Predicted ultrafine particulate matter source contribution across the continental United States during summertime air pollution events

The regional concentrations of airborne ultrafine particulate matter mass (D_p<0.1 µm; PM_0.1) were predicted in 39 cities across the United States (US) during summertime air pollution episodes. Calculations were performed using a regional source-oriented chemical transport model with 4 km spatial resolution operating on the National Emissions Inventory created by the U.S. Environmental Protection Agency (EPA). Measured source profiles for particle size and composition between 0.01 and 10 µm were used to translate PM total mass to PM_0.1. Predicted PM_0.1 concentrations exceeded 2 µg m⁻³ during summer pollution episodes in major urban regions across the US including Los Angeles, the San Francisco Bay Area, Houston, Miami, and New York. PM_0.1 spatial gradients were sharper than PM_2.5 spatial gradients due to the dominance of primary aerosol in PM_0.1. Artificial source tags were used to track contributions to primary PM_0.1 and PM_2.5 from 15 source categories. On-road gasoline and diesel vehicles made significant contributions to regional PM_0.1 in all 39 cities even though peak contributions within 0.3 km of the roadway were not resolved by the 4 km grid cells. Cooking also made significant contributions to PM_0.1 in all cities but biomass combustion was only important in locations impacted by summer wildfires. Aviation was a significant source of PM_0.1 in cities that had airports within their urban footprints. Industrial sources, including cement manufacturing, process heating, steel foundries, and paper and pulp processing, impacted their immediate vicinity but did not significantly contribute to PM_0.1 concentrations in any of the target 39 cities. Natural gas combustion made significant contributions to PM_0.1 concentrations due to the widespread use of this fuel for electricity generation, industrial applications, residential use, and commercial use. The major sources of primary PM_0.1 and PM_2.5 were notably different in many cities. Future epidemiological studies may be able to differentiate PM_0.1 and PM_2.5 health effects by contrasting cities with different ratios of PM_0.1∕PM_2.5. In the current study, cities with higher PM_0.1∕PM_2.5 ratios (ratio greater than 0.10) include Houston, TX, Los Angeles, CA, Bakersfield, CA, Salt Lake City, UT, and Cleveland, OH. Cities with lower PM_0.1 to PM_2.5 ratios (ratio lower than 0.05) include Lake Charles, LA, Baton Rouge, LA, St. Louis, MO, Baltimore, MD, and Washington, D.C.

1 Introduction

Airborne particulate matter (PM) has been linked with premature mortality and numerous other health risks in cities across the world (see, for example, references Dominici et al., 2006; Franklin et al., 2007; Pope et al., 2002, 2009; Ostro et al., 2006; Laden et al., 2000; Kheirbek et al., 2013; Aneja et al., 2017). Despite years of progress (EPA, 2017a), PM concentrations in many urban regions in the US still exceed health-based standards, resulting in an increase in non-accidental mortality (Franklin et al., 2007; Baxter et al., 2013). Toxicology testing suggests that ultrafine particles with a diameter <0.1 µm may be the most harmful size fraction within PM_2.5 (Li et al., 2003; Oberdorster, 2000; Ostro et al., 2015; Oberdorseter et al., 1995; Pekkanen et al., 1997). Initial attempts to analyze ultrafine particles in epidemiology studies have used particle number concentration as a surrogate for ultrafine particle exposure, but this approach has not found consistent relationships with health effects (HEI, 2013). In contrast, a recent epidemiology study based on ultrafine particle mass (PM_0.1) found significant associations with premature mortality (Ostro et al., 2015). In addition, ultrafine (UF) mass concentrations are highly correlated with particle surface area and can be a good metric for potential exposure to UF particles (Kuwayama et al., 2013; Ostro et al., 2015). Follow-up studies have also found significant associations between PM_0.1 and reproductive outcomes, including low birth weight and preterm birth (Laurent et al., 2016; Bergin et al., 1996). These findings have biological plausibility, since ultrafine particles may cross cell membranes and interfere with internal cell function (Sioutas et al., 2005). Ultrafine particles have greater surface area per volume due to the small particle diameter, making them more available for chemical reaction. Ultrafine particles can therefore have a larger impact when deposited deep into the lung cavity, from which they are not easily removed (Nel et al., 2006; Li et al., 2003).

