Department of Mechanical and Aeronautical Engineering, Department of Civil and Environmental Engineering, and Department of Land, Air and Water Resources, University of California, Davis, California, USA
 Approximately 236,000 single particle mass spectra were collected throughout the duration of the Pittsburgh Supersite experiment using the third-generation rapid single particle mass spectrometer (RSMS-3). The instrument was operated semicontinuously for 306 days, sampling particles with aerodynamic diameters in the range of 30–1100 nm and collecting both positive and negative ion spectra, particle size, and time of detection for each particle measured. The entire data set has been fully processed and analyzed. Spectra have been clustered into 20 distinct particle classes on the basis of the distribution of their positive ion mass peaks. Negative ion spectra were classified independently within each positive ion class. Frequency of occurrence versus particle size, month of the year, and wind direction has also been calculated for the full data set, as well as within each class. Results indicate a rich array of multicomponent ultrafine particles composed primarily of carbon and ammonium nitrate. Approximately 54% of all the particles measured fell into the carbonaceous ammonium nitrate (CAN) class. These particles were observed in all size bins and from most wind directions for the entirety of this study. Ubiquitous sources throughout the area, including vehicular emissions and secondary organic aerosol formation, are considered to be responsible for a larger fraction of these particles. In terms of particle number, metal containing aerosol dominated the remainder of the particle classes identified. These particles were rich in K+, Na+, Fe+, and Pb+ and to a lesser extent, Ga+ and Zn+. They tended to be smaller in size and were highly correlated with specific wind directions, facilitating the isolation of specific sources.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Numerous advances have been made in understanding the effects of anthropogenic particle emissions on the atmosphere including the mechanism and feedback of aerosol radiative forcing, the correlation between fine atmospheric particles and adverse health effects, visibility reduction due to particle loading and ecological degradation due to particle deposition. These advances have facilitated and necessitated the continual development of aerosol instruments capable of measuring particle concentrations, size distribution and chemical composition with increasing accuracy for smaller and smaller particle sizes.
 Conventional instruments for these measurements were designed with the idea of collecting particles onto a substrate, which can be analyzed later for mass and chemical composition using traditional techniques. Although these instruments are useful, they are limited by the amount of detail they can capture. Essential information, such as the original chemical composition of individual particles, the distribution of compositions among different particle sizes and time resolution of particle detection, is obscured or lost. Recent technological developments have allowed for improvements in the way that atmospheric aerosol is sampled. A new technique has been established that is capable of counting, sizing, and analyzing particles one by one in real time. The general design of an instrument implementing such a technique encompasses size-selective particle focusing, particle detection, laser ablation, mass spectrometry, and signal digitization [Wexler and Johnston, 2001]. Several groups around the world have built instruments based upon this method of single particle sampling [Thomson and Murphy, 1994; Prather et al., 1994; Hinz et al., 1994; Murphy and Thomson, 1995; Suess and Prather, 1999; Jayne et al., 2000]. One of the original prototypes [Johnston and Wexler, 1995; Carson et al., 1995] is referred to as a rapid single particle mass spectrometer (RSMS). Instruments like RSMS are clearly superior in their capacity to obtain the original composition of individual particles at the time of detection as well as counting and sizing particles of similar or different composition.
 One inherent drawback of these instruments is the use of light scattering to detect particles. Particles with diameters less than the wavelength of light cannot be easily detected. This puts a practical lower detection limit of approximately 200 nanometers on the instrument [Wexler and Johnston, 2001]. Unfortunately, some of the most fundamental physics and chemistry behind particle formation and evolution can only be captured in the analysis of ultrafine particles, typically defined as having aerodynamic diameters less than 100 nm. In addition, both toxicological and epidemiological studies show an increased risk of respiratory health effects associated with exposure to ultrafine particles, as compared to accumulation mode or coarse mode particulate matter (PM), making their characterization even more important [Peters et al., 1997; Pekkanen et al., 1997; Donaldson et al., 1998; Li et al., 2003]. The instrumental modifications necessary to overcome this obstacle have only recently been implemented though they have been known for several years. The essence of eliminating the need to detect a particle lies in the ability to fire the ablation laser with a constant energy output at a high frequency and to only collect spectra when a particle is present [Reents et al., 1995]. This technique was first deployed in the development of the second-generation rapid single particle mass spectrometer, RSMS-2 [Carson et al., 1997]. Further improvements to RSMS-2 were made by the inclusion of a size-selective focusing inlet, originally introduced by Mallina et al. . The design of the instrument and results from laboratory experiments have been described in detail elsewhere [Ge et al., 1998; Mallina et al., 2000; Johnston, 2000; Phares et al., 2002; Lake et al., 2003]. In addition, RSMS-2 has been successfully deployed in two major field studies, the 1999 Atlanta Supersite experiment [Rhoads et al., 2003] and the 2000 Houston Supersite experiment [Phares et al., 2003]. During the Atlanta study four different types of single particle instruments, including RSMS-2, were deployed simultaneously [Lee et al., 2002; Liu et al., 2003; Jimenez et al., 2003; Rhoads et al., 2003; Wenzel et al., 2003; Lee et al., 2003]. For an inter comparison of the measurements obtained from these techniques, see Middlebrook et al. .
