p                                                                EPA/#
p                                                             June 2000

:                Workshop on UNMIX and PMF

I                        As Applied to PM2>5

                            14-16 February 2000
                            U.S. EPA, RTP, NC

                               Final Report


                                 Robert D. Willis
                        ManTech Environmental Technology, Inc.
                          Research Triangle Park, NC 27709

                              Contract No. 68-D5-0049
                                 Project Officer
                                   Portia Britt

                             Work Assignment Manager
                                 Charles W. Lewis
                         National Exposure Research Laboratory
                    Human Exposure and Atmospheric Sciences Division
                         National Exposure Research Laboratory
                          Office of Research and Development
                          U.S. Environmental Protection Agency
                          Research Triangle Park, NC 27711

    The U.S. Environmental Protection Agency through its Office of Research and Development funded and
managed the research described here under Contract 68-D5-0049 to ManTech Environmental Technology. Inc.
It has been subjected to the Agency's peer and administrative review and has been approved for publication
as an EPA document.

                             3.RECIPIENTS ACCESSION NO.

Workshop on UNMIX and PMF as Applied to PM2.5
                                                         6 PERFORMING ORGANIZATION CODE
Robert D.Willis
                             8.PERFORMING ORGANIZATION
                             REPORT NO.

ManTech Environmental Technology, Inc.
Research Triangle Park, NC 27711
                             10J»ROGRAM ELEMENT NO.
                                                         11. CONTRACT/GRANT NO.


National Exposure Research Laboratory
Office of Research and Development
U.S. Environmental Protection Agency
Research Triangle Park, NC 27711
                             13.TYPE OF REPORT AND PERIOD

                             EPA Report; February 14-16,2000
                             14. SPONSORING AGENCY CODE

This report is the proceedings of a workshop convened to evaluate two new air quality receptor models,
Positive Matrix Factorization (PMF) and UNMLX. The workshop was held hi Research Triangle Park,
NC, during February 14-16,2000 and was sponsored jointly by EPA's Office of Research and Develop-
ment (ORD) and Office of Air Quality Planning and Standards (OAQPS). The workshop evaluation of
PMF and UNMLX was accomplished by examining the results of applying both models to two ambient
PM2.5 data sets, one real and one synthetically generated. Both data sets were supplied in advance to a
proponent of each model (UNMIX: Dr. Ron Henry, University of Southern California; PMF: Dr. Phil
Hopke, Clarkson University). Each brought to the workshop the results of independently applying their
model to both data sets. The report briefly summarizes the technical exchange and major conclusions
reached during the workshop.	  	

           19. SECURITY CLASS (This Report)

                                      20. SECURITY CLASS /This Faze)

                                            22. PRICE

                    GENERAL DISCLAIMER
    This document may have problems that one or more of the following disclaimer
                           statements refer to:
•  This document has been reproduced from the best copy furnished by the
   sponsoring agency.  It is being released in the interest of making
   available as much information as possible.

•  This document may contain data which exceeds the sheet parameters. It
   was furnished in this condition by the sponsoring agency and is the best
   copy available.

•  This document may contain tone-on-tone or color graphs, charts and/or
   pictures which have been reproduced in black and white.

•  The document is paginated as submitted by the original source.

•  Portions of this document are not fully legible due to the historical nature
   of some of the material. However, it is the best reproduction available
   from the original submission.

Volume I: Workshop Proceedings
Agenda	  1
Introduction 	  3

Session 1:   14 February, a.m	  4
    Opening Remarks	  4
    Session 1A:  UNMIX Methodology	  4
        UNMIX Results on Synthetic Data Set	  5
        UNMIX Results on the Phoenix Data Set	  6
    Session IB:  PMF Methodology	  6
        PMF Results on Synthetic Data Set	  8
        PMF Results on Phoenix Data Set	  8
    Session 1C:  Overview of Synthetic Data Set Results	  8

Session 2:    14 February, p.tn	 10
    Session 2A:  Description of the Synthetic Data Generation Process  	 10
    Session 2B:  Processing of Synthetic Data and Resulting Solutions for PMF	 10
    Session 2C:  Processing of Synthetic Data and Resulting Solutions for UNMIX 	 10
    Session 2D:  Description of Metric  of the Goodness of Fits of the Solutions and the Results
                of Applying the Metric  	 11

Session 3:   15 February, a.m	 13
    Session 3A:  Phoenix Source Apportionment Studies 	 13
    Session 3B:  Phoenix NERL Platform Studies—Data Quality issues and
                Supplementary Analyses	 14
    Session 3C:  PMF Analysis of Phoenix Data	 14
    Session 3D:  UNMIX Analysis of Phoenix Data	 15

Session 4:   15 February, p.m	 16
    Session 4A:  Reexamination of the  Synthetic Data Results	 16
    Session 48:  Demonstration of UNMIX Program 	 16
    Session 4C:  Demonstration of the PMF Program	 17
    Session 4D:  Potential Effects of Data Artifacts on Receptor Modeling Results	 17
    Session 4E:  Open Discussion  	 19

Session 5:   16 February- a.m	 20
    Session 5A:  Application of PMF in the Northern Great Lakes: A Tale of Two Studies	20
    Session 5B:  Discussion of FPEAK, Open Discussions, and Workshop Conclusion  	20

References 	 22
Attendees  	 24

Volume II: Appendices

Appendix 1 A:    UNMIX User's Manual and Presentation Materials for
                UNMIX Theory ai>d Applications	  jA-J
Appendix IB:    A Guide to Positive Matrix Factorization	  1B-1
Appendix 1C:    Presentation Materials forOverview of Synthetic Data Set Results	  1C-1
Appendix 2A:    Presentation Materials for Description of Synthetic Data Generation Process	2 A-1
Appendix 2B:    Presentation Materials for Processing of Synthetic Data and Resulting
                Solutions for PMF	  2B-1
Appendix 2C:    Presentation Materials for Processing of Synthetic Data and Resulting
                Solutions for UNMIX	  2C-1
Appendix 2D:    Presentation Materials for Description of Metric of the Goodness of Fits of
                the Solutions and the Results of Applying the Metric	2D-1
Appendix 3A    Presentation Materials for Phoenix Source Apportionment Studies	3A-1
Appendix 3B    Presentation Materials for Phoenix NERL Platform Studies  	  3B-1
Appendix 3C    Presentation Materials for PMF Analysis of Phoenix Data	  3C-1
Appendix 3D    Presentation Materials for UNMIX Analysis of Phoenix Data	3D-]
Appendix 4D    Presentation Materials for Potential Effects of Data Artifacts on
                Receptor Modeling Results	4D-1
Appendix 5A    Presentation Materials for Application of PMF in the Northern Great Lakes	SA>1
Appendix 5B    Presentation Materials for Discussion of FPEAK	  5B-I
Appendix 6      Journal Article Excerpt Concerning EPA's PAMS Guidance on Reporting of
                Low Concentration Data	6-1

    Volume I
Workshop Proceedings

                Agenda for Workshop on UNMIX and PMF as Applied to 5»M25
                                  Dates: 2/14/2000-2/16/2000
                Location: EPA Administrative Building Auditorium, RTP, NC

February 14,8:30 a.m. - 5:00 p.m.

    Morning Session: (Session 1)
    General presentations on the methodology behind the tools and a brief presentation of the solutions found
    for both the Phoenix and the synthetic data set. This session is geared toward a general audience with the
    purpose of giving an overview of the tools and the results from their applications. The following 4 sessions
    will go into the details and will be at an advanced technical level, thus not for a general audience.

    8:30-8:45   Welcome and Introductions (Chuck Lewis, ORD, and John Bachmann, OAQPS)
    8:45-10:00  Presentation on UNMIX methodology and results for Phoenix and synthetic data set (Dr.
                Ron Henry)
    10:00-10:15 Break
    10:15-11:30 Presentation on PMF methodology and results for Phoenix and synthetic data set (Dr. Phil
    i 1:30-12:00 Overview describing the synthetic data set and a pictorial presentation of how close the tools
                reproduce the "known" profiles (OAQPS)
    12:00-1:00  Lunch

    Afternoon Session: (Session 2)
    Thorough discussions of the results from the synthetic data set analysis. Includes description of the data
    generation, the metric used by EPA to determine how well the tools reproduced the "known" profiles, data
    preprocessing (e.g., outlier identification), selection criteria for which species to use in the models and the
    number of sources to try to fit, and a description of the solutions (identification of the fined sources and
    the uncertainties with these solutions).

    1:00-1:15   Description of the data generation process (OAQPS)
    1:15-2:00   Presentation of processing of synthetic data and resulting solutions  for PMF (Dr. Phil
    2:00-2:45   Presentation of processing of synthetic data and resulting solutions for UNMIX (Dr. Ron
    2:45-3:00   Break
    3:00-4:00   Description of metric of the goodness of fits of the solutions and the results of applying the
                metric (OAQPS)
    4:00-5:00   General discussion topics such as what it means to say that one solution is better than
                another, how to use "known" profiles to compare  with derived solutions for source
                identification, and whether it is realistic to have an automated source identification process
                (General discussion)

February 15,8:00 a.m.-5:00 p.m.                     *  •

    Morning Session: (Session 3)
    Thorough discussions of the results from the Phoenix analysis. Includes steps used to preprocess the data
    to identify' potential outliers, selection of species and number of sources used in the model, estimates of
    confidence (error bars) in the source compositions and contributions, and degree  of fit obtained.

    8:00-8:45   Results from other recent source apportionment studies in Phoenix (Mark Hubble, Arizona
                Department of.Environmental Quality')

    8:45-9:00   Data quality issues associated with Phoenix measurements used in current analyses, and
                supplementary analyses (SEM and trajectory analyses) performed to confirm sources (ORD)
    9:00-12:00  (Break when needed.) Presentations by Hopke and Henry on their respective Phoenix
                analyses, addressing the issues listed above.
    12:00-1:30  Lunch

    Afternoon Session: (Session 4)
    Thorough discussions on how the tools really work. In trying to use the tools over the past few months,
    EPA has had some questions about operating the tools and interpreting the output. This session will be
    a "question and answer" session, where many of the questions will have examples to illustrate them.

