weekdays and 1.25 ktons d−1 during weekends. Lane et al.  combined the predictions of a source-resolved version of PMCAMx with the results of source receptor analysis based on organic tracers for Pittsburgh and Atlanta to evaluate the OC and EC emission inventories. One of the conclusions of this study was that the EC emission rate by diesel engines was seriously overestimated, with nonroad emissions being the major problem. Following the suggestions of Lane et al.  the diesel EC emissions have been corrected here, resulting to an overall reduction of the total EC emissions by almost a factor of two. The corresponding emission rates of EC used in this study are 0.54 tons d−1 during weekdays and 0.45 ktons d−1 during weekends. Spatial and temporal improvements have been made to develop the Base E emissions inventory: the inventory uses the Carnegie Mellon University ammonia emissions [Pinder et al., 2004], MOBILE6 for vehicular sources [U.S. EPA, 2003], BIOME3 for biogenic emissions [Wilkinson and Janssen, 2001], and has reduced dust (soil-dust and road-dust) emissions. A different emission inventory is used for weekdays, Saturdays, and Sundays of each season.
4. Overview of Model Predictions
 The predicted average concentrations of PM2.5 sulfate, nitrate, ammonium, organic mass, elemental carbon, and total PM2.5 mass over the period of July 2001 are shown in Figures 1a, 2a, 3a, 4a, 5a, and 6a, respectively. The highest predicted concentrations for most inorganic PM2.5 species are over the Midwest during this period. High PM2.5 is also predicted for the New York and Boston areas. Average sulfate and ammonium levels are the highest of the year (up to 11.4 and 4.8 μg m−3, respectively, in Illinois, Indiana and Ohio). The elemental carbon and the organic mass peak values (3 and 7.5 μg m−3, respectively) are in the northeast United States and in the Midwest, respectively. Nitrate concentrations are generally low (less than 1 μg m−3 in most areas) with the higher values in Indiana, eastern Ohio, and Philadelphia area (up to 1.7 μg m−3) which is consistent with the observations. Overall, sulfate is predicted to account on average for around 40% of total PM2.5, followed by OM (22%), ammonium (13%), EC (4%), and finally nitrate (3%). The remaining 18% is crustal material and metal oxides.
 The results for October 2001 are shown in Figures 1b, 2b, 3b, 4b, 5b, and 6b. For this period, PM2.5 concentrations are predicted to be high in the New York City area (up to 28.6 μg m−3). Nitrate concentrations of 1.7 μg m−3 are predicted for the Midwest while the organic mass and elemental carbon are high both in the east coast and in the south (up to 11.4 and 2.6 μg m−3, respectively). The average PM2.5 mass in the entire domain consists of 26% sulfate, 26% organics, 11% ammonium, 6% nitrate, 5% elemental carbon and 26% crustal material and metal oxides. It is important to note here that ammonium is overpredicted in most of the model domain during this period and observations show that ammonium accounts only for 7% of total PM2.5 mass in the eastern United States during October 2001.
 The average predicted concentrations during January 2002 for sulfate, nitrate, ammonium, organic mass, elemental carbon, and PM2.5 mass are depicted in Figures 1c, 2c, 3c, 4c, 5c, and 6c, respectively. The concentrations of those species peak in the northeastern coastal areas (Baltimore to Maine) with values in the range of 30 μg m−3. These high PM2.5 concentrations are due to the accumulation of organic PM (up to 12 μg m−3 in those areas). Though, the organic mass is overpredicted by the model during January (the bias is 0.92 μg m−3) and consequently the predicted PM2.5 concentration is a little higher than the observed in those areas. Nitrate has the highest concentration of the year in that area, up to 2.8 μg m−3. In the southeast United States, the levels of sulfate and ammonium concentrations are also high (5 and 2.3 μg m−3, respectively). Overall, the PM2.5 mass is predicted to consist of 25% organics, 19% sulfate, 12% nitrate, 10% ammonium and finally 5% elemental carbon.
 The highest concentrations for PM2.5 during April 2002 are found in industrial areas (Baton Rouge, LA; Chicago, IL; New York, NY, etc) (Figure 6d) which is consistent with observations. In the northeast coast PM2.5 has the lowest average value of the year (less than 20 μg m−3). On the other hand, PM2.5 mass is predicted to be up to 27 μg m−3 in the state of Louisiana, which is the highest of the year in that area. In agreement with the measurements, the model predicts high nitrate concentrations in the Midwest (up to 3.2 μg m−3 in the lake of Michigan), while the elemental carbon is high mostly in the northeast coast (New York City) and the organic mass in the south (Baton Rouge). High values of sulfate (5 μg m−3) and ammonium (3.2 μg m−3) are predicted in the Midwest. In the entire model domain, sulfate accounts for 29% of total PM2.5 mass, followed by organic mass (22%), ammonium (13%), nitrate (11%) and elemental carbon (4%), while the crustal material accounts for the remaining 21%.
