Predicting geochemistry in geological samples using laser-induced breakdown spectroscopy: Effects of compositional and textural outliers

The chemistry of geologic targets on different planets is often determined from Laser-Induced Breakdown Spectroscopy (LIBS) data using multivariate, machine learning models calibrated on laboratory spectra from pressed pellets of ground rocks with known compositions. However, due to the diversity of geologic materials, LIBS may be used to predict the chemistry of targets that are distinct from a training dataset in their composition and form. For example, loose granular materials may introduce unknown spectral effects compared to solid rocks. We used a LIBS instrument to analyze mineral pellets and loose powders of olivine, labradorite, and augite in five grain size fractions ranging from <25 μm to 710–1000 μm. These minerals are compositional outliers relative to rocks, and powders are texturally distinct from pressed pellets. Outlier compositions adversely affected prediction accuracy in pellets and powders. Additionally, prediction accuracy was dramatically worse for powders <38 μm compared to either larger grain sizes (>63 μm) or pellets. Powders this fine coupled poorly to the laser and potentially had plasma confinement effects in deep, narrow ablation pits. Resulting spectra had low intensity atomic emissions and elevated noise when normalized. These outlying spectra introduce unexpected features in the model input, so the atomic relationships are not accurately interpreted and predictions trend towards what the model was scaled on, the mean of training data compositions. Despite best practices, quantification models for LIBS built on pellets of pressed rock powders cannot account for effects seen in loose fine powders. • Compositions of geologic samples analyzed by LIBS are predicted using PLS models. • PLS models are trained on pressed-powder pellets of rock standards. • Mineral pellets compositionally outside the training dataset had poor predictions. • Powders are texturally not in the training dataset and coupled poorly to the laser. • Predictions for powders <38 μm defaulted to the average composition of training data.