Molecular Characterization of Plastic Waste Using Standoff Photothermal Spectroscopy

An accurate molecular identification of plastic waste is important in increasing the efficacy of automatic plastic sorting in recycling. However, identification of real-world plastic waste, according to their resin identification code, remains challenging due to the lack of techniques that can provide high molecular selectivity. In this study, a standoff photothermal spectroscopy technique, utilizing a microcantilever, was used for acquiring mid-infrared spectra of real-world plastic waste, including those with additives, surface contaminants, and mixed plastics. Analysis of the standoff spectral data, using Convolutional Neural Network (CNN), showed 100% accuracy in selectively identifying real-world plastic waste according to their respective resin identification codes. Standoff photothermal spectroscopy, together with CNN analysis, offers a promising approach for the selective characterization of waste plastics in Material Recovery Facilities (MRFs).

Plastics are polymer materials consisting of repeating carboncontaining molecular groups that form long chains. 1 Their unique molecular structures give them many desirable properties such as durability, high strength-to-weight ratio, chemical stability, and water resistance.Because of their attractive physical and chemical properties and easy availability, plastics have become the primary building material for numerous disposable products 2,3 for day-to-day use.Paradoxically, these highly desirable properties also make plastic waste the most formidable environmental menace, primarily because of their recalcitrance to decomposition in a reasonable time. 4,5he extensive and ever-increasing societal dependence on disposable plastic products makes plastic pollution one of the most pressing environmental challenges facing us today.At a global scale, plastic production has witnessed a drastic increase, surging from 2.3 million tons in 1950 to 448 million tons in 2015, with a projected doubling by 2050, with only an 8.7% recycling rate. 6,7][16][17] Moreover, plastic production contributes to about 4.5% of global greenhouse gas emissions. 18Mitigating the environmental effects of waste plastic will require establishing a circular economy approach through efficient plastic recycling. 19,20Hence, increasing the recycling rate of plastic while maintaining the quality of recycled material is a vital step toward addressing the challenges posed by plastic waste.At the community level, plastic recycling is carried out in a Material Recovery Facilities (MRFs) where plastics, along with other wastes such as aluminum, cardboard, and paper, are sorted simultaneously. 21,22To avoid cross-contamination, plastic sorting in the MRFs is predominantly carried out manually, 23,24 which is timeconsuming and exposes the workers to toxic additives that might be present in the waste. 25The presence of additives and contaminants further complicates the sorting process, increasing the risk of false identifications. 26,27Therefore, there is an urgent need to develop a novel technique for plastic sorting that is highly selective and accurate.Such a technique would significantly increase the recycling rate while maintaining a high quality of the recycled material and hence improving the overall standard of the recycling process in the MRFs.
2][33][34] However, these techniques have their own limitations.For example, AI-based sorting robots lack high accuracy because of the lack of sufficient input data with high chemical selectivity.Although NIR is widely used for plastic identification, it encounters challenges due to the presence of overtones, which can lead to false identifications.6][37] On the other hand, the mid-infrared (mid-IR) region, known as the fingerprint regime, due to the absence of overtones, exhibits unique spectral signatures for different plastic materials and has the potential for effective classification of plastic waste at MRFs. 28,38 Infrared (IR) spectroscopy-based classification can be carried out in a standoff fashion by illuminating the sample with IR radiation from a distance and then by detecting and analyzing the scattered light from the sample.Standoff spectroscopy can be practiced in many different ways.In one approach, the IR wavelength is tuned over a range, and the scattered light intensity is plotted as a function of wavelength to obtain the IR spectrum of the sample.It can also be accomplished by illuminating the sample with a broadband source and detecting the scattered light using wavelength dispersive techniques.Since standoff detection is a noncontact method, it is ideal for in situ, real-time detection of plastic waste on moving conveyor belts in MRFs.Standoff identification of pure polymers using mid-IR spectroscopy in a narrow wavelength band has been demonstrated before. 39Long et al. have demonstrated detection of plastics by illumination with a broadband IR source and analyzing the scattered radiation using a wavelength dispersive detector followed by data analysis using Machine Learning (ML) technique. 29However, the demonstrated accuracy of this approach is limited.Stavinski et al. demonstrated classification of plastic waste by analyzing conventional ATR-FTIR spectra using machine z E-mail: yaolizha@buffalo.edulearning (ML) algorithms. 40However, the FTIR machine is bulky and expensive, which is not applicable for plastic waste classification in MRFs.
In this article, chemical identification of various real-world plastic samples, including those with surface contaminants, additives, and mixed plastics, using standoff mid-IR spectroscopy has been demonstrated.By analyzing the spectral data using ML techniques, six types of plastics, resin identification code (RIC) 1-6 (PET, HDPE, PVC, LDPE, PP, and PS), and mixtures were identified successfully.Furthermore, no additional preprocessing of the plastic spectra, even those with contaminants and additives, was required before applying the ML techniques for analysis.In contrast to the earlier studies, the present work can directly detect and analyze plastics in their original form, which reduces the time, effort, and potential errors associated with sample preparation.
In this standoff photothermal spectroscopy, a tunable mid-IR quantum cascade laser (QCL) was used to illuminate the plastic sample of interest.The scattered IR light is then detected using a thermally sensitive bi-material microcantilever.The bi-material microcantilever serves as a broadband IR detector by directly measuring the heat generated due to photon absorption. 41An uncooled bi-material microcantilever can detect temperature variations of 5-10 mK at ambient conditions. 42In theory, the typical bending, z, of the free end of a bi-material cantilever as a function of adsorbed IR power, P, can be expressed as: 43,44  where, α 1 and α 2 are the coefficients of the thermal expansion of the two layers of the microcantilever, t 1 and t 2 are the layer thicknesses, λ 1 and λ 2 are the thermal conductivities, and w is the width of the cantilever.The subscripts 1 and 2 refer to metal film and cantilever material, respectively.The thermal sensitivity parameter K is expressed as: Where E is Young's Modulus.Plotting the bending magnitude of the microcantilever as a function of the wavelengths of illumination reveals the mid-IR absorption spectrum of the target sample.The data collected in these experiments were categorized into three distinct groups.The first group comprises data collected from different types of plastics with resin identification codes 1-6 (PET, HDPE, PVC, LDPE, PP, PS).The second group consists of spectra obtained from mixed plastic types of samples (LDPE + PP, LDPE + HDPE), and the third group contains spectra of plastics with additives and surface contaminants (LDPE with various dyes, HDPE with milk residue, HDPE with cream residue, and PP with food contaminants).Over 50 spectra in the range of 775 cm −1 to 1900 cm −1 for each plastic sample with resin code 1-6 were collected.The collected spectral data were used for training the machine-learning model.The chosen machine learning model successfully identified various plastics based on their spectral features with an accuracy of 100%.These results show that the technique has the potential to assist in real-time sorting of plastic waste, further enhancing the accuracy of plastic identification and improving the overall quality of the recycling process.

