High-performance identification of insulating materials by using generalized spectrum in laser-induced breakdown spectroscopy

Junfei Nie^a, Furong Chen^a, Ting Luo^a, Jinke Chen^a, Jiapei Cao^a, Qiang Huang^a, Deng Zhang*^b and Zhenlin Hu*^c
aSchool of Electrical Engineering, Shaoyang University, Shaoyang, Hunan 422000, China. E-mail: njf@hnsyu.edu.cn
^bChina School of Computer and Electronic Information, Nanjing Normal University, Nanjing, Jiangsu 210023, China. E-mail: zhangd@nnu.edu.cn
^cState Key Laboratory of Ultra-Intense Laser Science and Technology, Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, Shanghai 201800, China. E-mail: huzhenlin@siom.ac.cn

First published on 21st August 2025

---

---

1. Introduction

Insulating materials play an indispensable foundational role in the fields of power, electronics, and energy. Their primary functions include isolating electrical currents, preventing leakage, and avoiding short circuits, thereby ensuring the safe operation and stability of power systems and equipment.¹ The performance of insulating materials not only affects the reliability and efficiency of devices, but also plays a decisive role in determining the operational lifespan of the entire system. With the rapid advancement of modern technologies and industrial applications, the use cases for insulating materials have become increasingly diverse, spanning high-voltage power transmission, electronic components, power generation equipment, and new energy systems.² The performance requirements of insulating materials often vary depending on the application scenarios, which further drives the diversification and specialization of insulating materials. For example, resins are widely used in electronic components and composite insulating structures due to their lightweight, high strength, and good processability.³ Rubber, with its flexibility and environmental resistance, plays a significant role in sealing and low-voltage insulating applications.⁴ While plastics, known for their excellent processing capabilities, lightweight, and chemical corrosion resistance, are extensively used in low-voltage devices and electronic components.⁵ These materials exhibit significant differences in chemical composition, physical structure, and electrical performance, leading to highly specialized functional characteristics. Therefore, the efficient classification of insulating materials is crucial for their recycling and utilization. It can effectively reduce resource waste, minimize environmental pollution, and enhance resource reutilization, thereby promoting sustainable development.⁶

Traditional classification methods of insulating materials typically rely on chemical composition analysis, physical performance testing, or empirical judgment. While these methods can meet certain requirements for material identification, their limitations have become increasingly evident. For example, chemical composition analysis often requires the destruction of samples and is time-consuming. Physical performance testing involves complex equipment and cumbersome procedures. Empirical judgment heavily depends on personal experience, making it less accurate and consistent.⁷ Particularly in the context of rapid industrial development, diverse demands, and large-scale production, traditional classification methods fail to meet the requirements for efficiency, precision, and real-time testing. Therefore, developing an efficient, precise, and rapid classification method for insulating materials is a key focus and hotspot in the field of material recyclin.⁸

With the rapid development of science and technology, advanced spectral analysis techniques are playing an increasingly important role in the classification and recycling of insulating materials. Li et al.⁹ utilized near-infrared spectroscopy to automatically sort and recycle plastics from waste household appliances and electronics. Hu et al.¹⁰ utilizes terahertz time-domain spectroscopy to achieve rapid and accurate identification of rubber. Differing from other techniques, laser-induced breakdown spectroscopy (LIBS) offers a completely new approach to the identification of insulating materials by analyzing their elemental composition. The basic principle of LIBS is simple, using a laser as an excitation source to ablate the sample and generate plasma. Further, by analyzing the radiation spectrum, the elemental composition and concentration of the sample can be determined.¹¹ With its unique technical principle, LIBS has plenty of advantages, such as no or simple sample preparation, quasi-nondestructive measurement, in situ and remote detection, rapid and real-time response, and all-element synchronous analysis.¹² Therefore, LIBS is highly suitable for the identification of insulating materials, and related research is increasingly gaining attention. Gundawar et al.¹³ achieved rapid and effective identification of plastics using a low-cost LIBS device. Chillu et al.¹⁴ successfully utilized LIBS to accurately classify polluted silicone rubber insulators. Unnikrishnan et al.¹⁵ successfully classified four types of plastic insulating materials based on LIBS spectra. These studies fully demonstrate the feasibility of LIBS in the field of insulating material identification.

