Forensic Chemistry 42 (2025) 100633

Contents lists available at ScienceDirect

Forensic Chemistry
journal homepage: www.sciencedirect.com/journal/forensic-chemistry

In silico created fire debris data for Machine learning

Michael E. Sigman a,b, Mary R. Williams a,* , Larry Tang a,c , Slun Booppasiri c , Nikhil Prakash a
a
National Center for Forensic Science, University of Central Florida, Orlando, FL 32816 USA
b
Department of Chemistry, University of Central Florida, Orlando, FL 32816 USA
c
Department of Statistics and Data Analytics, University of Central Florida, Orlando FL 32816 USA

A R T I C L E I N F O A B S T R A C T

Keywords: Attacking complex forensic problems, such as the classification of fire debris data as positive or negative for
Fire debris ignitable liquid residue (ILR), requires large amounts of training data if machine learning approaches are to be
Machine learning successful. This work examines the in-silico preparation of computed fire debris data for training a machine
Digital weathering
learning method to classify gas chromatography – mass spectrometry (GC–MS) data as positive or negative for
XGBoost
ILR, and reports the outcome of validation tests on a set of laboratory-generated fire debris samples with known
ground truth. A set of 240,000 total ion chromatograms (TIC) and total ion spectra (TIS) for fire debris (FD)
samples were calculated in silico (IS). The IS FD sample set was balanced with 50% of the samples containing
ignitable liquid residue (ILR) and substrate pyrolysis (SUB) contributions. The remaining 50% contained only
SUB components. The ignitable liquids incorporated into the samples containing ILR were digitally evaporated to
simulate weathering observed in experimental fire debris. The IS FD sample TIS were treated by principal
component analysis (PCA) with centering and variance scaling and retaining 90% of the variance. A set of 1,117
experimental FD samples were projected into the IS FD PCA model. The recovered experimental FD TIS were
compared to the TIS before projection by calculating the residual mean squared error (RMSE) for each sample as
a test of the IS FD samples representation of experimental samples. The range of the RMSE was [ 0.012, 0.127]
and the median RMSE was 0.029. Experimental FD samples where the recovered TIS had the larger RMSE values
were not well-represented by the IS FD samples. The IS FD samples were randomly split into balanced sets for
machine learning (ML) training (90%) and validation (10%). An XGBoost ML method, trained on the IS FD
training data, was validated on the testing IS FD data, giving a receiver operating curve (ROC) with area under
the curve (AUC) of 0.978. Validation of the model against the experimental FD data gave a lower ROC AUC of
0.845. Limiting the experimental data to samples in the lowest quadrant of RMSE values increased the ROC AUC
to 0.90.

1. Introduction machine learning in the classification of gas chromatography-mass

spectrometry (GC – MS) data from fire debris samples as positive or
Machine learning methods often require large amounts of training negative for the presence of ignitable liquid residue (ILR). The method of
data. The amount of training data that is required depends on the computationally combining ignitable liquid and substrate GC – MS data
complexity of the machine learning model (i.e., neural networks, sup­ is referred to as in silico (IS). The IS data can also be used to provide
port vector machines, linear discriminant analysis, etc.) and the analysts with visual representations of fire debris (FD) samples with a
complexity of the problem. If an insufficient amount of training data is prescribed composition. Currently, analysts can only mentally synthe­
available, data augmentation methods can be used to generate larger size an image of a FD total ion chromatogram (TIC) resulting from a
data sets, or transfer-learning can sometimes be used [1,2]. In computer partially evaporated ignitable liquid and pyrolysis of several substrates.
vision applications, augmented images are commonly used for machine The construction of IS FD allows for the generation of mixtures having
learning methods to boost generalization and improve robustness [3]. In user-specified composition, and allows for the development of training
this paper we describe a data augmentation method that has been used data for machine learning applications.
to computationally create a training data set to facilitate the use of The methodology reported here represents an expansion of IS

* Corresponding author.
E-mail address: [email protected] (M.R. Williams).

https://doi.org/10.1016/j.forc.2024.100633
Received 1 October 2024; Received in revised form 3 December 2024; Accepted 3 December 2024
Available online 6 December 2024
2468-1709/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

