Michael
Breuckmann
*a,
Georg
Wacker
abc,
Stephanie
Hanning
a,
Matthias
Otto
b and
Martin
Kreyenschmidt
a
aUniversity of Applied Sciences Münster, Department of Chemical Engineering, Laboratory of Instrumental Analytics, Stegerwaldstraße 39, 48565 Steinfurt, Germany. E-mail: michael.breuckmann@fh-muenster.de; hanning@fh-muenster.de; martin.kreyenschmidt@fh-muenster.de; Tel: +49-2551-962220
bTechnical University Bergakademie Freiberg, Faculty of Chemistry and Physics, Institute of Analytical Chemistry, Leipziger Straße 29, 09599 Freiberg, Germany. E-mail: matthias.otto@chemie.tu-freiberg.de
cVDM Metals GmbH, Plettenberger Str. 2, 58791 Werdohl, Germany. E-mail: georg.wacker@vdm-metals.com
First published on 3rd March 2022
A new approach to determine the elements carbon, hydrogen, nitrogen and oxygen (CHNO) in polymers by wavelength-dispersive X-ray fluorescence analysis (WDXRF) in combination with partial least squares (PLS) regression was explored. The quantification of CHNO was achieved by using the Rayleigh and Compton scattering spectra of an Rh X-ray tube from 84 different polymers. Concealed differences of the corresponding scattering spectra could be utilized to quantify CHNO in a multivariate manner. It was shown that the developed model was capable of determining these commonly non-measurable matrix elements in polymers using WDXRF. Furthermore, the influence of spectral resolution, which is given by the collimator and the crystal, on the prediction of CHNO was explored in this study. It was found that minimal spectral resolution led to the most accurate CHNO predictions. Information about matrix composition could be used to improve so-called semi-quantitative XRF methods based on fundamental parameters (FP) for the analysis of plastics, soil or other samples with high organic content.
Two approaches are employed for quantification in XRF: the empirical methods and the so-called semi-quantitative methods. The empirical method relates XRF signals to a known sample composition with certified reference materials (CRM). The so-called semi-quantitative methods are based on a theoretical approach to quantify elements using fundamental parameters (FP).13 In real-life applications, an enormous variety of polymer types with varying filler and additive combinations are employed to tailor the properties according to the application's demand. Unfortunately, so far, only a limited number of polymeric CRM are available, which are based on polyethylene (PE),14 unsaturated polyester resin (UP),15,16 polycarbonate (PC),17 polyvinyl chloride (PVC),18 polypropylene (PP)18 and acrylonitrile butadiene styrene (ABS).19,20 Furthermore, these CRM cover only limited element compositions and are restricted to small ranges of element concentrations. To perform quantification based on empirical methods, countless calibration standards based on suitable polymers and element compositions would be required.
A more applicable procedure to quantify elements in polymers is based on algorithms operating with FP (absorption coefficient, emission coefficient, etc.) of the elements.21,22 This approach models the interactions between elements and allows for the calculation of all element concentrations in a given sample. A prerequisite for obtaining reliable results by this theoretical quantification procedure is that all elements present in a sample have to be detectable by the XRF measurement.13
Most polymers contain the so-called low Z elements carbon, hydrogen, nitrogen or oxygen (CHNO) in total amounts greater than 95 wt%. These elements are often undetectable or insufficiently detectable by XRF. Subsequently, the quantification of elements in a polymeric matrix by employing FP methods often leads to inaccurate quantification. The common approach to overcome this problem is to add an assumed polymeric matrix for the FP method. When no further information concerning the material is available, CH2 (PE) is assumed as a sum parameter for plastic materials in software programs. When no information on CHNO matrix composition is available, the FP method cannot be applied to all polymers due to varying concentrations of CHNO within the polymers and additives (fillers, catalyst residuals, light stabilizers, heat stabilizers, UV stabilizers, plasticizers, antioxidants, etc.). Inaccurate matrix information may lead to great deviations for elements of higher atomic numbers, e.g. Ca, Ti, Fe or Zn, due to matrix effects. Several procedures were developed to determine the elemental composition of the polymers by considering low Z elements via the coherent scattering signal, incoherent scattering signals and ratios thereof. However, these approaches were not eligible to quantify the correct composition of the polymer matrix.23–27 Therefore, it is mandatory to identify the exact low Z element composition (CHNO) of the organic matrix to reliably quantify the additional elements present in plastic materials via FP methods.