However, LIBS spectra typically comprise tens of thousands of spectral lines, resulting in extremely high dimensionality.¹⁶ Meanwhile, due to the unique structure and composition of insulating materials, their spectra often exhibit characteristics such as sparsity, prominent background signals, and significant noise. Furthermore, with the continuous growth in demand for rapid and high-throughput analysis, the data volume is expected to increase exponentially. These spectral characteristics of insulating materials not only increase the complexity of data processing and storage but also make information extraction more challenging, significantly heightening the risk of losing critical details or introducing redundant information. Therefore, spectral dimensionality reduction has become an essential step in the LIBS analysis process.¹⁵

In the field of LIBS, most studies commonly employ artificial empirical selection, feature selection algorithms such as genetic algorithms and particle swarm optimization, as well as feature extraction algorithms like principal component analysis (PCA), independent component analysis, and locally linear embedding for dimensionality reduction.^17–19 While these methods can effectively reduce the complexity of spectra, the dimensionality reduction process inevitably leads to the loss of key information, such as weak spectral lines, intensity ratio, and spectral shape. Although these features may not be prominent in the overall spectrum, it is often crucial for high-performance analysis in specific applications and can significantly improve the analytical performance. Therefore, reducing the spectral dimensions while preserving as much critical information as possible has become a research hotspot in the field of spectral data processing.²⁰

In this study, to overcome this challenge, a novel approach called the generalized spectrum method (GSM-LIBS) was proposed. This method can extract more detailed features from the raw spectrum, such as peak intensity, integral intensity, intensity ratio, radiation background, and spectral shape, while reducing dimensionality, thus generating a more representative generalized spectrum. Therefore, compared to traditional dimensionality reduction techniques, GSM-LIBS not only incorporates spectral intensity but also combines additional potential physical, chemical, and geometric features, forming a higher-dimensional feature space that significantly enhances the richness of information. By enriching the effective features of spectra, this method improves the adaptability of analytical models to complex data, such as background noise, sparsity, and high dimensionality, while effectively improving analysis performance and generalization ability. To validate the effectiveness of this method, two traditional machine learning models, KNN and SVM, and one deep learning model, NN, were used. Further, the performance of these models was evaluated using relevant parameters, such as accuracy, macro precision, macro recall, macro F1-score, and confusion matrix. Finally, the experimental results fully confirm the feasibility of GSM-LIBS in effectively improving the performance of classification models.

2. Experimental setup and samples

2.1. Experimental setup

The experimental setup used in this study is depicted in Fig. 1. A laser beam (wavelength: 532 nm, pulse width: 8 ns, repetition: 10 Hz) is generated by a Nimma-400 series Q-switched Nd: YAG laser from Beamtech. And then, the laser beam is reflected by an optical system, and focused onto the sample, which absorbs the laser energy, rapidly melting, vaporizing, and ionizing, thereby generating plasma. Further, the photons emitted by plasma are collected by a light collector (Ocean Optics, 84-UV-25, wavelength range: 200–2000 nm), and then transmitted to a spectrometer (Avantes, AvaSpec-ULS4096CL-EVO, wavelength range: 200–900 nm, resolution: 0.01 nm), forming a LIBS spectrum. To ensure the laser beam ablates at different positions each time, the sample is placed on a three-dimensional displacement platform. The entire experimental setup is precisely controlled by a digital delay generator and managed by a computer, enabling the real-time acquisition of LIBS spectra.

All experiments were conducted under controlled laboratory conditions, with a temperature of approximately 25 °C and relative humidity of about 62% RH, at standard atmospheric pressure.

