A combination of spectrum selection and machine learning regression for minor element determination in gravel stones with LIBS

Abstract

The ability of laser-induced breakdown spectroscopy (LIBS) to directly analyze raw geological materials, such as rocks, often represents a decisive factor for its choice with respect to other techniques in field inspections, Mars exploration for example, or online analyses of ore transported by a conveyer belt for example. In such applications, the performance of LIBS is seriously affected by changes in the surface physical property and the posture of the materials to be analyzed when interacting with the laser, causing physical matrix effect in addition to the usual chemical matrix effect also encountered in the analyses of laboratory-prepared samples in the form of pellets. There can be application scenarios where concentration labels can be obtained for a collection of materials to be analyzed in their raw state, by using other characterization methods. Training prediction models using raw materials with concentration labels is possible and required, provided that the above-mentioned chemical and physical matrix effects are simultaneously and properly corrected. In this work, a machine learning regression model is combined with a spectrum selection procedure based on cosine similarity for especially reducing the influence of the physical matrix effect; while keeping a sufficient number of spectra for a regression based on random forest (RF), effectively correcting other perturbing effects including chemical matrix effect. Twenty-seven gravel stone samples prepared by crushing an equivalent number of natural rocks were involved in the experiment. LIBS spectra were directly acquired from gravel stones, while the laser beam was scanned over them. Four minor elements, Li, Rb, Sr and Ba, were quantified. The regression models were optimized by a control on the spectrum selection. Test of the optimized models by an ensemble of samples (six samples for each element) independent from the training ones leads to a final performance with root mean square errors of prediction (RMSEP) of 5.2, 36, 220 and 192 wt. ppm respectively for Li, Rb, Sr and Ba.