Aidene et al. used energy dispersive X-ray fluorescence (EDXRF) with X-ray tube excitation in combination with partial least squares (PLS) modeling to determine the hydrogen, carbon and oxygen (CHO) concentrations of various thermoplastic samples.28 In another study, the authors employed a monochromatic excitation to calculate the carbon and hydrogen concentrations.29 It was found that the precision was similar to polychromatic excitation.
The purpose of this study was to maximize the precision of CHNO quantification and assess the dependence on spectral resolution (i.e. crystal-collimator) used in WDXRF. The Rayleigh and Compton scattering spectra of different polymers were studied employing PLS regression to quantify CHNO concentrations utilizing WDXRF.
The incoherent or inelastic scattering of X-rays, known as Compton scattering, is also induced by an interaction of an incoming X-ray photon with an electron. In contrast to the Rayleigh scattering, the impact of the incoming photon leads to the removal of a loosely bound electron out of its orbital by transferring a part of its energy to this electron. This energy transfer induces a scattering of the collided photon and results in an increased wavelength.32–35
Compton scattering is inversely proportional to the mean atomic number of a given sample. Therefore, a sample composed of low Z elements yields a high intensity of Compton scattering. Thus, the shape or pattern of scattering spectra must contain information regarding the composition of elements with low atomic numbers.
![]() | (1) |
The interaction of X-ray photons with low Z elements, e.g. CHNO, mostly causes scattering of X-ray photons. The effect is increased at higher energies of involved X-ray photons, e.g. K-series of X-ray tube anodes like Rh (Fig. 1), Pd or W. This consideration implies that the composition of the elements CHNO is related to the specific X-ray scattering pattern of the respective polymer. Therefore, it may be possible to determine the composition of an organic polymer quantitatively. In this study, the material-specific scattering spectra are analyzed to quantify CHNO by exploiting concentration and X-ray energy-dependent attenuation of X-rays.
![]() | ||
Fig. 1 Rh-Kα X-ray tube scattering spectra (coherent and incoherent) of several materials of same sample masses. |
Thermoplastics | |
---|---|
10 samples of PA (polyamide) | 2 samples of PP (polypropylene) |
1 sample of POM (polyoxymethylene) | 1 sample of PR (phenolic resin) |
2 samples of PC (polycarbonate) | 5 samples of HD/LD-PE (high/low-density polyethylene) |
3 samples of ABS (acrylonitrile butadiene styrene) | 1 sample of HI-PS (high impact polystyrene) |
1 sample of PMMA (poly-(methyl methacrylate)) | 1 sample of TPU (thermoplastic polyurethane) |
1 sample of PET (polyethylene terephthalate) | 1 sample of PBT (polybutylene terephthalate) |
Thermoset plastics | |
---|---|
6 samples of EP (epoxy resin) | |
43 samples of UP (unsaturated polyester resin) | |
6 samples of PUR (polyurethane) |
All thermoplastic samples, including the thermoplastic polyurethane sample, were prepared by molding the polymers using an automatic mounting press equipped with a 40 mm mold cylinder (SimpliMet 3000, Buehler, Düsseldorf, Germany). The thermoset plastics (UP, EP) were made by mixing the acid and the polyolic component systems beforehand. Then, each mixture was cast in a customized aluminum mold equipped with two cavities. The samples were hardened in these cavities at standard room temperature (20°C). Photographs of the mold can be found in ESI (Fig. S1).† Then, a grinder–polisher (Alpha, Buehler, Düsseldorf, Germany) was applied to adjust the resulting sample thickness and mass. The produced plastic samples had masses of 5.0 ± 0.1 g.