    1:30-1:45   Reexamination of the synthetic data results (OAQPS)
    1:45-2:15   Demonstration of UNMIX Program (Dr. Ron Henry)
    2:15-2:45   Demonstration of PMF Program (Dr. Phil Hopke)
    2:45-3:15   Potential effects  of MDL on modeling results (Rich Poirot, Vermont Department of
                Environmental Conservation)
    3:15-5:00   Open discussions on how the tools really work. Questions of interest include:

    (1)  Can the tools identify a source that has a discrete profile change? How different do the before and
        after profiles have to be for the  tools to find two unique sources? (OAQPS has constructed an
    (2)  Should the measured total mass or the reconstructed mass (PM, 5) be included as a fining species or
    (3)  How to identify and handle outliers?
    (4)  UNMIX specific questions: What are the equations behind RA2 and strength/nofce? What do they
        measure? How are "edges" fit,  especially in light of errors? Do the interior (non-edge) points have
        any influence on the solution? Why is it that UNMIX uses at most -15 species and finds at most ~6
        sources? Why does UNMIX often find no feasible solution? How  does a user wisely use the new
        feature in UNM1X2 that allows for source compositions with very negative entries? Implications of
        not using MDLs and uncertainties (which is a continuation of the discussion started in (3))?
    (5)  PMF specific questions: What is rotmat and how can it be used to understand better how much
        rotation freedom there is in the solution? What is the appropriate FPEAK to use? Should multiple
        passes be made using various FPEAKS: one pass to improve source identification at the expense of
        the contribution component, and the second pass to accurately reflect the contribution component at
        the expense of source identification? How are FPEAK, FKEY, and GKEY implemented? Are they
        pan of the regularization component of Q? (OAQPS has constructed an example that shows slightly
        negative FPEAKs are preferable.)

February 16,8:30 a.m. -12:00 p.m.

    Morning Session: (Session 5)
    Discussion of general problems and potential solutions regarding issues such as treatment of secondary
    sources, regional vs local source identification, and recommendations for further research and testing of
    methods. Discuss why factor analysis is "ill-posed" (i.e., produces infinitely many solutions) and begin
    a discussion about how to use multiple receptors with these tools.

    8:30-9:15   Results from applying PMF to data from the Lake Michigan area (Dr. Kurt Paterson,
                Michigan Technological University)
    9:15-12:00  Work on issues listed  above.

    12:00   End of workshop

    This report provides a summary  of the Workshop  on
UNMIX and Positive Matrix Factorization (PMF) as Applied to
PM. .. This 2'/i-day workshop was held at the EPA administra-
tive building auditorium in Research Triangle Park, NC, during
14-16 February 2000. Sponsored jointly by EPA's  Office of
Research and Development (ORD) and Office of Air Quality
Planning and Standards (OAQPS), the workshop was intended
to facilitate an exchange of technical information on the use of
two source apportionment tools as applied to paniculate matter
(PM). PMF and UNMIX represent the current state of the art in
multivariate receptor modeling. Both methodologies attempt to
generate source contribution estimates as well as source compo-
sitions using only the ambient data.
    The workshop evaluation of PMF and UNMIX  was
accomplished by examining the results of applying both models
to two ambient PM:, data sets, one real and one synthetically
generated. Both  data  sets were  supplied  in advance  to a
proponent  of each model (UNMIX: Dr. Ron Henry, University
of Southern California: PMF:  Dr.  Phil Hopke,  Clarkson
University).  Each brought to the workshop the results  of
independently applying their model to both data sets. The source
contributions underlying the synthetic data set were of course
known to the EPA personnel who generated the data set, but this
information was not made available prior to the workshop.
    Approximately 40 attendees representing primarily EPA.
universities, and state environmental agencies attended the work-
shop. A list of attendees is provided at the end of this volume.
    The purpose of this report is to briefly summarize the
technical exchange and major conclusions reached during the
workshop. The organization of the report follows the workshop
agenda. The text of the report is intentionally brief to spare the
reader from overwhelming detail. Interested readers who seek
more detailed  information are referred  to  the  appendices
(Volume II) for hard copies of individual presentations and
supporting materials.
    The references given at the end of this report are intended
to provide  a complete list of all known publications relating to
the theory and application of PMF and UNMIX.
    In addition to this  report, the workshop was recorded on
videotape and the tapes are available for loan on request from
Dr.  Charles Lewis, EPA  (tel:  919-541-3154;  e-mail:  lewis

                                                  Session 1
                                            14 February, a.m.
Opening Remarks
Chuck Lewis (ORDj. John Bachmann (OAQPS). and Shelly
Eberly (OAQPSj

    Chuck Lewis opened the workshop by acknowledging the
efforts of Shelly Eberly who was the primary organizer of the
workshop  and who alternated with Chuck  Lewis as  session
moderator. Lewis stressed that the workshop was not intended
as a "shoot-out" between two competing receptor modeling
approaches in order to declare a winner. Rather, the intent was
to provide researchers with a better understanding  of the
methods in order to  assess the  potential of these tools for
regulatory  and research applications.
    Lev. is provided the following definition of receptor models:
  Receptor models are mathematical procedures for identi-
  fying and quantifying the sources of ambient air pollutants
  and their effects at a site (receptor)

  •   primarily on the basis of concentration measurements
      at the receptor, and

  •   generally without need of emissions inventories and
      meteorological data.
The two multivariate receptor models that are the subject of the
workshop are much more complicated to understand and use
than those presently in common usage. The potential reward for
the complexity is that these models "do it all." That is, they
generate both source contributions and source profiles, all from
ambient data.

    John Bachmann. Associate Director of OAQPS, stressed the
importance of receptor modeling from the regulatory perspec-
tive. Receptor  models can provide important scientific support
for current (or future) PM standards. In addition, receptor
models can be an important tool in understanding the associa-
tions among PM, visibility, and health effects, and in developing
regulatory control strategies. State-of-the-art tools  such  as
UNMIX and PMF, as well as experienced users of these tools,
will be needed to interpret the large quantity of data expected
from the PM,, Speciation Monitoring Network.

    Shelly  Eberly had  members  of the audience introduce
themselves  and briefly describe their experience in  receptor

    The remainder of Session 1 consisted of overviews of the
UNMIX  and PMF  models and  results by their principal
proponents, Drs. Ron Henry and Phil Hopke, respectively, and
an overview of the synthetic data set. Session 1 was intended as
a less technical summary of the methods and results for the
benefit of managers  and others who were unavailable for the
entire workshop.
Session 1A: UNMIX Methodology
Dr. Ron Henry, University of Southern California
(Full presentation is in Appendix 1 A.)

    Dr. Henry presented the theory1 of the UNMIX model from
a geometric perspective. The fundamental problem tor receptor
models  is posed  as  follows: Given  an  ambient  data  set,
find—with  as few assumptions as possible—the number of
sources, the composition and contributions of the sources, and
the uncertainties. However, the problem as presented  in the
conventional mass balance formulation is statistically ill-defined,
i.e., there exist an infinite number of solutions that have the same
root mean squared error and that satisfy the non-negativity
requirement for source compositions and contributions. The keys
to finding a unique solution are therefore (1) to determine the
number of sources in the data that are above the noise level, and
(2) to find additional constraints  that limit the number of

    The UNMIX model takes a geometric approach to these two
key problems that exploits the covarjance of the ambient data.
Simple two-element scatterplots of the ambient data provide a
basis for understanding the UNMIX model. For example, a
straight  line and high correlation for Al versus Si can indicate a
single source for both species (soil), while the slope of the line
gives information on the composition of the soil source. In the
same data set, iron does not plot on a straight line against Si,
indicating  other sources  of Fe  in  addition to soil.  More
importantly, the Fe-Si scatterplot reveals a lower edge. The
points defining this edge represent ambient samples collected on
days when the only significant source of Fe was soil. Success of
the UNMIX model hinges on the ability' to find these "edges"  in
the ambient data from which the number of sources and the
source compositions are  extracted.  UNMIX uses principal
component analysis to find edges in m-dimensional space, where
m is the number of ambient species. The problem of finding
edges is more properly  described as finding hyperplanes that
define a  simplex. The vertices at which the hyperplanes intersect
represent pure sources from which source compositions can be
determined. However, there is measurement error in the ambient
data that "fuzzes" the edges making them challenging to find.
UNMIX employs an "edge-finding" algorithm to find the best
edges in the presence of error. Once the edges are found, the
major issue remains of estimating the number of sources.
UNMIX  finds the number of sources  using a resampling
technique (NUMFACT algorithm) in which random subsets of
samples  are successively fit with  UNMIX.  Results  for major
sources change little during the resampling, while minor sources
show considerable variability. NUMFACT calculates a signal-
to-noise  (S'N) ratio for  each factor, and  results with real data
sets indicate that a S/N ratio >2  is an effective rule of thumb  in
estimating the number of quantifiable sources.
    Using only ambient data, UNMIX outputs the  following

        Number of sources

    •   Composition of each source

    •   Source contributions to each sample

    •   Uncertainties in the source compositions

    •   Apportionment of the average total  mass, if total mass
        is included in the model

    The major assumptions employed in UNMIX are as follows:

    •   Source compositions remain approximately constant.
    •   There are at least N • (N-1) points that have low or no
        impact from each of the N sources, i.e., need some
        points with one source missing or low.

    Advantages  of the  UNMIX  tool  were  given  as  the

    •   No assumptions about the number or compositions of
        sources are needed.

    •   No assumptions or knowledge of errors in the data are

    •   UNMIX automatically corrects source compositions for
        effects of chemical reactions.

    A major difference between UNMIX and PMF  is that
UNMiX does not make explicit use of errors or uncertainties in
the ambient concentrations. This is not to imply that the UNMIX
approach regards data uncertainty as unimportant, but rather that
the UNMIX model results implicitly incorporate error in the
ambient data.

