ATR-FTIR Spectroscopy Combined with Chemometric Methods for the Classification of Polyethylene Residues Containing Different Contaminants

Low density polyethylene (LDPE) and high density polyethylene (HDPE) are the principal plastics present in solid plastic waste and are found out as the main components of microplastics in marine and terrestrial environments. Currently, efforts have been made to develop new and effective methods to ensure the identification and separation of plastics in waste, ensuring the necessary purity to obtain quality and economically competitive recycled products. In this contribution, we investigated the usage of Fourier-Transform Infrared Spectroscopy in Attenuated Total Reflection mode (ATR-FTIR) combined with Principal Component Analysis (PCA), Linear Partial Least Squares Regression by Intervals (iPLS-R) and Competitive Adaptive Weighted Sampling (CARS/PLS-R) as chemometric methods to classify and determine the compositional fraction of the pristine and recycled mixtures of HDPE and LDPE from plastic waste in São Paulo, Brazil. Here, the 3D PCA plots were not effective in classifying the different polyethylenes and their polymer blends using the three main Principal Components (PC), but 2D PCA diagrams using PC1 and PC3 presented promising performance for this purpose. The iPLS-R presents the best predictive ability than CARS/PLS-R to determine the LDPE content in HDPE/LDPE recycled blends. However, the presence of different contaminants (in 5 wt%), such as silicon dioxide (SiO2), calcium carbonate (CaCO3), recycled polypropylene (PP), and recycled poly(ethylene terephthalate) (PET), reduces the potential usage of the iPLS-R models as identification tools for LDPE and HDPE sorting in industrial recycling processes.


Introduction
believe that this work is a way to contribute scientifically, even in a small way, in this very relevant field nowadays.
Independent of the recycling technology, separation, and sorting of plastic waste play an essential role in obtaining better quality recycled products and optimizing the industrial recycling process [23]. Fourier-Transform Infrared Spectroscopies (FTIR) are attractive vibrational spectroscopies because qualitative and semi-quantitative analyzes can be performed in a non-destructive way. When they are combined with multivariate analysis methods, as chemometric strategies, it is possible to identify and classify rapidly the plastic waste that is suitable to automatize the sorting at industrial scale in recycling processes [24], despite the negative influences caused by the high presence of pigments in black and dark-colored plastic residues [25]. The infrared spectroscopies can also be used to characterize and identify other solid waste types (not only plastic waste) [26] and used to evaluate the aging of plastics [27,28], a relevant factor that acts on polymeric materials limiting the mechanical recycling of their residues. Moreover, FTIR analysis can be performed in different wavenumber ranges [29]: Near Infrared (NIR, 12,800 -4000 cm -1 ; Medium Infrared (MIR, 4000 -400 cm -1 ); and Far Infrared (FIR, 400 -10 cm -1 ). Bekiaris and co-workers [30] successfully predicted the labile fraction of carbon in several composed and non-composed organic waste products using FTIR Photoacoustic Spectroscopy (FTIR-PAS) combined with Principal Component Analysis (PCA) and Partial Least Square Regression (PLS-R). Signoret et al. [31] applied Fourier-Transform Infrared Spectroscopy in attenuated total reflection mode (ATR-FTIR spectroscopy) in MIR spectral region to identify and differentiate various styrenic plastics in Waste of Electric and Electrical Equipment (WEEE). Lenz et al. [32] used FTIR measurements to construct PLS-R regression models as a multivariate analysis tool to control and monitor emissions involving leachate compounds from municipal solid waste (MSW) in