Materials and Methods
Sample preparation.-Theplastic samples used in the present study were day-to-day items, such as milk bottles, coffee cups, food containers, and plastic bags, representative of the plastics commonly encountered in recycling centers.Some of the items contaminated with food and other materials were also used in these studies.The nomenclature for plastic type and identifying information are given in Table I.Plastic mixtures analyzed in this study were prepared by juxtaposing two plastic pieces so that both plastics could be simultaneously exposed to the IR beam.
Experimental setup.-Themid-IR spectra of different plastic samples were recorded using a standoff detection system constructed on a standard optical table as illustrated in Fig. 1.A tunable Quantum Cascade Laser (QCL) with wave numbers ranging from 775 cm −1 to 1900 cm −1 was employed as the mid-IR source (Block Engineering, LaserTune™).The QCL operates at an average power level of 100 mW, ensuring sufficient intensity for effective probing of the samples.The IR beam from the QCL was chopped at 50 Hz by a mechanical chopper (Scitec Instruments, model 350CD).The use of chopped light and detection using a lock-in amplifier improved signal-to-noise ratio during data acquisition.A ZnSe aspheric lens was used for focusing the IR beam onto the sample placed 10 cm away.The scattered IR light from the sample was collected using a concave mirror and directed onto a bi-material (Au/Si) microcantilever.The distance between the sample and the cantilever was about 30 cm.Commercially available silicon microcantilevers with typical dimensions of 500 μm length, 100 μm width, and 1 μm thickness (Nanoworld, Switzerland) were used.The microcantilevers were made bi-material by depositing a 5 nm Cr adhesive layer and a 30 nm Au layer on one side.The choice of Au as the bi-metallic layer is justified by its high thermal diffusivity and chemical inertness, which ensures optimal spectral response data.The specific thickness of the 30 nm gold coating is intentionally chosen for experimental considerations.Absorption of IR light by the bimaterial microcantilever results in its bending due to the bi-material effect.As a result, the amplitude of microcantilever bending varies sensitively as a function of the IR absorption characteristics of the target surface when the wavelength of the IR source is tuned over the resonant wavelength range.The deflection of the bi-material microcantilever is monitored using an optical beam deflection method.In this technique, 634 nm light from a diode laser was directed onto the free end of the microcantilever, and the reflected light was detected by a position-sensitive detector (PSD).The PSD signal was monitored using a custom-made electronic box and was fed into a Lock-in Amplifier (Stanford Research Systems, model SR865 A) with the frequency of the mechanical chopper serving as the reference.
Data processing.-Atotal of 14 distinct databases were created based on the spectroscopic experiments conducted.Each spectrum consists of 385 data points spanning the wavenumber range of PS Plastic cutlery at a specific wavenumber.To establish a reference baseline, a background spectrum was initially collected when no sample was present on the sample holder in the setup.Subsequently, the background spectrum served as a reference to obtain the absolute spectrum of each sample.To eliminate any potential impact from different trials of experiments and to facilitate accurate ML analysis, all the collected spectra within the range of [0, 1] were normalized using the following method, represents the normalized spectra.The normalization process was executed on each of the spectra instead of the entire sample set in order to eliminate the errors on the baseline caused by different sets of experiments and experimental setups.
Machine learning (ML) techniques were employed to perform four tasks in this study.For each task, the data set was split into two distinct groups: the training set, consisting of 80% of the samples, and the test set, comprising of the remaining 20% of the samples.This division allowed for training the Ml models on the training set and evaluating their performance on the independent test set.Such an approach helps assess the models' generalization capabilities and provides an estimate of their performance on unseen data.
The analysis framework used in this study involved the implementation of a Convolutional Neural Network (CNN) architecture, encompassing three different architectures in total.The architecture was mainly composed of a 1D convolution layer, wherein filters were employed to convert IR spectral data into vectors within a specific range.These filters aim to identify and recognize specific attributes or patterns along each sequence, transforming them into numerical values.During the back-propagation process, the weights of the filters were evaluated and updated accordingly.
To optimize the performance of the CNN, the numbers and the sizes of the filters used in the architecture were fine-tuned.To leverage the advantages of deep neural networks, nonlinear transformations in the form of activation functions, specifically Rectified Linear Units (ReLU), were applied at the end of each convolutional layer.In each model, a pooling layer is added after the convolution layer to reduce the dimension of the data that is primarily expanded by the convolution layer.In this study, a 1D max-pooling layer with a size of 5 was deployed, where the maximum value within every five values in the vectors was extracted as the representative value.To enhance the overall robustness of the model, a dropout layer with a ratio of 0.5 was introduced.This layer randomly discards 50% of the output from the convolutional layer, thus increasing the model's resilience.Finally, the output layer was fully connected to the SoftMax activation function, while the categorical cross-entropy function was utilized as the loss function.Since classification was involved in the task, the output vector had the same number of dimensions for different types of materials, and each value in the output vector represented the likelihood of the input data belonging to the corresponding material type.