5. Model Performance Evaluation
 The model results are compared to daily average measurements from the U.S. EPA's STN monitoring network [U.S. EPA, 2002] and the IMPROVE network [IMPROVE, 1995]. Plots of the predicted values versus the measured values are shown in Figures 1–6 for the four seasons of the year (July, October, January and April, respectively) The fractional bias (FBIAS), fractional error (FERROR), mean bias (BIAS), and mean error (ERROR) were calculated (Tables 3–6) to assess the model performance:
where Pi is the predicted value of the pollutant concentration in a specific location, Oi is the observed value of the pollutant in the same location, and N is the total number of the predictions used for the comparison. The measurements (observed values) used for the model evaluation were taken from 100 sites (46 from the STN network and 54 from the IMPROVE network). The frequency of these measurements varies from 1 to 3 days resulting in an annual total of approximately 3,500 data points for each fine PM component. Hourly data from the Pittsburgh Supersite [Wittig et al., 2004] were also used.
Table 3. Comparison of PMCAMx Predictions With Daily Average Measurements During the Period of July 2001
 The performance of the model for sulfate is encouraging, especially during the periods of January and April 2002 (Tables 3–6). During July, in contrast to the other seasons, the model slightly underpredicts the concentration of sulfate (the overall fractional bias is −0.04). During October, PMCAMx overpredicts sulfate in most of the model domain. During the same period the model overpredicts the ammonia levels in the same areas. These high concentrations of ammonia increase the cloud pH to values over 5 accelerating the oxidation of SO2 to sulfuric acid by its reaction with dissolved ozone [Seinfeld and Pandis, 2006]. In all sites where ammonium was seriously overpredicted (Figure 3b), sulfate was also seriously overpredicted (Figure 1b) For example, during 10 October in Maryland the total (gas and aerosol) predicted ammonia is quite high (5.1 μg m−3) resulting in a simultaneous overprediction of sulfate (4.4 μg m−3 instead of the measured value of 1.5 μg m−3) and of ammonium (2.2 μg m−3 instead of 0.5 μg m−3).
 Errors in the timing of the predicted rainfall were identified as a major cause of some of the discrepancies between model predictions and measurements. For example, during 4 October the model underpredicts sulfate together with most of the major fine PM components in several stations in the northeast (Pittsburgh, Baltimore, etc.). Comparisons of the model predictions with continuous measurements can shed light to the causes of such discrepancies. The model performs quite well in Pittsburgh until the midnight of 3 October (Figure 7). Then, according to the input meteorological fields it rains in parts of the northeast, something that did not happen in reality. A significant fraction of the fine PM is removed from this part of the modeling domain and PMCAMx underpredicts its concentration for a couple of days. Sulfate is significantly underpredicted by the model during 4 October in Pittsburgh (3.3 instead of 7 μg m−3), together with the total PM2.5 (11.6 μg m−3 while the measured value was 26.7 μg m−3). Use of the measured instead of the predicted rainfall fields can correct these problems and improve the CTM model performance during such periods.
 The typical diurnal cycles are reproduced relatively well by the model during July and January (Figures 8 and 9) . During the summer the gas phase pathway dominates causing a peak in the afternoon. On the other hand during the winter the aqueous phase pathway is quite important and there is little diurnal variation of the average concentrations. PMCAMx is able to reproduce these features of the observed sulfate concentrations together with the average concentrations.
 PMCAMx has been used by Gaydos et al.  to simulate the July 2001 period. The simulation resulted in nitrate predictions that were a factor of five higher than the observations on average, and an order of magnitude higher on one third of the modeled days. In this work, after the improvement in the nighttime nitric acid chemistry this problem has been corrected.
 During the summer, 80% of the nitrate predictions diverge less than 30% or 0.5 μg m−3 from the measurements. While the model is successful in reproducing the relatively low nitrate levels in most of the domain, it tends to underpredict the nitrate levels. Considering that routine nitrate filter-based measurements are uncertain by roughly ±0.5 μg m−3, this discrepancy at low nitrate concentrations could be partially due to the measurements. However, comparisons of the predicted and measured diurnal nitrate profiles (Figure 8) suggest that the model underpredicts the nitrate mainly during the night (at least in Pittsburgh). This discrepancy is not the result of errors in the partitioning of the available nitric acid (the total ammonia is predicted reasonably well and the sulfate is actually underpredicted) but to an underprediction of the total nitric acid. The lack of an explicit description of the heterogeneous formation of nitric acid during the night biases the model predictions low.