landfills. Alassali et al. [33] demonstrated that NIR infrared spectroscopy and PLS-R can be employed to measure quantitatively and accurately the degradation of PE, PP, and PET after different aging conditions but the accuracy of this chemometric method depends on the plastic type. Nunes and co-workers [34] reported that FTIR measurements combined with PCA and PLS-R analyses are appropriate to identify the origin of the olive waste from different mono cultivar olive pomaces and these multivariate analyses present possible application in the quality control of suppliers. Zheng et al. [35] investigated the usage of NIR hyperspectral imaging data processed by PCA statistical method to characterize and sort several plastics in the solid waste, such as polystyrene (PS), PP, PE, PET, and poly(vinyl chloride) (PVC). The authors concluded that the similarity of infrared spectra from the plastics humper the increase or adjust properties to the polymer, and achieve a required performance over a specific application [6].
Low density polyethylene (LDPE) and high density polyethylene (HDPE) are the main polymers present in the plastic waste, followed by polypropylene (PP) and poly(ethylene terephthalate) (PET) [2,6]. Moreover, polyethylenes (PEs) are the main polymers identified as microplastics in marine and terrestrial environments [7][8][9][10], being able to remain in drinking water even after conventional wastewater treatment process [11].
Developed and developing countries have been applied different technological methods and politics for the management of plastic waste as a way to reduce the socio-environmental risks and impacts caused by this type of waste [6,[12][13][14]. Here, politics mean a set of laws, or a solid waste management plan (a technical document), with legal value adopted by a country or enterprise to manage its generated waste in an environmentally appropriate way. In this way, plastic waste recycling has been used as an eco-friendly option to reinsert plastic residues in industrial production or develop new products from polymer residues. Mechanical recycling has been widely used to reuse thermoplastic polymer waste in an economically inexpensive manner with a low investment cost, being applied mainly in underdeveloped countries. On the other hand, developed countries, such as the USA, Japan, and European countries, tend to opt for energy recycling which provides an immediate reduction in the volume of plastic waste, but it requires a high financial investment [15][16][17][18][19]. In this context, the recycling situation for polyethylenes is highly worrisome since they are polymer commodities with a relatively low price, accompanied by the high expenses (in cost/kg) for transporting the plastic waste, making their recycled products economically uncompetitive against the price of pristine polyethylenes. Consequently, the recycling processes for revalorization of polyethylene residues are generally considered uneconomic for several recycling industries [5]. Despite the recycling and revalorization of polymeric polyethylene waste is currently considered economically attractive in several countries, several researchers and international agencies [20,21], such as the United Nations Environment Program (UNEP) [22], have pointed out and warned that the Earth would not be able to support the current lifestyle concerning global plastic waste production and pollution, whose environmental impacts a long timing can be irreversible and lead to unpredictable damage to biota and fauna, including human life. In this way, humanity urgently needs to rethink its actions on the Earth's surface and the impacts caused by them, being necessary that we question and consider to what extent the economy is more important than life security or actions that contribute to health and quality of life. Although our research does not solve such problems, we

