Volume 11, Issue 4 p. 1100-1116
Research Article
Open Access

Quantifying Uncertainties of Ground-Level Ozone Within WRF-Chem Simulations in the Mid-Atlantic Region of the United States as a Response to Variability

Andrew Thomas

Andrew Thomas

Department of Meteorology and Atmospheric Science, Pennsylvania State University, University Park, PA, USA

Center for Advanced Data Assimilation and Predictability Techniques, Pennsylvania State University, University Park, PA, USA

Search for more papers by this author
Amy K. Huff

Corresponding Author

Amy K. Huff

Department of Meteorology and Atmospheric Science, Pennsylvania State University, University Park, PA, USA

Correspondence to: A. K. Huff and F. Zhang,

[email protected];

[email protected]

Search for more papers by this author
Xiao-Ming Hu

Xiao-Ming Hu

Center for Analysis and Prediction of Storms, and School of Meteorology, University of Oklahoma, Norman, OK, USA

Search for more papers by this author
Fuqing Zhang

Corresponding Author

Fuqing Zhang

Department of Meteorology and Atmospheric Science, Pennsylvania State University, University Park, PA, USA

Center for Advanced Data Assimilation and Predictability Techniques, Pennsylvania State University, University Park, PA, USA

Correspondence to: A. K. Huff and F. Zhang,

[email protected];

[email protected]

Search for more papers by this author
First published: 03 April 2019
Citations: 19

Abstract

Understanding forecast uncertainties and error growth dynamics is a prerequisite for improving dynamical prediction of meteorology and air quality. While predictability of meteorology has been investigated over the past few decades, the uncertainties in air quality simulations are less well known. This study explores the uncertainties in predicting ground-level ozone (O3) in the Mid-Atlantic region of the United States during June 2016 through a series of simulations using WRF-Chem, focusing on the sensitivity to the meteorological initial and boundary conditions (IC/BCs), emissions inventory (EI), and planetary boundary layer (PBL) scheme. The average uncertainty of ground-level maximum 8-hr average O3 mixing ratio (MD8-O3) was most sensitive to uncertainties in the IC/BCs, while uncertainty in the EI was of secondary importance, and was least sensitive was to the use of different PBL schemes. Updating the NO emissions in the EI had the greatest influence on the accuracy, with an estimated decrease of 0.59 ppbv/year in the root-mean-square error and an average decrease of 0.63 ppbv/year in the values of modeled MD8-O3. Our study suggests using perturbations in IC/BCs may lead to a more dispersive ensemble of O3 prediction than using different PBL schemes and/or different EI. However, considering the combined uncertainties from all three sources examined are still smaller than the averaged root-mean-square errors of predicted O3 against observations, there are apparent other sources of uncertainties not studied that need to be considered in future ensemble predictions of O3.

Key Points

  • Large uncertainties in simulations of daytime ground-level ozone derive from uncertainties in meteorological initial and boundary conditions
  • The average ground-level ozone mixing ratio (mean bias) is nevertheless influenced more by uncertainties in the emissions inventory than in the meteorological initial and boundary conditions
  • The error and bias of model-predicted ground-level ozone may be decreased as much as 0.59 ppbv per year due to progressive reduction over the past decade of emissions of nitrogen oxides (NOx), a chemical precursor of ozone

Plain Language Summary

Ozone, the primary pollutant in photochemical smog, is harmful to human health, particularly for children, senior citizens, and people with existing heart or lung diseases, like asthma. To protect public health, air quality forecasts of ozone are issued across the United States, primarily for metropolitan areas, where ground-level ozone tends to be the highest. When ground-level ozone is predicted to be exceed the daily health standard, people are advised to take steps to limit their outdoor activities. Operational air quality forecasters use predictions of ground-level ozone from numerical air quality models as guidance for their public forecasts. To assess the uncertainty in these model predictions of ground-level ozone, a series of simulations using the air quality model WRF-Chem were conducted in this study through changing the inputs to the model, including starting weather conditions, pollutant emissions inventories, and other model settings. Results show that the largest uncertainty in ground-level ozone was predicted by the model simulations using different starting weather conditions. Using the results of this study, the year-to-year decrease in the ground-level ozone was simulated, based on reduction in the source emissions. This study provides guidance regarding how numerical air quality models should be configured for future operational applications, which may lead to more accurate predictions of ground-level ozone and, thus, more accurate air quality forecasts.

1 Introduction

Ground-level ozone (O3), one of the U.S. Environmental Protection Agency (EPA)'s six criteria pollutants, has been linked to a litany of health problems, including cardiovascular disease (Azevedo et al., 2011), infant mortality (Bell, 2004), and asthma (Gent, 2003). Sensitive groups, including individuals with cardiovascular and pulmonary diseases, children, and senior citizens, are particularly susceptible to negative health effects associated with exposure to criteria pollutants. To protect the public from the harmful effects of ground-level O3 and the other criteria pollutants, the U.S. EPA has established National Ambient Air Quality Standards (NAAQS). Primary NAAQS project public health, and secondary NAAQS project public welfare. For O3, the current (as of 1 October 2015) primary and secondary NAAQS are both 70 ppbv, averaged over an 8-hr time period. Compliance with the O3 NAAQS is determined by the annual fourth-highest MD8-O3, averaged over 3 years; if this value for a given location is ≥71 ppbv, the location is designated a nonattainment area for O3.

To help protect the public from the adverse health effects of O3, air quality forecasts of ambient O3 are issued for major metropolitan regions and surrounding suburbs across the United States. These forecasts are communicated using EPA's the Air Quality Index (AQI), a color-coded, dimensionless scale (https://www3.epa.gov/airnow/aqi_brochure_02_14.pdf). An AQI of 101 or higher corresponds to an exceedance of the O3 NAAQS (termed “O3 exceedances” hereafter), therefore, it is most critical for air quality forecasters to issue accurate forecasts on days when the AQI is expected to reach 101 or higher.