 RSMS-3 is the latest version in the RSMS family. Improvements in RSMS-3 over earlier versions include: (1) the addition of a negative ion mass spectrometer to run in conjunction with the positive one such that both positive and negative ion spectra are obtained for each particle, (2) two A/D converters on each spectrometer to increase dynamic range, and (3) software improvements that enable automation and remote control of the instrument. RSMS-3 was specifically constructed for deployment in a yearlong air quality field study located adjacent to Carnegie Mellon University in Pittsburgh, Pennsylvania as a part of the U.S. Environmental Protection Agency (EPA) Supersites program. The purpose of RSMS-3 in this study was to provide semicontinuous, time and size resolved measurements of single particle composition for ambient air pollution in the Pittsburgh area. These measurements are extremely valuable not only in assessing Pittsburgh air quality but also for numerous other applications. Many issues of interest in this field, ranging from health affects to global climate, are believed to be strongly linked to the composition as well as the size of the particles involved. This paper, however, will focus strictly on statistical analysis of the data and the application of these statistics to initiating possible source-receptor relationships.
 As mentioned in the introduction, the details of RSMS have been described in previous work such that only the major points will be revisited here. RSMS-3 is depicted schematically in Figure 1. The measurement technique involves size-selective particle sampling, particle beam generation, laser desorption/ionization of individual particles and mass spectrometry. During this experiment, ambient air was drawn from outside the trailer at a height of 6.4 meters above ground through a 10 cm diameter stainless steel duct. Aerosol was sampled from the center of the duct through inches OD copper tubing and dried with a Nafion dryer (Perma Pure, Inc., Toms River, New Jersey) prior to entering the inlet. Using a computer controlled 10-position rotary valve (Valco Instruments Co., Houston, Texas), the particle flow is then directed through one of nine differently sized flow-limiting orifices contained in an orifice bank (O'Keefe Controls Co., Trumbull, Connecticut). The size of the orifice determines the pressure just upstream of the critical focusing orifice. This inlet pressure, in conjunction with the geometry of the critical orifice, determines the particle size focused and transmitted into the instrument. Aerodynamic diameter of the focused particles can be approximated by
where λ is the mean free path evaluated at the critical orifice conditions [Wexler and Johnston, 2001]. The maximum particle diameter that can be focused in this manner is given by the quantity within the inner brackets, where μ is gas viscosity, dn is the diameter of the critical orifice, ρp is particle density, Stkf is the stokes number that is focused (∼1.1) and Usonic is the speed of the gas through the critical orifice, which is sonic since the flow is choked. The particle beam then travels through several differentially pumped skimming stages in which the carrier gas is removed using both mechanical pumps (BOC Edwards, Wilmington, Massachusetts; Varian Inc., Lexington, Massachusetts) and split turbo pumps (Pfeiffer Vacuum, Nashua, New Hampshire). Particles then enter the source region where they are vaporized and ionized by a colinear, counter propagating and free-fired 193 nm excimer laser (GAM Laser, Orlando, Florida). Ions are initially accelerated by a dual gradient in the source region and then travel at constant velocity down their respective time of flight tubes until they impinge upon the microchannel plate detectors (Burle Electro-optics Inc., Sturbridge, Massachusetts). Current from each detector is recorded and digitized by two separate 500 MHz digitizer channels (Acqiris, Monroe, New York). Overlapping high and low sensitivity channels increases signal resolution and enhances dynamic range by optimizing the offset between signal saturation and limit of detection. Triggered by each laser pulse (50 Hz), the digitizer collects 5000 sample points with a sampling rate of 2 ns covering a mass-to-charge ratio range of ∼1–300 Da for each detector. As part of the data acquisition software, written in LabView™ (National Instruments Inc., Austin, Texas), signal intensity from each mass spectrum is checked against an experimentally determined threshold to ensure that only spectra from true particle hits are recorded. Peaks from either positive or negative ions can trigger the spectrum to be saved. Threshold values are different for different windows within the spectrum and certain sections were left out to avoid triggering on background peaks. Since the background peaks are associated with gas phase contamination originating from inside the instrument, their relative peak heights are independent of the particle size being sampled, and thus the same threshold values were used for all nine orifices. For each particle hit, positive and negative ion intensity, time of detection and particle size are recorded.
2.2. Sampling Protocol
 RSMS-3 was fully automated for the entirety of the Pittsburgh Supersite experiment, controlled solely by the data acquisition software, except during times of maintenance or special studies. The measurement protocol was based on sampling intervals starting every three hours for the first four months (September–December 2001) and every two hours for the remainder of the study (January–September 2002). Each sampling interval consisted of cycling through the nine flow-limiting orifices, corresponding to nine different particle sizes transmitted into the instrument ranging from about 30 nm to 1.1 micrometers. The instrument was operated at each orifice until either 10 min expired or 30 particles were sampled, whichever came first. On average, ∼1 hour was required to step through the entire orifice bank. This protocol was adopted to allow for instrument down time between sampling intervals to optimize the balance between instrument sustainability and robustness of the data set. Over the course of a day, the average number of single particle mass spectra acquired was ∼1100. Of these 1100, ∼15% had detectable signal from both positive and negative ions while the remainder had signal from positive ions only.