Sample Preparation
The pristine and recycled HDPE/LDPE blends were prepared in the Dynisco extruder (heating zones operating at 190 and 200 °C; screw rotation speed equal to 220 rpm), using twenty-one different compositions from 0 to 100 wt% of HDPE. Recycled PET, recycled PP, powdered silica (SiO 2 ), and powdered calcium carbonate (CaCO 3 ) were used as known contaminants in this research work. For this purpose, these materials were added to the HDPE/ LDPE polymeric blends based on the recycled polyethylenes (known contaminant content fixed on 5 wt%). Only the recycled HDPE/LDPE blends with PET were extruded with heating zones operating at 220 and 230 °C, corresponding to the minimum temperature range for plasticizing and processing PET.

Scanning Electron Microscopy (SEM) and X-ray Dispersive Energy Spectroscopy (EDS)
The SEM micrographs and the elementary analyzes by EDS spectroscopy were obtained by a Field Emission Gun Scanning Electron Microscope (FEI Company, model FEG -Inspect F50). For this, the polymers were covered with gold (thickness from 5 to 10 nm).

Surface Roughness by SEM
The SEM micrographs of the surface of the samples, in the form of pellets, were performed in the scanning electron clustering into distinctive groups, narrowing the application of this procedure to classify unknown plastics in the complex mixture waste.
In summary, PLS-R and PCA are mathematical-statistical methods practical to correlate the concentration of an analyte with a measurable property within a complex system, whose measured responses or analytical signals suffer direct or indirect action of different interference effects, which are intrinsical to the equipment or the measurement method or other physicochemical variables of the system itself [29,36]. While PLS-R permits predict the analyte content or property of the complex system by a mathematical regression model, PCA enables distinguishing the samples in distinctive groups by the FTIR spectral data decomposition in score plots of the several Principal Components (PCs) [35]. The main advantage of PLS-R is the easy treatment for interpretation of highly correlated data with noise, with physical-chemical interference, and with overlapping signals that are commonly observed in experimental data obtained from vibrational spectroscopies. PLS-R statistical modified versions have been proposed, such as Linear Partial Least Squares Regression by Intervals (iPLS-R) and Competitive Adaptive Weighted Sampling (CARS/PLS-R) [37]. However, few studies reported in the literature involve using these multivariate calibration methods to identify and classify plastic waste [38][39][40]. iPLS-R is a type of modified PLS-R regression, where the sample spectrum is previously subdivided into smaller equidistant intervals and, subsequently, PLS-R regression is applied to each one. In this way, it is possible to optimize prediction models because subintervals with significant interference or without relevant information can be disregarded, and, generally, iPLS-R models provide more robust prediction models with lower prediction errors [41,42]. The CARS/PLS-R algorithm was designed to interactively find the most minor "optimal" subset of measured experimental data to obtain the predictive model based on the PLS-R analysis with the lowest error values for prediction by a Monte Carlo sampling method that mimics the biological natural selection process [43].
This work aims to evaluate the potential and limitations associated with the use of ATR-FTIR spectroscopy to classify and analytically determine the compositional fraction of the recycled HDPE/LDPE blends. The PCA chemometric method was used as a classification tool, while iPLS-R and CARS/PLS-R were adopted as quantitative analytical implements. ATR-FTIR spectroscopy combined with multivariate analysis methods is a tool with potential application in the circular economy, being a helpful tool to close the production cycle (through reuse and proper recycling) of packaging, electronics, and automobile parts manufactured with mixtures of polyethylenes. before the CARS/PLS-R multivariate analyses (constant parameters: cross-validation = 15; the maximal number of latent variables for cross-validation = 20). These were the best conditions for iPLS-R and CARS/PLS-R regressions achieved in previous works involving the HDPE/LDPE polymeric system [38,39]. Figure 1 illustrates the procedure for the sample preparation, characterization, and ATR-FTIR data analysis using PCA and PLS-R multivariate analyses.
As a way to assess the performance and generality of prediction from PLS-R models, statistical parameters of mean cross-validation square error (RMSECV) and mean prediction square error (RMSEP) are used and calculated by Eq. (1) [38,48].
where z is the number of spectra, y i is the reference LDPE concentrations for the sample i, and y i is the predicted LDPE concentration of the sample in the calibration set (RMSECV) or the independent validation test set (RMSEP). The adjustment degree between the predicted values and the reference values calculated by the PLS-R methods is microscope (FEI Company, model FEG -Inspect F50). The samples were covered with gold (film with a thickness of 5 to 10 nm), using a BALZERS -SCD050 (BALTEC) table sprayer with a deposition time of 25 s. SEM micrographs (2000 times magnification) were made with secondary electrons, using a working distance of 11.14 mm and an angle of 90 ° between the electron beam and the sample. The mean quadratic roughness (R q ) was also obtained with the ImageJ program, but the micrographs were digitized in 32-bit grayscale, and the roughness was determined by the SurfCharJ plugin developed by Chinga and collaborators [44].

Thermogravimetric Analysis (TGA)
The thermal decomposition profiles of the polymers were identified by a thermogravimetric analyzer (TGA/SDTA 851e, Mettler Toledo. The samples were heated from 25 to 600 °C at the rate of 10 °C min −1 under a continuous flow of N 2 (50 mL min −1 ).

Attenuated Total Reflectance -Fourier Transform Infrared Absorption Spectroscopy (ATR-FTIR)
ATR-FTIR spectra of the HDPE/LDPE pellets were collected by an ATR accessory (ZnSe crystal) on a Thermo IS5 Nicolet spectrometer and collected from 600 to 4000 cm −1 at room temperature, 16 scans and spectral resolution of 4 cm −1 .

