1. Introduction
Raman spectroscopy is a vibrational spectroscopic technique that is fast emerging as an important analytical method in the field of forensic science. In recent years, Raman spectroscopy has proven advantageous for the investigation of a wide variety of forensic evidence types, including counterfeit pharmaceuticals, cosmetics, controlled substances, explosives, hair, fibers, paints, body fluids, and skeletal remains, among other topics [
1,
2,
3]. In addition to these many applications, Raman spectroscopy has also proven its application for evidence from crimes committed using firearms. According to the (Federal Bureau of Investigation) FBI Uniform Crime Report in 2020, 77% of all homicides in the United States were reported to be firearm-related. Thus, it is natural that firearm-related evidence holds particularly high significance for forensic scientists and researchers [
4]. Despite the importance of firearm-related evidence, there are limitations to current forensic firearm examination practices [
5,
6]. Matching ammunition cartridges to suspected firearms is common practice but often requires the recovery of both a suspected firearm and cartridge from a crime scene [
7]. Meanwhile, standard gunshot residue (GSR) analysis which involves the detection of inorganic particles produced during a firearm discharge event, can only establish the fact of shooting [
5,
8,
9].
In a typical (non-shotgun) ammunition cartridge, there are four major components of note: a projectile, primer, propellant (which is also called powder, charge, powder charge, or smokeless powder), and a cylindrical cartridge case which encompasses everything [
9]. Primer is reactive to percussive force [
9]. When a firearm discharges, the trigger is pulled, which causes the firing pin to strike the primer cup—this force causes the primer to produce a flame and burn [
9]. The reaction of the primer ignites the propellant, which forces the projectile from the cartridge case and through the firearm barrel [
9]. It is through this process that particles known as gunshot residue or GSR are produced [
9]. Gunshot residue may be divided into two distinct subclasses inorganic gunshot residue (IGSR) and organic gunshot residue (OGSR) [
9,
10]. IGSR and OGSR have distinct differences between them. IGSR ranges in size from 1–10 µm; meanwhile, OGSR is macroscopic [
11]. Most notably, the main source of IGSR is the bullet and primer, while the main source of OGSR is the propellant [
9].
The current American Society for Testing and Materials (ASTM) International standard for GSR analysis has focused on IGSR, using scanning electron microscopy with energy dispersive x-ray spectroscopy (SEM-EDS) to identify a specific list of elements termed as “characteristic” of GSR particles [
5,
8]. However, the majority of new research investigations into alternative methods for GSR analysis have focused on OGSR evidence [
12]. This is the result of many factors. First, while the ASTM international standard uses SEM-EDS for IGSR analysis, the standard is limited by its constraints, and as a result, this could lead to problems with the interpretation of GSR evidence [
5]. Secondly, the generation of OGSR at a crime scene is greater than that of IGSR due to the fact that more propellant is present in an ammunition cartridge than primer [
12]. Additionally, the larger size of OGSR makes it more likely to be seen and recovered at a crime scene [
11]. However, the key to why researchers are focusing efforts on the investigation of OGSRs lies in their forensic evidentiary value. OGSR analysis has shown the potential to help investigators differentiate between ammunition brands and calibers [
13,
14]. Since propellant is the major precursor to OGSR, it has become important for investigators to study its characterization and differentiation potential.
Historically, propellant has had two major forms—black powder and smokeless powders [
9]. However, smokeless powders are more commonly used in modern ammunition manufacturing, while black powder is reserved for specialized usages [
9]. Common components in smokeless powders include the stabilizers diphenylamine (DPA) and ethyl centralite (EC), nitrocellulose, and unique combinations of plasticizers, deterrents, lubricants, flash reducers, and other components [
9,
15]. Commercial smokeless powders have been well characterized and documented by the National Center for Forensic Science (NCFS) at the University of Central Florida since 2009 [
16]. Methods used for the creation of this database include mass spectrometry (MS), Fourier-transformation infrared spectroscopy (FT-IR), liquid chromatography (LC), and capillary electrophoresis (CE) [
15]. A few notable studies have focused on the analysis of smokeless powders with a variety of research aims and methodologies employed. In 2021, Lennert and Bridge studied the correlation of smokeless powder, residue, and lab-generated pyrolysis products using gas chromatography-MS (GC-MS) [
17]. This study highlighted that the applicability of direct comparisons between smokeless powders and their corresponding GSRs is composition-dependent; it was found that the sample smokeless powders containing DPA in combination with nitroglycerin and/or 2,4-dinitrotoluene, as well as smokeless powders containing dibutyl phthalate were much more similar to their GSR counterparts than those without these compounds present [
17].
