Raman Spectroscopy to Enhance Investigative Lead Information in Automotive Clearcoats

Introduction

Modern original equipment manufacturer (OEM) automotive paint is comprised of multiple layers. A typical layer sequence from top to bottom is clearcoat, colorcoat, primer–surfacer, and electrodeposition coat (e-coat). ¹ The colorcoat layer in an OEM automotive paint is thin to reduce the cost of the most expensive component in the paint. In addition, the colorcoat layer in a microscopic fragment of paint recovered at a crime scene is often so thin that it may be impossible to obtain accurate chemical information from it; the size of the fragment may also make it difficult to accurately compare the color with the manufacturer's paint standards. Fortunately, the make, model, and year of the vehicle from an OEM paint sample can be identified by combining chemical information obtained from the two primer layers (i.e., the primer–surfacer and e-coat) with the clearcoat layer as the paint chemistry of the topcoat and the two undercoat layers can often be associated with an individual assembly plant.

Several methods are available for the forensic analysis of an OEM paint sample including scanning electron microscopy ² supplemented with energy dispersive spectroscopy and backscatter electron imaging, pyrolysis gas chromatography, ³ X-ray fluorescence spectroscopy, ⁴ stereomicroscopy, ⁵ and Fourier transform infrared spectroscopy (FT-IR). ⁶ In North America, FT-IR is the most widely used analytical method for the examination and analysis of automotive paint. Infrared (IR) spectroscopy is fast, relatively inexpensive, and requires only a small quantity of sample. ⁷ In previously published studies, Lavine and coworkers have demonstrated that discriminants developed from the IR spectra of automotive clearcoats can identify the make and model of the vehicle.^8–12 Even in challenging trials where the samples evaluated were all from the same manufacturer within a limited production year range (2000–2006), discriminants were able to identify the specific assembly plant that manufactured the vehicle (and hence the make and model of the vehicle) from an OEM paint sample. For cases where there are no initial suspects, developing a hypothesis as to the origin of the evidence recovered from the crime scene (e.g., the make and model of the vehicle from which the paint sample originated) can lead to the identification of a suspect.

In this study, Raman spectroscopy has been investigated as an alternative to FT-IR for obtaining investigative lead information (e.g., make and model of the vehicle) from automotive clearcoats. Raman spectroscopy has several advantages over IR spectroscopy for automotive paint analysis. Raman bands are generally separated, whereas IR bands often overlap in the spectra. IR bands that are too weak to be observed may be sufficiently intense to be observed in the Raman spectra. However, only a few studies^13–17 have appeared in the literature on the application of Raman spectroscopy for the characterization of automotive clearcoats. The results reported by the authors in these studies were overly optimistic as most of the automotive clearcoat classes surveyed by the authors contained only a few samples. Furthermore, the clearcoat samples investigated in each study were manufactured in several countries including Japan, the United States of America, and Western Europe. Therefore, the results obtained by the authors were more likely indicative of differences in the supply chain for the automotive paint samples surveyed. By comparison, this study is restricted to 118 automotive clearcoats from General Motors (GM) vehicles manufactured in assembly plants in North America between 2000 and 2006. The forensic paint examination problem selected for this study was intentionally challenging as the samples were limited to a single manufacturer over a narrow production year range. Furthermore, there is a large sample test bed available to validate the results. Although several studies have appeared in the literature over the past five years on the application of Raman spectroscopy to automotive paint analysis and have been the subject of several reviews,^18,19 these investigations have largely focused on multilayered automotive paint chips. By comparison, this study is restricted to a single layer of automotive paint, the clearcoat layer. Because clearcoat formulations are often very similar, discriminating the different makes and models of automotive vehicles using the clearcoat layer alone can be challenging. Using Raman spectroscopy, we demonstrate that investigative lead information can be extracted from the clearcoat layer of OEM paint samples.

