We run the WRF-Chem model version 3.4 using a 12 km resolution grid with 444 rows, 336 columns, and 28 vertical layers. The modeling domain (see Fig. 1) covers the contiguous US, southern Canada, and northern Mexico. Previous studies (e.g., Appel et al., 2012; Yahya et al., 2014) have used 34 vertical layers; our choice of 28 vertical layers represents a tradeoff between vertical grid resolution and computational expense.

Within WRF-Chem, we use the Regional Atmospheric Chemistry Mechanism (RACM) (Stockwell et al., 1997) for gas-phase reactions and the Modal Aerosol Dynamics for Europe (MADE) (Ackermann et al., 1998) module for aerosol chemistry and physics. RACM and MADE were selected because of their relatively modest computational expense; at the time of this study, alternatives to RACM/MADE are impractical for large-scale simulations such as ours. We use the volatility basis set (VBS) (Ahmadov et al., 2012) to simulate formation and evaporation of secondary organic aerosol (SOA). The VBS approach differs from other SOA parameterizations in that it assumes that primary organic aerosol (POA) is semi-volatile. Meteorology options are set as recommended by the WRF user manual (Wang et al., 2012) and the WRF-Chem user manual (Peckham et al., 2012) for situations similar to those studied here. Table 1 summarizes the model options and inputs used. See supporting information for additional details.

We use results from the MOZART global chemical transport model (Emmons et al., 2010) as processed by the MOZBC file format converter (available at: http://web3.acd.ucar.edu/wrf-chem) to provide initial and boundary conditions for chemical species. Because the MOZBC boundary conditions for unclassified PM2.5 are unrealistic for the southeastern edges of the modeling domain – their use results in substantial PM2.5 overpredictions in the southeastern US – we set all initial and boundary concentrations to zero for unclassified PM2.5. As in Ahmadov et al. (2012), owing to uncertainty in secondary organic aerosol (SOA) concentrations over the open ocean, we assume that initial and boundary concentrations of SOA are zero. Data from the National Centers for Environmental Prediction (NCEP) Eta model (UCAR, 2005) provide meteorological inputs, boundary conditions, and, for the four-dimensional data assimilation (FDDA) employed here, observational “nudging” values.

We use the 2005 National Emissions Inventory (NEI) (US EPA, 2009) to estimate pollutant emissions. The NEI includes emissions from area, point, and mobile sources for year 2005 in the US, year 2006 in Canada, and year 1999 in Mexico. We use the model evaluation version of the NEI, which also includes hourly Continuous Emission Monitoring System (CEMS) data for electricity-generating units, hourly wildfire data, and biogenic emissions from the BEIS model (Biogenic Emission Inventory System; Schwede et al., 2005), version 3.14.

We prepare pollutant emissions at 12 km spatial resolution using the Sparse Matrix Operating Kernel Emissions (SMOKE) program (Houyoux and Vukovich, 1999), version 2.6, as bundled with the NEI data (available at: http://www.epa.gov/ttn/chief/emch/index.html), then we convert the emission files output by SMOKE to WRF-Chem format and apply a plume-rise algorithm (ASME, 1973, as cited in Seinfeld and Pandis, 2006) to estimate the mixing height of elevated emission sources and wildfires. Source code for the file format conversion and plume-rise program is available at https://bitbucket.org/ctessum/emcnv.

We simulate atmospheric pollutant concentrations for the period from 1 January through to 31 December 2005. We choose the year 2005 because at the time this study was performed it was the most recent year for which emissions data were available. For logistical expediency, we separate the year into eight independent model runs, each approximately 1.5 months in length plus a discarded 5-day model spin-up period. We run the simulations on a high-performance computing system consisting of 2.8 GHz Intel Xeon X5560 “Nehalem EP” processors with a 40 Gbit QDR InfiniBand (IB) interconnect and a Lustre parallel file system. Using 768 processors, each 1.5-month model run takes ∼19 h to complete (∼ 13 processor years for each annual model run).