Results and Discussion
Standoff photothermal detection of resin code identification.-Inthis study, three scenarios were considered to ensure the technique's validity: pure samples, samples with additives and contaminants, and samples with surface residues.Pure samples refer to cleaned samples without additives, contaminants, or residues.The samples with additives and contaminants were included to test the technique's ability to handle complex compositions.Additionally, samples with surface residues were analyzed to evaluate the performance of the technique for classification of real-world plastic waste.By examining these different scenarios, this study aimed to verify the technique's applicability and reliability across various practical circumstances.The analysis of the photothermal spectra using advanced ML algorithms allowed accurate identification and classification of different types of plastics under various conditions.
Table I shows the list of plastic samples used in the standoff photothermal spectroscopy experiments.Photothermal spectra were recorded for each of these plastic waste samples.The spectral data, after background subtraction, were analyzed using the principal components analysis (PCA) method.By performing PCA on the photothermal spectra, the dimensionality of the data was reduced while retaining the most significant features.Using this process, one can visualize the relationships between the plastic samples as well as identify the clusters or patterns corresponding to different types of plastics.The PCA analysis facilitated a comprehensive understanding of the spectral data and provided insights into the variations and similarities among the different plastic samples.This information was crucial for the subsequent development of machine-learning algorithms for plastic classification and identification.shows the PCA analysis of pure plastics (a) and plastic mixtures containing contaminants and additives (b).In scenario (a), the PCA analysis of pure plastics shows distinct clusters corresponding to different types of plastics, which indicates that pure plastics' photothermal spectra exhibit characteristic patterns that allow for effective differentiation and classification.In scenario (b), where plastic mixtures contain contaminants and additives, the clustering of spectra becomes more complex, giving more degree of overlap between clusters.This suggests that the presence of contaminants and additives complicates the sorting process and introduces additional variability in the spectral patterns.However, despite the challenges posed by contaminants and additives, the present study highlights the feasibility of employing simple Ml algorithms to classify plastic waste.By utilizing the insights gained from the PCA analysis, it is possible to develop Ml models that can effectively distinguish and identify different types of plastics, even in the presence of contaminants and additives.This finding is significant as it demonstrates that simple machine learning techniques on photothermal spectra acquired in the present work can be utilized in practical circumstances for plastic waste sorting, eliminating the need for complex methodologies.It offers a promising avenue for developing efficient and cost-effective plastic recycling and waste management solutions.
Detection of pure plastics-type 1 to type 6.-Figure 3a shows typical standoff photothermal spectra of pure plastic samples with resin codes 1 to 6 (PET, HDPE, PVC, LDPE, PP, and PS).The yaxis in Fig. 3a represent the normalized amplitude variations of the microcantilever bending as a function of the IR wavenumber (x-axis) used for illuminating the plastic samples.The extent of microcantilever bending in the spectra is proportional to the IR radiation scattered off by the plastics, with absorption peaks corresponding to each plastic's unique characteristics.These results reveal that the photothermal spectrum collected by a bi-material microcantilever effectively captures the characterization of vibrational bond information associated with each sample. 45able II presents the analysis of the observed bands in the photothermal spectrum of PET.The observed IR band at 795 cm −1 belongs to the vibrations of two adjacent aromatic H. 46,47 The photothermal band at 1342 cm −1 represents the characteristics wagging vibrational modes of the ethylene glycol segment while the band at 1453 cm −1 represents the bending of the C-O group.Finally, the band at 1730 cm −1 corresponds to the stretching of C=O of the carboxylic group. 46nalysis of the standoff photothermal spectra of HDPE, PVC, LDPE, PP, and PS shows excellent agreements with the corresponding vibrational spectra of the plastics.Detailed information on these modes is compiled and listed in Tables III-VII.
