Quantum cascade laser-based reflectance spectroscopy: a robust approach for the classification of plastic type

The identification of plastic type is important for environmental applications ranging from recycling to understanding the fate of plastics in marine, atmospheric, and terrestrial environments. Infrared reflectance spectroscopy is a powerful approach for plastics identification, requiring only optical access to a sample. The use of visible and near-infrared wavelengths for plastics identification are limiting as dark colored plastics absorb at these wavelengths, producing no reflectance spectra. The use of mid-infrared wavelengths instead enables dark plastics to be identified. Here we demonstrate the capability to utilize a pulsed, widely-tunable (5.59 7.41 μm) mid-infrared quantum cascade laser, as the source for reflectance spectroscopy, for the rapid and robust identification of plastics. Through the application of linear discriminant analysis to the resulting spectral data set, we demonstrate that we can correctly classify five plastic types: polyethylene terephthalate (PET), high density polyethylene (HDPE), low density polyethylene (LDPE), polypropylene (PP), and polystyrene (PS), with a 97% accuracy rate. © 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement


Introduction
The robust and rapid identification of plastic type is needed for environmental applications ranging from recycling facilities to understanding sources and sinks of plastics in the environment. For example, to understand plastic fate and transport in the environment, it is important to be able to robustly classify plastics found in locations ranging from land to the deep sea. Although hundreds of types of plastics exist with added complexity due to fillers, additives, and colorants, plastics are often identified by their recycling codes: #1 polyethylene terephthalate (PET), #2 high density polyethylene (HDPE), #3 polyvinyl chloride (PVC), #4 low density polyethylene (LDPE), #5 polypropylene (PP), #6 polystyrene (PS), and #7 other plastics.
Optical approaches for plastics classification allow samples to be identified in both stand-off configurations and in a non-destructive manner, with no damage to a sample. A range of optical approaches have been utilized for plastics classification [1], including attenuated total reflectance -Fourier transform infrared spectroscopy (ATR-FTIR) [2], laser induced breakdown spectroscopy (LIBS) (e.g. [3][4][5]), near-infrared reflectance spectroscopy (NIR) [6], and Raman spectroscopy [7]. Hybrid approaches, such as combining Raman and LIBS techniques, have also been applied to plastics identification [8]. There are challenges to some of these approaches; for example, NIR spectroscopy is limited for plastics identification, as the wavelengths cannot be used to identify black or dark (e.g. dark grey) plastics due to their low reflectance in the NIR spectral range. In the NIR region, other materials such as carbon black and soot both absorb completely [9]. For polymer sorting, Fourier transform infrared (FTIR) spectroscopy is typically too slow [9]. To address these limitations, it is important that a wide range of optical approaches for plastic identification be explored.
Mid-infrared (MIR) wavelengths are of particular interest for plastics identification, in particular, MIR quantum cascade lasers (QCLs), which are compact and can be made both high power and widely tunable. QCLs have been demonstrated to have applications ranging from trace gas detection (e.g. [10][11][12]) to explosives detection [13] to medically relevant compounds such as glucose, lactate and triglycerides [14][15][16]. The ability to be widely tunable makes them a viable source for covering a large spectral range and for measuring broadband absorbers. The compact design of the QCL and its ability to be used in a stand-off/remote operation make them a viable source for implementation in small, field portable sensors. Specular reflectance spectroscopy is a powerful and simple approach, only requiring optical access to a sample. Such an approach reduces the possibility of sample cross-contamination. For example, other MIR wavelength techniques such as FTIR and ATR-FTIR (e.g. [17][18][19]), require physical contact with the sample, which could result in cross-contamination if the sample, residues, or biofilms stick to the contacting crystal. Here we demonstrate the coupling of QCL-based reflectance spectroscopy with a classifier technique, linear discriminant analysis, for the accurate identification of plastic samples.