We compare WRF-Chem wind speed, air temperature, relative humidity, and precipitation predictions to data from the US Environmental Protection Agency (EPA) Clean Air Status and Trends Network (CASTNET) observations. We compare modeled ground-level concentrations of total PM2.5 to EPA Air Quality System (AQS) observations (US EPA, 2005) using 24 h average data (EPA parameter code 88101) and using the less extensive hourly measurement network (EPA parameter code 88502), which allows us to compare modeled vs. measured diurnal profiles. We compare WRF-Chem predictions of O3 to measurements from the AQS (EPA parameter code 44201) and CASTNET networks. We compare the predictions of PM2.5 subspecies to observation data from the EPA's Chemical Speciation Network (CSN) (US EPA, 2005) (formally called Speciation Trends Network (STN)) for organic carbon (OC, parameter code 88305), elemental carbon (EC, code 88307), particulate sulfate (SO4, code 88403), particulate nitrate (NO3, code 88306), and particulate ammonium (NH4, code 88301). We additionally compare predictions to data from the Interagency Monitoring of Protected Visual Environments (IMPROVE) network (University of California Davis, 1995) for particulate OC (code 88320), EC (code 88321), sulfur (code 88169), and NO3 (code 88306); and to CASTNET observations for particulate SO4, NH4, and NO3. WRF-Chem outputs organic aerosol (OA) concentrations, but methods for measuring organic aerosol only quantify OC. OC comprises a variable fraction of OA, but it is common to assume an OA : OC ratio of 1.4 (Aiken et al., 2008). Therefore, we divide WRF-Chem OA predictions by a factor of 1.4 for comparison with OC measurements. Finally, we compare WRF-Chem predictions of gas-phase sulfur dioxide (SO2) and nitrogen dioxide (NO2) to AQS observations. We remove from consideration those stations with ≥ 25 % missing data relative to the number of scheduled measurements during the simulation period. The fractions of excluded data for each type of comparison are in the Supplement.

WRF-Chem, as configured here, outputs instantaneous concentrations at the start of each hour, whereas the observation data are reported as hourly or daily averages. WRF-Chem calculates grid-cell-average concentrations, whereas observations generally represent concentrations at specific locations.

We compare measured and modeled values pair-wise at each time of measurement in the grid cell containing each measurement station. The 24 h average measurements are compared to the average of the modeled (hourly instantaneous) values within the same period. Comparisons are only made with observations that occur within the first (nearest to ground) model layer (height: ∼ 50–60 m). The source code for the program used to extract and pair model and measurement data is available at https://bitbucket.org/ctessum/aqmcompare.

In addition to reporting annual average model performance for the entire model domain, we also disaggregate results spatially and temporally. We evaluate performance using two spatial approaches. First, we use four regional subdomains: Midwest, Northeast, South, and West (basis: US Census regions (US Census Bureau, 2013); see Fig. 2). Second, we evaluate urban vs. rural (i.e., not urban) locations, also as defined by the US Census (US Census Bureau, 2014). CSN monitors tend to be placed in urban areas (85 % of 186 monitors are urban), whereas IMPROVE monitors tend to be placed in protected rural areas (10 % of 122 monitors are urban). All 67 monitors in the CASTNET network are in rural locations. We also split the analysis into four seasons: winter (January–March), spring (April–June), summer (July–September), and fall (October–December). Employing these time periods allows us to compare against previously published results (Appel et al., 2012).

After matching all measured values with their corresponding modeled values, and averaging modeled and measured values across the appropriate time period, we calculate metrics shown in Eqs. (1)–(8): MB=1n∑i=1n(Mi-Oi),ME=1n∑i=1n|Mi-Oi|,NMB=∑i=1n(Mi-Oi)∑i=1nOi×100%,NME=∑i=1n|Mi-Oi|∑i=1nOi×100%,MFB=1n∑i=1n2(Mi-Oi)Mi+Oi×100%,MFE=1n∑i=1n2|Mi-Oi|Mi+Oi×100%,MR=1n∑i=1nMiOi,RMSE=∑i=1n(Mi-Oi)2n, where i corresponds to one of n measurement locations, M and O are time-averaged modeled and observed values, respectively, MB is mean bias, ME is mean error, NMB is normalized mean bias, NME is normalized mean error, MFB is mean fractional bias, MFE is mean fractional error, MR is model ratio, and RMSE is root-mean-square error. We additionally calculate the slope (S), intercept (I), and squared Pearson correlation coefficient (R2) of a linear regression between modeled and measured values.

