Trends in concentrations of atmospheric gaseous and 1 particulate species in rural eastern Tennessee as related to 2 primary emissions reductions 3

11 Air quality measurements at Look Rock, Tennessee—on the western edge of the Great Smoky 12 Mountains National Park—were begun in 1980 and expanded during the 1980s to a National 13 Park Service (NPS) IMPROVE network station. Measurements were expanded again by the 14 Tennessee Valley Authority (TVA, 1999-2007) to examine the effects of electric generating unit 15 (EGU) emission reductions of SO 2 and NO x on air quality at the station. Analysis of temporal 16 trends (1999-2013) has been conducted at the site in collaboration with activities related to the 17 2013 Southeast Atmosphere Study (SAS) at Look Rock and other southeastern U.S. locations. 18 Key findings from these trend studies include the observation that primary pollutant levels have 19 consistently tracked emissions reductions from EGUs and other primary sources in the region but 20 reductions in secondary pollutants such as particulate sulfate and, specifically, ozone have been 21 smaller compared to reductions in primary emissions. Organic carbonaceous material (OM) 22 remains a major contributor (30-40 percent in the period 2009-2013) to fine particulate mass at 23 the site, as confirmed by ACSM measurements at the site in 2013. A large portion (65-85 24 percent) of carbon in OM derives from modern carbon sources based on 14 C measurements. 25 Important parameters affecting ozone levels, fine mass and visibility also include the specific 26 diurnal meteorology at this ridge-top site, its location in a predominantly mixed-deciduous forest,

Currently, special studies at the site include the deployment of an aerosol chemical speciation 1 quality assurance requirements as outlined in EPA NCore network strategy (National Ambient 1 Air Monitoring Strategy: http://www.epa.gov/ttn/amtic/files/ambient/monitorstrat). The 2 IMPROVE TEOM PM 2.5 mass had to be corrected to adjust for the fact that the NPS operated 3 the TEOM from the beginning at 30°C but used the calculation algorithm originally designed for 4 operation at 50°C. To avoid bias in reported mass (the lower temperature better preserves 5 ammonium nitrate and some semi-volatile OC) compared with standard filter-based PM 2.5 data, 6 TEOM values were corrected based on an algorithm developed to produce results which were 7 highly correlated with reported FRM PM 2.5 mass with essentially zero intercept 8 Annual quarters 1 (January-March), 2 (April-June), 3 (July-September) and 4 (October-9 December) were defined as winter, spring, summer and autumn, respectively, for determining 10 seasonal data trends for investigating meteorological influences. The approach permits spring 11 and summer to be consistently the periods with highest temperature (T), solar radiation, and leaf 12 coverage, while fall and winter have lower T and solar radiation and very little deciduous leaf 13 cover. Major differences in trends are unlikely when compared to seasonal data binned 14 according to solstice and equinox dates or using alternate definitions such as winter = December-15 February. 16

Influence of meteorological patterns 17
Temporal patterns in meteorology may influence Look Rock air quality due to the dependence of 18 atmospheric chemistry and pollutant transport on variations in temperature, solar radiation, cloud 19 cover, precipitation and wind patterns. Patterns were based on data collected at Look Rock and 20 the nearest National Weather Service (NWS) observations at the Knoxville airport. Data from 21 these two sites are described in Supplemental Material. Few significant (i.e., >95 percent 22 confidence) trends were identified over the time period from 1999 through 2013. 23 Atmospheric transport was inferred from computed three-dimensional air trajectories. Twenty-24 four hour air trajectories ending at Look Rock were computed using the NOAA HYSPLIT Model 25 (www.arl.noaa.gov/HYSPLIT_info.php; Draxler and Hess, 1998; Draxler, 1999; Draxler and 26 Rolph, 2014). Trajectories were based on the NOAA-NCAR Global Reanalysis meteorological 27 data set and used diagnosed vertical velocity to determine vertical air parcel motion. Trajectories 28 were derived for days when TEOM data indicated fine particulate levels in either the lowest or highest 5 percent of the distribution of all hourly PM 2.5 concentrations in each month. This 1 ensured that we examined trajectory classes that contributed to the full range of particulate levels 2 at Look Rock. Two or more trajectory analysis days were selected monthly during 2007- 2013 3 (ending in June of 2013) yielding a data set of 174 trajectories. Trajectories were computed to 4 arrive at 100, 500 and 1500 m above Look Rock at local midnight on the trajecgory analysis 5 days. Trajectories arriving at 1500 m agl were so similar to those for 500 m that they provided 6 little additional information and were not included in a subsequent cluster analysis. 