Plastic samples
Macroplastic samples from newly purchased consumer, laboratory, and hardware products with plastic type identified based on imprinted recycling code labels were selected (Table 1). Thirty samples of each of five types of plastics, PET, HDPE, LDPE, PP, and PS, for a total of 150 samples were selected (physical sample descriptions detailed in Appendix A, Table 6). The plastic samples included a range of color, opaqueness, and thickness. Thin film plastics were not selected due to the challenge they present with interference fringes from back-surface reflection; thus, all plastics selected were at least 0.13 mm thick. All samples were rinsed with deionized water and cut to a size of approximately 2 cm x 2 cm before analysis.

Optical set-up
A widely tunable (5.59 -7.41 µm / 1789.87-1350.07 cm −1 ) pulsed external cavity QCL (maximum average power 28 mW; Daylight Solutions Inc.) was selected based on its wavelength coverage of the key spectral peaks, identified previously by ATR-FTIR, of the five targeted plastics [2,18]. The QCL was pulsed at a 5.0% duty cycle, 100 kHz pulse repetition rate, with a 500 ns pulse width. A 45 degree fixed angle specular reflection accessory (Pike Technologies, 45Spec Accessory, 011-4500) with a 10 mm mask was utilized for sample analysis; samples were laid across the opening of the mask (Fig. 1). The laser beam has an ∼2.5 mm beam width at the 1/e 2 point. As a result, the laser beam diameter was less than the plastic area revealed through the mask, and the laser beam only interacted with the plastic sample and did not touch the mask. A gold mirror followed by a weight was placed on top of each sample to maintain or improve sample flatness as the plastic samples were often irregular in thickness. A 9 µm thermoelectrically-cooled mercury-cadmium-telluride (MCT) detector (Vigo -PCI-2TE-9) coupled to a pre-amplifier was used for collection of specularly reflected light. Two CaF 2 holographic wire grid polarizers (Thorlabs, WP25H-C) were placed in the beam path to reduce the amount of light received by the detector to avoid saturation. A lock-in amplifier (Zurich Instruments -HF2LI) was used for signal collection from the detector and data were recorded using MATLAB (R2018a). A background measurement using the gold mirror was collected prior to the measurement of every fifth plastic sample to monitor any changes in laser output power. For each mirror or plastic sample measurement, the QCL was scanned across its full tuning range 5 times. Each output reflection spectrum recorded was thus the average of 5 spectra. A single scan took approximately 5 seconds; therefore, the total time for the 5 spectra was 25 seconds. Each plastic sample was analyzed in triplicate, moving the sample between each measurement; resulting in 450 total spectra (three reflection spectra of each of the 150 plastic samples). Samples were placed onto a PIKE reflectance accessory (45°) along with a mirror and a weighted block. Two polarizers were used to limit the amount of light reaching the detector to avoid saturation. Specularly reflected light was collected by a detector and the signals recorded using a lock-in amplifier and a laptop running MATLAB (R2018a).

Data processing
The QCL spectrum of each plastic sample I(ν) was normalized by dividing by the background mirror spectrum I 0 (ν), and then was converted to the normalized specular absorbance spectrum A(ν) = -log(I(ν)/I 0 (ν)). This spectrum was then converted to its imaginary analytic signal using the Hilbert transform function in MATLAB (R2018a). The Hilbert transform was utilized as an alternative to the Kramers-Kronig [20]. The resulting spectrum was smoothed (moving average of 150) and then the 1697-1550 cm −1 region was removed from each spectrum due to a lack of identifying spectral features for plastics in this region. The removal of this region enabled a reduction in size of the spectral dataset and limited excess noise from entering the classification model.

Classification model
Linear discriminant analysis, a technique used to reduce the number of variables in a dataset, can be used as a classifier for modeling differences of groups. Linear discriminant analysis, implemented using the MATLAB Machine Learning toolbox, was utilized to develop a classification model. The dataset of 450 total spectra was split into a training set (two-thirds of the samples) and a holdout test set (remaining third of the samples). The holdout test set was chosen via a stratified random sample to ensure that the prediction accuracy on each class was equally weighted in the test accuracy. In order to capture any variance in the data, the training and test sets were randomly resampled ten times and the analysis was repeated as separate trials. The final reported test accuracy and confusion matrix results are the average of the ten repeated trials.
Linear discriminant analysis was then applied to only the HDPE and LDPE spectra to confirm that these plastic types could be classified correctly based on small differences in their two spectral peaks in the 1477-1458 cm −1 region. To confirm this, only HDPE and LDPE were used in a second classification model following the same approach as when all of the plastics were used. Two approaches were used 1) using the full spectral region (minus the omitted portion as described previously) and 2) using only the 1477-1458 cm −1 region of the HDPE and LDPE samples.

