Applying Raman and Infrared Spectroscopy in Forensic Paint Analysis

For "The Future of Forensic Analysis” series, we interviewed Barry Lavine, regents professor from The Department of Chemistry at Oklahoma State University in Stillwater, Oklahoma, to describe his most recent work in applying Raman and infrared (IR) spectroscopy in forensic paint analysis.

Barry Lavine of Oklahoma State University.

Would you explain your background and how you became interested in researching the application of spectroscopy in forensic paint analysis?

I am an analytical chemist by training with a specialization in chemometrics (that is, chemical data science), vibrational spectroscopy, and statistics. Analytical chemists are often in search of new opportunities to expand the reach of chemical analysis and develop new methods for solving problems of societal interest. My current interest in the application of spectroscopy in forensic paint analysis can be attributed to Scott Ryland (who is retired from the Orlando Crime Lab) who had contacted me about working with the Royal Canadian Mounted Police (RCMP) to evaluate the probative value of automotive clear coats for vehicle identification and assess their evidential significance assessment, both at the investigative lead stage and at the courtroom testimony stage. The publication in the journal Talanta on this subject in 2011 was the first study to demonstrate the importance of automotive clear coats in vehicle identification. Although I was the author of correspondence, Mark Sandercock (who is retired from the RCMP) is the individual who first recognized the importance of automotive clear coats in forensic automotive paint analysis and Scott Ryland is the individual who worked tirelessly to put together this collaboration. Both Mark and Scott should receive the lion’s share of the credit for this work. Both Mark Sandercock and Scott Ryland are groundbreaking scientists in forensic chemistry who have made important contributions to the field.

What are the biggest challenges in your work to match a collected accident or crime scene paint specimen with the available law enforcement database?

If a paint sample is not contained in an automotive paint database, a similarity search will be crucial for making a tentative identification. Currently, there are no algorithms commercially available to perform a similarity search. For this reason, much of the research performed in library searching of automotive paint databases by my research group has initially been focused on forward and reverse cross correlation library search algorithms. Another challenge in matching a paint sample to available law enforcement databases is the size of the hit-list, which can be large for some paint layers due to the limited number of suppliers of the components comprising the topcoat and undercoat layers of an original equipment manufacturer automotive paint. For this reason, search prefilters have been developed by my group using pattern recognition techniques to cull the spectra in the library to a specific assembly plant or plants.

Car painter applying green paint to car part with spray gun in workshop for car body repair Intricate vehicle body restoration. Generated with AI. | Image Credit: © Emin - stock.adobe.com.

What are the main challenges of using the laboratory paint matching results in a criminal court case?

Although I have never presented laboratory paint matching results to a jury in a criminal trial, it is my opinion that presenting these results in a way that a jury could understand would be a major challenge. As most jurors may not have taken a chemistry course (even those with a college degree), this task would require a major time investment as the basic principles underlying the analysis will need to be explained to the jury. Having taught a terminal general chemistry class to undergraduates who had not taken high school chemistry, I can appreciate the challenges involved, but a jury can be made to understand paint matching and why the identifications provided by our methods are reliable.

Can you elaborate on the specific advantages that Raman spectroscopy offers over Fourier transform infrared (FT-IR) in identifying and discriminating automotive clearcoats in forensic applications?

Search prefilters developed from Raman spectra of automotive clear coats for discrimination by assembly plant performed better than search prefilters developed from infrared spectra of automotive clear coats. We attribute this to the Raman bands being generally well separated, whereas IR bands overlap in the spectra. In addition, IR bands that contain discriminating information are often too weak to be observed whereas these bands may be sufficiently intense to be observed in the Raman spectra. When considering that clear coat formulations are often very similar, being able to discriminate among the different makes and models of automotive vehicles using the clearcoat layer is clearly a challenging problem. Nevertheless, improvements in the discrimination of vehicle assembly plants using the Raman spectra of automotive clearcoats is a truly impressive result.

How does the genetic algorithm for pattern recognition improve the discrimination power of Raman spectroscopy, and what role does model inference and sample error play in this process?

