postworkshop summaries

posters

os goals

os topics

poster abstracts

os agenda

guest contributor abstracts

contact us

      

Abstract - Topic #3

Organic Speciation Related To The Source – Receptor Modeling

Topic Leader: Eric Fujita
Contributors: Tad Kleindienst and Tim Larson

Receptor models have been widely used to apportion source contributions to ambient PM and visibility impairment. Adoption of the ambient standard for PM2.5 has increased attention on apportionment of the carbonaceous fraction of PM2.5 to diesel exhaust and other combustion sources. Particulate organic matter is a complex mixture of directly emitted particles from different combustion sources and secondary organic aerosols derived from emissions of volatile and semi-volatile organic compounds from both anthropogenic and natural sources. While adverse health effects have been associated with crude measures of exposures to combustion sources such as elemental carbon, no single compound or element has been found that is unique to diesel exhaust and other mobile sources. This has lead to potential exposure misclassification and is a constraint to understanding exposure-response relations in community and occupational exposure studies. The purpose of this session is to review applications of organic speciation in the attribution of ambient PM to sources.

An overview by topic leader Dr. Eric Fujita (Desert Research Institute, Reno, NV) will provide a framework for discussion by raising two general questions:

1) What organic compounds (or compound classes) are useful for source apportionment?

· Do we need more compounds? Do we need less? What are the needs of the modeling community?

2) What are primary versus secondary organics?

· How can additional measurements help apportion secondary materials?
· What are implications of using multivariate receptor models?

The second presentation by Dr. Tad Kleindienst (U.S. Environmental Protection Agency, Research Triangle Park, NC) will focus on laboratory and field studies to identify and quantify multifunctional organic compounds of low volatility that are expected to result from the atmospheric oxidation of biogenic and aromatic hydrocarbons. Recent studies have shown that these polar organic compounds are relatively important compounds in PM2.5 in ambient atmospheres. Implications for possible indicator compounds for the secondary organic aerosol (SOA) component in PM2.5 will be considered.

The third presentation by Dr. Tim Larson (University of Washington, Seattle, WA) will focus on the application of multivariate receptor models to apportion combustion-related sources of carbonaceous component of PM. The treatment in these models of the SOA component is described. These multivariate methods are also compared and contrasted to the chemical mass balance receptor model.

The presentations will be followed by open discussion aimed at clarifying the needs of receptor modeling community for organic speciation data that may provide more unambiguous apportionment of PM2.5.

Multivariate Receptor Models
Timothy Larson
University of Washington
March, 2004

Background

Receptor modeling has traditionally been used to estimate the contribution of various sources to measured airborne particulate matter concentrations [1-3] .  In the majority of applications, the receptor of interest is a single, fixed site sampler that captures the temporal variability of various chemical species.  In addition, however, one can use multiple sites to capture both spatial and temporal variability in a number of chemical species [4] , or simply the spatial variability of a single analyte [4-6] .  Here we limit our discussion to the single site receptor with temporally varying chemical composition.  The general mass balance receptor model [2] can be stated in terms of the contribution from p independent sources to all measured chemical species in a given sample as follows:

                                                                     (1)

where for airborne particles xij is the jth specie concentration (mg/m3) measured in the ith sample, gik is the particulate mass concentration (mg/m3) from the kth source contributing to the ith sample, fkj is the jth specie mass fraction in particles emitted from the kth source, and eij is the model residual.  Traditionally, the U.S. EPA has recommended using the effective variance weighted chemical mass balance (CMB) receptor model [7, 8] , in conjunction with emissions inventories, for making these source contribution estimates.  This approach requires knowledge not only of the number of sources contributing to the observed airborne concentration of particle mass and chemical species, but also the composition of the particles emitted from each source.  More recent applications of this method have explored the use of unique particulate organic tracers [9, 10] as well as combined particulate and gaseous tracers [11, 12] .  In its most basic form, this model assumes that the composition of the particles does not change from source to receptor.  Therefore its ability to resolve secondary particulate matter is limited.