Eulerian air quality modeling is a multifaceted problem that depends on economic forcing (Tong et al., 2016), meteorological conditions (Seaman, 2000), biogenic emissions (Bell & Ellis, 2004), and chemistry (Mar et al., 2016). Each of these processes is variable, which affects the practical predictability of O3. Practical predictability is defined as the uncertainty of modeling with errors in either initial state or process that are considered acceptable for operational uses. Practical predictability is contrasted against intrinsic predictability—the limit of the uncertainty of a nearly perfectly modeled system (Lorenz, 1969; Melhauser & Zhang, 2012; Zhang et al., 2006). Practical predictability studies evaluate the current state of the ability to accurately model a phenomenon or variable, with the intention of identifying factors that may improve accuracy. The practical predictability of O3 is essential for the proper implementation of models in air quality forecasting, since it allows model users to attribute error to each process. We are using the American Meteorological Society's definition of predictability (Predictability, 2012), which is “[t]he extent to which future states of a system may be predicted based on knowledge of current and past states of the system.” Since it is common to quantify predictability using model uncertainty (Zhang et al., 2006; Melhauser & Zhang, 2012; Houtekamer & Zhang, 2016; Bei et al., 2010), we will often refer to uncertainty (2012), expressed through a standard deviation, as a metric to describe the practical predictability. We will also use variability to refer to the uncertainty of a process. For this study, we will only examine model uncertainty as it pertains to practical predictability, since we want to target research areas, such as the emissions inventory or initial and boundary conditions, to maximize the accurate prediction of ground-level O3.

In this study, we used the Weather Research and Forecasting Model with Chemistry (WRF-Chem; Grell et al., 2005), rather than the Community Multiscale Air Quality Model (CMAQ; Byun et al., 1997), which is used by the National Air Quality Forecasting System (NAQFC; Otte et al., 2005), because WRF-Chem computes chemistry alongside the meteorology. In this way, WRF-Chem is not restrained to computing the chemical tendencies with meteorological information that is as old as the meteorological model output interval, as is the case with CMAQ (Grell et al., 2005). The downside is that the computational costs increase with the additional computations between output intervals (Grell et al., 2005). While another advantage of simulating chemistry alongside the meteorology is the possibility for chemistry to interact with model physics, we did not test such features in order to mitigate the role of errors in chemistry on physical and dynamical processes and to reduce the computational expenses attributed to modeling aerosols.

The model uncertainty of simulated ground-level O3 is influenced by several settings. The most notable settings that contribute to the model uncertainty are the physical parameterizations, initial and boundary conditions, and anthropogenic emissions inventory (Cuchiara et al., 2014; Mallet & Sportisse, 2006; Mena-Carrasco et al., 2009; Misenis & Zhang, 2010). Physical parameterizations, equations that approximately emulate subgrid phenomena, are methods representative of the current understanding of the physical process. The variety of solutions of the parameterizations is representative of the diverse understanding of those processes and is one aspect of practical predictability of simulated ground-level O3. Varying initial and boundary conditions are within the definition of practical predictability, for it is not reasonable to know the true state of the atmosphere, differentiating practical predictability from the intrinsic predictability (Melhauser & Zhang, 2012). The definition of practical predictability also contains the anthropogenic emissions inventory, as real-time point and area emissions are not available presently.

Several studies have been done on the model uncertainty of O3 due to variability of the physical parameterizations (e.g., Hu et al., 2012; Hu et al., 2013; Hu et al., 2013; Žabkar et al., 2013). Mallet and Sportisse (2006) indicated that O3 is most sensitive to the model uncertainty in the turbulence closure, which is embodied in the planetary boundary layer (PBL) scheme within WRF-Chem. Some studies indicate that the Yonsei University (YSU) PBL scheme shows the closest agreement between predicted hourly O3 and observed hourly O3 (Yerramilli et al., 2012; Cuchiara et al., 2014; Cheng et al., 2012). Although Yerramilli et al. (2012) claim that O3 is more sensitive to model uncertainty in the Land Surface Model (LSM), Pleim (2011) notes that there are deficiencies in that study, including the fact that the second version of the Asymmetric Convective Model (ACM2) at that time was not implemented for use with tracers within WRF-Chem. Since then, other studies, such as Cuchiara et al. (2014), have used ACM2 within WRF-Chem since diffusion was implemented in 2013. Hu, Klein, and Xue (2013) and Hu, Klein, Xue, Zhang et al. (2013) indicate that the mixing strength in the PBL scheme is critical to correctly simulate near-surface O3 and its dry deposition during night time, which will affect the amount of O3 in the convective boundary layer during the next day. Other studies, such as Misenis and Zhang (2010), corroborate the claim that O3 may be more sensitive to model uncertainty in the LSM than the PBL scheme, though Hodnebrog et al. (2011) suggested that the model uncertainty of O3 to variability in the LSM may be confined to certain regions. Additionally, the physical relationship between the LSM and O3 is not as clear as the physical relationship between O3 and the PBL. There are numerous settings specific to atmospheric chemistry that may impact the mixing ratios of O3 and its precursors but are limited to “on” and “off” within WRF-Chem, such as dry deposition.

While the variety of choices of the PBL scheme represents a large model variability of a solitary meteorological process, emissions inventory variability, which represents the variability of anthropogenic emissions, provides a large component of model uncertainty of O3. The emissions inventory includes, but is not limited to, emissions of carbon monoxide, volatile organic compounds (VOCs), and nitric oxide (NO) within the United States. O3 is formed from the reaction of atomic oxygen with molecular oxygen following a series of complex reactions involving sunlight; NO and nitrogen dioxide (NO2), collectively termed NOx; and VOCs (Calvert et al., 2015). Within the United States, anthropogenic NOx, which is emitted by high-temperature combustion processes, has been decreasing since at least 2005 due to regional emissions controls on large energy-generating units and mobile sources (Tong et al., 2015). Although NOx emissions overall have been decreasing, they retain day to day temporal and spatial variability. Variability in emissions inventories is associated with changes in emissions control devices, activity, fuel sources, and changes in emission factors, among other reasons (Anderson et al., 2014; Castellanos et al., 2009; Frost et al., 2006; Kim et al., 2006; Tong et al., 2015; Vijayaraghavan et al., 2012). Of the emission inventories available, the gridded version of the National Emissions Inventory (NEI) series from the U.S. EPA has 4-km resolution, the highest spatial resolution among readily available emissions inventories. This aspect is critical, since higher-resolution emissions inventories are important for local air quality studies (Hodnebrog et al., 2011). Additionally, the program that adapts the NEI data to the WRF-Chem grid implements an algorithm for subgrid buoyant plume rise and elevated point source emissions, which may not otherwise be resolved by WRF-Chem. The NEI gridded emissions inventories are updated approximately every 6 years, with the latest emissions inventory being representative of a typical July weekday in 2011 (NEI-11). The point sources are compiled by the EPA, according to the Air Emissions Reporting Rule. The mobile emissions sources were processed using the Motor Vehicle Emissions Simulator, which has the most variability according to Anderson et al. (2014). For more information on the construction of the NEI-11, please see EPA (2015).