A monitoring network for PM₁₀ and PM_2.5 has been operating throughout the continental US for almost 20 years. Multiple studies have performed source apportionment calculations for coarse and fine PM using these measurements (see, for example, Reff et al., 2009; Zhang et al., 2014; Zheng et al., 2002; Ham and Kleeman, 2011). In contrast, measurements of PM_0.1 are limited to focused field campaigns lasting for short time periods with even fewer studies attempting source apportionment calculations (Kleeman et al., 2009). Multiple barriers have prevented the widespread deployment of PM_0.1 monitoring networks including (i) the low concentration of PM_0.1 mass, which challenges the detection limits of analytical methods, (ii) the artifacts associated with collecting PM_0.1 samples, (iii) the additional workload involved in operating the collection devices, and (iv) the sharp spatial gradients of PM_0.1 concentrations. Expensive investments in PM_0.1 monitoring are unlikely to be purchased without compelling evidence linking PM_0.1 to public health. Early epidemiological studies for PM_0.1 must therefore use techniques other than direct measurements to calculate population exposure.

Various methods, such as the source-resolved PMCAMx chemical transport model, the chemical mass balance (CMB) model, photochemical box models, and land use regression (LUR) models, have been used to track source contributions to primary organic matter, elemental carbon, and in some cases particle number concentration (N_x) over areas in the Eastern US and parts of Europe and Asia (Lane et al., 2007; Posner and Pandis, 2015; Wang et al., 2011; Cattani et al., 2017; Wolf et al., 2017; Simon et al., 2018; Gaydos et al., 2005; Zhong et al., 2018). However, these methods are limited in one or more aspects of their ability to predict population exposure to ultrafine particles over large analysis domains. Source-resolved models, such as PMCAMx, have been used to resolve composition for N_x in the Eastern US but not for PM_0.1 (Posner and Pandis, 2015). CMB models need measurements of specific molecular markers at numerous sites to resolve the sharp spatial gradients of ultrafine particle source contributions. LUR models need comprehensive measurements that act as training data sets in order to extend throughout a modeling domain (Lane et al., 2007).

Hu et al. (2014) calculated population exposure to PM_0.1 in California using a regional source-oriented chemical transport model supported by measured profiles for the size and composition of particles emitted by dominant sources. Predictions were compared to all available fine and ultrafine particle measurements over the period 2000–2010 with good agreement observed for the dominant chemical components of PM_0.1 mass including organic aerosol, elemental carbon, and numerous trace metals (Hu et al., 2014). The 4 km spatial resolution used in these calculations supported multiple epidemiological studies based on spatial gradients of exposure (Ostro et al., 2015; Laurent et al., 2016). These encouraging results motivate the expansion of the PM_0.1 exposure technique to other locations.

Here we use the Eulerian source-oriented UCD–CIT chemical transport model to predict the concentration of PM_0.1 in 39 urban regions throughout the US during summer pollution events in 2010. The calculation tracks contributions from 15 primary particle sources through a simulation of all major atmospheric processes while retaining information about particle size, composition, and source origin (Hu et al., 2014). The results of this calculation reveal US national trends in PM_0.1 concentrations for the first time and suggest locations where the differential health effects of PM_0.1 and PM_2.5 can best be studied.

2 Methods

2.1 Simulation dates

A total of 39 of the largest cities in the continental US were selected as the primary target locations in the current study (Fig. 1). These cities have been used to characterize atmospheric reactivity across the US in previous air pollution studies (Carter, 1994, 2010; Venecek et al., 2018a, b). Simulations within each target city were carried out during peak summer air pollution events in 2010. Dates were selected based on an initial investigation of measured 1 h ozone (O₃) across all monitors in a core-based statistical area (CBSA). A CBSA is defined as a US geographical area that consists of one or more counties anchored by an urban center of at least 10 000 people plus adjacent counties that have a high degree of social and economic integration with the core as measured by commuting (United States Census Bureau, 2018).

