BACKGROUND LIBS technology is a non-destructive, high sensitivity, high resolution spectroscopy technology that can be used to analyze the composition and structure of chemical substances and materials. It has extensive application in fields such as chemistry, materials science, life science, and geological exploration, and its emergence has provided new methods and technologies for the development of these fields. LIBS technology can be used to non-destructively analyze the chemical composition of underground rocks and minerals, helping geologists to better understand the composition and properties of underground resources, thus providing better guidance for geological exploration and development. In recent years, scholars at home and abroad have been exploring LIBS technology constantly, and through improving the detection system and optimizing laser pulse parameters, high sensitivity LIBS analysis at extremely low concentration has been achieved. By using finer spectral lines, higher sampling rate, and more precise laser pulse control, high resolution LIBS analysis at nanoscale has been achieved. The combination of LIBS technology with multi-spectral image processing technology can integrate information from multiple spectral channels to achieve a more comprehensive analysis of samples. However, the existence of matrix effects and spectral fluctuations always affects the accuracy of LIBS quantitative analysis, and poor reproducibility and high detection limits also need to be solved.

OBJECTIVES To improve the accuracy of quantitative analysis of complex matrix samples.

METHODS A multi-layer classification model based on k-nearest neighbors (kNN) and support vector machine (SVM) algorithms was constructed to identify the rock type of samples. The samples were divided into two major categories of felsic rocks and mafic rocks using the kNN algorithm, and then six categories were formed by the SVM algorithm. Different element quantitative models were constructed for each rock type. The kNN algorithm was selected using cross-validation to determine the optimal k value, and the key punishment parameter C and RBF width parameter γ of the SVM algorithm were determined using a grid search method. Then, appropriate pre-processing methods were adopted to improve the stability of spectral data for different elements in different rock types. Compared to the traditional standard curve quantitative method, using the pre-classification method can reduce the influence of different rock matrices on each other, thus reducing errors caused by the non-uniform matrix of samples.

RESULTS Due to the influence of matrix effects, a single pre-processing method is not suitable for all elements in quantitative analysis. Therefore, in order to improve the accuracy and stability of quantitative analysis, different methods are used to pre-process the data. For different pre-processing methods, the R² values of four elements in six types of rock samples are mostly greater than 0.90, as shown in Table 2. After pre-processing, the correlation coefficients of the four elements are significantly improved, and they are all higher than 0.99. The correlation coefficients of Si, Ca, Mg, and K elements in the test set after quantitative analysis are increased from 0.664, 0.638, 0.461, and 0.231 to 0.999, 0.994, 0.999, and 0.996, respectively. In addition, it can be seen from the analysis of the data that the traditional quantitative analysis model has poor stability. The average relative standard deviation (RSD) of Si, Ca, Mg, and K elements in the test set are 3.4%, 10.7%, 48.2%, and 90.8%, respectively, while the RSD of four elements in the multi-layer model are 1.5%, 5.2%, 10.3%, and 17.4%, which shows a significant improvement in stability compared to traditional quantitative analysis models. At the same time, it can be more intuitively evaluated by comparing the average relative error between the predicted value and the target value of each element in the test set. As shown in Table 3, the prediction performance of Si element in the multi-layer model is the best, with an average relative error of only 4.65%. Although the average relative error of the other three elements is over 10%, it is significantly improved compared to the traditional standard curve model.

CONCLUSIONS By utilizing a multi-layer classification model for preliminary categorization, standard rock samples that match the matrix are obtained. Subsequently, quantitative analysis models are developed for samples with similar matrices. Employing distinct preprocessing methods for different elemental compositions within various rock types helps mitigate spectral discrepancies caused by matrix effects, reduces spectral fluctuations and data noise, and enhances the accuracy and stability of quantitative analysis. Standard curve models are then established for each element, enabling quantitative analysis of Si, Ca, Mg, and K elements in six categories of rock samples. Results demonstrate a notable improvement in the accuracy of quantitative analysis compared to traditional standard curve models. This model not only diminishes the impact of matrix effects on quantitative analysis but also corrects instabilities arising from hardware, environmental conditions, and sample variations. Furthermore, it alleviates the workload of data analysis, simplifying the analytical process and thereby boosting efficiency. However, the current multi-layer quantitative analysis model still exhibits some deviations in regard to different elements. In the future, a potential avenue is to consider integrating various algorithms to establish preliminary classification models, aiming for even better quantitative analysis outcomes.