The same model (equation 1) can be solved for G without prior knowledge of F using several different factor analytic algorithms.  In principle, there are an infinite number of possible solutions of equation (1), that is, the model is non-identifiable [13] .  To lessen this ambiguity, these multivariate algorithms impose positive constraints on Fand G.  The algorithms used in practice include the UNMIX algorithm [14, 15] , the positive matrix factorization (PMF) algorithm [16] , and the multi-linear engine (ME2) algorithm [17] .  The latter two algorithms provide a solution that minimizes an object function based upon the value of each observation and its corresponding uncertainty [16] .  In practice [18-26] ., the results of the latter two models are scaled to the measured mass concentration by a constant, sj as follows:

                                                  (2)

where sk is determined post hoc by regressing measured total PM2.5 mass concentration against gik.

Recent Developments

Recently, Wahlin [27] has developed a multivariate approach (COPREM) based on alternating least squares that allows specification of some elements of F while leaving others unspecified.  This approach has advantages over traditional CMB; the user may know the mass fraction of a given specie in one or several source profiles (e.g. unique organic tracers), but does not always know the mass fraction of a given specie in all source profiles used in the model.  In this way, COPREM imposes logical constraints on the solution of equation 1.  In the case where F is fully constrained, the model is similar to traditional CMB; in the fully unconstrained case, it is, for example, similar to PMF.  It should be noted that the ME2 program also allows for specification of upper and lower limits (‘hard’ constraints) on selected elements in selected profiles [17] .  Clearly, the future of these multivariate methods lies in their ability to provide meaningful and flexible constraints on F (and G).

To this end, we have been developing methods that include constraints on equation 1 in addition to the standard requirement that G and F be positive.  To add constraints in a flexible manner, we have been exploring the use of the multi-linear engine program [17] .  This program allows the user to apply hard constraints on selected elements of F and/or G.  In order to apply these constraints in a straightforward manner, we have not only used the measured chemical species in the mass balance equation, but also have included the unmeasured mass as an additional ‘specie’.  In this way, one can directly specify constraints on F in terms of mass fractions, providing a flexible approach similar to COPREM.  We are also developing methods that simultaneously apportion both particle mass and particle volume as measured by conventional sizing methods.   This approach appears promising and provides additional model constraints on G.

Although these constrained receptor models are potentially more useful than prior receptor models, they still do not fully exploit the constraints provided by meteorology.  In the simplest case, if the monitoring site is known to be upwind of the source in question on a given day, then that source cannot contribute to the measured mass on that day.  Paatero and Hopke and Kim et al [28, 29] used the ME2 program to utilize information on surface wind speed and direction at the monitoring site, as well as seasonal and weekday/weekend effects to help resolve sources and their impacts.  In general, one would also like to include the effects of atmospheric mixing, atmospheric chemical reactions that form particulate mass, and atmospheric processes that remove particulate mass.  A combined, or ‘hybrid’ model [30] would also serve to strengthen both component models: adding meteorological information provides additional constraints on the multivariate receptor model, whereas deducing details of the particle composition by source category can be used to improve the underlying assumptions of a meteorologically based model.  This work would necessarily involve the use of chemical transport models that apportion concentrations predicted at a given receptor to an upwind set of source locations and source types. [31-33] .

Equally as important as the predicted source contributions are the associated uncertainties in those predictions.  This uncertainty is composed of sampling uncertainty and, in the case of the non-identifiable factor analytic models, the uncertainty associated with this non-identifiability.  Assuming that the number of sources and their source profiles are correctly specified a priori (usually a big assumption), the traditional CMB model correctly deals with model uncertainty.  In contrast, the unconstrained PMF model estimates uncertainties for F assuming G is fixed and known, and for G assuming F is fixed and known.  One promising approach to better specify these uncertainties was initially proposed by Park et al [34-38]   In principle, their approach captures not only uncertainties associated with source profile species (expressed as a distribution rather than a fixed value), but also uncertainties due to model specification bias (e.g., assumed number of sources).  More work needs to be done in this area to better capture the overall uncertainties associated with these multivariate methods.