Figure 1 Map of 39 cities used for the prediction of PM_2.5 and PM_0.1 source contributions across the continental United States during summertime air pollution events.

The selected air pollution events within each CBSA typically had measured 1 h maximum O₃ concentrations greater than 70 ppb. Regional pollution events caused by atmospheric stagnation were selected whenever possible as opposed to special events caused by unusual occurrences such as wildfires that affected only one city at a time. The simulation dates in each city are listed in Table 1. Figure 2 illustrates the average 1 h maximum O₃ concentration across all monitors within each CBSA during the selected regional events. Simulation periods are organized in chronological order for the year 2010, and cities within the same geographical region are grouped together. Measured 24 h PM_2.5 concentrations during peak summer pollution events ranged between 3.2 and 30 µg m⁻³ depending on the location. The aggregation of these events across the US enables a comparison of typical summertime air pollution episodes within different cities.

Figure 2 Average 1 h maximum O₃ across all monitors in each domain. Cities are grouped by corresponding extreme O₃ dates (that averaged >70 ppb) and US geographical region.

Table 1 City, city code, simulation date, 2010 population, and geographical region.

2.2 Model description

The UCD–CIT model predicts the evolution of gas- and particle-phase pollutants in the atmosphere in the presence of emissions, transport, deposition, chemical reaction, and phase change (Held et al., 2005) as represented by Eq. (1):

$$\begin{array}{}\text{(1)}& \begin{array}{rl}{\displaystyle \frac{\partial C_i}{\partial t}}+\mathrm{\nabla }\cdot \mathit{u}{C}_i& =\mathrm{\nabla }K\mathrm{\nabla }{C}_i+E_i-S_i+R_i^{\mathrm{gas}}\left(C\right)\\ & +R_i^{\mathrm{part}}\left(C\right)+R_i^{\mathrm{phase}}\left(C\right),\end{array}\end{array}$$

where C_i is the concentration of gas- or particle-phase species i at a particular location as a function of time t, u is the wind vector, K is the turbulent eddy diffusivity, E_i is the emissions rate, S_i is the loss rate, $R_i^{\mathrm{gas}}$ is the change in concentration due to gas-phase reactions, $R_i^{\mathrm{part}}$ is the change in concentration due to particle-phase reactions, and $R_i^{\mathrm{phase}}$ is the change in concentration due to phase change (Held et al., 2005). Loss rates include both dry and wet deposition. Phase change for inorganic species occurs using a kinetic treatment for gas–particle conversion (Hu et al., 2008) driven towards the point of thermodynamic equilibrium (Nenes et al., 1998). Phase change for organic species is also treated as a kinetic process with vapor pressures of semi-volatile organics calculated using the two-product model (Carlton et al., 2010). More sophisticated approaches for secondary organic aerosol (SOA) formation (Cappa et al., 2016) were also tested in the current study but these required a larger number of assumptions and they did not produce higher SOA concentrations in the PM_0.1 size fraction.

Nucleation was included in the model using the ternary nucleation (TN) mechanism involving H₂SO₄–H₂O–NH₃ (Napari et al., 2002). A tunable nucleation parameter equal to 10⁻⁵ was used based on results from previous studies across California for the year 2012 (Yu et al., 2019). Yu et al. (2019) found good agreement between predicted and measured concentrations of daily-averaged PM_0.1 and N₇ source contributions in California. The current study expands these nucleation calculations to investigate new particle formation across all major US cities, but the data needed to evaluate the accuracy of these calculations are generally not available outside California, and particle number concentrations will not be a focal point of this work. The model spatial resolution was 4 km over the 4.2 million km² of simulated urban areas, so near-roadway concentrations of ultrafine particles on spatial scales of ∼0.1 km will not be presented.

A total of 50 particle-phase chemical species are included in each of 15 discrete particle size bins that range from 0.01 and 10 µm in particle diameter (Held et al., 2005). Artificial source tags are used to quantify source contributions to the primary particle mass and the secondary organic aerosol (SOA) mass for a specific bin size, thereby allowing the direct contribution of each source of PM_2.5 and PM_0.1 mass to be determined. Gas-phase concentrations of oxides of nitrogen (NO_x), volatile organic compounds (VOCs), oxidants, O₃, and semi-volatile reaction products were predicted using the SAPRC-11 chemical mechanism (Carter and Heo, 2013).