Accurate prediction of ground-level O3 is contingent upon accurate prediction of NOx. Errors in NOx emissions are particularly important for the Mid-Atlantic region, which is NOx limited (e.g., Butler et al., 2011; Duncan et al., 2010), meaning that during the summer O3 season, increases in ambient NOx emissions will increase production of O3. Travis et al. (2016) showed that the NEI-11 may be overestimating NOx emissions up to a factor of 2, demonstrating that the overestimation of NOx emissions leads to overestimation of O3 in NOx-limited environments. While Travis et al. (2016) agree with Anderson et al. (2014), the recommended decrease in NOx varies, indicating that there is ambiguity with the precise decrease needed. The NAQFC adjusted the NOx emissions for 2012 and found that the NOx bias decreased between 0.57 and 2.34 ppbv, while the decrease in O3 bias was between 0.92 and 1.87 ppbv (Pan et al., 2014). Other studies in different areas yielded varying results. Zhong et al. (2016) provides one example of the model uncertainty of O3 in China due to discrepancies between one regional and one global emissions inventory, with differences between 12 and 16 ppbv of O3 in certain locations, suggesting that the model uncertainty of O3 due to variability of the emissions inventory is location specific. Going further than simply comparing emissions inventories, some studies have examined the response of O3 due to certain sources (Vijayaraghavan et al., 2012) and temporal variability (Castellanos et al., 2009) of emissions.

While the emissions inventory is known to be a major source of model uncertainty for MD8-O3, the initial and boundary conditions are critical for determining the evolution of the model. Lorenz (1963) demonstrated the influence of perturbations of the initial state on the evolution of nonperiodic flow. This dependency on the initial conditions also includes the initial composition. While chemical initial and boundary conditions are important (Berge et al., 2001), the influence of chemical initial conditions can be constrained with a spin-up time of 48 hr, to reduce the correlation on initial conditions by 10% (Jiménez et al., 2007). Bei et al. (2010) suggests that meteorological initial conditions contribute more uncertainty to O3 mixing ratios in Mexico City, with the ensemble spread reaching 15 ppbv over Houston. Zhang et al. (2007) noted that, for a high O3 event in Houston, Texas, the spread of ensemble members with different initial (and boundary) conditions peaked with 40 ppbv of hourly O3 values over the Gulf of Mexico and 20 ppbv over one of the Texas stations. Gilliam et al. (2015) also observed a spread as high as 10–20 ppbv over the northeastern United States, while using the Short-Range Ensemble Forecasting system members as initial and boundary conditions. Hu et al. (2019) used a similar methodology as Gilliam et al. (2015) for the Dallas-Fort Worth area and found that the spread of the plume direction was most affected. Beekmann (2003) suggest that the O3 sensitivities may range from 4–10 ppbv over Paris. Discrepancies in the ranges of sensitivities may be a result of a variety of reasons, including the methods of perturbations of initial and boundary conditions, the quantity of O3 precursors emitted, the effect of local topography, and the modeled meteorological uncertainty.

We investigated and compared the relative impacts of the variability in the emissions inventory, the PBL scheme, and the meteorological initial and boundary conditions on the model uncertainty of MD8-O3 by running different WRF-Chem simulation experiments over the eastern United States for June 2016; the model domain is shown in Figure 1a. While the simulations were conducted over the eastern United States, the focus was on the urbanized Interstate-95 (I-95) Corridor, running from New York City to Washington, DC, as shown in Figure 1b. This region routinely observes the highest MD8-O3 in the Mid-Atlantic region. The goal was to quantify the model uncertainty of MD8-O3 along the I-95 Corridor, with the aim of making recommendations to improve air quality model guidance in an area that is highly susceptible to O3 exceedances. The motivation behind this research was to test potential settings needed to make near-continuous operational predictions of MD8-O3 using WRF-Chem for the Mid-Atlantic Region. Those potential settings for the simulations included the YSU, Mellor-Yamada-Janjic scheme (MYJ), and ACM2 PBL schemes, the NEI-11, NEI-05, NEI-14 emissions inventories, and Global Forecasting System (GFS), European Center for Medium Range Weather Forecasting Interim Reanalysis (ERA-Interim), and the second version of the Modern Era Retrospective analysis for Research and Applications (MERRA-2) meteorological initial and boundary conditions. The three analysis data sets differ in the forecast model, data assimilation methodology, and the amount of observations being assimilated, which can represent realistic uncertainties in the meteorological initial and boundary conditions.

Details are in the caption following the image
The simulated WRF-Chem domain (a), where the WRF-Chem simulations were conducted, and the analysis region (b), where our analysis occurs.

This study complements much of the existing literature and provides an evaluation of the relative importance of the choice of different model settings. This work also supplements previous studies regarding the model uncertainty of MD8-O3 as a response to variability in (a) the meteorological initial and boundary conditions, by testing different sources as a sampling method, and (b) the emissions inventory by testing qualitatively similar emissions inventories with the exactly the same meteorological scenarios. In this way, this research represents a nearly longitudinal modeling study of the trends of O3 due to changes in the emissions inventory.

2 Methods

2.1 The Climatology of O3 in the Mid-Atlantic Region

Historically, O3 exceedance days within the Mid-Atlantic region have been characterized by a well-defined synoptic scale pattern associated with the western edge of the quasi-stationary Bermuda High extending into the Mid-Atlantic region. This pattern includes a ridge of high pressure aloft, with the ridge axis over or west of the Mid-Atlantic, and slowly eastward migrating surface high pressure. This synoptic pattern is conducive to sunny skies, above average temperatures, stagnating surface winds, and regional transport of O3 precursor emissions from the historically NOx-rich Ohio River Valley source region (Ryan et al., 1998). Under these conditions, local and regional O3 formation was maximized, leading to multiday O3 exceedance events.

During June 2016, the weather patterns observed during the most widespread MD8-O3 exceedance days of 1 and 20 June were similar to those described by Ryan et al. (1998). On the days of 11, 15, 24, and 26 June, O3 formation was enhanced by mesoscale features, including Atlantic Ocean sea breezes, Chesapeake Bay breezes, and an Appalachian lee trough, all of which increased O3 production. An example of the enhancement of O3 by a sea breeze circulation in the Mid-Atlantic region was given in Stauffer et al. (2015). Another example is shown in Seaman and Michelson (2000), which examined the influence of an Appalachian lee trough on O3.