Each metric provides a useful and distinct evaluation of model performance. In general, metrics with “bias” in the name evaluate the accuracy of the model, whereas metrics with “error” in the name incorporate both precision and accuracy. Metrics that are in normalized or fractional form tend to emphasize errors where measured and observed values are relatively small, whereas non-normalized metrics tend to emphasize errors where measured and observed values are relatively large. We mainly focus here on MFB and R2 to evaluate performance as they facilitate direct comparisons among pollutants. Results for all combinations of time periods, measurement networks, spatial subdomains, and metrics are in the Supplement.

For O3, we calculate model performance via three model–measurement comparisons: (1) annual averages; (2) daytime-only (08:00–20:00 LT) annual averages, as in Appel et al. (2012); and (3) annual averages of daily peak concentrations, to match the epidemiological findings in Jerrett et al. (2009).

Model performance goals and criteria have been published for PM2.5 (Boylan and Russell, 2006). Goals reflect performance that models should strive to achieve; criteria reflect performance that models should achieve to be used for regulatory purposes. The goals and criteria suggested by Boylan and Russell (2006) vary with concentration: they are MFB less than ±30 and ±60 % and MFE less than 50 and 75 %, respectively, for most concentrations, but increase exponentially as concentration decreases below ∼3 µgm-3. To incorporate this aspect of performance evaluation, we calculate the fraction of observation stations for which our PM2.5 model results meet both the MFB and MFE performance goals (fG) and criteria (fC).

Figure 3 contains scatterplots comparing annual average observed and predicted values for meteorological variables and pollutant concentrations. The model tends to overpredict near-ground wind speed (Fig. 3a) and precipitation (Fig. 3d) relative to observations, whereas temperature (Fig. 3b) and relative humidity (Fig. 3c) predictions agree well with observations. Figures A2–A5 in Appendix A disaggregate model performance for meteorological variables by region (region boundaries are shown in Fig. 2) and by season; meteorological performance is relatively consistent among seasons and regions. Model–measurement comparisons provide important evidence on model performance but might overestimate model robustness for meteorological parameters because FDDA “nudges” model meteorological estimates toward observed values.

The annual average model–measurement agreement is good for total PM2.5 concentration (Fig. 3e, 94 % of measurements meet performance criteria), although the model tends to overpredict PM2.5 concentration at relatively high-concentration monitors (Fig. 3e). The model tends to generally overpredict O3 concentrations, with worse overpredictions for 24 h average concentrations (Fig. 3f) than for daily peak (Fig. 3g) and daytime average (Fig. 3f) concentrations.

Figure 4 shows the median and interquartile range for modeled and measured PM2.5 and O3 concentrations by hour of day (measurements of PM2.5 subspecies are only available as 24 h averages). For PM2.5, the model generally agrees with measurements, although on average it underpredicts concentrations at night and overpredicts during the day (Fig. 4a). For O3, on average the model overpredicts for all times of day but with a much lower fractional error during the day than during the night. For both pollutants, the model accurately captures the timing of diurnal trends, including the afternoon peak for O3 and the morning and evening peaks for PM2.5. As a result, when comparing the three averaging-time metrics for O3, we observe better model performance for the annual average of daily peak concentration (MFB = 11 %) and of average daytime concentration (MFB = 12 %) than for the overall annual average (MFB = 23 %). For O3, the first two metrics may offer greater relevance than the third. For example, the annual average of daily peak concentrations is more strongly correlated with health effects than are annual average concentrations (Jerrett et al., 2009); and, for comparisons to the 8 h peak concentration National Ambient Air Quality Standard (NAAQS), model performance is more important during daytime than at night.

Figures 5 and 6 disaggregate results by season and by location for total PM2.5 and daytime O3, respectively; analogous results are in Figs. 7–11 for PM2.5 subspecies, in Figs. A2–A5 in Appendix A for meteorological properties, in Figs. A6 and A7 for other O3 temporal summaries, in Fig. A8 for SO2, and in Fig. A9 for NO2. Daytime and peak O3 predictive performance does not exhibit obvious patterns among seasons or regions; MFB values range from -7 to 48 % (daytime; Fig. 6) and -12 to 29 % (peak; Fig. A7). The overprediction of PM2.5 concentrations at high-concentration monitors is more prevalent in the South and in urban areas, and is less prevalent in summer than in other seasons (Fig. 5). Model–measurement correlation for total PM2.5 is higher in summer (AQS R2 = 0.64) than in fall and winter (AQS R2 = 0.20 and 0.24, respectively), but overall PM2.5 concentrations are not higher in summer. Previous research has suggested that poor PM predictive performance in winter is common among CTMs and may be attributable to difficulty in reproducing the strongly stable meteorological conditions that are responsible for high winter PM concentrations (Solazzo et al., 2012). Annual average PM2.5 predictive performance in the West (AQS R2: 0.45 (summer), 0.13 (winter)) is worse than performance in the Northeast (AQS R2: 0.70 (summer), 0.37 (winter)). In the Northeast, performance is better in the summer (R2 = 0.69) than in other seasons (R2 = 0.30–0.40). Taken together, these findings suggest that there is an opportunity for future model development for PM2.5 to focus on winter or full-year simulations rather than summer-only simulations, and on the western US or the full contiguous US rather than just the Northeast.