Software, Multivariate Analysis, and Data Treatment
All ATR-FTIR data were smoothed using the Savitzky-Golay method [45], using polynomial order of 5 and window points of 10. Subsequently, the ATR-FTIR spectra were normalized before the multivariate analysis procedure. As an ordination method, PCA was carried out in Past4.01 software and Principal Component Analysis Tool v1.50 (MAT-LAB R2018b). CARS/PLS-R regressions pre-processed spectra were performed on MATLAB software (version R2015a) using libPLS 1.95 toolbox [46]. iPLS-R regressions were carried out in MATLAB (R2015a) environment using the iToolbox package [47]. The ATR-FTIR data from the recycled HDPE/LDPE blends without known contaminants were applied as a cross-validation set for the iPLS-R, while the spectral data from the contaminated recycled HDPE/LDPE blends were used as an independent prediction set.
For the iPLS-R algorithm, all the ATR-FTIR spectra were segmented in 30 equidistant intervals, processed using the autoscaling method, and the selected spectral region for the iPLS-R corresponds from 2800 to 3000 cm −1 . In contrast, the mean centering pretreatment method was used Material), the impurity content in the recycled polyethylenes is very low since the carbonaceous residues were 0.72 wt% and 0.78 wt% for pristine and recycled LDPE, respectively. At the same time, the residual mass is 0.68 wt% for pristine HDPE and 0.98 wt% for recycled HDPE. Moreover, the recycled and pristine polyethylenes present a similar mass loss curve with a single thermo-decomposition process with maximum rate, indicating the absence of another type of polymer in the recyclates.
The recycled LDPE also has impurities, as shown in Fig. 3, composed of Al, Fe, Mg, O, and Si. These elements are typical chemical constituents of silicates [54]. Since particle size is in the micrometer order, these silicates were probably applied as a filler or come from contamination after the disposal of the LDPE-based product. However, LDPE is typically used to manufacture food packaging, and nanometric clays can be introduced in LDPE and other polymer films to reduce oxygen gas diffusion into the product due to a barrier effect [55], reducing the speed oxidation of packaged food. In addition, nano-clays are used as reinforcement fillers in polymers, being widely applied in the development of polymer matrix nanocomposites [56,57].

ATR-FTIR Spectroscopy
The ATR-FTIR spectra of HDPE and LDPE, pristine and recycled, are shown in Fig. 4. The vibration band located at 705-735 cm −1 corresponds to the deformation in the plane (per rotation) of the connections in the methylene group (amorphous phase and crystalline). The balance-type deformations of the C-H bonds in the amorphous and crystalline phases are associated with the signal at 1450-1480 cm −1 . The signal at 2875-2770 cm −1 is related to the symmetrical stretching of CH 2 (amorphous and crystalline phases), while assessed by the correlation coefficient (R), given by Eq. (2) [40].
where y mean is the average result for the reference LDPE concentration of all samples in the cross-validation set (R cal ) or the independent/external validation set (R pred ).

