Combining prior knowledge with input selection algorithms for quantitative analysis using neural networks in laser induced breakdown spectroscopy.

Laser-induced breakdown spectroscopy (LIBS) is an emerging technique for the analysis of rocks and mineral samples. Artificial neural networks (ANNs) have been used to estimate the concentration of minerals in samples from LIBS spectra. These spectra are very high dimensional data, and it is known that only specific wavelengths have information on the atomic and molecular features of the sample under investigation. This work presents a systematic methodology based on the Akaike information criterion (AIC) for selecting the wavelengths of LIBS spectra as well as the ANN model complexity, by combining prior knowledge and variable selection algorithms. Several variable selection algorithms are compared within the proposed methodology, namely KBest, a least absolute shrinkage and selection operator (LASSO) regularization, principal component analysis (PCA), and competitive adaptive reweighted sampling (CARS). As an illustrative example, the estimation of copper, iron and arsenic concentrations in pelletized mineral samples is performed. A dataset of LIBS emission spectra with 12 287 wavelengths in the range of 185-1049 nm obtained from 131 samples of copper concentrates is used for regression analysis. An ANN is then trained considering the selected reduced wavelength data. The results are satisfactory using LASSO and CARS algorithms along with prior knowledge, showing that the proposed methodology is very effective for selecting wavelengths and model complexity in quantitative analyses based on ANNs and LIBS.