2.3. Data Processing and Analysis
 Remote control software, pcAnywhere™ (Symantec Inc., Cupertino, California), was used to monitor the performance of the instrument and to transfer data from the instrument's hard drive to University of California-Davis computers for storage, postprocessing and analysis. Each single particle mass spectrum collected has been processed from its original state as follows: (1) It has been time to mass calibrated according to the relation m/z = (at + b)1/2 where m/z is mass to charge ratio, a and b are experimentally determined constants and t is the time interval during which the ion current was detected. (2) A binned ion current for each m/z value was obtained by integrating ±0.5 Da about each integer m/z value. (3) The spectrum has been normalized across all m/z values according to a Euclidian norm. Calibration is the single most important step in this procedure. It is vital to the interpretation of the spectral peaks. Since the calibration constants are moderately dependent on the degree to which the particle and laser beams are coincident, they have a tendency to vary slightly from spectrum to spectrum. As a result, each spectrum has been inspected and calibrated individually to ensure the quality of the data processing and the integrity of the analysis.
 Owing to recent progress in digital capture, processing and storage technology, RSMS-3 presents a unique challenge to data analysis in its ability to amass enormous sets of data in relatively short time durations. For instance, over 200,000 dual-polarity single particle mass spectra were collected during this campaign alone. Confounding the issue is that each spectrum itself can have as many as 500 independent components, or dimensions, each with its own inherent uncertainty. Perhaps the single most challenging task in working with such data sets is creating an unbiased and self-consistent mechanism for reducing the amount of data without compromising resolution or robustness. An obvious starting point is to construct an algorithm that organizes individual data points (spectra) into unique clusters (particle classes) based upon some metric of similarity between them. The process should include a set of conditions that control the evolution of the classification while simultaneously allowing for convergence to a unique solution.
 Adaptive Resonance Theory version 2a (ART-2a) [Carpenter et al., 1991], originally introduced to the aerosol community by Hopke and Song , is an example of such a classification algorithm currently in use by a number of groups. Spectra obtained from laboratory generated aerosol of known composition were used to validate its application to single particle data [Phares et al., 2001]. It uses the vector dot product as its metric of similarity and is controlled by two parameters, one of which sets the condition for similarity and the other which determines the rate at which the algorithm learns. A new algorithm, based on ART-2a, has been developed specifically for the analysis of this data set encompassing several improvements. First, the process by which clusters are seeded has been reconstructed to allow for the identification of “natural seeds” in a data set, rather than picking random seeds. “Natural seeds” are defined as those spectra having the largest number of most similar partners. This makes the seeding process definitive, rather than random, and ensures that the outcome of the classification is no longer dependent on the choice of spectra used to seed it, since the same spectra will always be chosen. Secondly, the dynamics of the algorithm have been de-coupled from parameterization to allow for unbiased learning from iteration to iteration. Specifically, the learning rate parameter has been eliminated in favor of equal-weighted averaging; allowing the variance in the mass peaks of each cluster to be calculated. This ensures that the clustering is based on true statistics rather than an arbitrary parameter. Finally, a variance-weighted similarity metric, based on the geometric distance rather than the dot product, has been adopted. Both the dot product and the geometric distance are inherently a measure of the angle of separation between two vectors (spectra), but the geometric distance is proportional to the sine of the angle, rather than the cosine, and thus is more sensitive at small angles; similar vectors. Using either metric alone, however, assumes that each dimension is equally important when comparing spectra and therefore each dimension should be equally weighted. In attempts to relax this notion, a more realistic approach has been taken that weights each dimension by its own inherent variance. This allows clusters to adopt an ellipsoidal shape, rather than being strictly spherical. The algorithm has been thoroughly tested and validated on a well characterized subset of data collected during this experiment. Results indicate convergence from different sets of initial conditions, stability from iteration to iteration and reproducibility of results.
 In addition to data clustering, numerous other analysis techniques have been applied which sort, organize, correlate and count spectra. These have been used on the entire data set to determine the fraction of total particle hits by month, size and particle class. They were also applied independently to each particle class to establish the distribution of particle hits within the class by month, time of day, size and wind direction. Correlating single particle spectra with wind data is possible due to the time signature associated with each spectrum. It allows for the identification of the direction(s) from which each particle class is most frequently observed. This, in conjunction with knowledge of local industry, can be used to associate specific classes with nearby sources and has proven to be a very powerful technique for studying source attribution.