UNMIX Results on Synthetic Data Set
    Henry summarized his seven-source UNMIX model for the
synthetic data set. UNMIX source  apportionment results are
summarized in the following table:
Steel sinter
Residual oil
Palladium source
Asphalt roofing
Mean Source Contribution

                                         Steel   Sinter
                        0 .  	
                                       50      100     150    200    250    300     350    400
wind directions for these samples are normalized to the hourly
wind-direction data for all sanr-1 s and the relative frequency is
then plotted for each 10-degree wind sector. The plot shows that
on days when the steel sinter source has a high expected source
contribution, the winds are three times more likely to be from
200 to 220 degrees than the average frequency over all samples.

UNMIX Results on the Phoenix Data Set
    Dr. Henry presented a six-source UNMIX solution for the
Phoenix PM: < data set as summarized in the following table:
                              Mean Source Contribution
       Vegetative burning
Secondaries include  sulfaies  and organic carbon.  Source
compositions are shown in Appendix I A. It should be noted that
the "unexplained" source represents a real source (or mixture of
real sources) that was extracted by UNMIX but could not be
specifically identified.
    The identification of the "diesel" source hinged on the high
Mn concentration and the high OC and EC concentrations, as
well as the fact that this source contributed only one-fourth as
much on the weekends as on weekdays. Henry speculated that
the Mn is a fuel additive used (probably illegally) by diesel truck
operators to prevent engine fouling. Time-series plots for the
different sources are consistent with their identification, e.g.,
vehicle  source peaks during  the winter months, while the
secondary source peaks during the summer.

Session 1B: PMF Methodology
Dr. Philip Hopke, Clarkson University
(Full presentation is in Appendix IB.)

    PMF is a recently developed least squares formulation of
factor analysis with built-in non-negativity constraints. PM F was
developed by Dr. Pemti Paatero in Finland in the mid-1990s.
The tool is currently being refined jointly by Paatero and Hopke.
The following is excerpted from Hopke and Song. Appendix 2B:

    "Suppose X is a n by m data matrix consisting of
the measurements of n chemical species in m samples.
The objective of multivariate receptor modeling is to
determine the number of aerosol sources, p, the chemi-
cal composition profile of each source, and the amount
that each of the p sources contributes to each sample.
The factor analysis model can be written as:
where G is a  n by p matrix of  source  chemical
compositions (source profiles) and F is ap by m matrix
of source contributions (also called factor scores) to the
samples. Each sample is an observation along the time
axis,  so F  describes the temporal  variation  of the
sources.  E represents the pan of the data variance un-
modeled by the p-factor model.
    In PMF,  sources are  constrained to have non-
negative species concentration, and no sample can have
a negative source contribution. The error estimates for
each observed data point were used as point-by-point
weights.  The essence of PMF can thus be presented as:
    min Q(X,o,G.F)

with g,t 2 0 and fu 2 0 for k = l,...,p, and o is the known
matrix of error estimates of X. Thus, this is a least squares
problem with the values of G and F to be determined.
That is. G and  F are determined so that the Frobenius
norm of E divided by o (point-wise) is minimized. As
shown by Paatero and Tapper [1], h is  impossible to
perform factorization by using singular value decompo-
sition (SVD) on such a point-by-point weighted matrix.
PMF uses a unique algorithm  in which both G and F
matrices are varied simultaneously in  each least squares
step. The algorithm was described by Paatero [2].
        Application of PMF requires that error estimates for
    the data be chosen judiciously so that the estimates
    reflect the  quality and reliability of each of the data
    points. This feature provides one of the most important
    advantages of PMF, the ability to handle missing and
    below-detection-limit data by adjusting the correspond-
    ing error estimates. In the simulated data, there were
    some below-detection-limit values for different chemical
    species. As the input  to the PMF program, the  con-
    centration data and the associated error estimates were
    constructed as follows: For the measured data (above
    detection limit), the concentration values were  used
    directly, and the error estimates were built as the
    analytical uncertainty plus a quarter of detection limit.
    For the below-detection-limit data, half of the detection
    limit was used as the  concentration value, and as the
    error estimate as well. This strategy  [3] appeared to
    work well in the present study."

    Excerpt from Appendix IB:

       "Another important  aspect of weighting of  data
   points is the handling of extreme values. Environmental
   data typically shows a positively skewed distribution and
   often with a heavy tail. Thus, there can be extreme values
   in the distribution as well as true "outliers." In either
   case, such high values would have significant influence
   on the solution (commonly referred to as leverage).  This
   influence will generally distort the solution and thus an
   approach to reduce their influence can be a useful tool.
   Thus, PMF offers a "robust" mode. The robust factori-
   zation based on the Huber influence function [Huber,
   1981]  is  a  technique  of iterative  reweighing of the
   individual data values."

    A critical step in PMF analysis is the determination of the
number of sources. Plots of the scaled residuals for all species
can help determine the number of factors. It  is desirable to have
symmetric distributions and to have all the residuals within ±3
standard deviations. If there is asymmetry or a larger spread in
the residuals, then the number of factors should be reexamined.

    Note: The definition of F and G are interchanged throughout
this report. In some places F represents the source compositions
and G represents the source contributions and in other places F
represents the source contributions and G represents the source
compositions. From a mathematical perspective, this is permis-
sible, although it may lead to confusion for the reader. Most of
the current literature refers to F as the source composition matrix
and G as the source contribution matrix.

PMF Results on Synthetic Data Set
    Hopke presented a nine-factor solution for the simulated
data set as summarized in the following table:
Mt an Source Contribution
Source (ug'm')
Area source
Inner highway
Residual oil combustion
Steel sinter
Asphalt roofing
Municipal incinerator
Petroleum refinery
Lime kiln
Extra area source
    The major sources were the area  source and the inner
highway source. All factors showed a reasonable relationship to
the true source profiles provided to the modelers: For many
factors, concentrations of most species were within  1-sigma
uncertainty of the synthetic concentrations. Plots of residuals for
selected species were generally symmetric and were contained
within =2 sigma. Residual plots are a useful aid in deciding how
man>  factors are optimal. In the case of the synthetic data set,
residual peaks for some species were relatively broad and
asymmetric when  fewer than  nine factors were  used. A
scatterplot of the  modeled  mass versus  the synthetic mass
showed excellent agreement.

PMF Results  on Phoenix Data  Set
    PMF yielded a six-source model for the Phoenix PM,} data
set as summarized in the following table:
Biomass burning
Motor vehicles
Coal-fired power plant
Sea salt
Mean Source Contribution
    Motor vehicle emissions and biomass burning were the
major sources. It is noteworthy that PMF was able to extract the
sea-salt factor  even  though  concentrations  for the key
determining  species (Na and Ci) were mostly below their
respective detection limits. This source was not found with the
UNMIX model  because the Ma and Cl were not good-fining
species. Time-series plots for the six factors showed that most
source contributions generally peaked during the winter; how-
ever, the sea-salt source showed aperiodic episodes. Modeled
 mass and observed mass were generally in good agreement. PMF
 was also applied to the PM^ data and a five-factor model gave
 best results. The five sources were identified as (1) soil, (2) con-
 struction, (3) road dust, (4) sea salt, and (5) coal-fired power
 plant. Soil and construction were the major sources.

     In summary, Hopke cited the following advantages of PMF:

     •    PMF allows optimal weighting of individual data
         points. This in turn makes it possible to include less
         robust species (those with many missing  values or
,f        values below the detection limit) that may nevertheless
         define real sources.

    •    PMF provides for natural inclusion of non-negativity
         and other constraints.

    •   The PMF approach will allow future inclusion of better
        algorithms for finding die optimal number of factors.
 Session 1C: Overview of Synthetic
 Data Set Results
 Shelly Eberly, OAQPS
 (Full presentation is in Appendix 1C.)

     Ms. Eberly provided a brief overview of the synthetic data
 and a comparison  of the PMF  and UNMIX results to the
 synthetic data. Eberly's remarks addressed the following topics:

     •   A description of how the synthetic  data  set was

     •   Discussion of the 16 distinct sources that were input
        into the model. (Temporal modulation of the synthetic
        sources was critical in being able to resolve individual

     •   The geographic layout of "Palookaville."

     •   A summary of the average source contributions used to
        generate Palookaville's ambient data.

     •   Summary  of the materials provided to the analysts
        (Hopke and Henry).

     •   Summary of the materials received from the analysts.

     •   Side-by-side  comparison of the  sources identified
        by UNMIX and PMF  and the source  contribution

         Comparison of UNMIX and PMF results to the known
         results. This comparison is shown below:
               Comparison to Known Profiles
       Sources identified by both tool*
       (known / UNMIX / PMF;
       - Area /Soil /Area
       - mm? Hwr/WucH/ Inner Hwx
       - fto««u*! Oil Combustion
       - Mitn. Inon
OB/ 6/ IS
       - Asphah Roofing

       Source Identified by UNMIX only
       - Palladium fourct ;-3>
      Sources identified by PMF only
      (known /PMF)
      - P«tro Refm / Petro Refm           06/1
      - Urn* kim/ lime krin               06/5
      - Coal Comb / Enta Area            15/2
    •Note: Originally the above chart did not have the "muni-
    cipal incinerator" source in the category of "Sources identi-
    fied by both tools." UNMIX had identified the source, but
    under the label "Combustion source located toNE of site."

    •   Comparison of UNMIX and PMF residual oil com-
        bustion source profiles to the synthetic source profile.

    •   Scanerplots of  UNMIX source strength versus  true
        source strength  and PMF source strength versus  true
        source strength for the residual oil combustion source.

    Eberly offered the following conclusions:

    •   The  largest three known sources  were  correctly
        identified by both tools and the modeled  mass  was
        close to the simulated mass for all three sources.

    •   The fourth largest source (coal combustion, presence of
        source withheld from analysts) was not identified by
        either tool. PMF  found a  source similar to the coal
        combustion  source but identified it as an extra area
        source. UNMIX did not find the source.
chosen between 5% and 10%. These numbers were used as the
coefficients of variation (CVs) for log-normal distributions of
the measurement errors of the species. Daily random measure-
men^ error drawn from this distribution was applied after the
"true''  species concentration at the receptor was computed.
    An MDL for each species was provided. These MDLs were
computed as  a function of the average concentration and the
species' measurement error CV. Specifically, the MDL for each
species was computed as the maximum of  1.5 x CV  * (mean
concentration) and 0.001 ug/m'. The data below the MDL were
not modified in any way.
    As a consequence of not modifying the data below MDL,
Henry pointed out that scatterplots of certain species revealed an
unrealistic structure of sub-MDL data in the synthetic data set.
For example, although all values of iodine were below the MDL,
scanerplots of iodine values versus other selected species
showed high r values, indicating that the synthesized iodine data
were not truly noise.
    •   Three to four  smaller known point  sources  were
        identified but the estimated source contributions were
        larger than the true source strengths.