 There are several high nitrate concentration (above 2 μg m−3) measurements in the July data set. While a few of them are reproduced by PMCAMx, a number of them in sites such as Oakland, MI, Baltimore, MD, etc. are seriously underpredicted. Nitrate formation is limited by ammonia availability in those areas and in a lot of these cases ammonium is also underpredicted. As an example, during 18 July in Michigan the predicted total ammonia is 4 μg m−3. PMCAMx calculates that practically all of it should be in the particulate phase, resulting in predicted ammonium concentration of 4 μg m−3 instead of the measured 6.7 μg m−3. This modest error in ammonia availability results in significant error in the predicted nitrate (0.3 μg m−3 instead of the measured value of 4.4 μg m−3). Serious sulfate overprediction in a few cases also results in serious nitrate underprediction. The “extra” sulfate transfers all of the available NH3 to the particulate phase as ammonium sulfate or bisulfate resulting also in an underprediction of ammonium nitrate. For instance, during 21 July in St. Louis, MO sulfate is significantly overpredicted (22.8 instead of 10.8 μg m−3) resulting in an overprediction of ammonium (6.9 instead of 3.6 μg m−3) and in a significant underprediction of nitrate (0.2 instead of the measured value of 2.5 μg m−3).
 During October, January, and April the model tends to underpredict the concentration of nitrate compared to measurements in the mostly urban STN sites and to overpredict nitrate compared to the rural IMPROVE sites (Tables 4–6). The comparison of the predicted and measured diurnal variation of nitrate in Pittsburgh (Figure 9) suggests that the underprediction in the urban areas is probably due to the problems in describing the heterogeneous nighttime nitrate formation. Without the use of temperature-dependent gas phase reaction rate constant the model tends to seriously overpredict nitrate in the urban areas. The reasons for the small overprediction of nitrate in rural areas include overestimation of ammonia emissions (this is the case for the fall) and potential experimental errors.
 Ammonium predictions can be compared only against the STN measurements, because there are no measurements from the IMPROVE network. The ammonium predictions are quite sensitive to the ammonia emissions inventory, the predicted sulfate concentrations and secondarily the nitrate and chloride, sodium, etc., levels. The ammonium predictions during July, January, and April show little bias (Tables 3, 5, and 6) indicating that the CMU inventory [Pinder et al., 2004] used in this study captures well the seasonal variation of the corresponding emissions.
 During October 2001 the model overpredicts the concentrations of ammonium in most areas (Figure 3b and Table 4). These problems are probably related to the fall ammonia emission inventory. During this period the majority of emissions are from field applied manure [Pinder et al., 2004]. Uncertainties in farming practices, timing of manure application, the quantity of manure, and its volatilization are significant [Pinder et al., 2004]. Therefore the emission inventory of ammonia over agricultural areas in the fall should be reexamined, as the ammonia emissions in such areas are probably overestimated. These conclusions are consistent with the evaluation of the current ammonia emissions inventory by Pinder et al.  using the total (gas and aerosol) ammonia as an evaluation metric.
5.4. Organic PM
 The results from the evaluation of the model concerning the organic PM are expressed in μg m−3 (organic mass). A multiplier of 1.4 is applied to convert measured OC to organic mass (OM). The model predictions for the organic mass are generally in reasonable agreement with the measurements in most areas and seasons. This is rather surprising given the existing uncertainties in the primary OM emission inventory, the formation of SOA by anthropogenic and biogenic precursors included in PMCAMx, and also the fact that SOA formation by sesquiterpenes and isoprene and processes like oligomerization are not simulated at all. However, despite the above weaknesses the model captures relative well not only the daily average concentrations of OM, but also its average diurnal variation during all seasons (Figures 8 and 9) in Pittsburgh.
 During July, the model significantly underpredicts OM compared to the mostly urban STN stations and slightly underpredicts compared to the mostly rural IMRPOVE network measurements (Table 3). The STN network measurements have not been corrected for positive sampling artifacts, and this could explain part of the discrepancy. There are probably a number of other problems in both the primary and secondary OM but the different errors are probably offsetting each other. The agreement of the diurnal profiles for Pittsburgh (Figure 8) suggests that at least the total primary and secondary contributions (given their different timing) are close to reality.