HDPE and LDPE, having the same chemical components, exhibit overlapping peaks in the mid-IR spectrum.This poses a challenge for conventional plastic sorting techniques, including those utilizing AI/ML algorithms, as they struggle to differentiate between the two types accurately.However, since the standoff plastic detection technique uses a highly sensitive bi-material microcantilever as the readout system, LDPE and HDPE peaks are distinct even for the same vibration modes, which helps the machine learning technique in Fig. 3a.The identification accuracy of HDPE and LDPE can be further improved by increasing the tuning resolution of the wavelength emitted by the QCL.
ML algorithms were implemented to classify data obtained from the six types of recyclable plastics: PET, PVC, PP, PS, HDPE, and LDPE, as shown in Fig. 3b.The datasets comprised a total of 302 Table II.Vibrational mode of bonds for plastic type-1 PET. 48T Wavenumber (cm learning rate starting at 0.001.From the confusion matrix, the prediction accuracy on the entire dataset reached 100%.These findings highlight the potential of standoff photothermal spectroscopy using a bi-material microcantilever as a powerful tool for analyzing and distinguishing various plastic types based on their vibrational properties with a 100% accuracy rate, even at a standoff distance of 30 cm.Detection of plastic mixtures of different types.-Inapplications, plastics are often encountered as mixtures, especially in materials recovery facilities (MRFs).Figure 4a shows the standoff photothermal spectra obtained from a mixture of LDPE and PP.Compared to the spectra of individual LDPE and PP, the spectrum of the mixture exhibits a linear combination of their respective peaks.Notably, the peak at 1166 cm −1 , caused by CH bending, CH 3 rocking, and C-C stretching of PP, is clearly distinguishable in the mixture spectrum.Similarly, the peak at 997 cm −1 , which belongs to CH 3 rocking, CH 3 bending, and CH bending of PP, is also prominently observed in the mixture spectrum.These results demonstrate the standoff photothermal plastic sorting technique is not only capable of classifying different pure plastics but also identifying plastic mixtures.
Furthermore, the spectrum of the mixture retains the signature of LDPE, like the -C-H bending peak at 1473 cm −1 .To further demonstrate the efficiency and selectivity of this technique, standoff spectra of a mixture of HDPE and LDPE which have similar chemical compositions were collected.As shown in Fig. 4b the observed spectrum of the mixture displays the characteristic peaks of both HDPE and LDPE.Notably, LDPE has a small methyl umbrella mode (symmetric bending) peak at 1377 cm −1 from side chains, which is present in the spectrum of LDPE but absent in the HDPE spectrum.This peak is clearly distinguished in the spectrum of LDPE and HDPE mixture obtained in the present work as shown in Fig. 4b.In addition, two other prominent peaks are observed in the mixture spectrum, including the peak at 1473 cm −1 , which belongs to −C−H bending of LDPE, and the peak at 1464 cm −1 , which belongs to C-CH 3 symmetric bending of HDPE.
Since two types of mixture materials, LDPE with PP, and LDPE with HDPE, were involved in this work, the differences between the mixture and its components were analyzed by using an explicit model.To analyze the LDPE and PP mixture, the model was first trained using the entire data of pure materials, in this case, 48 samples of LDPE and 49 samples of PP.Then the data of the mixture material were fed directly into the model as the test data.Since this study used the SoftMax layer in the output layer of the model, the result was represented by vectors containing two values, indicating the likelihood of the input data being classified as LDPE or PP.As shown in Fig. 4c, on average, the mixed samples of LDPE with PP were classified as 45.1% likely to be LDPE and 54.9% likely to be PP, consistent with the experiments.Similarly, for the LDPE with HDPE mixture, the average likelihood of the mixture samples being classified as LDPE was 56.0%, and as HDPE was 44.0%, which aligns well with the experimental findings.
Next, the classification of pure plastics and plastic mixtures was performed together, as shown in Fig. 4d.In the mixed task, in addition to the existing six types, two more types were included: LDPE+PP and LDPE+HDPE.The dataset consisted of 399 samples, each with a (1, 385) shape; like before, 80% of the samples were used for training the model and 20% for testing.The model for the mixed plastic identification task comprises of three convolution layers, with 256, 256, and 128 filters and a filter size of 15, 10, and 3, respectively.The activation functions used were ReLU.The maxpooling layer, the dropout layer, and the optimizer were set the same as before.Based on the confusion matrix, only 5% of PP were misclassified, and the overall accuracy of the entire dataset was 98.375%.The effect of additives, contaminants, and residues on the detection.-Additivesand contaminants can significantly affect the properties of plastics, making them incompatible with each other during the recycling process.Different additives and contaminants require specific recycling processes for effective removal or neutralization of the additives.By identifying the additives and contaminants, recycling facilities can optimize their sorting processes, ensuring the effective removal of unwanted substances and minimizing the impact on the final recycled product.