 During this field study, RSMS-3 was housed at the Schenley Park site located adjacent to Carnegie Mellon University in Pittsburgh, Pennsylvania. Pittsburgh can be characterized as a largely industrialized and urbanized city surrounded by both suburbs and rural areas. As such, it was an excellent location to capture the true depth and complexity of urban aerosol. The list of emission sources within Pittsburgh is exhausting, let alone the possibility of regional transport from outside the city. Figure 2 depicts a map of some of the major PM2.5 point sources within 24 km (15 miles) of the site. Notice that the site is surrounded by coal fired power plants. To the northeast are Allegheny Power and the Duquesne Light Co., to the south are Elrama and Mitchell and to the northwest, not drawn on the map, are Orion Power Midwest LLC (Phillips Station, 28.3 km, 301°) and Bruce Mansfield (1255 metric tons/year PM2.5, 45.5 km, 298°). There are seven additional major coal fired power plants within 80 km (50 miles) of the site. All of these power plants are considered large sources of NOx and SO2, as well as moderate sources of trace metals. Steel mills and blast furnaces comprise the next largest industrial point source of PM in the area. USX Corporation (Edgar Thompson Works), USX Corporation (Irvin Plant), USS Clairton Works, Universal and Stainless Alloy Products and Shenango Neville Island Coke Works all use steel mills and blast furnaces. Other sources on the map include a landfill owned by Chambers Development Co., two cement processing facilities (possibly large sources of coarse mode calcium) and two glass plants. Not included on the map, but still of interest, are a hospital incinerator located adjacent to Shenango Coke Works (General Suburban Hospital, 11.9 km, 310°), Allegheny Ludlum Steel (410 metric tons/year PM2.5, 26.5 km, 42°) and Zinc Corporation of America (363 metric tons/year PM2.5, 41.8 km, 307°), a company dealing primarily in zinc and other nonferrous metals. Besides industry, a significant fraction of PM measured at the site can be attributed to numerous ubiquitous sources such as commuting, cooking and wood burning.
 RSMS-3 began taking measurements on 20 September 2001 and was successfully sampling for 306 of the possible 372 operation days. During this period, 236,286 single particle mass spectra were acquired, 185,244 (78.4%) of which had positive ion signal only and 51,042 (21.6%) that had both positive and negative ion signal. There were very few instances of spectra with only negative ion signal. Figure 3a shows the fraction of total particle hits obtained from each of the nine orifices for particles with positive ion spectra only (solid white) and those with positive and negative ion spectra (diagonal stripes). The sum of the two is the total fraction. The nominal aerodynamic diameter of particles focused by each orifice is labeled in nanometers on the x axis. Please note that this is not a true size distribution, particles are simply being counted at each orifice. There are many confounding factors involved with extracting true size distributions from this data. The difficulty lies in the fact that the hit rate efficiency of the instrument is a function of both particle size and composition, as has been discussed in detail by Kane and Johnston . It is also important to note that the existence of a complimentary negative ion spectrum is most commonly an indicator of atmospheric aging. This is simply due to the fact that the major negative ions detected are those of secondary particle components, specifically nitrates and sulfates. As a result, negative ion signal is seen predominantly in the spectra of particles from the larger size bins because there is simply more analyte in larger particles, making detection more likely. Note that small particles may have also undergone significant atmospheric transformations but this cannot be detected in the negative ion spectrum due to the combination of insufficient analyte and lower instrument sensitivities for these components in general [Kane et al., 2002]. The fraction of total particle hits by month of the year is shown in Figure 3b, normalized by the number of days within each month that the instrument was actually sampling. Once again, both the fraction of particles with positive ion spectra only (solid white) and the fraction of those with positive and negative ion spectra (diagonal stripes) have been plotted. Results indicate that the largest fraction of particles sampled, as well as the largest fraction containing negative ion spectra, was observed in the winter during the month of January. However, there are elevated levels in the summer months of June and July as well. Figure 3c displays frequency of occurrence versus wind direction for all wind observations greater than two meters per second, depicting major wind directions in the Pittsburgh area. Notice that the wind comes predominantly from the west to northwest and is almost never observed originating from anywhere inside the first quadrant. This is an important issue to consider when interpreting wind signatures associated with specific particle classes. Particles are observed in the first quadrant, but little statistical significance is given to these observations in the context of identifying source-receptor relationships.
 Data classification was performed in a series of steps designed to enhance the quality of the classification and make certain its validity. First, the entire data set was split into three groups based upon periods of similar operating conditions and instrument performance. The clustering algorithm was then applied independently to each of the three groups using positive ion spectra only. In total, over 500 clusters were isolated. Each cluster was individually inspected, peak assignments were made for all known peaks and the normalized peak heights recorded. This information was used in conjunction with a sorting algorithm to organize the clusters into approximately 100 subclasses. Sorting was performed under the condition that all clusters within each subclass contain the same six largest peaks in identical peak height progressions. Subclasses were then categorized into 20 distinct particle classes by grouping together those containing the same four dominant peaks, regardless of the distribution of peak heights. Results are summarized in Table 1. The first column contains a general description of each particle class, the second displays the percent of the total number of spectra which belong to that class, the third is the percent of spectra within the class observed to have negative ion signal, the fourth shows the percent of spectra within the class having detectable amounts of NO+, the fifth lists the subclasses within each class and the sixth column lists the subclass distribution within the class. Figure 3d depicts the monthly distribution of the ten largest particle classes listed in Table 1, along with the iron/cerium class. The number of spectra, belonging to a given class, within each month has been normalized by the total number of spectra collected during that same month. Notice that all of these classes, with the exception of unidentified organics and iron/cerium, were observed continuously throughout the duration of this study, and that a majority of the particles observed were carbonaceous in nature, both overall and within each month. A more detailed discussion of the temporal distribution of each class will be given below. Negative ion spectra were classified independently within each positive ion class during the final step of classification. Figure 4 depicts the distribution of all negative ion spectra among the identified negative ion classes. As stated above, sulfate and nitrate, secondary components, dominate the entire distribution. Note that every single negative ion class has some form of sulfate in it. The four largest classes contain nitrate/sulfate, sulfate only, chlorine/nitrate/sulfate and elemental carbon/organic carbon/sulfate.
