. ."aww
, -

.... 1 .«u ..
301th?
11....

I .4 t‘.’
7......» .
.

 

’ I&W
l. f. v 1 .
3, . .5 E...

..1.r 3.534% . . 1.. «$4wa ..
. . mmw. vim? a; 3%

2. 33...,

. a .S.

 

 

 

Q LIBRARY
, W. Michigan State
‘20 " University

This is to certify that the
thesis entitled

ASSOCIATION AND DISCRIMINATION OF DIESEL FUELS
USING CHEMOMETRIC PROCEDURES FOR FORENSIC
ARSON INVESTIGATIONS

presented by

LUCAS JAMES MARSHALL

has been accepted towards fulfillment
of the requirements for the

MS. degree in CRIMINAL JUSTICE

 

 

 

 

Major Proiessor’s Signature

20 TH HA7 1406’

Date

MSU is an affirmative-action. equal-opportunity employer

-.—.-.-.-p--.-.-.---1—.-.---.---.—-------.-—.-.-.-.-.------u—.-.—1-.-.-.-‘-.---v-.--v—-—.—-—

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 K:lProilAcc&Pres/ClRC/DaleDue.indd

ASSOCIATION AND DISCRIMINATION OF DIESEL FUELS USING
CHEMOMETRIC PROCEDURES FOR FORENSIC ARSON INVESTIGATIONS

By

Lucas James Marshall

A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

MASTER OF SCIENCE
School of Criminal Justice

2008

ABSTRACT

ASSOCIATION AND DISCRIMINATION OF DIESEL FUELS USING
CHEMOMETRIC PROCEDURES FOR FORENSIC ARSON INVESTIGATIONS

By
Lucas James Marshall

The identification of an ignitable liquid in fire debris is indicative of an
intentional fire and hence, is significant evidence in arson investigations. Currently, gas
chromatography-mass spectrometry (GC-MS) is the conventional analytical technique
used for the identification of ignitable liquids through chromatographic pattern matching.
Chemometric procedures such as Pearson Product Moment Correlation (PPMC)
coefficients and Principal Components Analysis (PCA) provide a more objective method
to statistically associate and discriminate burned and unburned diesel fuels based on
chemical composition in both the total ion chromatograrn (TIC), as well as extracted ion
profiles (EIP) corresponding to characteristic compound classes in the diesel samples.
Data pre-treatment options, such as retention time alignment and area normalization,
were also investigated in order to determine their effects on the Chemometric results.

The association and discrimination of burned and unburned diesels was also
examined. Diesels were spiked onto different matrices commonly found in the home
(cotton cloth, magazine, and carpet) and burned in order to simulate arson conditions.
The burned diesels were extracted using solvent extraction procedure and analyzed by
GC-MS. The data generated from the burnings was compiled into the same set as the
data generated from the neat diesels so that PPMC and PCA could be applied to the entire
data set. The potential for the association and discrimination of the burned diesels using

these procedures was investigated.

ACKNOWLEDGMENTS

I would first and foremost like to thank my family and my fiance Molly for
always being there to support me. It hasn’t been easy, but they have continued to
encourage me throughout the process, and without them, I would have never made it this
far.

I would also like to thank Dr. Ruth Smith for the opportunity to work with her as
a forensic chemistry student. She not only is a top-notch scientist, but she has also
become my friend. I know that I am lucky to have had daily contact with my advisor, but
again, without her, none of this work would be possible. She has constantly motivated
me to take the project to the next level, while along the way challenging me to think and
do for myself.

I would also like to thank Dr. Victoria McGuffin for her presence and advice
throughout this collaborative project. She always seemed to have the answers when I
needed help, and without her I would not be where I am as a confident scientist.

I would also like to thank Dr. Vince Hoffman for taking time out of his busy
schedule to sit on my thesis committee, and Dr. Kathy Severin for instrument time and
assistance.

Finally, I would like to thank the forensic science program at MSU, as well as all
the people in it. I have made some wonderful friends here, especially in the forensic
chemistry section and the McGuffin group. To Dr. Melissa Meaney, Amber Hupp, Sarah
Meisinger, Lisa LaGoo, Ruth Udey, Jamie Baemcopf, Patty Joiner, and John McIlroy, I

thank you. You guys have made my time here so much fun, and I will never forget you!

iii

TABLE OF CONTENTS

 

 

 

 

 

 

List of Tables - - - - _ - - .......... vi
List of Figures -- - - - - - -- - - ............ -- - _---ix
Chapter 1 - Introduction - - .......... . - 1
1.1 Arson .................................................................................................................... 1
1.2 Ignitable Liquids ................................................................................................... 1
1.3 Analysis of Fire Debris ......................................................................................... 2
1.4 Current Literature Review .................................................................................... 7
1.4.] Statistical and Chemometric Analysis of Petroleum Distillates ........................... 8
1.4.2 Chromatographic Pre-processing ....................................................................... 1]
1.4.3 Burn Conditions and Matrix Interferences ......................................................... 12
1.5 Research Objectives ........................................................................................... 14
Chapter 2 - Analytical and Chemometric Theory -- - - ...... . - 17
2.1 Gas Chromatography-Mass Spectrometry ......................................................... 17
2.2 Data Pre-treatment .............................................................................................. 21
2.2.1 Retention Time Alignment ................................................................................. 22
2.2.2 Area Normalization ............................................................................................ 26
2.2.3 Mean-Centering .................................................................................................. 27
2.3 Pearson Product Moment Correlation Coefficients ............................................ 30
2.4 Principal Components Analysis ......................................................................... 32
Chapter 3 - Association and Discrimination of Neat Diesels Using PPMC
Coefficients and PCA - - - - - - ..... 35
3. 1 Introduction ........................................................................................................ 3 5
3.2 Sample Collection and Analysis ......................................................................... 36
3.3 Data Pre-treatment .............................................................................................. 39
3.4 Data Analysis ...................................................................................................... 40
3.5 Results and Discussion ....................................................................................... 42
3.5.1 Initial Data Set .................................................................................................... 42
3.5.2 Data Set using Timed Injection .......................................................................... 48
3.5.3 Improving Abundance Levels of TIC and EICs ................................................. 51
3.5.4 Final Neat Data Set ............................................................................................. 54
3.5.4.1 RSD Values ........................................................................................................ 54
3.5.4.2 PPMC Coefficients ............................................................................................. 56
3.5.4.3 PCA .................................................................................................................... 59
Chapter 4 - Association and Discrimination of Burned Diesels Extracted from

Fire Debris Using PPMC Coefficients and PCA - - - -73
4.1 Introduction ........................................................................................................ 73
4.2 Analysis of New Set of Diesels .......................................................................... 75

iv

4.2.1 Procedure ............................................................................................................ 75

 

 

 

 

4.2.2 Results and Discussion ....................................................................................... 78
4.3 Efficiency of Solvent Extraction Procedure ....................................................... 92
4.3.1 Procedure ............................................................................................................ 92
4.3.2 Results and Discussion ....................................................................................... 94
4.4 Analysis of Burned Diesels ................................................................................ 99
4.4.1 Procedure ............................................................................................................ 99
4.4.2 Results and Discussion ..................................................................................... 102
4.4.2.1 Assessment of Potential Interferences from Unburned and Burned Matrices . 102
4.4.2.2 Solvent Extraction of Diesels from Unburned and Burned Matrices ............... 105
4.4.2.3 PPMC Coefficients of Neat and Burned Diesels .............................................. 108
4.4.2.4 PCA of Neat and Burned Diesels ...................................................................... 1 l 1
4.4.2.5 Identification of Blind Diesels .......................................................................... 124
Chapter 5 - Conclusions and Future Work - - - -- -- ........... 127
5.1 Conclusions ...................................................................................................... 1 27
5.1.1 Association and Discrimination of Neat Diesels Using PPMC Coefficients

and PCA ........................................................................................................... 127
5.1.2 Association and Discrimination of Burned Diesels Extracted from Fire

Debris Using PPMC Coefficients and PCA ..................................................... 129
5.2 Future Work ..................................................................................................... 130
Appendix A - ASTM Classification Scheme ............................................................. 135
Appendix B - PPMC Coefficients for the TIC and EIPs of Ten Neat Diesels
Analyzed in Triplicate -- - _ - 138
Appendix C - PPMC Coefficients for the TIC and EIPs of Five Neat Diesels
Analyzed in Triplicate -- - - - 145
Appendix D- PPMC Coefficients for the TIC and EIPs of Compiled Neat and
Burned Diesels - _ - - - - - 152
References - -- - - - _ 158

 

LIST OF TABLES

Table 1.1 Major Ions Included in Extracted Ion Profiles for Ignitable Liquids [12] ........ 6

Table 3.1 Diesel Collection Log ............... - ...................................................................... 37
Table 3.2 RSD Values for Eight Peaks Across the Retention Time Range for Ten

Diesel Samples Analyzed in Triplicate ........................................................................... 43
Table 3.3 PPMC Coefficients for Replicate Sets of Ten Diesel Samples Analyzed in
Triplicate ......................................................................................................................... 45
Table 3.4 RSD Values for Eight Peaks Across the Retention Time Range for Ten

Diesel Samples Analyzed in Triplicate Using a Timed Injection Method ..................... 49
Table 3.5 PPMC Coefficients for Replicate Sets of Five Diesel Samples Analyzed

in Triplicate Using a Timed Injection Method ............................................................... 50
Table 3.6 PPMC Coefficients for Individual m/z Values and Their Summed EIPs

for the Alkane, Aromatic, and Indane Profiles ............................................................... 53
Table 3.7 RSD Values for Eight Peaks Across the Retention Time Range for Final

Neat Diesel Data Set Analyzed in Triplicate Using a Timed Injection Method ............. 55
Table 3.8 PPMC Coefficients for TIC and EIPs for Replicate Sets of Ten Diesel
Samples (Final Data Set) ................................................................................................ 57
Table 4.1 Diesel Collection Log ..................................................................................... 76

Table 4.2 RSD Values for Eight Peaks Across the Retention Time Range for Five
Diesel Samples Analyzed in Triplicate ........................................................................... 79

Table 4.3 PPMC Coefficients for the TIC of Replicate Sets of Five Diesel Samples
Analyzed in Triplicate ..................................................................................................... 80

Table 4.4 PPMC Coefficients for TIC and EIPs for Replicate Sets of Five Diesel
Samples ........................................................................................................................... 82

Table 4.5 Extraction Recoveries from Cloth for Each Calibration Standard
(1,3,5-TMB = 1,3,5-trimethylbenzene, C10 = decane, C12 = dodecane, C14 =
tetradecane) ..................................................................................................................... 95

Table 4.6 Extraction Recoveries from Magazine for Each Calibration Standard

vi

(1,3,5-TMB = 1,3,5-trimethylbenzene, C10 = decane, C12 = dodecane, C14 =
tetradecane) ..................................................................................................................... 96

Table 4.7 Extraction Recoveries from Carpet for Each Calibration Standard
(1,3,5-TMB = 1,3,5-trimethylbenzene, C10 = decane, C12 = dodecane, C14 =

tetradecane) ..................................................................................................................... 98
Table A.l ASTM Ignitable Liquid Classification Scheme [4,5] .................................. 135
Table A2 Chromatographic and Mass Spectral Characteristics of ASTM Ignitable

Liquid Classes [4,5] ...................................................................................................... 136
Table 8.1 PPMC Coefficients for the TIC of Triplicate Analyses of Diesels 1-10 ...... 138

Table B.2 PPMC Coefficients for the Alkane EIP of Triplicate Analyses of
Diesels 1-10 .................................................................................................................. 139

Table 3.3 PPMC Coefficients for the Aromatic EIP of Triplicate Analyses of
Diesels 1-10 .................................................................................................................. 140

Table B.4 PPMC Coefficients for the Indane EIP of Triplicate Analyses of
Diesels 1-10 .................................................................................................................. 141

Table B.5 PPMC Coefficients for the OCP EIP of Triplicate Analyses of
Diesels 1-10 .................................................................................................................. 142

Table B.6 PPMC Coefficients for the PNA EIP of Triplicate Analyses of
Diesels 1-10 .................................................................................................................. 143

Table C.1 PPMC Coefficients for the TIC of Triplicate Analyses of Diesels 21-25....145

Table C.2 PPMC Coefficients for the Alkane EIP of Triplicate Analyses of
Diesels 21-25 ................................................................................................................ 146

Table C.3 PPMC Coefficients for the Aromatic EIP of Triplicate Analyses of
Diesels 21-25 ................................................................................................................ 147

Table C.4 PPMC Coefficients for the Indane EIP of Triplicate Analyses of
Diesels 21 -25 ................................................................................................................ 148

Table C.5 PPMC Coefficients for the OCP EIP of Triplicate Analyses of
Diesels 21-25 ................................................................................................................ 149

Table C.6 PPMC Coefficients for the PNA EIP of Triplicate Analyses of
Diesels 21-25 ................................................................................................................ 149

vii

Table D.1 PPMC Coefficients for the TIC of Neat and Burned Diesels (Cloth A-C =
Replicates of Burned Diesel 21 from Cloth, Mag A-C = Replicates of Burned

Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,

B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet) ............. 152

Table D2 PPMC Coefficients for the Alkane EIP of Neat and Burned Diesels (Cloth
A-C = Replicates of Burned Diesel 21 from Cloth, Mag A-C = Replicates of Burned
Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,

B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet) ............. 153

Table D.3 PPMC Coefficients for the Aromatic EIP of Neat and Burned Diesels (Cloth
A-C = Replicates of Burned Diesel 21 from Cloth, Mag A-C = Replicates of Burned
Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,

B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet) ............. 154

Table D4 PPMC Coefficients for the Indane EIP of Neat and Burned Diesels (Cloth
A-C = Replicates of Burned Diesel 21 from Cloth, Mag A-C = Replicates of Burned
Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,

B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet) ............. 155

Table D5 PPMC Coefficients for the OCP EIP of Neat and Burned Diesels (Cloth

A-C = Replicates of Burned Diesel 21 from Cloth, Mag A-C = Replicates of Burned
Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,

B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet) ............. 156

Table D6 PPMC Coefficients for the PNA EIP of Neat and Burned Diesels (Cloth

A-C = Replicates of Burned Diesel 21 from Cloth, Mag A-C = Replicates of Burned
Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,

B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet) ............. 157

"Images in this thesis are presented in color.

viii

LIST OF FIGURES

Figure 2.1 The Effects of a Retention Time Alignment Algorithm on the Decane
Peak of Diesels 1-5 ......................................................................................................... 23

Figure 2.2 The Effects of Area Normalization on the Decane Peak of Diesels 1-5 ....... 28
Figure 2.3 The Application of Mean-Centering to the Decane Peak of Diesels 1-5 ...... 29

Figure 3.1 Distribution of Select Peak Areas Among Replicate Injections of
Diesel 9 ........................................................................................................................... 47

Figure 3.2 Scores Plot - TIC (a) PCl v. PC2 and (b) PCl v. PC2 v. PC3 ...................... 60
Figure 3.3 Scores Plot - Alkane EIP (a) PCl v. PC2 and (b) PC 1 v. PC2 v. PC3 .......... 62
Figure 3.4 Scores Plot - Indane EIP (a) PCI v. PC2 and (b) PCl v. PC2 v. PC3 .......... 63
Figure 3.5 Scores Plot - PNA EIP (a) PC] v. PC2 and (b) PCl v. PC2 v. PC3 ............. 64

Figure 3.6 Scores Plot - Aromatic EIP (a) PC] v. PC2 and (b) PC 1 v. PC2 v. PC3 ...... 66

Figure 3.7 Scores Plot - OCP EIP (a) PC] v. PC2 and (b) PCl v. PC2 v. PC3 .............. 67
Figure 3.8 Loadings Plots of First Two Eigenvectors from PCA of the TICs of Ten

Diesels Analyzed in Triplicate ........................................................................................ 69
Figure 3.9 Loadings Plots of First Two Eigenvectors from PCA of the Alkane ElPs

of Ten Diesels Analyzed in Triplicate ............................................................................ 71
Figure 3.10 Misaligned Peak in the Alkane EIPs of Diesel 1 Replicates ....................... 72
Figure 4.1 Scores Plot for the TIC of the Neat Diesels .................................................. 84
Figure 4.2 Scores Plot for the Alkane EIP of the Neat Diesels ...................................... 85
Figure 4.3 Scores Plot for the Aromatic EIP of the Neat Diesels ................................... 87
Figure 4.4 Scores Plot for the Indane EIP of the Neat Diesels ....................................... 88
Figure 4.5 Scores Plot for the OCP EIP of the Neat Diesels .......................................... 89
Figure 4.6 Scores Plot for the PNA EIP of the Neat Diesels .......................................... 91

ix

Figure 4.7 Calibration Curve for the Neat Calibration Standards for Cloth Matrix

Recovery Study ............................................................................................................... 95
Figure 4.8 Calibration Curve for the Neat Calibration Standards for Magazine

Matrix Recovery Study ................................................................................................... 96
Figure 4.9 Calibration Curve for the Neat Calibration Standards for Carpet Matrix
Recovery Study ............................................................................................................... 98
Figure 4.10 Representative Chromatograms of Unburned Matrices: (a) Cloth, (b)
Magazine, and (c) Carpet .............................................................................................. 103
Figure 4.11 Representative Chromatograms of Burned Matrices: (a) Cloth, (b)
Magazine, and (c) Carpet .............................................................................................. 104
Figure 4.12 Representative Chromatograms for Unburned Matrices Spiked with

Diesel 21: (a) Cloth, (b) Magazine, and (c) Carpet ....................................................... 106
Figure 4.13 Representative Chromatograms for Burned Matrices Spiked with

Diesel 21: (a) Cloth, (b) Magazine, and (c) Carpet ....................................................... 107
Figure 4.14 Representative Chromatograms for Matrices Spiked with Diesel 21

and then Burned: (a) Cloth, (b) Magazine, and (0) Carpet ........................................... 109
Figure 4.14 Scores Plot for the TIC of the Neat and Burned Diesels ........................... 1 13

Figure 4.15 Representative PCA Scores Plots for (a) TIC and (b) Alkane EIP for
the Burned Data Scored with the Eigenvectors from the Neat Data ............................. 1 15

Figure 4.16 Scores Plot for the TIC of the Neat Diesels and the Projected Burned
Diesels After a 10X Correction ..................................................................................... l 17

Figure 4.17 Scores Plot for the Alkane EIP of the Neat Diesels and the Projected
Burned Diesels After a 10X Correction ........................................................................ l 19

Figure 4.18 Scores Plot for the Aromatic EIP of the Neat Diesels and the Projected
Burned Diesels After a 10X Correction ........................................................................ 120

Figure 4.19 Scores Plot for the Indane EIP of the Neat Diesels and the Projected
Burned Diesels After a 10X Correction ........................................................................ 122

Figure 4.20 Scores Plot for the OCP EIP of the Neat Diesels and the Projected
Burned Diesels After a 10X Correction ........................................................................ 123

Figure 4.21 Scores Plot for the PNA EIP of the Neat Diesels and the Projected
Burned Diesels After a 10X Correction ........................................................................ 125

Chapter 1
Introduction

1.1 Arson

Arson is defined as the deliberate setting of a fire with the malicious intent to
cause damage [1]. The United States Fire Administration (USE A) estimates that more
than 30,000 intentional structural fires were set in 2006, which resulted in more than 300
civilian deaths [2]. These intentional fires were reported to have caused approximately
$755 million in property damage. An estimated 20,000 vehicle fires were also set. which
caused an additional $134 million in damages. Arson is a destructive and expensive
crime, yet only a small number of cases ever result in an arrest or a conviction. The
Federal Bureau of Investigation (FBI) reports that in 2002 only 16.5% of arson cases
were closed [3]. This low rate of closure demonstrates the difficulties encountered in
arson investigations and fire debris analysis, and indicates that research into the
improvement of the methods in place is essential for advancements in this field.
1.2 Ignitable Liquids

Ignitable liquids are commonly used as accelerants in arson cases. The presence
of an ignitable liquid, which is used to increase the rate and spread of a fire, may indicate
that the fire was intentionally set. Common ignitable liquids frequently used as
accelerants include petroleum distillates, such as gasoline, kerosene, and diesel fuel, and
other hydrocarbon products, such as aromatic products and oxygenated solvents.

The American Society of Testing and Materials (ASTM) has developed a
classification scheme for ignitable liquids based on chemical composition [4,5]. Nine

classes have been differentiated for the characterization of ignitable liquids based on

composition: (1) gasoline-all brands, including gasohol, (2) petroleum distillates, (3)
isoparaflinic products, (4) aromatic products, (5) naphthenic paraffinic products, (6)
normal alkanes products, (7) de-aromatized distillates, (8) oxygenated solvents, and (9)
others-miscellaneous. These classes are further divided into three subclasses based on
the range of carbon content present in the liquid. The first subclass covers the light
carbon range of C4-C9, with no components present above C”. The medium carbon
range includes C3-C13 compounds, with none present below C7 or above C14. The heavy
carbon range includes compounds at C9 and above, with the typically observed range of
C9-C23. It should be noted that these demarcations are not rigid, and that often samples
are reported as “light to medium” or “medium to heavy” when the observed carbon range
fits both profiles. The details of the ASTM classification scheme, including examples of
common ignitable liquids that are representative of each class, can be found in Appendix
A [4,5].
1.3 Analysis of Fire Debris

At a fire scene, ignitable liquid residues (ILRs) are most likely to be found in the
area where the fire started, known as the origin. ILRs are also commonly found in pour
patterns known as trailers, which occur when an ignitable liquid is poured in a constant
stream from room to room in order to force the fire to spread throughout the entire
structure. For evidence collection, the fire debris most likely to contain ILRs are those
with porous surfaces, because the liquid may soak into them and thus be protected to
some extent from the heat of the fire. Similar debris types that are not suspected to
contain ILR are collected as control samples, which are necessary to prove that, if an ILR

is detected, it is not a natural component or pyrolysis product of the debris matrix itself.

The laboratory analysis of fire debris encompasses the identification of both
suspected neat liquids and ILRs extracted from burned and unburned substrates in the
debris. The current protocols in place in forensic laboratories follow the ASTM
standards that have been established and are maintained by the technical committee
E3001, which is the criminalistics subcommittee of the committee on forensic sciences
[6]. ASTM standards are in place for both the analysis of ignitable liquids and the
extraction of ILRs from debris.

In order for a suspected neat ignitable liquid to be analyzed, it is simply diluted in
an acceptable solvent and then analyzed by gas chromatography-mass spectrometry (GC-
MS). The analysis of fire debris for the presence of ILRs is similar to the analysis of a
neat liquid. The difference arises in the need to extract the ILR from the debris in order
for it to be analyzed by GC-MS. A variety of extraction techniques can be used, though
each has its advantages and disadvantages. Five common extractions are listed by ASTM
for the separation of ILRs from debris: steam distillation, solvent extraction, passive
headspace concentration with activated charcoal, dynamic headspace concentration, and
passive headspace concentration with solid-phase microextraction (SPME) [7-11].
Currently, the most frequently used extraction technique in crime labs is passive
headspace concentration with activated charcoal strips, though solvent extractions are
still valuable in some cases.

The ASTM standard E1412-07 details the procedure for the passive headspace
extraction of ILRs from fire debris using activated charcoal [9]. The activated charcoal is
suspended inside the submitted evidence container, and the container resealed. The

container is placed in an oven and heated to a temperature of 50-80 °C for 2-24 hours.

The temperature and time heated is sample dependent, as higher temperatures and longer
durations may be necessary to promote sufficient adsorption of less volatile compounds
in the debris. After the adsorption step, the activated charcoal is washed with 50-1000 pL
of an appropriate elution solvent, typically carbon disulfide, n-pentane, or diethyl ether.
The eluate is collected and then analyzed by GC-MS. This procedure is capable of
extracting ILRs across a wide range of concentrations with an extremely high level of
sensitivity. It is considered a nondestructive technique for the extraction of ILRs from
fire debris.

