Statement of Novelty

The research novelty is based on the development of automatic recognition strategies based on laser induced fluorescence (LIF), as non-destructive and non-invasive technique, combined with machine learning, for the identification of black polymers from plastic packaging waste, in view of its potential application at recycling plant scale.

The promising results demonstrated as LIF technology has great potentialities for black polymers sorting at recycling plant scale, solving one of the critical issues currently existing in plastic waste sorting for the production of high-quality secondary raw materials, i.e., constituted by a single polymer. The proposed approach can allow to increase the number of recycled products, reducing waste sent to incinerators and landfills, in agreement with the principles of circular economy and the sustainable use of resources.

Introduction

Plastic materials are widely used in modern society and their demand is constantly growing. Consequently, the amount of plastic waste tends to increase over time, causing serious environmental problems and reaching unsustainable rates. In 2020, for example, more than 29 million tonnes of post-consumer plastic waste were collected in Europe [1]. A critical aspect of polymer waste management is related to plastic colors. Colorless or white plastic samples have the highest market value as recycled products, as they can be re-colored with any color [2]. Plastics of other colors, especially black ones, have a lower economic value than colorless or white plastic samples. On the other hand, black samples can only be recycled into black plastics [2]. The color of plastics is due to organic or inorganic colorants, that could affect the plastic recycling process depending on their characteristics and on processing strategies. In order to obtain recovered plastic flow streams easy to handle for the further recycled product manufacturing plastics of different colors should be separated during the recycling process [3]. The fulfilment of this goal supports the substitution of virgin polymer with secondary raw materials, thus, reducing the environmental impacts, such as greenhouse gas emissions and microplastics diffusion, and the consumption of non-renewable primary resources, according to circular economy principles [4].

Mechanical recycling is currently the most environmentally and economically sustainable process to perform plastics recycling. In plastic recycling, separation of the different polymers is a fundamental step to obtain high-quality secondary raw materials [5]. This separation step is more and more carried out by sensor-based sorting technologies, allowing to improve both quality and quantity of recovered plastics [6]. More in detail, automated sorting of polymers is usually carried out by optical sensors working in the Near Infrared and Short-Wave Infrared spectral ranges (NIR and SWIR: wavelengths from 1000 to 2500 nm) [7]. Many studies have demonstrated the possibility offered by NIR spectroscopy to classify different typologies of polymers from different waste sources [8,9,10]. Despite the good performance of this technique for the identification of different types of plastic waste, its efficiency is reduced due to the presence of black-colored plastics [11]. More in detail, the main limit of such technology is that classification of black plastics by polymer type cannot be achieved due to their strong light absorption, producing, as a consequence, low reflectance flat spectra in the NIR range [12]. Black plastic products represent about 15% of the household plastic waste stream, most of which are single-use packaging [13]. Most black plastics are colored by adding carbon black, a group of industrial coals created by the partial combustion of various hydrocarbons. Carbon black is cheap to produce and has high covering power, solvent resistance and ultraviolet stability, acting as a reinforcing filler and antioxidant [14]. The characteristics of carbon black make it suitable for a wide range of plastic products especially in outdoor applications or where strength, conductivity or thermal stability is required. Carbon black is also used for aesthetic reasons of color, such as in toys, jewels and food packaging [15]. Nevertheless, black plastic waste poses critical challenges in the recycling process and there is the need to find an effective technological solution for their identification and sorting based on polymer type at recycling plant scale [16].