2.3 Model inputs

Anthropogenic emissions were generated using the Sparse Matrix Operator Kernel Emissions (SMOKEv3.7) modeling system applied to the 2011 National Emissions Inventory. The NEI reports county-wide emission totals from all 50 states that are then mapped using spatial surrogates. Temporal profiles are also used to account for variation by time of month, week, and day; however, the NEI does not account for “no-burn” days that would impact residential wood combustion or precipitation events that would impact paved–unpaved road dust. These default profiles may result in larger model performance bias when comparing predictions to measured values. Emissions from each of the four major source sectors (area, mobile, non-road, and point) were tagged to create 15 different emissions groups: on-road diesel, on-road gasoline, off-road diesel, off-road gasoline, biomass, cooking, natural gas, process heaters, distillate (oil), aviation, cement, coal, steel foundries, paper products, and all other emissions. Size- and composition-resolved source profiles were then assigned to the PM emissions within each of these groups using the UCD–CIT emissions processor based on the most recent measurements available in the literature (Robert et al., 2007a, b; Kleeman et al., 2008). Some of the 15 source categories were represented using weighted-average source profiles from multiple sources as described in Table S1 in the Supplement.

Daily values for 2010 wildfire emissions were generated using the Global Fire Emissions Database (GFED) (Giglio et al., 2013). Biogenic emission rates were generated using the Model of Emissions of Gases and Aerosols from Nature (MEGANv2.1). The gridded geo-referenced emission factors and land cover variables required for MEGAN calculations were created using the MEGANv2.1 preprocessor tool and the ESRI_GRID leaf area index and plant functional type files available at the Community Data Portal (Guenther et al., 2012).

Meteorology parameters used to drive the UCD–CIT chemical transport model (CTM) and the MEGANv2.1 biogenic emissions were generated using the Weather Research and Forecasting model (WRFv3.6) and WRF preprocessing system (WPSv3.6). Meteorological fields were created within three nested domains with horizontal resolutions of 36, 12, and 4 km. Each domain had 31 telescoping vertical levels up to a top height of 12 km. Four-dimensional data assimilation (FDDA) or “FDDA nudging” was used to anchor meteorological predictions to measured values (Hu et al., 2010). Meteorological data and gridded map projections needed for 2010 emissions modeling were taken from the corresponding WRF simulations using the meteorology–chemistry interface processor (MCIP).

2.4 Supporting measurements

Ambient hourly O₃ measurements and daily PM_2.5 measurements were obtained from the Environmental Protection Agency (EPA) AQS API/Query AirData (EPA, 2017b). Model predictions are compared to these measurements to build confidence in the accuracy of the overall modeling system since PM_0.1 measurements are not available during any of the peak summer pollution events studied here.

3 Results

Predicted maximum 1 h O₃, NO₂, SO₂, and CO concentrations were compared to measurements at all available monitors within each study CBSA to indirectly evaluate the accuracy of the emissions inventories and meteorology fields. Many of the sources that emit O₃ precursors also emit ultrafine particles. Likewise, meteorological parameters like wind speed and mixing depth influence the concentrations of all pollutants, including ultrafine particles. Successful prediction of gas-phase species is therefore a necessary step in the accurate prediction of ultrafine particle concentrations during summer photochemical smog episodes. Predicted 24 h PM_2.5 concentrations were also compared to measurements at all available monitors within each study CBSA. Many of the combustion sources that emit primary particles within the PM_2.5 size fraction also emit primary PM_0.1 and/or precursor gases that can condense into the PM_0.1 size range. The Chemical Speciation Monitoring Network (CSN) operated by the U.S. Environmental Protection Agency (EPA) measures PM_2.5 mass and chemical composition at more than 260 sites throughout the US, including many of the 39 cities studied in the current analysis (Solomon et al., 2014). Full monitor information including latitude, longitude, and total number of available measurements for comparison within the simulation period are show in Tables S2–S6.