 During January, the model overpredicts the organic mass especially in urban areas in the northeast United States (New York, Baltimore, Pittsburgh, etc.). In these areas, according to the model more than 90% of the OM is primary, so the overprediction is related to errors in the OM emissions. The analysis of Lane et al.  suggested that the wood burning emissions (residential fireplaces, wood stoves, and wood-fired boilers) have probably been overestimated for this area during the winter.
 During October the OM predictions by PMCAMx show on average relatively little fractional bias in both urban and rural areas (Table 4). However, there is a small tendency toward overprediction in the northeast and toward underprediction in the south. During April, the model predicts high levels of organic mass in south Louisiana (Figure 4d), which are not in agreement with the measurements from this area (STN stations). The industrial OM emissions are probably overestimated here.
 The primary organic aerosol is predicted to account on average for 60% (in July) to 80% (in January) of the predicted OM (Figure 10). The predicted contributions in the spring and fall are between the July and January values. Therefore, in all four seasons most of the organic mass is predicted to be primary. The emission rate of organic PM in the inventory for the whole modeling domain is 1.28 ktons d−1. Even if there are probably significant errors in the contributions of individual sources, the total emissions based on this evaluation appear to be quite reasonable. There is little variation of these assumed emissions from day to day and this simplification is probably at least partially responsible for the scatter in Figure 4.
 The model assumes that the primary and the secondary organic compounds participate in the same solution. The case of two separate organic phases (one for the primary and one for the secondary OM) has also been examined. This resulted in relatively small changes of the predicted OM with the most significant change in the northeast United States (on average a 10% decrease in OM). The changes were a lot smaller in the south (a 3% decrease in OM on average) where most of the OM is predicted to be biogenic SOA.
5.5. Elemental Carbon
 The model during July tends to overpredict EC with a mean bias of 0.24 μg m−3 when compared to the STN measurements and is unbiased when compared to the IMPROVE network. Overall, the model still performs the worst, in all four months compared to the STN measurements (urban areas) where the NIOSH TOT protocol [National Institute of Occupational Safety and Health (NIOSH), 1999] is used for measuring EC. The comparison with the IMPROVE measurements (rural areas) is better, where the TOR protocol [Chow et al., 1993] is used to analyze the filters. Overall the performance of the model is quite encouraging suggesting that the EC emission corrections suggested by Lane et al.  bring the total emissions closer to reality. The predicted EC diurnal variation is quite consistent with the measurements in Pittsburgh for both July and January (Figures 8 and 9). The model predicts that the elemental carbon does not grow to the coarse mode (PM2.5–10 EC is predicted to be less than 0.05 μg m−3 during all seasons) which is in general consistent with the findings of Venkataraman and Friedlander  and Cabada et al. .
5.6. PM2.5 Concentration
 PM2.5 mass is equal to the sum of the above components, with the addition of crustal material and metal oxides. As with most of the individual species, the model predictions generally agree with the measurements during the four seasons of the year (Tables 3–6). During July 2001, the model tends to underpredict the PM2.5 levels (fractional bias of −0.18). Part of this underprediction is due to the measurements of water by the FRM PM2.5 method [Rees et al., 2004]. In October despite the agreement between the model and the observations in most of the measurement stations, there are areas where the model over (northeast coast) or underpredicts (Midwest) the total PM2.5 mass. In winter the model predictions generally agree with observations, though, the model overpredicts the PM2.5 mass in some areas (Bronx, Queens, Pittsburgh, etc.). During April the good performance of the model for most species leads to a good performance for the overall PM2.5 mass too (Table 6).
 The model predictions for total PM2.5 mass are also compared to hourly measurements taken during the Pittsburgh Air Quality Study (PAQS) [Wittig et al., 2004]. Although the model predictions generally agree with the observations in Pittsburgh, there are periods where the model significantly under or over predicts the concentration of all species (Figure 7). During the periods of 18–19 and 22–24 July and 3–5 October the model underpredicts the concentration of all species. The opposite phenomenon occurs during 24 October and 13 April. This systematic and simultaneous over or under prediction in all species indicating that there are periods where the model fails to capture the timing of rainfall. During January 2002 there is a systematic overprediction of the total PM2.5 mass in Pittsburgh (Figure 7). This is observed also in other areas in the northeast United States. This overprediction is the result of the overprediction of the organic PM discussed in previous sections but also from the high predicted crustal material concentrations in these areas in the winter. For instance in Pittsburgh, PM2.5 mass is predicted to consist of 6.1 μg m−3 crustal materials and metal oxides while the measured PM2.5 mass contained only 2 μg m−3.
