Quantitative analysis and classification of uranium polymetallic ores using femtosecond laser-induced breakdown spectroscopy combined with machine learning algorithms

Abstract

Uranium polymetallic ores (UPOs) are strategic emerging mineral resources that possess both significant economic value and strategic importance. In this study, femtosecond laser-induced breakdown spectroscopy (LIBS) was combined with four machine learning algorithms—partial least squares regression (PLSR), principal component analysis (PCA), support vector machine (SVM), and linear discriminant analysis (LDA)—to perform quantitative analysis of uranium (U) concentration and classification of UPO samples. The concentrations of U in six UPO samples were determined using a high-purity germanium gamma ray spectrometer, which served as reference values. Three spectral normalization methods were applied to preprocess the raw spectra and assess the impact of preprocessing on model performance. Due to the significant matrix effect, the intensities of U characteristic emission lines did not exhibit a clear linear correlation with U concentration, making the univariate analysis impossible. However, the multivariate regression model based on PLSR algorithm was employed to mitigate the matrix effect, enabling accurate U quantification in UPOs. The leave-one-out cross-validation results showed that, with the exception of sample 1#, the combination of femtosecond LIBS and PLSR reliably predicted U concentration in the other samples, with relative standard deviation and mean relative error maintained below 9.48% and 7.30%, respectively. Furthermore, PCA was applied to the whole LIBS spectral dataset for dimensionality reduction and feature vector reconstruction. The resulting principal components were used as inputs for SVM and LDA classification algorithms to distinguish among the six ore types. The SVM model achieved a classification accuracy of 91.67%, while LDA demonstrated a superior classification performance with 100% accuracy. Overall, this study demonstrates that the combination of femtosecond LIBS with machine learning algorithms enables effective quantitative analysis of U and accurate classification of UPOs.