Figure 3 Model performance statistics for predicted maximum 1 h O₃ against measured values. The red line represents performance criteria of 0.15 for NMB and 0.25 for NME. NMB and NME were calculated for available measurements against predictions at every monitor in the CBSA based on the U.S. EPA AQ Data Mart. Monitor number (horizontal axis), latitude and longitude, name, MO, MP, NMB, NME, FB, and FE values are available for all monitors in the Supplement.

Figure 3 illustrates the normalized mean bias (NMB) and normalized mean error (NME) for predicted 1 h maximum O₃ against measured 1 h maximum values for each monitor within a specific modeling domain. Figure S1 in the Supplement illustrates the fractional bias (FB) and fractional error (FE) for predicted 1 h maximum CO, NO₂, and SO₂ against measured 1 h maximum values. Figure 4 illustrates the NMB and NME for 24 h average predicted PM_2.5 concentrations against measured 24 h average PM_2.5 concentrations at each available monitor over the specific simulation period. A time series of predicted vs. measured O₃ concentrations is displayed in Fig. S2.

Figure 4 Model performance statistics for predicted 24 h average PM_2.5 against measured values. The red line represents performance criteria of ±0.30 for NMB and 0.50 for NME. Normalized mean bias and normalized mean error were calculated for available measurements against predictions at every monitor in the CBSA based on the U.S. EPA AQ Data Mart. Monitor number (horizontal axis), latitude and longitude, name, MO, MP, NMB, NME, FB, and FE values are available for all species in the Supplement.

Table 2 Percent of monitors throughout the entire US domain that met performance criteria for normalized mean error (NME).

Table 2 summarizes the total number of available monitors for a comparison of measured values vs. predicted values for O₃ and PM_2.5. Emery et al. (2017) recommend model performance criteria for 1 h O₃ NMB less than or equal to ±0.15 and NME less than or equal to 0.30. The 24 h PM_2.5 model performance recommendations, also based on Emery et al. (2017), are NMB less than or equal to ±0.30 and NME less than or equal to 0.50 (Emery et al., 2017). Table 2 displays the percentage of measured vs. predicted comparisons that met the performance criteria for NME over the entire US modeling domain. In summary, 95 % of all locations met NME performance criteria for O₃ predictions, and 85 % of all locations met NME performance criteria for PM_2.5 predictions.

Elemental carbon (EC) and organic carbon (OC) are the chemical components most relevant for both the PM_2.5 and the PM_0.1 size fractions. Figure 5 illustrates predicted vs. measured 24 h PM_2.5 EC and OC concentrations for all 39 cities. Primary organic matter tracked by model calculations is converted to OC by dividing by a factor of 1.2 (Russell, 2003). Secondary organic aerosol tracked by model calculations is converted to OC by dividing by a factor of 1.5. In general, the model slightly underpredicts PM_2.5 EC, OC, and mass with regression slopes ranging from 0.62 for EC to 0.71 for OC. The negative bias in model predictions may stem from the 4 km spatial averaging inherent in the calculations vs. the influence of sources closer than 4 km to the measurement site in urban environments, such as highways and restaurants. Model performance statistics for PM_2.5 predictions are summarized in Table S6.

Figure 5 Predicted average 24 h vs. measured 24 h average (a) organic carbon and (b) elemental carbon (µg m³). Predicted OC was converted from predicted organic matter (OM) and secondary organic aerosol components using a ratio of 1.2 and 1.5, respectively (Russell, 2003).

PM_0.1 measurements are not available for model evaluation in the 39 cities across the US in 2010 at the core of the current study, but measurements are available in California in the years 2015 and 2016 that can be used to evaluate similar modeling procedures. Yu et al. (2019) compared PM_0.1 concentrations in Los Angeles, Fresno, East Oakland, and San Pablo, California, predicted using the UCD–CIT air quality model to receptor-based source apportionment calculations based on measured concentrations of molecular markers in the ultrafine particle size fraction (Xue et al., 2018). Good agreement was found between predictions for PM_0.1 concentrations associated with gasoline engines, diesel engines, cooking, wood burning, and “other sources” from these two independent techniques. Further details on this comparison are provided by Yu et al. (2019). This evaluation of the modeling procedures builds confidence in the PM_0.1 source predictions across the US in the current study, but new measurements would be helpful to fully evaluate model predictions in the future.