XRF spectrometer | S4 Pioneer (Bruker AXS) | |
X-ray tube target (anode) | Rhodium (Rh) | |
Voltage/current | 60 kV/50 mA | |
Energy range | 17–24 keV | |
Crystal/collimator | Low resolution | LiF (200)/0.46° |
Medium resolution | LiF (220)/0.23° | |
High resolution | LiF (420)/0.12° | |
Acquisition time | Low resolution | 80.6 s |
Medium resolution | 176 s | |
High resolution | 241 s | |
Measuring spot (mask) | 34 mm | |
Measuring mode | Vacuum |
For the investigation of the scattering spectra, a PLS algorithm was applied (Statistics Toolbox of MATLAB®, R2009a, The MathWorks Inc.). The base for multivariate regression modeling is the multiple linear regression approach (eqn (2)).
Y = X·B + E | (2) |
The matrix Y is composed of CHNO concentrations data, and the matrix X corresponds to the intensities of the recorded WDXRF scattering tube spectra. The regression coefficients are given by matrix B. These regression coefficients can be used to interpret the created regression model, because they are associated with corresponding photon energy, respectively. The regression residuals are given by the matrix E. In PLS regression, the matrices X and Y are decomposed into smaller submatrices for scores and loadings. For this step, the data are transformed on calculated PLS components. These PLS components can be seen as a new coordinate system. Concerning the scores, the new axes capture maximum variance in the spectral data X and the CHNO concentrations Y. Furthermore, the new axes maximize the covariance and, thus, the correlation between X and Y. The determination of the optimal PLS component number is crucial in preventing the overfitting of the training data. Thus, deviations between predictions of validation data and corresponding reference data can be minimized.37 Detailed descriptions of the PLS method can be found in the literature.38–41
Pre-processing of data records may have a significant influence on the regression's performance.7 The PLS algorithm, used in MATLAB®, applied mean-centering of each feature, i.e. energy, prior to PLS decomposition. Further pre-processing, e.g. z-scaling,38 was not necessary for this study. When z-scaling was applied, the R2 metrics for testing data were less than or equal compared to unscaled data.
A single PLS model was created to quantify CHNO concentrations instead of utilizing individual PLS models for C, H, N and O concentrations. This single PLS model procedure is advantageous due to improved model interpretability, e.g. concerning regression coefficients, and especially when target element concentrations are correlated.40
The calibration model quality on training data and predictions of hold-out testing samples is given by the root mean square error (RMSE, eqn (3)). The optimal number of PLS components was determined with the aid of RMSE and leave-one-out (LOO) cross-validation (CV). A schematic representation of the CV procedure can be found in ESI (Fig. S2).† The testing set was not used for CV and was completely hold-out during training of the model. Therefore, it was possible to see how well the cross-validated PLS model could generalize the data by prediction of the testing set.
![]() | (3) |
In RMSE, elemental analysis concentrations are represented by yi and predicted concentrations via the PLS regression model are represented by ŷi. The number of considered samples equals N. The arithmetic average RMSE (avg. RMSE) gives a single metric to assess the model performance for a given spectral resolution of CHNO concentration predictions (eqn (4)). Additionally, a weighted average RMSE (wavg. RMSE) is calculated (eqn (5)).
![]() | (4) |
![]() | (5) |
In the wavg. RMSE, the individual RMSEi for CHNO quantifications are weighted concerning the corresponding average mass fraction ȳi in the training set or the testing set, respectively. Therefore, in wavg. RMSE the individual RMSE, especially for N concentrations, are down-weighted due to comparatively low N concentrations in all 84 plastic samples. The wavg. RMSE is a more important metric compared to avg. RMSE when assessing the CHNO concentration prediction quality. For a correct ascertainment of μm (eqn (1)), high accuracy in the determination of the CHNO elements is more important for elements with high concentrations in the polymer than for elements with very low concentrations.