 Figures 5–16 summarize the results of the 10 major particle classes (>1% of the total number of particles, see Table 1) and two minor classes (Ce/Fe and Cr/Mo/W). Each figure contains five graphs, a–e, that illustrate different ensemble characteristics of the particles found in that class. The first graph (a) is a spectral representation of the class showing normalized peak area versus mass to charge ratio (m/z), the average of all spectra belonging to the dominant subclass. The bars at the top of each mass peak represent the standard deviation in the average. The second plot (b) depicts both the fraction of particle hits having positive ion signal only (solid white) and the fraction having positive and negative ion signal (diagonal stripes) versus aerodynamic diameter. Third is a wind rose (c) illustrating frequency of observation of the class versus wind direction for all wind speeds greater than 2 m/s. Real time data acquisition allows each single particle spectrum to be associated with a specific wind speed and direction. As a result, particles can be counted for a given wind direction interval. For the analysis described in this paper, 5° intervals were used. Values obtained for each interval have been normalized by the total number of wind observations, depicted in Figure 3c, for that same interval. This has been done in attempts to normalize out the major wind directions, isolating regions where particles were observed more frequently than the wind, relative to the other directions. However, as mentioned above, the wind is almost never observed originating from anywhere inside the first quadrant and thus particle observations within this quadrant will tend to be exaggerated, relative to the other three quadrants, during the normalization process. In this sense, peaks occurring in the first quadrant are considered statistically insignificant and will be disregarded. Next is a graph (d) showing the fraction of particle hits by month for spectra containing positive ion signal only and those containing both positive and negative ion signal. The fractions reported have been obtained by normalizing the number of particle hits, belonging to the class, that were observed during each month by the total number of particles observed during that same month. The final plot (e) is a pie chart displaying the distribution of particles, for which negative ions were observed, among the identified negative ion classes.
3.1. Carbonaceous Ammonium Nitrate (CAN)
 CAN is by far the largest and perhaps most complex class of particles identified in Pittsburgh. For a large majority of these particles, it is not possible to determine whether the carbon is elemental or organic solely from the peaks in their mass spectra. The laser energy required to vaporize and ionize particles is too large to retain any significant information about the original structure of most organic molecules. As a result, unless stated otherwise, the carbon in these particles can be elemental, organic or a combination of the two. The spectral representation shown in Figure 5 is that of an ammonium nitrate dominant subclass containing smaller carbon peaks (44% of class). Closely related is the carbon dominant subclass containing smaller ammonium nitrate peaks (28%). It is important to note that making a distinction between these two subclasses is a bit arbitrary. In reality, the ratio of signal intensities, NO+/C1+, is continuously distributed over a range of ∼0.1–1.5 [Zhao et al., 2005]. Therefore it is more appropriate to think of only one subclass that displays a wide range of relative peak heights. The other three subclasses, comprising the remaining 28% of the class, have the same four dominant peaks (C1+, C2+, C3+ and NO+) as the previous two, but contain additional peaks, facilitating their isolation as different particle types. Two of them contain carbon that is distinguishably organic and appears to be highly oxygenated (oxy-OCAN), as evidenced by the presence of signature peaks for: CO+, C2H3O+, C4H7O+, etc. The remaining one, as well as one of the oxy-OCAN subclasses, contains distinct positive ion sulfate peaks (SO+, HSO2+, HSO3+ and NH4·HSO3+) in addition to carbon and ammonium nitrate (CANS, oxy-OCANS), possibly indicating a greater amount of sulfate in these particles. The purpose to making a distinction between these subclasses is to accentuate the fact that this class does not represent a single particle type, but rather a distribution of many carbonaceous aerosols containing various combinations of numerous organic species, elemental carbon, nitrate and sulfate.
 Looking at the fraction of total particle hits by size indicates that these particles have undergone a considerable amount of atmospheric aging and thus tend to be skewed toward the larger size bins. This is also evident by the overwhelming amount of ammonium nitrate and sulfate found in both the positive and negative ion spectra. The organic carbon itself is most likely from combustion or secondary organic aerosol. From the wind rose there appear to be numerous sources, as was expected given the complexity of the class and the nature of the particle. The monthly distribution shows the particle fraction slowly climbing during the late fall, peaking during the winter months of January and February, and then falling off as spring and summer approach. This trend is most likely associated with a reduction in the vapor pressure of ammonium nitrate accompanying lower temperatures. The reduction in vapor pressure will force the ammonium nitrate into the particle phase and thus increase the frequency with which this class of particles is observed. However, notice that there are elevated levels in the month of June, falling outside the trend. Obviously nitrate and sulfate dominate the negative ion speciation within this class. A bit more surprising is the significant amount of chlorine found in these particles, whose source is unknown.