SEM and EDS Analyses
In the SEM micrographs of the recycled HDPE (Fig. 2), it is possible to identify impurities on the surface of this polymer. The EDS spectra indicate that these contaminants are constituted by K, O, Si, Na, Mg, Fe, S, Cl, Ca, and P chemical elements. These chemical elements can be related to different additives present in HDPE, as detailed in Table 1 [ [49][50][51][52]. The presence, in greater concentration, of carbon is due to the chemical composition of the PE polymer chains. Besides being associated with the additives mentioned previously, the oxygen element can also be associated with oxidative degradation of the polymer during its processing and life service time. These EDS elemental signals cannot be associated with residual Phillips catalyst (based on chromium oxide) supported on silica or alumina and residual Ziegler-Natta catalysts based on alkyl aluminum compounds and a salt (containing Ni, Co, Zr, or Ti) [53] due to their low residual concentrations for the synthesis of HDPE, since these chemical elements were not identified in the pristine HDPE. According to TGA results (Supplementary  1340-1390 cm −1 . These infrared alterations are associated with the differences in the content of amorphous and crystalline phases occasioned by the increase of CH 3 groups in LDPE due to the highest quantity of polymer branches in this polyethylene type than in HDPE [63]. The characteristic absorption signals for the chemical groups and their vibrational modes in PP [64], PET [65][66][67], silica (SiO 2 ) [68], and calcium carbonate (CaCO 3 ) [66,69] from the ATR-FTIR spectra (Fig. 5) are detailed in Table 2. The FTIR spectrum from recycled PET does not indicate the presence of functional groups due to the presence of contaminants. Vaterite (space group Pbnm), calcite (space group R − 3 C) and aragonite (space group Pmcn) are the three CaCO 3 polymorphic crystalline phases based on hexagonal, trigonal, and orthorhombic crystal systems, in that order [70,71]. Silicon dioxide (silica) also displays polymorphic behavior, presenting eight different crystalline phases containing a different number of SiO 2 groups per unit cell [72]: α-quartz (trigonal cell unit with space group P3 2 21), α-cristobalite (tetragonal cell unit with space group P4 1 2 1 2), α-tridymite (orthogonal cell unit with space group C222 1 ), coesite (monoclinic cell unit with space group C2/c), keatite (tetragonal cell unit with space group P4 3 2 1 2), tridymite (hexagonal cell unit with space group P6 3 /m mc), β-quartz (hexagonal cell unit with space group P6 2 22), and β-cristobalite (cubic cell unit with space group Fd-3 m). Highlighting that calcite and α-quartz are the crystalline phases most stable at ambient temperature and pressure conditions, while the space group of vaterite is not well established in the literature [73]. The ATR-FTIR spectra suggest the calcite phase in the CaCO 3 particles due to characteristic absorption bands at 874 and 2512 cm −1 .
HDPE, LDPE, and PP recyclates seem to have a low degradation degree since there is not infrared signal at 1170 and 1167 cm −1 , corresponding to the formation of C-O (carbonyl) from ester groups due to the aging of polyolefin residues after their end-of-life and mechanical recycling steps [27].