Recommendations for Future Work

  1. Develop methods that better capture the uncertainties inherent in these multivariate predictions.
  2. For the same measurement set(s), compare the predictions from various algorithms including not only those that are unconstrained (e.g. PMF, UNMIX), but also those with prior constraints on the source profiles (e.g., CMB, COPREM, ME2 and MCMC).
  3. Develop ‘hybrid’ models that incorporate multivariate receptor models with chemical transport models; compare the ‘hybrid’ model predictions with those from each individual model.

References

  1. Seigneur, C., et al., Review of Air Quality Models for Particulate Matter. 1997, American Petroleum Institute Document No. CP015-97-1b.
  2. Hopke, P.K., Receptor Modeling for Air Quality Management. Environ. Sci. Technol., 1997. 8: p. 95-117.
  3. Henry, R.C., History and fundamentals of multivariate air duality receptor models. Chemometrics and Intelligent Laboratory Systems, 1997. 37(1): p. 37-42.
  4. Henry, R.C., Y.J. Wang, and K.A. Gebhart, The Relationship between Empirical Orthogonal Functions and Sources of Air-Pollution. Atmospheric Environment Part a-General Topics, 1991. 25(2): p. 503-509.
  5. Poirot, R.L. and P. Wishinski, Visibility, Sulfate and Air-Mass HIstory Associated with the Summertime Aerosol in Northern Vermont. Atmos. Environ., 1986. 20: p. 1457-1469.
  6. Paatero, P., et al., Advanced factor analysis of spatial distributions of PM2.5 in the eastern United States. Environmental Science & Technology, 2003. 37(11): p. 2460-2476.
  7. Miller, M.S., S.K. Friedlander, and G.M. Hidy, A chemical element balance for the Pasadena aerosol. J. Colloid Interface Sci., 1972. 39: p. 165-176.
  8. Dunker, A.M., A method for analyzing data on the elemental composition of aerosols. General Motors Research Laboratories Report, 1979. MR-3074(ENV-67).
  9. Schauer, J.J., et al., Source apportionment of airborne particulate matter using organic compounds as tracers. Atmospheric Environment, 1996. 30(22): p. 3837-3855.
  10. Zheng, M., et al., Source apportionment of PM2.5 in the southeastern United States using solvent-extractable organic compounds as tracers. Environmental Science & Technology, 2002. 36(11): p. 2361-2371.
  11. Schauer, J.J. and G.R. Cass, Source Apportionment of Wintertime Gas-Phase and Particle-Phase Air Pollutants Using Organic Compounds as Tracers. Environ Sci Technol, 2000. 34: p. 1821-1832.
  12. Schauer, J.J., et al., Source reconciliation of atmospheric gas-phase and particle-phase pollutants during a severe photochemical smog episode. Environmental Science & Technology, 2002. 36(17): p. 3806-3814.
  13. Henry, R.C., Current Factor-Analysis Receptor Models Are Ill-Posed. Atmospheric Environment, 1987. 21(8): p. 1815-1820.
  14. Henry, R.C., Multivariate receptor modeling by N-dimensional edge detection. Chemometrics and Intelligent Laboratory Systems, 2003. 65(2): p. 179-189.
  15. Henry, R.C., E.S. Park, and C.H. Spiegelman, Comparing a new algorithm with the classic methods for estimating the number of factors. Chemometrics and Intelligent Laboratory Systems, 1999. 48(1): p. 91-97.
  16. Paatero, P., Least squares formulation of robust non-negative factor analysis. Chemometrics and Intelligent Laboratory Systems, 1997. 37(1): p. 23-35.
  17. Paatero, P., The multilinear engine - A table-driven, least squares program for solving multilinear problems, including the n-way parallel factor analysis model. Journal of Computational and Graphical Statistics, 1999. 8(4): p. 854-888.
  18. Poirot, R.L., et al., Comparative application of multiple receptor methods to identify aerosol sources in northern Vermont. Environmental Science & Technology, 2001. 35(23): p. 4622-4636.
  19. Kim, E., et al., Source identification of PM2.5 in an arid Northwest US City by positive matrix factorization. Atmospheric Research, 2003. 66(4): p. 291-305.