Solvent extractions, though less frequently used, can provide significant
additional information in analyses when activated charcoal extractions are limited by
volatility. Solvent extractions offer a more complete extraction that is representative of
the entire range of compounds in a sample, whereas activated charcoal extractions are
biased against less volatile compounds and those that are not selectively adsorbed onto
the carbon strip. The ASTM standard E1386-00(2005) details the specifics of solvent
extractions [8]. A suitable organic solvent, typically carbon disulfide, pentane, petroleum
ether, or diethyl ether, is added to the debris. The solvent-debris mixture is thoroughly
agitated for approximately one minute. The solvent is poured off and then filtered if
particulates remain in the decantate. The extract is evaporated to approximately 1 mL
and then analyzed by GC-MS. This technique is very sensitive, and is especially useful
for the extraction of nonporous surfaces or when a small amount of sample needs to be
extracted. The major disadvantage of solvent extraction is the concurrent extraction of
interferences from the burned matrix. It must also be noted that some of the lighter

classes of ignitable liquids may be lost in the evaporation phase of the procedure. The

ASTM standard suggests that, because solvent extraction is a destructive technique. it
only be used in tandem with another extraction procedure. For example, if the results of
an extraction with a charcoal strip indicate a heavy petroleum distillate, then a solvent
extraction may also be performed in order to overcome the volatility bias in the charcoal
strip extraction and to obtain a more representative chromatogram.

Both neat ignitable liquids and ILRs extracted from fire debris are analyzed by
GC-MS per ASTM Standard El6l8-O6el [4]. The GC-MS is operated under conditions
that are capable of adequately separating a test mixture consisting of common ignitable
liquid constituents. The typical test mixture contains the even-numbered normal alkanes
from C3 to C20, as well as toluene, p-xylene, o-ethyltoluene, m-ethyltoluene, and 1,2,4-
trimethylbenzene. This test solution is usually prepared at a concentration of 0.005%
(v/v) per component to ensure the sensitivity of the instrument [4,5].

Once a chromatogram has been obtained of neat ignitable liquid or extracted ILR.
the pattern of its peaks, as well as the relative peak ratios, are used to determine whether
or not it is consistent with a common ignitable liquid. In order for an ignitable liquid to
be identified based on the aforementioned ASTM classification scheme, certain criteria in
the chromatographic patterns must be met for each ignitable liquid class [4,5].
Information about the specific characteristics observed in the chromatographic profiles
for each defined ASTM class can be found in Appendix A [4,5].

In addition to the chromatographic patterns, the mass spectral data are also used to
identify the questioned sample. Extracted ion profiles (EIP) are generated for specific
compound classes that are common to ignitable liquids. Table 1.1 lists the major ions

that are used to create these summed profiles [12]. The patterns of the EIPs are also

Table 1.1 Major Ions Included in Extracted Ion Profiles for Ignitable Liquids [12]

 

 

 

 

 

 

 

 

 

 

 

 

 

Compound Class Ion Mass-to-Charge (m/z) Ratio
Alkane 43, 57, 71, 85
Cycloalkane and alkene 55, 69
n-Alkylcyclohexane 82, 83
Aromatic 91, 105, 119; 92, 106, 120
Indane 117,131;118,132
Alkylnaphthalene 128, 142, 156, 170
Alkylstyrene 104, 117, 118, 132, 146
Alkylanthracene 178, 192, 206
Alkylbiphenyl/acenaphthene 154, 168, 182, 196
Monoterpene 93, 136
Ketone 43, 58, 72, 86
Alcohol 31, 45

 

utilized in the identification of the questioned sample.

The method described above for the identification of ignitable liquids is
subjective in nature, and is often based on the analyst’s experience. The pattern
recognition is not statistically based and frequently is not straightforward [4,5]. The
questioned sample and reference samples used for comparison seldom correlate perfectly.
In fact, the chromatograms of the questioned samples can be skewed due to weathering or
insufficient recovery. Intense heat exposure from the fire can cause the loss of more
volatile components at the beginning of the chromatogram. Interferences from the
pyrolysis of substrate materials indigenous to the collected fire debris can create
additional peaks in a questioned chromatographic profile. A fire debris analyst must be
aware of these possible complications when making ignitable liquid identifications in
order to prevent false positive identifications.

1.4 Current Literature Review

Fire debris analysis and ignitable liquid characterization is well represented in the
literature [13-25]. Because the current method of ignitable liquid identification is
subjective, research into more objective methods, such as association and discrimination
by Chemometric procedures, is essential for the advancement of fire debris analysis.
Current research primarily discusses the utility of multivariate statistical methods in the
association and discrimination of ignitable liquids [16-20, 23]. The importance of data
processing prior to Chemometric analysis is also discussed [26,27]. In addition to
chemometrics, interferences and burning conditions have also been studied to determine
their effects on fire debris analysis [28-32]. Although gasoline is the most commonly

encountered ignitable liquid in arson cases [13] and most often discussed in the literature,

diesel fuel was selected as the ignitable liquid for this research project. Diesel was
chosen because it is a complex petroleum distillate with a number of components that
span the boiling point range, and it is less frequently discussed in the literature.
1. 4.1 Statistical and Chemometric Analysis of Petroleum Distillates

Statistical comparisons of chromatograms produced from the analysis of various
ignitable liquids have been limited to an evaluation of peak area ratios of specific
components or to Chemometric procedures performed on small sections of the
chromatogram. Mann employed a method of comparing peak area ratios for eight
components from the n-pentane to n-heptane region of the chromatogram to successfully
discriminate gasoline samples [13]. Barnes et al. utilized a similar approach in which
normalized peak area ratios of various aliphatic and aromatic constituents present in the
headspace (e. g. substituted cycloalkanes, alkylbenzenes, and alkylnaphthalenes) were
used to associate unevaporated gasoline samples to those from the same source that had
been evaporated to 75% and 50% [14]. More recently, Sigman and co-workers described
a statistical technique called covariance mapping in which gasoline samples could be
distinguished based on distance metrics calculated between covariance matrices of
replicates of both the same and different samples [15]. It should be noted that in a blind
study of two unknown samples from a set of ten, one of the two was determined to be
statistically different from its known source. It was suggested that evaporation of the
sample between analyses could explain this Type I error. However, the presence of a
Type I error, which is defined as a difference identified by the technique when one is not
actually present, significantly limits the utility of the proposed identification

methodology.

More advanced multivariate statistical techniques have also been utilized to
associate and discriminate ignitable liquid samples. Sandercock and Du Pasquier used
Chemometric procedures such as linear discriminant analysis (LDA) and principal
components analysis (PCA) on GC data collected using selected ion monitoring (SIM)
for the C0-C2 naphthalene components in order to group similar gasoline samples based
on grade, country of origin, and the season in which the sample was collected [16-17].
They employed similar methods to link evaporated samples to unevaporated samples
[18]. Tan et a1. utilized PCA and a soft independent model classification analogy
(SIMCA) approach to differentiate GC-MS-SIM data of various ignitable liquids spiked
onto wood and carpet as background matrices [19]. Successful classification was
achieved for all ignitable liquids in the presence of wood and carpet using the developed
SIMCA model. Doble et al. demonstrated the use of artificial neural network (ANN)
algorithms to successfully classify 88 gasoline samples as either regular or premium
grade based on the percent peak area of 44 compounds that were identified in all samples
[20]. Similar ANN algorithms were also reported to distinguish the gasoline samples
based on their season of collection with a 97% success rate.

While the majority of the forensic literature focuses on the identification and
classification of gasoline samples as ignitable liquids, other more complex petroleum
distillates such as diesel fuel have been discussed in the environmental science literature
with respect to the association of samples collected from oil spills to their likely source.
Environmental literature tends to focus on the use of unique classes of compounds in
diesels in order to discriminate samples. In two recent studies, Wang et al. researched the

ability of the sesquiterpane and diamondoid compound classes to potentially discriminate

diesel samples by analyzing more than 100 crude oils and refined products by GC-MS
[21,22]. Diagnostic peak area ratios of several compounds from both the sesquiterpane
and diamondoid classes were determined in an effort to identify the ratios that would
provide the most discriminatory information. Ultimately these diagnostic ratios were
effectively applied to a case in which the source of an oil spill was determined based on
these two compound classes alone. The discriminatory potential of the sesquiterpane
class was further investigated by Gaines et al. in which fourteen diesel samples were
analyzed by GC-MS and extracted ion chromatograms (EIC) for 22 different
characteristic compound classes were generated [23]. The peak ratios of several
compounds within each of the 22 classes were determined and PCA was performed in
order to select the ratios that generated maximum discrimination among samples. Results
illustrated that full discrimination of all samples was possible with only nine peak ratios
from the alkylbenzene, alkylphenanthrene, and sesquiterpane classes, and two samples
from an actual oil spill were differentiated by the model developed.

More novel analytical methods have also been discussed in oil spill identification.
Another study by Gaines et al. employed a qualitative and quantitative GC x GC
approach to analyze a controlled oil spill [24]. By comparing two-dimensional ordered
chromatograms, they observed that the patterns of alkanes, cycloalkanes, alkylbenzenes,
alkylnaphthalenes, and anthracenes/phenanthrenes were useful in comparing the oil spill
to one of its two potential sources. Quantitatively, four panels of integrated peak area
ratios of the abovementioned compound classes were used to compare samples. Using
both these methods a controlled oil spill was correctly matched to one of two potential

sources. A study by Wang et al. showed a novel procedure for the visualization of GC-

10

MS data as a two-dimensional separation that is valuable in specific compound class
separation when a nonfragmentation ionization technique is used [25]. This method was
demonstrated on complex petroleum fractions, so that even compound classes at low
abundances can be utilized for characterization.

1. 4. 2 Chromatographic Pre-processing

The need for pre-treatment of the chromatographic data prior to statistical and
Chemometric analysis has been more recently discussed in the literature, although it has
yet to be applied specifically in a forensic science context [26]. Some pre-treatment
processes, such as area normalization and mean-centering prior to PCA have long been
demonstrated [27]. Another more significant pre-treatment is that of retention-time
alignment of chromatograms to correct retention time drift anomalies caused by column
degradation and random fluctuations in analysis conditions. Retention time alignment is
an important step prior to PCA, since PCA maximizes the variance among a sample set.
If the sample chromatograms contain multiple shifts in retention time, the PCA will focus
on those differences in the samples instead of actual chemical variation.

Johnson et al. proposed a peak matching algorithm for the retention time
alignment of GC peaks in order to reduce retention time shifts among several
chromatograms [26]. This approach applies an estimation of the first derivative
throughout the chromatograms and searches for zero crossings to identify peaks in a
specified target chromatogram and the sample chromatograms. Peaks in the samples are
matched to those in the target within a set window size and are then interpolated to
include either more or less data points in the retention time axis so that the retention times

of the peaks in the sample will be equal to those in the target. The authors demonstrated

11

the utility of this approach by analyzing a set of diesel samples by GC with a flame
ionization detector (FID) and then subjecting them to the proposed algorithm followed by
PCA. The unaligned set of 60 chromatograms exhibited differences in retention times up
to 300 ms, but after alignment, the variation was only 17 ms with a relative standard
deviation of 0.003%. PCA demonstrated that the alignment did not cause a decrease in
chemical selectivity, but instead the resolution of clusters of different samples was shown
to be greatly improved after alignment [26].

I. 4. 3 Burn Conditions and Matrix Interferences

The analysis of ILRs extracted from fire debris is not straightforward. In fact. the
burning conditions and contributions from the burned matrix can hinder chemometric
analysis. Artifacts present in the chromatogram from sample weathering or matrix
interferences may cause samples that are actually similar to be difficult to associate
because of the extraneous components. These issues have been moderately discussed in
the literature, though most studies only look at the matrices alone without any ignitable
liquid present, much less an ignitable liquid that has been burned as would be the case in
an actual arson situation [28-32].

Bertsch examined potential interferences from carpet samples in the identification
of ILRs, specifically gasoline, from fire debris [28]. It was observed that when carpets
and carpet paddings were burned that some amounts of alkylbenzenes and naphthalenes
were produced. These same compounds are typically used as markers in the
identification of gasoline. It was determined, however, that by observing EICs of the C2-
C5 alkylbenzenes (m/z 106, 120, 134, and 148) and the methylnaphthalene isomers (m/z

142), the relative amounts of these compounds observed in a gasoline versus that

12

produced by the pyrolysis of carpet is distinguishable. It was noted, however, that when
the carpet has been freshly manufactured some remnants of petroleum-based products
used in the manufacturing process may still be detectable. In addition to these remnants,
carpets treated with petroleum-based water-proofing agents or insecticides may also
produce a gasoline-like chromatogram. The use of EICs in this case is not beneficial for
the identification of gasoline because the petroleum product is inherent to the carpet
itself. In a similar study, Cavanagh et al. examined carpets and mats from cars of
unknown history, and a small percentage of these carpets exhibited a chromatographic
profile similar to that of evaporated gasoline [29]. The authors suggested that inherent
interferences from car carpets are uncommon, which would increase the evidentiary value
of the presence of a more concentrated ILR sample or one that does not compare with an
evaporated profile.

Lentini et a1. presented a study in which several commonly encountered materials
were shown to produce petroleum-like chromatograms [30]. Several samples of different
matrix types were analyzed using passive headspace concentration followed by GC-MS.
The matrix types analyzed included clothing, shoes, household products, building
materials, paper products, cardboard, and adhesives. Many of the substrates examined
exhibited Tle or EIPs very similar to those of ignitable liquids, though patterns
consistent with gasoline were much less frequently observed than those of medium of
heavy petroleum distillates. It was concluded that, because of these potential sources of
interference that are likely due to the regular use of petroleum products in the
manufacturing, cleaning, and treating processes, control samples are crucial as a

reference in fire debris analysis. In another similar study, Almirall and Furton analyzed

13

common household products that had been burned in controlled conditions and also
reported consistency in composition with ignitable liquids [31]. They determined the five
most frequently encountered compounds in these substrates that are also observed in
ignitable liquids are toluene, styrene, naphthalene, benzaldehyde, and ethylbenzene. In
addition, Almirall and Furton analyzed several of the same samples using pyrolysis GC-
MS to characterize the products formed during burning. These pyrolysis products, most
frequently toluene, naphthalene, styrene, and ethylbenzene, may serve as another source
of interference in the identification of ignitable liquids extracted from fire debris.

Borusiewicz et al. studied several factors that may influence the collection and
detection of ILRs, and, unlike the previous studies mentioned, actually included the
spiking of ignitable liquids onto the matrices being analyzed [32]. By reproducing burn
conditions and visually comparing chromatograms, the authors were able to examine how
the ignitable liquid type, burned matrix type, burn time, and air availability affect the
ability to detect ILRs. They concluded that the material being burned (e. g. wood, carpet.
etc.) is the most significant factor in ILR analysis. The other parameters were determined
to be less significant, even when compared to random variables in fire debris analysis that
are difficult to replicate in a laboratory setting, such as ignitable liquid dispersion or the
arrangement of the material when burned. The disadvantage of this study is that the
ignitable liquids analyzed were not actually burned; instead, they were only spiked onto
already burned matrices.
1.5 Research Objectives

In the preliminary work of this project, conducted throughout the summer and fall

of 2006, neat diesels were analyzed by GC-MS and characterized using PPMC and PCA

14

[33]. 25 diesels from eight different brands of service stations were collected for analysis
and characterization. In addition to the TICs, EICs were generated for representative
compounds of both aliphatic (rn/z 57) and aromatic (m/z 91 and m/z 141) constituents
present in the diesel samples. PPMC coefficients were generally observed to be higher
between diesels from the same brand and lower between those from different brands. A
wider array of PPMC coefficients were observed in the EICs, especially for m/z 91,
which suggests that the aromatic composition of the diesel may be the source of the
variance between those diesel samples. PCA scores plots typically showed four diesel
sample clusters, three that contained only samples from one brand and one with the
remaining samples. This trend was generally observed for both the TIC and EIC data.
The PCA loadings plots, which demonstrate the contribution of each component to the
variance in the data set, suggested that the relative aliphatic and aromatic composition,
similar to that observed by PPMC calculations, was the most significant source of
variation among the samples.

The fundamental objective of this research project is to expand on the previous
work and study the possibility of linking ignitable liquid residues extracted from fire
debris to a neat liquid sample. Again, diesel fuel was selected as the ignitable liquid for
this project because it is more chemically complex than other ignitable liquids such as
gasoline, and because it is less frequently discussed in the literature. Basic statistical and
more advanced chemometric procedures, such as PPMC coefficients and PCA, are
implemented to associate and discriminate both neat diesels and those that have been
extracted from burned debris based on their chemical composition. Data pre-treatment

issues, which include chromatographic alignment and normalization, are also considered.

15

This project expands on the current research described in the literature by
focusing primarily on the ability to link an unburned diesel sample to a diesel sample that
has been extracted from burned debris. This project examines data from the entire
chromatogram instead of just peak ratios or sections of the chromatogram as reported in
the literature. Using the entire chromatogram may increase discrimination capability, and
no information needs to be known or assumed about the samples in question. EIPs of
several potential characteristic compound classes (e. g. aliphatics, aromatics, polynuclear
aromatics, and indanes) are also used for comparison. Again, this project implements the
entire EIP to further enhance discrimination. The increased chemical information from
the full TICs and EIPs may potentially show more significant association and dissociation
of diesel based on chemical composition in chemometric procedures. The chemometric
procedures may overcome the changes produced in the diesel sample due to burning as
well as matrix interference contributions so that a burned sample can be accurately
associated with its neat counterpart, which minimizes the risk of false positive ILR
identifications. This research, which is applicable to almost any petroleum product with
potential as an ignitable liquid, serves as the groundwork for future projects. and could

ultimately conclude with scientific validation for use in a forensic laboratory setting.

16

Chapter 2
Analytical and Chemometric Theory
2.1 Gas Chromatography-Mass Spectrometry

Gas chromatography-mass spectrometry (GC-MS) is a well developed analytical
technique that has become commonplace in forensic laboratories because of its wide
application range. The combination of chromatographic and mass spectral results offers
a multidimensional analytical tool. It is commonly used in forensic laboratories for drug
analysis, toxicology, environmental contamination, and explosives detection. It is also
the standard technique used for ignitable liquid identification and fire debris analysis. In
fact, as previously described in Chapter 1, a standard method has been developed by
ASTM for the specific use of GC-MS in the analysis of ignitable liquid residues in fire
debris extracts [4].

Chromatography is a technique in which separations of mixtures are driven by
interactions of individual components in the mixture with a mobile and a stationary
phase. In gas chromatography, the mobile phase is an inert carrier gas, typically helium
or nitrogen. A GC system consists of several individual components: a heated injection
port, through which the sample is introduced, a column housed in a temperature regulated
oven, where the separation takes place, and a detector. Several different types of
columns and detectors are available for use based on the sample and the type of
separation required. In GC-MS, the typical column is a capillary column with the
stationary phase coated on the inner walls, and the mass spectrometer serves as the
detector. In order to analyze a sample by GC-MS, it must be volatile and thermally

stable.

l7

In a GC—MS system, the sample must be introduced in the vapor phase. Liquid
injections, however, are commonly performed, because the injection port is sufficiently
heated (250 °C and above) to immediately vaporize the sample. The mobile phase
constantly flows through the system and sweeps the vaporized analytes onto the column.
The column, as mentioned before, contains the chromatographic stationary phase to cause
separation. The stationary phase is usually coated directly on the interior surface of the
capillary column, though it can also be coated onto particles that are used to pack the
column. The affinity of the analyte for this stationary phase dictates its retention on the
column. For example, if a component has a significant affinity for the stationary phase
(e. g. a polar component with a polar stationary phase), it will interact more with it than a
component with a lesser affinity (e. g. a non-polar component with a polar stationary
phase). The component with a higher affinity will be more retained and will elute later
than the component with the lesser affinity. The parameters of the column, such as
length, diameter, and stationary phase thickness, as well as the composition of the
stationary phase itself, have an effect on retention and ultimately the quality of the
separation.

The positioning of the column in a temperature regulated oven allows for it to be
heated to facilitate analyte separation. In an isothermal analysis, the oven is held at a
constant temperature throughout the course of the run. This temperature is typically at or
just above the boiling point of the sample to be analyzed. For samples that contain
components with a broad range of boiling points, a temperature programmed analysis
provides better resolution, quicker analysis time, and minimal band broadening in

comparison to isothermal conditions. In a temperature program, the oven is held at some

18

initial temperature in order to initiate the separation as well as to prevent condensation of
the analytes on the top of the column. The temperature is then ramped over the course of
time to a final temperature. The rate of the ramp is dependent on the sample. A slower
ramp rate is ofien used when components that have similar boiling points need to be
separated, such as in a complex mixture like diesel. Faster ramp rates, however, provide
quicker analysis times, thus ramp rates are often selected to provide sufficient resolution
in the minimal amount of time. The oven is then held at the final temperature to ensure
complete analyte elution.

The column passes through a heated transfer line into the ionization source of the
mass spectrometer detector. The transfer line must be maintained at a high temperature,
typically 300 °C, in order to prevent condensation of the separated analytes during
transfer into the director. From the transfer line, the analytes are introduced directly into
the ionization source where they are ionized and fragmented. Several types of ionization
sources are available, each with its advantages and disadvantages. Electron impact (El)
ionization is the most common in most bench-top GC-MS instruments used in forensic
laboratories for fire debris analysis. In El, a filament with electric current running
through it is heated in order to produce high-energy electrons (70 eV). These electrons
interact with the vaporized analyte molecules to produce molecular ions that are
fragmented through further collisions. These positive molecular and fragment ions are
then transported by a series of focusing lenses into the mass analyzer, where they are
separated based on their individual mass-to-charge (m/z) ratio.

As with ionization sources, several types of mass analyzers are available for use,

but the most common in bench-top instruments is the quadrupole mass analyzer. The

19

quadrupole consists of four parallel conducting rods, with one pair connected to the
negative terminal of a DC source and the other pair connected to the positive end. Each
pair is also connected to a radio frequency (RF) AC source. These four rods define a
cylindrical region through which the ions travel. Under the influence of the electric field.
the positive ions are attracted to the negatively charged rods. As the ion begins to move
toward the negative rod, the potentials on the rods are switched, and the ion moves
toward the opposite rod that is now negative. As the potential is varied, the ion travels in
an oscillating pathway through the quadrupole and on to the detector. At a given DC/RF
ratio, only ions of a narrow m/z value will travel an oscillation path that allows them to
pass through the region defined by the rods without hitting the rods. Ions of other m/z
values will travel a wider path that will cause them to come in contact with the charged
rods, become neutralized, and therefore not be detected. A full mass scan is obtained by
increasing the DC and RF potentials while still maintaining a constant ratio between the
two. As the ions pass through the quadrupole, they are directed to a transducer, typically
a continuous conversion dynode. The positive ions contact the highly negatively charged
conversion plate, which causes the release of secondary particles, of which electrons are
included. These electrons then enter the continuous dynode electron multiplier, where
they strike the multiplier plates, which are coated with a substance the readily emits
secondary electrons. This release of electrons caused by the contact with the plates
results in a cascade of electrons that reach the end of the multiplier where they are
detected.