A new method to determine the make and model of a vehicle from an automotive paint sample recovered at the crime scene of a vehicle-related fatality such as a hit-and-run using Raman microscopy has been developed. This study entails the discrimination of Raman spectra of OEM automotive clearcoats from six GM vehicle assembly plants, that is, Moraine (Ohio), Lansing (Michigan), Fort Wayne (Indiana), Fairfax (Kansas City, Kansas), Doraville (Georgia), and Arlington (Texas). Raman spectra were collected from 118 automotive paint samples from six GM vehicle assembly plants to investigate the discrimination power of Raman spectroscopy for automotive clearcoats using a genetic algorithm (GA) for pattern recognition that incorporates model inference and sample error in the variable selection process. Each vehicle assembly plant pertained to a specific vehicle model. The spectral region between 1802 cm^–1 and 697 cm^–1 was found to be supportive of the discrimination of these six GM assembly plants. By comparison, only one of the six automotive assembly plants could be differentiated from the other five assembly plants using FT-IR (which is the most widely used analytical method for the examination of automotive paint) and the GA for pattern recognition. The results of this study indicate that Raman spectroscopy in combination with pattern recognition methods offers distinct advantages over FT-IR for the identification and discrimination of automotive clearcoats. Parts of this study were presented at the 2021 Current Trends in Forensic Trace Analysis Online Forensic Symposium ²⁰ and in the final report for the National Institute of Justice entitled, “Application of Raman and Infrared Microscopy for the Forensic Examination of Automotive Clear Coats and Paint Smears.” ²¹

Materials and Methods

One hundred and eighteen GM automotive paint samples (Table I) from six vehicle assembly plants (Tables S1–S6, Supplemental Material) were selected to investigate the discrimination power of Raman spectroscopy and FT-IR for OEM automotive clearcoats. The paint samples used in this study were provided by the Royal Canadian Mounted Police (RCMP) from their paint data query ²² paint sample collection. Prior to Raman analysis, the surface of each multilayered automotive paint chip was cleaned with American Chemical Society grade methanol (Fisher Scientific). After the particulate matter was removed from the surface of the paint chip, the clearcoat layer was gently scrapped off the sample using a sharp knife under a stereo microscope, and the scrapings were transferred onto a microscope glass slide (Globe Scientific Inc.) covered with aluminum foil which offered a low and almost featureless background for Raman analysis.

OPEN IN VIEWER

Table I.

Assembly plants used in the Raman and FT-IR studies.

Assembly plant	Plant ID number	Number of samples
Arlington	1	33
Doraville	4	5
Fairfax	5	20
Fort Wayne	8	16
Lansing	14	22
Moraine	18	22

Raman spectra were acquired using a WITec alpha300 R confocal micro-Raman system equipped with a charge-coupled detector (CCD) operated at −60 °C and a 532 nm neodymium-doped yttrium aluminum garnet laser. A (111) silicon wafer was used as the calibration standard. The spectral calibration was done with reference to the first-order Raman scattering peak at 520.5 cm^–1.

Each Raman spectrum was acquired using an integration time of 200 s (sum of 20 spectra, each collected for 10 s). For some samples, peak shifts (of >2 cm^–1) were observed due to photothermal heating. To address this problem, the laser power was lowered from 40 to 10 mW and the integration time was increased to 900 s (sum of 30 spectra, each collected for 30 s). For some clearcoat samples, cosmic spikes were generated by the CCD. To eliminate a spike, typically corresponding to a single CCD pixel (wavenumber data point), the spike intensity was replaced by the average of intensities corresponding to adjacent CCD pixels (wavenumber data points). For other paint samples, excessive autofluorescence was observed. This fluorescence was probably due to pigments transferred from the colorcoat layer when the clearcoat layer (which was thin for these samples) was scrapped off the paint chip in the process of isolating the clearcoat layer from the other layers. To address this problem, photobleaching was performed. The sampled spot was exposed to a 40 mW laser beam for about 5 min prior to the Raman acquisition (Fig. 1).

OPEN IN VIEWER

Figure 1.

A representative Raman spectrum of automotive paint clear coat: (a) before photobleaching and (b) after photobleaching for 5 min.

