Quantitative analysis of ferromanganese crusts (Fe-Mn crusts) using laser-induced breakdown spectroscopy combined with machine learning.

Background: Rapid and on-board elemental analysis on the mineral deposits taken from the deep seabed are of great importance for the deep-sea mineral resource survey. Traditional geochemical tools are often time-consuming that require transferring the samples from the vessel to laboratory, therefore make a quite long time for acquiring the information of the mineral deposits after taking them from the deep seabed. There is a need to develop a rapid and environmentally friendly method, which is more important for the on-board mineral analysis during the deep-sea mineral resource survey.

Results: We evaluated the potential of LIBS combined with machine learning as a rapid tool for the quantitative analysis of ferromanganese crusts (Fe-Mn crusts). Both PLS and CNN models were built for the quantification of Fe, Mn, Ti, and Fe-Mn ratio in Fe-Mn crusts. With the full spectrum as input variables, PLS shows an overfitting behavior, while CNN exhibits superior generalization ability and robustness. Thereafter, feature selection of SD-SKB was applied on the broadband spectra and compared with Grad-CAM feature visualization within CNN. The performances of the feature-based models are superior or comparable with the full-spectrum models, while the model complexity and computational costs are significantly reduced and the interpretabilities of the models are improved. The predictive performance of CNN with the selected variables is clearly better than PLS with the selected variables, with the RMSE values of 0.422 wt% (Fe), 0.532 wt% (Mn), 0.045 wt% (Ti), and 0.031 (Fe-Mn ratio) for the feature-based CNN model.

Significance: This work demonstrated the capability of LIBS combined with machine learning that could be potentially used for the on-board mineral analysis during the deep-sea mineral resource survey.