Laser-induced breakdown spectroscopy combined with multi-task convolutional neural network for analyzing Sm, Nd, and Gd elements in uranium polymetallic ore

Abstract

When analyzing rare earth elements in uranium polymetallic ores using laser-induced breakdown spectroscopy (LIBS), issues such as spectral interference and overlap significantly reduce the accuracy of elemental analysis. A multi-task convolutional neural network (CNN) model based on an uncertainty-weighted loss function is developed, with raw LIBS spectral data as the input. Conduct experiments using pure oxides under the same experimental conditions to identify the characteristic spectral lines. The results show that, compared to normalized and feature-extracted data processed with partial least squares (PLS), random forests (RF), and single-task CNN models, multi-task CNN model with uncertainty loss achieves the best quantitative results for Sm and Nd elements (Sm: R² = 0.9939, RMSE = 0.6373; Nd: R² = 0.9987, RMSE = 0.3040), and also demonstrates excellent performance in quantifying Gd (R² = 0.9949, RMSE = 0.6106). The multi-task CNN model based on an uncertainty-weighted loss function indicates great potential applications for end-to-end processing of LIBS spectra from uranium polymetallic ore. © 2001 Elsevier Science. All rights reserved.