7 Cluster analysis is a useful aggregation tool for classifying data into groups with similar 8 locations in K-dimensional space (Wilks, 2006). Hourly upwind trajectory coordinates were 9 available for each combination of trajectory arrival date and trajectory height. Trajectory latitude 10 and longitude coordinates at 12 and 24 hours upwind and at 100 and 500 m arrival heights were 11 the focus of a cluster analysis designed to identify similar atmospheric transport pathways 12 approaching Look Rock. In this case K=2 and spatial coordinates were defined using latitude 13 and longitude transformed into orthogonal variables with unit variance and mean of zero. The 14 transformation was made using principal component analysis (Preisendorfer, 1988)  Even so, trend values for individual clusters are not as important as cluster trends sorted by PM 2.5 25 concentration groupings. A comparison across PM 2.5 groups indicated that temporal trends were 26 significant and downward only for clusters that fell into the high PM 2.5 group. In the subsequent 27 discussion, the trajectory cluster number refers to the hierarchy of clusters starting with the 28 cluster (assigned a value of one) having the most trajectories for which some predetermined 29 criteria are met that define a property of the trajectories. In the present study, Ward's method of have fewer trajectories. Clusters are built starting with each trajectory in its own cluster and 1 aggregating trajectories on subsequent passes through the data. This process can continue only 2 until all trajectores are clustered into one group (such an endpoint has no value, however). The 3 analyst must select some number of clusters >1 for which the highest ranked clusters contain a 4 useful amount of trajectories that are classified. Once a trajectory is placed into a cluster it is not 5 removed. 6

Trend determination 7
In the subsequent discussion, a determination of "significance" when evaluating the presence of 8 a trend, when comparing changes in emissions and air pollutant concentrations or when 9 comparing air quality and meteorology was based on a statistical association computed (using 10 least-squares regression) to have a p value <0.05. Temporal changes in both domain-wide annual 11 NO x and SO 2 emissions (all sources) were significant with p <0.001 (i.e., confidence exceeding 12 99.9 percent). Seasonal emissions were not estimated (seasonal emissions for non-EGU sources 13 must be modeled and this adds another level of complexity and uncertainty to the trends 14 analysis). Trends in total seasonal emissions are expected to be similar to annual trends with 15 some seasons experiencing more changes than others depending on the timing of specific 16 regulatory drivers. This assumption of similarity is used in all subsequent comparisons between 17 seasonal air quality values and emissions and contributes to uncertainty when comparing them. 18 The convention for comparing emissions, air quality and meteorological values averaged over 19 time (annual or seasonal) involved converting actual values into deviations from the 15-year 20 average and then normalizing the deviation by the standard deviation of the 15-year value: 21 In Eq. (1) the overbar denotes the multi-year average of variable x while σ x denotes its 23 corresponding standard deviation. This scaling of x reduces the impact of outliers (<5 percent of 24 all air quality values of x fell outside ±2σ x ) and allows for a straightforward comparison between 25 different data sets. Thus, when comparing two variables x 1 and x 2 , a significant association 26 with mean regression slope Δ x 2 /Δ x 1 =1 implies that a one-standard deviation change in variable 27 associations in this manner allows a direct comparison between variable sensitivities (e.g., 1 Δ x 2 /Δ x 1 versus Δ x 3 /Δ x 1 ). In the current context, x represents the annual or single-season 2 average (or sum as in the case of precipitation) of a variable such as ozone concentration or 3 temperature. indistinguishable from that for the third quarter (July-August) labeled "summer". The trends are 11 similar to that observed for O 3 using the National Ambient Air Quality Standard (NAAQS) 12 metric-based on the annual fourth highest of the maximum daily 8-hr average (not plotted here) 13 and indicate only a slightly negative trend for 1999-2011. This change is much more modest 14 (about 0.3 ppbv yr -1 or <1 percent yr -1 ) than the trend in domain emissions of NO x (an O 3 15 precursor) for the 2002-2011 period. The O 3 trends for the second (-0.5 ppbv yr -1 ; -0.9 percent 16 yr -1 ; -0.13σ yr -1 ) and third quarters (-1.0 ppbv yr -1 ; -1.9 percent yr -1 ; -0.15σ yr -1 ) were significant 17 at p<0.05 but this was not true of the annual (-0.3 ppbv yr -1 /-0.11σ yr -1 ), first-quarter (0.2 ppbv 18 yr -1 /0.09σ yr -1 ) and fourth-quarter (0.1 ppbv yr -1 /0.03σ yr -1 ) changes. Likewise, the association 19 between quarterly O 3 and annual NO x emissions was positive and highly significant for the 20 spring and summer quarters (0.7< Δ x O3 /Δ x NOx <0.8), when high photochemical reactivity should 21 enable regional emissions to have the greatest impact on ozone observed at Look Rock, but not 22 for the other quarters when ozone had small, positive changes. 