2.2 WRF-Chem Simulations

A control simulation, designated CNTL, using the WRF-Chem version 3.6.1 (Grell et al., 2005) with the regional atmospheric chemistry mechanism (RACM; Stockwell et al., 1997), was performed with meteorological initializations starting every day at 12 UTC from 1 June to 29 June, with a spin-up period on 30 and 31 May. A 2-day spin-up period was chosen because there were precedents for it, or for even fewer hours, without extensive explanation of the accuracy of the chemical initial and boundary conditions (Jiménez et al., 2007; Tie et al., 2010; Zhang et al., 2006). Each initialization contained 48 hr of forecasts, such that each day in June 2016 was simulated. The chemical fields were reinitialized with the previous day's prediction of the chemical fields, such that the chemical fields for the first 24-hr period of each initialization was considered continuous. The reinitialization of the meteorological fields is assumed to have a small effect on the chemical tendencies, with only a hypothetical jump in the photolysis rates. The tendencies from the chemical mechanism are considered autonomous because the initialization of the meteorological fields was a dry start, so clouds were still forming, but since the initialization time was shortly after sunrise, the inconsistency in the photolysis rate is considered minimal. Inconsistencies in the horizontal advection and vertical mixing tendencies may have occurred due to changes in the momentum fields, but they are considered (a priori) to be small due to the timing of when reinitialization occurred. A preliminary test was conducted and showed that the inconsistencies between the soil data and meteorological data had a negligibly small effect on the result (not shown). The CNTL sensitivity experiment used the YSU (Hong et al., 2006) PBL scheme and the NEI-11 emissions inventory (EPA, 2015). In addition to the model settings listed in Table 1, all sensitivity experiments had a domain spanning from the Mississippi River to southern Maine, as is depicted in Figure 1a. The vertical model structure was the same as described by Hu et al. (2012) with 12-km grid spacing. In Figure 1b, major cities are marked and labeled in the analysis region which contains the urbanized I-95 Corridor, marked in red. We made the WRF-Chem domain more expansive than the analysis region to account for anthropogenic sources of NOx upwind.

Table 1. CNTL Model Settings
Setting Choice Reference
Longwave radiation RRTM Mlawer et al. (1997)
Shortwave radiation Dudhia Dudhia (1989)
Cumulus Grell-Devenyi Grell and Dévényi (2002)
Microphysics WSM6 Hong and Lim (2006)
Land surface model NOAH Tewari et al. (2004)
Chemistry RACM Stockwell et al. (1997)
Biogenic emissions MEGAN Guenther et al. (2006)
Photolysis Madronich Madronich (1987)

Besides the CNTL sensitivity experiment, a total of six sensitivity experiments were performed, each of which varied either the PBL scheme, emissions inventory, or the meteorological initial and boundary conditions from that used in the CNTL sensitivity experiment. All sensitivity experiments, including the CNTL, are detailed in Table 2, along with the naming convention for each subgroup of sensitivity experiments.

Table 2. Sensitivity Experiment Names and Changed Model Settings
Sensitivity experiment PBL Meteorological initial and boundary conditions Emissions
CNTL YSU GFS 2011
ACM2 ACM2 GFS 2011
MYJ MYJ GFS 2011
NEI-05 YSU GFS 2005
NEI-14 YSU GFS 2014
MERRA-2 YSU MERRA-2 2011
ERA YSU ERA 2011
  • Note. ACM2 = Asymmetric Convective Model; CNTL = control simulation; ERA = European Center for Medium Range Weather Forecasting Interim Reanalysis; GFS = Global Forecasting System; MERRA = Modern Era Retrospective analysis for Research and Applications; MYJ = Mellor-Yamada-Janjic scheme; NEI = National Emissions Inventory; PBL = planetary boundary layer; YSU = Yonsei University.

The first subgroup of sensitivity experiments explored two different emissions inventories, the 2005 version of the National Emission Inventory (NEI-05) and a version of the NEI-11 updated to reflect emissions of NO in 2014 (NEI-14) in order to capture the annual variability of the emissions of NO and the corresponding model uncertainty of O3. Utilizing the method of Tong et al. (2015), 2014 was the most recent year for which NO emissions could be updated. The weekday emissions of the NEI-05 were used to provide continuity with the NEI-11, since only the weekday emissions of the NEI-11 were available. The NO point emissions of the NEI-14 were updated for each hour by using the average of the Clean Air Market data (https://ampd.epa.gov/ampd/), which is provided by power plants to EPA (EPA: Air Markets Program Data, 2018). The area emissions were updated for each state by using state-wide ratios of NOx emissions in 2014 to NOx emissions within the NEI-11, which were contributed by the National Oceanic and Atmospheric Administration (Tong et al., 2015). The creation of the NEI-14 loosely follows the methodology used to update the NAQFC O3 model (Pan et al., 2014). Figure 2 demonstrates the changes to each emissions inventory in the analysis region. Daily NO emissions decreased by 27.4% from the NEI-05 to the NEI-14 and daily emissions of VOCs decreased by 39.6% from the NEI-05 to the NEI-11; emissions of VOCs were unchanged by design from the NEI-11 to NEI-14.

Details are in the caption following the image
Average hourly emissions of NO and nonmethane volatile organic compounds from each emissions inventory used in this study.

The next subgroup of sensitivity experiments focused on two different PBL schemes, the ACM2 and MYJ schemes. As reviewed in Hu et al. (2010), the CNTL sensitivity experiment uses the YSU scheme, which is a nonlocal parabolic K-scheme that defines entrainment. The ACM2 is a hybrid local/nonlocal PBL scheme, which treats upward mixing nonlocally and downward mixing locally. In contrast to the YSU scheme used by the CNTL sensitivity experiment, the MYJ scheme is a local K-scheme. As noted by Skamarock et al. (2008), the WRF model constrains some PBL schemes to certain surface layer physics. Therefore, the sensitivity experiments that utilized the MYJ PBL scheme used the Eta similarity surface layer physics (Monin & Obukhov, 1954), while the ACM2 and YSU PBL schemes use the revised MM5 surface layer physics (Jiménez et al., 2012).