GC-MS results are two-fold. The retention time of a component, as well as its

mass spectrum, are used in making comparisons among samples. In a conventional GC-

20

MS analysis, the entire mass range is scanned during detection and a total ion
chromatogram (TIC) is generated. The TIC is a plot of intensities of m/z values in each
scan versus time. From this TIC, a specific m/z ratio can be selected as a filter so that
only the signal from that ion is retained while all signal from all other ions is eliminated.
The result is an extracted ion chromatogram (EIC) in which the abundance for a
particular ion is plotted against retention time. Ele can be useful to reduce highly
complex and convoluted samples to a more manageable format for interpretation. In fire
debris analysis, EICs are beneficial to reduce a petroleum distillate, which is a very
complex mixture of many components, to only those components of a specific compound
class, such as aliphatics or aromatics. The amount of aliphatic and aromatic components
present in an ignitable liquid is valuable information for ignitable liquid classification. In
fire debris analysis, several individual EICs are summed to generate an extracted ion
profile (EIP) of multiple representative ions of a characteristic compound class. These
EIPs are at higher abundance levels and contain more characteristic features than 15le
alone.
2.2 Data Pre-treatment

Data pro-treatment is necessary prior to the application of chemometric
procedures in order to correct for any variation among the diesels not actually due to
differences in chemical composition. For example, minor shifts in retention time can
occur between sample injections due to instrumental drift in the GC-MS. Retention time
alignment procedures can correct for these nominal variations in retention time. Another
source of variation among the samples includes slight differences in injection volume. A

normalization process can correct for these slight differences. Data pre-treatment

21

procedures are essential in order to ensure accurate results from the PPMC coefficients
and PCA.
2. 2.1 Retention Time Alignment

The first data pro-treatment step that must be taken after the raw data is collected
is a retention time alignment. The alignment is necessary to account for normal
instrumental drift, such as slight changes in temperature, pressure, and flow rate, from
run to run. Retention time shifts can also be caused by the natural aging process of the
column, which degrades over the course of several runs. A retention time alignment
algorithm can correct for these inevitable sources of retention time drift. The algorithm
used in this work was developed specifically for diesels and is available in the literature
[26]. Figure 2.1, which highlights the decane peak for Diesels 1-5, illustrates the effect
of the application of the retention time alignment algorithm to the raw chromatographic
data. The decane peak in Diesel 2 is severely shifted when compared to the other diesels,
but the algorithm is capable of correcting for that shifi.

In the algorithm a representative target chromatogram is chosen by the user from
the sample set to be aligned. The first algorithm function is to perform a baseline
correction on the target and sample chromatograms. The baseline correction factor is
determined by a linear regression through the first few points and the last few points of
the chromatograms. This correction is applied to account for minor changes in the
baseline from run to run, though it is contained in a sub-routine in the algorithm and is
easily edited or even removed if deemed unnecessary. The algorithm then cycles through
each chromatogram in turn and estimates the first derivative of the chromatogram by

calculating the difference in signal strength between consecutive data points. Once this

22

 

200000 fl

 

 

 

 

 

 

 

 

 

 

 

 

 

Unaligned
8
E
N
'6
fl
5
1D
4
o
9.35 Retention Time (min) 9.60

200000 1

Aligned l
3
:1
CB
1:
:1
:1
.c
<2

l

O A

9.35 Retention Time (min) 9.60

Figure 2.1 The Effects of a Retention Time Alignment
Algorithm on the Decane Peak of Diesels 1-5

23

difference exceeds a threshold for baseline noise defined by the user, the data points are
considered to be the leading edge of a chromatographic peak. The algorithm then
searches for a zero crossing, or a change in sign, within this first derivative estimation.
which indicates the apex and tailing edge of a chromatographic peak. The algorithm then
interpolates the subsequent retention time at which the first derivative is equal to zero.
rounds that retention time to the nearest integer value, and then adds that retention time to
a list being generated for that chromatogram. Once the retention time tables for each
chromatogram have been compiled, the algorithm compares the peaks present in each
sample with the peaks identified in the target chromatogram. If the sample
chromatogram contains a peak within a user-defined window size to that of the target.
then it is considered a match and the retention time axis will be interpolated to include
more or fewer data points so that the sample peak has the same retention time as the
target peak and the peaks align. If the target chromatogram contains a peak that the
sample chromatogram does not, or vice versa, those peaks are not considered in the
retention time axis interpolation, which allows the alignment to accommodate samples
that contain varying numbers of peaks.

Prior to implementation, however, certain user-specified parameters of the
alignment algorithm need to be investigated in order to ensure optimal alignment. The
selection of the target chromatogram must be done carefully so that it contains as many
peaks present in the chromatograms to be aligned as possible. It is also important that the
peaks in the target chromatogram are relatively centered in terms of retention time and
peak symmetry in comparison to the peaks from the sample chromatograms. A poor

target selection will lead to a greater propensity for misalignments. Another factor

24

defined by the user that must be considered is the baseline noise threshold for peak
identification. The noise of the chromatogram is determined from the first several data
points in the chromatogram by default. However, because diesel chromatograms contain
features within those first points, in this research the algorithm was amended to calculate
the noise from the last several points in the chromatogram where a baseline was
discernible. The standard threshold for determining the presence of a peak is five times
the standard deviation of the calculated noise. The last user-defined variable that requires
investigation is the window size. This last parameter dictates the magnitude of retention
time shift of a peak in a sample chromatogram that is still determined a match to the same
peak in the target chromatogram. Ideally, the window size, which the algorithm defaults
as five, is large enough to account for normal retention time drift, but smaller than the
average distance between adjacent peaks. A brief examination of these parameters can
lead to fewer errors and improved alignments.

Even after optimization of these variables, the alignment algorithm is not devoid
of limitations. Peaks with small signal-to-noise ratios are often not identified because
they still fall below the optimal baseline noise threshold. These smaller peaks also tend
to be noisier, which can result in the determination of multiple zero crossings and leads to
peak misidentifications. Another limitation of the alignment algorithm is that the peak
identification method of first derivative estimations is entirely dependent on the scan rate
of the mass spectrometer. Smaller peaks are often overlooked by the algorithm’s peak
finding function when the scan rate is too rapid because the first derivative is not
substantial enough to exceed the noise threshold. The alignment algorithm also lacks any

method for the determination and alignment of shouldered peaks. Their identification

25

depends on the size of the shoulder in comparison to its parent peak, so that some
shoulders are well aligned while others are not. Despite these limitations, the alignment
algorithm has been demonstrated in the literature to be adequate for GC peak alignment
of diesel samples [26]. The user, however, must be aware of the consequences of these
limitations not only on the alignment itself, but on the ensuing chemometric analyses as
well.
2. 2. 2 Area Normalization

In addition to retention time alignment, other pre-treatment steps are necessary to
minimize sources of variation in the chemometric analyses introduced by the analytical
technique or injection method. Minimizing artificial variation ensures that the results of
Pearson product moment correlation (PPMC) coefficients and principal components
analysis (PCA) are based on the true sample variation. Normalization is performed in
order to account for minor differences in injection volume from run to run. In a basic
normalization process, each chromatographic peak is ratioed to a reference that is present
at a consistent concentration throughout every analysis. In many cases, the reference
peak is a purposely added component known as an internal standard that is not inherent to
the sample being analyzed. However, because diesels are complex and contain numerous
components that cover a broad range of boiling points, the choice of an internal standard
that would not co-elute with a native component is difficult. Therefore in this research.
an area normalization process was used instead. In this method, the areas of all peaks
present in a chromatogram are summed, and the individual peak areas are then divided by
the total area. In this research, however, the entire chromatogram is analyzed, not just the

areas of each peak. In order to area normalize the individual data points of the full

26

chromatogram, the total area underneath the chromatographic curve is calculated to serve
as the reference. The individual data points are then divided by the total area. The data
points are then multiplied by an average of the total areas of all sample chromatograms in
the data set in order to return the chromatograms to the original order of magnitude so as
to maintain discernible features for further data analysis.

Figure 2.2, which shows the same decane peak of the five diesels as before,
demonstrates the results of the area normalization process. In the aligned data only. it
appears that little similarity exists among the five diesels in the concentration of decane
present. However, once the data is normalized, clear associations can be made for
Diesels 3 and 4 as well as for Diesels l, 2, and 5. These two sets are easily differentiated
based on the concentration of decane present.

2.2.3 Mean-Centering

Another pre-treatment step, mean-centering, is performed specifically in
preparation for PCA. In this process, the mean of an individual variable (in this case. the
abundance level at a single retention time) is calculated across the entire data set. This
mean is then subtracted from each value so that the resultant sum across the sample set
equals zero. Figure 2.3 shows the same decane peak of the five diesels as before, but
now the data has been aligned, normalized, and mean—centered. The same diesels are
associated as before, only in the mean-centered plot, Diesels 1, 2, and 5, which were
lower in decane concentration, exhibit a negative peak. Diesels 3 and 4, on the other
hand, remain positive. It should be noted that the two positive peaks are greater in
magnitude than the three negative peaks, a characteristic that is observed because the

means have been mathematically forced to sum to zero.

27

 

200000

 

 

 

 

 

 

 
 
 

 

 

 

 

 

 

 

 

 

. l
Aligned l
3 m
= l
a i
-o l
:1
:3
.o
<1:
\
0 \
9.35 Retention Time (min) 9.60
150000
Aligned and Normalized
Diesels 3 and 4 }
3
1:
fl
'5
g Diesels 1, 2, and 5 }
<11
0
9.35 Retention Time (min) 9.60

Figure 2.2 The Effects of Area Normalization on the Decane Peak of Diesels 1-5

28

 

40000 We a:

Diesels 3 and 4 }

    

Abundance
P ‘?

 

 

 

-40000 '
Retention Time (min)

Figure 2.3 The Application of Mean-Centering to the Decane Peak of Diesels 1-5

29

Mean-centering is an important step prior to PCA to ensure that the maximum
variance described by the principal components selected does not contain any deviations
from the mean. Oftentimes when PCA is performed on data that has not been mean-
centered, the primary principal component reflects how the samples deviate from the
mean, not necessarily how they inherently vary. Upon mean-centering, the variance
becomes centered around the origin so the principal components are allowed to describe
actual sources of variation among the samples.

2.3 Pearson Product Moment Correlation Coefficients

The PPMC coefficient (r) indicates the degree of linear correlation between two
samples. The correlation coefficient is calculated by dividing the covariance between the
two samples by the product of each sample’s standard deviation as shown in Equation

2.1:

ZtX, 500’, 47)

r:
[201 -§)2‘/Z(Y. 4‘02

Values for r range between -1 and l, where a correlation of 1 indicates a perfect positive

 

 

Equation 2.1

linear relationship and a correlation of -1 indicates a perfect negative linear relationship.
It has been suggested that correlation coefficients between 0.8 and 1 indicate a strong
correlation between samples, while coefficients between 0.5 and 0.8 indicate a medium
correlation. Coefficients less than 0.5 imply a weak correlation [34]. Correlation
coefficients close to zero signify a lack of any correlation.

To assess the correlation between two diesel samples, coefficients are calculated
as in Equation 2.1, where the X variables denote abundances at individual retention times

for one diesel, and the Y variables correspond to abundances at the respective retention

30

times for the other diesel. In order to determine the strength of association between two
diesels, the standard convention mentioned above is not applicable because of the level of
inherent similarity of the samples being compared. The conventional model relates to
samples that range significantly in their chemical composition. Diesels, however, are
very similar in their basic composition, so their correlation coefficients would be
expected to be closer to unity. Therefore, although differences in correlation coefficients
between diesels will be smaller, these differences may still be significant.

A more appropriate guideline for samples like diesel that are already so similar is
the range of coefficients between several replicate analyses of the same sample.
Replicate coefficients, which represent the highest possible level of similarity, should be
close to unity. If correlation coefficients between samples that are not replicates fall
within the same range as the replicates, then those samples must be considered very
similar. Diesels with correlations outside that range can be considered less similar due to
differences in chemical composition. With the application of this guideline, PPMC
coefficients are useful in assessing the association and discrimination of diesels based on
chemical composition.

Another application of PPMC coefficients is method precision. As previously
mentioned, replicate analyses should result in correlation coefficients of unity if the
instrumental technique and manual injection procedure are repeatable. Because of
random instrumental and analyst variations, perfect correlation coefficients "are unlikely;
however, coefficients very close to one should be attainable between replicate analyses.
Therefore, average correlation coefficients of replicates can serve as an indicator of how

precise the instrument and the manual injection process are.

31

2.4 Principal Components Analysis

PCA is a multivariate statistical technique most commonly used to reduce larger
data sets to dimensions that can be easily visualized by standard procedures while still
retaining the most significant information. PCA, on the most basic level, determines the
sources of maximum variation among samples in a data set and clusters the samples
based on their relative contribution to these sources of variation. It is a clustering
technique that is capable of associating and discriminating samples based on how they
are correlated.

PCA consists of several steps of complex matrix mathematics. The first step is to
generate the covariance matrix, which consists of covariance calculations between all
dimensional pairs in the data set. The covariance between a pair is calculated by

Equation 2.2:

cov(X, Y): Z (X’ _ YXY‘ _ )7) Equation 2.2
(n - 1)

The covariance matrix for a data set of n dimensions contains n!/(n-2)!*2 different
covariance values, as well as commutative repeats [cov(X,Y) = cov(Y,X)] and covariance
identities [cov(X,X)], which are equal to the variance of that variable [35]. These
commutative repeats make the covariance matrix symmetric about the forward diagonal.
The next step in PCA is to calculate the eigenvectors and eigenvalues of the
covariance matrix. Eigenvectors are derived from the basic properties of transformation
matrices. A vector that, when multiplied by a square matrix, results in a reflection of
itself is considered an eigenvector of that matrix. Any multiple of that vector is also
considered an eigenvector because the directionality of the vector is not affected, only the

length. In simpler terms, an eigenvector a is a vector that satisfies the following

32

relationship with a square matrix A: O. * A = k * a, where it is a nonzero number known
as the eigenvalue. In PCA of a multidimensional data set, computerized iterative
algorithms are often employed to calculate all the eigenvectors and their associated
eigenvalues of a covariance matrix. The eigenvectors are then ranked in descending
order by their respective eigenvalue. The maximum eigenvalue corresponds to the most
significant eigenvector. It should be noted that eigenvectors are orthogonal to one
another, and that n eigenvectors are calculated for a matrix of n x n dimensions.

For PCA, the eigenvector with the largest eigenvalue is the first principal
component (PCl). This primary eigenvector can be thought of as the axis on which the
original data can be projected that results in the maximum amount of spread among the
samples in the data set. Additional PCs, positioned orthogonally in the sample space to
the previous PC, describe lesser amounts of variation corresponding to the magnitude of
their eigenvalues. The amount of variation afforded by each PC is represented as a
percentage of its eigenvalue to the sum of all eigenvalues. Often, only two to three PC 3
are needed to describe 85-90% of the variance among the data [27]. The dimensional
reduction property of PCA is reflected in the significant eigenvalues. If a complex data
set of many dimensions can be accurately described by only two or three dimensions.
which are capable of being plotted and visually observed, the analysis of that data set
becomes much easier by conventional methods.

Once the eigenvectors and eigenvalues have been determined, they are utilized to
calculate scores for each sample. Scores are calculated for each sample by a simple
matrix multiplication of the original data (in columns) by the eigenvectors (in rows).

This multiplication results in a single score for each sample for each principal

33

component. A scores plot usually consists of an x-y scatterplot of the PCI score versus
the PC2 score for each sample. Three-dimensional plots of the scores for the first three
PCs are also frequently used. In these scores plots, the samples will cluster based on how
they are each affected by the eigenvectors selected. For diesels, the eigenvectors
correspond to individual chemical components that are the most variable among them.
Therefore, the scores plots cluster diesels based their similarities and differences in
chemical composition. A plot of the individual eigenvectors, known as a loadings plot,
indicates those components that contribute the most to the variance. If a component
loads high, then it contributes more significantly to the score of a sample than other
components do.

Although clustering patterns observed in scores plots can be used for sample
association, PCA is an unsupervised technique, because the method is not guided by or
modeled after a known training set. Instead, PCA clusters samples in a data set solely on
how those particular samples vary. As samples are added or removed, the clustering
results may change. For this reason, empirical classification by PCA is difficult; instead,
it is more useful to discuss how the samples fall into the clustering patterns by degrees of

association with and discrimination from one another.

34

Chapter 3
Association and Discrimination of Neat Diesels Using PPMC Coefficients and PCA
3.1 Introduction

The fundamental objective of this research is to investigate the potential of linking
ILRs extracted from fire debris to the corresponding neat liquid sample using
chemometric procedures, specifically PPMC coefficients and PCA. The initial step,
however, is to demonstrate that similar neat ignitable liquids, diesel fuel samples from
different service stations in this case, can be associated and discriminated based on
differences in their chemical composition. This preliminary work is essential in order to
show that PPMC and PCA are capable of distinguishing diesel samples that are known to
be very similar in their basic chemical composition. If the neat diesels cannot be
differentiated, then burned diesels will be nearly impossible to discriminate, because they
will have lost the more volatile components, and hence some discriminatory information,
during the burning process.

Although diesels have a similar basic composition, some amount of variation is
introduced through the refining process. Diesel is a heavy distillate fraction collected
from the refining of crude oil. Oil refining is a multistep process that includes a
temperature-based fractional distillation, several chemical processing steps (thermal or
catalytic cracking, and hydrocarbon unification or alteration), and multiple purification
and product blending phases [36]. Refineries may also include various additives in their
final product blends to improve their quality and performance. Refineries differ in how
they complete these processes to produce their final diesel fuel product, which adds some

amount of variation to a distillate fraction that is generally similar in composition.

35

The differences in the refining process allow for potential association and
discrimination of diesels based on how they were produced. However, it must also be
noted that the refining process itself is variable, so that even products from the same
refinery may vary. In addition to random variations in the refining process, diesel fuel
products must be transported to a distributor and then finally a service station for
purchase by the consumer. While some service station brands such as Marathon refine
their own oil to sell, other brands such as Speedway or Meijer purchase their products
from a regional distributor [37]. Some stations have long-term contracts with distributors
so that they always receive similar products, while others purchase their products on a
need basis from the most cost efficient distributor. Therefore, it is impossible to generate
a database of diesel samples from different service stations because of these constant
changes. This variation must be considered in the interpretation of the PPMC and PC A
results for the association and discrimination of diesel samples.

In order to examine a neat diesel data set, diesel samples were collected, analyzed
by GC-MS, and subjected to both PPMC and PCA in order to assess the degree of
association and discrimination achieved among them.

3.2 Sample Collection and Analysis

Ten diesel samples were collected from various service stations located within the
surrounding area of Lansing, Michigan. Table 3.1 lists the samples, which were
numbered consecutively, and the details of their collection. It should be noted that five
brands of service station (Sunoco, Speedway, Meijer, Marathon, and Mobil) are
represented in the sample set. Diesel samples were collected from two different locations

of each of the five brands in order to have a representative but diverse data set.

36

 

 

323% Eoomocosc SBNBO onco>< $3M 955 $3 Q3woomw 3 Emma
3o=o> Eoomocosc 88:8 osco>< «Emzcnchom 550m Ev 55.982 a $35
326» Eoomocozc 33 too god $583 33 “we? Odom 8:352 w 385
3oz?A Eoomocosa 8% _\oo god @553 8:3 $8 :82 N. 335
as? asasé 85% 33. as: 32 32.68% e 335
30:03 Eoomohosc SR too gem “Ho—mam 82 ESE m Emma
3233 E8383“ no: Coo 3:22. 83% 2:80 ”.33 mmom 8.302 V 335
30:?A 3838:: noRoBo 98M :5: own 8.322 m 385

30:?A Bag 3698 98m 3x055 ooom 825m N .085

30:9A 23 59 SRO veom finial aim 805% ~ 335

noflaum
553.30 no 25,—. «a .830 253—30 he 82— not—3:1— oomtom can—am

 

we.— .aeoseu .085 em 2...;

37

Approximately one-half gallon of diesel was collected from each station in a five gallon
yellow diesel container (Wedco, Willowbrook, IL). The containers were cleaned with
hexane (reagent grade, Mallinckrodt, Phillipsburg, NJ) between sample collections.

After collection, the diesels were brought to the laboratory, transferred to
previously acid washed 250 mL amber bottles, and stored in a refrigerator. Immediately
prior to analysis, the diesels were diluted 200:1 with dichloromethane (C H2C 12,
spectrophotometric grade, Sigma-Aldrich, St. Louis, MO) in clear glass vials (Kimble
Chase, Vineland, NJ). One uL of the sample was injected into the GC-MS running the
Diesel.M method, which employs the following temperature program: initial temperature
50 °C, 2 °C/min ramp to 150 °C, 3 °C/min ramp to 280 °C, hold for 15 minutes. For the
MS, a heated transfer line (300 °C) was positioned in an electron impact (El) ionization
source operating at 70 eV. A quadrupole mass analyzer scanning the 40-550 m/z range at
a rate of 2.94 scans per second. Each sample was analyzed in triplicate to assess the
precision of the instrument as well as the injection technique.

For each diesel sample, the data files of the TIC were extracted as Comma
Separated Values (CSV) files, along with the data files of the individual EICs for
aliphatic compounds (m/z 57) and aromatic compounds (m/z 91). Once all diesel
samples were analyzed, their individual data files were compiled into one set for the TIC
and one set for each EIC using Microsoft Excel spreadsheet software (Redmond, WA).
The Diesel.M method contains a 15 minute hold at the final temperature to serve as a
cleaning step between runs and to prevent carryover. Prior to data analysis, the data
points corresponding to this extended hold, from 80.51 minutes to 108.33 minutes. were

removed from the files since they contained no information about the samples

38

themselves. Once truncated, each chromatogram, both TIC and ElC, contained 1.3.531
data points, all of which were used in subsequent data analysis.
3.3 Data Pre-treatment

The compiled data sets for the TICs and each EIC were taken through a series of
pro-treatment steps prior to data analysis. These pre-treatments served to minimize any
artifacts in subsequent data analysis procedures resulting from retention time shifts and
slight differences in injection volume.

First, individual data sets were subjected to a peak matching alignment algorithm
available in the literature [26]. A discussion of how the alignment algorithm functions
can be found in Chapter 2. The alignment was applied in order to account for minor
drifts in retention time from run to run that can be caused by routine column degradation
and nominal differences in the flow rate of the carrier gas. The alignment algorithm
required a target chromatogram to be selected from the data set. The target was chosen at
random to eliminate any potential bias from its selection. Once chosen, the target
chromatogram was inserted into the data spreadsheet, and the entire file imported into
Matlab (Version 7.4.0, The MathWorks, Natick, MA). After the file was imported, the
target and data columns were designated variables within the Matlab workspace and the
algorithm was performed. The aligned data was extracted from Matlab into spreadsheet
form in Microsoft Excel for further pre-treatment.

After alignment, the data sets were normalized to overcome for slight differences
in injection volume from run to run. An area normalization procedure, as described in
Chapter 2, was selected as the most practical for the diesel samples. For each diesel

sample, the abundance values at each retention time were summed in Microsoft Excel to

39

obtain'the total area under the chromatogram. These sum totals were then averaged to
obtain a mean area. Then, for each diesel sample, every individual data point was
divided by its respective area sum, and then multiplied by the mean area sum. The
multiplication by the mean was performed only to return the abundance values to the
original order of magnitude.

A final pre-treatment step, mean-centering, was performed on the aligned and
normalized data specifically prior to PCA. Mean-centering is applied so that the
variation described by the calculated principal components is centered about the mean as
discussed in Chapter 2. The aligned and normalized data was taken in Microsoft Excel,
and the mean was calculated for the abundances of all samples at each individual
retention time. The respective means were then subtracted from each data point to obtain
the mean-centered data.

3.4 Data Analysis

Once the data sets were pre-treated, several data analysis steps were performed to
determine the precision of both the instrumental technique and the injection method.
Precision was assessed through both relative standard deviations (RSD) of select peak
areas and PPMC coefficients.

For RSD calculations, eight peaks across the entire retention time range were
selected to represent the wide array of boiling points of components contained in a diesel
sample. These peaks were identified using a mass spectral search program from the
National Institute of Standards and Technology (N IST, Version 2.0d, Gaithersburg, MD).
The peaks were integrated in order to determine their area, which corresponds to the

concentrations of the components present. The peak areas were then imported into

40

Microsoft Excel and normalized to the total area under the chromatographic curve, and
the individual RSD values were calculated using the mean and standard deviation for
each peak. For replicate analyses, peak areas should be consistent with less than 5%
relative standard variation [3 8].

PPMC coefficients were also used to gauge precision. In order to calculate these
coefficients, the aligned and normalized data spreadsheet was imported into Statistical
Analysis Software (SAS, Version 9.1, The SAS Institute, Cary, NC) and the default
internal PCA function was run. This default function performs PCA on the correlation
matrix of the data instead of the usual covariance matrix, and part of the output of the
function is a table of PPMC coefficients. This table was exported into Microsoft Excel
while the rest of the data was discarded. For smaller data sets (less than 15 samples), the
PPMC coefficients were also calculated in Microsoft Excel using the CORREL function.