In recent years, some new approaches have been developed for post-consumer black plastic waste sorting in the recycling industry. Rozenstein et al. (2017) demonstrated the possibility to identify several black plastics (i.e., PET, PP, PS, PLA, and PE) using a device working in the Mid Infrared (MIR) range [10]. The MIR region is less affected by plastics surface morphology and color (including black color) than the NIR range, as it is sufficiently distinct from the visible light region [17,18,19]. However, further studies in this field are needed to better understand the effectiveness of MIR technology for plastic sorting [18]. Küter et al. (2018) presented the development of a THz-domain-based device that is able to measure slight differences in broadband spectral signatures of black plastics. Different materials can be distinguished using this technique combined with a machine learning approach [19]. Shameem et al. (2017), showed an approach based on Laser-Induced Breakdown Spectroscopy (LIBS) and Raman spectroscopy combined with machine learning techniques to recognize black complex mixture of different plastics (i.e., PET, PE, PP and PS) [20]. Despite the potential of these techniques, i.e., THz, LIBS and Raman, for the identification of black plastics, they are still under development and can be quite expensive [21]. Indeed, THz does not seem mature enough for industrial application. LIBS is affected by the limited number of chemical elements that can be measured. Raman spectroscopy is much less used than infrared spectroscopy for plastic identification, due to the low intensity of its signal (much lower than Rayleigh scattering), and it is difficult to obtain a competitive online acquisition using such technology [22]. Other methods, such as Attenuated Total Reflectance—Fourier Transform Infrared spectroscopy (ATR-FTIR) combined with deep learning technique, allow the identification of plastics with an accuracy value close to 100% [23]. However, the measurement conditions and the amount of sample that can be analyzed do not allow to implement real-time or online sorting strategies. In this scenario, Laser Induced Fluorescence (LIF) can allow to perform a significant progress in plastic sorting specifically with reference to black plastics. LIF is based on the excitation of a molecule by laser radiation, i.e., when a UV photon hits a molecule, it passes from the fundamental level to an excited level and after a decay time (in the order of nanoseconds/microseconds), it returns to the fundamental state, emitting light in the UV, visible (VIS) and NIR. Since the decay time and the energy jump between the intermediate level and the fundamental level are characteristic of the molecule, it is possible to characterize them based on their emission spectrum and their characteristic decay time [24, 25]. Furthermore, several studies have shown the potential of LIF technique for the recognition of organic materials and in particular plastics using punctual ad imaging devices [26,27,28,29,30,31,32,33]. However, despite the potential of this technique, there are no studies on LIF application for the recognition of black plastics. For this reason, in this work, an innovative strategy based on the application of LIF combined with machine learning techniques, was developed and tested to identify and classify black plastics from packaging waste. In fact, LIF could represent an attractive solution for classification and on-line sorting of black plastics in recycling plants, being a non-destructive and non-invasive technique and, to the best of our knowledge, the possibility of applying LIF for black plastic classification in recycling plants has never been proposed. The study is specifically focused on the classification by LIF of black plastics from packaging waste, selected among the most diffused in the market,, such as HDPE (high-density polyethylene), PP (polypropylene), EPS (expanded polystyrene) and PS (polystyrene) [14, 15]. HDPE and PP were considered as they are among the main polymers used for the production of bottles and caps of detergents and cleaning products. In addition, EPS and PS have been considered as they are widely used in food packaging [14]. In agreement with the literature, the presence of black plastics reduces the efficiency of NIR-based sorting as the type of polymer cannot be identified [11, 34]. Different combinations of preprocessing sets and Principal Component Analysis (PCA) were applied to improve and explore the collected spectral characteristics. Finally, a hierarchical modelling approach, based on Partial Least Square-Discriminant Analysis (PLS-DA) was used to obtain efficient identification of different black plastic waste starting from their collected fluorescence spectra.

Materials and Methods

Investigated Samples

Twelve black plastic packaging samples were selected among those most commonly found in household waste, such as plastic bottles, food containers and bottle caps, constituted by EPS, PS, PE and PP (Table 1).

Table 1 Description of the investigated black plastic waste samples from household waste
Full size table

Attenuated Total Reflection—Fourier Transform Infrared spectroscopy

Attenuated Total Reflection—Fourier Transform Infrared spectroscopy (ATR-FTIR) working in the range of 3800–550 cm−1, was utilized to control the correctness of the polymer identification codes, and to validate the classification results obtained by LIF. To reach this goal a Shimadzu spectrophotometer (IRAffinity-1S) was utilized. For every item listed in Table 1, a small sample 0.7 cm × 0.7 cm was obtained and analysed, setting 45 scan per measure. For every sample, both internal and external face were tested, in order to exclude the presence of thin coating layers able to misrepresent the declared material during LIF measurements. The obtained ATR-FTIR spectra were automatically processed and compared with the polymer database make available by Shimadsu.