23 Zero or slightly upward O 3 patterns during the first and fourth quarters-when photochemistry is 24 low and contributions from local and regional sources are at a minimum-have occurred despite 25 reductions in regional NO x emissions. These observations may be due in part to reduced emitted 26 NO x titrating less O 3 during the winter and autumn periods. If true, this pattern implies the 27 relative importance of long-range (i.e., background) ozone transport during the cooler seasons. improvements in the worst O 3 levels at a faster rate than seen at most rural sites nationally. 18

19
Annual SO 2 concentrations, based on high-sensitivity hourly pulsed fluorescence data measured 20 at Look Rock, are available from 2007 to present and show about a 25 percent yr -1 decrease over 21 the period, compared to a mean reduction rate of 11 percent yr -1 since 1999 based on CASTNET 22 weekly SO 2 data ( Table 2). This reflects the large reductions in SO 2 emissions in the domain. 23 Compare the relative reduction rate of SO 2 with the reduction rate expressed as a fraction of a 24 standard deviation in the concentration ( Table 2): 11 percent yr -1 is equivalent to an annual 25 decline of 0.42σ yr -1 (i.e., almost a half of a standard deviation per year). 26 Annual averages of SO 2 concentrations from 1999 to the present from CASTNET weekly data 27 and continuous hourly pulsed fluorescence data are shown in Figure 4 along with a comparison 28 with NEI and EGU emissions of SO 2 in the domain (see discussion below). Hourly, continuous ( Figure 5) due to a combination of seasonal changes in emissions, gas-to-particle conversion and 1 meteorology-driven scavenging rates. Monthly comparison with hourly SO 2 data is less than 2 ideal because CASTNET SO 2 data are averaged weekly but, given the similarity in changes, it is 3 reasonable to expect that the conclusions would be the same.  year (trend ≈-0.4 ppbv yr -1 independent of season). This result is similar to that for regional NO x 17 annual emissions. However, despite similar tendencies, ambient NO y was not significantly 18 associated with annual NO x emissions except during the third and fourth quarters. This is likely 19 due to the extreme nonlinearity in NO-NO 2 -NO z (NO z =NO y -NO x ) chemistry and the spatial 20 inhomogeneity of NO y . It is possible that the significant associations during two quarters were 21 coincidental or represented a serendipitous agreement between regional emissions trends and 22 trends for sources directly impacting Look Rock air quality during those periods. 23 In addition, O 3 was not significantly associated with NO y except during the third (summer) 24 quarter when a positive association was determined (Δ x O3 /Δ x NOy =0.76). Again, this may be 25 because Look Rock NO y data do not reflect regional NO x emissions. Look Rock ozone appears 26 to only respond significantly (in a positive sense) to local NO y levels during the summer quarter 27 when weak transport conditions and high photochemical rates coincide. In that quarter, NO y was 28 significantly linked to only two variables: wind speed (positive sense) and cloud cover (negative quarters despite the fact that regional NO x emissions were in strong decline. The conclusion 1 from these results is that local conditions at Look Rock that are influenced strongly by 2 atmospheric photochemical processes (especially ozone and secondary organic aerosol 3 formation) are only characterized by Look Rock measurements during the summer quarter. Data 4 from other times of the year indicate a blend of impacts from local and more distant emissions, 5 consistent with the stronger atmospheric transport and/or slower photochemical processing 6 typical for those seasons.  Table 2 and range from 2.0 percent to 3.6 percent per yr -1 20 since 1988. Note that the levels of OC as measured by the IMPROVE method actually increased 21 at Look Rock from 1988 to 1998 then decreased from 1998 to the present at a net rate of 2.7 22 percent yr -1 . 