methods previously reported from our laboratory [4–9]. As described in “burned” using the 2 min modified destructive distillation method
previous work, each single data record for FD visualization is generated (MDDM) [21]. The experimental FD data consisted of the 1117 samples
from a linear combination of GC – MS data from a single ignitable liquid from the Fire Debris Database [22]. Sample files were converted from
(IL) with the GC – MS data obtained from pyrolysis and/or partial their original Agilent files to common data format (.cdf). File mixing in
combustion of some number, n, of common building materials and the IS process is described in the following section.
household furnishings. In this work, a set of substrate pyrolysis samples
are combined into a single IS FD sample and collectively referred to as 3. Theory/Calculations
SUB (upper case) to differentiate the set from individual substrates
comprising a SUB, which are referred to as “subi ” (lower case), (i.e., The computational procedure used in this work is outlined in Fig. 1.
SUB = {subi , i = 1, 2⋯n}, where in this work n = 3). The linear com­ The procedure involved computationally generating the recipes for
bination model is supported by results from co-pyrolysis of polymers, many IS FD samples, and then using the compositional recipes to
biomaterials and mixtures of these materials [6,10–12]. generate and save the total ion chromatographs (TIC) and total ion
This work expands on previous efforts in two important respects. spectra (TIS) for each sample. The TIC and TIS can subsequently be
Firstly, in this work the GC – MS signal is deconstructed into contribu­ tested as training data for machine learning applications to fire debris.
tions from background and instrumental factors (e.g. solvent tailing, The TIC is calculated by summing the ion intensities across all mass
background signal and column bleed), designated as “noise”, and MS channels at each RI to produce a “chromatogram” for the mixture. The
response attributed to the analytes (e.g. IL or SUB) and referred to here TIS is calculated by summing across all RI values to produce a “mass
as “signal”. Separating the two responses allows the signal-to-noise ratio spectrum” for the mixture.
to be simulated over a relevant range. Secondly, single IL samples are The first step was to define the parameters for the set of IS FD to be
digitally weathered to allow representations at many different extents of generated. The total number of IS FD in a single set was specified
evaporation (also known to fire debris analysts as weathering). Sepa­ (typically 60,000) along with the fraction of those samples that would
ration of the signal components from the noise allows the IL component only contain SUB. The remaining fraction of the samples would contain
to be digitally weathered independent from the SUB and noise. When an IL, as well as SUB, and a stipulation of the fraction of IL-containing
IL is present in a fire, the recovered residue may exhibit partial or high samples drawn from each of the ASTM E1618 classes: normal alkanes
extents of evaporation. Physically evaporating an IL in the laboratory is (NAL), aromatics (AR), gasoline (Gas), isoparaffinic (ISO), miscella­
an arduous and time-consuming task. Digital evaporation, on the other neous (MISC), naphthenic paraffinic (NP), petroleum distillates (PD)
hand, can be performed reproducibly and rapidly. In this work, digital and oxygenated solvents (OX). Within ASTM E1618, the PD class is often
evaporation is applied to unevaporated IL data as part of the data broken into subclassifications of light (LPD), medium (MPD) and heavy
augmentation process [13–18]. Previous papers describing the creation (HPD) distillates based on carbon range of the components. In this work
of IS FD samples was limited to the use of samples from the Ignitable a 0.5 fraction of the samples contained IL and the remaining 0.5 fraction
Liquids Reference Collection and Database (ILRC) that had been contained only pyrolyzed substrates (SUB). The fraction of IL-containing
experimentally evaporated to remove 25, 50, 75, 90 or 99 percent of the samples drawn from each ASTM class was evenly distributed across all 8
original liquid [4,5,7–9,19,20]. classes.
In this work, the fraction of IL remaining at each RI value after digital The second step in the procedure was to computationally generate a
evaporation was calculated by Equation (1), adapted from [18]. In Eq. list of sample compositions. Each IL-containing sample composition
(1), RI corresponds to the retention index and t corresponds to the specified a single unevaporated IL and a randomly determined number
evaporation time. Calculating FRI by Eq. (1) for a set of retention indices of specified substrates, subi (i = 1 − 3), for mixing. The relative
at a selected evaporation time results in an evaporation curve. The curve compositional contribution of IL (i.e., IL fraction) in each sample was
can then be multiplied by the unevaporated IL total ion chromatogram determined from a uniform random distribution in the range of 0.1 to
(TIC) represented on the same retention index scale to produce an es­ 0.9. The fractional contribution of SUB was adjusted to bring the total
timate of an evaporated TIC. When the targeted fraction of evaporative sum of IL and SUB to one. The fractional contribution of SUB was
loss is specified, the evaporation time is varied until the ratio of the divided among the randomly selected number (n) of subi=1..n by drawing
integrated area for the digitally evaporated TIC divided by the inte­ n values on the range 0 to 1 from a uniform random distribution. The
grated area of the unevaporated TIC equals the targeted value. For relative substrate contributions were adjusted to keep the total IL and
example, the evaporation curve resulting from Eq. (1) would be calcu­ SUB fractions at one. In this work the maximum number of substrates
lated at arbitrary values of t until the integrated area of the unevapo­ was set at three. Finally, the signal-to-noise ratio, as discussed above,
rated liquid had been decreased by a targeted fraction (i.e., 0.75 for 75 was selected from a random uniform distribution on the range of 1 to 30.
% evaporation). As previously noted, the chromatogram for a sample Samples that did not contain an IL component were constructed by the
digitally evaporated to a targeted fraction will differ from the chro­ same procedure, except the IL fractional contribution was set to zero.
matogram of the liquid physically evaporated to the same targeted The seed for the random number generator within R was set by the user
fractional volume. The relationship between digital and experimental before generating the set of compositions. If additional sets were
evaporation is also examined in this work. created, the seed was changed before the new list was created. In this
( ( ) ) work, four sets of 60,000 IS FD each, were generated to give a total of
FRI = exp(− exp − 1.04 × 10− 2 RI + 6.07 t ) (1)
240,000 IS FD samples.
Step 3 of the computational process involved sequentially reading
In addition to generating a set of 240,000 IS FD samples, the accuracy of
the composition for each IS FD sample and proceeding to retrieve and
the IS data in representing experimental FD is evaluated. Finally, we test
mix the specified files according to the recipe, as illustrated in Fig. 1.
the IS FD samples as training data for a single machine learning model to
Each IL or subi file was read from storage and processed by super­
classify experimental FD as positive or negative for the presence of ILR.
imposing the GC – MS data onto a matrix composed of 2,800 rows,
corresponding to RI values, and 271 columns, corresponding to nominal
2. Materials and methods
m/z values in the range of 30 – 300. The conversion from retention time
to RI was computed using a spline function in the R software [23]. The
The IL data utilized in the IL FD construction were from 669 une­
RI values were rounded to integers in the range 1 – 2,800. Slight vari­
vaporated IL in the Ignitable Liquids Reference Collection and Database
ations in MS scan times resulted in a few replicate RI values in some
(ILRC) [19]. The subi components data used in the calculations corre­
samples after rounding. In these cases, the MS scan with the largest sum
sponded to the 299 samples from the Substrates Database that had been
across all m/z (i.e., the largest TIC intensity) was retained and the other