To test the efficiency of the standoff spectroscopy using microcantilevers for classifying plastics with additives, LDPE samples with different dyes, were investigated.LDPE samples with three different types of dyes, including black plastic, which is challenging to differentiate by conventional spectroscopy methods, were investigated.Black plastics are harder to sort with reflectance/transmittance spectroscopic methods like NIR. 36,37 The carbon particles that create the black color absorb all NIR radiation, and as a result, no detectable signal can be obtained for material separation. 50However, the proposed technique in this paper, based on MIR spectroscopy, can detect black plastics.This has the potential to improve MPW (mixed plastic wastes) characterization accuracy which is crucial in the plastic recycling process as a whole.Figures 5a-5d compares the spectrum of LDPE with the spectra of LDPE samples containing different dyes.Despite the presence of dyes, all spectra exhibit the signature peak of -C-H bending at 1473 cm −1 , characteristic of LDPE.Added black dye introduces peaks around 900 cm −1 and 1050 cm −1 , while green dye generates prominent peaks at 1400 cm −1 and 1690 cm −1 .Yellow dye results in additional peaks at 900 cm −1 , 1200 cm −1 , and 1520 cm −1 .It is clear that the standoff spectroscopy method demonstrated in the present work can detect black plastics and has the potential to improve mixed plastic wastes (MPW) characterization accuracy which is crucial in the plastic recycling process as a whole.
Advanced machine learning was employed to analyze plastics containing dyes.As shown in Fig. 5e, the output includes four different types of materials: LDPE, LDPE with green dye, LDPE with red dye, and LDPE with black dye.The model used for analyzing plastics with dye had two convolution layers, with 128, and 64 filters, and a filter size of 15 and 5, respectively.Both activation functions were ReLU.The max-pooling layer, the dropout layer, and the optimizer were set the same as before.From the confusion matrix, the prediction accuracy on the entire dataset was 100%.
Plastics that are used for packaging food, beverages, or personal care products may retain residual materials such as food particles, liquids, or chemicals.These residues can contaminate the plastic, making it more challenging to recycle.So efficient identification of plastics with contaminants can help facilities take appropriate measures to remove or separate the contaminated materials, ensuring the quality of the separated plastics.
Figure 6 shows standoff spectra of plastics with food residues, a) PP with food residues, c) HDPE with milk residues, and e) LDPE with cream residues.It is evident that additional peaks other than the characteristic peaks of respective plastics appeared in midinfrared spectra when contaminants were present.In Fig. 6e, the characteristic peak of LDPE at 1473 cm −1 due to the -C-H bending cannot be seen when cream residues are present on the surface.This may be due to the intensity difference between cream and the characteristic peak of LDPE.Machine learning algorithms were applied to discern the type of plastic with surface contaminants and can achieve 100% accuracy.Figure 6b shows the successful identification of PP with food residues, while Figs.6d and 6f demonstrate the accurate identification of HDPE with milk residue and LDPE with cream residue, respectively.
From the results presented above, it is evident that standoff photothermal spectroscopy, combined with machine learning techniques, can provide highly accurate identification of waste plastics even in the presence of additives and surface contaminants.This approach can be implemented in MRFs to enhance the efficiency of plastic sorting for recycling.However, there are potential limitations and challenges associated with its practical implementation in large-scale recycling facilities.Some of the key challenges are the need for a high rate of data collection and the availability of cost-effective tunable IR sources.Addressing these challenges will be crucial for the scalable and practical implementation of the standoff photothermal technique in MRFs.Efforts are presently underway to optimize the technique's collection rate as well as to enhance its adaptability to handle mixed plastic wastes and for its successful integration into the recycling industry.