Laser-Induced Fluorescence

LIF is a non-invasive and non-destructive molecular spectroscopy technique, based on the study of the effects of ultraviolet (UV) laser radiation interaction with materials [35]. Samples are irradiated by a laser beam, an optical system collects induced fluorescence emissions and send them to a detector. For wavelength-resolved fluorescence spectroscopy, the back-scattered radiation is typically dispersed and detected by a spectrometer or by a monochromator coupled with photomultipliers or with an intensified charge-coupled device (ICCD) camera [36].

LIF data acquisition was performed at the Diagnostics and Metrology Laboratory (DIM) of Frascati ENEA (Italian National Agency for New Technologies, Energy and Sustainable Economic Development) Research Centre, using a punctual scanning LIF system called Lifart developed at the DIM Laboratory [29]. The used instrument works remotely up to 10 m distance from the target and is equipped with a diode pumped solid state laser emitting in the UV at 266 nm. In order to transmit the exciting radiation and to receive the fluorescence signals from the studied target, a coaxial optical design is used. A mirror with two rotating servo controls operating at high accuracy allowed to obtain the point scanning and the scan of a selected area. Practical aperture of the scanning is limited by the mounting assembly to approximately ± 45° in both the horizontal and vertical axes. The fluorescence radiation is focused on the entrance of a fiber optic linked to a compact QE-Pro spectrometer from Ocean Optics, allowing acquisitions in the range of 250–950 nm with a 2.5 nm bandwidth.

On each plastic sample, two measurement runs of 4 × 4 pixels on two different areas of about 4 cm2 were carried out, in order to verify and take into account any non-homogeneity of the sample. In this way, for every item, 32 fluorescence spectra were obtained. In total 384 fluorescence measurements were acquired and divided into a calibration (256 LIF spectra) and validation dataset (128 LIF spectra) to build and evaluate the classification performances of the Hi-PLS-DA model. For the processing of data, only the channels corresponding to the spectral range 270—750 nm were considered: before excitation wavelength no signal can be present, after 750 nm the fluorescence emission ends and only interferent signals are present. However, after 550 nm no signal was found. For this reason, in order to reduce the number of wavelengths to be processed, the selected range for this case study was from 270 to 550 nm.

Data Processing

Data processing was performed at the Raw Materials Laboratory (RawMaLab) of the Department of Chemical Engineering, Materials & Environment of Sapienza University of Rome, using the PLS_toolbox (ver. 9.0 Eigenvector Research, Inc.) running in the Matlab environment (version R2021b, The Mathworks, Inc.).

Different preprocessing algorithms were preliminary applied in order to highlight differences among the spectra of the different polymers. More in detail, all data were preprocessed using normalization algorithm in order to reduce the effect of scattering [37] and then applying first derivative algorithm to emphasize the spectral characteristic and to reduce the additive effects [30, 31]. Finally, all data are mean centred for PCA [38].

PCA was chosen to perform exploratory analysis about the spectral variability of data. PCA allows the decomposition of pre-processed spectral data into linear combinations of the original spectral data, i.e., Principal Components (PCs), collecting the spectral variations in reduced set of factors. The first PCs were used to analyse the common characteristics of samples and their grouping, as the samples characterized by similar spectral signatures tend to aggregate in the score plot of the first two or three components [32, 39].

Finally, three different PLS-DA models (Rule 1, 2 and 3) were assembled in a hierarchical structure, constituting a single classification model to optimize the classification results. PLS-DA is a supervised technique that needs a prior knowledge of the data [40], which is based on the reduction of dimensionality through Partial Least Squares Regression (PLS-R) with discriminating characteristics [41, 42].