2
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

Fig. 1. Outline of 3-step process of developing fire debris samples in silico.

scan data was not incorporated into the matrix. Keeping the largest sum software written in-house. The software for defining the compositions
gave the best reproduction of the instrument-generated TIC. The “noise” and building the IS FD was written using the R statistical package [23].
contributions contained in columns corresponding to intensities for m/z
32, 76, 207 and 281, were replaced with columns of zeroes to produce a 4. Results
“signal” component matrix. The noise ions correspond to a minimal air
background, CS2 solvent and column bleed. The intensities of the noise 4.1. Digital and experimental weathering
ions were added to a new 2,800 x 271 matrix which contained intensities
of zero in all entries before addition of the noise ions. Each of the signal As previously noted by Smith and coworkers, the TICs from digital
and noise matrices were scaled so that the maximum intensity in the TIC and experimental evaporation of an IL to the same fraction (i.e. 50 % loss
corresponding to each would have a value of one. The signal and noise in integrated signal from digital evaporation and 50 % volume loss by
matrices for the IL and each substrate were then scaled by the fractional physical evaporation) are not equivalent [14,15,17]. Furthermore, the
contribution prescribed in the recipe. The scaled signal matrices were extent of the difference between TICs of experimental and digital
combined by index-wise addition, as were the noise matrices, and both evaporation to a specified fraction could vary across IL classes, as
resulting summed matrices were normalized to give a maximum TIC defined in ASTM E1618 – 19 [24]. Experimental evaporation percent­
intensity of one. The signal matrices were then multiplied by the spec­ ages for samples in the ILRC correspond to the fraction of the volume
ified signal-to-noise ratio. The sample and background matrices were removed, whereas digital percent evaporation corresponds to the frac­
subsequently added index-wise, and the resulting matrix normalized tion of the chromatographic integrated area digitally removed. Gener­
again to provide a TIC with maximum intensity of one. ating digitally weathered samples that approximate the volume-fraction
In step 3, the intensity-normalized TICs were multiplied by the weathering given in the ILRC requires a reasonable model between
fractional mixing contributions, rather than using the integration- digital and experimental evaporations. The model need only be accurate
normalized TIC profiles to provide more “visually” consistent data. enough to produce reasonable evaporated profiles to include in digital
Using the ratios of integrated areas, while perhaps more chemically training data. Higher accuracy would be required if the purpose of the
correct, would result in computational FD where the visual contribution machine learning were extended to predicting the extent of evaporation.
of IL relative to the SUB, was less intuitive. For example, in a case where A simple linear model was tested to investigate the relationship between
a normal alkane was mixed with a carpet pyrolysis sample. Since the experimental and digital evaporation fractions.
normal alkane typically has only three or four peaks, the alkane peaks A set of TICs from unevaporated ignitable liquids were converted to
would appear very large compared to the substrate, even when the RI and digitally evaporated by Eq. (1) to achieve the best match between
fractional contribution of the alkane was small. The compositions of fire the digital evaporation TIC and the TIC of the same liquid that had been
debris samples that did not contain ILR were specified in the same way experimentally evaporated to a specified volume reduction. Prior to the
as those that did contain ILR, except that an IL and the IL fractional digital evaporation, the three largest peaks remaining in the evaporated
contribution were not designated. sample TIC were aligned with the same peaks in the unevaporated
The computational mixing was performed as described above using sample TIC using the previously reported procedure [25]. The best