    Following Eberly's presentation, the session was opened
for questions to any of the previous presenters. Eberly was asked
how the synthetic uncertainties and minimum detection limits
(MDLs) were determined. Response: Each of the 50 species had
a single MDL and a single uncertainty, which were fixed across
the entire data set. For each species a number was randomly

                                                 Session 2
                                           14 February, p.m.
    Drs. Hopke and Heniy described in more detail their PMF
and UNMIX solutions for the synthetic data set. Dr. Basil
Coutant discussed goodness of fit (GOF) metrics for evaluating
receptor model solutions and  the results of applying GOF
metrics to the PMF and UNMIX solutions.
Session 2A: Description of the
Synthetic Data Generation Process
Dr. Basil Coutant, OAQPS
(Full presentation is in Appendix 2A.)

    Dr. Coutant provided a more detailed description of how
the synthetic data set was generated. Sixteen distinct source
profiles were used in Palookaville—nine point sources, four
industrial complexes, one area source, and two highways. The
area profile was a mixture of dust and road profiles. All source
profiles with the exception of the petroleum refinery were
fixed. The latter profile had some built-in variability (coef-
ficient of variation of approximately 25%). Temporal modula-
tion of the source strengths (50% CV for most) was found to
be essential in being able to resolve the sources by PMF or
UNMIX. A total of 366, 24-h samples were generated at the
receptor site.
    There was further discussion regarding MDLs. Data below
the MDL should be noise with no structure. What does it mean
to quote a value below the MDL? Some laboratories report
values and uncertainties only  for data above the MDL, while
other labs (and the IMPROVE network on occasion) report
values below  MDL. Lewis  presented EPA documentation
reflecting the EPA view that it is perfectly allowable to report
sub-MDL values (at least in the AIRS database for VOCs). See
Appendix 6, quote from JAWMA 4£, 71 (1998).
Session 2B: Processing of Synthetic
Data and Resulting Solutions for PMF
Dr. Phil Hopke
(Full presentation is in Appendix 2B.)

    Dr. Hopke described how the synthetic data set was analyzed.
Initial trials with PMF yielded low Q values indicative of incorrect
weighting of the  data. Alternative data weights were evaluated
until the Q values became more reasonable (approximately equal
to the sample  size). At this point, plots of residuals are very
helpful in determining the optimum number of factors. Generally.
residual peaks that are broad for a whole suite of elements imply
die need  for more factors; residual  peaks that are positively
skewed imply the need for another factors); residual peaks that
are negatively skewed imply the need for fewer factors. PMF with
nine factors seemed to yield the  best results. Trials with eight
factors left some residual peaks with positive tails, while PMF
with 10 factors failed to extract a physically interpretable 10th
factor. Scatterplots of predicted mass versus the actual mass reveal
whether PMF results consistently underpredict or overpredict the
known mass and may provide additional guidance on whether the
optimal number of factors has been used. The PMF model was run
multiple times starting with totally random source profiles to
ensure there was a robust solution.
Session 2C: Processing of Synthetic Data
and Resulting Solutions for UNMIX
Dr. Ron Henry
(Full presentation is in Appendix 2C.)

    Dr. Henry typically begins an UNMIX analysis with
graphical analysis of the data. UNMIX provides the ability to

view scanerplots of the data. Scatterplots of all species versus
mass are very useful in choosing those species that influence the
mass and should be included in the analysis. Henry looks for
straight lines between species, which can suggest a common
source. He also tries to select species whose scanerplots yield
well-defined edges. Scanerplots can also  be  used1 to identify
outliers in the data, which can be removed  if desired.
    Henry typically runs UNMIX multiple times, varying the
fitting  species and'oi the number of factors. UNMIX will
consistently extract the major sources, but the minor'sources
come and go during successive runs. Wind-frequency plots can
be helpful in locating and identifying sources, even weak sources
that cannot be quantified. Based on these plots, Henry located
his Palookaville sources as follows: residual oil combustion
(10-30 degrees): incineration combustion (broad, 30-50 and
60-80): Se source (broad,  20-40); steel sinter (200-220);
aircraft jet fuel (200-220): asphalt roofing (210-230); Pd source
(260-280): Mg source (215-235). Interestingly, the location for
the airport (aircraft jet fuel source)  determined  by Henry
disagreed with the airport location as shown on the Palookaville
map (see Appendix 1C), which placed the  airport north of the
receptor. Subsequent  examination of  the synthetic  data set
simulation by OAQPS revealed that the  airport, asphalt roofing
manufacture, and steel sinter sources were, in fact, inadvertently
located in the same place—about 200 degrees from north, just as
found by Henry and in subsequent wind-direction analyses by

Session 20: Description of Metric of
the Goodness of Fits of the Solutions
and the  Results of Applying the Metric
Dr. Basil Coutani, OAQPS.
(Full presentation is in Appendix 2D.)

    Dr. Coutant discussed  goodness  of  fit (GOF) metrics
developed by EPA to determine how well the tools reproduced
the "known" profiles and'br contributions. Ideally, one would
like a  single GOF  number that can  indicate how closely the
model  results approximate the profile matrix or the contribution
matrix. Two GOF metrics were described—a mean based and a
median based, both of which measure the  relative error in the
apportioned species mass from a source. Both metrics sum these
relative errors for the largest three sources only.
    Both  metrics  were applied to the PMF and UNMIX
synthetic data set solutions. The mean- and median-based GOFs
yielded substantially different results. In particular, the mean-
based  metric is very  sensitive to the largest relative errors. In
these metrics developed  by Coutant,  all  species are treated
equally (no weights). There was some discussion as to the merits
of (1)  unequal weighting of species and (2) making the metrics
independent of the number of fitting species. In addition to GOF
merries for the source profiles, Coutant  described  GOF metrics
for (1) the source contribution matrix and (2) the raw data, and
discussed the results of applying these metrics to the PMF and
UNMIX solutions. Coutant presented an algorithm intended to
automatically identify source profiles generated by UNMIX or
PMF. For a given source profile, the algorithm finds the best
match from a list of candidate profiles. (These might come from
the SPEC1ATE  source profile library,  for example). The
automated profile identification algorithm was applied to the
PMF source profiles with promising results. The  algorithm
works better as more species are included. A minimum of 30
species  is recommended. Some of the  audience  expressed
concern about making such a too)  available to inexperienced
receptor modelers, while others felt that such a tool could assist
even experienced receptor modelers in coming up with a short
list of potential source identifications. There followed  some
discussion of the quality and reliability of SPEC1ATE source
profiles. SPEC1ATE profiles certainly have error associated with
them; are these  errors considered  in the spectral matching
algorithm? In some cases, automated source identification using
the SPECIATE library might be a step backward compared to
reliance on knowledge of local sources. Coutant concluded that
"the  profile  GOF metrics have worked well:  they let one
objectively identify sources, [and] they provide a systematic way
of measuring the overall quality of the fit."

    Session 2 concluded with a general discussion and questions
for the presenters. Henry responded to a question about physical
constraints in UNMIX. UNMIX presently does not allow the
user to impose constraints on the source profiles (e.g., the user
may know from experience that a certain species is absent from
a source), but this could be implemented in future versions. PMF
presently has only non-negativity constraints built-in, but it is
possible through the regularization  functions to force specific
source contributions or profile components toward zero. Henry
expressed his concern that the errors in both tools are not being
properly estimated. As a next step in model validation, Henry
proposed development of a synthetic data set with variable
source profiles and more realistic error structure. UNMIX and
PMF should be run on 1000 different data sets and  the errors
estimated by the models should  be compared with the standard
error of the synthetic data set to see if the model error estimates
are realistic.
    There was some discussion regarding how well the models
deal 'with secondary' aerosols.  Basically, secondaries  are a
challenge for the models. In the case of regional transport, one
might be able to combine UNMIX or PMF with back-trajectory
methods or regional transport models. Stratifying the ambient
data  set by season  and'or wind direction  may  improve the
apportionment of secondaries; however, one must be careful not
to make the data sets too small in the process.
    Asked how Henry and Hopke  view each other's model,
Henry reiterated his philosophy that it is best to do  as little as
possible to the data and let the  data  speak  for  itself. He

expressed his concern that by weighting the data as PMF does,
one runs the risk of puning additional  distance between the
statistical model and the physical reality. Hopke argues that the
ability to weight individual data points allows the modeler to
extract the most information from the data.

                                                Session 3
                                           15 February, a.m.
    Session 3 began with a description of the Phoenix area and
results from three earlier source apportionment studies. This was
followed by results from an independent analysis of the same
Phoenix data set to which the UNMIX and PMF models were
applied. The session concluded with thorough discussions of the
UNMIX and PMF results from Phoenix, including steps used to
preprocess the data to identify potential outliers, selection of
species and number of sources used in the model, estimates of
confidence (error bars)  in the source compositions and con-
tributions, and degree of fit obtained.

Session 3A: Phoenix Source
Apportionment Studies
Mark Hubble, Arizona Department of Environmental Qualirj,'
(Full presentation is in Appendix 3A.)