For the actual association and discrimination of diesels based on their chemical
composition, the PPMC coefficients calculated for the precision determination were
utilized. In addition to PPMC coefficients, PCA was performed on the data sets, since
PPMC coefficients only allow a pairwise comparison, while PCA allows the full data set
to be assessed simultaneously. The data was aligned, normalized, and mean-centered
prior to PCA. The pre-treated data was then converted from spreadsheet form to a C SV
form and imported into a temporary workspace on the Michigan State University High
Performance Computing Center (HPCC) servers. From this workspace, Matlab was
accessed on a powerful cluster of computers from which the default PCA function (using
the covariance matrix) was performed on the data. Use of the HPCC was necessary due

to the large CSV file size generated from several replicate samples, each containing more

41

than 13,000 data points. Once PCA had been performed on the data, the scores,
eigenvalues, and first three eigenvectors were saved in ASCII format in the HPCC
temporary workspace. Only the first three eigenvectors could be saved due to storage
space complications. The first three PCs, though, typically describe 85-90% of the
variance of the sample set; thus, they are sufficient to allow association and
discrimination among the samples. Once saved, the PCA files were transferred back to a
personal computer where they were converted into a usable spreadsheet form in
Microsoft Excel. The scores values for PCI and PC2 for each sample were then plotted
using an x-y scatterplot in Microsoft Excel. For 3D scores plots, Origin Pro graphing
software (OriginLab, Version 8, Northampton, MA) was used to plot the scores values
for PCI, PC2, and PC3 for each sample.
3.5 Results and Discussion
3. 5. 1 Initial Data Set

Initially, the ten diluted diesel samples were analyzed in triplicate in order to
examine the reproducibility of the analytical method. RSD values were first calculated to
assess precision. For normal error distribution, a 5% error rate indicates that the
measurement falls within two standard deviations from the mean. Therefore, a threshold
of 5% RSD or less is typically considered acceptable for analytical precision [38]. Peak
areas that fall outside of the normally accepted range indicate imprecision in either the
instrument or the injection method.

Table 3.2 lists the RSD values for eight peaks across the retention time range
from the triplicate analysis of the ten diesel samples. Though some RSD values fall

below the normally accepted 5% error threshold, several do not. GC-MS is widely

42

 

Table 3.2 RSD Values for Eight Peaks Across the Retention Time Range
for Ten Diesel Samples Analyzed in Triplicate

 

Rt t'
”n 10“ Potential Peakldentity Diesell Diese12 Diesel3 Diesel4 Diesels Diesel6 Diesel7 Diese18 Diesel9 Diesel 10 Average

 

 

Time (min)
20.784 dodecane 7.7% 5.9% 5.7% 0.6% 6.9% 3.3% 12.7% 4.6% 22.1% 4.8% 7.4%
26.207 1’2’3’4‘tetrahydm'8' 5.9% 4.4% 13.3% 6.8% 2.1% 3.3% 9.6% 3.2% 13.2% 5.5% 6.8%
methylnaphthalene
27.064 tridecane 4.7% 2.9% 2.7% 2.2% 4.0% 0.7% 5.8% 2.5% 11.1% 0.4% 3.7%
31.782 2,6,10-trimethyldodecane 4.2% 15.6% 12.4% 11.9% 10.2% 6.0% 14.9% 11.7% 9.3% 8.5% 10.5%
33.251 tetradecane 3.4% 2.2% 2.4% 2.4% 6.9% 2.4% 6.8% 3.6% 3.9% 2.3% 3.6%
39.211 pentadecane 2.0% 7.9% 4.6% 2.1% 4.2% 3.5% 4.0% 4.4% 10.5% 1.4% 4.5%
44.905 hexadecane 3.3% 6.9% 2.8% 2.5% 3.9% 2.5% 12.2% 4.8% 14.4% 8.2% 6.2%
50.327 heptadecane 4.9% 5.8% 9.8% 3.5% 7.1% 2.7% 16.9% 4.9% 15.9% 4.2% 7.6%

 

43

 

 

known for its analytical precision, as its instrumental errors tend to be insignificant.
Therefore, it is more likely that the error observed in the RSD values is being introduced
in the manual injection process.

An alternative method to examine reproducibility is to consider PPMC
coefficients among replicates. These correlations offer some measure of the similarity
between two samples. For two replicate injections of the same sample, correlation
coefficients should be very close to one if both the instrument and injection method are
precise. Table 3.3 shows the combinations of PPMC coefficients for the replicate
analyses of the ten diesels. It should be noted that the correlation coefficients range from
0.7769 to 0.9886, with an average of 0.9581 :t 0.0519. While the minimum correlation is
low (0.7769) for a replicate analysis, the average is reasonably acceptable by normal
convention, in which a coefficient greater than 0.8 indicates a strong correlation [34].
But upon further consideration, this mean indicates that replicate injections of the same
sample are only 95.81% similar on average, when in fact a higher degree of correlation
should be attainable.

The unacceptable RSD values combined with poor correlation coefficients among
replicates suggest that the analytical methodology is not sufficiently precise. Because of
the well known precision of GC-MS, it is feasible that variation in the method of
injection is responsible for the lack of precision. Further evidence that the injection
procedure was imprecise showed in a preliminary scores plot from PC A of the ten
triplicate diesels. The three replicates of each diesel should have been tightly clustered;
instead, considerable spread was observed in the replicate clusters (data not shown). The

consistently poor precision of the data analysis results compelled a closer look at the

44

Table 3.3 PPMC Coefficients for Replicate Sets of
Ten Diesel Samples Analyzed in Triplicate

 

 

 

 

 

 

 

 

 

 

 

 

 

Diesel Sample Replicate Pair PPMC Coefficient
1 and 2 0.9815
Diesel 1 1 and 3 0.9748
2 and 3 0.9808
1 and 2 0.9768
Diesel 2 1 and 3 0.9824
2 and 3 0.9854
1 and 2 0.9540
Diesel 3 l and 3 0.9585
2 and 3 0.9877
1 and 2 0.7896
Diesel 4 1 and 3 0.7769
2 and 3 0.9868
1 and 2 0.9805
Diesel 5 1 and 3 0.9747
2 and 3 0.9845
1 and 2 0.9673
Diesel 6 1 and 3 0.9867
2 and 3 0.9664
1 and 2 0.9326
Diesel 7 1 and 3 0.9466
2 and 3 0.9688
1 and 2 0.9785
Diesel 8 1 and 3 0.9886
2 and 3 0.9881
1 and 2 0.8903
Diesel 9 l and 3 0.9465
2 and 3 0.9650
1 and 2 0.9761
Diesel 10 1 and 3 0.9797
2 and 3 0.9876
0.9581
Standard Deviation 0.0519

 

45

chromatograms of the replicate samples themselves. Figure 3.1 depicts two sections (9-
10 minutes and 50-51 minutes) of three overlaid chromatograms, all of which are
replicate injections of Diesel 9. The chromatograms have been aligned and normalized.
In the 9-10 minute range of the chromatogram (Figure 3.1a), the components in the
second replicate are most abundant, while those in the first replicate are the least
abundant. In the 50-51 minute range (Figure 3.1b), however, the opposite trend is
observed. This change in the composition distribution among replicates is surprising,
since component levels should remain constant throughout the entire range of the
chromatogram. It is possible that, as seen with the high RSD values and low PPMC
coefficients, an error in the manual injection method is causing this undesirable change in
composition distribution. Hence, further investigation of the injection method was
warranted.

Because the components found in diesels cover a very broad range of boiling
points (approximately 70-400 °C), their vaporization differs in the injection port. In a
normal injection technique, one pL of sample is drawn up into the syringe, which is
placed into the inlet, and the sample immediately injected. This rapid injection does not
allow the syringe needle to be sufficiently heated to the temperature of the inlet, which
causes the less volatile components to be insufficiently vaporized. For these reasons,
using the normal injection method for diesel samples introduces variability in the
distribution of the compounds present by preferentially loading the column based on
volatility. For example, the more volatile components in the diesel may be completely
vaporized and passed onto the column while the less volatile components are only

partially vaporized. The vaporization of all components is linked to both the current

46

-- Replicate 1 -- Replicate 2 — Replicate 3

 

 

 

 

 

 

 

 

50000 .
(a) 9-10 mmutes
8
E
a
'6
E
I
Q
<
. K l
0 “MA A A i v I . ' 4:. ‘

 

9.0 Retention Time (min) 10.0
150000

 

(b) 50-51 minute

 

Abundance

 

 

 

50.0 Retention Time (min) 51.0

Figure 3.1 Distribution of Select Peak Areas Among
Replicate Injections of Diesel 9

47

temperature of the injection port and the time the needle spends in the port, so that the
loading preference varies with each injection as demonstrated in Figure 3.1. To
overcome the volatility bias, the injection procedure was amended to draw one uL of air
first, then one 11L of sample, followed by another 11L of air. These surrounding air
pockets protect the plug of sample from the heat of the injection port until it is actually
injected. The injection procedure was also changed to include a two second delay after
the insertion of the needle in the inlet prior to actual injection. This timed injection
allows the needle of the syringe to come to temperature without beginning to vaporize
any of the sample, thus ensuring reproducible vaporization.
3. 5.2 Data Set using Timed Injection

After it was determined that the original method of injection resulted in
unacceptable precision in the first data set, five of the diesels were re-analyzed in
triplicate using the timed injection method described above. The data was pre-treated and
analyzed in the same manner as before in order to investigate the precision of the
modified injection method.

The RSD values for the data from the timed injection method are listed in Table
3.4. For the same eight peaks that were examined previously, the RSD values were
improved for the timed injection compared to those from the previous injection
procedure. Only the RSD value for the 2,6,10-trimethyldecane peak at 31.782 minutes
exceeded the 5% error threshold in all five diesels, whereas for the previous injection
procedure, five of the eight peaks had unacceptable RSD values.

In addition to RSD values, PPMC coefficients were also calculated. Table 3.5

lists the correlations among replicates for the five diesels. The range of correlation

48

 

 

.32 .33 .342 .32 333 .32 2488232 Room
33am .32 3mm .33 .32 .32 6:88.28; 83.3.
.33 33mm .32 333.3 .3: .33 6:333:63 2 5.3
.393. 333 3E .32 .33 .33. 288832 EN?
.3: 33:: .33 .333 ..3o.3 333.3 memoaeoeiaeeewea Se. 3 m
.32 .32 333. .32 .33 .33. 23823 3.8.8
.33.... .32 .33 .32. .33 .33. 8235233353565 SEN
336335.13;
.33 .32 33% .33 .33 .33 888% 433.8
«aw—954 m Emu—G v Ema—G m .935 N .3105— — 33.: 55:93 :3.— EE—oeom 3:5 25,—.
. . . . . . . Season

 

9.592 558.2: con—E. « ”Em: Snow—AF; E poi—«.3. moan—am .085 5B. you
owns— ofifi. not—.33— 05 29.34 $133— Ewfi— .8.— 823» 9mm in 03:.

49

Table 3.5 PPMC Coefficients for Replicate Sets of Five Diesel Samples
Analyzed in Triplicate Using a Timed Injection Method

 

 

 

 

 

 

 

 

Diesel Sample Rgrlicate Pair PPMC Coefficients
1 and 2 0.9884
Diesel 1 l and 3 0.9870
2 and 3 0.9878
1 and 2 0.9898
Diesel 2 1 and 3 0.9852
2 and 3 0.9831
1 and 2 0.9870
Diesel 3 1 and 3 0.9659
2 and 3 0.9599
1 and 2 0.9821
Diesel 4 1 and 3 0.9700
2 and 3 0.9445
1 and 2 0.9850
Diesel 5 1 and 3 0.9721
2 and 3 0.9846
Average 0.9782
Standard Deviation 0.0131

 

50

values among replicates, 0.9445 to 0.9898, is narrower compared to that observed in the
data collected using the previous injection procedure. The average correlation value,
0.9782 3: 0.0131, indicates that the timed injection generally provides improved precision
throughout replicate analysis. However, an average correlation coefficient of 0.9782 is
still lower than would be expected among replicates. Although the timed injection did
improve the precision of replicates, any manual injection carries the opportunity for
analyst error.
3. 5.3 Improving Abundance Levels of TIC and EICs

Once the precision of the injection method was established, further data analysis
was performed in an effort to begin associating and discriminating the diesels based on
their chemical compositions. However, through the course of this data analysis, it
became obvious that the level of association and discrimination demonstrated in the
PPMC coefficients and PCA results was being limited by the relatively low abundances
of the chromatograms. The TICs, though low in comparison to abundances achieved
through standard crime lab protocols (~106-107) [39], were at least sufficient for analysis
(~104-105), but the real limitation occurred in the data analysis of the EICs. Because ions
of a single m/z value were extracted from an already low TIC, it became difficult to
differentiate peaks in the EIC from the noise level in the background. When abundances
are significantly high, the variation in the baseline is insignificant, but at these low levels,
any variation can create artifacts in the subsequent data analysis. In order to achieve
meaningful results from the PPMC and PCA, especially from the EICs, as well as
potentially improving further upon the precision of the technique, adequate abundance

levels must be reached.

51

In order to improve the abundance levels in the chromatograms, two steps were
taken. First, the dilution factor was decreased from 200:1 to 50:1 to prepare the diesel
samples for GC-MS analysis. This more concentrated sample yielded TICs
approximately five times more abundant than the previous 200:1 dilution. These
increased abundances in the TIC allowed for more abundant EICs, though some amount
of variation was still discernible in the baseline. Therefore, the second step to improve
abundances was taken specifically for the EICs. The particular EICs were chosen in
order to represent compound classes that are characteristic of petroleum distillates. It was
determined that instead of using only a single m/z value from each compound class,
multiple values from each characteristic class could be used to improve the signal-to-
noise ratio for the extracted ions. The many individual ions from each class were
summed to create extracted ion profiles (EIPs) for the specific compound classes. The
following ions were summed to create the EIPs for their respective classes: alkane — 57 +
71+ 85 + 99; aromatic— 91 + 105 +119 + 133; indane— 117 +131 +145 + 159; olefin-
cycloparaffin (OCP) - 55 + 69 + 83 + 97; and polynuclear aromatic (PNA) — 128 + 142 +
156.

The EIPs offer a significant amount of additional information in comparison to a
single EIC. Table 3.6 illustrates this fact by showing PPMC coefficients between single
extracted ions from the same compound class and their summed EIPs for the alkane.
aromatic, and indane profiles. The correlations among the individual alkane ions, as well
as the correlations among the individual ions and the summed EIP, are all relatively hi gh,
which indicates that few differences exist between them and additional information is not

obtained using the alkane EIP. On the other hand, the correlations for the individual

52

Table 3.6 PPMC Coefficients for Individual m/z Values and Their

Summed EIPs for the Alkane, Aromatic, and Indane Profiles

 

 

 

 

 

 

 

 

 

Alkane Profile m/z Values
57 71 85 99 EIP
57 1.0000 0.9917 0.9876 0.9689 0.9971
71 1.0000 0.9897 0.9758 0.9972
85 1.0000 0.9794 0.9949
99 1.0000 0.9797
EIP 1.0000
Aromatic Profile m/z Values
91 105 119 131 EIP
91 1.0000 0.4333 0.2943 0.2558 0.7288
105 1.0000 0.2255 0.1249 0.8040
119 1.0000 0.2875 0.6470
131 1.0000 0.4171
EIP 1.0000
Indane Profile m/z Values
117 131 145 EIP
117 1.0000 0.3990 0.2892 0.7160
131 1.0000 0.2846 0.7599
145 1.0000 0.7429
EIP 1.0000

 

53

aromatic ions and the summed EIP are very low, which suggests that the EIP contains
more information than the single ions themselves. The same trend is observed for the
individual indane ions and the corresponding EIP. These differences confirm that
additional information concerning chemical composition is garnered from the EIP in
comparison to a single EIC. These two final improvements to the method resulted in a
sufficiently optimized analytical methodology and data analysis procedure.
3. 5. 4 Final Neat Data Set

The modifications made to the experimental and data analysis methods
throughout the course of the preliminary trials were taken into account when a final neat
diesel data set was analyzed. By applying the improved method, any random variations
or artifacts in the results of the PPMC calculations or the PCA would be negligible. The
final neat data set consisted of all ten diesels, previously described in Table 3.1, re-
analyzed in triplicate. The data was compiled for the TICs, as well as the EIPs for the
five characteristic compound classes (alkanes, aromatics, indanes, OCPs, and PNAs).
For the TICs, RSD values were calculated of the same eight peaks as before. For the TIC
and each EIP, PPMC coefficients were calculated and PCA was performed.
3.5.4.1 RSD Values

Table 3.7 shows the RSD values for the ten diesels analyzed in triplicate. As
previously, RSD values for all eight peaks except the 2,6,10-trimethyldecane peak at
31.782, were lower than 5%. In fact, the RSD values for the less dilute samples were, on
the whole, even lower than before. It is possible that the shape and size of the 2,6,10-
trimethyldecane peak is more variable than the other peaks investigated because it is a

small, shouldered peak, which would explain RSD values greater than 5%.

54

Table 3.7 RSD Values for Eight Peaks Across the Retention Time Range
for Final Neat Diesel Data Set Analyzed in Triplicate Using a Timed Injection Method

 

 

 

132:?222) Potential Peak Identity Diesel 1 Diesel 2 Diesel 3 Diesel 4 Diesel 5 Diesel 6 Diesel 7 Diesel 8 Diesel 9 Diesel 10 Average
20.784 dodecane 2.6% 3.7% 1.2% 2.8% 5.2% 1.6% 4.2% 3.7% 4.7% 7.2% 3.7%
26.207 1’213’4'tetrahydm'8' 1.6% 0.6% 1.6% 3.2% 6.1% 2.6% 2.0% 2.2% 1.5% 3.3% 2.5%

methylnaphthalene
27.064 tridecane 3.3% 1.7% 0.7% 2.1% 0.4% 0.5% 1.7% 0.5% 0.7% 0.4% 1.2%
31.782 2,6,10—trimethyldodecane 20.0% 8.5% 1.6% 5.8% 1.7% 0.7% 11.8% 2.3% 4.6% 2.7% 6.0%
33.251 tetradecane 2.1% 0.7% 0.6% 2.4% 2.1% 1.0% 2.5% 1.4% .3% 1.1% 1.4%
39.211 pentadecane 1.5% 0.8% 0.8% 1.6% 2.3% 0.7% 2.3% 1.5% 4.2% 1.0% 1.7%
44.905 hexadecane 1.3% 1.2% 0.4% 3.3% 2.2% 0.6% 4.1% 0.3% 2.1% 3.5% 1.9%
50.327 heptadecane 5.6% 0.5% 1.2% 8.2% 5.5% 0.2% 7.6% 1.8% 3.5% 2.5% 3.7%

 

55

 

 

3. 5. 4.2 PPMC Coefficients

The PPMC coefficients for triplicate analysis of the ten diesel samples were
utilized to determine the range of correlations that corresponds to diesels of known
similar origin. Correlation coefficients between two samples of different origin must fall
outside this range in order to be considered statistically distinguishable from the
replicates. Table 3.8 lists average correlation coefficients and observed ranges among
replicates for the TICs and each of the EIPs. It should be noted that correlations are
highest among replicates in the TICs (0.9895 :t 0.0072), which may suggest that its
potential for discrimination is higher than that of the individual EIPs. The average
correlation coefficient for the TICs is higher than the previous correlation coefficient for
the TICs (0.9782 i 0.0131), which indicates that the amendments to the injection method
and the dilution factor improved precision. The average correlation coefficients for the
EIPs are lower than that of the TIC, which could be caused by a number of factors. It is
possible that the abundance levels are still problematic though they have improved. From
later data analysis, it became clear that the lower correlations for the ElPs are, however,
more likely due to errors in the alignment algorithm, as discussed in the next section.

The average and standard deviation of correlation coefficients among replicates
serves as a benchmark in the interpretation of PPMC coefficients for samples that are not
of similar origin. Using these values, a 90% confidence interval (CI) can be calculated
for the means of the PPMC coefficients for replicate samples as well as for sample pairs
of different origin in order to assess statistical similarities among the diesels [27]. A
confidence level of 90% was selected instead of the more conventional 95% due to the

small sample size. If the 90% CI calculated for the PPMC coefficients for different

56

Table 3.8 PPMC Coefficients for TIC and EIPs
for Replicate Sets of Ten Diesel Samples (Final Data Set)

 

Standard

 

Data Set Average . . Minimum Maxium
Devratron

TIC 0.9895 0.0072 0.9649 0.9963
Alkane 0.9824 0.0146 0.9299 0.9945
Aromatic 0.9791 0.01 15 0.9376 0.9922
Indane 0.9690 0.0171 0.9157 0.9881
OCP 0.9631 0.0142 0.9271 0.9778
PNA 0.9513 0.0389 0.8163 0.9884

 

57

diesels overlaps the CI for the replicate samples,then those samples are considered the
same statistically at a 90%confidence level. If the CI does not overlap that of the
replicates, then the samples are considered statistically different. The CI for the PPMC
coefficients is more useful than the empirical range of correlation coefficients for all
replicate measurements. The minimum correlation coefficient calculated between two
replicates for the TIC, for example, is 0.9649, which is well outside the defined
confidence interval (09800-09989). With the opportunity for occasional disparity in the
manual injection, outliers, though rare, have to be anticipated.

Using the confidence intervals for each data set, the PPMC coefficients for
samples of known different origin could be assessed for any potential association and
discrimination capabilities. The complete PPMC tables for each data set are shown in
Appendix B. Again, for some sample pairs, such as Diesels 6 and 8, the correlation for
the TICs (09875-09944) is similar statistically to that of the replicates (09800-09989),
which indicates that those two samples are not statistically different. However, Diesels 2
and 3 (08859-08976) are less correlated than the replicate TIC samples, and those are
statistically different. It should be noted that the range of correlation coefficients
observed are typically wider for the EIPs. For example, the minimum correlation among
the TICs is 0.8668. While the minimum for the alkane EIP is similar (0.8727), the
minima for the other EIPs are considerably lower (aromatic —- 0.7440, indane — 0.7594.
OCP — 0.8380, PNA — 0.6675). This wide range of correlation coefficients may indicate
that these compound classes are more discriminatory than the TIC as a whole. However.
the average correlation coefficients for replicates based on these EIPs are much lower

than that of the TIC, which limits the statistical discrimination capabilities to those

58

samples that fall below that decreased range. These two contradicting properties place
some amount of question on the validity of the application of PPMC coefficients to ElPs.
If still low abundance levels or misalignments are indeed the sources of error for the
EIPs, then it is possible that they can be corrected and more reliable information can be
gained from the EIPs. Currently, however, any data interpretation is limited to the lower
average correlation thresholds established by the replicates.
3. 5. 4.3 PCA

PCA was also performed on the final data set using the TIC and each EIP. The
scores plots for the TIC are shown in Figure 3.2. In Figure 3.2a, PCl is plotted against
PC2, where PCl accounts for 35% of the variance, and PC2 31%. This plot affords a
moderate amount of discrimination among the samples while still maintaining tight
clustering of replicate samples. Diesels 1 and 2 are separated from the rest of the large
cluster of samples in PC2, as PC1 appears to only distribute the samples while exhibiting
very little discrimination. The separation afforded in PC2 indicates that explicit
differences in chemical composition contribute to the variance described by that principal
component, whereas the even distribution in PCl indicates that a uniform range of
variables contributes to its described variance. However, even within PC 1 , some amount
of sample discrimination can be determined. For example, although Diesels 5 and 6 fall
within that large group, they are significantly different in the first principal component.
Other samples, however, such as Diesels 9 and 10, are very close to one another within
that large grouping, so it can be reasonably concluded that these samples are not
sufficiently distinguishable. With the addition of PC3 in the three-dimensional (3D) plot.

firrther discrimination is observed. The third principal components accounts for 16% of

59

A Diesel 1 ADiesel 2 A Diesel 3 Diesel 4 A Diesel 5
ADiesel 6 3. Diesel 7 ADiesel 8 A Diesel 9 A Diesel 10

 

 

 

       

1500000 —
AA
A
r; AA
9\
a l A A A l
m'leOOOO :5. EA .1.“ 1200000
A
-1500000 -
(a) - PC1(35%)
I 000000
.3
3
I
E’- 0
1500000
0
9
CI] o[o\
0 0 \
60/11} 101-9
(b) 1000000

Figure 3.2 Scores Plot - TIC (a) PC 1 v. PC2 and (b) PCl v. PC2 v. PC3

60

the variance, for a total of 82% described by all 3 PCs. Diesels 1 and 2 are still separated,
but the addition of PC3 allows for differentiation within the larger cluster observed in the
two-dimensional (2D) scores plot. Diesels 5, 6, and 8 are moderately separated in PC3.
On the other hand, Diesels 3 and 7 remain closely associated with one another, as do
Diesels 4, 9, and 10. The diesels in these two clusters are not distinguishable from one
another, and thus are considered to be very similar in their TIC content.