2.2. Samples and sample preparation

In this study, 7 different insulating materials were utilized, namely rubber (category 1), silicone (category 2), mica (category 3), ultra-high molecular weight polyethylene (category 4), polyoxymethylene (category 5), nylon (category 6), and polyvinyl chloride (category 7), all sourced from the local market in China. To ensure the reliability of the spectral data, all samples were cleaned prior to analysis to eliminate the influence of surface dust. Before analysis, the samples were gently wiped with a lint-free cloth, rinsed with deionized water, and air-dried at room temperature to ensure complete removal of surface contaminants. Meanwhile, to obtain high-quality spectra, the experimental parameters were optimized, and the gate delay and gate width of the spectrometer were ultimately set to 9 μs and 2.5 μs, respectively. For each type of sample, 100 spectra were obtained, resulting in a total of 700 spectra, for subsequent research on insulating material identification.

3. Analytical methods and evaluation indexes

3.1. Principle of generalized spectrum method

Due to the high resolution and rich detailed information of LIBS spectra, the data dimensions are typically very large, often containing tens of thousands of features. While these high-dimensional data provide abundant information, they also increase the complexity of data processing. Therefore, how to effectively extract key information and reduce redundant data has become an important challenge in LIBS.²¹ There are various traditional dimensionality reduction methods, including feature selection methods such as genetic algorithms, and particle swarm optimization, as well as feature extraction methods like PCA, independent component analysis, and locally linear embedding.^17–19 However, these traditional dimensionality reduction methods typically focus only on the intensity information of the spectra, often failing to preserve all key spectral features, such as intensity ratio, radiation background, and spectral shape, which may be overlooked or lost during the dimensionality reduction process. These features are crucial for precisely characterizing the chemical composition or physical properties of substances. For instance, intensity ratio can reflect the relative concentrations of different elements or molecules, while radiation background and spectral shape can provide insights into physical properties like plasma temperature and electron number density.^22–25 Therefore, while traditional dimensionality reduction methods can effectively reduce data dimensionality and complexity, it may lead to the loss of these essential spectral features, thus impacting the performance and reliability of subsequent analysis and modeling. To solve this problem, this study proposes an innovative method named GSM-LIBS. This method extracts multiple key features from the spectra, including peak intensity, integral intensity, intensity ratio, radiation background, and spectral shape. Furthermore, these features are then combined to form a generalized spectrum. In this way, GSM-LIBS can effectively reduce the spectral dimensionality while preserving as much critical information as possible, thus providing a more comprehensive and effective feature representation for subsequent analysis. The flow chart of GSM-LIBS is shown in Fig. 2, and the specific implementation process will be introduced in detail below.

The first step of GSM-LIBS is to extract the peak intensity of the spectral line from the raw spectrum, as shown in Fig. 3(a). The peak intensity of spectral lines is one of the most intuitive features of the raw spectrum, reflecting the radiative intensity of the corresponding element, and is directly related to the composition and concentration of the sample. Therefore, extracting peak intensity can capture the critical information of the absolute elemental content from the spectrum, serving as the foundation for subsequent spectral analysis.²⁶ In this process, various existing methods can be employed. For instance, the peak intensity of high-quality spectral line can be manually selected based on the NIST database. Additionally, genetic algorithms, particle swarm optimization, and other intelligent optimization methods can be used to automatically select the most representative spectral line.^17,18

The second step of GSM-LIBS is to extract the integral intensity from the raw spectrum, as shown in Fig. 3(b). The integral intensity not only includes the integral intensity of individual spectral line, but also encompasses the integral intensity over the ultraviolet, visible, infrared, and the entire spectral range. Compared to peak intensity, integral intensity not only provides the overall energy information of a spectral line, but also effectively smooths the impact of noise, avoiding errors that may arise from a single peak. This improves the stability and robustness of spectral feature identification to some extent, especially in complex backgrounds or weak signals, where it offers a higher signal-to-noise ratio and better discriminative ability.²⁷ Additionally, the integral intensity over the ultraviolet, visible, infrared, and the entire spectral range can also reflect the fluctuation behavior of plasma, which can provide reference information for the spectral analysis process, thereby improving its reliability.²⁸