3
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

match was judged based on the largest Pearson product moment cor­ Table 1
relation coefficient (PPMC) that could be obtained. After the best PPMC IL class, ILRC sample reference numbers (SRNs) from the unevaporated (SRN0)
was determined, the fractional loss from digital evaporation was and evaporated (SRN1) liquids, experimental fraction evaporated by volume
determined. Table 1 provides the IL class, ILRC sample reference reduction (FExp ), carbon range for the unevaporated liquid, fractional reduction
numbers (SRNs) from the unevaporated (SRN0) and evaporated (SRN1) in integrated area of digitally evaporated sample relative to the unevaporated
sample (FDig ), and the largest PPMC between the experimental and digitally
liquids (see [19]), experimental evaporated volume reduction, carbon
evaporated TICs.
range for the unevaporated liquid, fractional reduction in integrated
area of digitally evaporated sample relative to the unevaporated sample, Class SRN0 SRN1 FExp Low C High C FDig PPMC
and the largest PPMC between the experimental and digitally evapo­ AR 5 724 0.25 8 12 0.32 0.94
rated TICs. AR 5 725 0.50 8 12 0.54 0.93
As described above, the ILRC provides chromatographic data for AR 5 726 0.75 8 12 0.76 0.91
AR 5 727 0.90 8 12 0.83 0.83
samples that have been evaporated to a fraction of their initial volume. AR 5 728 0.95 8 12 0.88 0.90
However, the chromatogram for a sample that has been digitally evap­ AR 6 1125 0.50 9 13 0.45 0.87
orated to remove a specified fraction of the integrated area (e.g., 75 %) AR 6 1126 0.75 9 13 0.72 0.93
will not have the same profile as a sample that has been physically AR 6 1127 0.90 9 13 0.73 0.95
AR 74 1147 0.75 8 12 0.74 0.86
evaporated to remove the same fraction of the volume (e.g., 75 %). The
Gas 105 733 0.25 6 13 0.26 0.84
best linear fit through the data is given by Equation (2), where FDig Gas 105 734 0.50 6 13 0.41 0.88
represents the fraction of the TIC removed by digital evaporation and Gas 105 735 0.75 6 13 0.49 0.84
FExp represents the experimental fraction of the volume removed by Gas 105 736 0.90 6 13 0.71 0.82
weathering. The linear relationship explains 97 % of the variance in the HPD 14 807 0.25 8 17 0.25 0.93
HPD 14 808 0.50 8 17 0.45 0.90
data. The intercept of the linear fit described in Eq. (2) was not statis­ HPD 14 810 0.90 8 17 0.85 0.94
tically different from zero. HPD 50 966 0.25 8 26 0.20 0.97
ISO 89 886 0.25 11 16 0.33 0.95
FDig = 0.89 × FExp (r = 0.98, r2 = 0.97, n = 67) (2) ISO 89 887 0.50 11 16 0.54 0.95
ISO 89 888 0.75 11 16 0.75 0.95
The data from Table 1 is plotted in Fig. 2 with the best linear fit shown as ISO 89 889 0.90 11 16 0.82 0.95
a solid line and the 95 % confidence bands shown as dashed lines. From ISO 89 890 0.95 11 16 0.82 0.91
ISO 119 679 0.25 9 12 0.33 0.98
the data in Table 1, there is no clear and consistent trend in the ordering
ISO 119 680 0.50 9 12 0.38 0.96
of most samples from the various ASTM E1618 classes. The MISC class ISO 119 681 0.75 9 12 0.53 0.97
samples appear to be at the top of the digital evaporation range at each ISO 119 682 0.90 9 12 0.67 0.96
experimental evaporation. The ISO class samples tend toward the bot­ ISO 119 683 0.95 9 12 0.81 0.96
LPD 35 742 0.25 6 10 0.29 0.90
tom of the digital evaporation range at each experimental evaporation
LPD 35 743 0.50 6 10 0.45 0.92
and follow a non-linear relationship. LPD 35 745 0.90 6 10 0.84 0.95
Fig. 3 shows box plots for the optimal PPMC values for comparison of LPD 35 746 0.95 6 10 0.93 0.92
the experimental and digitally evaporated chromatograms (from LPD 92 1155 0.50 7 9 0.23 0.83
Table 1), as observed for each ASTM E1618 Class. As shown in Fig. 3, the MISC 16 775 0.25 8 26 0.42 0.89
MISC 16 776 0.50 8 26 0.56 0.93
isoparaffinic (ISO) class had the highest average PPMC and the medium
MISC 16 777 0.75 8 26 0.77 0.90
petroleum distillate (MPD) class had the lowest PPMC. The heavy pe­ MISC 39 983 0.25 10 13 0.29 0.87
troleum distillate (HPD) and miscellaneous (MISC) classes had the MISC 39 984 0.50 10 13 0.38 0.90
broadest PPMC interquartile ranges and the MISC PPMC values are MISC 39 985 0.75 10 13 0.66 0.89
negatively skewed. A trend of decreasing PPMC with increasing exper­ MISC 39 986 0.90 10 13 0.84 0.89
MPD 4 1030 0.25 8 12 0.42 0.88
imental evaporation can be seen for the HPD class in Table 1. A similar MPD 4 1033 0.90 8 12 0.9 0.94
trend is not established for the MISC class, as the two samples with the MPD 4 1034 0.95 8 12 0.91 0.94
lowest PPMC and the three with the highest PPMC originate from MPD 21 670 0.25 8 13 0.28 0.85
different unevaporated IL (i.e., SRN0 39 and 16, respectively). A more MPD 21 671 0.50 8 13 0.44 0.85
MPD 24 1132 0.50 8 11 0.36 0.84
important factor is retention time shifts between the unevaporated and
MPD 24 1133 0.75 8 11 0.35 0.87
evaporated TIC. Chromatographic shifts are due in part to the conver­ MPD 24 1134 0.90 8 11 0.52 0.83
sion of the retention time for each mass scan to the retention index scale NAL 77 688 0.25 10 14 0.13 0.83
with rounding of the RI to the nearest integer value. An addition factor NAL 77 689 0.50 10 14 0.24 0.87
contributing to the peak shift in RI may be the evaporation of com­ NAL 77 691 0.90 10 14 0.58 0.85
NAL 192 825 0.25 12 15 0.28 0.94
pounds thus reducing the intensity which sometimes narrows the peak. NAL 192 828 0.90 12 15 0.80 0.92
The simple linear relationship in Eq. (2) allows for an approximate NP 53 697 0.25 9 16 0.22 0.87
conversion between FDig and FExp . The IS FD samples prepared in this NP 53 698 0.50 9 16 0.46 0.91
work were created to a specified digital evaporation fraction. The NP 53 699 0.75 9 16 0.73 0.94
NP 53 700 0.90 9 16 0.86 0.96
intended purpose in examining the match between digital and experi­
NP 53 701 0.95 9 16 0.93 0.90
mental evaporation is to determine an approximate relationship that NP 140 141 0.25 9 15 0.29 0.93
allows identification of TICs which approximately match the analyst’s NP 140 142 0.50 9 15 0.47 0.94
expectation based on experience stemming from laboratory evapora­ NP 140 143 0.75 9 15 0.68 0.93
tions. The simple linear relationship in Eq. (2) is sufficient to meet this NP 140 769 0.90 9 15 0.86 0.87
OXY 182 1085 0.25 5 9 0.35 0.83
need. OXY 182 1086 0.50 5 9 0.42 0.92
Furthermore, the data in Table 1 demonstrates the application of OXY 182 1087 0.75 5 9 0.80 0.92
Eq.1 to experimental data ranging from an RI = 500 to an RI = 2600. The OXY 182 1088 0.90 5 9 0.86 0.94
data demonstrates the use of Eq. (1) to digitally evaporate the chro­ OXY 182 1089 0.95 5 9 0.89 0.93
OXY 218 852 0.25 7 9 0.00 0.90
matographic profile of an unevaporated liquid to give a best match with
an experimentally evaporation of the same liquid. While the matches
between the experimentally and digitally evaporated chromatographic