![]() | ||
Fig. 2 XRF scattering spectra of 84 polymer samples, (a) low spectral resolution, (b) medium spectral resolution and (c) high spectral resolution, random colors to show minor intensity differences between spectra (see Table 2 for spectral resolutions). |
![]() | ||
Fig. 3 Concentrations of CHNO in the 84 prepared calibration samples (CHN by elemental analysis, O balanced to 100 wt%). |
To obtain an optimal PLS model, CV was applied for the training set data. After CV, the final model was fitted with the optimal number of PLS components. The model predicted the CHNO concentration for training and testing data. The predictions were compared with reference values in terms of avg. RMSE and wavg. RMSE.
The CV results (Fig. 4) show that for increasing spectral resolution, the RMSE for training data decreased for all numbers of PLS components, respectively. This supports the hypothesis that a better spectral separation of coherent and incoherent scattering signals may lead to improved performance in the prediction of CHNO concentrations. For the validation data, this tendency was not observed, but the inverted tendency. For reducing spectral resolution, the RMSE (CV) was decreasing. This implies that the generalization capability of the corresponding PLS model was improved for lower spectral resolutions. It should be noted that this was the result of CV. In CV, samples are used for training the PLS model as well as for validating the PLS model to optimize for the number of PLS components. Thus, information from validation samples was also used for training in other CV folds (Fig. S2).† An improved approach, which avoids this information reuse, was based on evaluating predicted CHNO concentrations from the hold-out testing data by the final PLS models (Table 3). The evaluation of CHNO predictions from testing data enabled an improved analysis because the model did not adapt to testing data.
Spectral resolution | Element | Training set (wt%) | Testing set (wt%) | ||||
---|---|---|---|---|---|---|---|
RMSE | avg. RMSE | wavg. RMSE | RMSE | avg. RMSE | wavg. RMSE | ||
Low | C | 1.8 | 1.5 | 1.6 | 1.8 | 1.5 | 1.6 |
H | 0.22 | 0.3 | |||||
N | 2.4 | 2.4 | |||||
O | 1.7 | 1.3 | |||||
Medium | C | 1.2 | 1.0 | 1.1 | 2.0 | 1.7 | 1.9 |
H | 0.17 | 0.4 | |||||
N | 1.4 | 2.5 | |||||
O | 1.2 | 2.0 | |||||
High | C | 1.3 | 1.7 | 1.4 | 2.0 | 1.8 | 1.9 |
H | 0.25 | 0.54 | |||||
N | 3.2 | 2.6 | |||||
O | 2.2 | 2.0 |
The regression coefficients of the trained model enabled the interpretation of energy regions that are significant for extracting the respective CHNO concentration (Fig. 5, spectra in grey). Energies associated with high absolute values of regression coefficients substantially influence the particular target element concentration. The algebraic sign of regression coefficients indicates the direction of influence, i.e. positive coefficients are positively related with target concentrations and vice versa. It was shown that the quantification of C concentration was mainly influenced by the inflection points of the incoherent Rh-Kα signal, i.e. higher number of recorded counts affected an increase in C concentration. In contrast, the energy region around the incoherent scattering peak maximum 18.75–19.5 keV showed the inverted effect; thus, a higher number of recorded counts affected a decrease in C concentration. The magnitude of the decreasing effect was smaller due to lower absolute regression coefficients. In contrast to the quantification of C concentration, other energy regions influenced the quantification of N. The PLS predictions of N concentrations were positively correlated with the incoherent Rh-Kα scattering peak and coherent Rh-Kα scattering peak. This was given by high regression coefficients in these energy regions. On the other hand, the incoherent Rh-Kβ peak and coherent Rh-Kβ peak were negatively correlated with the PLS predictions of N concentrations.