The final subgroup of sensitivity experiments used two different initial and boundary conditions: the second version of the MERRA-2 by the National Aeronautics and Space Administration and the ERA-Interim reanalysis from the European Centre for Medium-Range Weather Forecasts. MERRA-2 is a reanalysis data set that utilizes the Goddard Earth Observing System model with a 3D-Variational data assimilation system. The approximate resolution of MERRA-2 is 0.5° × 0.625° with 72 vertical levels (Gelaro et al., 2017). The ERA-Interim reanalysis uses the European Centre for Medium-Range Weather Forecasts model with T255 resolution (approximately 80 km) with a 4D-Variational data assimilation system (Dee et al., 2011). The soil temperature, moisture, and depth for all sensitivity experiments were taken from the GFS, such that only meteorological data was varied among this subgroup. The potential inconsistencies were found to be of little importance after comparing the ERA sensitivity experiment with GFS soil information and a simulation with the ERA initial and boundary meteorological and soil conditions (not shown). All inputted data sets were temporally interpolated, to the extent that the boundary conditions were updated hourly.

As mentioned previously, the focus of the study was on the urbanized I-95 Corridor. This entire region, encompassing parts of eastern Pennsylvania, New Jersey, Delaware, Maryland, and northern Virginia, is in nonattainment for O3, based on the 2008 O3 NAAQS of 75 ppbv (since nonattainment is determined by a 3-year average of O3 observations, nonattainment areas based on the 2008 NAAQS are the most recent data currently available). In the I-95 Corridor analysis region, we calculated the sensitivity of MD8-O3, expressed by the standard deviation among the three sensitivity experiments of each subgroup. We also examined the vertical distribution of MD8-O3, and the modeled uncertainty thereof, along I-95. Additionally, we evaluated the effect of the changing emissions inventory on predicted MD8-O3.

3 Results and Discussion

3.1 Average Group Differences

Figure 3 shows the temporal average of ground-level MD8-O3 for each sensitivity experiment in the analysis region, as compared to the similarly processed observations. The model-interpolated location of I-95 is marked in white. Each sensitivity experiment overproduced MD8-O3 in the analysis region, but the ERA sensitivity experiment overproduced O3 to the largest extent.

Details are in the caption following the image
Temporal average of observed maximum daily 8-hr average ozone (MD8-O3) during the modeling period, representing the average modeled and observed MD8-O3.

Among the sensitivity experiments, the ERA sensitivity experiment produced the most O3, even more than the NEI-05 sensitivity experiment. The cause of this overprediction was not well understood, though one possible explanation is that the ERA sensitivity experiment inadequately modeled vertical transport of O3 and O3 precursors from the residual layer. The choice of PBL scheme led to less noticeable differences in the average MD8-O3 compared to the subgroups with different initial and boundary conditions or different emissions inventories. This result is different from previous studies, which found that the PBL scheme selection was more critical under different circumstances, such as Yerramilli et al. (2012), Cuchiara et al. (2014), or Cheng et al. (2012).

More quantitatively, Figure 4 and Table 3 show the mean bias and error metrics of each sensitivity experiment during the modeling period. The NEI-05 and ERA sensitivity experiments had the most overprediction of MD8-O3, while the NEI-14 experiment overpredicted less MD8-O3 relative to the other sensitivity experiments in the emissions inventory subgroup (Table 3). The reduced overproduction of the NEI-14 sensitivity experiment suggests that an updated emissions inventory results in more accurate predictions of MD8-O3. Also, the difference between the unbiased root-mean-square error (RMSE) of the day 1 and day 2 predictions of MD8-O3 is statistically significant at the 95% confidence interval. While this is worth noting, the limited number of models used, as well as the fact that none of the other error tables within section 3.5 displays similar statistical significance, suggests that this is a Type 1 error.

Details are in the caption following the image
Mean bias of maximum daily 8-hr averaged ozone (MD8-O3) for modeling period.
Table 3. Error Metrics of Maximum Daily 8-hr Averaged O3 (MD8-O3) for Each Sensitivity Experiment
Sensitivity experiment Mean bias RMSE Unbiased RMSE*
Day 1 Day 2 Day 1 Day 2 Day 1 Day 2
CNTL 14.23 14.3 20.1 19.55 14.2 13.33
NEI-05 17.59 17.56 23.41 22.7 15.45 14.39
NEI-14 12.14 12.38 18.08 17.78 13.4 12.75
ERA 17.49 16.29 23.63 21.52 15.89 14.06
MERRA2 13.84 13.65 19.43 19.34 13.65 13.7
ACM2 13.98 13.96 19.66 18.95 13.82 12.82
MYJ 13.32 13.08 19.68 18.72 14.48 13.39
  • Note. One asterisk (*) signifies that the day 1 error metric is statistically significantly greater than the day 2 error metric, using a 95% confidence interval for the Student's T test. ACM2 = Asymmetric Convective Model; CNTL = control simulation; ERA = European Center for Medium Range Weather Forecasting Interim Reanalysis; MERRA = Modern Era Retrospective analysis for Research and Applications; MYJ = Mellor-Yamada-Janjic scheme; NEI = National Emissions Inventory; RMSE = root-mean-square error.

The temporal average of the cross-sectional MD8-O3 over I-95 is displayed in Figure 5. The latitudes of major cities are marked and labeled by dashed lines. The vertical coordinate was computed using the hypsometric equation of the averaged hydrostatic pressure and temperature during the same period that the ground-level MD8-O3 was modeled. In the southern part of I-95, the NEI-05 sensitivity experiment showed the most overproduction of MD8-O3, with peak values of MD8-O3 in rural and suburban areas. The overproduction of MD8-O3 in the NEI-05 sensitivity experiment in the southern analysis region was most prominent 0.5–1 km aloft, near the height of the boundary layer, where the average mixing ratio reached over 80 ppbv. Additionally, the ERA sensitivity experiment showed that the overproduced O3 was enhanced aloft by 5 ppbv over the cities of Washington D.C., Baltimore, and Philadelphia, as compared to both the MERRA-2 and CNTL sensitivity experiments. The overprediction of the ERA sensitivity experiment is likely due to its inadequate simulation of moist convection (not shown), leading to the enhancement of MD8-O3. Thus, this cross section analysis demonstrates the impact that emission reductions can have on the vertical profile of MD8-O3, as well as the uncertainty due to variability in the meteorological initial and boundary conditions.

Details are in the caption following the image
Average cross section of the maximum daily 8-hr average ozone (MD8-O3) along I-95 during the modeling period.