Additional information about how the individual diesel samples cluster can be
garnered from the scores plots of the five different EIPs. The scores plots for the alkanes
are shown in Figure 3.3. The 2D plot, in which PCs 1 and 2 describe 45% and 12% of
the variance respectively, does not offer any further discrimination than the TIC. While
Diesels 1 and 2 are still reasonably separated in both PC] and PC2, the rest of the
samples are scattered amongst themselves. Even the replicates are less tightly clustered,
as PC2 tends to spread them. For example, the replicates of Diesels 4 and 7 are so widely
spread, it would difficult to even associate them with one another. When PC3 is also
included for the 3D plot, an additional 9% of the variance is described, for a total of 66%.
PC3 does nothing to further associate or discriminate samples. In fact, it causes
additional spread in the replicate diesels. Similar trends are observed in both the 2D and
3D scores plots for the indane and PNA profiles, which are shown in Figures 3.4 and 3.5.
Diesels 1 and 2 are still somewhat separated from the main cluster, but the spread in the
replicates is even worse (again, likely due to misalignments in the EIPs). It is possible
that the diesel samples do not vary greatly in their alkane, indane, or PNA content, which
would explain the lack of discrimination in the scores plots. A brief visual comparison of

overlaid chromatograms for different diesels for each profile indicated that some amount

61

A Diesel 1 A Diesel 2 A Diesel 3 Diesel 4 ADiesel 5
A Diesel 6 .11. Diesel 7 A Diesel 8 A Diesel 9 A Diesel 10

 

 

 

  
  

500000 -
g: A A
V r— w—%' A ' 1
N i ‘ 11_
2500000 ‘A a“ f 500000
(0)
-500000 -
PC1(45%)
200000
3
e 0
8
D-
-20000()
-60(l()00
3,!
‘6‘ I)
1."
1.0
(b) -300000 YC

Figure 3.3 Scores Plot - Alkane EIP (a) PC 1 v. PC2 and (b) PCI v. PC2 v. PC3

62

ADiesel l ADiesel 2 ADiesel 3 »“ Diesel 4 A Diesel 5
ADiesel 6 A Diesel 7 ADiesel 8 ADiesel 9 ADiesel 10

 

 

100000 1
l A
k l
3° 45““ A
e; .3, 2, I 2 A; -32., ,2. A 3, ”AL .2.
8 ‘ A‘A‘A‘.-.
”250000 [ A 250000
A I
A‘ A
((1)
-100000
PCI (58%)

100000
:3
t:
n
o
D:

-100000

fl

C},-

33, 0
’9 °/)

(b) 100000 -100000 v00 °

Figure 3.4 Scores Plot - Indane EIP (a) PC 1 v. PC2 and (b) PC 1 v. PC2 v. PC3

63

ADiesel 1 A Diesel 2 A Diesel 3 Diesel 4 A Diesel 5

 

60000

 

ADiesel6 iDiesel7 ADiesel8 ADiesel9 ADiesel 10
50000 1
A
3 i ‘A A A
E A 11 A A A
r. . A A A '
2-60000 A
A A
(a) A
A -50000 —
PC] (46%)
50000

pc3 (17%)
o

    
   

-50000
-50000

0
9It"
1. 01
(b) 50000 -50000 Yc

Figure 3.5 Scores Plot - PNA EIP (a) PC] v. PC2 and (b) PCl v. PC2 v. PC3

64

of variation is noticeable, while some areas are very similar. It is likely that the
differences are not significant enough to be isolated by PCA, which would imply that
these compound classes are not useful for determining association and discrimination
among the diesels.

The other two compound classes, however, do provide discrimination of the
diesels in their scores plots. The scores plots for the aromatic profiles, Figure 3.6, show
tight clustering for replicates. In the 2D plot, Diesels 1 and 2 are still isolated from the
main cluster, but in this case, that main cluster is not spread about the first principal
component as it was in the TIC. Instead, the eight other diesels are very tightly clustered.
which suggests that they are all very similar in their aromatic content. The 3D plot
demonstrates the same trend, though PC3 spreads the tightly clustered eight diesels, as
Diesels 5 and 7 begin to be separated from the larger group. The differentiation in PC 3 is
difficult to discern, however, it only describes a small amount of the total variance (3%)
in comparison to PCs I and 2, 75% and 8% respectively. The scores plots from the OCP
profile, shown in Figure 3.7, afford a considerable amount of discrimination than the
other profiles. Replicates are still closely associated in the 2D plot, with the exception of
Diesels 2 and 8, which are spread in comparison to the other replicates. This spread
could be caused by a number of different factors, though the most likely source is
possible misalignment. Individual diesel samples, most notably Diesels 1 and 2, are
beginning to be isolated in the 2D plot as well, though it should be noted that their
separation is only due to 48% of the total variance of the data. The 3D plot, which
includes an additional 10% of the variance, exhibits similar traits to the 2D plot.

Replicate spread worsens in PC3, though on a minor level, some diesels are more

65

A Diesel 1 A Diesel 2 A Diesel 3 A Diesel 4 A Diesel 5
A Diesel 6 A Diesel 7 A Diesel 8 A Diesel 9 A Diesel 10
350000 -

(8%)
D
4"

 

 

 

 

 

 

 

 

m"-350000 ‘A 350000
(11)
3500001
PC1(75%)
100000 ,1
A
A
. ‘ t‘
O
‘3 ° ’ 4 "at:
8 A
m J
“‘ 100000
A
o \"
400088 ’ 5°.
- 000 a.
(b) 0 0)) A
P C1 (75%) 100000 '1 ”00

Figure 3.6 Scores Plot - Aromatic EIP (a) PC] v. PC2 and (b) PCl v. PC2 v. PC3

66

ADiesel l ADiese12 ADiesel 3 Diesel 4 ADiesel 5
ADiesel 6 Diesel 7 ADiesel 8 ADiesel 9 ADiesel 10

 

 

150000-—
§ 1 ‘11 r A
g 1——% Aa’i I
N A
0 -150000 A 150000
On A A
A A
A
a
() -150000 —
PC] (26%)
100000

PC3 (I 00/0)
o

 

      

- I 00000
- 120000
46
9r I 00000
O) 0
(b) v01 K“

- I 00000

Figure 3.7 Scores Plot - OCP EIP (a) PCl v. PC2 and (b) PC] v. PC2 v. PC3

67

separated. For example, in the 2D plot, Diesels 4 and 8 are closely associated, but in the
3D plot, they are more separated in PC3. Based on these results, the aromatic and OCP
profiles are more useful in associating and discriminating diesels than the alkane, indane.
and PNA profiles.

In addition to the scores plots, PCA also provides information about which
components in the samples contributed most to the determination of each principal
component. This eigenvector, which can be displayed in a loadings plot, determines how
the score of a sample is calculated. The loadings plots for the first and second principal
components of the TICs are shown in Figure 3.8. It is clear that, while the alkane
components dominate the loadings for the first PC (Figure 3.8a), a mixture of compounds
from other characteristic classes is responsible for separation along the second PC (Figure
3.8b). The fact that the alkanes dominate PCl does not necessarily indicate that they are
the primary source of variation among the diesels. In fact, the diesels are not clearly
separated along PC] in the TIC as shown in Figure 3.2; they are relatively evenly
distributed by their differences in alkane content. The lack of sufficient clustering of the
diesels in the scores plot of the alkane profiles also indicates that the alkanes alone are
not adequate for associating and discriminating the samples. The presence of
components other than alkanes at lower levels in PCI and at higher levels throughout
PC2 implies that some combination of aliphatics and non-aliphatics is responsible for the
discrimination among the diesels demonstrated in the scores plots for the TIC (Figure
3.2).

The loadings plot is not only useful for determining the components contributing

the most to the eigenvectors, it can also highlight any anomalies caused by the analytical

68

Principal Component 1

 

0.20
(a)
0.15

0.10‘

0.05 *

 

 

 

 

 

 

 

0.00 ~~

 

 

 

-0.05
Retention Time (min)

Principal Component 2

 

0.10

(b)

0.05 --

 

 

 

 

 

 

0.00m w. .. i 1.11 1

 

 

 

 

 

 

 

-0.05
Retention Time (min)

Figure 3.8 Loadings Plots of First Two Eigenvectors from PCA of the
TICs of Ten Diesels Analyzed in Triplicate

69

technique or the data pre-treatment steps that can potentially contribute to the sources of
variation observed in scores plots. For example, in the loadings plot for the first two PC 5
of the alkane EIP, shown in Figure 3.9, several derivative-shaped curves are noticeable.
In fact, these derivative-shaped peaks dominate the second eigenvector (Figure 3.9b).
Not normally observed in loadings plots, these peaks are indicative of an artifact in the
samples. Because these peaks are not observed in the loadings plots for the TIC s, it is
possible that the artifact causing them has been introduced in the data pre-treatment steps.
The most likely source of error is in the alignment procedure. Figure 3.10 overlays three
replicate chromatograms of Diesel 1 around the peak with a retention time of
approximately 33 minutes, which corresponds to derivative-shaped curves in the first two
PCs as shown in Figure 3.9. It is obvious that this peak has been misaligned for this
sample. Further inspection revealed that this peak has been misaligned throughout the
alkane EIP data set, and this introduction of error in the data pre-treatment stage has
propagated through to the data analysis results. The increase in range for the correlation
coefficients of the replicates, as well as the spread observed among the replicates in the
scores plots, can most likely be attributed to significant misalignments. It can be
concluded that the alignment algorithm is by no means perfect, and misalignments must
be considered when the PPMC and PCA results are being interpreted.

In the analysis of a neat diesel data set, several obstacles were encountered and
overcome. A precise manual injection method was developed, and sufficient abundance
was achieved for accurate PPMC coefficients and PCA results. It was then demonstrated
that PPMC coefficients and PCA were sufficient for the association and discrimination of

neat diesels based on chemical composition.

70

Principal Component 1
0.20 2

(a)

 

0.10

 

 

 

 

 

Lil _ .11

E

0.00 i -

 

 

-O. 10 i
Retention Time (min)

Principal Component 2

 

 

 

 

 

 

 

 

 

 

 

0.20
(b)
0.10 a
0-00 We LL WWL—mwj T r s .
-0.10 ‘-
-0.20

 

Retention Time (min)

Figure 3.9 Loadings Plots of First Two Eigenvectors from PCA of the
Alkane EIPs of Ten Diesels Analyzed in Triplicate

71

200000

Abundance

0

33.0
Figure 3.10 Misaligned Peak in the Alkane EIPs of Diesel 1 Replicates

 

 

'F—'

 

 

 

 

Retention Time (min)

72

33.5

Chapter 4
Association and Discrimination of Burned Diesels Extracted
from Fire Debris Using PPMC Coefficients and PCA
4.1 Introduction

The substantiation of the principle that diesels can be associated and
discriminated based on variation in chemical composition by chemometric procedures
(PPMC coefficients and PCA) was demonstrated in Chapter 3. These procedures,
however, possess the potential for applications past the association and discrimination of
neat diesels alone. The next step in the research then is to investigate the associative and
discriminatory potentials of these chemometric procedures for diesels that have been
burned and extracted from fire debris. These conditions are more similar to those
encountered in an actual arson case, and are therefore more practical in a forensic
laboratory setting.

When ignitable liquids are utilized as accelerants, they are often poured
throughout the structure and then ignited in order to start the fire and guarantee its spread.
The ignitable liquid can be affected in many ways both during the fire as well as while
the fire is being extinguished. While much of the ignitable liquid is consumed in the
burning process, it can also be absorbed by the substrate on which it was poured, where it
remains protected to some extent from the heat of the fire. More porous substrates, such
as wood and carpet, have a greater potential for retention of ignitable liquids, and these
substrates are frequently sampled in fire debris analysis.

Although the ignitable liquid can be absorbed by porous substrates and shielded

from the fire, the intense heat is still capable of causing the loss of the more volatile

73

components present in the ignitable liquid. For these reasons, the chromatograms
obtained from the GC-MS analysis of ignitable liquid residues (ILRs) extracted from fire
debris are often skewed toward the less volatile components—that is, the chromatograms
contain more of the less volatile components and little to none of the more volatile
components. This distortion in chemical composition may lead to a misidentification of
an ignitable liquid when a visual assessment of chromatographic patterns is performed.
However, the application of chemometric procedures such as PPMC coefficients and
PCA to associate and discriminate the burned ignitable liquids is more objective in
nature, which could potentially overcome the subjectivity issues of a visual assessment.
These procedures can be performed on not only the TIC, which may potentially be too
skewed to be associated to its neat counterpart, but also on the ElPs of the characteristic
compound classes of ignitable liquids. The classes that consist of mostly low volatile
compounds may be more discriminatory than the TIC or other compound classes that are
more significantly affected by the burning process.

In order to investigate the potential for associating a burned diesel extracted from
fire debris to its unburned counterpart, a complete set of burning experiments was
performed on a newly collected set of diesels. Since a new instrument was used for this
part of the research, the precision of the instrument as well as the injection method was
re-assessed with the analysis of a new set of neat diesels. These diesels were analyzed in
triplicate to examine the association and discrimination observed among the new sample
set. Prior to conducting the burning experiments, Diesel 21, chosen at random, was
spiked onto three different matrices (cotton cloth, magazine, and carpet) and a series of

spike and recovery studies were performed to examine the efficiency of the solvent

74

extraction procedure selected. Then, Diesel 21 was subjected to a series of burning
conditions on the three different matrices. For each matrix, five analyses were performed
in triplicate: unburned matrix, burned matrix, unburned then spiked (50 uL) matrix,
burned then spiked matrix, and finally a spiked then burned matrix to simulate actual
arson conditions. The final phase was to perform a blind association study in which two
diesels that had been previously analyzed were treated as unknowns. The blind diesels
were spiked separately onto each of the three matrices and then burned. The debris was
extracted and analyzed in an attempt to associate the blind diesel sample to its neat
counterpart. PPMC coefficients and PCA were used to associate and discriminate the
diesels, both burned and unburned, based on chemical composition.

4.2 Analysis of New Set of Neat Diesels

4. 2.1 Procedure

Five of the previous ten diesel samples (Table 3.1) were re-collected from various
Lansing service stations during the winter. Table 4.1 lists the samples and the details of
their collection. After collection, the diesels were stored as described previously in
Chapter 3.

The five neat diesels were analyzed in triplicate by GC-MS, using a newly
purchased instrument. The GC was the same make and model number (Agilent 6890
GC) as previous, while the MS was a newer version of the previously used MS (Agilent
5975 MSD compared to Agilent 5973). After initial analyses of the freshly collected
diesel samples, it was observed that, in order to obtain chromatograms of similar
abundances to before, a different dilution (10:1 in CHzClz instead of 50:1) of the neat

diesels was necessary. The higher abundance levels in the TIC are essential in order for

75

 

 

BozoNA 308825 w? _ Q 3 Rod 323: com M .502 mm E35
26:?» Eoomocosm on mho ozco>< Bid 285 $3 .333on um Emma
26:0» Eoomocosc waQS osco>< 32¢ 285 3.03 mmom 8:02 mm 385
>>o=oNA 2nd mokmio 93% $0355 ooom 895m mm .035
25:0NA 2838:: M55: 5 80M ago—mm: mm 5 895m _N .035
553m
cacao—EU we 25,—. as .830 533—30 we 8.5 553::— 3_>..om o-QEam

 

”3 559:5 .32: E. as;

76

the peaks present in the EIPs to be sufficiently distinguished from baseline noise. The
sensitivity discrepancies between the two instruments are most likely due to minor
discrepancies in the mass spectrometer, such as the position of the column in the
ionization source or the voltages of the accelerating lenses. It was also concluded that the
previous temperature program used was too long (108.33 minutes) to be practical in a
forensic laboratory setting. The temperature program employed by the National Center
for Forensic Science (NCFS) and the Technical Working Group for Fire and Explosions
(TWGFEX) to generate the Ignitable Liquid Reference Collection (ILRC) was used
instead [40]. This temperature program (NCFS.M) was as follows: initial temperature
50 OC, hold for 3 minutes, 10 °C/min ramp to 280 °C, hold for 4 minutes. The total run
time was reduced to 30 minutes. All other instrumental parameters were the same as
described in Chapter 3.

Preliminary assessments of the chromatograms, both of the TIC and each EIP,
indicated that the change to the temperature program did not noticeably sacrifice any
chemical features. The increase to the ramp rate, however, offered a significant
improvement in total analysis time which, in turn, allows for the methodology to be
practical in the traditionally backlogged setting of an operating forensic laboratory. This
compromise between improved analysis time and retention of discriminatory chemical
features is a necessary consideration in the development of this methodology.

Once the neat diesel data were collected and analyzed, the TIC and the ions used
to generate the five EIP sets were extracted into Microsoft Excel as described in Chapter
3. The shorter temperature program, however, eliminated the need for truncating the end

of the chromatograms. For the diesels analyzed by the NCFS.M method, each

77

chromatogram consisted of a total of 4702 data points (compared to 13,000 data points
previously), all of which were considered in subsequent data analysis. The data were pre-
treated as described in Chapter 3. RSD values and PPMC coefficients were also
calculated for the TICs in order to demonstrate the precision of the new GC-MS
instrument. The TIC and EIPs were also analyzed by PPMC coefficients and PC A in
order to evaluate the association and discrimination of the neat diesels based on chemical
composition.
4. 2.2 Results and Discussion

Initially, RSD values were calculated to demonstrate the precision of the
instrument. Table 4.2 lists the average RSD values for eight peaks throughout the
chromatographic range for triplicate analyses of the five diesels. For the same eight
peaks as selected in Chapter 3, average RSD values were well below the accepted
threshold of 5% [3 8]. For previous average RSD values for the timed injection, one of the
eight peaks still exhibited an RSD of greater than 5%. The lack of precision for that peak
has been eliminated in this data set. It should be noted that the RSD values for Diesel 21
are higher than those of the other samples, which could indicate an aberration in the
timed injection procedure for one of the replicate analyses for Diesel 21. However, all
the average RSD values were below 3% which indicates that the instrumental technique
and the injection method are precise.

PPMC coefficients were also calculated to assess precision among replicate
analyses. Table 4.3 lists the PPMC coefficients for the replicate analyses of the five
diesels based on the TIC. Correlation coefficients for the replicate analyses ranged from

0.9780 to 0.9975, with an average of 0.9921 i 0.0061. This correlation is higher than

78

 

 

:2. £2 £2 :3 £3 $3 88833; w 5.:
xi :2 £3 :33 $3 £3. 888st 5.2
$3 $3 :2 £3 :3 :3 858368 m3:
:3 as: .x. _ .o $3 :2. 3% assuage 8v:
:2 $2 :3 $2 s2 :3 mascogéeaséeqm 2 a:
as: :2 :2 :3 :3 £2 05825 Ed
saw $3 :3 some as? sow osoafisegsae $0.:
-waifieeeafi
:3. $3 :2 :2 as? $3 03088 83:
owauo>< mm .895 vu .329 mm 385 NN 385 mm 335 @553 :3“— _a=:3om AEEV 25H.
. . . . . . . 2:53.

 

83:3; 5 c-ba=< main—am .33: 25— he

omnam 08:. 55:33— 2.. 233.. 9.3m Ewa— .8._ 335» Gwy— Né 035,—.

79

Table 4.3 PPMC Coefficients for the TIC of Replicate Sets

of Five Diesel Samples Analyzed in Triplicate

 

 

 

 

 

 

 

 

Diesel Sample Replicate Pair PPMC Coefficients
1 and 2 0.9941
Diesel 21 l and 3 0.9962
2 and 3 0.9975
1 and 2 0.9879
Diesel 22 1 and 3 0.9841
2 and 3 0.9899
1 and 2 0.9908
Diesel 23 1 and 3 0.9836
2 and 3 0.9780
1 and 2 0.9969
Diesel 24 l and 3 0.9967
2 and 3 0.9969
1 and 2 0.9955
Diesel 25 1 and 3 0.9973
2 and 3 0.9956
Average 0.992 1
Standard Deviation 0.0061

 

80

that of a previous data set run on a different instrument (0.9895 :1: 0.0072) as described in
Chapter 3. This high correlation indicated that the new instrument was precise.

As discussed in Chapter 3, 90% confidence intervals (CI) for the average PPMC
coefficient for replicate analyses for the TICs and EIPs were used as the threshold for the
comparison of samples from different origins [27]. If a CI for a diesel pair of different
origin overlaps with the CI defined by the replicates, then that pair would be considered
statistically indistinguishable. On the other hand, if the CI for that pair did not overlap
the CI as determined by the replicates, then the diesel pair could be considered
statistically distinguishable. The average correlation coefficient and respective standard
deviations are given in Table 4.4 for the TIC and each EIP set for triplicate analyses of
the five diesels. It should be noted that, on the whole, a higher level of association
among replicates for the EIPs is observed as compared to previous data sets (Chapter 3).
This narrow range for replicates improves the discrimination capability afforded by each
EIP.

Data analysis was initially performed on the neat diesel data set in order to
determine the sources of variance in chemical composition among the new set of five
diesels. PPMC coefficients were calculated for all diesel pairs in the TIC data set as well
as each EIP compilation (alkane, aromatic, indane, OCP, and PNA). Each data set was
also subjected to PCA, and scores plots were generated to assess the natural clustering of
the diesels based on their inherent chemical composition.

The complete PPMC coefficient tables for each data set are shown in Appendix C.
The 90% CI for the replicates for the TIC was calculated to be 0.9813 — 1.0029. When

examining Cls for different diesel pairs, Diesels 21 and 23 (0.9688 — 0.9934), Diesels 21

81

Table 4.4 PPMC Coefficients for TIC and EIPs
for Replicate Sets of Five Diesel Samples

 

 

Data Set Average Standard Minimum Maximum
Dev1at10n

TIC 0.9921 0.0061 0.9780 0.9975
Alkane EIP 0.9729 0.0384 0.8788 0.9971
Aromatic EIP 0.9813 0.0145 0.9597 0.9984
Indane EIP 0.9960 0.0019 0.9916 0.9987
OCP EIP 0.9869 0.0086 0.9682 0.9973
PNA EIP 0.9910 0.0064 0.9772 0.9985

 

82

and 24 (0.9692 — 0.9818), and Diesels 23 and 25 (0.9809 — 0.9889) exhibited overlapping
ranges with the CI established by the replicates, which indicated that these three pairs
were statistically similar. All other sample pairs did not exhibit CIs that overlapped with
the replicates, which indicated that they are statistically different. Similar observations
concerning statistical similarities between diesel pairs can be made for the ElPs as well.