By selecting the truly informative features in the Raman spectra for assembly plant discrimination using a genetic algorithm in the development of the search prefilter, we can improve the performance of the search prefilters for assembly plant. When sample error is incorporated into the analysis, the performance of the pattern recognition genetic algorithm (GA) can be improved to identify the truly informative wavelengths in the Raman spectra.

Given that your study focused on samples from six General Motors (GM) vehicle assembly plants, how broadly do you think your findings can be generalized to other manufacturers and vehicle models? Are there any steps being taken to expand this research?

The six GM assembly plants that were selected for this study possessed similar IR spectra for the clear coat layer. Given the challenging nature of this spectral discrimination problem, the significance of these results is that they constitute direct evidence of the potential advantages offered by Raman for forensic automotive paint analysis. These results are contrary to what has previously been reported about the use of Raman (which should be limited to the color coat layer and the pigments in other layers) in forensic automotive paint analysis.

How does the incorporation of an ultramicrotome enhance the capability of FT-IR imaging microscopy in the analysis of small paint chips, particularly those that are 1 mm or less in size?

Paint fragments that are too small (1 mm or less in size) to be hand sectioned under a stereo microscope or too small to be cross sectioned using a conventional microtome can be analyzed by a transmission infrared imaging microscope equipped with an ultramicrotome to cross section minute paint chips. Thus, the different layers of a very small paint chip can be exposed to and then analyzed by transmission infrared microscopy. Small paint chips are typically found on the clothing of a pedestrian injured or killed in a hit-and-run.

Can you explain the challenges faced when extracting IR spectra from thin layers (less than one micron) of automotive paint chips using alternating least squares, and how modified alternating least squares address these challenges?

The obvious challenge when extracting IR spectra from thin paint peels using alternating least squares is obtaining a representative spectrum of each layer from the line map of the infrared image. Unfortunately, problems may occur when resolving basis vectors characteristic of each paint layer comprising the sample when the data has low signal to noise and is nearly collinear. For this type of data, alternating least squares often fails. However, modified alternating least squares (in most cases) can produce a satisfactory solution because of the use of ridge regression that allows the least squares solution to be robust towards collinear data and has a term to compensate for the bias that occurs when using ridge regression. Modified alternating least squares is also superior to alternating least squares for determining the spectra of minor components. Clearly, modified alternating least squares is a better solution to the constrained non-negative least squares optimization problem than alternating least squares. Nevertheless, the performance of modified alternating least squares for spectroscopic imaging problems, like alternating least squares, is also influenced by the spatial resolution between sample components since low spatial resolution produces a greater mixing of components within the image.

In what ways does the new sample preparation technique combined with multivariate curve resolution methods improve the quality of IR spectra obtained from small automotive paint fragments, and how does this compare to conventional Fourier transform IR spectroscopy?

Obtaining reliable IR spectra of the different layers of automotive paint from a very small chip can be problematic. This new sample preparation method, which is geared to very small samples (because it was developed for tissue and biopsy specimens) offers a potential solution to this problem when coupled with transmission infrared imaging and modified alternating least squares. The thin peels generated by the ultramicrotome are ideal for transmission as the throughput is higher than a peel created by a conventional microtome. However, fringing can be a problem with thin peels which can be remedied through the appropriate data preprocessing of the spectra in the sample image line map.

What are the main challenges forensic automotive paint examiners face when analyzing paint smears found at crime scenes, and how does the lack of reference materials in databases exacerbate these issues?

Paint smears have proven challenging to analyze by transmission infrared spectroscopy because of the mixing of the various automotive paint layers. Furthermore, transferred paint may undergo chemical changes during smear formation, which must be considered in the analysis of its IR spectrum. Potential mechanisms responsible for these changes include oxidation or other thermally activated chemical reactions due to local temperature rise by friction. For example, CH₂ sites can oxidize to C═O moieties. This modification may be monitored from the attenuation of C─H stretching peaks with a simultaneous increase of the C═O stretching peaks in the range of 2850–2950 and 1650–1720 cm⁻¹ respectively. There is also mechanochemistry in the presence of high velocity gradients and shear, where tension develops along the backbone of the polymer chains leading to scission. The resultant radical pair can recombine, disproportionate, react with oxygen, or attack another polymer chain and lead to crosslinking. The cleaved aliphatic chain ends may also transform to ─HC═CH₂. Hence, increased C═C stretching may be discernible in the vibrational spectrum.