    Mark Hubble  described the Phoenix geography, meteor-
ology, and major  emissions sources. Hubble also presented
results from three source apportionment studies carried out in the
Phoenix area:

    1.   1989-1990 Urban Haze Study (principal investigators:
        John  Watson  and Judith Chow, Desert  Research

    2.   1994-1995 Maricopa Association of Govemments/DRJ
        Brown Cloud Analysis (principal investigators: Tom
        Moore et al.. Arizona Department of Environmental
        Quality, and Eric Fujita, Desert Research Institute)

    3.  1994-1996 ADEQ/ENSR Analysis (principal investi-
        gators:  Tom  Moore et al,  Arizona Department of
        Environmental Quality, and Steven Heisler, ENSR)

The first two studies were conducted during the fall and winter,
while the last study was conducted  during all seasons. The
Urban Haze Study used conventional chemical mass balance
(CMB7) to apportion fine mass (PM2S) and light extinction to
source categories. Local motor vehicle and geological source
profiles were generated. The Brown Cloud Study used con-
ventional and extended CMB to apportion fine mass only. The
extended CMB included selected semivolatile  organic com-
pounds and polycyclic aromatic hydrocarbons to separately
apportion gasoline and diesel  combustion. The ADEQ/ENR
Study used conventional CMB to apportion fine mass and light
    Results from the first two studies were in general agree-
ment and showed that motor vehicles contributed the bulk of
PM, 5 (in the range of 44-75%) and that geological sources
were typically the second most abundant source of (PMj,),
accounting for approximately  10-20% of PMi5. Ammonium
nitrate and  ammonium sulfate were smaller but significant
contributors to PMJS. The third study differed  from the first
two  studies in  that  it was conducted year-round  and it
attempted  to apportion  vegetative  burning using soluble
potassium. The apportionment results showed a significant
increase in vegetative burning (11-17%  of PM25) and
geological sources (26-33%) at the expense of motor vehicles
(typically <40%of PMJ5). However, the vegetative burning
source is probably overestimated since the model indicates that
it contributes 15-20% of PM35 during the summer months,
when there should be little vegetative burning.

    In conclusion:

    • ' All studies show that most fine mass comes from

    •    All show similar proportions between geological and
        combustion source categories.

    •   All show rather low contributions  from secondary
        nitrate and sulfate.

Session 3B: Phoenix NERL Platform
Studies—Data Quality  Issues and
Supplementary Analyses
Dr. Can- Harris. KERL
(Full presentation is in Appendix 3B.)

    Dr. Morris discussed the following topics in regard to the
Phoenix NERL platform data:

    •   NERL Platform data (measurements, sampling equip-

    •   Receptor modeling results

    •   Scanning electron microscopy results

    •   Health effects studies

    The NERL monitoring platform in Phoenix provided data
that was submitted to Drs. Henry and Hopke for the UNMIX and
PMF analyses. The data consisted of collocated measurements
from a dual fine-panicle sequential sampler(DFPSS), a dichoto-
mous sampler, TEOMs, and a 10-m meteorological tower. The
DFPSS data were the subject of analysis unless otherwise noted.
The data were collected between 1 February J995 and 30 June
1998. Norris et al. carried out their own chemical mass balance
receptor modeling study, which has recently been submitted for
publication.  This study attributes 42.2% of PM2} to  motor
vehicles, 24.5% to road dust, 17% to secondary organics, 9.5%
to ammonium bisulfate, 5.4% to wood smoke, and 1.4% to
marine aerosol. Norris suggested that secondary organics may
represent a positive artifact on the quartz filter, which may
accour  for some of Hopke's "biomass burning" source and
Henry ,* "secondary" source.
    Scanning electron  microscopy was used to validate  the
receptor model results and to provide evidence for additional weak
sources. For example.back-trajectories pointing toward the Pacific
combined with SEM images of sah aerosols provided confirmation
of the marine source. SEM also identified particles suggestive of
smelting operations and an unrelated source(s) of Pb particles.
    Health effects  associated with the Phoenix aerosol were
analyzed in a recent study by Mar et al. (Associations between
Air Pollution and Mortality in Phoenix, 1995-1997). Cardio-
vascular mortality was significantly  associated with  PM;},
coarse PM, and elemental carbon. Factor analysis revealed that
combustion-related pollutants  (motor vehicle exhaust and
vegetative  burning) and  secondary aerosols (sulfates) were
associated with cardiovascular mortality'.
Session 3C: PMF Analysis of Phoenix Data
Dr. Phil Hopke
(Full presentation is in Appendix 3C.)

    Dr. Hopke discussed his PMF analysis of the Phoenix data.
Hopke found a six-source model for Phoenix. In order of
descending mass contribution, these sources were  biomass
burning, motor vehicles, coal-fired power plant, soil, Cu smelter,
and sea salt. Time-series plots of the six sources showed reason-
able seasonal trends. Sea salt and soil were episodic in nature;
motor vehicles, biomass burning,  and perhaps the Cu smeller
source appear to peak in winter. Wind-directional analysis of the
copper smelter  source  might clarify  whether this  is  being
transported across the Mexican/US, border. Because PMF
allows the user to fill in missing data or replace sub-MDL data.
Hopke was able to use Na, CI, and Cu species to advantage in
extracting the sea-salt and copper smelter sources, in contrast to
the UNMIX solution.
    Determining the number of factors to include in the  model
is a multistep process. After obtaining a trial PMF solution, the
total mass (PM.,) is regressed on the source contributions to
apportion the mass  to  each  of the sources. If any of the
coefficients in this regression are negative, then there likely are
too many factors in the model. Another technique for evaluating
the number of factors is to examine the standardized residuals by
species. If these residuals are not  symmetric  or if there are a
number of residuals more than three standard deviations from
the mean, this may indicate there are too many or too  few factors
(although it may also indicate that the uncertainties provided to
PMF by the user are not appropriate).
    Once the number of factors has been determined, then the
correct rotation for the solution needs to be determined. One
easy way to rotate  the solution is through the  parameter
FPEAK. Graphing Q against different values of FPEAK is a
useful diagnostic for selecting the appropriate rotation. As a
general rule of thumb,  one should increase  FPEAK until Q
starts to rise.
    Although the selection of the number of factors and the
appropriate rotation are presented here as independent steps,
they, in fact, interact. For example, after selecting FPEAK, one
should reexamine the residuals to be sure they are still small and
symmetric and reexamine the regression coefficients to be sure
they are still non-negative.
    As an aside, Hopke separately applied PMF to data col-
lected with the DFPSS and to data  collected with the collocated
dichot sampler. The results lent support to the modeling  results
since the resulting source profiles  for the two samplers looked
very similar with the exception  of sea salt and soil.  These
typically represent coarse-fraction intrusion and were affected by
the different inlet efficiencies for the two sampling  systems.

    Note: An eight-source model, whose results differ con-
siderably from the six-source model presented at the workshop
and are much more similar to the UNMIX  results,  has been
submitted for publication (Ramadan et al., JAW MA, in press).

Session 3D: UNMIX Analysis of Phoenix Data
Dr. Ron Henry
(Full presentation is in Appendix 3D.)

    Dr. Henry discussed his six-source solution for Phoenix
using UNMIX. He excluded Na, Cl, and Cu from the list of
fining species because scatterplots versus mass indicated that
little mass was associated with these species. Also, there were a
large  number of measured values below the detection limit.
Henry's six sources in order of decreasing mass contribution
were non-diesel vehicles (37%), secondaries (20%), soil (15%),
diesels (10%), vegetative burning (5%), and unexplained (12%).
In contrast to Hopke, Henry used soil-corrected potassium as a
fining species. The correction was made by using the lower edge
in the potassium versus silicon scatterplot as an estimate of the
soil potassium. Non-soil potassium proved to be very important
in being able to extract the weak vegetative burning source. The
secondaries source was high  in S and  organic carbon. The
unexplained source, distinguished by Br and OC, is probably a
mixture of sources according to Henry- (Phoenix has a surprising
number of local OC sources according to  Henry,  although
regional transport of OC is another possibility'.) Several factors
supported  Henry's labeling of the diesel source. First was the
high EC component. Second, Henry compared the diesel
contributions on weekdays versus weekends and found nearly a
factor of 4 decrease on the weekends, consistent with com-
mercial truckers' reluctance to work on weekends. (The other
sources, if anything, may have shown a tendency toward higher
contributions on the weekends.) Third, some research on the
Internet indicated that it is common  practice among truckers
(though possibly illegal) to add MMT(an octane-enhancing fuel
additive) to their fuel to minimize engine fouling. This could
then explain the targe Mn component in the diesel source profile.
Unfortunately, no  traffic count  data were  available in the
Phoenix area  showing the  number  of diesel  vehicles on
weekends  versus weekdays. Henry also presented the 1-sigma
source composition errors that can be generated by UNMIX. By
dividing each contribution in the source profile matrix  by its
associated error, one calculates the normalized signal-to-noise
values for the source profiles. With the exception of vegetative
burning (the weakest source), the great majority of these values
are greater than 2.
    Time-series plots of the six sources showed reasonable
seasonal  cycles.  Vegetative burning  and non-diesel vehicle
sources peaked in winter, while the secondaries peaked in
September-October. "Unexplained"' had no discernible pattern.
 In contrast to the synthetic data set, wind-directional  plots
 showed little directionality to the sources, and any directional
 trends that did show  up were probably driven by seasonal
 changes in wind direction. (Winds are more likely to come from
 the north during the winter and the top 10% samples for the
 vehicle source are most likely to occur in the winter, so the
 wind-direction plot for the vehicle source will be skewed toward
 the north.)
    As an aside, Henry included PM10 and PM2 5 masses from
 collocated TEOM samplers in the UNMIX model and generated
 a seven-source  solution. Six of  the sources reproduced the
.previous six-source solution very well. In addition, the DFPSS
 fine mass and the TEOM fine mass apportioned to each of the
 six sources  were in remarkable agreement. The additional
 seventh source appeared to be associated with PM,0.

    Session 3 concluded with a brief comparison of the UNMIX
 and PMF solutions to the Phoenix data as summarized in the
 following table:
Biomass burning
Motor vehicles
Coal-fired power
Sea salt
Vegetative burning
There were some major differences in the two solutions. The
largest source in the PMF solution was biomass burning,
accounting for nearly  35% of the mass.  By  comparison,
UNMIX's vegetative burning accounted for only 5% of the
mass. It is worth noting that Henry used non-soil K to extract his
vegetative  burning source,  while Hopke did  not.  Hopke
speculates that his biomass burning source may be a combination
of Henry's diesel and unexplained sources, which account for
about 22% of the mass.  Motor vehicles account for about 28%
of the mass in PMF versus 47% in UNMIX (combining diesels
plus non-diesel). Based on profile similarities, Hopke's coal-
fired power plant source, accounting for about 17% of the mass,
appears to be equivalent to Henry's secondaries source, repre-
senting 20% of the mass. Hopke's soil source accounts for about
15% of the mass, the same as Henry's soil source estimate.