  20. Kim, E., P.K. Hopke, and E.S. Edgerton, Source identification of Atlanta aerosol by positive matrix factorization. Journal of the Air & Waste Management Association, 2003. 53(6): p. 731-739.
    21. Maykut, N., et al., Source apportionment of PM2.5 at an urban IMPROVE site in Seattle, WA. Environ Sci Technol, 2003. 37: p. 5135-5142.
  21. Polissar, A.V., et al., Atmospheric aerosol over Alaska - 1. Spatial and seasonal variability. Journal of Geophysical Research-Atmospheres, 1998. 103(D15): p. 19035-19044.
  22. Ramadan, Z., X.H. Song, and P.K. Hopke, Identification of sources of Phoenix aerosol by positive matrix factorization. Journal of the Air & Waste Management Association, 2000. 50(8): p. 1308-1320.
  23. Song, X.H., A.V. Polissar, and P.K. Hopke, Sources of fine particle composition in the northeastern US. Atmospheric Environment, 2001. 35(31): p. 5277-5286.
  24. Xie, Y.L., et al., Identification of source nature and seasonal variations of arctic aerosol by positive matrix factorization. Journal of the Atmospheric Sciences, 1999. 56(2): p. 249-260.
  25. Yakovleva, E., P.K. Hopke, and L. Wallace, Receptor modeling assessment of particle total exposure assessment methodology data. Environmental Science & Technology, 1999. 33(20): p. 3645-3652.
  26. Wahlin, P., COPREM-A multivariate receptor model with a physical approach. Atmospheric Environment, 2003. 37: p. 4861-4867.
  27. Paatero, P. and P.K. Hopke, Utilizing wind direction and wind speed as independent variables in multilinear receptor modeling studies. Chemometrics and Intelligent Laboratory Systems, 2002. 60(1-2): p. 25-41.
  28. Kim, E., et al., Incorporation of parametric factors into multilinear receptor model studies of Atlanta aerosol. Atmospheric Environment, 2003. 37(36): p. 5009-5021.
  29. Lewis, C.W. and R.K. Stevens, Hybrid receptor model for secondary sulfate from an SO2 point source. Atmos. Environ., 1985. 19(6): p. 917-924.
  30. Hakami, A., M.T. Odman, and A.G. Russell, High-Order, Direct Sensitivity Analysis of Multidimensional Air Quality Models. Envrion Sci Technol, 2003. 37: p. 2442-2452.
  31. Dunker, A.M., et al., The Decoupled Direct Method for Sensitivity Analysis in a Three-Dimensional Air Quality Model-Implementation, Accuracy and Efficiency. Environ Sci Technol, 2002. 36: p. 2965-2976.
  32. Dunker, A.M., et al., Comparison of Source Apportionment and Source Sensitivity of Ozone in a Three-Dimensional Air Quality Model. Environ Sci Technol, 2002. 36: p. 2953-2964.
  33. Park, E.S., P. Guttorp, and R.C. Henry, Multivariate receptor modeling for temporally correlated data by using MCMC. Journal of the American Statistical Association, 2001. 96(456): p. 1171-1183.
  34. Sato, M., Importance of identification of the model: some comments on the article 'Bilinear estimation of pollution source profiles and amounts by using mulitvariate receptor models. Environmetrics, 2002. 13(7): p. 799-800.
  35. El-Shaarawi, A.H., Comments on the article 'Bilinear estimation of pollution profiles and amounts by using multivariate receptor models'. Environmetrics, 2002. 13(7): p. 801-802.
  36. Hopke, P.K. and P. Paatero, Remarks on the article 'Bilinear estimation of pollution source profiles and amounts by using multivariate receptor models'. Environmetrics, 2002. 13(7): p. 803-805.
  37. Park, E.S., C.H. Spiegelman, and R.C. Henry, Remarks on the article 'Bilinear estimation of pollution source profiles and amounts by using multivariate receptor models' - Authors' response. Environmetrics, 2002. 13(7): p. 807-809.

 

back to top

 

 

 

 

 

 

 

 

 

 

 

 

 


 

 
 
 
The Organic Speciation International Worskhop is sponsored by the Western Regional Air Partnership/Western Governors Association. APACE is seeking support from the US Dept. of Energy, US EPA Office of Air Quality Planning and Standards, and the National Science Foundation.