As for the lack of reference materials in an automotive paint database (for example, samples from paint smears), this can impair the accuracy of paint smear analysis. Furthermore, paint smear reference materials are crucial for the creation of proficiency tests for forensic laboratories to evaluate their methodology and can be used in the training of forensic chemists.

Could you elaborate on how the impact tester procedure simulates paint transfer between vehicles during collisions, and how varying the conditions of the test affects the number of distinct paint layers obtained in smears?

The generation of a paint smear as paint transfer evidence in a vehicle–vehicle collision can be simulated using two wedge shaped steel blocks. The automotive paint sample is mounted on the upper wedge, whereas the paint-transfer substrate is mounted on the lower wedge. The upper wedge is released from a specified height and accelerated by gravity towards the lower wedge. The upper wedge slides vertically on two guiding rods, but it transfers both vertical and horizontal momentum to the lower wedge upon collision, because the collision surfaces (that is, wedge surfaces) are at 450 to the horizontal/vertical. Because of this design, the contact force, which is normal to these surfaces, has both vertical and horizontal components. Upon collision, energy and momentum is transferred to the lower wedge, as it is allowed to move vertically and horizontally on the rail guide. As a result, the lower wedge surface also slides against the upper wedge surface, creating friction. The frictional work is responsible for the transfer of the paint to the substrate as a smear.

Paint smear samples were generated with varying levels of the elastic force constant and the damping coefficient. These collision parameters were controlled using replaceable springs of two different stiffnesses and by changing the damping level of the shock absorbers. As expected, higher stiffness springs caused more severe collisions, with higher collisional forces and greater frictional work. By exploring the damping levels and the spring stiffness, various automotive paint smears of varying complexities could be successfully simulated.

How does attenuated total reflection infrared microscopy, combined with alternating least squares, assist in isolating and identifying individual paint layers from smears, and how reliable is this method in matching paint smears to their corresponding original equipment manufacturer (OEM) paint samples?

Attenuated total reflection, unlike transmission infrared spectroscopy possesses the necessary spatial resolution to decouple the individual paint layers comprising the paint smear. When the spectral image is coupled to alternating least squares, it is possible to recover the individual layers that comprise the smear. In some cases, four layers can be recovered by alternate least squares (ALS), whereas in other cases it may be only one, two, or three layers of automotive paint. The number of layers that can be recovered depends upon the conditions used to generate the smear (spring stiffness and damping level of the shock absorbers compromising the assembly used to simulate smear generation).

Could you explain how alternate least squares (ALS) reconstructions of IR spectra enhance the accuracy and speed of forensic automotive paint examinations? What role does ML play in this process?

ALS is necessary to recover spectra of the individual automotive paint layers from a line map of the infrared image of the paint sample. To recap, in infrared imaging, the time necessary for data collection is minimized by collecting concatenated IR data from all paint layers in a single analysis by scanning across the cross-sectioned layers of a paint sample using an IR imaging microscope. Once the data has been collected, it then undergoes decatenation using alternating least squares to obtain a “pure” IR spectrum of each paint layer. This approach not only eliminates the need to analyze each layer separately, but also ensures that the recovered IR spectrum of each layer is “pure” and not a mixture of spectra which occurs at the interface between the layers. Minimizing the probability of collecting a mixed IR spectrum can result in a considerable time savings as well as objectively ensuring that only a “pure” IR spectrum from each layer has been collected and is used in subsequent comparisons or database searches. By comparison, hand sectioning each paint layer and collecting transmission spectra of each separated layer by placing it between two diamond anvils followed by performing a search of the Paint Data Query (PDQ) database against each separated paint layer using a text-based search system can be time consuming.

Using machine learning, search prefilter have been developed from the IR spectral library of the PDQ automotive paint database to identify the vehicle manufacturer and the assembly plant of the vehicle from which the paint sample originated. By applying the search prefilter directly to the ALS recovered spectra of the clear coat and the two undercoat layers, the PDQ library can be culled to a subset of spectra. Matching against this subset is performed using a library search algorithm based on cross correlation.

