Abstract

The trading price and taxes of copper-concentrate exports depend on the chemical composition. Valuable elements such as Cu, Ag, and Mo increase the price while the presence of As reduces it. The quantification of these elements by laser induced breakdown spectroscopy (LIBS) in mineral ores with varied concentrations spanning from ppm to percentage is challenging. Several factors such as matrix effects, signal fluctuations (laser energy fluctuations, plasma instability, sample inhomogeneity), and self-absorption effects can contribute to analytical errors. Although advanced chemometric methods like artificial neural networks (ANN) can account for these limitations, tuning too many parameters may lead to overfitting of the results. Hence, to avoid overfitting, meaningful information should be fed to the model. In this work, a systematic procedure for getting meaningful LIBS spectral data i.e. "judicious data processing" is proposed. It consists of spectra normalization, feature selection and stratified data division. First, the normalization is done by using an internal standard (ISSN) considering emission lines (Al, Ca, Fe, Zn, and Si) of the main components of the matrix. Second, the feature selection is performed by identifying linearly correlated wavelengths (LCW) with the concentration of each target element. Third, stratified data division in the ANN regression is implemented. This approach, involving LIBS-ANN regression, is applied for multi-elemental quantification of valuable (Cu, Ag, Mo) and penalized (As) elements in copper ores. We compared the ANN models, considering features selection by LCW and prior knowledge spectral lines (PKSL), which are usually utilized in the literature. In terms of analytical figures, the root mean square error of predictions (RMSEP) using LCW improved for Cu (1.45% to 1.04%), Ag (8.5 mg.kg(-1) to 6.0 mg.kg(-1)), and As (0.16% to 0.06%), while it remained unchanged for Mo (0.04%). These improvements can be attributed to the decrease of spectral interferences, caused by the selection of the least affected lines and the compensation of signal fluctuations by spectral normalization.