                                                  Session 4
                                            15 February, p.m.
    Session 4 included a reconsideration of the synthetic data
results, discussions of how the tools really work,  and live
demonstrations of PMF and UNMIX by Hopke and Henry.
Session 4A: Reexamination of
the Synthetic Data Results
Shelly Eberly

    Eberly reviewed the PMF and UNMIX results for the
synthetic data set and  made some corrections. Specifically,
UNMIX identified four sources larger than noise, including the
municipal incinerator source  (identified by  Henry as solid
material combustion) with an estimated strength of 4 ug/m'.
Both tools tended to overestimate the contributions from the
minor sources. Henry  explained that this is simply  a con-
sequence of the fact that both tools attempt to explain all of the
observed mass with only seven or nine sources rather than the 16
sources that were used to generate the synthetic data. Therefore,
some of the  source  contributions will  necessarily be over-
estimated. Henry emphasized the need  to put error bars on
estimated source contributions when comparing results from
different tools.
    Other issues pertaining to the synthetic data results included
the actual location of the airport in Palookaville. With regard to
putting labels on sources, Henry encouraged modelers to provide
a one-sentence justification for each source label so that readers
will understand how the sources were identified.
 Session 4B: Demonstration
 of UNMIX Program
 Dr. Ron Henry

     Dr. Henry presented a live demonstration of the UNMIX
 program. UNMIX is copyrighted to Henry. The current version
(UNMIX2.1) is available at no charge from Dr. Henry, v/no
requests that users not distribute the program to others. E-mail
Dr. Henry at rhenry@usc.edu to request a copy. In addition to
the program, users will receive a user's manual (PDF format)
and some test input files. Users must have MatLab 5.3 in order
to run UNMIX.
    Ambient data is input to UNMIX as a flat ASCII text file with
column headings. UNMIX has a user-friendly Windows interface.
UNMIX provides some statistical measures  to guide the user
toward the best solution. These include minimum r-square (r) and
minimum signal-to-noise (S *N). Recommended values are r3 > 0.8
and S/N > 2. UNMIX allows the user to set one species as a
"tracer" if desired. This forces all measured mass for that species
into one source. UNMIX has an option for displaying scatterplots
of any species against any other species.  This is very usefiil in
selecting fining species. In the same plots one can identify outliers
and remove them (temporarily) from the data set. One can also
display "edge"  plots. Henry recommends  this as a good way to
find out which species are important. In selecting fitting species,
Henry had the following suggestions: (1) Major species must be
included or die model won't be able to find a solution. (2) Select
"robust" species—i.e., those with few missing or sub-MDL values.
(3) Use as few species as possible, since each additional species
adds error to the analysis and usually degrades the S/N. UNMIX
outputs include the source composition matrix and the source
contributions. Additionally, UNMIX can  estimate errors in the
UNMIX source compositions using a bootstrap approach in which
the model is applied to 100 random subsets of the data. "UNMIX
overnight" is another useful feature that allows the user to try all
possible subsets of a selected set of fitting species in order to find
the optimal solution. This can be a lengthy process and the user
will probably want to limit the number of candidate species to
seven or less.
    Future improvements that Henry would like to see include
(1) a stand-alone version that would not require MatLab and
could potentially run  much faster,  (2) the ability to input
constraints on source compositions, and (3) the ability to save

"fining sessions" with all pertinent information so users can
remember where they've been or reproduce earlier analyses.
Asked whether the quoted uncertainties in the ambient data
could be used to some advantage, Henry reiterated his philo-
sophy that it is best to assume that you know nothing about the
data and that, in his experience, uncertainties are often meaning-
less. Nevertheless, Henry did not entirely rule out the possibility'
that future versions of UNMIX may try to use the information
present in the quoted uncertainties.
Session 4C: Demonstration
of the PMF Program

    The PMF programs are available from Dr. Pentti Paatero via
the ftp site rock.helsinki.fi/pub'misc/pmf. First-time users can
get PMF for a 6-month free trial period after which there is a
license fee. PMF is still primarily a research tool and does not
have a nice graphical interface. Researchers interested in learn-
ing to use PMF are invited to spend a week with Hopke at
Clarkson University.
    PMF can be run through a programmer's file editor (PFE),
which is free shareware downloadable from the Internet. Every
PMF job begins by setting up an '.INI file, which contains all
the parameters needed for the analysis, including the file names
and paths for input data files.
    The output of PMF includes three matrices: the G matrix of
source contributions, the F matrix of source compositions, and
the matrix of residuals. PMF also outputs a text file containing
a log of the current analysis session. The G matrix can be input
to a statistics program in order to cany out the regression versus
mass to get the scaled source contributions. The  PMF program
has no built-in diagnostic tools (e.g., for displaying residual
    Looking to the future, the PMF program may not be refined.
Instead, programming efforts may be directed entirely into the
Multilinear  Engine (ME) program,  which Hopke sees as
replacing PMF (Paatero,  1999). ME is considered more flexible
in its ability to handle the imposition of physical constraints. A
wish list for future versions of ME includes a much more user-
friendly graphical interface,  the ability to input fixed  source
profiles or ratio constraints (e.g., Al:Si ratio), and a stand-alone
version with built-in diagnostics (e.g., residual plots), which will
obviate the need to export results to other software packages.
Hopke speculated that it might be possible to automate to some
extent the search for the optimal FPEAK by, for example,
increasing FPEAK until there is a substantial rise in Q.

    Further discussion of MDLs revealed a general consensus
that there is  considerable lack of agreement on the meaning of
MDLs and  how they are  reported by various  labs. Lewis
provided the  following definitions  of the limit of detection
(equivalent to the MDL) and limit of quamitation:
  From Lloyd Currie, pg. 289, in uX-Ray Fluorescence
  Analysis of Environmental Samples," T.G. Dzubay,
  ed., Ann Arbor Science (1977):

  Limit of Detection      = 3.29 o0
  (fdse positive risk = 5%,
  false negative risk = 5%)

  Limit of Quantitation   *=*= 10 f O0
  (RSD of measured
  concentration = J0%)

  where o0 «= (1.0 - 1.4) x standard deviation of blank
  and    f - 1
    It was noted that the above definitions define methodltmhs,
as distinguished from sample limits. The latter vary from sample
to sample and are more realistic limits because they include the
effects of spectral interferences due to other analytes present in
the particular sample. Some labs report the fixed-method MDLs,
and some report variable-sample MDLs. Also, some labs report
values below the MDL, while others do not. Some statisticians
argue for reporting only raw values plus uncertainties and
dispense  with  the concept of MDLs.  Hopke is currently
investigating the use of a statistical method known as "multiple
imputation" as a way to use existing data to impute missing data,
but this research is in a preliminary stage. The discussion did not
lead to any resolution of the difficult issue of how best to handle
and report nondetected values.
Session 4D: Potential Effects of Data
Artifacts on Receptor Modeling Results
Rich Poirot, Vermont Department of Environmental
(Full presentation is in Appendix 4D.)

    Data artifacts, which can include measurement errors, uncer-
tainties, and various hole-filling replacements for nondetects, can
interfere with the identification of real sources. Poirot discussed
his experience with UNMIX and dealing with nondetect data.
There are two choices for dealing with nondetects: one can censor
die input data to screen out all nondetects, or one can use some
hole-filling techniques to replace nondetects. The former approach
can create a small and biased  subset of the original data. Poirot
discussed the results of using  various hole-filling techniques to
modify the input data for UNMIX calculations. In the end, Poirot

fell that simple replacement of nondetect values with zeros (or
some small constant) yielded the most consistent and imerpretable
UNMIX results.
    Poirot showed a series of slides lending support to those who
mistrust reported uncertainties and MDLs. For example, Ni and As
measurements at Lye Brook, VT, are totally uncorrelated and yet
the reported uncertainties exhibit a significant positive correlation
(top figure below). Also, he said, "although concentrations of Ni
and As are uncorrelated, their MDLs are highly correlated, both
as a function of three methods changes in different time periods,
and  also  within  each of three different  reporting  periods"
(bottom figure below).
                                  3 "Sources" of Arsenic and Nickel« Lye Brook, VT V81-5/99
                                  Time Series of Arsenic « Nickel MDL * Lye Brook, VT: 8/91-5/99
                           0.7 i
                               1   51  101 151 201  251  301 351  401  451  501 551  601  651
                                                       Sample* (1-693)