Figure 6 (a) PM_2.5 and (b) PM_0.1 24 h average mass (µg m⁻³) during summer air pollution event. Scale drawn to highlight all areas of the US. Actual maxima: (a) 94.25 µg m⁻³; (b) 9.43 µg m⁻³.

Figure 6 illustrates a composite representation of PM_2.5 and PM_0.1 mass across the US during the summer pollution episodes listed in Table 1. The spatial plot in Fig. 6 is constructed using the intermediate 12 km simulation results from multiple simulations stitched together to cover a broader geographical area. Regional PM_0.1 concentrations reach a maximum value of 5 µg m⁻³ in a few isolated grid cells with wildfires, but concentrations generally exceed 2 µg m⁻³ in major urban regions across the US, including Los Angeles, the San Francisco Bay Area, Houston, Miami, and New York. The comparison between PM_2.5 mass (Fig. 6a) and PM_0.1 mass (Fig. 6b) shows that predicted PM_0.1 spatial gradients are sharper, with fewer regional contributions between “hot spots”. Locations in the Midwestern and Eastern US outside cities with high PM_2.5 concentrations due to secondary formation (sulfate and secondary organic aerosol) did not have corresponding high concentrations of PM_0.1. Most major urban centers had noticeable peaks of both PM_2.5 and PM_0.1. This pattern presents a challenge for epidemiological studies seeking to differentiate the effects of PM_2.5 and PM_0.1 because the locations with differential exposure (high PM_2.5 but low PM_0.1) have a low population density, which will reduce the power of the analysis.

Figure 7 PM_0.1 source contribution for 39 cities across the continental US.

The UCD–CIT model explicitly tracks source contributions to particle mass in each size bin using artificial source tags. Pie charts of PM_2.5 and PM_0.1 source contributions are illustrated in Fig. 6 for selected major cities. Pie charts for PM_0.1 source contributions in all 39 US cities are shown in Fig. 7. The detailed source profiles within each city are based on the nested 4 km simulation results during the pollution events listed in Table 1. Source contribution spatial plots for the entire US are shown in Figs. S3 through S5, and pie charts for PM_2.5 source contributions in all 39 US cities are shown in Fig. S6. On-road gasoline and diesel vehicles made significant contributions to regional PM_0.1 in all 39 cities even though peak contributions within 0.3 km of the roadway were not resolved by the 4 km grid cells. Cooking also made significant contributions to PM_0.1 in all cities, but biomass combustion was only important in locations impacted by summer wildfires. Residential wood combustion is not typically a strong source in the summer due to warmer temperatures; however, in the wintertime biomass would most likely be a dominant source. Aviation was a significant source of PM_0.1 in cities that have airports within their urban footprints. Industrial sources including cement manufacturing, process heating, steel foundries, and paper and pulp processing impacted their immediate vicinity but did not significantly contribute to PM_0.1 concentrations in any of the target 39 cities. Natural gas combustion made significant contributions to PM_0.1 concentrations due to the widespread use of this fuel for residential, commercial, and industrial applications. Natural gas contributions were especially significant in locations with high levels of industrial use, such as chemical refineries, and in locations with significant levels of natural-gas-fired power plants.

The major sources of primary PM_0.1 and PM_2.5 were notably different in many cities (compare Fig. 6a and b). The sources that contribute most strongly to PM_2.5 are on-road diesel, gasoline, cooking, coal, and “other”, which includes brake and tire wear from mobile sources and dust. Natural gas combustion makes minor contributions to primary PM_2.5 mass since particles from this source have a mass distribution peaking at ∼0.05 µm in particle diameter (Chang et al., 2004), with all of the emitted mass in the PM_0.1 size fraction. In contrast, other combustion sources using more complex fuels, such as on-road vehicles, have a mass distribution peaking at ∼0.1 µm, with at least half the emitted mass outside the PM_0.1 size fraction (Robert et al., 2007a, b). Likewise, cooking contributes strongly to PM_2.5 concentrations, but the emitted particle mass distribution peaks at 0.2 µm, with the majority of the mass outside the PM_0.1 size fraction.