Raman spectroscopy is a non-destructive, sensitive, and specific analytical method in which monochromatic laser light is focused on a sample, and the resulting inelastically scattered radiation is analyzed [
18]. Raman spectroscopy’s success in GSR analysis has already been well documented [
19]. A key first example of this comes from Bueno et al.’s work in which Raman microspectroscopy was combined with advanced statistics to differentiate between GSR particles originating from ammunition of different calibers and manufacturers [
13]. Bueno et al. have also managed to detect IGSR and OGSR particles on a strip of tape and showed the potential of using a unique “vibrational fingerprint” to identify GSR particles using FT-IR [
10,
20]. Meanwhile, Karahacane et al. showed the ability of Raman spectroscopy to distinguish between the OGSR spectra of two different cartridges using DPA or EC [
8]. Despite the ongoing interest in OGSR and Raman spectroscopic analyses of GSRs, no exhaustive larger study has been conducted to differentiate OGSRs based on manufacturer and caliber. Even fewer studies have been concerned with the analysis of smokeless powders via Raman spectroscopy. In 2011 López-López et al. performed a comparative analysis of smokeless powder samples using both Raman spectroscopy and Fourier transform infrared spectroscopy [
21]. In this study, different smokeless powders were analyzed, and their spectra were compared visually and using discriminant analysis techniques. In this case, it was reported that statistical analysis was successful in helping classify the smokeless powders and might help identify unknown samples. In 2012 López-López et al. peripherally addressed this question by comparing the Raman spectra of OGSRs and their respective smokeless powders and noting similarities between the two [
14]. In 2016 López-López et al. utilized surface-enhanced Raman spectroscopy (SERS) to analyze smokeless powders and OGSRs. In the latter study, 21 smokeless powder samples were analyzed, and Raman peaks attributed to major stabilizers were identified [
22].
OGSR analysis using Raman spectroscopy has been shown to have the potential to offer investigators a wealth of information about manufacturer and caliber origins, but delving into a large-scale study of OGSR differentiation via Raman spectroscopy without assessing the best way to model such data would be unwise. Differentiation between OGSRs has been achieved on a small scale with a very small dataset. Indeed, the problem is complex and not easy to solve. However, studying the Raman spectroscopic analysis of smokeless powders from a variety of ammunitions provides a potential framework to determine how to begin building statistical models to help differentiate specific parameters within ammunitions.
Since smokeless powders are the major precursor to OGSR, it is crucial to first determine how well the differentiation between smokeless powders can be achieved. As such, the goal of this investigation was to develop a method to differentiate between the Raman spectra of 12 smokeless powder samples from three different manufacturers (Winchester, Remington, and Federal), with two different calibers represented within each manufacturer (9 mm and .38). All smokeless powder samples were dissolved and spotted onto slides to create cast films and analyzed via Raman spectroscopy. The obtained spectra were analyzed using simple chemometric methods to differentiate between smokeless powders stemming from different manufacturers and calibers. Simple chemometric techniques were utilized to create a modeling system that might be easy and accessible for forensic investigators to implement. A hierarchical modeling framework in which samples were modeled by the manufacturer first and then by caliber within each manufacturer was determined to be the most effective differentiation methodology. This study demonstrates a proof-of-concept with a diverse sample set of smokeless powders and shows that a hierarchical modeling system presents an efficient and effective way to differentiate between smokeless powder manufacturers and specific calibers. The work presented here will be used to guide the statistical analysis of the Raman spectra from the relevant OGSR counterparts of the smokeless powders assessed in this study.
