The influence of multivariate analysis methods and target grain size on the accuracy of remote quantitative chemical analysis of rocks using laser induced breakdown spectroscopy

Abstract

Laser-induced breakdown spectroscopy (LIBS) was used to quantitatively analyze 195 rock slab samples with known bulk chemical compositions, 90 pressed-powder samples derived from a subset of those rocks, and 31 pressed-powder geostandards under conditions that simulate the ChemCam instrument on the Mars Science Laboratory Rover (MSL), Curiosity. The low-volatile (<2 wt.%) silicate samples (90 rock slabs, corresponding powders, and 22 geostandards) were split into training, validation, and test sets. The LIBS spectra and chemical compositions of the training set were used with three multivariate methods to predict the chemical compositions of the test set. The methods were partial least squares (PLS), multilayer perceptron artificial neural networks (MLP ANNs) and cascade correlation (CC) ANNs. Both the full LIBS spectrum and the intensity at five pre-selected spectral channels per major element (feature selection) were used as input data for the multivariate calculations. The training spectra were supplied to the algorithms without averaging ( i.e. five spectra per target) and with averaging ( i.e. all spectra from the same target averaged and treated as one spectrum). In most cases neural networks did not perform better than PLS for our samples. PLS2 without spectral averaging outperformed all other procedures on the basis of lowest quadrature root mean squared error (RMSE) for both the full test set and the igneous rocks test set. The RMSE for PLS2 using the igneous rock slab test set is: 3.07 wt.% SiO 2, 0.87 wt.% TiO 2, 2.36 wt.% Al 2O 3, 2.20 wt.% Fe 2O 3, 0.08 wt.% MnO, 1.74 wt.% MgO, 1.14 wt.% CaO, 0.85 wt.% Na 2O, 0.81 wt.% K 2O. PLS1 with feature selection and averaging had a higher quadrature RMSE than PLS2, but merits further investigation as a method of reducing data volume and computation time and potentially improving prediction accuracy, particularly for samples that differ significantly from the training set. Precision and accuracy were influenced by the ratio of laser beam diameter (∼490 μm) to grain size, with coarse-grained rocks often resulting in lower accuracy and precision than analyses of fine-grained rocks and powders. The number of analysis spots that were normally required to produce a chemical analysis within one standard deviation of the true bulk composition ranged from ∼10 for fine-grained rocks to >20 for some coarse-grained rocks.