4
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

IL, the resulting matrix was renormalized to give a maximum TIC in­
tensity of one. The evaporated IL was mixed in a 1-to-1 ratio with the
single substrate. The intensity-normalized TICs of the IL and SUB are
shown in the top row and second from top row in Fig. 4. The noise
contributions from the IL and SUB were also mixed in a 1-to-1 ratio,
renormalized to give a TIC with maximum intensity of one and then
divided by the signal-to-noise ratio of 15. The resulting noise TIC is
shown in the third from top TIC in Fig. 4. The bottom TIC in Fig. 4
represents the IS FD achieved by summing the IL and SUB signals with
the noise.
Fig. 5 shows the construction of the same IS FS sample from Fig. 4. In
Fig. 5, the graph in the lower lefthand corner shows the full GC – MS data
sets for the IL and SUB, along with the noise contribution as a contour
plot. The upper graph shows the IS FD TIC that results from summing the
ion intensities at each RI value. The graph on the right shows the TIS that
comes from summing each m/z across all RI values, with subsequent
intensity normalization.

4.3. Testing the representative Nature of the IS FD

One of the motivations for generating an IS FD data set is to use that

data to train machine-learning method. The resulting models should be
able to accurately predict the presence or absence of ILR in a “experi­
Fig. 2. Linear model of FDig as a function of FExp . The linear relationship from
mental” FD sample. Applying the IS FD for machine learning (ML)
Eq. (2) is shown by the solid line (r = 0.94, r2 = 0.88, n = 50) with the 95 %
training requires that the IS data and the experimental FD data are from
confidence bands shown as dashed lines.
the same population, otherwise the resulting model will not accurately
predict the presence or absence of ILR in experimental samples. As a
means of testing the IS FD to determine if it was representative of
experimental FD, ions representative of IL classes from Table 1 of ASTM
E 1618–19 were selected from the IS FD data. The selected ions were
indexed in increasing order as Index(m/z): 1(31), 2(43), 3(45), 4(55), 5
(57), 6(58), 7(69), 8(71), 9(72), 10(82), 11(83), 12(85), 13(86), 13(91),
15(92), 16(99), 17(105), 18(106), 19(117), 20(118), 21(119), 22(120),
23(128), 24(131), 25(132), 26(134), 27(142), 28(156). The selected
data was subjected to principal components analysis (PCA) to reduce the
dimensionality of the data and reduce noise in the data while keeping
90 % of the variance (14 principal components). PCA also eliminates
correlation among factors by producing an orthonormal set of eigen­
vectors and representing each IS FD sample with a set of scores. The
extracted ions from the IS FD data were centered, and each factor was
scaled by the associated variance prior to PCA. The experimental FD
data was then centered, scaled by the variances determined for each
factor in the IS FD, and then projected into the IS FD PCA space. The
scores resulting from projection of the experimental FD samples were
then utilized to recover experimental FD. The variance scaling and mean
centering were reversed, and the recovered data was compared to the
projected data by calculating the root square mean error (RMSE) for
each experimental FD sample. Fig. 6 shows the RMSE as a function of the
sample Index for each of the experimental FD samples. The range of the
Fig. 3. Box plots for the PPMC values observed for each IL Class in Table 1. RMSE was [ 0.011, 0.304] and the median RMSE was 0.055. Fig. 7 shows
a comparison of experimental FD TIS (positive-going) and the recovered
profiles are not perfect (PPMC range 0.7–0.98), the results are deemed sample (negative-going) for (a) a sample with minimal RMSE and (b) the
sufficient to implement the method in the development of in-silico sample with the maximum RMSE.
computed fire debris data for testing as machine learning training data. The IS FD PCA space does not fully encompass the experimental FD
data; however, the samples clustered around the median RMSE of 0.055
4.2. Example IS FD construction result are represented within the PCA space, as indicated by the low RMSE
between the projected and recovered TIS, see Fig. 7(a). Experimental FD
An example IS FD sample and the set of TIC used to create the sample samples with an RMSE between the test and recovered TIS that ap­
are shown in Fig. 4. The sample is composed from a gasoline sample, proaches the maximum RMSE are not well represented within the IS FD
SRN 258 from the ILRC [19], and a single olefin – polyester carpet PCA space and are unlikely to be modeled well by a ML routine trained
substrate, MRN 751 in the Substrate database [21]. Both samples were on the IS FD PCA scores.
decomposed into the signal and noise contributions. The ignitable liquid The experimental fire debris sample which demonstrated the largest
signal TIC was digitally evaporated to 50 % of the integrated intensity of RMSE when the extracted TIS ion intensities were compared with the
SRN 258. After the evaporation time was determined from the TIC, each recovered ions was a fire debris sample containing only pyrolysis
m/z channel was evaporated by the application of Eq. (1) using the products from a single substrate. The substrate corresponded to an air-
established evaporation time. Following the digital evaporation of the freshener (automotive accessory), MRN 1780 in the Substrates