 In urban and residential settings, biomass particles are typically associated with burning wood for energy or recreation. Rural areas can also be a large source of biomass particles, typically through heating, recreation and waste disposal. The signature spectrum, shown in Figure 6, includes a potassium ion in conjunction with multiple carbon peaks (C1+–C4+). Similar to CAN, the K+/C1+ ratio exhibits a rather continuous distribution. Also similar to CAN, it is not possible to discern whether the carbon is elemental or organic from these mass peaks. A majority of the particles measured had aerodynamic diameters in the range of 100–200 nm, consistent with a combustion source. The wind rose suggests numerous sources and the monthly distribution indicates that these particles were observed largely during the late fall and early winter, although there are slightly elevated levels in July as well. The negative ion spectra contain predominantly nitrate and sulfate, again with a significant amount of chlorine. Chlorine in this class is most likely associated with wood burning, as chloride ions have been observed in several studies on wood smoke emissions [Kleeman et al., 1999; Schauer et al., 2001]. Depending on the type of wood that is burned, chlorine comprises as much as 0.13–1.7%, by mass, of the fine particulate emissions [Schauer et al., 2001].
3.3. Elemental Carbon/Organic Carbon (EC/OC)
 Typical spectra in this class contain only the first three carbon cluster ions (C1+–C3+), see Figure 7. Once again, it is not immediately clear from this series of peaks whether the carbon is elemental, organic, or a combination. However, the sizes of the particles suggest that they are combustion particles since the majority of those detected were in the 75–125 nm range. The wind rose suggests that this class of particles has been routinely detected from all directions, which is consistent with multiple sources; such as vehicles and cooking. One of the major differences between this class and CAN is the absence of detectable amounts of ammonium nitrate in the positive ion spectra. Notice that the observed monthly distribution for this class is largely anticorrelated with that of CAN, further evidence for the seasonal trend in ammonium nitrate discussed earlier. Carbon particles without ammonium nitrate are observed less frequently in the winter months, when the temperatures are colder, and more frequently in warmer months, with the exception of August and September. Nitrate is still observed in the negative ion spectra, along with sulfate and EC/OC, but note that, similar to CAN, only ∼20% of the particles in the class have detectable negative ion signal.
3.4. Sodium/Potassium (Na/K)
 Easily distinguished by the combination of Na+, K+, Na2+ and NaK+ ions, as depicted in Figure 8, this is the largest class of non-carbon-based particles. Sodium and potassium are very common earth alkali metals emitted by a variety of sources. Potassium was discussed briefly in the context of biomass burning, but has also been identified in both sea salt and road salt particles, similar to sodium. However, given that the majority of the particles detected were in the size range of 75–200 nm, this class is inconsistent with what one would expect to see for sea salt or road salt particles, suggesting an alternate source. This is supported by the fact that the nearest ocean is almost 500 km away and the majority of Na/K particles were observed in the summer months, when the roads are not being salted. On the other hand, besides the nitrate/sulfate class, the negative ion speciation is dominated by chlorine (33%), implying that these particles are, in fact, composed primarily of NaCl and KCl salts. From the wind rose, it is clear that these particles originate largely from somewhere inside the second quadrant. There are numerous sources in this general area so it is hard to attribute the class to any one in particular.
3.5. Unidentified Organics
 This is a very interesting class of particles. As the name of the class states, the chemical composition of these particles is yet to be completely identified. For this reason, no peak assignments have been made in Figure 9; only the m/z values of the dominant and reoccurring peaks have been listed. Notice the structure that begins to develop at m/z 81. Successive peaks further downfield are separated by 14 Da (81, 95, 109, 123, 135, 149, etc.). This kind of structure suggests large organic carbon chains perhaps with high degrees of functionality. From the plot of particle fraction versus size, these particles tend to be larger in size, peaking around 200–300 nm, with significant fractions of secondary components in the largest size bins. Looking at the wind rose, there is a very well-defined wind signature for this class of particles. Notice that the direction coincides exactly with Chambers Development Co. at 108°. This is a major landfill and a moderate source of volatile organic compounds (16 metric tons/year). In fact, these particles were seen almost exclusively in the summer months when the VOC emissions from the landfill were probably at a maximum. Figure 9 includes an additional plot (f) showing the fraction of total particle hits by hour of the day. Notice that this particle type is observed almost entirely during the nighttime hours, after the sun has gone down and the atmosphere begins to cool. Combining this evidence we posit that the sun drives the volatilization of VOCs from the landfill during the day, but the atmosphere cools at night and these vapors condense on preexisting particles. Another possibility, however, must also be considered. Photochemical oxidation may suppress the detection of these particles during the day, but when the sun goes down and the chemistry shuts off, the particles reappear. Both mechanisms most likely play a role in the observed diurnal variation.