23 Comparing Look Rock PM 2.5 trends with those from other locations is more complicated than for 24 O 3 because aerosols impacting a given site often come from a different mix of pollution source 25 types and pollutant precursor species. Across the eastern U.S. the predominant source of PM 2.5 26 has been sulfate from SO 2 emitted by fossil fuel combustion. However, some locations are more 27 influenced than others by biomass burning, biogenic organic aerosols, windblown dust and so same regions that impact Look Rock). Trends in these different source types influence overall 1 trends in PM 2.5 pollution. The U.S. EPA tracks trends in annual average PM 2.5 concentrations at 2 both the national and regional level (www.epa.gov/airtrends/pm.html#pmreg). Data from 2000- Since 1999 the changes in aerosol mass and its major constituents are all in a narrow range from 12 3.1 to 4.2 percent yr -1 despite significant inter-annual variability. Comparing the results reported 13 in Table 2  that in 2013 NH 4 NO 3 varied between 4 (summer) and 9 (winter) percent of total PM 1 mass 27 ("PM 1 " denotes aerosols with aerodynamic diameters <1 µm) mass. 28 Comparisons of aerosol components against emissions revealed that sulfate, OC and EC (and by 29 extrapolation, PM 2.5 ) were significantly associated with one or both emission species for at least The strong covariance between SO 2 and NO x emissions produces cross-correlations with any 1 aerosol component that is significantly associated with at least one of the emissions species. 2 However, sulfate was more closely associated (higher r 2 and lower p values; p <0.001) with SO 2 3 emissions for each quarter of the year (0.60<Δ x sulf /Δ x NOx <0.76 and 0.76<Δ x sulf /Δ x SO2 <1.0). 4 Conversely, OC was strongly associated with NO x emissions during the spring (p <0.01) and 5 summer (p < 0.001) quarters and OC was associated to a lesser degree with spring, summer and 6 One 7 relationship that stands out is between OC and ambient NO y(g) . Similar to what was reported in 8 the previous section regarding ozone, OC was significantly associated with NO y(g) only during 9 the third quarter (Δ x OC /Δ x NOy =0.86), perhaps reflecting a unique combination of factors 10 occurring only during summer. Finally, EC was more closely linked with SO 2 emissions than 11 NO x emissions for each quarter (p ≤0.001 for SO 2 in all seasons) although the SO 2(emiss) :NO x(emiss) 12 covariance muddles interpretation of the data. 13 Another notable change in aerosol composition is reflected in the relative abundances among the 14 primary species measured at Look Rock. The change in relative abundances of the major fine 15 particle types is best seen by examining the IMPROVE data from 1999 and 2013 (

Relationship to long-term trends in meteorology 26
It is reasonable to consider whether these trends in aerosol and trace gas levels are associated 27 with meteorological trends at Look Rock. Data from the NWS at the nearby Knoxville airport in either annual or seasonal precipitation or temperature trends, nor are these meteorological 1 variables correlated with measured air quality changes (details on the long-term trends in 2 temperature and precipitation and the year-to-year variability in those trends at Look Rock are 3 given in the Supplemental Material). Likewise, no significant 15-yr trends in wind direction 4 frequency, cloud cover or solar radiation have been detected. A slight upward trend in solar 5 radiation since 2007 was found which would suggest an increase in photochemistry, but the net 6 effect on secondary aerosol production is unclear. 7 A small decline (not statistically significant) since 1999 in the number of high moisture hours 8 (i.e., those with relative humidity >90 percent) was observed during all seasons but autumn. 9 Considering that the inter-annual variations in this parameter were large, the effect on observed 10 aerosol trends (such as through reduced heterogeneous chemistry) is likely to be small. The only 11 statistically significant, incontrovertible meteorological trend is that of wind speed which has 12 declined steadily since 1999 at Look Rock. This is consistent with wind speed measurements Some meteorological factors, when viewed on a quarterly basis, were found to be significantly 24 associated with measured air quality values although they were not contributory to long-term 25 trends. Table 3 summarizes the notable linkages between air quality and meteorology at Look 26 Rock (in all cases p <0.05). Details of these associations are described in Supplemental Material 27 to allow comparisons of relative sensitivities across different species and meteorological factors. 