5
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

Fig. 4. Example TIC comprising an IS FD sample. The sample is composed of a gasoline sample, SRN 258 from the ILRC, digitally evaporated to 50% and a single
olefin – polyester carpet substrate, MRN 751 in the Substrate database. The signal-to-noise ratio for the mix is 15:1.

Database. This substrate was not included in the 240,000 IS FD samples. was done by a grid search, coded in house, running on a MacBook Pro
The experimental fire debris sample demonstrating the second largest equipped with 8 GB memory and Apple M2 chip. The IS FD validation
RMSE (see Fig. 6) also contained a single substrate. In this case, the ROC curve is shown in Fig. 9. The predicted probability that a sample
pyrolyzed substrate was MRN 1371 from the Substrate Database. The contained IL given the TIS evidence, p(IL|E), was converted to the
substrate was composed of apparel (worn clothing) and MRN 1371 loglikelihood ratio using Eq. (3), which is a rearranged odds form of
appeared 857 times in combination with other substrates and evapo­ Bayes’ equation. The base rates, p(IL) and p(IL), were both 0.5 based on
rated liquids in the 240,000 IS FD training samples. Boxplots showing the balanced training set. The loglikelihood ratio was used as the score
the distributions of the RMSE for each IL class and the SUB class in the for plotting the ROC curve. The area under the ROC curve was 0.965 and
experimental FD is shown in Fig. 8. The data depicted in Fig. 8 dem­ the operational loglikelihood ratio was 0.06 based on a Neyman-Pearson
onstrates a lower overall RMSE for experimental FD samples containing false positive rate of 0.05 [26,27]. The
HPD, and the two largest RMSE values were outliers in the distribution ( ) ( )
of SUB (samples containing no IL residue). p(E|IL) p(IL|E) p(IL)
log10 = log10 × (3)
p(E|IL p(IL|E) p(IL)
4.4. Testing the IS FD as a ML training data set The experimental FD data was projected into the PC space and the scores
were used to predict the probability that the sample contained IL. An
Four sets of IS FD, each containing 60,000 records, were combined to ROC curve based on the loglikelihood ratios of the predicted probabil­
give 240,000 total IS FS records. Ten percent of the data (24,000 re­ ities had an AUC of 0.85, which is substantially lower than the IS FD
cords, 12,000 containing IL and 12,000 containing only substrates) was validation. Limiting the experimental FD samples to those in the lowest
randomly selected and removed to serve as an IL FD validation. Two quartile of the RMSE values gave an ROC with an AUC of 0.899. As the
background ions (m/z = 32 and 76) were removed from the IS FD TIS second and third quartiles of the RMSE were included in the predictions,
spectra, along with all other ions not selected from Table 1 of ASTM the AUC decreased to the value of 0.85 when all four quartiles were
E1618-19 (see Section 4.3). The training data was subjected to PCA for included, see Fig. 10. This result demonstrates that even with 234,000 IS
further feature reduction. Keeping 90 % of the variance resulted in 33 FD training samples, some of the experimental FD samples are not from
principal component scores comprising the IS training data. the same population as the IS FD samples and were not predicted well by
The XGBoost algorithm was applied to the IS FD training data scores. the XGBoost method. One approach to address this problem would be to
The model was optimized by adjusting the tree depth 1 – 13 in steps of 4, increase the size of the IS training set. Another approach would be to
the number of training rounds on the range of 1 – 150 in steps of 50, the pre-select experimental samples that could be modeled by calculating
learning rate on the range of 0.1 – 0.5 in steps of 0.2, and the regula­ the RMSE between the experimental sample and the sample recovered
rization was optimized on the range 1 – 5 in steps of 2. The optimization

6
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

Fig. 5. A display of the same IS FD sample from Fig. 4. In this display, the top plot shows the composite TIC, the right spectrum shows the composite TIS, and the
lower left plot shows the full GC – MS data set from the IS FD sample.