The third step of GSM-LIBS is to extract the intensity ratio from the raw spectrum, as shown in Fig. 3(c). This feature mainly includes the intensity ratio between spectral lines of the same element and different elements. Among them, the intensity ratio between lines of the same element can be used to monitor nonlinear effects in the plasma, such as self-absorption effects and spectral line overlap.²⁹ Meanwhile, it can also reflect the changes of plasma temperature, providing more precise information about the plasma state.³⁰ And the intensity ratio between lines of different elements can reflect the relative concentrations of elements in the sample and reveal the relative abundance distribution between elements.²²

The fourth step of GSM-LIBS is to extract the radiation background from the raw spectrum, as shown in Fig. 3(d). The radiation background originates from the plasma, ambient light, and the instrument itself. Although it is typically regarded as an interference factor, it also contains crucial information about the plasma state, experimental conditions, and instrument performance. Therefore, effectively utilizing these background signals can deepen the understanding of plasma characteristics and improve the reliability of experimental results.^24,25 In this process, many methods such as polynomial fitting and wavelet transform can be used to extract the radiation background. Subsequently, traditional dimensionality reduction methods such as PCA can be applied to efficiently extract key information from the radiation background.³¹

The fifth step of GSM-LIBS is to extract the spectral shape from the raw spectrum, as shown in Fig. 3(e)–(g). The spectral shape includes several key features, such as full width at half maximum (FWHM), as well as the first and second derivatives of the spectral line. Among these features, the FWHM can reflect important physical properties of the plasma, such as temperature, pressure, electron number density, and self-absorption effect.³² The first derivative can reveal regions of rapid variation, helping to identify the intensity change trend and positional shift of the spectral line. The second derivative can capture the curvature changes, highlighting features such as the peak, width, symmetry, and potential overlap or distortion of the spectral line.^33,34

Furthermore, the extracted key features from the raw spectrum are processed and concatenated. Since there may be significant differences in the magnitudes of different features, all features should be first normalized to ensure consistent weighting in the spectral analysis. This process not only effectively reduces scale differences between features but also preserves the essential characteristics of each type of information. After normalization, these features are concatenated to form a new generalized spectrum. Finally, the generalized spectrum is input into the model for training and modeling, achieving high-performance analysis and prediction.

3.2. Evaluation indexes

To evaluate the performance of different classification models, multiple parameters, accuracy, macro precision (MacP), macro recall (MacR), macro F1-score (MacF), and the confusion matrix, were utilized in this study. Among them, macro precision is the arithmetic average of the precision (P) of each class. Macro recall is the arithmetic average of the recall (R) of each class. And macro F1-score is the weighted harmonic average of the precision (P) and recall (R) of each class. Their expressions are as follows:^35,36

	(1)

	(2)

	(3)

	(4)

where represents the class label; represents the total number of classes; and represent the true predictions and false predictions of the positive samples; and represent the true predictions and false predictions of the negative samples, respectively.

4. Results and discussion

As shown in Fig. 4, the spectra of insulating materials have very high dimensionality, reaching 24564 dimensions. However, only a small portion of these dimensions contains meaningful information, such as the spectral lines of elements like C, H, O, Ca, Mg, and Na, while the majority may consist of noise or redundant data. Therefore, when performing data analysis and modeling, it becomes crucial to extract features with practical significance from these high-dimensional spectra. In this study, GSM-LIBS was proposed and further applied to the task of insulating material identification, demonstrating its high-performance classification capabilities. In this method, key features from the raw spectrum, including peak intensity, integral intensity, intensity ratio, radiation background, and spectral shape, were extracted to form the generalized spectrum. Further, the classification models of insulating materials were constructed based on the generalized spectrum. To comprehensively evaluate the performance of GSM-LIBS, both traditional machine learning and deep learning approaches were employed. To make the results universal, two traditional machine learning models, KNN and SVM, and one deep learning model, NN, were used. Before modeling, the generalized spectra were randomly split into the training set and prediction set in a 7:3 ratio. Among them, the training set was used to train the models and optimize their hyperparameters, while the prediction set was used to evaluate the performance of the optimized models. To eliminate the impact of randomness in sample division, 100 random segmentation operations were performed, and the average evaluation indexes of these models were obtained. To thoroughly validate the effectiveness of GSM-LIBS, it is necessary to compare this method with traditional methods. In this study, the commonly used dimensionality reduction method, PCA, was chosen for comparison. The experimental results of PCA-LIBS and GSM-LIBS based on different classification models were analyzed in detail below.