In micro-Raman spectroscopy, the highest signal-to-noise ratio is usually acquired with a high numerical aperture lens (e.g., 0.90) and diffraction-limited laser spot size which is less than a micron when visible laser excitation is employed. However, the small laser spot size may also be disadvantageous as it maximizes the laser intensity and may lead to photodegradation and photothermal artifacts. Furthermore, the small laser spot may not allow for sufficient statistical averaging for a sample not sufficiently homogenous. For the automotive clearcoats, sample heterogeneity was a problem when collecting Raman spectra of these polymers. To improve statistical averaging, a 20× objective lens (Nikon) with a numerical aperture of 0.40 and a laser spot size of either 5 or 10 µm (depending on the sample) were employed. Additionally, we used the lowest available confocality with a 100 µm diameter fiber (serving as the pinhole) that ensures the signal is collected from the largest area of ∼10 µm diameter circle. As a result, the signal is integrated from a larger area with a higher level of statistical averaging. An increase in the laser spot size beyond 10 µm resulted in a significant reduction of the signal counts. The objective lens of the Raman microscope was focused just below the air/polymer interface of each clearcoat, not on the surface exposed to the environment. Finally, for each clearcoat sample, Raman spectra were acquired at five different sites and averaged to obtain a representative averaged spectrum with a higher signal-to-noise ratio. With these procedures, we acquired highly reproducible spectra for the OEM clearcoat samples.

The Raman acquisitions were performed with a high spectral resolution (1.08 cm^–1) 1800 lines/mm diffraction grating. Although this grating is limited to a wavenumber range of only 1105 cm^–1 in a single scan, it covers most of the fingerprint region which is sufficient for identifying the make and model of the automotive vehicle from the paint sample. Although the same Raman microscope also has a 600 lines/mm grating, enabling a larger wavenumber range (3867 cm^–1), its lower resolution (3.78 cm^–1) was a limitation for multivariate analysis. For this reason, the 1800 lines/mm grating was employed in this study.

Clearcoat IR spectra for this study were previously collected by the RCMP. For each clearcoat, 4 μg of the layer were placed in a high-pressure diamond cell and measured in transmission mode at 4 cm^–1 resolution using either a Thermo Nicolet 6700 FT-IR spectrometer or a Bio-Rad 40 A/60A spectrometer, both equipped with a deuterated triglycine sulfate detector and a Harrick 6x beam condenser (Thermo Nicolet 6700 FT-IR spectrometer) or a Harrick 4× beam condenser (Bio-Rad 40A and Bio-Rad 60A instrument). Each IR spectrum was normalized to the helium-neon laser frequency of 15 798.0 cm^–1 using OMNIC (Thermo-Nicolet).

Pattern Recognition Analysis

Spectral Preprocessing

All Raman spectra were preprocessed using routines in the PLS toolbox (Eigenvector Technology) to eliminate problems of noise and background (e.g., fluorescence) that have plagued previous workers. ²³ First, each spectrum was baseline corrected using a Whitaker filter (λ = 100 000). Next, the spectra were smoothed using a Savitzky–Golay filter (five-point window and polynomial order 1). The baseline corrected and smoothed Raman spectra were then normalized to unit length. A representative Raman spectrum before and after preprocessing is shown in Fig. 2.

OPEN IN VIEWER

Figure 2.

A representative Raman spectrum of automotive paint clear coat: (a) before baseline correction, smoothing, and normalization to unit length and (b) after baseline correction, smoothing, and normalization to unit length.

Outlier Analysis

For each assembly plant (i.e., class), outlier analysis was performed on both the Raman and IR spectra as the presence of sample outliers can adversely influence the performance of principal component (PC) analysis and other eigenvector-based methods. ²⁴ The generalized distance test ²⁵ was performed on each class at the 0.01 level using SCOUT. ²⁶ Six samples were flagged as outliers. Three samples were from the Fairfax assembly plant (Table S3, Supplemental Material: CONT01472, CONT01475, and UNVL00007) and the other three (Table S5, Supplemental Material: CONT00985, CONT01049, and UVAC00134) were from Lansing. A visual examination of the spectra of these six samples revealed small but noticeable differences compared to the other samples in their respective class.