3.2 Model Uncertainty of MD8-O3

Figure 6 displays the average standard deviation of near-surface MD8-O3 for each sensitivity experiment subgroup. The model-interpolated location of I-95 is denoted by the white line. This figure illustrates that the meteorological initial and boundary conditions subgroup had a larger modeled uncertainty of MD8-O3 than the uncertainties of modeled MD8-O3 due to variabilities in the emissions inventory and PBL parameterizations for the study period. Variations in using different initial and boundary conditions led to peak modeled uncertainty (mean standard deviation among the subgroup) for O3 of 5.2–6.1 ppbv, which is considerably higher that the peak sensitivities of about 3.4–4.3 ppbv due to changes in either the emissions inventory or the PBL parameterization scheme. This implies that the meteorological conditions are a primary source of O3 modeling uncertainties, which is consistent with findings in Zhang et al. (2007). Despite having a similar peak model uncertainty of MD8-O3, the variations in the emissions inventory experiments have a greater impact over a larger area on variations of MD8-O3 than do variations due to the PBL schemes. Although the model uncertainty of MD8-O3 may be largest for the initial and boundary conditions, it does not necessarily indicate that the model uncertainty was largest at every point. For example, MD8-O3 in southern New Jersey was more sensitive to variability in the emissions inventory than variability induced by using different meteorological initial and boundary conditions (by approximately 1 ppbv). Additionally, the largest uncertainties due to different meteorological initial and boundary conditions may not be collocated with the highest values of MD8-O3. For example, I-95, which connects major cities in the analysis region, runs along the outside edge of the peak uncertainties. This highlights the model uncertainty of the transport of near-surface O3 resulting from NOx plume sources, such as I-95.

Details are in the caption following the image
The average model uncertainty of the maximum daily 8-hr average ozone (MD8-O3), expressed by the temporal average of the standard deviation of each subgroup, during the modeling period.

Figure 7 illustrates the vertical profile in the lowest 3-km altitude above I-95 for the temporally averaged model uncertainty of MD8-O3. While the MD8-O3 aloft was most sensitive to uncertainties in the meteorological initial and boundary conditions, the model uncertainty due to variability in the emissions inventory was larger aloft over the southern part of the analysis region. This result agrees with Travis et al. (2016) that the emissions inventory represents a large source of O3 modeling error for the southeastern United States. Also, the most model uncertainty of MD8-O3 associated with the emissions and initial/boundary conditions subgroups was located near the top of the boundary layer. This increase in the sensitivity model uncertainty of O3 with height is consistent with the hypothesis that the ERA sensitivity experiment inadequately modeled convective vertical transport of O3 and O3-precursors.

Details are in the caption following the image
The average model uncertainty, expressed by the temporal average of the cross section of the standard deviation, of maximum daily 8-hr average ozone (MD8-O3) during the 20% of days with National Ambient Air Quality Standards exceeding ozone observations. PBL = planetary boundary layer. ICBC= meteorological initial and boundary conditions.

Figure 8 shows which days have the most model uncertainty of MD8-O3, by displaying the average standard deviation of each subgroup at each observing location. Consistent with Figure 6 and Figure 7, the initial and boundary conditions subgroup frequently had the largest uncertainties among the three sensitivity subgroups. However, the average standard deviation of each subgroup is considerably smaller than the RMSE, which is defined as the square root of the average squared distance between the observations and predictions, among all sensitivity experiments. This result also shows a smaller standard deviation than the findings of Gilliam et al. (2015), but this may be due to the use of only three sensitivity experiments in each subgroup, as well as the use of days with low O3, which is inherent in the utilization of the entire month of June. By comparison, Gilliam et al. (2015) used 10 Short-Range Ensemble Forecasting-initialized WRF-CMAQ simulations, as compared to our three multimodel initialized WRF-Chem simulations. Our results for the model uncertainty in MD8-O3 due to variability in the initial and boundary conditions over land were of similar order to those presented in the “half-run” by Zhang et al. (2007). The similarity of these results implies that the initial differences between the CNTL, ERA, and MERRA-2 sensitivity experiments were representative of about half the variability (i.e., model uncertainty) of the initial ensemble employed by Zhang et al. (2007).

Details are in the caption following the image
A time series of model uncertainty of the maximum daily 8-hr averaged ozone (MD8-O3) for each O3 observing monitor, expressed by the spatially averaged standard deviation, in June 2016. The beginning of each line segment indicates the uncertainty associated with the day 1 prediction of MD8-O3, while the end represents the day 2 uncertainty. PBL = planetary boundary layer.ICBC = meteorological initial and boundary conditions

Although the initial and boundary conditions have the most model uncertainty among the subgroups, the initial and boundary condition subgroup is still underdispersive, suggesting other sources of uncertainties beyond what are examined in this study also need to be considered simultaneously in probabilistic O3 modeling. This is evidenced by Figure 9, which shows a time series of MD8-O3 of each subgroup at the Rutgers University O3 monitor, which is representative of the New York City metropolitan area. The Rutgers University monitor was chosen because it observed 4 days with NAAQS exceedances of MD8-O3 during the study period. Since many of the observations were outside the ranges of each subgroup, all of the subgroups are underdispersive, which is an important consideration when data assimilation of air quality models is conducted (Houtekamer & Zhang, 2016). Additionally, Figure 9 supports the idea that older emissions inventories produce more MD8-O3, enhancing the positive bias. Figure 10, which shows a rank histogram of MD8-O3 within the entire analysis region, broadens this idea of underdispersion beyond a single monitor. Even with the bias removed, Figure 10 shows signs of underdispersion.

Details are in the caption following the image
Time series of the maximum daily 8-hr averaged ozone (MD8-O3) at Rutgers University.
Details are in the caption following the image
Rank histogram for biased (a) and unbiased (b) predictions of the maximum daily 8-hr averaged ozone (MD8-O3) within the entire analysis region.

3.3 Uncertainty and Error for Each Forecast Hour

Figure 11 shows the RMSE for O3 predicted each hour within the analysis region for each subgroup. For the first few hours, the model error increases greatly. The ERA sensitivity experiment shows the largest error. The error also decreases to a relative minimum near hour 24. This minimum is similar to the minimum found at hour 0. This result may be caused by the daily reinitialization, but that does not sufficiently explain the drastic decrease in RMSE preceding reinitialization. However, one possibility is that diurnal pattern of the errors may be associated with the diurnal pattern of the sensitivity, related to the treatment of the PBL scheme. Figure 11 also shows that the errors in day 2 predictions of hourly O3 may be due to the a priori assumption of the small impact of reinitialization error. This explanation is particularly relevant to the ERA and NEI-05, which have the most amount of error in all of the sensitivity experiments. Moreover, this pattern may be broadened to the idea that variability in the meteorological initial and boundary conditions are as important as variability in the emissions inventory.