Conclusions
The proper recycling of plastics is a big challenge in the 21st century, and the solution begins with developing newer and more efficient techniques for plastic waste sorting.A standoff technique, based on mid-IR photothermal spectroscopy, has been used for the chemical identification of plastic waste.Photothermal spectra, collected as the bending of a bi-material microcantilever, show characteristic vibrational peaks of the plastics.The standoff spectra collected in the wavelength window of 5-12 μm were analyzed using convolutional neural network (CNN) framework to identify plastics according to their respective resin codes with 100% accuracy.The CNN analysis of spectral data from plastic waste with additives and common surface contaminants, as well as black plastics also shows 100% accuracy in plastic identification.These findings offer a promising solution for enhancing the efficacy of the plastic waste sorting process by accurately characterizing them using a sustainable and environmental-friendly approach.

Figure 1 .
Figure 1.Experimental setup of standoff detection of mid-IR photothermal spectroscopy using a bi-material microcantilever for plastic waste detection.It consists of a light source (QCL), mirror, chopper, lens, Lock-in amplifier, and a cantilever deflection monitoring system.(Position sensitive detector (PSD) and a red laser).

Figure 2 .
Figure 2. Scatterplots of first, second, and third principal components (PCs).(a) pure plastics (b) plastics with additives, contaminants, and plastic mixtures.X, Y, and Z in these plots indicate the first, second, and third principal components.

