This study presents source apportionment results for PM2.5 from applying positive matrix factorization (PMF) to a 32-month series of daily PM2.5 compositional data from Denver, CO, including concentrations of sulfate, nitrate, bulk elemental carbon (EC) and organic carbon (OC), and 51 organic molecular markers (OMMs). An optimum 8-factor solution was determined primarily based on the interpretability of the PMF results and rate of matching factors from bootstrapped PMF solutions with those from the base case solution. These eight factors were identified as inorganic ion, n-alkane, EC/sterane, light n-alkane/polycyclic aromatic hydrocarbon (PAH), medium alkane/alkanoic acid, PAH, winter/methoxyphenol and summer/odd n-alkane. The inorganic ion factor dominated the reconstructed PM2.5 mass (sulfate + nitrate + EC + OC) in cold periods (daily average temperature < 10 °C; 43.7% of reconstructed PM2.5 mass) whereas the summer/odd n-alkane factor dominated in hot periods (> 20 °C; 53.1%). The two factors had comparable relative contributions of 26.5% and 27.1% in warm periods with temperatures between 10 °C and 20 °C. Each of the seven factors resolved in a previous study (Dutton et al., 2010b) using a 1-year data set from the same location matches one factor from the current work based on comparing factor profiles. Six out of the seven matched pairs of factors are linked to similar source classes as suggested by the strong correlations between factor contributions (r = 0.89 − 0.98). Temperature-stratified source apportionment was conducted for three subsets of the data in the current study, corresponding to the cold, warm and hot periods mentioned above. The cold period (7-factor) solution exhibited a similar distribution of reconstructed PM2.5 mass as the full data set solution. The factor contributions of the warm period (7-factor) solution were well correlated with those from the full data set solution (r = 0.76 − 0.99). However, the reconstructed PM2.5 mass was distributed more to inorganic ion, n-alkane and medium alkane/alkanoic acid factors in the warm period solution than in the full data set solution. For the hot period (6-factor) solution, PM2.5 mass distribution was quite different from that of the full data set solution, as illustrated by regression slopes as low as 0.2 and as high as 4.8 of each matched pair of factors across the two solutions.

Abstract.A Positive Matrix Factorization receptor model for aerosol pollution source apportionment was fit to a synthetic dataset simulating one year of daily measurements of ambient PM 2.5 concentrations, comprised of 39 chemical species from nine pollutant sources. A novel method was developed to estimate model fit uncertainty and bias at the daily time scale, as related to factor contributions. A circular block bootstrap is used to create replicate datasets, with the same receptor model then fit to the data. Neural networks are trained to classify factors based upon chemical profiles, as opposed to correlating contribution time series, and this classification is used to align factor orderings across the model results associated with the replicate datasets. Factor contribution uncertainty is assessed from the distribution of results associated with each factor. Comparing modeled factors with input factors used to create the synthetic data assesses bias. The results indicate that variability in factor contribution estimates does not necessarily encompass model error: contribution estimates can have small associated variability across results yet also be very biased. These findings are likely dependent on characteristics of the data.

To evaluate the utility and consistency of different speciation data sets in source apportionment of PM(2.5), positive matrix factorization (PMF) coupled with a bootstrap technique for uncertainty assessment was applied to four different 1-year data sets composed of bulk species, bulk species and water-soluble elements (WSE), bulk species and organic molecular markers (OMM), and all species. The five factors resolved by using only the bulk species best reproduced the observed concentrations of PM(2.5) components. Combining WSE with bulk species as PMF inputs also produced five factors. Three of them were linked to soil, road dust, and processed dust, and together contributed 26.0% of reconstructed PM(2.5) mass. A 7-factor PMF solution was identified using speciated OMM and bulk species. The EC/sterane and summertime/selective aliphatic factors had the highest contributions to EC (39.0%) and OC (53.8%), respectively. The nine factors resolved by including all species as input data are consistent with those from the previous two solutions (WSE and bulk species, OMM and bulk species) in both factor profiles and contributions (r = 0.88-1.00). The comparisons across different solutions indicate that the selection of input data set may depend on the PM components or sources of interest for specific source-oriented health study.

To identify the sources of PM2.5 – bound carbonaceous species and examine the spatial variability of source contributions in the Denver metropolitan area, positive matrix factorization (PMF) was applied to one year of every sixth day ambient PM2.5 compositional data, including elemental carbon (EC), organic carbon (OC), and 32 organic molecular markers, from four sites (two residential and two near-traffic). Statistics (median, inner quantiles and 5th – 95th percentiles range) of factor contributions, expressed as reconstructed carbonaceous mass (EC + OC), were estimated from PMF solutions of replicate datasets generated by using a stationary block bootstrap technique. A seven-factor solution was resolved for a set of data pooled across the four sites, as it gave the most interpretable results and had the highest rate of neural network factor matching (76.9%). Identified factors were primarily associated with high plant wax, summertime emission, diesel vehicle emission, fossil fuel combustion, motor vehicle emission, lubricating oil combustion and wood burning. Pearson correlation coefficients (r) and coefficients of divergence (COD) were used to assess spatial variability of factor contributions. The summertime emission factor exhibited the highest spatial correlation (r = 0.74 – 0.88) and lowest CODs (0.32 – 0.38) among all resolved factors; while the three traffic dominated factors (diesel vehicle emission, motor vehicle emission and lubricating oil combustion) showed lower correlations (r = 0.47 – 0.55) and higher CODs (0.41 – 0.53) on average. Average total EC and OC mass were apportioned to each factor and showed a similar distribution across the four sites. Modeling uncertainties were defined as the 5th – 95th percentile range of the factor contributions derived from valid bootstrap PMF solutions, and were highly correlated with the median factor contribution in each factor (r = 0.77 – 0.98). Source apportionment was also performed on site specific datasets; the results exhibited similar factor profiles and temporal variation in factor contribution as those obtained for the pooled dataset, indicating that the four sites are primarily influenced by similar types of sources. On the other hand, differences were observed in absolute factor contributions between PMF solutions for the pooled versus site-specific datasets, likely due to the large uncertainties in EC and OC factor profiles derived from the site specific datasets with limited numbers of observations.