     Poirot went on to say that "(this is] possibly due to common
 interferences or instrumental drift, but not due to changing
 ambient concentrations. Generally, in most long-term measure-
 ment programs both ambient concentrations and detection limits
 are likely to decrease over time, creating the possibility of false
 positive correlations between source activity for some elements
 and lab activity for other elements." This latter point was
 elegantly demonstrated by a plot of same-day, above-MDL As
 concentrations at Acadia and Mt. Rainier IMPROVE sites. The
 measured concentrations exhibit no correlation (as expected
 given the continental distance between sites). However, same-
 day As MDLs for these  sites are correlated, generally due to
 "progress"' (improving detection limits over time) in the 10+
 year IMPROVE network.  Poirot also  provided evidence for
 "misquamified" MDLs for Al in IMPROVE data. He presented
 some encouraging results, which showed that despite wide
 differences in data preprocessing and model input, both UNMIX
 and PMF identified three common sources in an IMPROVE-like
 data set. However, artifacts associated with changes in Se MDLs
 due to a change from P1XE to XRF analysis during the sampling
 period clearly influenced  the UNMIX and PMF results in
 different ways.
    Poirot concluded  by saying. "Data Artifacts, including
 MDLs and uncertainties as reported by labs and/or as processed
by data analysts, can and do influence receptor model results."
Session 4E: Open Discussion
    There was further discussion of the MDLs. It was not
known whether the EPA PM. 5 Speciation Monitoring Network
will report the single method-based MDLs or the daily-varying
sample MDLs. Henry  reconsidered his distrust of reported
MDLs and uncertainties and found it to be justified. In situations
where one cannot afford to lose data due to nondetects, Henry
recommends just replacing the nondetects with zero or a small
    Important  MDL-relaied questions  include the following:
How are MDL and uncertainty values determined by analytical
laboratories? Do these reported values have the same meaning
at different labs or in different measurement programs? How
have analytical methods and the resultant data changed over the
course of a measurement program? And finally, what are the best
ways of processing th is information as input to different receptor
    In response to the question of whether or not to use mass as
a  fining  species, Henry and Hopke  expressed  different
philosophies. Henry likes to include the mass so that the total
mass is apportioned just like the  species mass. Hopke has
traditionally kept the mass separate and likes to use the results of
the mass regression analysis as an added check on the validity of
the model results.
    Henry expressed his concern that the errors reported in both
UNMIX and PMF have not been given adequate scrutiny. Hopke
believes that the error estimates in  PMF are almost certainly
    Several members of the audience commented on the dreaded
UNMIX message informing the user that there was "no feasible
solution" to a problem. Henry responded that rather than viewing
this as a bug or deficiency  in UNMIX, it should instead be
viewed as a valuable feature in that a bad solution is worse than
no solution.
    There was some discussion about dealing with outliers.
Henry relies heavily on UNMIX scatterplots to identify outliers.
He urged caution in eliminating suspected outliers because, if
real, they can provide very important information about source
compositions. Hopke typically does a principal components
analysis of the data and plots factor scores to identify outliers.
The "robust mode" option in PMF automatically down weights
outliers  (but does not eliminate them) so that they do not exert
too much influence. If the user knows that a certain sample is an
outlier (e.g., fireworks on the Fourth of July), then it is  best to
remove  that data point before performing UNMIX or PMF
    The interpretation of source profiles remains one  of the
biggest  challenges  in using these tools. Receptor  modeling
should not be done in a vacuum. Ideally, the modeler will have
intimate knowledge of the modeled airshed, or will work closely
with someone who  does. Rich Poirot  suggested creating an
informal, unofficial bulletin board or site where modelers could
share  source  profiles (accompanied  by  some descriptive
information) generated by UNMIX or PMF. Lewis would like
modelers to show their profiles in publications. With emphasis
on PM: „ there is likely to be  increased mass being apportioned
to regional sources, which are typically dominated by secondary
species.  It would be useful  to compile a library of regional
"fingerprints." Such a library could be helpful in proper  source
identification. There are some good tools such as residence time
analysis, back-trajectory analysis, and partial source contribution
function (PSCF) analysis for identifying and  quantifying
regional impacts. Hopke showed how PSCF was able to trace a
Ni-V factor in Vermont back to residual oil combustion in the
Eastern urban corridor.

                                                 Session 5
                                           16 February, a.m.
Session 5A: Application of PMF
in the Northern Great Lakes:
A Tale of Two Studies
Dr. Kurt Paterson, Michigan Technological University
(Full presentation is in Appendix 5A.)

    Dr. Paierson presented  an overview  of two  studies
conducted in the Northern Great Lakes in which PMF was
applied. The first study involved source apportionment of a
mixture of trace gases and paniculate matter in order to identify
the sources that influence air quality in the northern Great Lakes.
PMF extracted three sources identified by Paterson as biogenic
(defined by isoprene). local, and regional transport. Paterson
combined PMF with residence time analysis, met data analysis,
and time-series  analysis to confirm the identification of the
sources. In the second study PMF was used on panicle size
distribution data, not to apportion sources, but to extract distinct
factors that could reveal the dynamics of different panicle
modes. The original data comprised  100 size ranges from 5 nm
to 7.5 urn and 1046 half-hour samples. PMF collapsed this data
into six factors, which fell out into distinct panicle size ranges
and which exhibited different dynamic properties. Two factors,
for example, showed strong diurnal cycles, Two factors were
most influenced by long-range transport. And PM, < mass was
most influenced by panicles in the size range 220-800 nm. The
chemical composition data for these samples are now available
and Paterson will repeat these analyses, adding in the com-
position data and using both PMF and UNMIX.
Session SB: Discussion of FPEAK, Open
Discussions, and Workshop Conclusion
    Depending on the input data  set, PMF may generate
multiple solutions that are all equally valid within the rotational
ambiguity of the PMF model. Somehow the user must decide
which rotation is the best. FPEAK is one parameter available in
PMF that allows the  user to try "arious rotations. Positive
FPEAK values force the source composition matrix toward more
extremes (zeros for some species and large percentages for other
species) and  the  source  contribution matrix  toward less
extremes, while negative FPEAK values produce the opposite
effect. Eberly presented a simple example (seven samples, three
species, two sources) to show  the  effect  of FPEAK (see
Appendix 5B). PMF was executed with FPEAK values of-0.5,
0.0. and 0.5, and the resultant source composition and con-
tributions were presented. All three of these possible solutions
are consistent with the measurements recorded at the receptor,
that is, the masses balance. Examination of the solutions shows
that (1) for the negative FPEAK value, the source contributions
are the most extreme, including some days when one source is
not contributing, and (2) for the positive FPEAK value,  the
source compositions are the most extreme, including a species
whose proportions are 0.01  and 0.85.
    UNMIX was also run on the simple example and the results
were presented. UNMIX produces  only one solution and this
solution had compositions  and contributions similar to those
from PMF where the FPEAK value was -0.5. The reason for this
is that the UNMIX algorithm assumes there are days when each
source is not contributing to the receptor. That is, UNMIX seeks
sources for which there are some contributions near zero, and
this is similar to what PMF  does with negative FPEAK values.
    As mentioned, a requirement of UNMIX is that there must
be sampling days when each source disappears or is insig-
nificant. How does UNMIX handle a source like motor vehicles
in Washington, DC, which never turns off? Henry responded
that this  was the reason for putting the "tracer" option in
UNMIX. This option allows the user to select one species as a
tracer. This constrains the UNMIX solution by forcing all of the
tracer species mass into one source. For motor vehicles, Henry
recommended using CO as a tracer (not perfect, but usually good
enough). Without a tracer in this  case, UNMIX may not find a
feasible solution.
    Is there a rule of thumb for the number of samples needed
by UNMIX or PMF? It is really a signal-to-noise problem. PMF
has been applied to as few  as 40 samples, but typically there is
not enough variability present in so  few data points to be able to
pull out distinct factors. Recent work by John Ondov (PM2000

 Charleston Conference) has shown that by sampling with high
 time resolution (half-hour) one can dramatically improve the
 signal 'noise for sources with temporal variability. Henry offered
 the following rule of thumb for UNMIX: 200-300 samples may
 get  you five sources; 2000-3000 samples may be needed to
 extract 9-10 sources.
     How can receptor modelers take advantage of the EPA
 Speciation Monitoring Network now coming online? What tools
 are  available to  interpret  these data? Instead of modeling
 multiple species at a single site, one can model a single species
 across  multiple sites.  In this way, one can extract spatial
 concentration gradients, which, combined with wind-direction
 analysis, can identify source locations.  Alternatively, one can
 model multiple species at multiple sites using three-way factor
 analysis (Hopke et al.. 1998).
     A member of the audience pointed out the discrepancies in
 the UNMIX  and PMF solutions for the Phoenix data, most
 notably the mass apportioned to vegetative burning in the two
 models. Are such discrepancies the result of applying different
 models, or the result of different people interpreting the same
 information? Hopke responded by saying that the modeler needs
 to tap into the local expertise to help identify' important sources
 and to screen out unreasonable solutions. It is always good to
 come at a problem with as many tools as possible. If one can get
 similar  solutions using  both PMF and UNMIX, this adds
 confidence to the results.
    Henry proposed a strategy that combines UNMIX and PMF
 and  should yield defensible solutions.  In this combined ap-
 proach, the modeler might  start with UNMIX to estimate the
number of factors  and to get good starting source profiles.
 UNMIX profiles could be used as  starting profiles for PMF,
 since PMF is particularly good at finding smaller sources and
 including additional species. (This will shorten the PMF analysis
 since the model does not have to start with random profiles.)
 Applying other information such as wind-direction plots, one
 can probably come up with 10 or more sources.  The abi'ity to
 look at residuals in PMF can be very helpful as a quality' check
at the end of the modeling process.
    Kurt Paterson suggested that it would be very useful to have
thorough training tutorials for both PMF and UNMIX showing
detailed applications of the tools in actual case studies.
    There was a broad discussion regarding the roles of regional
planning bodies and state regulatory agencies in dealing with
 compliance issues. Within a few years, regulatory agencies will
 need to address reductions at both the regional and local levels,
 with the regional planning bodies probably taking the lead. Will
 the state and regional  agencies have the resources and the
 expertise to utilize the latest modeling tools? How can PMF and
 UNMIX be used to separate the regional from the local sources?
 Can the IMPROVE and Speciation networks be combined in
 some way to help separate regional sources from local sources?
 Henry responded that there presently exist  a handful of good
 tools for dealing with regional sources. The  challenge is for
someone to put these tools together and make people aware that
they exist. Perhaps the EPA regional offices can play a role in
disseminating information about these tools or providing training
to state agencies.

PMF Publications
Anttila, P., P. Paatero. U. Tapper, and 0. Jarvinen (1995)
    Application of Positive Matrix Factorization to Source
    Apportionment:  Results of a Study of Bulk Deposition
    Chemistry in Finland. Aimos. Environ. 29:1705-1718.

Chueinta. W., P.K. Hopke. and P. Paatero (2000) Investigation
    of Sources of Atmospheric Aerosol Urban and Suburban
    Residential Areas in Thailand by Positive Matrix Factori-
    zation. Atmos. Environ. In press.

Garrido Frenich. A.. M. MartinezGalera, J.L. Martinez Vidal,
    D.L. Massan. J.R. Torres-Lapasio. K. De Braekeleer, J.H.
    Wang, and P.K.  Hopke (2000)  Resolution of  Multi-
    component Peaks  by OPA, PMF and ALS,  Anal. Chim.