Hierarchical modelling can be a powerful solution when multi-classes classification problems have to be faced [43]. The hierarchical structure is usually shown through a dendrogram, allowing to identify the existing relationships between data [8]. Through hierarchical modelling it is thus possible to solve the nodes/rules of the dendrogram, starting from the simplest steps up to the most complex, obtaining a single prediction result. Venetian blind with a number of data splits equal to 10 was used as cross-validation method, in order to evaluate the complexity of the model and to select the appropriate number of latent variables (LVs).

In Fig. 1, the dendrogram developed for the classification of the 4 different polymers by 3 different rules is shown. In Rule 1, two classes were identified: (1) EPS/PS and (2) PP/HDPE; in Rule 2, the spectra of EPS/PS were processed, utilizing their original raw spectral signatures, to identify the two classes: (1) EPS and (2) PS; finally, applying Rule 3, the raw spectra of PP/HDPE were treated to separate PP from HDPE.

Fig. 1
figure 1

Hi-PLS-DA structure developed for the classification of black polymers based on LIF spectra

Full size image

A pixel-based approach was used for the evaluation of the classification model. In a pixel-based logic, the performances of the Hi-PLSDA were evaluated in terms of false color classified data for both the training and the validation sets and in terms of statistical parameters: sensitivity and specificity in calibration (Cal), cross-validation (CV) and prediction (Pred) [41, 44]. Sensitivity expresses the model ability to correctly recognize samples belonging to the considered class and is defined by Eq. (1), in which TP represents the total number of True Positive and FN the total number of False Negative; Specificity describes the model ability to correctly reject samples belonging to all the other classes and is defined by Eq. (2), in which TN represents the total number of True Negative and FP the total number of False Positive. Both Sensitivity and Specificity can assume values between 0 and 1, the latter being the ideal value for a prediction model.

$$\mathrm{Sensitivity }=\frac{\mathrm{True\,Positive}}{(\mathrm{True\,Positive }+\mathrm{ False\,Negative})}$$
(1)
$$\mathrm{Specificity}\hspace{0.17em}=\hspace{0.17em}\frac{\mathrm{True\,Negative}}{(\mathrm{True\,Negative }+\mathrm{False\,Positive})}$$
(2)

Experimental Results and Discussion

FT-IR Spectroscopy Results

The FT-IR/ATR spectra of the investigated samples were shown in Fig. 2, confirming the polymer identification codes found on samples. In more detail, the IR spectrum in red, showed the characteristics of EPS, such as bands at 693 and 723 cm−1 (aromatic CH out-of-plane bend bands), 1452 cm−1 (CH2 bend), 2854 cm−1 (C–H stretch) and 3026 cm−1 (aromatic C-H stretch) [45]. The IR spectrum in light blue showed typical bands of PS, such as 696 and 758 cm−1 (aromatic CH out-of-plane bend), 1454 cm−1 (CH2 bend), 1604 cm−1 (aromatic ring stretch), 2848 cm−1 (C–H stretch) and 3028 cm−1 (aromatic C–H stretch) [46]. The IR spectrum in green showed bands typical of PP, such as 2950, 2915 and 2838 cm−1 (C–H stretch), 1454 cm−1 (CH2 bend), and 1377 cm−1 (CH3 bend) [46, 47]. The IR spectrum in blue showed bands attributable to HDPE, such as 2915 and 2845 cm−1 (C–H stretch), 1472 and 1462 cm−1 (CH2 bend), 731 and 718 cm−1 (CH2 rock) [46, 48].

Fig. 2
figure 2

ATR-FTIR spectra of the investigated black plastic samples, showing the absorption bands typical of EPS, PS, PP, and HDPE

Full size image

LIF Results

Average Raw and Preprocessed LIF Spectra of the Different Black Polymers

The average raw fluorescence spectra of the 4 plastic products (i.e., EPS, PS, PP and HDPE) are reported in Fig. 3. The raw spectra were preliminarily analyzed to detect and compare their characteristics. The UV range provides important information about polymers, as the UV spectral fingerprint can be related to the double bonds that absorb ultraviolet light [49]. As shown in Fig. 3a, b, EPS and PS showed a distinct fluorescence emission around 320 nm. Inside the emission peak of EPS and PS, two superimposed bands were observed at 320 and 325 nm, correlated to the spectral features of the carbonyl groups. Furthermore, the fluorescence intensity was uniform for all EPS and PS samples.