Figure 3 .
Figure 3. (a) Standoff Mid-IR photothermal spectra of PET, HDPE, PVC, LDPE, PP, and PS.The (a), (b), (c), and (d) labels in the figures mean characteristic peak information corresponding to the plastics.The peak information of the (a), (b), (c), and d labels were listed in the tables from Tables II-VII.(b) Confusion Matrix of pure plastics.The accuracy of classification of pure plastics can reach 100%.The values for all other cells are all zero.
two aromatic H in aromatic bands b 1342 wagging vibrational modes of the ethylene glycol segment c 1453 bending of the C-O group d 1730 Stretching of C=O of the carboxylic acid group ECS Sensors Plus, 2023 2 043401 samples, each with a shape of (1, 385).The model designed for the pure material classification task consisted of two convolution layers with 128 and 64 filters, respectively, and filter sizes of 20 and 5.The ReLU activation function was employed in both cases.A maxpooling layer with a size and stride of 5 was added, maintaining the same dimension through padding.A dropout ratio of 0.5 was applied to enhance model robustness.The optimizer used was Adam, with a

48 PP 48 PS
bending Table VI.Vibrational mode of bonds for plastic type-5 PP.Table VII.Vibrational mode of bonds for plastic type-6 PS.

Figure 4 .
Figure 4. (a) Standoff Mid-IR photothermal spectra of PP, LDPE, and their mixture.(b) Standoff Mid-IR photothermal spectra of HDPE, LDPE, and their mixture.(c) Prediction results of the plastic mixture.(d) Confusion Matrix of pure plastics and plastic mixture.Only 5% of PP were misclassified, and the overall accuracy of the entire dataset was 98.375%.

Figure 5 .
Figure 5. Standoff Mid-IR photothermal spectrum of LDPE with (a) yellow dye, (b) green dye, and (c) black dye.(d) Standoff Mid-IR photothermal spectrum of LDPE.(e) Confusion Matrix of LDPE and LDPE with different dyes.The accuracy of classification of plastics with different dyes can reach 100%.

Figure 6 .
Figure 6.(a) Comparison of standoff Mid-IR photothermal spectra and (b) Confusion Matrix of PP and PP with food residue.(c) Comparison of standoff Mid-IR photothermal spectra and (d) Confusion Matrix of HDPE and HDPE with milk residue.(e) Comparison of standoff Mid-IR photothermal spectra and (f) Confusion Matrix of LDPE and LDPE with cream.The overall accuracy of the classification can reach 100%.

Table I .
Plastic samples used for recording the standoff photothermal spectra.

Table III .
48brational mode of bonds for plastic type-2 HDPE.48

Table IV .
49brational mode of bonds for plastic type-3 PVC.49

Table V .
48brational mode of bonds for plastic type-4 LDPE.48