4.1. Results of machine learning models

A detailed analysis was conducted on the results of the optimal classification models based on two machine learning methods, KNN and SVM, as shown in Fig. 5. First, the classification models based on PCA-LIBS were analyzed. The KNN algorithm, due to its simplicity and reliance on distance calculations in the feature space, is highly sensitive to the scale and distribution of features and is easily influenced by noise and outliers, resulting in relatively poor performance. The accuracy, MacP, MacR, and MacF of the KNN classification model on the prediction set were 0.935, 0.943, 0.935, and 0.930, respectively. Compared to KNN, the SVM algorithm can effectively handle high-dimensional data and is more robust to noise and outliers, usually having better analytical performance. Therefore, the performance of the SVM classification model for insulating materials was significantly improved, with an accuracy of 0.965, MacP of 0.969, MacR of 0.965, and MacF of 0.965 on the prediction set. However, the PCA dimensionality reduction method only focuses on spectral intensity, overlooking many important features, such as integral intensity, intensity ratio, radiation background, and spectral shape. Therefore, there is still room for improvement in the performance of the classification models built on PCA-LIBS. Further, a detailed analysis of the classification models based on GSM-LIBS was conducted. The experimental results show that both the KNN and SVM models have better analytical performance, compared to the results of PCA-LIBS. The accuracy, MacP, MacR, and MacF of the KNN classification model on the prediction set increased from 0.935, 0.943, 0.935, and 0.930 to 0.979, 0.980, 0.979, and 0.979, respectively. Among all the classification models, the SVM model based on GSM-LIBS had the best performance, with accuracy, MacP, MacR, and MacF all reaching 0.996.

Furthermore, the confusion matrix of each classification model was analyzed, as shown in Fig. 6. It can be seen that the classification models based on PCA-LIBS performs poorly in distinguishing nylon from other insulating material. The KNN model achieves a recognition accuracy of only 0.623 for nylon, while the SVM model performs better with an accuracy of 0.859. After dimensionality reduction by GSM-LIBS, the accuracy of the classification models in distinguishing nylon has significantly improved, with both the KNN and SVM models achieving 1.000. Therefore, these results preliminarily demonstrated that GSM-LIBS can effectively extract critical features from the raw spectrum, significantly improving the analytical performance of classification models.

4.2. Results of deep learning model

To further validate the wide applicability of GSM-LIBS, a detailed analysis was performed on the optimal classification models based on the deep learning method, NN, with the results presented in Fig. 7. The accuracy, MacP, MacR, and MacF of the classification model based on PCA-LIBS on the prediction set were 0.984, 0.985, 0.984, and 0.984, respectively. After dimensionality reduction by GSM-LIBS, the analytical performance of the classification model was significantly improved. The accuracy, MacP, MacR, and MacF on the prediction set were improved to 0.994, 0.995, 0.994, and 0.994. Therefore, based on the above results, it can be observed that GSM-LIBS can effectively reduce the spectral dimensions while significantly improving its analytical performance by preserving the critical features of the raw spectrum.