PCA was also performed on the neat data set using the TIC and each EIP. Only
2D scores plots are shown for these data sets for ease of comparison. The scores plot for
the TIC is shown in Figure 4.1. In this plot, PCI accounts for 53% of the variance, while
PC2 accounts for 26%. The replicates for each diesel are well clustered, though the
replicate clusters for Diesels 21 and 23 exhibit minor spread in PC2. For the five diesels,
89% of the variance for the first two PCs reasonably separates all five samples. Diesel 22
is significantly positive in PC], while Diesel 24 is negative. The other three diesels are
all relatively close to zero in PCI. Diesels 22 and 24 are negative in PC2, and Diesels 23
and 25 are positive. The combination, then, of PCI and PC2 is sufficient for separation
of the replicate clusters of each diesel.

Additional information about how the individual diesels cluster can be garnered
from the scores plots of the five different EIPs. Figure 4.2 shows the scores plot for the
alkane EIP, in which PCl accounts for 51% of the variance and PC2 24%. The same
general clustering trend as observed in the TIC scores plot can be seen in PC 1 in the
alkane plot. Diesel 22 is positive, Diesel 24 is negative, and Diesels 21, 23, and 25 are
close to zero. However, in PC2, most sample scores are close to zero, with the exception
of one triplicate analysis of Diesel 23, which is significantly distant from its replicate

cluster. As discussed previously, this separation of one replicate analysis from the others

83

ADiese121 ADiese122 ADiese123 ‘ Diesel 24 ADiese125

 

 

6000000 4
‘ A

g A

«‘3

v 1 ‘ 1

N

2 -6000000 A 6000000
-6000000 —

PC] (53%)

Figure 4.1 Scores Plot for the TIC of the Neat Diesels

84

PC2 (24%)

ADiese121 ADiese122 ADiesel 23 Diesel 24 ADiesel 25
3000000 a

 

r I

-3000000 A 3000000

 

30000004
PC] (51 %)

Figure 4.2 Scores Plot for the Alkane EIP of the Neat Diesels

85

is often due to a misalignment introduced in the data-pretreatment steps. The alkane EIP
has been prone to misalignments as seen in Chapter 3, and the fact that PC2 describes
24% of the variance indicates that a considerable misalignment, or potentially more than
one misalignment, has been introduced in the data pre-treatment steps for that particular
replicate sample. Therefore, association and discrimination for the alkane plot can only
be accurately assessed in the first principal component.

The aromatic scores plot, shown in Figure 4.3, indicates that Diesel 22 differs
significantly in aromatic content than the other four diesels. In fact, the other four diesels
appear to be highly similar in their aromatic content. It should be noted that the
clustering of the samples in the aromatic EIP is almost completely dictated by PC 1. Very
little spread is observed in PC2. Because of the considerable amount of variance
described by PC] (87%) as compared to PC2 (5%), it can be assumed that PC2 does little
to separate the diesels based on chemical composition. Therefore, clustering in the
aromatic profile is only considered as based on PCI.

The indane scores plot, shown in Figure 4.4, shows a similar trend in PC 1 to the
aromatic scores plot. Diesel 22 is largely positive in PCI (which describes 83% of the
variance), while the other four diesels are each slightly negative and are not well
separated from one another. However, unlike in the aromatic scores plot, PC2 (which
describes 8% of the variance) contributes to the natural clustering of the samples. Diesels
23 and 25 are separated from each other in PC2, as well as from Diesels 21 and 24;
however, due to minor spread in the replicate clusters in PC2, Diesels 21 and 24 cannot
be reasonably separated from one another and are likely similar in indane content.

The OCP scores plot, shown in Figure 4.5, is similar in its cluster patterns to the

86

ADiesel 21 ADiese122 ADiese123 ‘ Diesel 24 ADiesel 25

 

 

2000000 .
3" A
g , 4 ‘ Iii—fl
2 -2000000 2000000
-2000000 _.
PC1(87%)

Figure 4.3 Scores Plot for the Aromatic EIP of the Neat Diesels

87

PC2 (3%)

ADiese121 ADiese122 ADiese123 Diesel 24 ADiesel 25

200000 7

A
A
‘ 1

 

-750000

750000

A
A

 

-200000 —
PC] (83%)

Figure 4.4 Scores Plot for the Indane EIP of the Neat Diesels

88

PC2 (14%)

A Diesel 21

500000 ~

ADiese122 ADiese123 ' Diesel 24 ADiese125

fl

 

1

-1000000

A
A

 

-500000 J
PC] (62%)

Figure 4.5 Scores Plot for the OCP EIP of the Neat Diesels

89

1000000

alkane scores plot (Figure 4.2). In PC] (which describes 62% of the variance), Diesel 22
is positive, Diesel 24 is negative, and the other three diesels center around zero.
However, in the alkane plot, differentiation in PC2 is potentially due to a misalignment
instead of actual chemical variation. In the OCP plot, differentiation in PC2 is not
characteristic of a misalignment, and it can be assumed that any discrimination is likely
caused by differences in chemical composition. In PC2, Diesels 21, 23, and 25 are
separated, while Diesels 22 and 24 are similar in PC2. Again, the combination of PCI
and PC2 allows for moderate separation of the five diesels in the OCP scores plot.

In the PNA scores plot, Figure 4.6 PCI accounts for 50% of the variance, and
PC2 accounts for 17%. The five diesels are somewhat separated in PCI, where Diesel 22
is negative, Diesel 21 is around zero, Diesels 24 and 25 are slightly positive, and Diesel
23 is slightly more positive. PC2 also causes separation of the samples, with Diesels 22.
23, and 25 being negative and Diesels 21 and 24 positive in PC2. When combined with
the separation caused by PC 1 , the five diesel replicate clusters are reasonably separated
from one another.

It should be observed that, with the exception of the misalignment in PC2. the
clustering of the diesels in this data set is significantly improved over prior data sets,
most likely due to improved precision in the manual injection method that comes with
analyst experience. For example, in the final neat data set described in Chapter 3, the
alkane, indane, and PNA plots were deemed less useful because they demonstrated
spread in the replicates and a lack of natural clustering among the samples. In this data
set (analyzed using a new instrument), however, the TICs and EIPs provide considerable

association among replicates and discrimination among samples for all five diesels. This

90

PC2 (17%)

ADiesel 21 ADiese122 ADiesel 23 ' Diesel 24 ADiese125
100000 7

A

 

I I

-200000 % AA 200000

AA

 

-100000 J
PC] (50%)

Figure 4.6 Scores Plot for the PNA EIP of the Neat Diesels

91

association and discrimination capability demonstrates the utility of PCA, not only for the
TIC but also for characteristic EIPs, for comparing neat diesel samples.
4.3 Efficiency of Solvent Extraction Procedure
4.3.] Procedure

After the analysis of the neat diesels, the next step was to perform the burned
studies. First, however, a method to extract the diesel residue from the charred debris
was investigated. Of the ILR extraction procedures maintained by ASTM, passive
headspace using activated charcoal is more commonly used [9]. However, because it is a
headspace method, it can be biased against heavier, less volatile components. Diesel is
considered a heavy petroleum distillate because it contains high carbon number
compounds; therefore a headspace method has the potential to disproportionately extract
the less volatile diesel components. For this reason, a solvent extraction is the preferred
extraction method for diesel and thus was selected as a more representative procedure.

Via preliminary studies, the solvent extraction method chosen was to add 5 mL of
carbon disulfide (C82, spectrophotometric grade, Sigma-Aldrich, St. Louis, M0) to the
matrix to be extracted in a 50 mL round bottom flask, which was then attached to the
rotary arm of a Rotovap (R110, Btichi, Switzerland). The arm, which during normal
operation of the Rotovap sits in a heated water bath, was positioned outside the water
bath to prevent evaporation. The rotary function was adjusted to a setting of 4 out of 10.
and the extraction mixture was allowed to agitate for 5 minutes. The solvent was then
removed with a glass Pasteur pipet and, if necessary, filtered through a disposable syringe
(Becton Dickinson & Co., Franklin Lakes, NJ) with a 25 mm, 0.2 pm pore size PTF E

membrane syringe filter (Grace Davison Discovery Sciences, Deerfield, IL). The filtrate

92

was then analyzed by GC-MS using the same temperature program and instrument
parameters described previously (Section 4.2.1).

In order to demonstrate the efficiency of the solvent extraction method, a spike
and recovery study was performed for the three matrices (cotton cloth, magazine, and
carpet) to be considered in the burned studies. The cotton cloth was taken from a white
cotton wash cloth (100% cotton, Wholesale Merchandisers, Grand Rapids, MI). The
magazine was cut from a copy of ESPN The Magazine (20 pages including cover and
glossy pages, ESPN, New York, NY). The carpet was taken from a roll of unused nylon
carpeting (origin unknown). A series of standard solutions containing 1,3,5-
trimethylbenzene (98%, Aldrich, Milwaukee, WI), decane (99+%, Aldrich, Milwaukee,
WI), indane (97%, Aldrich, Milwaukee, WI), dodecane (99+% olefin free, Matheson,
Coleman, & Bell, Norwood, OH), and tetradecane (99%, Aldrich, Milwaukee, WI) were
prepared in dichloromethane (CH2C12, spectrophotometric, Sigma-Aldrich, St. LOuis,
MO) at different concentrations: 0.01 %, 0.03 %, 0.05 %, 0.07 %, and 0.10 % (v/v).
These components were selected to represent the common compounds classes present in
diesels that span the retention time range. One mL of each calibration standard was first
spiked with no matrix present, i. e. straight into the round bottom flask. It was then
subjected to the extraction procedure as if a matrix were present and analyzed by GC-
MS. The peak areas were integrated using the ChemStation software (Agilent, Santa
Clara, CA) and plotted against mass injected (calculated using density) to create a
calibration curve for each component in the calibration standard. The same extraction
procedure was then performed with the matrix present for all concentrations of the

calibration standard solutions. The mass of each component recovered was determined

93

by substituting the corresponding peak area into the appropriate calibration equation.
The calculated mass was then divided by the theoretical mass in order to determine a
percent recovery and hence assess the efficiency of the extraction procedure for each
component at each of the spiked concentrations.
4.3. 2 Results and Discussion

The first spike and recovery study was performed for the cotton cloth matrix. The
calibration curve for the neat standards by mass injected (no matrix present) is shown in
Figure 4.7 for all five components of the calibration standard. All five linear regressions
show R2 values of greater than 0.9900. Table 4.5 shows the percent recovery values of
each calibration standard from the cloth. With the exception of the 0.01% (v/v) standard
solution, all recovery values are approximately 90% with a standard deviation of less than
0.5%. The higher than average recovery for the 0.01% (v/v) standard is most likely due
to the fact that it is at the lower end of the linear range, where the theoretical mass to be
recovered is already small. In fact, the linear regression equations alone nominally
accounted for a significant portion of the recovery because the x-intercept values for each
were close to the theoretical recoveries for the 0.01 % (v/v) standard. It should also be
noted that the low standard deviations for each calibration standard indicate that the
solvent extraction procedure does not preferentially extract one type of component.

Figure 4.8 shows the calibration curves for the neat standards by mass injected
(no matrix present) prepared for the day in which recoveries were examined for the
magazine matrix. Again, R2 values of greater than 0.9900 indicate a linear calibration
curve for the standard solutions. Table 4.6 lists the recovery values for each calibration

standard from the magazine. Similar trends to the cloth recovery values are observed.

94

 

4.0E+07

 

 

 

 

y = 2272551954253x - 10764415738 A
R2 = 0.9922
y --— 1 1 17822794745x - 3668916148
R2 2 09928
a y : 2268134387453x — 10617742459
if, R2 = 0.9925 ‘
a 5":1434206771519x - 4820168852 9
3 R2 r 099119
On
' y = l609814l3.3864x - 6200999180
.1 R2 = 0.9917
0.0E+00 I T
0.000 0.060 0.120 0.180

Mass Injected (ug)

01,3,5-Trimethylbenzene I Decane A Indane X Dodecane O Tetradecane

Figure 4.7 Calibration Curve for the Neat Calibration Standards
for Cloth Matrix Recovery Study

Table 4.5 Extraction Recoveries from Cloth for Each Calibration Standard

(1,3,5-TMB = 1,3,5-trimethylbenzene, C10 = decane,

C12 = dodecane, C14 = tetradecane)

 

 

 

Standard Concentration! Percent Recovery Average Recovery
(°/o v/v) 1,3,5-TMB C10 Indane C12 C14 (V0)
0.01 106.33 101.89 105.16 101.27 101.76 103.28 :1: 2.30
0.03 87.86 87.92 87.54 87.27 87.19 87.56 :1: 0.33
0.15 87.65 88.29 88.02 88.33 87.71 88.00 i 0.31
0.07 87.35 87.96 88.11 87.26 87.21 87.58 :t 0.42
0.10 92.97 93.09 92.78 92.97 93.00 92.96 i 0.11

 

 

 

 

 

 

 

 

95

 

6.0E+07

 

Peak Area

 

= 0.9929

R2 = 09943

= 0.9936

R2 = 0.9942

y = 3126603510607x - 12120031639
R2
y = 1630024456546): - 364969.795]

y = 317995571 .5255x - 11718932541

R2
y = 19(3662438.281(.1x - 6160627869

A
O

y = 231112863.6301x - 963208.1393
R2 =0.9942

,/

 

0.0E+00
0.000

1

0.060

Mass Injected (ug)

1

0.120

 

0.180

01,3,5-Trimethylbenzene I Decane A Indane XDodecane O Tetradecane

Figure 4.8 Calibration Curve for the Neat Calibration Standards
for Magazine Matrix Recovery Study

Table 4.6 Extraction Recoveries from Magazine for Each Calibration Standard

(1,3,5-T MB = 1,3,5-trimethylbenzene, C 10 = decane,

C12 = dodecane, C14 = tetradecane)

 

 

 

 

Standard Concentrat’ Percent Recovery Average Recovery
(70 V/V) 1,3,5-TMB C10 Indane C12 C14 (%)
0.01 115.78 111.98 114.44 110.62 116.21 113.81 $2.43
0.03 98.35 97.74 97.85 98.28 99.53 98.35 :1: 0.71
0.05 86.35 86.66 85.83 86.38 87.29 86.50 a: 0.53
0.07 104.22 105.04 104.54 105.56 106.04 105.08 :1: 0.74
0.10 99.13 99.95 99.51 99.86 100.54 99.80 a: 0.52

 

 

 

 

 

 

 

96

 

The 0.01% (v/v) standard is higher than average, while the rest of the standards show
recoveries all greater than 85%. Even though the recovery value for the 0.07% (v/v)
standard is slightly higher than 100%, it still falls within an acceptable range for
recoveries (5% error, or a range of 95-105% recovery) [3 8]. As before, the low standard
deviations indicate no compound bias in the solvent extraction procedure from the
magazine matrix.

Figure 4.9 shows the calibration curves for the neat standards by mass injected
(no matrix present) prepared for the day in which recoveries were examined for the carpet
matrix. R2 values are also above 0.9900 for the linear regression equations. Table 4.7
lists the recovery values for each calibration standard from the carpet matrix. Again, the
same trends as for the cloth and magazine are observed. In this instance, however, the
recovery value for the 0.05% (v/v) standard is low by comparison with other recovery
values. In the corresponding calibration curves, however, it appears that the peak areas
for the 0.05% (v/v) standard are lower than the expected peak areas determined by the
regression equations. It is possible that the standard was poorly prepared. It could have
been omitted from the calibration curve entirely, but because acceptable R2 values were
still attained, the 0.05% (v/v) standard values were retained even with the disparity. Since
the same standard that was used for the neat calibration curve was also used for the spike
and recovery from the matrix, it is feasible that, if poorly prepared, this standard would
project low in the recovery rates as well.

The results of the spike and recovery studies using a representative standard
solution indicate that a five minute extraction by agitation in five mL CS; is likely to be

sufficient for the extraction of diesel from the three matrices investigated. The recoveries

97

 

 

7.0E+07

 

 

 

y = 4093101469953x - 28838442623
R2 = 0.9960
y 2 2139902108691x — 11201185410
R2 =. 0.9968
,3 y = 4158075864954x - 32094293607
2 R2 = 0.9962
<3 .. - , _
:4 y =— ZED/8401115843.»; - 14706172541
3 18 = 0 9965
On
, _. y = 3008998562243x - 17269372951
, r "J i ' R2 = 0.9970
0.0E+00 ‘ ' Y 7
0.000 0.060 0.120 0.180

Mass Injected (ug)
O 1,3,5-Trimethylbenzene I Decane A Indane >< Dodecane O Tetradecane

Figure 4.9 Calibration Curve for the Neat Calibration Standards
for Carpet Matrix Recovery Study

Table 4.7 Extraction Recoveries from Carpet for Each Calibration Standard
(1,3,5-TMB = 1,3,5-trimethylbenzene, C 10 = decane,
C12 = dodecane, C14 = tetradecane)

 

 

 

 

Standard Concentrat Percent Recovery Average Recover
(% vlv) 1,3.5-TMB C10 Indane C12 C14 (W
0.01 135.02 133.22 136.10 137.25 138.01 135.92 :1: 1.89
0.03 97.39 98.13 97.37 99.53 100.41 98.57 :1: 1.35
0.05 73.40 74.78 73.10 75.29 75.55 74.42 t 1.11
0.07 90.72 92.25 89.77 92.84 93.43 91.80 :1: 1.52
0.10 81.68 82.39 80.34 83.43 84.20 82.41 :t 1.50

 

 

 

 

 

 

 

 

98

are certainly not optimal; however, they are adequate for the purposes of this research
project, especially since fire debris analysis is typically qualitative and not quantitative in
nature. The more important feature of the solvent extraction procedure is that it is not
biased toward any particular component or compound class in diesel, as any preference
would skew the results of the PPMC coefficients and the PCA. Due to the sufficiency of
the results of the spike and recovery studies, the solvent extraction procedure was deemed
appropriate for subsequent extractions of spiked diesels from burned matrices.

4.4 Analysis of Burned Diesels

4.4.1 Procedure

Three matrices (cotton cloth, magazine, and carpet) were investigated for the
extraction and analysis of burned diesels. Each matrix was subjected to a series of
burning conditions, with and without diesel present, in order to examine the matrix
interferences and burning effects on the diesels. All substrates were cut to three
centimeter squares, placed in a glass petri dish, spiked with diesel as necessary, and then
burned. All GC-MS analyses were performed using the previously described NCFS.M
method.

First, the unburned matrix was extracted using the solvent extraction procedure
and analyzed in order to identify the presence of any potential interferences inherent to
the matrix itself. Next, the matrix was burned, extracted, and analyzed to determine any
pyrolysis products created during the burning process that could also act as potential
interferences. To burn the matrix, a butane grill lighter (BIC, Shelton, CT) was used to
ignite it, and it was allowed to burn until self-extinguished. If after extinguishing, the

majority of the surface was not charred, the substrate was re-ignited until significant

99

charring had occurred. Next, 50 11L of Diesel 21 was spiked onto the unburned matrix,
which was then extracted and analyzed in order to examine recovery of the diesel and
potential interferences. The same conditions were repeated for a burned matrix, i.e. the
diesel was spiked onto the burned matrix, in order to assess the efficiency of the diesel
recovery from the burned matrix. The final conditions examined were those most similar
to an actual arson case. Diesel 21 was spiked, again 50 11L, onto the matrix, which was
then ignited. The debris was extracted and analyzed by GC-MS in order to assess the
potential for the association of the burned diesel to its unburned counterpart.

After Diesel 21 was used to carry out each burning series for all three matrices.
two other previously analyzed samples were treated as blind unknowns and analyzed in
an attempt to assess the potential for the association of a burned diesel extracted from fire
debris to its unburned counterpart. A colleague selected two diesels and placed them in
vials labeled “A” and “B”. These two blind samples were then spiked (50 11L) onto each
of the three matrices, burned, and finally extracted and analyzed by GC-MS.

For the data sets collected for this portion of the research, the need for data pre-
treatment is the same as described before, though some minor adaptations were necessary
to accommodate the alterations to the chromatograms effected by both the burning
process and the extraction procedure, such as loss of volatiles and variation in extraction
efficiency. The TICs from the burned diesel data set (the samples that were spiked onto
the unburned matrix and then burned) were similarly compiled and retention time
aligned, though a window size of five was insufficient for their alignment (the algorithm
failed with a window of five). For that reason, the burned TIC data set was aligned using

a window size of seven. Targets were selected at random for each data set. The ElPs

100

were similarly compiled and aligned. For some EIPs, the same parameters used to align
the TIC were sufficient for alignment. However, other EIP sets required a manual
selection of the noise threshold for peak identification as opposed to using the default
calculation applied by the alignment algorithm. Both the indane and PNA profiles
required a threshold of 100, while the OCP profile required a threshold of 300.
Normalization and mean-centering were also performed in the same manner as detailed in
Chapter 3, though each data set was treated separately. For example, the burned diesels
were normalized and mean-centered separately from the neat diesels. Separate treatments
were necessary because of the differences in abundances between a 1:10 dilution of a
neat diesel and an extracted 50 11L diesel spike that had been burned.

After pre-treatment, the TICs and EIPs were then analyzed as described in
Chapter 3. PPMC coefficients were used to determine the level of correlation among
sample pairs. PCA was performed in order to observe the natural clustering of the diesels
based on chemical composition. It should be noted that PCA was not actually performed
on the bumed diesels. Instead, the eigenvectors calculated for the neat diesel data set
were used because they represent the chemical components that are the sources of
maximum variance among the five diesels. The eigenvectors were used to calculate a
score for the diesels that were extracted from the burned matrices. This score was
calculated in Microsoft Excel by multiplying each point in the eigenvector by its
respective point in each mean-centered chromatogram. These products were then
summed to calculate the score for that diesel corresponding to that specific principal
component. The objective was to associate and discriminate the samples based on their

chemical composition, not on changes introduced in the burning and extraction process.

101

If the diesels from the burning trials were included in the PCA, then the PCA would
specifically search for any differences among the neat and burned samples instead of
associating and discriminating them only on their inherent composition.
4. 4.2 Results and Discussion
4. 4. 2.1 Assessment of Potential Interferences from Unburned and Burned Matrices

Figure 4.10 shows sample chromatograms from the extraction of the three
unburned matrices. In the cloth matrix, no peaks of interest were observed—in fact. the
chromatogram resembled that of a blank solvent injection. In the magazine extract,
however, a relatively significant peak eluted at approximately 16 minutes. A mass
spectral library search tentatively identified this peak as a long chain ester, likely from
the ink used to print the magazine. Other peaks at lower abundance are also present in
the unburned magazine. In the carpet extract, two significant peaks were observed after
20 minutes. Mass spectral library searches indicated that these two peaks were bi-
naphthalene products, potentially from the carpet backing, adhesives, or a chemical
treatment such as water repellant or stain resistant. Again, smaller peaks are also
observed throughout the chromatogram. These smaller peaks, however, are at such an
insignificant abundance compared to the abundances expected for the diesel components;
therefore, it is unlikely that these smaller matrix peaks will interfere in anyway with the
results of the chemometric procedures.

Figure 4.11 shows sample chromatograms from the extraction of the three burned
matrices. Again, the cloth extract contains no significant peaks. The magazine and
carpet extracts contain the same peaks as observed in the unburned extracts, though at

higher abundance. These increased levels could be due to variation in the extraction

102

10000 ..___._.._- -- 7 A --.,.

Abundance

 

0 1 1 ‘1 1 llJILLlilutrril 1 LAMA

0 5 10 15 20 25 30
Retention Time (min)

40000 ——-—— -—» —-»--

(b

 

 

 

 

 

 

g I

" 1
1:1

:1

=
.n
<

0 up i “1.; LI; -11 ULL [111.11. ‘1“ ILL—.1.“
0 5 10 15 20 25 30
Retention Time (min)
40000 --———-—— ._ - i ——
(C)

0

8

8

1:

:1

.n

<

J 1
1 11 l. LAgulAAllllil A. 1 JUL."

 

 

 

0 5 10 1 5 20 25 30
Retention Time (min)

Figure 4.10 Representative Chromatograms of Unburned Matrices:
(a) Cloth, (b) Magazine, and (c) Carpet

103

10000 -————-- 777 77 7-.- -2.