Figure 3i–m illustrates model performance for annual average concentrations of PM2.5 component species. In all cases, > 65 % of locations meet performance criteria for at least one of the three observation networks.

The model overpredicts particulate SO4 (CSN MFB = 34 %, IMPROVE MFB = 40 %, CASTNET MFB = 36 %) (Fig. 3i) and SO2 (MFB = 51 %) (Fig. 3n). This finding (overprediction of total sulfur) agrees with prior research for multiple CTMs (McKeen et al., 2007). Particulate SO4 prediction performance does not vary much by region; as with total PM2.5, performance is worse in winter (CSN MFB = 59 %) than in summer (CSN MFB = 10. %) (Fig. 7).

WRF-Chem as configured here performs well in predicting observed particulate NH4 concentrations, with 99 % of locations meeting performance criteria (Fig. 3j). Similar to total PM2.5, performance for particulate NH4 is worst in the urban areas in the West region (Fig. 8), where a number of monitors report relatively high measured concentrations but modeled concentrations are relatively low.

Particulate NO3 concentrations are consistently underpredicted (annual average MFB = -110 %) (Fig. 3k). Figure 9 shows that these underpredictions are more severe in some seasons and regions than in others. The best predictive performance is for the Midwest in summer (MFB =-39 %) followed by the Northeast in summer (MFB =-47 %). NO3 predictions in the West region are poor for all seasons (MFB = -148 %). As with other PM2.5 species, there is an opportunity for future development and evaluation of models for particulate NO3 prediction to focus on seasons and regions other than summer in the Northeast. Predictions of gas-phase NO2 (Fig. 3o) agree relatively well with observations (MFB = 4 %) but, as with other species, the model tends to overpredict NO2 concentrations in areas where measured concentrations are relatively high. This effect is especially prominent in the West and in urban areas (Fig. A9).

Model–measurement agreement for EC concentrations is relatively good (Fig. 3l), with 96 % of monitor locations meeting performance criteria. As with other comparisons, for EC the model tends to overpredict concentrations for monitors with relatively high concentrations, especially in urban areas (Fig. 10).

Model predictions of OC concentrations (Fig. 3m) are biased low compared to CSN (MFB = -55 %) but agree relatively well with IMPROVE (MFB = 15 %). Mean bias values given here are within the range of values reported by a previous publication using the VBS SOA formation mechanism (Ahmadov et al., 2012). As shown in Fig. 11, the differences in model–measurement agreement between the two networks do not appear to be dependent on urban vs. rural monitor location. Instead, they may reflect between-network differences in sampling or analysis; different analysis techniques are known to produce widely varying OC concentrations (Cavalli et al., 2010).

Table 2 compares performance of WRF-Chem as configured here to that of the CMAQ model in a similar modeling effort by Appel et al. (2012). In this table, CMAQ as configured by Appel et al. (2012) in most cases predicts O3 observations with greater accuracy and precision than does WRF-Chem as configured here, while WRF-Chem in most cases does a better job predicting PM2.5. However, given the many differences in physical and chemical parameterizations and input data (including a difference in simulation year), the observed differences may or may not be generalizable. Instead, our conclusion from Table 2 is that the models are generally comparable in performance.

Table A2 compares WRF-Chem results from this study to results from Yahya et al. (2014) for a 12-month, contiguous US WRF-Chem simulation with a 36 km horizontal resolution spatial grid. NME results from the simulation performed here are lower (i.e., better) than those reported by Yahya et al. (2014) for most pollutants and measurement networks, but NMB results are more mixed. As horizontal grid resolution, input data, and model parameters all differ between the two studies, we are not able to determine the cause of the differences in results.