5. Conclusion

In this study, to achieve high-performance identification of insulating materials, a novel method named GSM-LIBS, was proposed. This method can extract various critical features from the raw spectrum, such as peak intensity, integral intensity, intensity ratio, radiation background, and spectral shape. Unlike traditional dimensionality reduction methods, GSM-LIBS can effectively reduce data dimensions while preserving key information, providing more precise details on plasma state, element concentration, and spectral characteristics, thereby significantly improving the performance of classification models. To verify the effectiveness of GSM-LIBS, this method was applied to the classification study of seven types of insulating materials and compared with PCA-LIBS. To ensure the general applicability of this study, two traditional machine learning models, KNN and SVM, and one deep learning model, NN, were used. For the machine learning models, the accuracy of the KNN and SVM classification models on the prediction set improved from 0.932 and 0.956 to 0.976 and 0.995, respectively. For the deep learning model, the accuracy of the NN classification model on the prediction set improved from 0.956 to 0.995. The relevant results fully prove the effectiveness and wide applicability of GSM-LIBS. Therefore, this study provides an effective spectral analysis method for LIBS, improving its analytical stability, robustness, and reliability, while also driving its further development in industrial applications.

Author contributions

Junfei Nie: conceptualization, methodology, investigation, writing – original draft. Furong Chen: data curation, formal analysis, writing – review & editing. Ting Luo: software, visualization, validation. Jinke Chen: resources, investigation. Jiapei Cao: formal analysis, experimentation. Qiang Huang: project administration, supervision. Zhenlin Hu: validation, writing – review & editing. Deng Zhang: supervision, conceptualization, funding acquisition, writing – review & editing, corresponding author. All authors have read and approved the final version of the manuscript.

Conflicts of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

The data supporting this study's findings are available from the corresponding author upon reasonable request.

Acknowledgements

This research was financially supported by, National Key Research and Development Program of China (No. SQ2024YFB2505200), the National Natural Science Foundation of China (No. 52207167, No. 52477112, No. 62405336), the Postdoctoral Fellowship Program of CPSF (No. GZB20230791), and the Hunan Provincial Department of Education General Project (No. 24C0392).