![]() | ||
Fig. 5 PLS regression coefficients and XRF spectra for low spectral resolution (grey lines, right axis). This figure reveals energy regions that are used to extract corresponding CHNO concentration. |
![]() | ||
Fig. 6 Recovery plot for CHNO quantification by PLS models, results for training and testing data are shown, diagonals plotted to guide the reader's eye (see Table 2 for spectral resolutions). |
The dependence of CHNO quantifications on spectral resolution was investigated. The avg. RMSE and wavg. RMSE of predicted CHNO concentrations and elemental analysis were calculated for training and testing data to assess the PLS models' prediction capabilities (Table 3). It was found that avg. RMSE and wavg. RMSE for training data was less than for testing data concerning high spectral resolution and medium spectral resolution. For low spectral resolution avg. RMSE and wavg. RMSE were about the same. It can be concluded that for high and medium spectral resolution the RMSE for training data is overly optimistic because the RMSE for training data was lower compared to those of testing data. For low spectral resolution, there was no difference in the evaluation of training or testing data.
The higher spectral resolution, i.e. improved separation of coherent scattering signals and incoherent scattering signals, did not improve CHNO predictions. Besides the spectral resolution, other essential factors describe the quality of spectra for CHNO predictions, e.g. the total counts recorded. The counting statistics are more precise when more counts are recorded.30 The emission and detection of X-ray photons can be described by a Poisson distribution.30,42 Therefore, the standard deviation of detected photons due to statistical fluctuation in spectra can be estimated by the square root of the number of detected photons, respectively. To compare the 3 spectral resolutions in terms of counting statistics, total signal intensity was calculated by summation over the energies and all samples for each spectral resolution. At low spectral resolution, the total signal intensity of scattering spectra for all investigated samples was 2.08 × 109 counts. For medium spectral resolution, the total signal intensity was 1.64 × 109 counts and for high spectral resolution signal intensity was 0.48 × 109 counts. To compare the counting statistics for spectral resolutions, the square roots and ratios thereof were formed. The ratio of roots was 2.1 to 1.8 to 1 in the order of low, medium and high spectral resolution. Additionally, the ratio of wavg. RMSE for CHNO concentration predictions and elemental analysis of the testing set was 0.83 to 0.97 to 1, also in the order of low, medium and high spectral resolution. As a result, the lowest spectral resolution was best suited for predictions of testing data due to the increased precision in counting statistics and reduced wavg. RMSE for testing set data. Thus, the lowest spectral resolution could generalize the data best and had the best performance in the prediction of CHNO concentrations from WDXRF scattering spectra.
Various combinations of CHNO concentrations may result in similar mass absorption coefficients μm. However, the presented PLS regression model can resolve the varying concentrations. The arithmetic mean μm in the Rh-Kα scattering region (17–24 keV) for all samples were calculated based on only the CHNO compositions to illustrate this. The total cross sections (Compton and Rayleigh cross-sections from Elam et al.43 and photoionization cross sections from Kissel44) were calculated using xraylib45 4.0.0. The mean μm, reference CHNO concentrations and predicted CHNO concentrations for the low spectral resolution were plotted. An interactive plot is given in the ESI.†
In this study, only undoped polymers were investigated. Elements with atomic numbers greater than 8 have a considerable influence on the scattering spectra. In plastics, these elements are commonly applied, e.g. as components in additives to manufacture certain material properties. To incorporate this influence on the scattering spectra, a further study was carried out.36 In this further study, additional polymers containing varying amounts of elements in a broad range of atomic numbers are considered for PLS regression modeling. The CHNO quantification is also possible in doped plastics when the doped plastics are also considered in the training data. Besides the CHNO quantification based on WDXRF spectra, the proposed procedure is also applicable for EDXRF spectra (including a micro XRF device and a handheld XRF device).36
In FP-based quantification procedures, the quantification of CHNO could be employed to analyze low Z material, e.g. plastics or biomass. The matrix composition could directly be used for semi-quantitative determination of higher Z elements, e.g. Ti, Fe, Zn, Br, Pb or Cd, from the same scan measurement.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: 10.1039/d1ja00412c |
This journal is © The Royal Society of Chemistry 2022 |