Laser-induced breakdown spectroscopy combined with multi-scale convolutional neural network and attention mechanisms for analyzing Dy, Nd, and Yb in uranium polymetallic ores.

BACKGROUND: Complex matrix effects and spectral interference severely limit the capability of laser-induced breakdown spectroscopy (LIBS) for quantitative analysis of rare earth elements in uranium polymetallic ores. A multi-scale dual-attention convolutional neural network (MSDA-CNN) is proposed for spectral analysis. The model employs parallel multi-scale convolutional branches to extract local details and broader-scale spectral trends, while channel and spatial attention mechanisms enable the model to adaptively focus on key wavelength regions, amplify informative features, and suppress redundant interference. RESULTS: The MSDA-CNN model achieves coefficient of determination (R2) values above 0.98 and root mean square error (RMSE) below 0.89 for quantifying Dy, Nd, and Yb. Module-level comparison experiments confirm the critical role of the multi-scale architecture and attention mechanisms in improving model performance. Visualization of attention weights further indicates that the model effectively focuses on characteristic spectral regions. It exhibits differentiated weight distributions for different elements, enhancing interpretability. Compared with other models, MSDA-CNN demonstrates advantages in both predictive accuracy and stability. SIGNIFICANCE: Integrating LIBS with the MSDA-CNN model provides a solution for high-precision quantification of rare earth elements in uranium polymetallic ores, with potential applications in mineral exploration and resource utilization.