GRSG Conference 2022: Orbit to Outcrop
Title: Laser-induced breakdown spectroscopy coupled with machine learning techniques for a high-level characterization of materials in various deposit types
Author: Feven Desta
In mining, accurate and reliable resource estimation is required, as it is advantageous for the efficient extraction of materials. Sensor technologies can provide rapid and accurate data for the reliable estimation of resources. The use of sensors along the mining value chain requires test case-specific calibration models. This study describes an innovative data-driven methodology for the characterization of raw materials at different levels using laser-induced breakdown spectroscopy (LIBS).
The LIBS data were acquired using 180 samples from various deposit types and IRIS SPECTRAL spectrometer with a 1064 nm pulsed laser. A methodological approach was developed to access the usability of the LIBS spectral data for the identification, classification, or semi-quantification of the target elements (Li, Cu, Pb, Zn, and Fe) in different matrices.
The approach commences with data exploration and pre-processing, followed by a usability assessment of the technique for the identification of the elements and feature extraction, and culminates with data modelling and model validation. The data exploration task includes outlier detection and pattern recognition. Following this, data pre-processing was performed using the data filtering techniques.
The pre-processed data were subsequently used to extract important variables (peaks wavelength location that corresponds to the specific element and correlates with the concentration of the elements). The extracted features were used to develop classification (support vector classification ─ SVC and linear discriminant analysis – LDA) and prediction (partial least square regression – PLSR and support vector regression ─ SVR) models for each target element, separately.
The results of this study show that the LIBS data include relevant information that can be employed in the identification, classification, and semi-quantification of the elements of interest. For each element identified in the LIBS data, the intensity of at least three diagnostic and relatively well-isolated emission lines was used. This approach allowed the identification of the target elements in most of the analyzed samples. However, peak overlapping was one of the challenges in assigning peaks.
The experimental results show that the LIBS data can be employed in the discrimination of ore–waste materials. For example, at a 0.5 % Cu cut-off grade, the best achieved SVC model resulted in a correct classification rate of 94 %, whereas the result from the LDA at the same cut-off grade is 97.6 %. At the 20 % Fe cut-off grade, the LIBS data LDA and SVC models yielded correct classification rates of 92 % and 79 %, respectively. For the classification of Fe, Cu, and Zn ore–waste materials the performance of the LDA model is superior to the SVC model. Whereas, for the classification of Pb and Li ore–waste materials the SVC model yielded better results than the LDA.
This shows the importance of model choice based on commodity. It is also evident that the choice of the pre-processing technique could influence the classification and prediction accuracies of the models. For example, for the prediction of Fe, the best-performing model was achieved using PLSR after standard normal variate (SNV) was applied to the dataset (RMSEP = 5.2, R2 = 0.90). The best prediction of Li concertation was achieved by the PLSR model (R2 = 0.73).
The use of LIBS data for semi-quantitative analysis of the elements resulted in promising performance. It is also evident that the prediction performance of the models depends on the deposit type in which the element occurs.
Quantitative analysis of elements using LIBS data is challenging due to various reasons. Such as matrix effect, particle size, and peak overlapping. Thus, the development of a good calibration model is challenging for most of the elements. However, other methodological approaches such as calibration of the models using known standards could be one of the possible solutions for a better determination of elemental concentration in the analyzed samples.
For the classification of the materials into ore and waste, both the linear and non-linear techniques provided good and acceptable results. Although the acquired prediction accuracies are lower than those of the standard laboratory-based techniques, the proposed method is suitable for the rapid in-situ characterization of materials along the mining value chain.