3.6. Elemental Carbon (EC)
 This is the classic elemental carbon particle, also called black carbon. It is easily identified by the long series of carbon cluster ions, typically from C1+–C11+, seen in Figure 10. Approximately 63% of these particles also contain detectable amounts of ammonium nitrate, as evidenced by the presence of NO+ peaks. Carbon has already been discussed extensively, so this class will only be touched on briefly. Similar to the EC/OC class, the size of these particles tend to be skewed toward the smaller size bins (75–125 nm), consistent with a combustion source. The wind rose shows multiple sources for this particle, also consistent with combustion. They were observed continuously throughout the campaign and the negative ion classification reveals a significantly larger fraction of EC/OC, in addition to nitrate and sulfate.
 This is the single largest class of metal particles identified in the data set. Sodium and potassium have already been discussed, but only in the context of biomass burning and salts. The spectral representation, shown in Figure 11, contains the Si+, K+, SiO+, Fe+ and Ga+ peaks. Some of the subclasses within this class contain trace amounts of additional metal, such as aluminum, sodium, lithium and lead, but the dominant ions remain the four listed above. Notice that the majority of the particles fall in the size range of 100–300 nm, implicating combustion or a high temperature furnace as the source. Also notice that there is significant negative ion fraction in each size bin, indicating extensive atmospheric processing and suggesting a more distant source. A very pronounced wind signature for this class is centered at ∼305°. Numerous possible sources exist in this direction, but not all are equally probable. Shenango Coke plant (297°, 13 km) and the hospital incinerator (303°, 11.9 km) seem too close to the site to be consistent with the amount of aging, while the Zinc Corporation of America (307°, 42.2 miles) deals only in nonferrous metals. Gallium, the least commonly seen of the metals in this class, is actually a trace metal found in coal. As there are very few sources of gallium in the atmosphere, it is most likely that these particles originated from coal combustion. It is also common to see other crustal metals like aluminum, silicon, potassium and iron in coal. Supporting this hypothesis, is the existence of two large coal fired power plants precisely in the direction of interest: Orion Power (Phillips Station, 301°, 28.3 km) and Bruce Mansfield (298°, 45.5 km, 1255 metric tons/year PM2.5).
 Besides being composed of different metals, this class of particles closely resembles the previous one. A majority of the particles are in the size range of 100–300 nm, all size bins contain some fraction of negative ion spectra and there is a well resolved wind signature at ∼300° (Figure 12). Given this information, and the composition of the class, the most likely source of this particle is Zinc Corporation of America. It is interesting to note, however, that almost half of the negative ions contain chlorine, suggesting that a significant amount of the metal in these particles is bound as a salt.
 Except for the addition of lithium, these particles are almost identical to those of the Na/K class discussed above, see Figure 13. They were observed consistently throughout the study and a majority of them are in the size range of 75–200 nm. Once again, this is more indicative of a combustion process or high temperature furnace rather than a sea salt particle or crustal material. However, the wind signatures are fewer and more resolved. Of particular interest is the well developed peak at ∼120°; the peak near 80° is not statistically significant. Notice that this coincides exactly with the USX Corporation (Edgar Thompson Works, 128°, 8.5 km, 280 metric tons/year PM2.5), a steel manufacturer housing blast furnaces and steel mills. Almost half of the negative ion spectra contain chlorine, suggesting predominantly NaCl and KCl particles. Sodium and potassium chloride are both used as hot molten alkali salts in salt pots for heat treating steel, suggesting that these particles originate from within the framework of steel processing. Since 48% of the total PM2.5 (6368 metric tons/year) emitted within 80 kilometers of the site is directly a result of the steel industry (8 separate companies), this could help explain the overwhelming amount of ultrafine sodium and potassium observed in Pittsburgh aerosol.
 Iron is one of the most commonly observed transition metals in urban particles and such is also the case in Pittsburgh. Consisting primarily of the iron isotope peaks (54, 56 and 57), the spectral representation of this class is shown in Figure 14. Notice from the size plot that the particle fraction distribution tends to be skewed toward the larger size bins, peaking around 200–300 nm, possibly the tail end of a distribution of mechanical abrasion particles, but more likely combustion or high temperature furnace related. Looking at the wind rose, there is a predominant wind signature centered at ∼125°, again pointing almost directly at Edgar Thompson Works (USX Corporation). This makes sense given that, besides steel, this company also manufactures iron products. Blast furnaces, used in the casting houses of this facility, are the most probable source of these particles. From the negative ion speciation, it is apparent that, to some degree, all seven negative ion classes are present. However, notice the increased fractions of both fluorine and chlorine in these particles. One possible explanation is the existence of hydrofluoric and hydrochloric acid on the interior wall of the emissions stack. If this is the case, then as iron is emitted it would sequester these acids, eventually transferring the fluorine and chlorine into the particle phase as salts.