Notes and references

1. M. Pfundstein, R. Gellert, M. Spitzner and A. Rudolphi, Insulating Materials: Principles, Materials, Applications, DETAIL-Institut für internationale Architektur-Dokumentation GmbH & Co. KG, 2008 Search PubMed .
2. S. Li, S. Yu and Y. Feng, High Volt., 2016, 1, 122–129 CrossRef .
3. A. Küchler and A. Küchler,High Voltage Engineering: Fundamentals-Technology-Applications, Springer, 2018, pp. 301–354 Search PubMed .
4. H. Niu, Y. Ren, H. Guo, K. Małycha, K. Orzechowski and S.-L. Bai, Compos. Commun., 2020, 22, 100430 CrossRef .
5. B. Björkner, M. Frick-Engfeldt, A. Pontén and E. Zimerson, Contact Dermatitis, Springer, 2011, pp. 695–728 Search PubMed .
6. L. Yang, L. Zhen, T. Guilan and L. Ying, Proc. 2011 Third International Conference on Measuring Technology and Mechatronics Automation, IEEE, 2011, pp. 973–976 Search PubMed .
7. M. E. Grigore, Recycling, 2017, 2, 24 CrossRef .
8. D. Achilias, Material Recycling: Trends and Perspectives, 2012 Search PubMed .
9. X. Wu, J. Li, L. Yao and Z. Xu, J. Clean. Prod., 2020, 246, 118732 CrossRef CAS .
10. J. Hu, W. Liu, L. Yang, H. Lv, C. Zhan and P. Qiao, Infrared Phys. Technol., 2024, 139, 105324 CrossRef .
11. M. Cui, G. Shi, L. Deng, H. Guo, S. Xiong, L. Tan, C. Yao, D. Zhang and Y. Deguchi, J. Anal. At. Spectrom., 2024, 39, 1361–1374 RSC .
12. C. Sun, Y. Tian, L. Gao, Y. Niu, T. Zhang, H. Li, Y. Zhang, Z. Yue, N. Delepine-Gilon and J. Yu, Sci. Rep., 2019, 9, 11363 CrossRef PubMed .
13. R. Junjuri and M. K. Gundawar, Waste Manag., 2020, 117, 48–57 CrossRef CAS PubMed .
14. K. Sanjana, M. S. Babu, R. Sarathi and N. Chillu, IEEE Access, 2022, 11, 1752–1760 Search PubMed .
15. V. Unnikrishnan, K. Choudhari, S. D. Kulkarni, R. Nayak, V. Kartha and C. Santhosh, RSC Adv., 2013, 3, 25872–25880 RSC .
16. P. Pořízka, J. Klus, E. Képeš, D. Prochazka, D. W. Hahn and J. Kaiser, Spectrochim. Acta Part B At. Spectrosc., 2018, 148, 65–82 CrossRef .
17. Z. Gajarska, A. Faruzelová, E. Képeš, D. Prochazka, P. Pořízka, J. Kaiser, H. Lohninger and A. Limbeck, J. Anal. At. Spectrom., 2024, 39, 3151–3161 RSC .
18. A. K. Myakalwar, N. Spegazzini, C. Zhang, S. K. Anubham, R. R. Dasari, I. Barman and M. K. Gundawar, Sci. Rep., 2015, 5, 13169 CrossRef PubMed .
19. D. Pokrajac, A. Lazarevic, V. Kecman, A. Marcano, Y. Markushin, T. Vance, N. Reljin, S. McDaniel and N. Melikechi, Appl. Spectrosc., 2014, 68, 1067–1075 CrossRef CAS PubMed .
20. A. Gracia, S. González, V. Robles and E. Menasalvas, Inf. Sci., 2014, 270, 1–27 CrossRef .
21. T. Zhang, H. Tang and H. Li, J. Chemom., 2018, 32, e2983 CrossRef .
22. A. Stankova, N. Gilon, L. Dutruch and V. Kanicky, Fuel, 2010, 89, 3468–3474 CrossRef CAS .
23. M. L. Najarian and R. C. Chinni, J. Chem. Educ., 2013, 90, 244–247 CrossRef CAS .
24. A. De Giacomo, R. Gaudiuso, M. Dell'Aglio and A. Santagata, Spectrochim. Acta Part B At. Spectrosc., 2010, 65, 385–394 CrossRef .
25. J. Hermann, D. Grojo, E. Axente and V. Craciun, Spectrochim. Acta Part B At. Spectrosc., 2018, 144, 82–86 CrossRef CAS .
26. D. W. Hahn and N. Omenetto, Appl. Spectrosc., 2010, 64, 335A–366A CrossRef CAS PubMed .
27. J. Yan, F. Fan, Q. Li, Z. Tang, W. Chen, W. Jiang, Q. Li, L. Yu and X. Li, Spectrochim. Acta Part B At. Spectrosc., 2022, 190, 106375 CrossRef CAS .
28. J. Guezenoc, A. Gallet-Budynek and B. Bousquet, Spectrochim. Acta Part B At. Spectrosc., 2019, 160, 105688 CrossRef CAS .
29. J. Hou, D. Zhang, Z. Feng, J. Zhu and L. Zhang, Opt. Express, 2023, 31, 34404–34412 CrossRef CAS PubMed .
30. N. Konjević, M. Ivković and S. Jovićević, Spectrochim. Acta Part B At. Spectrosc., 2010, 65, 593–602 CrossRef .
31. B. Tan, M. Huang, Q. Zhu, Y. Guo and J. Qin, Spectrochim. Acta Part B At. Spectrosc., 2017, 138, 64–71 CrossRef CAS .
32. H. Li, H. Wang, S. Zhang, S. Kang, F. Wang, L. Fang and X. Li, Optoelectron. Lett., 2021, 17, 558–563 CrossRef .
33. I. G. Roy, J. Comput. Phys., 2015, 295, 307–321 CrossRef .
34. E. Koryukina and G. Revalde, Russ. Phys. J., 2006, 49, 435–441 CrossRef CAS .
35. A. Gong, L. Guo, Y. Yu, Y. Xia, X. Deng and Z. Hu, Inf. Fusion, 2024, 101, 101981 CrossRef .
36. M. Grandini, E. Bagli and G. Visani, arXiv, 2020, preprint, arXiv:2008.05756.05756,  DOI:10.48550/arXiv.2008.05756.

This journal is © The Royal Society of Chemistry 2025