Wavelet Preprocessing

Both the IR and Raman spectra were further preprocessed using wavelets. ²⁷ In this study, the Symlet mother wavelet (sixth smallest filter size, eighth level of decomposition) was chosen for preprocessing. The discrete wavelet transform was applied to each IR and Raman spectrum. This transform involves two filters: a low-pass filter that produces “approximation” coefficients and a high-pass filter that produces “detail” coefficients that are equal in number to the approximation coefficients (e.g., if 177 approximation coefficients are generated by the low-pass filter, 177 detail coefficients will be generated by the high pass filter). The original spectra serve as input for the first level of decomposition. The approximation coefficients generated at each level of decomposition become the input for the next level of decomposition (e.g., the second level of decomposition uses the first-order approximations) with this process continued until the eighth level of decomposition is reached.

Genetic Algorithm for Variable Selection and Classification

An approach based on identifying the smallest set of wavelet coefficients that optimize the separation of the classes in a plot of the two or three largest PCs of the data using a GA was employed. Because PCs maximize variance, the bulk of the information encoded by these coefficients will be about differences between the classes in the data set. Although a PC plot is not a shark knife for discrimination, if we have a PC plot that shows clustering then our experience is that we will be able to predict robustly using this coefficient set.

For this study, the model inference was incorporated into the PCKaNN fitness function^28–33 to identify variables that minimize the error across the entire model, which is the PC score plot of the wavelet coefficients selected by the pattern recognition GA. This was accomplished by assessing the uncertainty of the scores for each sample in the PC plot using the jackknife ³⁴ to generate estimates of dispersion. During each generation, the fitness function of the pattern recognition GA evaluated thousands of PC plots, one for each feature subset (i.e., chromosome) from the population of potential solutions. For each PC plot, the corresponding training set samples were removed one at a time, and the score and loading matrices for the resampled (i.e., jackknifed) training set were recomputed. Due to the rotational ambiguities of PC analysis, the loading matrix for each resampled training set was rotated using a Procrustean rotation ³⁵ to match the loading matrix associated with the score plot containing all the samples. For each training set sample, the scores for each sample across all leave-one-out score plots are projected onto the original PC plot of the wavelet coefficient subset which is then evaluated using PCKaNN. Thus, information about the level of confidence in the classification of each sample in the training set is directly incorporated into the variable selection process, with the jackknifed scores for each sample effectively constituting an error cloud to depict the uncertainty for each training set sample.

Discriminant Analysis

The sample cohort (118 IR and Raman spectra) was divided into 16 training set/validation set pairs (Table S7, Supplemental Material) with each sample present in only one of the validation sets. For each training set/validation set pair, the data were autoscaled. Classifiers were developed for the training sets and then tested on the validation sets. The cross-validation procedure used in this study ²⁸ differed from the well-known procedure of Wood et al. ³⁶ In this study, the pattern recognition GA selected the wavelet coefficients for each training set, whereas, in traditional cross-validation, the wavelet coefficients for each training set are the same and are determined using the entire data set prior to dividing the data into training set/validation set pairs. For this reason, error rates reported in this study were less biased than traditional cross-validation for computing classification success rates of discriminants.

Results and Discussion

Due to the similarity of the Raman spectra, it was necessary to use pattern recognition methods to discriminate the spectra by the assembly plant. We wanted to assess whether the spectral differences between assembly plants were larger than the differences between spectra within an assembly plant. For pattern recognition analysis, the spectral range 1802 cm^–1 to 697 cm^–1 was used for both the Raman and IR spectra, thereby restricting the analysis to the binder present in the clearcoat layer while ignoring the fillers, extenders, and other inorganic constituents. Previous studies have shown that restricting the analysis to the spectral region corresponding to only the binder in an OEM clearcoat is sufficient to determine the make and model of the vehicle.^9–12,37 Three specific questions were addressed in this study: (i) Can discriminants be developed from the Raman and IR spectra that differentiate the six assembly plants? (ii) Can these discriminants predict the assembly plant of an unknown paint sample? (iii) Is there any advantage with Raman in lieu of FT-IR for the analysis of automotive clearcoats?