Fig. 6. The root mean square error (RMSE) is show for the comparisons of
experimental FD samples projected into the IS FD PCA space and then recov­
ered. The solid line at 0.055 corresponds to the median RMSE. Fig. 7. (a) Comparison of the positive-going experimental TIS and the negative-
going recovered TIS for the experimental FD sample with the minimum RMSE.
from the PCA space determined from the training data. Only experi­ (b) Comparison of the positive-going experimental TIS and the negative-going
mental samples with very low RMSE values would be predicted with recovered TIS for the experimental FD sample having the maximum RMSE.
confidence.

7
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

Fig. 10. The upper ROC curve corresponds to the IS FD validation ROC shown
in Fig. 8. The other curves correspond to the ROC curves resulting from
Fig. 8. Boxplots showing the distributions of the RMSE for each IL class and the experimental FD sample LLR calculation using the model based on IS FD
SUB class in the experimental FD. samples. Keeping only the quartile of samples having the lowest RMSE results in
the highest AUC of 0.899. Keeping all experimental FD samples results in a
lower AUC or 0.85.

represented by the IS data. Projecting experimental FD test data into the

IS FD PCA space and then examining the RMSE between the recovered
sample and the projected sample provides insight into how well the
experimental sample is covered by the IS FD samples. Consequently, the
projection-and-recovery method gives some confidence in the outcome
of the classification based on XGBoost.

CRediT authorship contribution statement

Michael E. Sigman: Writing – original draft, Validation, Supervi­

sion, Resources, Project administration, Methodology, Investigation,
Funding acquisition, Formal analysis, Conceptualization. Mary R. Wil­
liams: Writing – review & editing, Supervision, Resources, Methodol­
ogy, Funding acquisition, Data curation, Conceptualization. Larry
Tang: Writing – review & editing, Validation, Investigation, Funding
acquisition, Formal analysis. Slun Booppasiri: Writing – review &
editing, Validation, Investigation, Formal analysis. Nikhil Prakash:
Data curation.

Declaration of competing interest

Fig. 9. ROC curve for the test IS FD validations samples. The vertical line
The authors declare that they have no known competing financial
corresponds to a false positive rate of 0.05, which intersects the ROC curve at a
interests or personal relationships that could have appeared to influence
loglikelihood ratio (LLR) of 0.06. The intersection represents the operational
the work reported in this paper.
LLR for making a classification decision on an IS FD sample as positive for ILR
while keeping the FPR at or below 5%. The horizontal arrow represents TPR of
0.87 that would be observed for IS FS test samples using this decision threshold Acknowledgements
LLR 0f − 0.1.
This research was supported by Award Number 15PNIJ-22-GG-
5. Conclusions 04421-SLFO awarded by the National Institute of Justice, Office of
Justice Programs, U.S. Department of Justice. The opinions, findings,
The in-silico method of calculating samples that are representative of and conclusions or recommendations expressed in this presentation are
fire debris provides a method of generating large training sets for ma­ those of the authors and do not necessarily reflect those of the Depart­
chine learning. Incorporating digital weathering into the fire debris ment of Justice.
calculation facilitated the generation of realistic models. The IS FD
samples were representative of many of the test experimental FD sam­ Data availability
ples; however, some experimental samples appear to have not been well
Data used in this study, including the training data computed in silico