What challenges arise due to peak shifts in IR spectra when comparing ALS reconstructed spectra and in-house library spectra? How does the correction algorithm developed by your laboratory address these issues?

An examination of the in-house library spectra and the reconstructed IR spectra of the same paint sample often reveals large peak shifts (approximately 10 cm^-1) for some vibrational modes. For library search algorithms utilizing correlation to assess spectral similarity, these shifts diminish the hit quality index value of the match. These frequency shifts may pose an even greater problem for machine learning techniques used in classification, as the metric computed to optimize discrimination minimizes the within source variability (which is linked to similarity) and maximizes the between source variability (i.e., the differences in the spectral profile of the various assembly plants). Because of the even more stringent criteria for classification than spectral matching, higher quality data containing fewer artifacts is required.

To solve the problem of the frequency shifts, an algorithm previously developed in our laboratory which converts transmission spectra to ATR spectra was adapted to standardize the IR transmission spectra from the in-house spectral libraries collected using a high-pressure diamond transmission cell and from the transmission IR microscope collected at ambient pressure using a BaF₂ cell. The ATR simulation algorithm used to standardize the residual polarization of the IR beam for both the in-house library spectra and for the ALS and multi-angle light scattering (MALS) reconstructed IR spectra is based on solving sequentially a set of six equations to standardize the real and imaginary components of the refractive index and to compute the coefficients of the s- and p- polarized light at each wavenumber. To develop this algorithm, thirteen transmission spectra of the clear coat layer were selected from the PDQ database. attenuated total reflection/reflectance (ATR) spectra of the same samples (clear coat layer) were measured using the ATR accessory of the Nicolet iS-50 FT-IR spectrometer. A self-consistent set of parameters for the correction algorithm (for example., thickness and refractive index of the polymer film and the internal reflection angle of the diamond crystal) referenced to the iS-50 FT-IR spectrometer were determined by continually adjusting for the in-house library spectra and comparing the thirteen simulated ATR spectra with the ATR spectra measured on the iS-50 in order to obtain the best match in ordinate intensities and peak positions between the two sets of spectra. Thus, the polarization of the IR beam is adjusted to that of the iS-50 spectrometer by developing a set of self-consistent parameters (thickness and refractive index of the sample, and internal reflection angle of the diamond crystal) for the ATR simulation which is then applied to all ALS and MALS reconstructed spectra and to all the spectra in the in-house library.

In your study, how effective were the ML methods in classifying the make, model, and assembly plant of vehicles based on the IR spectra of automotive paint samples? Could you elaborate on the validation process using the twenty-six OEM paint samples?

The search prefilters developed from residual polarization corrected IR spectra using machine learning methods were effective as they correctly predicted the make, model and assembly plant of the vehicle from which the 26 cross-sectioned OEM paint samples originated. From the perspective of each search prefilter, these 26 OEM paint samples constitute an external prediction set which is the true litmus test for any classifier.

Can you elaborate on the challenges posed by epoxy infiltration in automotive paint chip analysis and how it impacts the accuracy of IR spectral matching against automotive paint libraries?

To collect FT-IR spectra from a paint chip using an IR imaging microscope, it is common practice to cast the paint chip in an epoxy followed by cross sectioning with a microtome to reveal the individual layers of the paint. Ideally, the epoxy should present little or no spectral interference. Unfortunately, our previous work has revealed that in some cases the epoxy as it cured infiltrated specific layers in some paint samples, contaminating their spectra and thereby prevent accurate spectral library searching against an automotive paint database.

Your study highlights the benefits of cross sectioning automotive paint chips without embedding them in epoxy. What specific improvements in sample preparation and spectral quality did you observe with this method?

One would like to cross section a paint chip without casting it in epoxy or any other embedding media because it would make sample preparation faster and easier and more importantly, eliminate interfering peaks due to the embedding media that otherwise could be present in the FT-IR spectra of the different layers of an automotive paint sample. The specific improvements observed in spectral quality when not embedding paint chips in an epoxy resin prior to cross sectioning included higher quality matches and correct identification of the source of the paint fragment (i.e., the line and model of the vehicle from which the paint sample originated) for all samples. This was not always the case when paint chips were cast in an epoxy resin prior to cross sectioning them to expose the different layers of the paint sample to the IR beam from the IR transmission microscope.