Classification of the Raman and IR spectra was performed in two steps. Because of the similarity of the clearcoat spectra, the approach taken for classification focused on the variable selection using a GA that identified the informative wavelet coefficients by sampling key feature subsets in the population of potential solutions, scoring their PC plots, and tracking those assembly plants and/or spectra that were difficult to classify. The boosting routine of the pattern recognition GA used this information to steer the population to an optimal solution. After only 200 generations, the pattern recognition GA was able to identify a set of wavelet coefficients that contain information about the classification problem of interest, which is the identity of the vehicle assembly plant from which each paint sample originated.

In the first step, the Arlington assembly plant was discriminated from the other five assembly plants (Moraine, Lansing, Fort Wayne, Fairfax, and Doraville) using the pattern recognition GA to identify the wavelet coefficients characteristic of the Arlington assembly plant. For each training set/validation set pair, the validation set spectra were projected onto the PC plot of the spectra comprising the training set and the wavelet coefficients identified by the pattern recognition GA. For the validation set samples to be correctly classified, the spectra representing the Arlington clearcoats must lie in a region of the PC plot with the other Arlington paint samples, and the spectra representing the Doraville, Fairfax, Fort Wayne, Lansing, and Moraine clearcoats must lie in a region of the PC plot containing the samples from these five assembly plants.

In the second step, the spectra from the other five assembly plants (Moraine, Lansing, Fort Wayne, Fairfax, and Doraville) were discriminated using the pattern recognition GA to solve a five-way classification problem. The samples projected onto the region of the PC plot encompassing the Moraine, Lansing, Fort Wayne, Fairfax, and Doraville assembly plants in the first step were passed to the discriminant in the second step. Validation set samples from Moraine, Lansing, Fort Wayne, Fairfax, and Doraville classified as Arlington in the first step were not passed to the discriminant (i.e., PC score plot) developed for the five-way classification (Doraville, Fairfax, Fort Wayne, Lansing, and Moraine assembly plants) in the second step. If a Doraville validation set sample from the first step was passed to the discriminant in the second step and was correctly classified, the sample would lie in a region of the PC plot containing the other Doraville samples.

Results for the first Raman and IR training set/validation set pair (excluding the sample outliers) are summarized in Figs. 3–6. For the first Raman training set/validation set pair, the two Arlington samples comprising the validation set were correctly classified in the first step (Fig. 3). As for the second step (Fig. 4), all four samples from Moraine, Lansing, Fort Wayne, and Fairfax were also correctly classified. By comparison, only one of the two Arlington samples was correctly classified in the first step (Fig. 5) for the first IR training set/prediction set pair, and only one of the four samples from Fairfax, Fort Wayne, Lansing, and Moraine was correctly classified in the second step (Fig. 6).

OPEN IN VIEWER

Figure 3.

PC plot of the first training set/prediction set pair: Arlington versus Doraville, Fairfax, Fort Wayne, Lansing, and Moraine. Prediction set samples (red) projected onto the PC plot developed from the training set samples and the wavelet coefficients identified by the pattern recognition GA. 1 = Arlington, 4 = Doraville, 5 = Fairfax, 8 = Fort Wayne, 14 = Lansing, and 18 = Moraine.

OPEN IN VIEWER

Figure 4.

PC plot of the first training set/prediction set pair: five-way classification study (Doraville, Fairfax, Fort Wayne, Lansing, and Moraine) for the prediction set. Prediction set samples (red) projected onto the PC plot developed from the training set samples and the wavelet coefficients identified by the pattern recognition GA. 1 = Arlington, 4 = Doraville, 5 = Fairfax, 8 = Ft. Wayne, 14 = Lansing, and 18 = Moraine.

OPEN IN VIEWER

Figure 5.

PC plot of the first training set/prediction set pair: Arlington versus Doraville, Fairfax, Fort Wayne, Lansing, and Moraine. Prediction set samples (red) projected onto the PC plot developed from the training set samples and the wavelet coefficients identified by the pattern recognition GA. 1 = Arlington, 4 = Doraville, 5 = Fairfax, 8 = Fort Wayne, 14 = Lansing, and 18 = Moraine.