Details are in the caption following the image
Root-mean-square error (RMSE) of hourly O3 as a function of forecast hour.

Figure 12 shows the station-averaged uncertainty (root-mean-square difference or spread among members) of O3 at each hour for each of the three subgroups as a response to variability. All three subgroups display a diurnal pattern. The uncertainty of the PBL subgroup shows negatively skewed behavior with peaks during the nighttime. The uncertainty of the emissions inventory subgroup is the smallest but shows a positive skewness and peak during the morning hours. The uncertainty of the emissions inventory subgroup also shows a small decrease in the afternoon and evening, implying a trend toward increased confidence. The meteorological initial and boundary conditions uncertainty is slightly larger in the daytime hours, but the diurnal pattern is not as pronounced compared to the emissions inventory or PBL subgroups.

Details are in the caption following the image
Station-averaged uncertainty of hourly O3 as a function of forecast hour. PBL = planetary boundary layer. ICBC= meteorological initial and boundary conditions.

3.4 Emissions Inventory Analysis

Since predicted MD8-O3 has a clear positive bias among all sensitivity experiments, as seen in Figure 4, and the newer emissions inventories decrease error, as described previously in section 3.1, we estimated the average decrease of the RMSE due to emissions inventory updates. Figure 13 illustrates the slope of the least squares line of the RMSE of MD8-O3 as a response to the emissions inventory year. There was one caveat: since only the emissions of NO were updated in the NEI-14, the changes in emissions of VOCs may alter the finding, but because O3 in the Mid-Atlantic region is NOx-sensitive, the result of changing emissions of VOCs will likely be small. Based on this analysis, an average decrease of approximately 0.59 ppbv/year of the RMSE and mean bias of MD8-O3 was estimated.

Details are in the caption following the image
Slope of the least squares line regression of the root-mean-square error (comparing model predictions against observations) of the maximum daily 8-hr average ozone (MD8-O3) per emissions inventory year during the 20% of days with the most National Ambient Air Quality Standards exceeding observations of ozone as a response to the different emissions inventory years.

Using a methodology similar to that used to estimate the average decrease of the RMSE, the average decrease of MD8-O3 was also found, as shown in Figure 14. Predictions of MD8-O3 decreased across the analysis region with each progressive emissions inventory. Over land, the average decrease of MD8-O3 was 0.63 ppbv/year.

Details are in the caption following the image
Slope of the least squares line regression for the modeled maximum daily 8-hr average ozone (MD8-O3) as a response to the different emissions inventory years.

3.5 Analysis of Errors for Other O3 Metrics

The results described to this point have focused on analysis of model-predicted MD8-O3, due to its practical relevance for ambient air quality. However, analysis of other model-predicted metrics, such as daily average O3, daily maximum O3, and daily minimum O3, can also provide important information and thus are also examined. For brevity, here only the errors of a few of such additional O3 metrics are discussed. The model prediction period is 12 to 12 UTC daily, so there will be a difference in each metric when compared to the corresponding metric based on local standard time. This is not usually an issue for daily maximum O3, but it can lead to differences for daily minimum O3 and daily average O3.

Table 4 lists the errors for daily maximum O3, which are similar to results to Table 3, with a few exceptions. One difference is that the unbiased RMSE for day 1 predictions is not (statistically significantly) greater than the RMSE for day 2 predictions. Another difference is that the magnitude of the errors of MD8-O3 (Table 3) are greater than the magnitude of the errors of daily maximum O3 (Table 4), except for the mean bias of both days of the CNTL, NEI-14, and day 1 predictions of the ERA and ACM2.

Table 4. Error Metrics of Daily Maximum O3 for Each Sensitivity Experiment
Sensitivity experiment Mean bias RMSE Unbiased RMSE
Day 1 Day 2 Day 1 Day 2 Day 1 Day 2
CNTL 14.12 14.27 21.94 21.83 16.8 16.51
NEI-05 18.32 18.67 26 25.98 18.45 18.07
NEI-14 11.62 11.94 19.63 19.71 15.82 15.68
ERA 17.38 16.4 25.42 23.91 18.55 17.4
MERRA2 14.18 14.25 21.81 22.13 16.57 16.93
ACM2 13.71 14.06 21.25 21.12 16.23 15.76
MYJ 14.35 14.31 22.75 21.65 17.65 16.24
  • Note. ACM2 = Asymmetric Convective Model; CNTL = control simulation; ERA = European Center for Medium Range Weather Forecasting Interim Reanalysis; MERRA = Modern Era Retrospective analysis for Research and Applications; MYJ = Mellor-Yamada-Janjic scheme; NEI = National Emissions Inventory; RMSE = root-mean-square error.

Table 5 lists the errors in daily average O3. In general, the mean bias of daily average O3 was greater than that of MD8-O3 (Table 3), with the exception of the ACM2 experiment. Both the RMSE and unbiased RMSE are greater for daily average O3 compared to MD8-O3. This increase in RMSE and mean bias with a widening averaging gap supports the idea that WRF-Chem predicts maximum O3 most accurately, with errors in daily average O3 inflated due to large errors in prediction of nocturnal O3 (Hu, Klein, & Xue, 2013). It is also worth noting that the unbiased RMSE for daily average O3 is smaller than that of the other O3 metrics.

Table 5. Error Metrics of Daily Average O3 for Each Sensitivity Experiment
Sensitivity experiment Mean bias RMSE Unbiased RMSE
Day 1 Day 2 Day 1 Day 2 Day 1 Day 2
CNTL 17.02 16.81 21.02 20.51 12.34 11.75
NEI-05 19.24 18.92 23.28 22.59 13.1 12.33
NEI-14 15.86 15.73 19.82 19.46 11.89 11.45
ERA 19.42 18.48 23.46 22.07 13.16 12.07
MERRA2 16.84 16.84 20.65 20.71 11.96 12.06
ACM2 13.61 13.7 19.02 19.09 13.29 13.3
MYJ 15.03 14.8 19.56 19.21 12.52 12.24
  • Note. ACM2 = Asymmetric Convective Model; CNTL = control simulation; ERA = European Center for Medium Range Weather Forecasting Interim Reanalysis; MERRA = Modern Era Retrospective analysis for Research and Applications; MYJ = Mellor-Yamada-Janjic scheme; NEI = National Emissions Inventory; RMSE = root-mean-square error.