Attenuated total reflectance-Fourier transform infrared (ATR-FTIR) spectroscopy
Five representative samples were selected for ATR-FTIR analysis (PET02, PP22, HDPE28, LDPE17, and PS05, detailed in Appendix A, Table 6) to determine peak position for comparison with the QCL-reflectance data. ATR-FTIR was performed on these five samples in triplicate. These measurements were made using an Agilent Technologies Cary 630 FTIR spectrometer coupled to a D-ATR diamond crystal accessory with a single reflection sensor and a sample press. Absorbance spectra were collected using 32 scans at a 2 cm −1 resolution measuring between 4000 -650 cm −1 . A background atmospheric spectrum was subtracted from all sample spectra.

Fourier transform infrared (FTIR) spectroscopy
A Fourier transform infrared spectrometer (Bruker Vertex 80) was utilized in reflectance mode using the 45 degree fixed angle specular reflection accessory (Pike Technologies, 45Spec Accessory, 011-4500) in the sample compartment. A broadband mid-infrared globar along with a KBr beamsplitter, and a liquid nitrogen-cooled MCT detector with a ZnSe window that covers the 12,000 cm −1 to 600 cm −1 region. Each plastic was placed on top of the reflection accessory and a gold mirror was used to calculate a background spectrum. Spectra were collected using 32 scans with a spectral resolution of 2 cm −1 measuring between 1300 -1800 cm −1 . The Hilbert transform was applied using MATLAB (R2018a), and spectra were smoothed with a moving average of 10.

QCL reflectance spectra reveal distinct features for different plastic types
The QCL reflectance spectra showed clear peaks corresponding to known distinct features for all five of the plastics, with similarities in peak location for HDPE and LDPE (Fig. 2). Significant spectra-to-spectra variability among replicate runs of the same plastic sample existed, which we attribute to changes in reflection due to how the sample was placed on the reflection accessory, as each sample was moved between replicates. However, the key spectral features did not vary in location. Variability existed between different samples of the same type, which we additionally attribute to variations in plastic formulation (e.g. stabilizers, fillers, colorants, and additives) and physical variability of the samples (e.g. differences in smoothness, shininess, opaqueness, and color) (Fig. 3). Interference fringe patterns were observed in some spectra, which we hypothesize is due to refractive index differences with these plastics. Despite the differences between spectra of the same plastic type, the distinct spectral features identified for each type of plastic appeared in 95% of all spectra. Only in 21 measurements out of the 450 measurements were spectral features not clearly identifiable.  To examine the influence of color on spectra, we compared a black HDPE sample and a white HDPE sample analyzed using the QCL reflectance setup. The spectra show clearly visible peaks at 1473 cm −1 and 1463 cm −1 in both spectra (Fig. 4). This ability to analyze dark plastic samples is a key advantage of the utilization of mid-infrared wavelengths instead of near-infrared for plastics identification.

Comparisons of peak locations and ease of analysis for QCL, ATR-FTIR and FTIR
The spectral peaks present in the QCL reflectance spectra were compared to ATR-FTIR spectra reported in the literature as well as spectra taken in the laboratory using both ATR-FTIR (Fig. 5) and FTIR-reflectance spectroscopy (Fig. 6, Table 2). Limitations in these techniques must be noted as ATR-FTIR requires that the sample be physically in contact for the measurement and our FTIR measurements required the use of a liquid nitrogen-cooled detector. For ATR-FTIR, the peak locations reported in the literature [2] aligned with those we measured ( Fig. 5; Table 2). ATR-FTIR is routinely used for plastics analysis including microplastics analysis. Differences in peak location however were observed between the ATR-FTIR measurements when compared to those seen in the FTIR-reflectance and QCL-reflectance data, both of which were in agreement with each other. At the longer wavelengths measured, ATR spectral peaks are often shifted towards lower frequencies (shift in peak position) when compared to transmission or reflectance spectra [21,22]. Since plastics are routinely identified spectrally in the infrared region by their characteristic peaks, it is important to recognize these shifts.