Abundance

 

0 5 10 15 20 25 30
Retention Time (min)

 

 

150000 -- --

(b)

Abundance

 

 

AA7A

0 5 10 15 20 25 30
Retention Time (min)

 

 

50000 — -- -———

Abundance

 

 

 

 

O 3 1. It.“ illibilkul . 3M

0 5 10 15 20 25 30
Retention Time (min)

Figure 4.11 Representative Chromatograms of Burned Matrices:
(a) Cloth, (b) Magazine, and (c) Carpet

104

efficiency, or even differences in injection volume. It is possible that these peaks could
interfere in the chemometric results if they remain significant in abundance in
comparison to the presence of a diesel sample.

It should be noted that no significant peaks that could be attributed to pyrolysis
products (that is, peaks present in the burned matrix but not in the unburned) were
observed for any of the three matrices. It is possible that pyrolysis product formation is
limited in these three matrices, or again that the pyrolysis products are volatilized and lost
to the environment during the burning process. It is also possible that the solvent
extraction procedure is not selective for any pyrolysis products present, i. e. the pyrolysis
products may not be soluble in the solvent. Whatever the reason, the lack of interfering
pyrolysis products is advantageous for the accuracy of the results of both the PPMC
coefficients as well as PCA.

4. 4. 2. 2 Solvent Extraction of Diesels from Unburned and Burned Matrices

Figure 4.12 shows sample chromatograms from the extraction of the three
unburned matrices spiked with 50 1.1L of Diesel 21. All three chromatograms show
similar diesel chromatographic patterns. No extraneous peaks are observed that could
potentially interfere with the chemometric results. Figure 4.13 shows sample
chromatograms from the extraction of the three burned matrices spiked with 50 1.1L of
Diesel 21. Again, a similar chromatographic pattern is observed, and no interference
peaks are obvious. The lack of any interference peaks in either the unburned or burned
matrices indicates that the diesel is present at a sufficient concentration that inherent
hydrocarbons from the matrices do not make significant contributions to the

chromatograms. The consistency in the diesel peak pattern from the extracts of both

105

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3000000 ~——— 7777 722-- -
(a)
8
:1
a
11
7 1
3
< 1
O .. .u. 1 T 1 ‘ I
0 5 10 15 20 25 30
Retention Time (min)
3500000 ~7b—)-—— i
8
:1
a
e
:1
:1
.1:
<
0 - . . , , J
0 5 10 15 20 25 30
Retention Time (min)
2000000 .77 — .__---._-____-
(C)
0
‘e’
S
5
.E
<
0 — 7 7

10 15 20 25 30
Retention Time (min)

0
M

Figure 4.12 Representative Chromatograms for Unburned Matrices Spiked
with Diesel 21: (a) Cloth, (b) Magazine, and (c) Carpet

106

3000000 ‘--~———* - c .

Abundance

 

 

 

 

 

 

0 S 10 15 20 25 30
Retention Time (min)

 

4000000 -—————

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

l
8
E
a
a
G
5
.n
<
I
0 1 , 7
0 5 10 15 20 25 30
Retention Time (min)
2000000 — -— - e ~~ M-
(C) a
1
8 l
E
I
11
E
=
.D
<
0 ' T . 1 J
0 5 10 15 20 25 30

Retention Time (min)

Figure 4.13 Representative Chromatograms for Burned Matrices Spiked
with Diesel 21: (a) Cloth, (b) Magazine, and (c) Carpet

107

burned and unburned matrices also signifies that the presence of the matrix does not
negatively affect the solvent extraction of the broad range of diesel components.
4. 4. 2. 3 PPMC Coefficients of Neat and Burned Diesels

Sample chromatograms for the extraction of spiked and burned diesels from each
of the three matrices are shown in Figure 4.14. PPMC coefficient tables for the compiled
neat and burned data can be found in Appendix D. For the TIC PPMC coefficients. the
burned Diesel 21 replicates (spiked on unburned matrix, then burned) were assessed for
precision of the burning process for all three matrices. The average correlation
coefficient for each matrix was as follows: cotton cloth — 0.9847, magazine — 0.9344. and
carpet — 0.8411. These correlations, which are lower than observed for neat diesel
replicates, indicate the degree of uncertainty introduced by the burning process, which
appears to be matrix dependent. The cotton cloth yields higher PPMC coefficients for
Diesel 21 when it is burned and extracted in comparison to the other two matrices. The
higher level of correlation for cotton cloth indicates that either the burning process is
more reproducible for the cloth matrix, or that the solvent extraction procedure is the
most efficient for diesel from the cloth. The replicate spiked and burned extracts from
the magazine are less correlated, most likely because of the way the magazine itself
burns. Unlike the cloth, which only chars, the magazine becomes ashen. This difference
makes the surface area of the burned matrix variable, and thus introduces variability in
the PPMC coefficients. The magazine is also less absorbent than the cloth, and is less apt
at shielding the diesel from the burning process. The carpet yielded the least precise
correlation coefficients for the replicate spiked and burned extracts of Diesel 21. This

imprecision is most likely due to the relatively high retention of the diesel by the carpet.

108

1000000 ~"—-—W~- --* A 7777. 77 3,

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(a)

8 ';
:1 1
a 1

'5 l
‘ l
=
.9
< 1

l
0 5 10 15 20 25 30
Retention Time (min)

1000000 a
.. l
I t
a 1
u
E
:1
.1:
<

I
O ‘ 1 1 7
0 5 10 15 20 25 30
Retention Time (min)
300000 T——~w~-————-- 77v .
l
(c)
l
3
= 1
I 1
'° 1
1: .
5 1
a l
< 1
l
1
l
0 , . ' ' 7 1 . -1
0 5 10 15 20 25 30

Retention Time (min)

Figure 4.14 Representative Chromatograms for Matrices Spiked with Diesel 21
and then Burned: (a) Cloth, (b) Magazine, and (c) Carpet

109

This increased retention may protect the diesel from the fire, but it can also hinder the
solvent extraction procedure. The carpet also has a greater surface area than the two
other matrices due to the height of the carpet fibers from the carpet backing, as well as
the presence of multiple fibers. This additional surface area can lead to a more variable
burning process. The difficulty in extraction and imprecision in the burning process may
prevent replicate burnings of diesel on carpet from being precise.

In addition, PPMC coefficients were utilized to associate the burned Diesel 21 to
its neat counterpart. However, the highest correlation for none of the three matrices
corresponded to Diesel 21. The cloth extracts were on average more closely associated
with Diesel 23 (0.8840) than Diesel 21 (0.8657), the magazine extracts from the
magazine with Diesel 25 (0.8183) instead of Diesel 21 (0.7977), and the carpet extracts
with Diesel 22 (0.6203) instead of Diesel 21 (0.6028). These differences, however, are
not statistically significant at the 90% confidence level, which indicates that the extracts
are as well correlated with Diesel 21 as with the other neat samples. Several steps were
taken in order to improve the correlation coefficients between Diesel 21 and the three
matrix extracts, which included the selection of different target chromatograms for the
alignment algorithm, the alteration of the user defined window size for the alignment
algorithm, and the statistical examination of replicate samples for outliers. These
changes, however, did not affect the trends observed for the correlation coefficients
calculated for the matrix extracts.

For the blind samples, it should first be noted that the correlations for the TIC s of
the magazine and carpet extracts for Blind B (Appendix D) are all either poor or negative.

which indicates a complete lack of correlation to any of the neat diesels. It is possible

110

that these samples are insufficient for comparison due to poor recovery by the solvent
extraction procedure. The cloth extract for Blind B has highest average correlation with
Diesel 25 (0.9076). The three extracts for Blind A all have maximum correlations with
Diesel 22 (cloth — 0.8665, magazine — 0.8575, and carpet — 0.8339). This regularity in
maximum correlation indicates that Blind A is most consistent with Diesel 22.

Closer examination of the PPMC coefficients for the EIPs reveals similar trends
to those observed in the TIC. The burned Diesel 21 extracts are somewhat correlated
with neat Diesel 21. However, extracts from all three matrices for Blind A consistently
have a maximum correlation with Diesel 22, and the cloth extract for Blind B
consistently maximizes when paired with Diesel 25.

The PPMC coefficient results for the burned diesels indicate that the burning
process itself is highly variable. This variability inevitably affects the precision of the
PPMC coefficients and can impede the association of burned diesels to their unburned
counterparts. This variability also seems to be dependent on the features of the matrix
being burned, which include how the burning affects the surface area of the matrix, and
how the matrix retains the diesel and shields it from the heat of the fire. On the other
hand, a consistency was observed for the maximum correlations of the blind diesel
samples to a neat diesel sample, which implies that some amount of association is still
possible even after the effects of the burning process, even though association was not
demonstrated for the replicates of Diesel 21.

4. 4. 2. 4 PCA of Neat and Burned Diesels
Initially, the burned data was concatenated with the neat data, all of which was

carried through the previously described data pre-treatment processes. The mean-

111

centered data was then subjected to PCA and a scores plot generated for the TIC s. This
scores plot, shown in Figure 4.15, demonstrates the data set dependency of the PCA
procedure. PCA is designed to calculate the variance among the given data set. PCA as
an unsupervised procedure is incapable of associating a neat diesel and one that has been
altered during burning based on chemical composition, because obvious differences exist
between those chromatograms. In the scores plot, the neat diesels are clustered on the
negative side of PCI, while the burned diesels are clustered on the positive side. Thus.
PCA is not a feasible option for a compilation of neat and burned samples.

When considering PCA as an associative and discriminatory tool, the nature of
the samples to be associated and discriminated must be determined. In this research, the
objective is two-fold: to discriminate burned diesels of different origin from one another,
and to associate burned diesels to their unburned counterparts. In order to achieve this
objective, the factors that are capable of associating and discriminating neat diesels based
on chemical composition (i. e. the eigenvectors from the neat diesel PCA) can also be
applied to burned diesels from the same data set in order to associate and discriminate
them based on chemical composition. The assumption is that sufficient significant
features of the diesels persist through the burning process to still associate and
discriminate them based on the eigenvectors calculated for the neat diesel set. It is also
possible that a specific compound class is more likely to remain after burning. so that the
EIPs may provide more discriminatory information than the TIC alone.

Once the scores were calculated for the first two principal components for all the
burned samples for the TIC and the EIPs, they were added to the scores calculated for the

neat diesel set, and a new scores plot was generated. Each blind sample was consolidated

112

A Diesel 21 A Diesel 22

 

 

A Diesel 23 - Diesel 24
A Diesel 25 A Cloth - Diesel 21
A Magazine - Diesel 21 A Carpet - Diesel 21
A Blind A - Cloth, Magazine, Carpet A Blind B - Cloth, Magazine, Carpet
10000000 1
A
,3
:2. A n
N r é V fi‘ 1
2 -20000000 0 A 20000000
.10000000 J
PCI (92%)

Figure 4.15 Scores Plot for the TIC of the Neat and Burned Diesels

113

into a set of scores for all three matrices. However, an initial assessment of the resulting
scores plots indicated that the scores for the burned diesels all seemed to center around
the origin. Exemplary scores plots for the TIC and alkane EIP can be found in Figure
4.16. It was first thought that this location of the scores for the burned data was
indicative of a lack of chemical features in the burned chromatograms. However, a closer
look at the scores calculation method revealed that significant disparity in relative
abundances between the neat and burned diesels would lead to a difference in the
magnitude of their respective scores values. More explicitly, the eigenvectors were
calculated for the normalized abundances of the neat diesels. These same eigenvectors
were then multiplied by the less abundant chromatograms of the burned diesels and
summed to calculate a score. Therefore, these scores were lower in magnitude based on
the differences in abundances between the neat and burned diesels.

This discovery is an important revelation when considering the applicability of
this method to actual arson cases. It seems some method of correction is necessary for
samples that differ greatly in abundance levels in order to apply pre-determined
eigenvectors to calculate scores. In this case, it appeared that the neat diesel
chromatograms were approximately an order of magnitude larger in abundance than the
burned diesel chromatograms, so a correction factor of ten was applied. The normalized
burned data was multiplied through by ten, and the data was again mean-centered and the
scores calculated. It was observed that this multiplication of ten to the normalized data
led merely to an approximate ten-fold multiplication of the calculated scores values, so
henceforward the already calculated scores were simply multiplied through by ten. The

score plots were re-generated, and observations were made about their probative value

114

A Diesel 22

A Diesel 21

A Diesel 23 Diesel 24

A Diesel 25 A Cloth - Diesel 2]
A Carpet - Diesel 21

A Magazine — Diesel 21
.15. Blind A - Cloth, Magazine, Carpet
6000000

A Blind B - Cloth, Magazine, Carpet

A
A
.“91 .. A
. “g;

 

 

 

.3
\c
S 1 " 1
§-6000000 A " 6000000
4A
(0)
-6000000
PC1(53%)
3000000 ~
? A A
a: ‘ A A?
c . “In, 1
93500000 A‘ A 3500000
‘ A
b
() ,

 

-3000000 ~
PC1 (51 %)

Figure 4.16 Representative PCA Scores Plots for (a) TIC and (b) Alkane EIP for the
Burned Data Scored with the Eigenvectors from the Neat Data

115

for the association and discrimination of the neat and burned diesels.

Figure 4.17 shows the TIC scores plot containing the neat and burned diesels.
The first observation is the obvious spread in the replicates of the burned diesels as
compared to the replicate analyses of the neat diesels. The spread occurs in both PC1 and
PC2, and because these eigenvectors correspond to actual chemical differences as
suggested by the discrimination of the neat diesels, it is likely that the spread in the
replicates is caused by genuine dissimilarities in their chromatograms. These
discrepancies can be attributed to variability in both the burning process and extraction
efficiency; therefore, any interpretation of the association and discrimination among
samples must be prefaced with this potential for variability.

The scores for Diesel 21 spiked onto the three matrices and extracted are only
moderately close to the scores for neat Diesel 21. Two replicates of Diesel 21 spiked
onto a magazine and burned are more closely associated to the neat diesel than those
extracted from the other two matrices, though one cloth replicate is close as well. The
carpet replicates are not remotely close to the neat diesel. This trend is sensible, as the
magazine is less absorbent than the cloth and carpet matrices. Neither of the blind
samples can be logically associated with a specific neat diesel cluster. For Blind A, the
carpet extract is located in the proximity of the Diesel 22, which is consistent with the
PPMC coefficients results, while the magazine extract is closer to the Diesel 21 cluster.
The cloth extract is located between the other two extracts and cannot reasonably be
associated with any neat diesel cluster. For Blind B, the cloth extract is located within
the cluster for Diesel 21, which is inconsistent with the PPMC coefficients results, while

the magazine and carpet extracts are far from any neat diesel cluster in the upper left

116

A Diesel 21. A Diesel 22

 

 

A Diesel 23 . Diesel 24
A Diesel 25 A Cloth - Diesel 21
A Magazine - Diesel 21 A Carpet - Diesel 21
A Blind A - Cloth, Magazine, Carpet A Blind B - Cloth, Magazine, Carpet
12000000 1
A
A A
S; A A
\D
9' ' if A .
N -12000000 12000000
2 ‘ A1“ A A‘ A A
A
-12000000 J

 

PC1 (53%)

Figure 4.17 Scores Plot for the TIC of the Neat Diesels and the
Projected Burned Diesels After a 10X Correction

117

quadrant of the scores plots.

The alkane scores plot for the neat and burned data, shown in Figure 4.18.
exhibits similar traits to those observed in the TIC scores plot. The replicate analyses for
burned Diesel 21 are inconsistent and barely associated with their neat counterpart. For
Blind A, the cloth and the carpet extracts are associated with the clusters for Diesels 21,
23, and 25, while the magazine extract is more closely linked with Diesel 24. For Blind
B, the cloth extract is associated with the Diesel 24 cluster, while the magazine and
carpet extracts are again not remotely close to any other clusters. In fact, they are not
shown because their scores values are considerably different from the scores values of the
rest of the data set.

At this point, it became obvious that the magazine and carpet extracts for Blind B
were extremely different from the cloth extract and the rest of the data set being analyzed.
An examination of their chromatograms revealed exceptionally low abundances as
compared to the other samples. The TIC for the cloth extract of Blind B exhibited an
abundance level of approximately 500,000, whereas the TICs for the magazine and carpet
extracts exhibited abundance levels of approximately 70,000. The burned samples are
already low in abundance, and these aberrant values for the magazine and carpet extracts
prevent them from being scored on a similar level as the other samples. It is possible that
these two samples were poorly extracted. It does not seem feasible to use a different
correction factor for these samples in order to maintain consistency for all the samples
throughout the data analysis procedures. Whatever the reason, any interpretation of their
association to a neat diesel will henceforth be omitted.

The aromatic scores plot for the neat and burned diesels is shown in Figure 4.19.

118

A Diesel 21

A Diesel 23

A Diesel 25

A Magazine - Diesel 21

.35. Blind A - Cloth, Magazine, Carpet

A Diesel 22
Diesel 24
A Cloth - Diesel 21
A Carpet - Diesel 21
A Blind B - Cloth, Magazine, Carpet

 

 

3000000 1
3 A ‘ A At». ‘A
$5 sh___A._ , _.A 1‘ A .
§-4500000 ‘1‘ {‘9‘ 4500000
A
30000009

PC1 (51%)

Figure 4.18 Scores Plot for the Alkane EIP of the Neat Diesels and the
Projected Burned Diesels After a 10X Correction

119

A Diesel 21 A Diesel 22

 

 

 

A Diesel 23 Diesel 24_
A Diesel 25 A Cloth - Diesel 21
A Magazine - Diesel 21 . A Carpet - Diesel 21 _
A Blind A - Cloth, Magazme, Carpet A Blind B - Cloth, Magazme, Carpet
2mmm7
.3 at
In * A .
I: a 5% A A 93‘ m
3 -2000000 A - ‘ 2000000
-2000000 4
PC1 (87%)

Figure 4.19 Scores Plot for the Aromatic EIP of the Neat Diesels and the
Projected Burned Diesels After a 10X Correction

120

Greater precision among the burned replicates is observed than for the other scores plots.
and the burned Diesel 21 scores for all three matrices fall in the four diesel cluster
observed in the neat aromatic scores plot that contains Diesel 21 (as well as Diesels 23,
24, and 25). The cloth extract for Blind B also falls in that same cluster, which also
contains Diesels 23, 24, and 25. An interesting feature to observe is the location of the
scores for the three extracts of Blind A. The three extracts are separated from the four
diesel cluster along PC1 toward Diesel 22. The carpet extract is the closest to Diesel 22.
while the cloth extract is the furthest away. The relative precision among replicates
within this plot suggests that the identity of Blind A is potentially Diesel 22, whereas no
reasonable conclusions could be drawn from the other scores plots.

The indane scores plot for the neat and burned diesels is shown in Figure 4.20.
As before, significant spread is observed among the burned replicates, though in this case
the majority of the spread is along the second principal component. Again, little
consistency with the neat sample is observed for the burned Diesel 21 replicates for all
three matrices, though the magazine extracts are closest yet again. For Blind A, the
extracts from the three matrices are generally closest to the cluster for Diesel 22, though
by no means are they close enough to be considered statistically similar. Then again,
Diesel 22 is the only neat sample that is positive in PC 1 , and the Blind A extracts are all
significantly positive in PC1; therefore, it is plausible that some degree of association can
be suggested by these similarities between the two samples. The carpet extract is once
again the closest to Diesel 22. The cloth extract for Blind B is again located in the large
four diesel cluster centered about the origin.

The OCP scores plot for the neat and burned diesels is shown in Figure 4.21. The

121

A Diesel 21

A Diesel 23

A Diesel 25

A Magazine - Diesel 2]

A Blind A - Cloth, Magazine, Carpet

PC2 (8%)

A Diesel 22
Diesel 24
A Cloth - Diesel 21
A Carpet - Diesel 21
A Blind B - Cloth, Magazine, Carpet

 

 

‘ 600000 1
A
A
A‘
r b“ i a
-1200000 A A 1200000
A

 

-600000 J

‘A

PC1 (83%)

Figure 4.20 Scores Plot for the Indane EIP of the Neat Diesels and the
Projected Burned Diesels After a 10X Correction

122

A Diesel 21

A Diesel 23

A Diesel 25

A Magazine - Diesel 21

A Blind A — Cloth, Magazine, Carpet

 

A Diesel 22
Diesel 24
A Cloth - Diesel 21
A Carpet - Diesel 21
A Blind B - Cloth, Magazine, Carpet

 

1000000 -
A
A A A
i: q. A A
:3“ r * § A A '
3 -1200000 A 1200000
A A
A A
-1000000
PC1(62%)

Figure 4.21 Scores Plot for the OCP EIP of the Neat Diesels and the
Projected Burned Diesels After a 10X Correction

123

same trends are observed for the known burned Diesel 21 scores. Few are actually
associated with Diesel 21, and the replicate analyses are imprecise as observed by the
lack of clustering. The extracts for the three matrices for Blind A are again most closely
associated with Diesel 22, with the cloth and carpet extracts located the closest. The
cloth extract for Blind B is most closely associated with the clusters for Diesels 21 and
23.

The PNA scores plot for the neat and burned diesels is shown in Figure 4.22, and
again the same trends are observed for the known burned Diesel 21 scores. The blind
samples, however, are more informative in this plot. The cloth and carpet extracts for
Blind A are reasonably associated with the Diesel 22 cluster. The magazine extract is
located farther from the Diesel 22 cluster, though it is still negative in PC 1 just like the
other Diesel 22/Blind A scores and unlike the other neat diesel samples. The cloth
extract for Blind B is not associated with any neat diesel cluster.

4. 4. 2. 5 Identification of Blind Diesels

Through the PPMC coefficients and minor corroborating evidence from the PC A
scores plots, it was determined that the identity of Blind A was Diesel 22, and that the
identity of Blind B was Diesel 25. Although these were the correct assignments, the
identifications were not capable of being made with sufficient statistical confidence. Too
much variation is present throughout the PPMC coefficients and the PCA results in order
to conclusively identify the blind samples. Instead, only the most likely candidates could
be determined. However, with the considerable level of variability involved in the
extraction procedure and the burning process itself, the identification of a likely candidate

among samples that are so inherently similar is useful as a preliminary assessment. The

124

A Diesel 21

A Diesel 23

A Diesel 25

A Magazine - Diesel 2]

A. Blind A - Cloth, Magazine, Carpet

A Diesel 22
Diesel 24
A Cloth - Diesel 21
A Carpet - Diesel 21
A Blind B - Cloth, Magazine, Carpet

 

 

200000 —
A ‘ A
a A AA
0\
5: a.
X ' .52, ' A“ n. '
0 -200000 e - 200000
a. ‘A A
.. A
A
A
-200000J

PC1 (50%)

Figure 4.22 Scores Plot for the PNA EIP of the Neat Diesels and the
Projected Bu med Diesels After a 10X Correction

125

potential exists for a greater confidence level when associating a burned diesel to its
unburned counterpart if diesels were compared with a larger set of different classes of

ignitable liquids due to more significant differences in chemical composition.

126

Chapter 5
Conclusions and Future Work
5.1 Conclusions
5.1.1 Association and Discrimination of Neat Diesels Using PPMC Coefficients and PCA

Throughout the initial analysis of the neat diesel samples, several obstacles were
overcome in order to improve the precision of both the analytical method and the
chemometric procedures, as well as the significance of the results obtained. It was
determined that a two-second timed injection of a one nL sample encased in a pL air
pocket is necessary to ensure consistent vaporization of all components in the diesel in
the heated injection port. This consistency is imperative to the precision of the injection
method. It was also determined that a less dilute sample is essential in order to achieve
sufficient abundances in the chromatograms, so as to ensure that any variation in the
instrumental background is inconsequential in comparison to the variation among the
diesels. However, poor abundances were still observed in some of the EIPs, which
indicates that further investigation into the abundance issue is necessary. In addition to a
less dilute sample, it was determined that summed EIPs are more informative than a
single BIC for improved signal-to-noise.