Fig. 3
figure 3

Average raw (a, b, c, d) and pre-processed (e) fluorescence spectra of the calibration dataset; average raw (f, g, h, i) and pre-processed (l) fluorescence spectra of the validation dataset of the studied black plastic samples

Full size image

HDPE and PP samples showed two distinct fluorescence emissions around 355 and 440 nm. The first sharp emission at 355 nm can be attributed to the ketone carbonyl, while the fluorescence emission at 440 nm is due to a charge-transfer complex of HDPE and PP with oxygen [28]. The main spectral differences between HDPE and PP were detectable between 280 and 350 nm. The fluorescence intensity showed a remarkable variability for all PP and HDPE samples.

Principal Component Analysis and Hierarchical Model Building

PCA was used for the definition of the three rules defined for the Hi-PLS-DA model. The three PCA score plots and the corresponding loadings plots, related to the 4 classes of black polymers (EPS, PS, PP and HDPE), are shown in Fig. 4 with reference to rule 1 (Fig. 4a, b), rule 2 (Fig. 4c, d) and rule 3 (Fig. 4e, f) of the Hi-PLS-DA model. The description of the three rules is reported, in the following:

Fig. 4
figure 4

PCA score (a) and loadings (b) plots of PC1 and PC2 related to rule 1 (i.e., EPS, PS, PP and HDPE). PCA score (c) and loadings (d) plots of PC1 and PC2 related to rule 2 (i.e., EPS and PS). PCA plot (e) and loadings (f) plots of PC1 and PC3 related to rule 3 (i.e., PP and HDPE)

Full size image
Rule 1

PCA score and loadings plots of the LIF data for Rule 1 are reported in Figs. 4a, b. PCA score plot showed as PC1 and PC2 captured most of the variance (96.93%). In detail, two main clouds related to different plastics are shown: (1) EPS and PS (I and IV quadrants); (2) PP and HDPE (II and III quadrants).

Positive values of PC1 for EPS and PS allow their separation from PP and HDPE, and PC2 highlights sample variability, but does not allow a further separation between classes. Consequently, in order to obtain an optimal classification of these black plastics, 2 classes were defined for Rule 1 of the Hierarchical PLS-DA model. The loadings plot of the first 2 PCs showed the main spectral regions contributing to the identified differences. The highest contribution to the definition of PC1 was given by positive signals around 310 nm (carbonyl group) and negative ones at 350 nm (ketone carbonyl). The highest contribution to the definition of PC2 was mostly influenced by positive signals before 320 nm and negative ones at 310 nm and 425 nm.

Rule 2

PCA score and loadings plots of the LIF data for rule 2 are reported in Fig. 4c, d. PCA score plot showed as PC1 and PC2 captured most of the variance (99.89%). In detail, PC1-PC2 score plot highlighted the presence of cluster separation between EPS and PS data. The positive values of PC1 allowed the separation of EPS (II and III quadrants) from PS (mainly concentrated in the I and IV quadrants). PC1 highlighted the difference in variability between PP and HDPE classes. As a consequence, to obtain an optimal classification of these polymers, 2 classes (end nodes) were defined for Rule 3 of the Hierarchical PLS-DA model. The loadings plot of the first 2 PCs indicated the main spectral regions contributing to the identified differences. The highest contribution to the definition of PC1 was given by positive signals around 290 nm and negative ones at 350 nm. The highest contribution to the definition of PC2 was shown by positive signals around 360 nm and negative ones at 280 nm. Both loadings showed how the influence of the spectral features of the carbonyl groups allowed the separation between EPS and PS, highlighting further variability within each class.