8
M.E. Sigman et al. Forensic Chemistry 42 (2025) 100633

and the validation experimental fire debris data are available from [14] V.L. McGuffin, R.W. Smith, A unified kinetic and thermodynamic model of evaporation
for forensic applications, Forensic Chem. 23 (2021) 100304, https://doi.org/
https://ilrc.ucf.edu.
10.1016/j.forc.2020.100304.
[15] V.L. McGuffn, R.W. Smith, A Evaporation Kinetic Model Based of on Gas
References Chromatographic Retention Index, in: N. Grinberg, P.W. Carr (Eds.), Advances in
ChromatographyVolume 59, CRC Press, Boca Raton, FL, USA, 2022, pp. 33–95.
[1] K. Maharana, S. Mondal, B. Nemade, A review: Data pre-processing and data [16] J.W. McIlroy, R.W. Smith, V.L. McGuffin, Fixed-and variable-temperature kinetic
augmentation techniques, Global Transitions Proc. 3 (2022) 91–99. models to predict evaporation of petroleum distillates for fire debris applications,
[2] T.Y. Huang, J.C. Chung Yu, Assessment of artificial intelligence to detect gasoline in fire Separations 5 (2018) 47, https://doi.org/10.3390/separations5040047.
debris using HS-SPME-GC/MS and transfer learning, J. Forensic Sci. 69 (2024) [17] R. Smith, V.L. McGuffin, G.P. Jackson, Investigating Kinetic and Thermodynamic
1222–1234, https://doi.org/10.1111/1556-4029.15550. Approaches to Predict Evaporation of Gasoline at Elevated Temperatures for Fire Debris
[3] Rebuffi, S.-A., et al., Data augmentation can improve robustness. in: M. Ranzato, A. Applications, Accessed online September 2024, Office of Justice Programs’ National
Beygelzimer, Y. Dauphin, P.S. Liang and J. Wortman Vaughan (Eds.) Advances in Criminal Justice Reference Service (2022).
Neural Information Processing Systems, NeurIPS 2021 2021, pp. 29935-29948. [18] R.W. Smith, et al., Mathematically modeling chromatograms of evaporated ignitable
[4] A. Akmeemana, M.R. Williams, M.E. Sigman, Convolutional Neural Network liquids for fire debris applications, Forensic Chem. 2 (2016) 37–45, https://doi.org/
Applications in Fire Debris Classification, Chemosensors 10 (2022) 377, https://doi. 10.1016/j.forc.2016.08.005.
org/10.3390/chemosensors10100377. [19] Ignitable Liquids Reference Collection and Database. National Center for Forensic
[5] A. Allen, M.R. Williams, M.E. Sigman, Application of likelihood ratios and optimal Science, University of Central Florida, accessed online September 2024, https://
decision thresholds in fire debris analysis based on a partial least squares discriminant ilrc.ucf.edu.
analysis (PLS-DA) model, Forensic Chem. 16 (2019) 100188, https://doi.org/ [20] S.P. Allen, M.R. Williams, C. Bryant, D. Byron, J. Cerven, B.D. Cooper, D.
10.1016/j.forc.2019.100188. C. Hilliard, J. Hoffmann, J. Kwast, S.A. Thomas, C.M. Whitcomb, The National
[6] Lingam, B., Mass Spectral Studies to Investigate Butylbenzene Fragmentation Pathway Center for Forensic Science ignitable liquids reference collection and database, For. Sci.
and Pyrolysis Products. 2015. Electronic Theses and Dissertations. 1278.https://stars. Comm. 8 (2006).
library.ucf.edu/etd/1278. [21] S. Database, National Center for Forensic Science, Univerity of Central Florida,
[7] M. Lopatka, et al., Class-conditional feature modeling for ignitable liquid classification accessed online September 2024,, https://ilrc.ucf.edu/substrate/, 2024.
with substantial substrate contribution in fire debris analysis, For. Sci. Internat. 252 [22] Fire Debris Database. 2024, National Center for Forensic Science, Univerity of
(2015) 177–186, https://doi.org/10.1016/j.forsciint.2015.04.035. Central Florida, accessed online September 2024, https://ilrc.ucf.edu/firedebris/.
[8] M.E. Sigman, M.R. Williams, Assessing evidentiary value in fire debris analysis by [23] R Core Team, A Language and Environment for Statistical ComputingR Foundation for
chemometric and likelihood ratio approaches, For. Sci. Internat. 64 (2016) 113–121, Statistical Computing, Vienna, Austria, 2021, https://www.R-project.org/.
https://doi.org/10.1016/j.forsciint.2016.03.051. [24] ASTM, ASTM E1618-19 Standard Test Method for Ignitable Liquid Residues in Extracts
[9] N.A. Thurn, et al., Classification of ground-truth fire debris samples using artificial from Fire Debris Samples by Gas Chromatography-Mass Spectrometry. 2019, ASTM
neural networks, Forensic Chem. 23 (2021) 100313, https://doi.org/10.1016/j. International: 100 Barr Harbor Drive, PO Box C700, West Conshohocken, PA,
forc.2021.100313. 19428-2959 USA.
[10] S.C. Moldoveanu, Analytical pyrolysis of synthetic organic polymers, Elsevier, 2005. [25] S.C. Tran, M.R. Williams, M.E. Sigman, Hydrocarbon retention on activated carbon:
[11] P. Bhattacharya, et al., Wood/plastic copyrolysis in an auger reactor: Chemical and Preservation of fire debris evidence, Forensic Chem. 39 (2024) 100579, https://doi.
physical analysis of the products, Fuel 88 (2009) 1251–1260, https://doi.org/ org/10.1016/j.forc.2024.100579.
10.1016/j.fuel.2009.01.009. [26] X. Tong, Y. Feng, J.J. Li, Neyman-Pearson classification algorithms and NP receiver
[12] A. Ephraim, D.P. Minh, D. Lebonnois, C. Peregrina, P. Sharrock, A. Nzihou, Co- operating characteristics, Sci. Adv. 4 (2018), https://doi.org/10.1126/sciadv.
pyrolysis of wood and plastics: Influence of plastic type and content on product yield, gas aao1659.
composition and quality, Fuel 231 (2018) 110–117, https://doi.org/10.1016/j. [27] X. Tong, et al., Neyman-Pearson classification: parametrics and sample size
fuel.2018.04.140. requirement, J. Mach. Learn. Res. 21 (2020) 1–48.
[13] N.K. Eklund, B.A. Capistran, V.L. McGuffin, R.W. Smith, Improvements in a kinetic-
based model to predict evaporation of gasoline, Forensic Chem. 17 (2020) 100194,
https://doi.org/10.1016/j.forc.2019.100194.