Hopke, P.K.,P. Paatero. H. Jia, R.T. Ross, and R.A. Harshman
    (1998) Three-Way (PARAFAC) Factor Analysis: Exami-
    nation  and Comparison of  Alternative  Computational
    Methods  as Applied to Ill-Conditioned Data,  Chemom.
    Intell. Lab. Syst. 43:25-42.

Hopke. P.K., Y. Xie, and P. Paatero (1999) Mixed Multiway
    Analysis  of Airborne  Panicle  Composition  Data, J.
    Chemom.  13:343-352.

Hopke, P.K. (2000) A Guide to Positive Matrix Factorization,
    report prepared for the Office of Air Quality Planning anc*
    Standards, U.S. Environmental Protection  Agency, under
    Contract No. 9D-1808-NTEX, January 2000.

Hopke,  P.K.  (2000)  Application of Source Apportionment
    Methods  to the  State Implementation Planning Process,
    report prepared for the Office of Air Quality Planning and
    Standards, U.S. Environmental Protection Agency, under
    Contract No. 9D-1808-NTEX, January 2000.
Huang. S., K.A. Rahn, and R. Arimoto (1999) Testing and
    Optimizing Two Factor-Analysis Techniques on Aerosol at
    Narragansett, Rhode Island,X/mos. Environ. 33:2169-2185.

Juntto, S., and P. Paatero (1994) Analysis of Daily Precipitation
    Data by Positive Matrix Factorization, Environmetrics

Lee, E., C.K.  Chan, and P.  Paatero (1999) Application of
    Positive Matrix Factorization in Source Apportionment of
    Paniculate Pollutants in  Hong Kong, Atmos. Environ.

Paatero, P. (1997) Least Squares Formulation of Robust, Non-
    Negative  Factor Analysis,  Chemom. Intell  Lab.  Syst.

Paatero, P. (1999) The Multilinear Engine—A Table-Driven
    Least Squares Program for Solving Multilinear Problems,
    Including the  n-way Parallel Factor Analysis Model, J.
    Computational and Graphical Stat. 8:1-35.

Paatero, P., and U. Tapper (1993) Analysis of Different Modes
    of Factor Analysis as Least Squares Fit Problems, Chemom.
    Intell. Lab. Syst. 18:183-194.

Paatero, P., and U. Tapper (1994) Positive Matrix Factorization:
    A Non-negative Factor Model with Optimal Utilization of
    Error Estimates of Data Values, Environmetrics 5:\ 11-126.

Paterson, K.G., J.L. Sagady, D-L. Hooper, S.B. Bertman, M.A.
    Carroll, and P.B. Shepson (1999) Analysis of Air Quality
    Data Using Positive Matrix Factorization, Environ. Sci.
    Technol. 33:635-641.

Polissar, A.V., P.K. Hopke, W.C. Malm, J.F. Sisler (1996) The
    Ratio of Aerosol Optical Absorption Coefficients to Sulfur
    Concentrations, as an Indicator of Smoke from Forest Fires
    when Sampling in Polar Regions, Atmos.  Environ. 30:

 Polissar, A.V., P.K. Hopke, W.C. Malm, J.F. SisJer (1998)
    Atmospheric Aerosol over Alaska: 2. Elemental Com-
    position and Sources../. Ceophys. Res. 103:19,045-19,057.

 Polissar, A.V., P.K. Hopke, P. Paatero. YJ. Kaufman, O.K.
    Hall. B.A. Bodhaine, E.G. Dutton, and J.M. Harris (1999)
    The Aerosol at Barrow, Alaska:  Long-Term Trends and
    Source Locations, Atmos. Environ. 33:2441-2458.

 Ramadan. Z., X.-H. Song. P.K. Hopke (2000) Identificaton of
    Sources of Phoenix Aerosol by Positive Matrix Factori-
    zation. JA WMA. In press.

 Xie. Y.-L.. P.K. Hopke. and P. Paatero (1998) Positive Matrix
    Factorization  Applied to Curve  Resolution Problem, J.
    Chemom. 12:357-364.

 Xie, Y.-L., P.K. Hopke. P. Paatero, L.A. Barrie, and S.-M. Li
    (I999a) Identification  of Source Nature and Seasonal
    Variations of Arctic Aerosol by the Multilinear Engine,
    Atmos. Environ. 33:2549-2562.

Xie. Y.-L., P.K. Hopke, and P. Paatero (1999c) Calibration
    Transfer as a Data Reconstruction Problem, Anal.  Chim.
    Acta 384 -.193-205.

Xie. Y. L., P. Hopke.  P. Paatero, L.A. Barrie, and S.M. Li
    (1999b)  Identification of Source Nature and Seasonal
    Variations of Arctic Aerosol by Positix-e Matrix Factori-
    zation, J. Atmos. Sci. 56:249-260.

Yakovleva, £., P.K. Hopke, and L. Wallace (1999) Receptor
    Modeling  Assessment of PTEAM Data, Em-iron. Sci.
    Technol. 33:3645-3652.
Henry, R.C. (1997) History and Fundamentals of Muhivariate
    Air Quality Receptor Models, Chemom. Intell. Lab. Syst

Henry, R.C., and C. Spiegelman (1997) Reported Emissions of
    Volatile Organic Compounds  are not Consistent with
    Observations, Proc. Nat. Acad Sci. 94:6596-6599.

Henry, R.C., E.S. Park, and C.H. Spiegelman (1999) Comparing
    a New Algorithm with the Classic Methods for Estimating
    the  Number of Factors,  Chemom. Intell.  Lab.  Syst.  48:

Kim, B.-M., and R.C. Henry (1999) Extension of Self-Modeling
    Curve  Resolution  to Mixtures of More  Than  Three
    Components. Pan  2: Finding  the  Complete  Solution,
    Chemom. Intell. Lab. Syst. 49:67-77.

Kim, B.-M., and R.C. Henry (2000) Application of the SAFER
    Model  to  Los Angeles PM10 Data, Atmos.  Environ.

Kim, B.-M., and R.C. Henry (2000) Extension of Self-Modeling
    Curve Resolution  to Mixtures of More  Than  Three
    Components. Part 3, Chemom. Jntell. Lab. Sysl. Submitted.

Park. E.S., C. Spiegelman, and  R.C. Henry (2000) Bilinear
    Estimation of Pollution Source Profiles and Amounts by-
    Using Receptor Models,  Communications in Statistics,
    Simulation & Computation, 29(3). In press.
UNMIX and SAFER (the Parent
Model of UNMIX) Publications

Henry, R.C. and B.- M. Kim (1989) A Factor Analysis Model
    with Explicit Physical Constraints, Transactions AirPollut.
    Control Assoc. 14:214-225.

Henry, R.C.,and B.-M. Kim (1990) Extension of Self-Modeling
    Curve Resolution to Mixtures of More Than Three Com-
    ponents.  Part 1: Finding the  Basic  Feasible  Region,
    Chemom. Intell. Lab. Syst. 8:205-216.

Henry, R.C., C.W. Lewis, J.F. Collins (1994) Vehicle-Related
    Hydrocarbon Source Composition from Ambient Data: The
    GRACE/SAFER Method, Environ. Sci.  Technol 28:823-

UNMIX/PMF Receptor Modeling Workshop Attendees
             14-16 February 2000
              U.S. EPA, RTP, NC
Charles Leu is
Philip K. Hopke
Ron Henry
John Bachmann
Scon Kegler
Mark Hubble
Lara Awry
Donna Kenski
Tom Pace
Lam Cox
N. Dean Smith
Basil Couum
Alan Vene
Anne Rea
Ten Conner
Clarkson University
University of Southern
California, Los Angeles
Arizona DEQ
2 1 3-740-0596
mhl Sev.suie.az.us
autn .laragepa gov
'(tnski.donna 2 epa.gov
pace .tomg epa.gov
Mailing Address
NERL (MD-47)
RTP, NC 277 11
Clarkson University
Ptisdam, NY 1 3655-58 !"

RTP, NC 277 11
RTP, NC 27711
Arizona DEQ
3033 N. Central Ave.
Phoenix. AZ 85012
RTP, NC 277 11
EPA Region 5
77 W. Jackson Blvd.
Chicago, JL 60604
RTP, NC 277 11
NERL (MD-75)
RTP, NC 277 11
RTP, NC 277 11
NERL (MD-14)
RTP, NC 277 11
RTP, NC 277 11
RTP, NC 277 11
NERL (MD-46)
RTP, NC 2771)

Tom Braverman
Allan Marcus
Shao-Hang Chu
Ned M ever
Peter Frechtel
Eugene Kim
Tom Rosendahl
Cynthia Howard
Barbara Parzygnai
Bob Willis
Shaibal Mukerjee
John LangsiatT
Melissa Gonzales
Tom Coulter
Rich Poirot
Sieve Fudge
University of Washington
f.S. EPA
ManTech Environmental
New York University


parzygnat.barbarag epa.gox
mukerjee.shaibal gepa.gov
Mailing Address
NER1. (MD-14)
RTP.NC 277 11
NERL (MD-52)

OAQPS (MD- 15)
12201 Sunrise Valley Dr.
555 National Center
Reston. VA 20192
RIP, NC 277 11
ManTech Environmental
P.O. Box 123 13
RTP.NC 27709
New York University, Env. Med.
57 Old Forge Rd.
Tuxedo, NY 10987
RTP.NC 277 11
Chapel Hill, NC 275 14
RTP.NC 27711
RTP.NC 277 11
103 South Main St.
Waterbury, VT 05671
1129 Weaver DairvRd.
Chapel Hill, NC 27514

Shell) Eberly
Jonj; Hoon Lee
Barbara Turpin
Kurt Paierson
Rutgers University
Rutgers University
Michigan Tech University
9) 9-54 Ml 28
Mailing Address
RTF, NC 277 II
Rutgers University
14 College Farm Rd.
New Brunswick, NJ 08901
Rutgers University
14 College Farm Rd.
New Brunswick, NJ 08901
Michigan Tech University
Dspi. Civil & Env. Engineering
1400 lownsend Drive
Houghtcn, MI 49931