Table 6 lists the errors in daily minimum O3. Both the mean bias and RMSE of daily minimum O3 are greater than the corresponding values for MD8-O3, maximum O3, and daily average O3, with the exception of the mean bias of the ACM2 experiment. The difference between the mean bias of the ACM2 experiment and the other sensitivity experiments suggests that the treatment of the nocturnal boundary layer in the ACM2 experiment produces less of a biased prediction, but the unbiased RMSE value indicates that this does not necessarily correspond to a more accurate prediction. Hu, Klein, & Xue (2013) provides a more in depth discussion on the treatment of the PBL scheme in relation to nocturnal O3.

Table 6. Error Metrics of Daily Minimum O3 for Each Sensitivity Experiment
Sensitivity experiment Mean bias RMSE Unbiased RMSE
Day 1 Day 2 Day 1 Day 2 Day 1 Day 2
CNTL 22.28 21.91 27.37 27.07 15.89 15.9
NEI-05 22.45 22.15 28.11 27.79 16.91 16.78
NEI-14 22.44 22.04 27.22 26.89 15.42 15.41
ERA 23.29 23.14 28.42 28.23 16.29 16.18
MERRA2 22.25 22.14 27.66 27.6 16.45 16.48
ACM2 13.26 13.41 26.07 25.45 22.45 21.63
MYJ 19.38 19.01 25.31 25.03 16.29 16.28
  • Note. ACM2 = Asymmetric Convective Model; CNTL = control simulation; ERA = European Center for Medium Range Weather Forecasting Interim Reanalysis; MERRA = Modern Era Retrospective analysis for Research and Applications; MYJ = Mellor-Yamada-Janjic scheme; NEI = National Emissions Inventory; RMSE = root-mean-square error.

3.6 Limitations

One important limitation of this study was the number of sensitivity experiments conducted. As with most research involving groups of simulations, the number of simulations used was limited by computational resources. Since WRF-Chem was computationally more expensive than the stand-alone WRF, and we ran WRF-Chem for 1 month, we limited the number of sensitivity experiments in consideration of finite computing resources. Additionally, we wanted to compare the same number of sensitivity experiments, and we had three emissions inventories available, so we were limited by our choice to test the model uncertainty of ground-level O3 to the variability of the emissions inventory. This limitation may be one reason the model uncertainty of all subgroups was less than the error of the observations, which further suggests there are other sources of uncertainties, such as biogenic emissions, dry deposition, chemical mechanisms, photolysis parameterizations, and convective parameterizations, in O3 modeling over the study region. This limitation on the number of sources of sampled uncertainties are not trivial. For example, Mar et al. (2016) showed that difference in chemical mechanisms can lead to differences in predicted O3 of 5-10 ppbv. These additional uncertainties will be subject of future research.

The role of both observational errors and unresolved emissions is also of note. Evaluating each sensitivity experiment based on observations of O3 is subject to factors such as representativeness and measurement uncertainty, and in most cases the measurement uncertainty is considered of negligible influence. Moreover, we are comparing errors between models, and in this way, our study can also be interpreted as comparing the representativeness of the different model settings.

4 Summary and Conclusion

Predictability within the field of meteorology has been comprehensively investigated in the past few decades and understanding of modeled uncertainties and error growth dynamics has proved critical for improving weather forecasting. In contrast, this topic in air quality modeling has been less explored. Uncertainties of simulated near-surface O3 in the eastern United States and its predictability were investigated in this study. We assessed the predictability of O3 by running seven WRF-Chem experiments with different meteorological initial and boundary conditions, emissions inventories, and PBL schemes over the Mid-Atlantic region for June 2016. Our results demonstrate that these particular sources of model uncertainty or possibly the methods in sampling the variability, even without air quality data assimilation, are underdispersive. Simulated ground-level O3 was most sensitive to variability in the meteorological initial and boundary conditions. This heightened model uncertainty emphasizes the need for accurate data assimilation of meteorological data, to ensure accurate modeling of ground-level O3. In addition, our results demonstrate how the choice of initial and boundary conditions may result in systematic biases for prediction of ground-level O3. Therefore, the choice of meteorological initial and boundary conditions, such as the choice of GFS or ERA, may influence the findings of other research. Of the three initial and boundary conditions tested, the MERRA-2 sensitivity experiment had the smallest bias and RMSE, suggesting it is most appropriate for operational use, though in the future we will examine other operational analysis and forecast datasets as initial and boundary conditions. The enhanced model uncertainty due to variability in the initial and boundary conditions also suggests that, pending other findings, air quality ensembles for modeling O3 should consist of perturbations in initial and boundary conditions.

Furthermore, our results show a correspondence of emissions inventory age to RMSE in predicted O3. While newer emissions inventories lead to more accurate results than older emissions inventories, with an improvement of approximately 0.59 ppbv/year, emissions inventories are updated infrequently. Therefore, updates in emissions modeling, or inverse modeling algorithms to remedy outdated emissions—especially in NO and VOCs—may improve O3 model accuracy.

To further generalize the findings from this study, future research is needed to extend the experiments to different months and years, and with combinations of simultaneously changing emissions inventories, PBL schemes, and meteorological initial and boundary conditions. Further research will also focus on determining the underlying reason why ground-level O3 is more sensitive to variability in the initial and boundary conditions rather than variability in the emissions inventory or the PBL scheme.

It is also worth noting that, given the total uncertainties from the three selected sources (emissions inventories, PBL schemes, and meteorological initial and boundary conditions) are still on average smaller than the RMSEs of predicted O3 verifying against ground observations, the current study does not fully capture all sources of O3 forecast uncertainties. Future work will investigate differences in O3 tendencies, including, but not limited to, differences in the photolysis rate, chemical tendency, horizontal transport, and photostationary O3. Additional sources of uncertainty will also be investigated, such as the chemical mechanism and photolysis parameterization. Similar methods may be applied to trace chemicals beyond O3, such as O3 precursors or the various VOCs within the chemical mechanism.

Acknowledgments

The data used for providing initial and boundary conditions is obtained from publicly accessible data archives at NCAR and NASA. The authors thank Daniel Tong for his help with developing the NEI-14 emissions inventory. The authors thank the two anonymous reviewers for their comments that improved the manuscript, and the editors for their support. Partial financial support for this study was provided by the Pennsylvania Department of Environmental Protection. Computing was performed at the Texas Advanced Computing Center, an NSF HPC Center, where the modeling output used in this study is archived and accessible.