Rule 3

PCA score and loadings plots of the LIF data for Rule 3 were reported in Fig. 4d, e. PCA score plot showed as PC1 and PC3 captured most of the variance (80.63%). In detail, PC1-PC3 score plot highlighted the presence of cluster separation between PP and HDPE data. In detail, PC1 highlighted how HDPE samples showed high variability compared to PP samples which were mainly concentrated in the central portion of the III and IV quadrants. In fact, the HDPE samples were distributed throughout the I and II quadrants. The positive values of PC3 for HDPE allowed its separation from PP. Consequently, in order to obtain an optimal classification of such polymers, 2 classes (end nodes) were defined for Rule 2 of the Hierarchical PLS-DA model. The loadings plot of the first 2 PCs indicated the main spectral regions contributing to the identified differences. The highest contribution to the definition of PC1 derived from positive signals around 288 nm and negative ones at 316 nm. The highest contribution to the definition of PC3 was given by positive values around 280, 330 and 510 nm and negative ones at 295 and 370 nm. All loadings showed how the influence of the ketone carbonyl combined with charge-transfer complex of HDPE and PP with oxygen spectral signal allowed the separation between PP and HDPE, highlighting a variability of HDPE samples.

Application of the Hierarchical PLS-DA Classifier

Starting from the information displayed and explored through PCA, a four-class Hi-PLS-DA model was defined.

The PLS-DA model used for Rule 1, that is the separation of polymers in 2 classes, EPS/PS and PP/HDPE, allowed to capture 99.65% of the variance with seven latent variables. The PLS-DA model for Rule 2, that is the separation between EPS and PS, allowed to capture 99.43% of the variance with five latent variables. The PLS-DA model for Rule 3, that is the separation between PP and HDPE, allowed to capture 96.07% of the variance with five latent variables.

The results of Hi-PLS-DA applied to the 128 LIF spectra constituting the validation set of the four classes of black polymers (32 spectra each) are reported in Fig. 5, in terms of predicted class value for each measured spectrum. From Fig. 5, it can be noticed that all the four different black classes of polymers were correctly identified and classified. The correctness of the classification model was confirmed by the Sensitivity and Specificity values in calibration, cross-validation, and prediction, with performance values equal to 1 except for the Sensitivity of EPS and Specificity ofPS in cross-validation, which showed a value equal to 0.984.

Fig. 5
figure 5

Predicted class value of EPS, PS, PP, and HDPE classes for each measured LIF spectrum of the validation set

Full size image

Conclusions

In this work a novel strategy for the classification of post-consumer black polymers by LIF technology coupled with machine learning was presented, in order to face this critical issue of plastic recycling industry, contributing to increase quality and quantity of recycled plastics. More in detail, after LIF spectra acquisition of the most representative black polymers from packaging waste, a Hi-PLS-DA classification approach was developed and applied. The results were evaluated in terms of efficiency and reliability, with the specific aim to design and set up multivariate processing/recognition strategies maximizing black polymers identification.

Combining LIF-based detection technique with chemometric logics (i.e., Hi-PLS-DA), several goals were achieved, that is:

  • acquisition of LIF spectral information, in the UV–VIS range, allowing to perform black plastics discrimination, through a pre-processing strategy able to highlight polymer spectral differences (i.e., training set definition);

  • definition of hierarchical modelling, allowing the fast identification of different black polymers at the same time;

  • identification of black polymers, i.e., EPS, PS PP and HDPE, an issue not easy to reach with conventional NIR sensor-based sorting technologies applied at recycling plant scale.

Based on the promising results obtained in this work on the potential of LIF technique for the classification of black polymers, future studies will be carried out to increase the typologies of black plastics (such as LDPE), including biodegradable polymers. Finally, it is important to underline that the proposed methodological approach could be profitably applied, in the near future, not only for plastic waste quality control, but also for black polymer sorting at recycling plant scale, thanks to the growing of both LIF imaging technologies and computing power.