The fraction of PM that is primary within each CBSA is listed in Tables S7–S16. Averaged across the US, PM_2.5 was found to be approximately 62 % primary material, while PM_0.1 was found to be approximately 87 % primary material.

Figure 8 Population-weighted average source contribution across the 39 major cities in the continental US for (a) PM_2.5 and (b) PM_0.1.

4 Discussion

Figure 8 illustrates the population-weighted average PM_0.1 source contributions across all 39 study cities shown in Table 1. These predictions are based on source profile measurements for wood burning, cooking, mobile sources, and nonresidential natural gas combustion reported in multiple peer-reviewed studies (Taback et al., 1979; Cooper, 1989; Houck et al., 1989; Hildemann et al., 1991a, b; Harley et al., 1992; Schauer et al., 1999a, b, 2001, 2002a, b; Kleeman et al., 2008, 2000; Robert et al., 2007a, b). In addition, new measurements made by Xue et al. (2019) were conducted to confirm previous measurements of the particle size distribution associated with natural gas and biomethane combustion particles.

Figure 9 (a) Natural gas compressor stations and pipelines across the US and (b) PM_0.1 natural gas combustion concentrations (µg m⁻³).

Figure 10 Scatter plot showing correlation between 24 h average PM_2.5 and PM_0.1 for the 39 cities.

Figure 11 PM_0.1∕PM_2.5 ratio for each city.

Figure 12 PM_0.1∕PM_2.5 ratio across the US.

The results summarized in Fig. 8 highlight the importance of natural gas combustion particles in the PM_0.1 size fraction and the minor role that these natural gas combustion particles play in the PM_2.5 size fraction. Natural gas typically consists of +93 % methane, with the balance of the fuel made up by higher-molecular-weight alkanes and trace impurities. In addition to background sulfur compounds in the natural gas, sulfur-containing odorants such as mercaptans are commonly added to aid in leak detection. Natural gas combustion does not emit high amounts of particulate matter per joule of energy in the fuel, but the widespread use of natural gas suggests that it could still contribute significantly to ambient PM_0.1 concentrations. Natural gas combustion accounted for 29 % of total US energy consumption in 2016 (U.S. Department of Energy, 2017). In contrast, gasoline combustion accounted for 17 % of US energy consumption, and diesel fuel combustion accounted for approximately 6 % of US energy consumption in 2016. Less than half of the PM emitted by gasoline and diesel fuel combustion is in the PM_0.1 size fraction (Robert et al., 2007a, b), whereas all of the PM emitted by natural gas combustion is in the PM_0.1 size fraction (Chang et al., 2004). Taken together, these facts support the potential importance of natural gas combustion for ambient PM_0.1 concentrations.

The five states with the highest consumption of natural gas in 2016 were Texas (14.7 %), California (7.9 %), Louisiana (5.7 %), New York (5 %), and Florida (4.8 %). These consumption patterns are reflected in the natural gas distribution system (Fig. 9a) and the predicted PM_0.1 concentration field associated with natural gas combustion (Fig. 9b). Natural gas end use included electric power generation (36 %), industrial applications (34 %), residential use (16 %), commercial use (11 %), and transportation (3 %).

Lane et al. (2007) used a source-resolved version of PMCAMx and individual emission inventories to determine source contributions of primary organic material (POM_2.5) (Lane et al., 2007). Lane et al. (2007) note that POM_2.5 associated with natural gas sources ranged from 0.1 to 0.8 µg m⁻³. Chang et al. (2004) measured emitted particle size distributions for gas-fired stationary combustion that fell between 10 and 100 nm. The combination of these two results indicates that the natural gas mass component of POM_2.5 predicted by Lane et al. (2007) is consistent with the magnitude of the PM_0.1 mass associated with natural gas combustion found in the current study. Lane et al. (2007) were not studying PM_0.1, so the major role of natural gas in this size fraction was not identified.

