Laser-Induced Breakdown Spectroscopy and a Convolutional Neural Network Model for Predicting Total Iron Content in...

Area of Science:

• Analytical Chemistry
• Geochemistry
• Materials Science

Background:

• Laser-induced breakdown spectroscopy (LIBS) offers rapid total iron (TFe) detection in iron ores.
• Univariate regression analysis in LIBS faces limitations in accuracy due to laser fluctuations and spectral interference.
• Advanced multivariate analysis and feature selection are crucial for precise TFe quantification.

Purpose of the Study:

• To develop and evaluate a Convolutional Neural Network (CNN) model for accurate TFe content prediction in iron ores using LIBS data.
• To compare the performance of CNN-assisted LIBS with traditional and variable importance-based methods (VI-RF, VI-BP-ANN).
• To assess the potential of machine learning and deep learning for enhancing LIBS applications in the mining industry.

Main Methods:

• Collected 2034 LIBS spectra from 339 iron ore batches across five countries.
• Developed a CNN model for TFe content prediction.
• Compared CNN with Variable Importance Random Forest (VI-RF) and Variable Importance Back Propagation Artificial Neural Network (VI-BP-ANN) models.
• Evaluated models using R², RMSE, MRE, and modeling time.

Main Results:

• Variable importance methods enhanced quantitative accuracy and reduced modeling time compared to standard BP-ANN and RF.
• The CNN model outperformed VI-BP-ANN and VI-RF, demonstrating the shortest modeling time, highest R², and lowest RMSE and MRE.
• CNN's architecture (weight sharing, local connection) is well-suited for high-dimensional LIBS data analysis.

Conclusions:

• Machine learning and deep learning approaches significantly improve LIBS accuracy for TFe prediction in iron ores.
• CNN-assisted LIBS offers a powerful and accurate method for TFe content determination.
• The developed CNN-LIBS method shows great potential for practical applications in the mining industry.