The remarkable intensity of the C=O stretching vibration from carboxylic acid groups at wavenumber 1715 cm −1 indicates that the PET recyclates present a high degradation degree [27,74], justifying the elevated value for the MFI (14 g/10 min at 255 °C) of the recycled PET. During the mechanical recycling, PET residues can degrade by thermal-oxidative and hydrolytic action combined with water and oxygen molecules, both involving breaking the ester bonds between the terephthalic acid and diethylene glycol units along the polymer backbones. The thermal-oxidative degradation leads to rupture of the polymeric chains, forming carboxyl acid and vinyl ester end groups in the PET. On the other hand, hydrolysis also causes the PET polymer the vibrational band at 2980-2875 cm −1 refers to the asymmetric stretching of CH 2 (amorphous phase). The absorption bands with low intensity at 1306-1351 cm −1 were observed and are connected with the out-of-plane deformations (torsion and swing type) of the CH 3 groups. A weak absorption at 1340-1390 cm −1 is observed in the polyethylene spectra, also due to the balance-type deformation of the CH 3 groups [58][59][60][61][62]. The main differences between the ATR-FTIR spectra from HDPE and LDPE correspond to variations in the absorption intensity of infrared radiation in the 1450-1480 cm −1 and 2980-2875 cm −1 bands. Also, there are differences in the ATR-FTIR spectra region at

Principal Component Analysis (PCA)
The 2D and 3D diagrams of the PC scores obtained by the PCA statistical analysis of the ATR-FTIR data from the pristine and recycled polyethylenes are presented in Fig. 6. The PC1 and PC2 scores correspond to 72.9% and 8.6% of the variance among the infrared spectral data from the polyethylenes, respectively. However, the PC scores do not enable the complete distinction of the pristine and recycled polyethylenes without the 5 wt% of contaminants (pHDPE, pLDPE, rHDPE, and rLDPE samples) due to the overlap of the 95% confidence ellipses. Moreover, the PC1 score evidently has less representativeness over the ATR-FTIR spectra of LDPE. In contrast, the PCA plot using PC1 and PC3 scores permitted the separation of some plastics, such as pristine polyethylenes, pristine HDPE and recycled LDPE (with or without 5 wt% of contaminants). Also, the contaminant presence reduces the confidence ellipse area for the recycled HDPE, enabling a more effective separation between the recycled HDPE and the samples containing LDPE.
The first three PC scores (PC1, PC2, and PC3) from PCA analysis represent 76.1% of the total variability for the HDPE/LDPE blends based on recycled and pristine plastics, as can be seen in Fig. 7. Obviously, the addition of the contaminants modifies the format of the confidence ellipses, but the recycled and pristine HDPE/LDPE without contaminants present a similar confidence ellipse. The presence of 5 wt% of silica in the recycled HDPE/LDPE substantially increases the variation of the calculated PC scores, and consequently, the confidence ellipse area is enlarged. According to the 2D diagrams in Fig. 8, the PC1 and PC2 scores do not make it possible to separate the HDPEs, LDPEs, and HDPE/ LDPE blends effectively since the groups are not clustered by the PCA analysis, indicating a high similarity of their ATR-FTIR spectra [35]. The 2D diagrams of PC1 and PC2 scores from PCA analysis using two different software in Figs. 7a and 8c are very similar, indicating a good reproductivity of this statistical analysis with the ATR-FTIR data from the HDPE/LDPE blends.