Posner and Pandis (2015) utilized PMCAMx with the LADCO 2001 BaseE source-resolved mass emissions inventory for a July 2001 prediction of N_x over the Eastern US with 36 km resolution (Posner and Pandis, 2015). Posner and Pandis used a “zero-out” method in combination with source-specific size distribution to study the percent contribution of six major sources (on-road gasoline, industrial, non-road diesel, on-road diesel, biomass, and dust) of N_x. They found that N_x was made up of 36 % on-road gasoline, 31 % industrial, 18 % non-road diesel, 10 % on-road diesel, 1 % biomass burning, and 4 % long-range transport (Posner and Pandis, 2015). The emissions particle number inventory was normalized based on PM₁₀ mass from each source and particle emissions from natural gas sources were assumed negligible, which effectively removed natural gas sources from the simulation. This has minor effects on PM_2.5 and PM₁₀ predictions, but the results of the current study suggest that natural gas combustion significantly contributes to ultrafine particle concentrations.

Future epidemiological studies may be able to differentiate PM_0.1 and PM_2.5 health effects by contrasting cities with different predicted ratios of PM_0.1∕PM_2.5. Although the current study does not calculate the annual average concentrations that would be needed for such an analysis, the results for the peak photochemical episodes may provide some useful insights to guide future studies. Figure 10 illustrates the correlation between predicted PM_2.5 and PM_0.1 concentrations in the 39 cities considered in the current analysis, Fig. 11 illustrates the ratio of PM_0.1∕PM_2.5 for each city, and Fig. 12 illustrates a field plot showing the ratio of PM_0.1∕PM_2.5 across the continental US. Cities with higher PM_0.1∕PM_2.5 ratios include Houston, TX, Los Angeles, CA, Salt Lake City, UT, Cleveland, OH, and Bakersfield, CA. Cities with lower PM_0.1 to PM_2.5 ratios include Lake Charles, LA, Baton Rouge, LA, St. Louis, MO, Baltimore, MD, and Washington, D.C. Measurements should be conducted in these locations to verify the contrast in PM_0.1∕PM_2.5 concentrations in preparation for future exposure analysis.

5 Conclusion

The UCD–CIT regional chemical transport model was used to predict source contributions to PM_0.1 across the continental US during peak photochemical smog periods in the year 2010. Performance for PM_2.5 and O₃ predictions met or exceeded the criteria typically used for regional air quality model applications, building confidence in the emissions inputs and meteorological fields used to drive the calculations. Similar model exercises carried out for episodes in California in 2015 and 2016 find good agreement between predicted PM_0.1 source contributions and receptor-based PM_0.1 source contributions calculated using measured concentrations of molecular markers (Yu et al., 2019). In the current study, predicted regional PM_0.1 concentrations exceeded 2 µg m⁻³ during summer pollution episodes in major urban regions across the US including Los Angeles, the San Francisco Bay Area, Houston, Miami, and New York. Predicted PM_0.1 spatial gradients were sharper than predicted PM_2.5 spatial gradients due to the dominance of primary aerosol in PM_0.1. This finding suggests that the PM_0.1 measurement networks needed to support epidemiology must be denser than comparable PM_2.5 measurement networks. Nonresidential natural gas combustion was identified as a major source of PM_0.1 across all major cities in the US. On-road gasoline and diesel vehicles contributed on average 14 % to regional PM_0.1 even though peak contributions within 0.3 km of the roadway were not resolved by the 4 km grid cells. This is consistent with other studies that have found an exponential decrease in ultrafine particle concentrations downwind of major roadways (Wang et al., 2011) due to the sharp gradient of PM_0.1. Cooking also made significant contributions to PM_0.1 in all cities, but biomass combustion was only important in locations impacted by summer wildfires. Aviation was a significant source of PM_0.1 in cities that have airports within their urban footprints. The major sources of primary PM_0.1 and PM_2.5 were notably different in many cities. Future epidemiological studies may be able to differentiate PM_0.1 and PM_2.5 health effects by contrasting cities with different ratios of PM_0.1∕PM_2.5 sources.