 As part of one of the most unanticipated classes observed during this campaign, these particles contain the rare earth metal cerium. From Figure 15, it can be seen that a majority of them occupy the largest size bins, 300–1100 nm, unlike any other class. Also unique to this class is the fact that the particles were only observed for a relatively short period of time. Notice from the monthly distribution that they begin to appear in December and then practically disappear by the end of January. As a result, there were not enough data points with wind speeds greater than 2 m/s to obtain a good wind signature, making source attribution difficult. However, the two most common uses of cerium are:(1) as part of catalytic crackers used in oil refineries to break down large hydrocarbons in crude oil to smaller, higher grade, hydrocarbons and (2) as an additive to many oxidation and heat resistant alloys. Rare earth metals like cerium, lanthanum and yttrium are added to alloys (microalloying) specifically for oxidation and scaling resistance. Cerium is also found naturally in the Earth's crust, as a trace component, and has been observed in certain types of mineral dust [Utsunomiya et al., 2004].
 This is the last class of particles that will be addressed in this paper. A spectral representation, showing the Cr+, Mo+, MoO+ and W+ peaks, can be seen in Figure 16. Similar to many of the other metal-based classes, the majority of the particle fraction resides in the smaller size bins, peaking at ∼75 nm. In addition, there is only a very small fraction of negative ions across all size bins. Together, this indicates that the particles originated from a high temperature furnace located relatively close to the site. Incorporating data from the wind rose, it is clear that they are most frequently observed from the southwest, at ∼235°. This coincides almost exactly with Universal Stainless and Alloy Products (237°, 16.9 km, 15 tons/year PM2.5). Not surprisingly, this company is primarily involved with melting and refining alloys, including stainless steel, using equipment like electric arc furnaces. Note that chromium and molybdenum are two major constituents (along with nickel, iron and silicon) of almost all alloys. Chromium is added for oxidative resistance and strength, while molybdenum is added for strength and weldability. Tungsten, on the other hand, is a major constituent of superalloys, added for additional strength. It is also interesting to note that the other dominant wind signature is aligned with the USX Corporation (Edgar Thompson Works), also stainless steel manufacturers. Chromium is of special interest in this class since, depending on its oxidation state (Cr[III] versus Cr[VI]), it can be either a nutrient or a carcinogen. Unfortunately, RSMS-3 is not able distinguish between the two species [Neubauer et al., 1995].
 RSMS-3, operating semicontinuously for 306 days and sampling ambient aerosol within the size range of ∼30–1100 nm, collected approximately 236,000 single particle mass spectra as part of the EPA funded Supersites experiment in Pittsburgh, Pennsylvania. Although negative ion mass spectra were collected for every particle measured, very few (∼22%) had detectable amounts of negative ion signal. This can be attributed to the fact that the sensitivity of the instrument to negative ions increases with increasing particle size. Since atmospheric number concentrations are typically much smaller for larger particles, the hit rate of the instrument is significantly lower at this size range. Consequently, fewer particles are sampled and the fraction of spectra containing negative ion signal is small. For this reason, the negative ion mass spectrometer is able to characterize larger aerosol, but becomes less informative as the size of the particles decrease, especially in the ultrafine regime.
 All particles measured during this campaign have been clustered into particle classes based on the spectral distribution of their positive ion mass peaks. Twenty classes, and over 100 subclasses, were identified during this analysis. Negative ion spectra were then classified independently within each positive ion class. Characteristic size distributions, monthly distributions and wind signatures were also obtained within each class. These data were used not only to assess the dominant sizes and composition of urban aerosol in Pittsburgh, but also to construct possible source-receptor relationships between local industry and classes of observed particles.
 Particulate pollution in Pittsburgh was found to be predominantly carbonaceous in nature with ∼79% of the particles measured containing some form of carbon. These particles were observed in all size bins and from almost every direction for the duration of this study. Numerous ubiquitous sources scattered throughout the area, such as vehicular traffic, biomass burning and secondary organic aerosol formation, are most likely responsible for a large fraction of these carbon particles. In addition, there was a significant amount of ammonium nitrate observed in these particles.
 Besides carbon and secondary components, metals were recognized as the next largest constituent of Pittsburgh aerosol. Although a rich array of multicomponent metal particles was identified, the most commonly observed ions were K+, Na+, Fe+, Pb+, and to a lesser extent, Ga+ and Zn+. These particles were typically smaller in size, ranging from about 75–300 nm, and tended to be associated with specific wind directions. The analysis of the wind roses for individual classes has facilitated the isolation of specific local industries to which the observed metal-based classes may be attributed. Results from this analysis alone indicate that high temperature furnaces are the single largest source of ultrafine metal particles in the Pittsburgh area.
 Although the research described in this paper has been funded by the U. S. Environmental Protection Agency through grants to the Pittsburgh and Baltimore Supersites, it has not been subjected to the agency's required peer and policy review and therefore does not necessarily reflect the views of the agency. No official endorsement should be inferred. The authors acknowledge the assistance of Allen Robinson in compiling emissions inventory for industry-based air pollution sources in the Pittsburgh area, from which emission rates and source locations reported in this paper have been obtained. The authors also acknowledge Jon Ondov and his group for supplying the map of Pittsburgh used in the construction of Figure 2.