OPEN IN VIEWER

Figure 6.

PC plot of the first training set/prediction set pair: five-way classification study (Doraville, Fairfax, Fort Wayne, Lansing, and Moraine) for the prediction set. Prediction set samples (red) projected onto the PC plot developed from the training set samples and the wavelet coefficients identified by the pattern recognition GA. 1 = Arlington, 4 = Doraville, 5 = Fairfax, 8 = Ft. Wayne, 14 = Lansing, and 18 = Moraine.

The combined Raman and IR results for the 16-training set/prediction set pairs (which do not include the six samples flagged as outliers) are summarized in Table II. One hundred and five of the 112 Raman spectra or 93.75% of the validation set were correctly classified. By comparison, only 75 of the 112 IR spectra or 66.96% of the validation set were correctly classified. These results demonstrate that information derived solely from the spectra of clearcoats can categorize paint samples as to the assembly plant of the vehicle from which the automotive paint sample originated. Furthermore, Raman spectroscopy appears to be a better solution than IR spectroscopy for obtaining investigative lead information from automotive clearcoats.

OPEN IN VIEWER

Table II.

Summary of results for the hierarchical classification study (outliers excluded).

	Raman (outliers removed)		IR (outliers removed)
Arlington versus others	Correct	Wrong	Correct	Wrong
Total (first step)	27	6	18	15
Linear separability (Doraville, Fairfax, Fort Wayne, Lansing, and Moraine)	Correct	Wrong	Correct	Wrong
Total (second step)	78	1	57	22
Total (first step + second step)	105	7	75	37

Table III summarizes the combined Raman and IR results with the six sample outliers included in the 16 training set/prediction set pairs. For the Raman spectra, 101 of the 118 samples (or 85.6%) were correctly classified, whereas, for the IR spectra, only 69 of the 118 samples (or 58.5%) were correctly classified.

OPEN IN VIEWER

Table III.

Summary of results for the hierarchical classification study (outliers included).

	Raman (outliers included)		IR (outliers included)
Arlington versus others	Correct	Wrong	Correct	Wrong
Total (first step)	25	8	18	15
Linear separability (Doraville, Fairfax, Fort Wayne, Lansing, and Moraine)	Correct	Wrong	Correct	Wrong
Total (second step)	76	3	51	28
Total (first step + second step)	101	11	69	43

The chemical formulation of the automotive clearcoats used in this study is acrylic melamine styrene. For melamine and styrene, the Raman and IR spectra have distinct and recognizable peaks for each of these monomers. The situation regarding the acrylate monomer is more complex as there are several acrylate monomers present in a typical clearcoat formulation. ³⁸ The informative bands for acrylates in the IR spectra of clearcoats are the carbonyl stretch and the C─O─C antisymmetric stretch. The latter band may be characteristic of specific acrylate monomers and although allowed in IR, is too weak to be observed unequivocally in the IR spectra of clearcoats due to bands from other components. Raman spectra, on the other hand, also contain bands characteristic of acrylates including the carbonyl stretching mode and the C─O─C symmetric stretch, which is also characteristic of specific acrylate monomers. The C─O─C symmetric stretch is less overlapped by neighboring bands in the Raman spectra of clearcoats, is more intense, and is more accessible after wavelet preprocessing. ³⁹

Conclusion

To further enhance the general discrimination power of automotive clearcoats, Raman spectroscopy has been investigated as an alternative to IR spectroscopy for extracting investigative lead information from automotive clearcoats. Given the challenging nature of the spectral discrimination problem investigated in this study, these results constitute direct evidence of the potential advantages offered by Raman for forensic automotive paint analysis. The Raman clearcoat study described here is timely and addresses several important questions crucial for assessing the potential of Raman spectroscopy for the forensic analysis of automotive paints. Since 2014, the Bundeskriminalamt (BKA) has been collecting automotive paint samples to analyze by Raman spectroscopy. At present, the BKA continues its systematical collection of Raman spectra of automotive clearcoats. The BKA has yet to perform an evaluation of the forensic value of their Raman automotive paint database.

Supplemental Material