28 The quarter of the year when a significant association was identified is noted in the table for each 29 combination of air quality and meteorological factor. The summer (third) and autumn (fourth) 30 quarters had the most significant linkages. Wind speed was most frequently associated with air quality (positively in all cases). Thus, speed is not an indicator of dilution but rather something 1 else, perhaps an increasing linkage between Look Rock air quality and local emissions with an 2 increasing role for "Look Rock domain" emissions as speed declines. Ozone, OC and EC are 3 occasionally negatively associated with precipitation which acts as a scavenger of pollutants. 4 Ozone and OC are both positively linked with temperature during summer with higher 5 temperatures associated with more biogenic precursor emissions (Lamb et al., 1993) and faster 6 photochemical reactions (e.g., Alley and Ripperton, 1962;Seinfeld and Pandis, 1998). Cloud 7 cover and solar radiation generally are anti-correlated. Insolation (reduced by cloud cover) is a 8 driver for photochemistry although neither it nor clouds appear as significant drivers during all 9 quarters. Insolation was negatively associated with sulfate during the spring quarter for reasons 10 that are not apparent. Also unexpected is the negative association between ozone/OC/EC and 11 high relative humidity conditions (i.e., frequency of humidity levels >90 percent) at Look Rock. 12 Typically, clouds impact the monitoring site directly (i.e., place the site in fog) under such humid 13 conditions. The occasional negative association between some air quality levels (i.e., ozone and 14 OC) on one hand and cloud cover and high humidity on the other might represent the same effect 15 on photochemistry but the physical basis for an impact on EC is not obvious.  Table 3-may also come into play. This is supported by 26 seasonal trends data for EC and BC illustrated in Figure 11. "modern" (f m ) and fossil carbon can be determined because 14 C in fossil sources of carbon has 4 decayed to zero (Stuiver and Polach, 1977). With respect to control strategies for atmospheric 5 aerosol carbon in PM, this estimation is complicated by the fact that anthropogenic activities 6 such as agricultural burning and prescribed forest burns contribute modern aerosol carbon as do 7 vegetative emissions. An additional complication is the unexpected correlation of f M with certain 8 anthropogenic compounds (e.g. CO). Biomass burning samples present a challenge in 9 interpretation of 14 C content due to the varying age of carbonaceous material in biomass fuel and 10 due to the fact that atmospheric 14 C content is decaying to its cosmic ray background from higher 11 levels caused by nuclear bomb tests in the 20th century. One approach to resolving this problem 12 (not used herein) would involve using different correction factors for modern aerosol carbon  Table 4). As shown in Figure 6

Atmospheric transport and aerosol trends 1
Trends in PM 2.5 were examined within trajectory coordinate clusters (Section 2.3) to determine 2 which transport scenarios contributed most to the observed overall negative PM 2.5 trends 3 previously described (Table 5). Cluster populations and the fraction of total variance (in 4 transformed cluster coordinates) decrease as cluster number increases. Also, the lower cluster 5 numbers usually included data from the largest range in years and, thus, most often provided 6 statistically significant (with a minimum 95 percent confidence) trends in PM 2.5 . Locations of 7 the top 10 clusters (based on cluster-averaged 12-hr upwind trajectory locations) are illustrated in 8 Figure 12. Numbered ellipses contain the 100-m and 500-m trajectory centroids for each cluster. 9 Shaded clusters are those associated with significant downward trends in PM 2.5 concentrations. 10 Symbols denoting the locations of Look Rock and Knoxville are plotted for spatial reference. 11 Table 5 lists the PM 2.5 changes over time for the top eight 12-and 24-hr trajectory clusters (parts 12 a and b, respectively). These clusters accounted for >60 percent of the total variance in cluster 13 location for each upwind period and more than half the total number of trajectories analyzed. 14 Although some clusters experienced increased PM 2.5 over time, the weighted trend remains 15 negative when higher-order/ lower numbered clusters are added together (see last column in each 16 part of the table). The PM 2.5 trend across all trajectories (there is no difference in trajectory age 17 because all are associated with the same PM 2.5 data) was -1.35 µg m -3 yr -1 which is significant at 18 the 99 percent confidence level. 