Plastic-type identification using QCL reflectance spectroscopy
The classification model using linear discriminant analysis resulted in a 97% correct identification rate for the 150 samples analyzed. The variability between spectra for the same plastic type is hypothesized as the cause of some misidentifications. All PET samples were correctly identified (Table 3), due to the strong spectral feature at 1736 cm −1 . For each of the other four plastic types, the model was also highly successful, resulting in at most 9 plastic samples being misidentified during a single model run, with most misclassifications occurring between HDPE and LDPE. For example, during one model run, 6 HDPE samples were incorrectly classified as LDPE. HDPE and LDPE both have a spectral peak at 1463 cm −1 and a closely spaced second peak at 1473 cm −1 and 1472 cm −1 for HDPE and LDPE, respectively. To confirm that the slight peak difference allows for the discrimination of HDPE and LDPE, linear discriminant analysis was then run on only HDPE and LDPE. When the full spectral region (minus the chopped portion as described previously) was included, the success rate for identification between HDPE and LDPE was 88 +/-4% (Table 4). When only the peak region (1477 -1458 cm −1 ) was utilized, the success rate increased to 97 +/-3% (Table 5). Therefore, this suggests that the small difference in the HDPE and LDPE peaks allows for the discrimination to take place.
a Values are the average and standard deviation of 10 repeated random splits of the data using test sets containing 30 measurements of each plastic type.

Conclusion
QCL-based MIR reflectance spectroscopy coupled to a classification model using linear discriminant analysis was demonstrated to be a successful approach for rapid and robust identification of plastic type with a 97% correct identification rate. Each set of five spectra taken took only 25 seconds to acquire. In future set-ups, this could be reduced to one spectrum, resulting in a very rapid 5-second analysis time. While five different types of plastics were selected that had strong spectral features in the 5.59 to 7.41 µm region, other plastics, such as polyvinylchloride (PVC), were not included in this study due to the lack of strong spectral features in this region. However, due to the ability to design and fabricate QCLs at specific wavelengths, other plastics should also be identifiable using this same approach by selecting a QCL with the appropriate wavelength region. The use of widely-tunable QCLs (e.g. [24]), multiple QCLs or QCL arrays [25] would allow a broader range of plastic types to be distinguishable. QCL beam diameters are typically on the order of ∼3mm in diameter [26] but can be focused down to reduce the beam to less than 300 µm (e.g. [27,28]), and some calculations point to beam diameters achieved as small as ∼20 µm [26]. Although macroplastic samples were utilized here, the use of a smaller diameter beam would make it a viable approach for the analysis of smaller (<100 µm) plastic samples including microplastic (<5 mm) samples.
QCLs are tiny sources that can be designed to operate at mid-infrared wavelengths and at the same time can be made widely tunable. Incorporating a QCL into a small sensor, that does not require physical contact with the plastic sample, could have broad applications for the identification of plastic for recycling and environmental applications. If an environmental application was sought, future studies would be needed to examine plastic samples collected from the environment, which have been chemically and physically weathered by environmental processes. This weathering, which could occur in both terrestrial and aqueous locations, has the potential to alter spectral peaks [29]. Other characteristics of samples found in the environment (e.g. wetness of sample, presence of water or fluids in a sample, presence of thin plastic or paper labels, curvature of surfaces) may introduce additional classification issues. Further laboratory studies would also be needed to fully understand these impacts and to design methodologies that allow for correct identification. The samples used in this study were newly acquired plastic samples and offer an important first-step in presenting the ability to use a QCL to identify plastic type.

Disclosures
The authors declare that there are no conflicts of interest related to this article.