3. Results and Discussion
Selected preprocessed individual Raman spectra from each sample class can be seen in the
Supporting Information in Figure S1; meanwhile, the averaged spectral dataset for all samples is presented in
Figure 2. The Raman spectra show distinct bands in the fingerprint region which are consistent with peaks found in the literature pertaining to OGSRs and smokeless powder compositions. All Raman spectral peak assignments based on literature can be seen in
Table 2.
We initially assessed the obtained spectra visually. Although some differences between the spectra were evident, we needed to employ statistical analysis in order to achieve informative differentiation between the classes. Cross-validation (CV) was used to validate the built models. As long as the hyperparameters are tuned using a sensible procedure (e.g., cross-validation), some machine learning algorithms are likely to provide good results with a low risk of overfitting, even in small data sets. SVM and PLS techniques have advantages over other techniques when analyzing relatively small sample sizes.
We constructed the first model to differentiate the Raman spectra by ammunition manufacturer. The averaged Raman spectra for the manufacturer classes, and their respective standard deviations can be seen in
Figure 3. The manufacturer differentiation modeling was done using an SVM-DA model with 4 LVs using internal cross-validation with Venetian blinds. SVM algorithms are a good option when one is working with smaller datasets that have hundreds or thousands of features. They typically perform better when compared to other algorithms because of their ability to handle small, complex datasets [
30]. The original variables were first introduced to feature extraction using a PLS algorithm. The feature extraction based on PLS compresses the explanatory variables X by creating LVs, that was subsequently used to create an SVM-DA model. A variety of SVM-DA models were tested to distinguish between the spectra from different manufacturers. The final SVM-DA model with a radial basis function kernel was validated by a Venetian blind CV process in which ten test subsets were created by selecting every 1st–10th object in the data set. The assignment of all samples was based on the classification of all sample spectra; the majority percentage determined the assigned class. If no majority percentage was found, then a sample’s assignment was considered as ‘unassigned.’ The results from the manufacturer differentiation, which detail the percentages of spectra assigned by the model, can be seen in
Table 3. All samples were assigned correctly based on the majority of their spectra except samples 3 and 4 from the Winchester manufacturer, which were misclassified as Remington samples. It is notable that sample 3 was only misclassified due to a single-spectrum misclassification, and 40% of the spectra from this sample were correctly classified as Winchester spectra. It is also important to note that samples 3 and 4 are very different from the other samples in the Winchester dataset as they are lead-free and lead bullet-containing, respectively. It is perhaps due to these unique factors that these particular samples were difficult to classify correctly. A summary of the overall results of manufacturer differentiation using this model can be found in
Table 4.
The average and standard deviations for the Raman spectral datasets pertaining to each specific caliber within the manufacturer classes can be seen in
Figure 4. It was found that PLS-DA modeling using differing amounts of LVs was the most effective way to differentiate between the calibers. When considering datasets with fewer observations but a higher number of features like spectral data, usually a linear machine-learning algorithm exhibiting high bias but a low variance is recommended. Performance metrics for assessing a model justify the choice of an algorithm. In the case of caliber differentiation, a PLS algorithm was selected and used to train three separate models for each manufacturer. The PLS-DA models were constructed to differentiate between the different calibers for each manufacturer. Although we tested SVM-DA models, we found in this case that PLS-DA models performed with more accuracy for differentiation. The number of latent variables sufficient to capture variance in the data set used in each model is specified in
Table 5 and
Table 6. The Winchester model used 6 LVs, the Remington model used 10 LVs, and the Federal model used 8 LVs. The need for different LVs for each specific model was expected as the manufacturer classes are different from one another. All three models provided a clear separation between the two calibers. Across all models, a total greater than or equal to 90% of the spectra was correctly classified. The PLS-DA models were validated by a Venetian blind CV process with 10 data splits. The predictions were made for each spectrum from all test subsets. Detailed classification results for the three caliber differentiation models can be seen in
Table 5. The overall percentage of correctly classified spectra can be seen in
Table 6.