How does the process of unfolding IR image maps using an oblique transit and reconstructing spectra with alternating least squares contribute to obtaining high-quality spectral matches? Can you discuss the comparative results of using epoxy versus non-epoxy methods in your study?

For an unembedded paint chip, the shape of the angled chip can create problems for the construction of accurate line maps due to difficulties encountered in positioning the paint sample for scanning. The use of an oblique transit to fully bisect the chip appears to obviate the effects of skewed sample positioning. The resulting image line maps are of sufficient quality to ensure that IR spectra of the different layers of paint can be recovered by alternating least squares. The starting point and endpoint for each transit in the IR image for an unembedded paint sample is the unoccupied region of the BaF₂ disk adjacent to the paint sample. For paint chips embedded in epoxy, the starting point and endpoint is the IR spectrum of the pure epoxy used to cast the sample. A large fraction of the paint chip when cast in an epoxy resin and cross sectioned is barely visible and most of the clear coat and e-coat layer are sometimes buried under epoxy. Collecting IR spectra indicative of the different paint layers using the appropriate transit through the image for some embedded cross sectioned samples can be problematic.

References

(1) Affadu-Danful, G. P.; Zhong, H.; Dahal, K. S.; Kalkan, K.; Zhang, L.; Lavine, B. K. Raman Spectroscopy to Enhance Investigative Lead Information in Automotive Clearcoats. Appl. Spectrosc. 2023, 77 (9), 1064–1072. DOI:10.1177/00037028231186838

(2) Zhong, H.; Donkor, E.; Whitworth, L.; White, C. G.; Dahal, K. S.; Fasasi, A.; Hancewicz, T. M.; Uba, F.; Lavine, B. K. Application of Ultramicrotomy and Infrared Imaging to the Forensic Examination of Automotive Paint. J. Chemom. 2023, 37 (8), e3509. DOI: 10.1002/cem.3509

(3) Affadu-Danful, G. P.; Kalkan, A. K.; Zhang, L.; Lavine, B. K. Analysis of Automotive Paint Smears Using Attenuated Total Reflection Infrared Microscopy. Appl. Spectrosc. 2023, 77 (3), 281-291. DOI:10.1177/00037028221136122

(4) Kwofie, F.; Perera, N. U. D.; Dahal, K. S.; Affadu-Danful, G. P.; Nishikida, K.; Lavine, B. K. Transmission Infrared Microscopy and Machine Learning Applied to the Forensic Examination of Original Automotive Paint. Appl. Spectrosc. 2022, 76 (1), 118-131. DOI:10.1177/00037028211057574

(5) Kwofie, F.; Perera, N. U. D.; Allen, M. D.; Lavine, B. K. Application of Infrared Microscopy and Alternating Least Squares to the Forensic Analysis of Automotive Paint Chips. J. Chemom. 2021, 35 (1), e3277. DOI: 10.1002/cem.3277

About the Interviewee

Barry K. Lavine is a Professor of Chemistry at Oklahoma State University. Lavine’s research interests encompass many aspects of chemical analysis including optical sensors, vibrational spectroscopy, infrared and Raman imaging, and chemometrics. Lavine has published more than 115 research papers, 21 book chapters, 16 review articles, and was editor for three ACS monographs. Lavine is on the editorial board of several journals including the Journal of Chemometrics, Microchemical Journal, Analytical Letters, and Molecules. He has served as Chair of the Northern New York (1997–2004) and Oklahoma (2006–2008) sections of the ACS. Lavine has also been Program Chair for several meetings including SCIX (1992), Northeast Regional ACS Meeting (1999), and the Pentasectional Meeting of the local Oklahoma Sections of the ACS (2005). Lavine is the recipient of the 2015 Kowalski prize, the 2017 Chemometrics Award (sponsored by the Eastern Analytical Symposium) and is a Fellow of the Society for Applied Spectroscopy.