5.7. Overall Performance
 Four levels of model performance goals/criteria, from problematic to excellent, for fractional bias and error have been proposed to evaluate model performance [Morris et al., 2005b]. The results of this evaluation of PMCAMx using the above criteria are shown in Table 7. The organic mass performance is average while the sulfate performance varies from average to good for all seasons. Nitrate underpredictions during the summer remain a challenge. PMCAMx performance for ammonium is excellent in the spring and good in the winter. Though, the ammonia model predictions are not as accurate during the summer and have serious problems during autumn when there are fundamental problems with the ammonia emission inventory. Elemental carbon concentrations are reproduced well by the model indicating that the zeroth-order correction proposed by Lane et al.  is in the right direction and could be used to guide the reevaluation of the corresponding inventory. Finally, the model reproduces the total PM2.5 concentrations well in all four seasons except winter where the performance is considered average.
 Wet deposition is an important removal process for particles, so precipitation is expected to have a significant effect on aerosol concentrations resulting in an over or underprediction of the aerosol concentrations from the model. In order to investigate the effect of precipitation and the wet deposition parameterization on the results of this work a sensitivity analysis was conducted for the October 2001 and April 2002 simulations increasing the scavenging coefficients by 50%. Wet deposition changes tended to have rather modest effects on the average PM2.5 mass concentrations, which is consistent with the findings of Dawson et al.  for the months of July 2001 and January 2002. As expected, PM2.5 concentrations decreased with increased precipitation rate, and the sensitivity to this change was larger in July (16.7 ng m−3 %−1) than in January (1 ng m−3 %−1), October (1.6 ng m−3 %−1) and April (2.5 ng m−3 %−1).
 A three-dimensional chemical transport model, PMCAMx, is presented and applied to four periods (July 2001, October 2001, January 2002, and April 2002) in the eastern United States. The current performance of PMCAMx for the major aerosol components and PM2.5 during all seasons is encouraging. During July the model shows little bias for the concentration of most PM2.5 species. The only exception is nitrate which is underpredicted by the model. In January, nitrate is also underpredicted by the model (fractional bias of −0.06) while the model overpredicts the concentration of the rest of the PM2.5 species by 0.03 μg m−3 (for ammonium) to 0.71 μg m−3 (for organic mass). The model has the best performance statistically during April. The model performs the worst during October 2001. The concentration of sulfate is overpredicted with a mean bias of 0.71 μg m−3, while the average of the measurements is 2.36 μg m−3. Though, the biggest problem is the overprediction of ammonium with a fractional bias of 0.78. This overprediction indicates that the emission inventory of ammonia over agricultural areas this season should be reexamined. The rest of the PM species are predicted with satisfactory accuracy (fractional bias for nitrate, organic mass and elemental carbon is −0.04, −0.07 and 0.21, respectively).
 The improvement of the model performance for July compared to Gaydos et al.  was mainly due to the improvement of the description of the nighttime nitrate chemistry. Additional improvements for nitrate especially during the summer and spring require better descriptions of the heterogeneous nighttime nitric acid formation processes.
 The model performance for EC has also improved after following the recommendations of Lane et al.  for the reduction of the diesel EC emissions (especially the nonroad contribution). The used average EC emission rate for the whole domain is approximately 0.55 ktons d−1 (0.45 ktons d−1 during the weekends). As a result, 40% of the predictions diverge less than 30% from the measurements. A slight overprediction occurs during all seasons (the fractional bias is between 0.16 and 0.29), while the model used to predict the EC concentrations two or three times higher than the measurements before the change [Lane et al., 2007]. This overprediction comes mainly from the comparison with the STN measurements (urban areas), where the NIOSH TOT protocol [NIOSH, 1999] is used for measuring EC. The comparison with the IMPROVE measurements (rural areas) is better, where the TOR protocol [Chow et al., 1993] is used to analyze the filters.
 The model performance for organic PM is average during all four seasons of the year. The model is able to reproduce not only the daily average concentrations of OC but also its diurnal variation patterns in Pittsburgh. The predicted organic PM during the summer is on average 60% primary and 40% secondary. The average predicted secondary contribution to organic PM drops to around 20% during the winter. The discrepancies during the colder periods appear to be due to an overprediction of the wood-burning emissions in the inventory used.
 The evaluation suggests that use of the measured rainfall fields instead of the predicted by MM5 or other meteorological models would result in improvements of the model performance.
 This research was supported by the National Science Foundation (ATM-336296). V. Karydis and A. Tsimpidi are supported by a scholarship from the Molina Center for Energy and the Environment.