Partial Least Squares Linear Regression Modified by Competitive Adaptive Reweighted Sampling (CARS/ PLS-R)
The value obtained for RMSEP for the CARS-/PLS-R model is very high (Fig. 9), equal to 39.712 wt% of LDPE, and the value of R pred is equal to 0.331. The predicted results form a straight line with an inclination equal to 1, but with a linear coefficient other than zero as a constant value shifted the model's prediction line, 40 wt% of LDPE, concerning the regression line (built in the multivariate calibration of chains, but carboxyl acid and hydroxyl ester end groups are generated by this degradation process [75]. As the ATR-FTIR spectroscopy analyses are performed on the samples' surface, it is of great importance to evaluate their surface roughness. Thus, the R q roughness of the HDPE and LDPE mixtures was determined from SEM micrographs using computational analysis. This type of essay is based on local brightness variations in the SEM micrograph (i.e., changes in grayscale uniformity) directly associated with the topography of the sample [44,81]. According to Fig. 10, the R q roughness results of the HDPE pellets are 100% higher than R q values for the LDPE pellets, independently of the plastics being recycled or pristine even though these plastics were extruded under similar conditions. Thus, R q roughness emphasizes the weight of the topographical peaks and valleys in the results of the surface roughness. It is observed in Fig. 10a that the roughness of the pellets of the pristine LDPE/HDPE blends increases as the HDPE concentration increases in the polymer mixture. When the HDPE content is equal to or higher than 50 wt%, there is a significant increase (> 25%) in the R q roughness of the pellets of recycled LDPE/HDPE blends (Fig. 10b).
the CARS/PLS-R prediction model). According to Miller et al. [76], this type of non-agreement between calibration and prediction values may be connected with systematic or random errors in the spectrometer's background signal. Therefore, this high discrepancy suggests that the CARS/PLS-R model is susceptible to slight variations in ATR-FTIR spectral signals related to the heterogeneity of recycled plastics. The high prediction error and low calibration error for the CARS/PLS-R predictive model indicate that the CARS/ PLS-R multivariate calibration method can obtain highly calibrated predictive models with low capacity to predict the LDPE content in plastic waste containing HDPE and LDPE mixtures.
The Competitive Adaptive Weighted Sampling (CARS) algorithm forces the PCR regression to calibrate with a low error, but it makes the prediction models show slight effectiveness to predict the data of an external set used in the prediction test. It is probably due to the small number of wavelengths from FTIR spectra used by the CARS-PLS predictive model, which is caused by the reduction of spectra data information in the predictive model forced by the CARS algorithm to reduce machine time for data processing. Consequently, the prediction performance of the CARS-PLS model is more susceptible to variations of the FTIR spectra, for example, the spectrometer's background signal or infrared laser scattering by the surface roughness. Then, the low performance of the CARS/PLS-R model is also associated with the increase in the surface irregularity of the recyclates' pellets since HDPE increases the viscosity of the HDPE/LDPE blend in the molten state because HDPE has a lower melt flow index than the LDPE at the same temperature. The rise in melt viscosity of HDPE, LDPE, and HDPE/LDPE polymer systems intensifies the shear rate on the surface of the molten material during its passage in the extruder matrix and, consequently, there is the formation of surface defects due to plastic deformation in the extruded polymer. If the shear stress in the extruder matrix is much higher than the melt strength, the defects are no longer superficial and the melt fracture, which is a macroscopic defect, occurs. In addition, the elevation in melt viscosity results in more outstanding adhesion with the machinery, which also intensifies the occurrence of surface defects and fracture of the cast of the extruded profile. In addition, the mixtures of HDPE with LDPE can form immiscible multiphase systems in which there are morphological changes (e.g., deformation, rupture, and coalescence) and interfacial processes (e.g., interfacial relaxation and sliding) between the dispersed polymer droplets and the polymeric matrix under stress shear; both phenomena significantly influence the HDPE/LDPE melt flow within the extruder matrix, leading to changes on the extrudate surface [77][78][79][80]. Due to the highest predictive performance, the predictive capability of this iPLS-R model to determine the composition of polymeric blends of recycled HDPE/LDPE containing other materials in its composition was evaluated. In this work, we chose the contaminant concentration of 5 wt% for PP, PET, calcium carbonate and silica, since PP and PET are two of the most present thermoplastics in solid urban waste generated worldwide. Calcite and silica are often used in polymers as fillers to improve the dimensional stability of plastic products to reduce the product's final cost, as these inorganic materials are generally cheaper than the polymers available on the market. The ATR-FTIR spectra of the uncontaminated and contaminated HDPE/LDPE blends obtained with the recycled plastics are present in Supplementary Material.
As shown in Fig. 11, the prediction errors of the iPLS-R model increase significantly when the HDPE/LDPE blends present 5 wt% of PET, PP, and carbonate of calcium applied as contaminants in this recycled polymeric blend system. The minor prediction error was obtained in tests involving polymeric blends with silica gel (RMSEP = 7.771 wt% of LDPE) and a prediction error exceeding 16 wt% of LDPE when the plastic samples contain the other three contaminants investigated in this work. In addition, the pronounced calibration and prediction errors are also associated with the increase in the surface roughness of the plastic pellets (Fig. 12). The increase of the surface roughness of the polyethylene pellets due to the elevation of the viscosity of the plastic melt, which is caused by contaminant addition in the recycled polyethylene, leads to more significant uncertainty in the ATR-FTIR spectral measurements.
According to Fig. 13, there is a significant increase in the roughness of the recycled LDPE pellets when it is used silica and PP as contaminants. Silica and calcium carbonate substantially increase the roughness of the recycled HDPE, while PET and PP seem not to affect it. This phenomenon is because the contaminants (principally SiO 2 and CaCO 3 ) decrease substantially the melt fluidity of the PE-based systems, increasing their viscosity in the molten state, and consequently, the pellets' roughness is raised as discussed so far.