Once these improvements were applied, a final neat diesel data set was analyzed
to investigate the potential for association and discrimination based on chemical
composition. PPMC coefficients illustrated that, on a pairwise basis, the diesel samples
can either be associated or discriminated based on a comparison of their correlation
coefficient to that of replicate samples. The ranges of correlation coefficients for both the

replicates and for samples that are different in origin based on ElPs extend further than

127

those for the TIC, which suggests that the EIPs could provide additional discriminatory
information. The PCA results for the TICs demonstrated that the diesels are relatively
well differentiated, and that replicate samples are tightly clustered. The PCA of the
alkane, indane, and PNA profiles did not provide any additional discrimination
capabilities compared to the TIC. It should be noted that errors in the retention time
alignment were observed in both the scores plots and the eigenvectors for the
aforementioned profiles. These errors cause the results to misrepresent the variation
among the samples based on chemical composition alone, so that the discrimination
capabilities of these profiles carmot be fairly gauged until these errors are corrected. The
PCA results for the aromatic and OCP profiles did, however, offer some additional
association and discrimination among the samples. These profiles did not exhibit the
same alignment errors as were observed in the alkane, indane, and PNA profiles, which is
most likely due to higher abundances and improved peak shapes for the aromatic and
OCP EIPs as compared to the other EIPs.

The results from the analysis of this neat diesel data set have shown that it is
possible to associate and discriminate diesel samples based on their chemical
composition using both PPMC and PCA. The fact that any amount of discrimination is
possible among diesels, which are so inherently similar in chemical composition. is
significant for the progress of this research. This potential for association and
discrimination indicates that it may be possible to apply the same instruniental techniques
and chemometric procedures to associate and discriminate diesels that have been burned

and extracted from fire debris.

128

5.1.2 Association and Discrimination of Burned Diesels Extracted from Fire Debris
Using PPMC C oeflicients and PCA

The objective of this part of the research was to examine the potential of PPMC
coefficients and PCA for the association and discrimination of burned diesels that have
been extracted from fire debris. In order to accomplish this objective, neat diesel samples
were analyzed for comparative purposes. This sample set was analyzed on a new
instrument, and it was determined the sensitivity of the new instrument required a
dilution of 10:] instead of 50:1 in order to achieve comparable abundance levels to prior
data sets collected on another instrument. It was also determined through RSD
calculations and PPMC coefficient determinations that the precision of the neat diesel set
was improved over prior data sets. The five diesels in the data set were associated and
discriminated based on chemical composition in not only the TIC, but in each of the EIPs
as well, which indicates the utility of both the TIC and EIPs in association and
discrimination determinations.

Once the neat diesel set had been analyzed, a series of spiking and burning
experiments was performed on three common household matrices: cotton cloth,
magazine, and carpet. A solvent extraction, consisting of agitation of the matrix in five
mL of C82 for five minutes, was selected first based on a set of spike and recovery
studies that demonstrated acceptable efficiency of the extraction. It was then determined
that, although some unburned and burned matrices exhibit potential interference products,
in the presence of a diesel, the interference peaks are masked by the components of the
diesel itself. Finally, a set of spiked and burned diesels (two of which were treated as

blind samples) was analyzed in an effort to simulate actual arson conditions. First, severe

129

spread was observed among replicate burnings in both the PPMC coefficients and scores
plots from PCA, which confirms that the burning process is variable. Secondly, for PCA.
it was determined that abundance plays a significant role when attempting to project
scores for burned samples from eigenvectors calculated for a neat data set. In this case. a
correction factor was applied to remedy the disparity in abundances between the neat and
burned diesels. Ultimately, PPMC coefficients were a better marker for the identification
of the blind samples, as differences in abundances did not affect the correlation
coefficients. While it was possible to associate burned diesels to their neat counterparts
using PCA, the association could only be made with minimal confidence. The PCA
results for the EIP sets also showed promise for the association of burned samples to their
neat counterparts. Overall, it is important to note that the diesels contained in the data sat
are, by nature, very similar chemically. As a result, the ability to begin to make
associations between neat and burned diesels, even at low confidence levels, is promising
for the utility of the described methodology. It is likely that associations between burned
samples and their neat counterparts can be made with higher confidence when the
methodology is applied to a range of ignitable liquids that greatly differ in chemical
composition.
5.2 Future Work

The future direction of this project lies mostly in the optimization of the already
developed methodology. A few issues in both the analytical technique and the data pre-
treatment and analysis phases must be addressed. It was concluded that a reproducible
injection method is crucial to the precision of the technique, and consequently the

accuracy of the results of the chemometric procedures. PCA is a procedure that focuses

130

on differences present in the chromatograms, which should be only due to actual
differences in chemical composition. An imprecise injection method can introduce
artificial variations in the data set, and thus lead to erroneous results. Although an effort
was made to standardize a manual injection method, an investigation into a programmed
injection method performed by an autosampler could improve precision and further
reduce the chance for the introduction of artifacts.

The other obvious issue that needs to be addressed is the retention time alignment
procedure. The current alignment algorithm is prone to misalignments, especially in the
EIPs due to lower abundances as compared to the TIC. This particular peak-matching
algorithm could be amended, with full optimization of all user-defined parameters, in
order to be more appropriate for these specific diesel data sets. However, the fault in the
alignment may be in the method with which the alignment itself is performed. More
recent work published on retention time alignments suggests that piecewise alignment
algorithms or warping algorithms may be more effective and less likely to result in
misalignments than more basic peak-matching algorithms as was used in this research
[41-44].

The last issue to be dealt with is the abundance problem in the scores projections
for the burned diesel samples to be associated and discriminated. It was determined that
significant disparities in the chromatographic abundance levels between the burned,
extracted diesels and the neat diesels from which the eigenvectors were calculated can
cause the scores for the burned data to be significantly different from the neat diesel
scores. In this case, a correction factor was applied to the scores for the burned diesels

based on average differences in abundance levels. However, no investigation was made

131

into the correction factor process, and in order for it to be a valid step, further study is
required. Different or even multiple correction factors may be necessary in order to
better represent the burned data, especially in actual arson cases when the abundances for
the chromatograms collected from extracted fire debris will be so variable.

Once these aspects of the research are sufficiently remedied, more broad
objectives can be addressed. Diesel was the ignitable liquid chosen for this study because
of its complex chemical composition, but in reality diesel is rarely seen in arson cases.
Typically, the lighter petroleum distillates like gasoline and lighter fluid are the more
frequently encountered ignitable liquids. When applied to different classes of ignitable
liquids, the methodology could prove more useful as an associative and discriminatory
tool. Further investigation into matrix interferences is also essential before this
methodology can be adapted into a forensic laboratory setting.

With further research, this methodology could be beneficial in the fire debris
analysis section of a forensic laboratory. At the moment, ignitable liquid identifications
are purely subjective in nature. The chemometric procedures offer a more objective
approach for the determination of ignitable liquids. The methodology has the potential
for complete automation, from the actual instrumental analysis all the way through the
data analysis stage. Coupled with an extensive reference collection, this methodology
could provide a powerful statistical method for the identification of ignitable liquid class.
and potentially even the association of a burned ignitable liquid to its unburned

counterpart.

132

APPENDICES

133

Appendix A

ASTM Classification Scheme

134

Table A.l ASTM Ignitable Liquid Classification Scheme [4,5]

 

Medium (Cg-C 13)

Heavy (Cs-(320+)

 

 

 

 

 

 

 

 

 

Class Light (C4-C9)
Gasoline-all
brands, . . . .
including Fresh gasollne lS typically 1n the range of C4-C .2
gasohol
Petroleum Ether Some Charcoal Kerosene
. Starters D1esel Fuel
Petroleum Some Cigarette . .
. . . . Some Paint Thmners Some Jet Fuels
Distillates L1ghter F lu1ds .
. Some Dry Cleaning Some Charcoal
Some Camping Fuels
Solvents Starters
Some Charcoal
Isoparaffinic Aviation Gas Starters Some Commercial
Products Specialty Solvents Some Paint Thinners Specialty Solvents
Some Copier Toners
Some Paint and Some Automot1ve
. Parts Cleaners . .
Varnish Removers S ecial Cleanin Some Insect1c1de
Aromatic Some Automotive p SotK/en ts g Vehicles
Products Parts Cleaners . . Industrial Cleaning
Some Insectlctde
Xylenes, Toluene- . Solvents
based Products Veh1cles
Fuel Additives
Some Charcoal . .
. Some Insect1c1de
Naphthemc Starters .
. Cyclohexane-based . . Veh1cles
Paraffimc Some Insectimde .
Solvents/Products . Lamp 0118
Products Veh1cles .
. Industrial Solvents
Lamp Olls
n-Alkanes Solvents Some Candle Oils Some Candle OHS
Pentane, Hexane, . Carbonless Forms
Products Cop1er Toners .
Heptane Cop1er Toners
De- Some Charcoal Some Charcoal
Aromatized Some Camping Fuels Starters Starters
Distillates Some Paint Thinners Odorless Kerosene
Alcohols, Ketones Some Lacquer
Some Lacquer Thinners
Oxygenated Thinners Some Industrial
Solvents Fuel Additives Solvents
Surface Preparation Metal Cleaners/
Solvents Gloss Removers
Single Component .
Products Turpentme Products Some Blended
Some Blended
Others- Some Blended Products
. Products . .
Miscellaneous Products . . Various Specralty
Various Specralty
Some Enamel Products
Products
Reducers

 

135

Table A.2 Chromatographic and Mass Spectral

Characteristics of ASTM Ignitable Liquid Classes [4,5]

 

 

 

 

 

 

 

 

 

Class Alkanes Cycloalkanes Aromatics Polynuclear
Aromatics
affirm-a“ Present, less Present, less
. . ’ abundant than abundant than Abundant Present
including . .
aromatlcs aromatics
gasohol
Present
Abundant, Present, less Present, less (depending on
Petroleum . . .
Dis tilla tes Gaussran abundant than abundant than b0111ng range),
d1str1butron alkanes alkanes less abundant
than alkanes
Branched
alkanes
Isoparaffinic abundant, n-
Products alkanes absent Absent Absent Absent
or strongly
diminished
A t' Abundant
roma 1c Absent Absent Abundant (depending on
Products . .
borlmg range)
Branched
Naphthenic $323123? n-
Paraffinic ’ Abundant Absent Absent
alkanes absent
Products
or strongly
diminished
n-Alkanes Abundant Absent Absent Absent
Products
De- Abundant, Present, less Absent or Absent or
Aromatized Gaussian abundant than strongly strongly
Distillates distribution alkanes diminished diminished
nglgfelmtsw Composition may vary, presence of oxygenated organic compounds

 

136

Appendix B

PPMC Coefficients for the TIC and EIPs of Ten Neat Diesels Analyzed in Triplicate

137

 

 

DIA
DlB
Dl(‘
DZA

Table B.1 PPMC Coefficients for the TIC of Triplicate Analyses of Diesels 1—10

 

 

 

 

DlA

D6C
D7A
D7B
D7C
DSA
D813
D8C
D9A
D9B
D9C
DIOA
D10B
DIOC

139

Table B.2 PPMC Coefficients for the Alkane EIP of Triplicate Analyses of Diesels 1-10

 

 

 

DlA

D10C

Table B.3 PPMC Coefficients for the Aromatic EIP 0f Triplicate Analyses of Diesels 1-10

 

 

 

   

 

Table B.4 PPMC Coefficients for the Indane EIP of Triplicate Analyses of Diesels 1-10

D9B
D9C
DIOA
DIOB
DIOC

 

 

 

 

 

 

Table B.5 PPMC Coefficients for the OCP EIP of Triplicate Analyses of Diesels 1—10

 

 

 

 

 

DIOA
DlOB
DlOC

Table B.6 PPMC Coefficients for the PNA EIP of Triplicate Analyses of Diesels 1—10

 

143

 

 

Appendix C

PPMC Coefficients for the TIC and EIPs of Five Neat Diesels Analyzed in Triplicate

144

Umuo

Eman—

<m~o

Uvnn— MVND (van UMND Emma (8ND UNND mNNn— <NND 0::
3-: €85 .9 32.22. 33:3; 2 u: a: E £29580 02.: 5 as;

man—

 

<—ND

145

Umnn

Emma—

<MND UVND M—VND <VNQ UMND DMND <MND UNND mNNG <NNG 035 m an
3-: 232: cc use“: saith. a. .5 2.5.2 a: .8 383580 02,: 8 asap

 

<—ND

146

Umno

ammo

<mNG UVNQ mvufl <VNG UmND manna <mNn UNNQ mNNQ <NNQ U—NQ man
3-: 2295 ca 83.2% seize .3 EH ease? 2: E 353580 02.: no 2.3

 

(SNA—

147

Oman—

mMN:

(mNQ UVND MVNQ <VND UmND MMNQ <m~fl UNNQ mNND <NND U—NG m—NQ
3-: 2935 a 33“: age; a .5. 2.2:: a: E ease-ego 02,: to 2.3

 

«1ND

148

 

Umnn

mmNn

<mNQ UVNG mvnn <v~G UMNQ mmun (nun UNNn mNNQ <NNQ U—Nn m— _ ND
3-: some: a 3223.. 33:3; 2 AE— muo 2: 3. assuage 02.: no 2...:

 

<3:

149

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

dddd._ ndaad adaad monad chad mooad manad wvnad mnnad aanad mwmad waNad dnoad Nwead mmoad
ndaad ddddg Saad vooad nmnad _nnad Vmwad mnnad Swad mam-ad nvMad Novad a_nad nmnad Nmnad
adaad Saad dado; omoad wmnad mvnad _anad omnad nmnad Svad _Nmad vnmad Snad cdnad Enad
monad vocad cmoad coco; NNwad Nnnad Nanad wnnad nomad ddmad mmmad mnmad avwad amwad Cwad
mwoad nmnad mmnad NNwad doom: Ewad nmnad amnad omnad damad dmmad Namad mdaad dawad mvwad
mooad _nnad mvnad mnnad Emad dddo; mdwad agad Nomad dwmad wmmad aamad 3 mad aanad owmad
manad Vmwad _anad Nanad nmnad Nomad dado; omaad mwaad m ~ mad wmmad nomad :wad aanad a_ wad
wvnad mnnad cmnad wnnad amnad aEad emaad dddd.— _naad nomad oamad owmad ownad donad wdwad
Nnnad Swad nmnad nomad dmnad Nomad mwaad _naad dddo._ mdmad mead owmad dmwad momad mmwad
mead mamad Q wad domad dwmad dwmad m _ mad nomad mdmad ddddé mmaad nwwad _avad vaad wavad
mwNad anad _Nmad mmmad dmmad wmmad wmmad camad emmad mNaad ddddé wmwad anvad novad Smad
mamad Novad vnmad mnmad Namad aamad mdmad owmad cwmad nwwad wwmad coca: mwvad mnvad mdmad
oncad a_nad Snad anad moaad Ewad _ Sad ownad dmwad _avad anvad mwvad .dddoA mwaad Nvaad.
Nwoad nmnad odnad amwad oawad aanad aanad ocnad mdwad vwvad novad mnvad mwaad ddddg nmaad
mmoad Nmnad Snad m-wad mvwad owwad a_wad wdwad mmwad wavad Smad mdmad .Nvaad nmaad econ:
UmND MMNG <mNG UVNG DVNA— <VNA— Oman mnNA— <MNG UNNQ mama— <NNQ 059 m5: «Jun—

mNAN m—omoma we moan—.23.. exam—nth .«e mun..— <Zm 2: .5.“ 3:395er 02.: v.0 «Ea-H

Oman
mmun—
«‘an
UVNG
QVNQ
<VN:
Unma—
finan—
(dun—
UNNG
mun:
<NNG
Unna—
m una—
«SNA—

150

Appendix D

PPMC Coefficients for the TIC and EIPs of Compiled Neat and Burned Diesels

151

 

Table D.1 PPMC Coefficients for the TIC of Neat and Burned Diesels (Cloth A-C = Replicates of Burned Diesel 21 from
Cloth, Mag A-C = Replicates of Burned Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from Carpet,
B-A = Blind A, B-B = Blind B, Cl = Cloth, Mag = Magazine, Car = Carpet)

D22C

 

 

 

 

 

ESI

Table D2 PPMC Coefficients for the Alkane EIP of Neat and Burned Diesels (Cloth A-C = Replicates of Burned Diesel 21
from Cloth, Mag A-C = Replicates of Burned Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from
Carpet, B-A = Blind A, B—B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet)

D23A

B—A

B-A

 

D21A
D21 B
D21C
022A
D228
022C
D23A
D238
D23C
D24A
D24B
D24C
DZSA
DZSB
D25C
Cloth A
Cloth B
Cloth C
Mag A
Mag B
Mag C
Car A
Car 8
Car C
B-A Cl
B—B Cl

B-B
B-A Car
B-B Car

 

 

 

 

 

Table D.3 PPMC Coefficients for the Aromatic of Neat and Burned Diesels (Cloth A-C = Replicates of Burned Diesel 21
from Cloth, Mag A—C = Replicates of Burned Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from
Carpet, B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet)

 

 

 

 

991

Table D4 PPMC Coefficients for the Indane EIP of Neat and Burned Diesels (Cloth A-C = Replicates of Burned Diesel 21
from Cloth, Mag A-C = Replicates of Burned Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from
Carpet, B-A = Blind A, B-B = Blind B, C] = Cloth, Mag = Magazine, Car = Carpet)

D25A D25B

 

 

 

 

9‘51

 

 

Table D.5 PPMC Coefficients for the OCP EIP of Neat and Burned Diesels (Cloth A-C = Replicates of Burned Diesel 21
from Cloth, Mag A-C = Replicates of Burned Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from
Carpet, B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet)

D22A D228 D23A

  
 
   
 
 
   
  
 
 

Cloth 8
Cloth C
Mag A
Mag 8
Mag C
Car A
Car 8
Car C
B—A Cl
B-B CI
B—A
B—B
B-A
B-B

B-B
B-A Car
8—8 Car

 

 

 

 

ASI

 

 

Table D6 PPMC Coefficients for the PNA EIP of Neat and Burned Diesels (Cloth A—C = Replicates of Burned Diesel 21
from Cloth, Mag A-C = Replicates of Burned Diesel 21 from Magazine, Car A-C = Replicates of Burned Diesel 21 from
Carpet, B-A = Blind A, B-B = Blind B, C1 = Cloth, Mag = Magazine, Car = Carpet)

D218 D238 D24C DZSA
D21A

 

B—A Car
B—B Car

 

 

 

[8]

[9]

[10]

[11]

[12]
[13]

[14]

[15]

[16]

References

Saferstein, R. Criminalistics: An Introduction to Forensic Science, 9th ed. Upper
Saddle River, NJ: Prentice Hall, 2006.

http://www.usfa.dhs.gov/statistics/arson/index.shtm.
http://www.fbi.gov/ucr/cius_02/html/web/offreported/02-narsonl 1 .html.

American Society for Testing and Materials, ASTM E1618-O6e1. Annual Book of
ASTM Standards 14.02.

American Society for Testing and Materials, ASTM E1387-Ol . Annual Book of
AS T M Standards 1 4. 02.

http://www.astm.org/COMMIT/COMMITTEE/E30.htm.

American Society for Testing and Materials, ASTM E1385-00. Annual Book of
AS TM Standards 14. 02.

American Society for Testing and Materials, ASTM E1386-00(2005). Annual
Book of AS TM Standards 1 4. 02.

American Society for Testing and Materials, ASTM E1412-07. Annual Book of
AS TM Standards 1 4. 02.

American Society for Testing and Materials, ASTM E1413-O7. Annual Book of
AS T M Standards 14. 02.

American Society for Testing and Materials, ASTM E2154-01. Annual Book of
AST M Standards 14.02.

Smith, R.M. Analytical Chemistry 1982, 54, 1399A-1409A.
Mann, D.C. Journal of Forensic Sciences 1987, 32, 606-628.

Barnes, A.T.; Dolan, J .A.; Kuk, R.J.; Siegel, J .A. Journal of Forensic Sciences
2004, 49, 1018-1023.

Sigman, M.E.; Williams, M.R.; Ivy, R.G. Analytical Chemistry 2007, 79, 3462-
3468.

Sandercock, P.M.L.; Du Pasquier, E. Forensic Science International 2003, 134, 1-
10.

158

[17]

[18]

[27]

[28]

[29]

[30]

[31]

[32]

Sandercock, P.M.L.; Du Pasquier, E. Forensic Science International 2004, 140.
71-77.

Sandercock, P.M.L.; Du Pasquier, E. Forensic Science International 2004, 140,
43-59.

Tan 8.; Hardy, J.K.; Snavely, R.E. Analytica Chimica Acta 2000, 422, 37-46.

Doble, P.; Sandercock, M.; Du Pasquier, E.; Petocz, P.; Roux, C.; Dawson. M.
Forensic Science International 2003, 132, 26-39.

Wang, Z.; Yang, C.; Fingas, M.; Hollebone, B.; Peng, X.; Hansen, A.B.;
Christensen, J .H. Environmental Science & Technology 2005, 39, 8700-
8707.

Wang, Z.; Yang, C.; Hollebone, B.;Fingas, M. Environmental Science &
Technology 2006, 40, 5636-5646.

Gaines, R.B.; Hall, G.J.; Frysinger, G.S.; Gronlund, W.R.; Juaire, K.L.
Environmental Forensics 2006, 7, 77-87.

Gaines, R.B.; Frysinger, G.S.; Hendrick-Smith, M.S.; Stuart, J .D. Environmental
Science & Technology 1999, 33, 2106-2112.

Wang, F.C.; Qian, K.; Green, L.A. Analytical Chemistry 2005, 77. 2777-2785.

Johnson, K.J.; Wright, B.W.; Jarman, K.H.; Synovec, R.E. Journal of
Chromatography A 2003, 996, 141-155.

Brereton, R.G. Applied Chemometrics for Scientists. Hoboken, NJ: John Wiley &
Sons, Ltd, 2007.

Bertsch, W. Journal of Chromatography A 1994, 674, 329-333.

Cavanagh, K.; Du Pasquier, E.; Lennard, C. Forensic Science International 2002.
125, 22-36.

Lentini, J.L.; Dolan, J .A.; Cherry, C. Journal of Forensic Sciences 2000, 45, 968-
989.

Almirall, J .R.; Furton, K.G. Journal of A nalytical and Applied Pyrolysis 2004, 7/ .
51-67.

Borusiewicz, R.; Zieba-Palus, J .; Zadora, G. Forensic Science International 2006.
160,115-126.

159

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44]

Hupp, A.M.; Marshall, L.J.; Campbell, D.I.; Smith, R.W.; McGuffin, V.L.
Analytica Chimica Acta 2008, 606, 159-171.

Devore, J .L. Probability and Statistics for Engineering and the Sciences, 4th ed.
Belmont, CA: Duxbury Press, 1995.

Smith, L.I. A tutorial on Principal Components Analysis.
http://www.seas.harvard.edu/courses/cs171/lectures/principal__
componentspdf, 2002.

Freudenrich, C.C. Science Channel: How Stufl Works.
http://science.howstuffvvorks.com/oil-refining.htm.

“Advanced Fire Debris Analysis Workshop.” Presented by F 81 Laboratory
Explosives Unit. Indianapolis, IN, April 15-17, 2008.

Skoog, D.A.; Holler, F .J .; Crouch, S.R. Principles of Instrumental Analysis. 61}]
ed. Belmont, CA: Thomson, 2007 .

“Advanced Fire Debris Analysis Workshop.” Presented by Midwestern
Association of Forensic Scientists. Traverse City, MI, September 24.
2007.

Ignitable Liquid Reference Collection. http://www.ncfs.org/ilrc1.html.

Nielsen, N.V.; Carstensen, J .M.; Smedsgaard, J. Journal of Chromatography A
1998, 805, 17-35.

Tomasi, G.; van den Berg, F.; Andersson, C. Journal of Chemometrics 2004, 18,
231-241.

van Nederkassel, A.M.; Daszykowski, M.; Eilers, P.H.C.; Vander Heyden. Y.
Journal of Chromatography A 2006, I I I 8, 199-210.

Pierce, K.M.; Hope, J.L.; Johnson, K.J., Wright, B.W.; Synovec, R.E. Journal of
ChromatographyA 2005, 1096, 101-110.

160