19 A plot of PM 2.5 changes over time versus cluster-averaged PM 2.5 ( Figure 13) indicates that the 20 rate of change becomes increasingly negative as PM 2.5 exceeds 12 µg m -3 . The rate of change is 21 slightly positive when PM 2.5 <12 µg m -3 . The trajectory age (i.e., 12 or 24 hours upwind) makes 22 no difference. The downward trend in PM 2.5 mass is caused principally by reductions in aerosol 23 mass when conditions are such that aerosol levels are high (i.e., there is no evidence of a 24 downward trend when aerosol levels are low), and that may be caused by a combination of 25 meteorological factors and emissions changes. As shown in Table 5, cluster 1 for both 12-and 26 24-hr upwind had the highest average PM 2.5 and the largest decreases in PM 2.5 over 2007-2013. 27 Note that the 12-and 24-hr (upwind) trajectory clusters were not the same because of differences 28 in upwind trajectory age and other factors. Also, most PM 2.5 temporal changes within clusters 29 were not significant at the 95 percent confidence level, but when combined into a larger data set that included all trajectories the preponderance of declining PM 2.5 levels indicates a significant 1 negative trend. northern Alabama (-12 hr), and northern Mississippi (-24 hr). This analysis cannot assign 5 importance to changes in specific aerosol components due to a lack of data, but given the relative 6 contributions to PM 2.5 we know that a trend must be associated with some combination of 7 changes in sulfate and organic particle mass. 8

Conclusions 9
As stated in the beginning, the three objectives of this effort were to (1) identify trends in 10 percursor emissions and air quality, (2) identify the degree of association between emissions and 11 air quality trends and (3) identify whether trends in meteorology influenced air quality trends. 12 All three objectives were met-as summarized here-and the only potential association between 13 air quality and meteorological trends (objective 3) appears to be small. There have not been any significant changes in the modest diurnal patterns of major gaseous or 22 particulate sulfate over the past decade. It should be noted that there are large year-to-year 23 changes in most measured species concentrations which appear to be driven principally by 24 changes in meteorology-rainfall, clouds, solar radiation and temperature-and this is reflected 25 by some of the seasonal associations identified between air quality and meteorology. 26 The more pronounced reduction in SO 2 levels from 2007 to 2013-compared with changes in 27 sulfate levels over the same period (85 percent for SO 2 and 58 percent for sulfate based on 28 annual averages)-indicates that the spatial footprint of primary emissions of SO 2 in the region is likely smaller than that for the largely secondary sulfate species. It appears that reductions in 1 SO 2 emissions have produced primary SO 2 concentration reductions closer to the sources than 2 those for sulfate whose levels depend on atmospheric SO 2 -to-sulfate conversion processes that 3 respond to changes in precursor emissions over a larger region. 4 Statistical modeling revealed that a one standard deviation (1σ) decline in annual SO 2 emissions 5 in the emissions domain was associated with a 0.62σ decline in annual-average ambient SO 2 . By 6 comparison, a similar emission decline was associated with 0.97σ decrease in spring and summer 7 and a 0.76σ decrease in winter. The OC association with NO x emission was smaller with a 1σ 8 drop in emissions linked to a 0.44σ OC decrease in winter and 0.76σ OC decrease in summer. 9 Note that these statistical relations are all based on variances and do not reflect absolute 10 concentration responses to emissions. Changes in Look Rock air quality (1999-2013) appear to 11 be influenced little by meteorological trends (unless a slow but steady decrease in winds 12 somehow plays a role) but are instead primarily controlled by changes in precursor emissions. 13 Much of those changes have occurred nearby (as represented by the Look Rock emission domain 14 defined here) but some of those changes are also due to emissions changes farther upwind. 15 Author contributions 16 R. Tanner was the primary architect of this study and performed most of the air quality data 17 analysis. S. Bairai operated the Look Rock site and was primarily responsible for instrument 18 maintenance, data quality assurance and data archival. S. Mueller provided the meteorological, 19 trajectory modeling and statistical trends analyses. All three authors contributed to manuscript 20 preparation. 21

Acknowledgements 22
We are indebted to the emissions data provided by the Environmental Protection Agency and the 23 meteorological and air quality data collected and archived by the National Park Service. We      oxides (e.g., iron oxide) associated with soils and "Unknown" represents mass whose 16 composition is indeterminate (most likely some water along with mass associated with OC 17 that is not included in the 1.8 adjustment factor applied by IMPROVE to OC to estimate OM). 18 Numbers associated with each pie section denotes the annual mean mass (µg m -3 ) of that 19 chemical constituent. 20