Final Remarks
Automatized and efficient processes to classify plastics are essential for the recycling industry to improve the recycling of the principal plastics in the waste generated worldwide. However, there are several difficulties in producing recycled polyethylene residues with technologically attractive properties and competitive prices, such as, for example, the degradative processes that act on the polymer (during and after its lifetime) and the presence of contaminants

Partial Least Squares Linear Regression (iPLS-R)
The iPLS-R algorithm found the minimum calibration and prediction errors in the range of 2875-2980 cm −1 to construct the iPLS-R model, suggesting a close relationship between composition and the amorphous fraction of the HDPE/LDPE blends. The results of the prediction test of the iPLS-R model are shown in Fig. 11. As can be identified, the RMSEP error is equal to 6.479 wt% of LDPE, and the R pred is 0.995, while there is a low adjustment factor for the recycled blends containing less than 10 wt% of LDPE.   6 2D (a and c) and 3D (b) diagrams of PC scores from PCA analysis (with 95% confidence ellipses) for recycled HDPE (rHDPE), recycled (LDPE), pristine HDPE (pHDPE), pristine LDPE (pLDPE), and polyethylene recyclates with 5 wt% of the different known contaminants (silica, calcium carbonate, recycled PP, and recycled PET) that often prevents the separation of solid polymeric waste (SPWs) [17,82]. In addition, the quality control of the separation step, that is, assessing the purity degree of the different plastics separated to be mechanically recycled, plays a fundamental role in obtaining good quality recycled plastic products. However, quantifying the composition of mixtures of HDPE with LDPE is not a simple task since they are only constituted by hydrogen and carbon atoms connected by saturated chemical bonds.
In Brazil and other underdeveloped countries, plastics are recycled mainly by mechanical recycling, which is the most used process to reuse plastic residues [60]. However, the complete separation of the mixtures of HDPE and LDPE residues by the method of density difference, using alcoholic and aqueous solutions, after manual separation, is challenging. In addition, the presence of fillers (and other additives), plastic mixtures, and multilayer polymeric packaging in the plastic waste make this conventional sorting method unfeasible [83]. It is worth mentioning that this separation procedure is currently the cheapest and, consequently, it is the most used for the separation of polyolefin waste in the mechanical recycling process in Brazil [53].
Another aggravating factor for the proper identification and separation of plastics in polymeric packaging waste by regulatory laws. In Brazil, the correct identification of plastic-type in solid plastic waste by ABNT NBR 13.230 does not exceed 30% of the total food and non-food products sold on the market [84], making the processes characterization and sorting involved in plastic recycling even more difficult.
In our work, the identification of polyethylene recyclates using ATR-FTIR spectroscopy combined with PCA and different PLS-R statistical analyses were investigated in-depth. This methodology would be helpful for the fast characterization and sorting of plastic pieces in a large-scale recycling process involving solid plastic waste. Although the sample needs to be placed directly in contact with the ATR crystal, the collection of infrared absorption spectra by ATR-FTIR spectroscopy is not time-consuming for sample preparation, taking a few minutes to collect the infrared spectra of each sample faster than other FTIR methods. However, several limitations were identified in this proposed procedure due to the great complexity associated with the heterogeneity observed for plastic residues, which affect the ATR-FTIR signal associated both by the composition and the surface morphology of the sample.
The objective of the present work was to point out the applicability and the barriers to the usage of the ATR-FTIR spectroscopy, which is a technique that makes it possible to characterize analytically and quickly solid samples, such as solid plastic waste, being suitable to be applied in different industrial sectors. The PCA and PLS-R multivariate analyses show appropriate development of fast, analytical,  limitations, the analytical iPLS method proposed here presents more excellent reliability for mixtures of LDPE/HDPE with silica impurities under the evaluated conditions of tests. The PCA method using the PC1 and PC3 plot can be helpful to separate pristine and recycled LDPE and HDPE, depending on the waste composition, showing a latent capacity to be applied as a tool to identify polyethylene-based plastic packages and support the management of the circular economy processes.

Conclusions
In this paper, PLS-R and PCA multivariate analyses were investigated for the ATR-FTIR identification and sorting of different mixtures of recycled HDPE and LDPE with or without PP, PET, silica gel, and calcium carbonate applied as known contaminants. The HDPE, LDPE, and their mixtures were not clustered into distinct groups using their ATR-FTIR spectral characteristic signals using PC1 and PC2 scores in two-dimensional and three-dimensional PCA plots. Only the PCA plot with PC1 and PC3 scores enables grouping the pristine LDPE and pristine HDPE, as well as recycled polyethylene mixtures with 5 wt% of the known contaminants for identification and separation of these plastics.
The HDPE and LDPE recycled plastics have different impurities but in low concentrations (< 1 wt%). In tests using ATR-FTIR spectral data from recycled HDPE/LDPE blends containing 5 wt% of PP, PET, silica, and calcium carbonate, the predictive models based on iPLS-R showed prediction errors between 11 wt% and 19 wt% of LDPE in the whole LDPE content range (from 0 to 100 wt%). Only for samples containing silica, the iPLS-R predictive model exhibited RMSEP close to 9 wt% of LDPE. iPLS-R enables quantifying the composition of recycled HDPE/ LDPE blends containing contaminants (in 5 wt%) in the whole LDPE content range (0-100 wt%) a root mean square error below 7 wt%. This paper shows that the CAR/PLS-R approach does not guarantee effective classification and correct composition determination of HDPE and LDPE mixtures in plastic waste.

Electronic Supplementary Material
The online version of this article (doi:https://doi.org/10.1007/s10924-022-02396-3 contains supplementary material, which is available to authorized users. and automated protocols for classifying polyethylenes and determining the LDPE content in HDPE/LDPE mixtures using infrared measurements. However, the presence of 5 wt% of a contaminant, such as another polymer with different chemical composition or microparticles of oxides, restricts the capacity and generality of the PCA and PLS-R statistical models for sorting highly heterogeneous mixtures of plastic residues found in real plastic waste. Despite these