Machine Learning-Enhanced Evaluation of Handheld Laser-Induced Breakdown Spectroscopy (LIBS) Analytical Performance for Multi-Element Analysis of Rock Samples

Handheld laser-induced breakdown spectroscopy (hLIBS) can be considered one of the most recent techniques for rock characterization in situ. Handheld LIBS devices are useful tools for providing “fit for purpose” qualitative and quantitative geochemical data. The analytical performance of hLIBS instruments varies significantly between similar instruments from different manufacturers. This study employed two commercial hLIBS instruments, both making use of noise reduction and multivariate partial-least-squares (PLS) calibration. Model validation was performed using the Leave-One-Out Cross-Validation (LOOCV) method. The Random Forest (RF) and Artificial Neural Network (ANN) algorithms were also employed as complementary approaches to PLS modeling, with the goal of exploring potential nonlinear relationships between spectral intensities and reference analyte concentrations. A comparison was also made with the most basic and commonly used approach, univariate analysis, demonstrating that multivariate methods achieve superior performances. To evaluate the predictive performance and quantification capability of the acquired LIBS spectra, the Pearson’s coefficient (R²) and root-mean-square error (RMSE) were employed in the analysis of 21 diverse certified geochemical reference materials (CRMs). The results achieved suggested that the spectral resolution was the key factor determining the performance of multivariate LIBS calibrations. The PLS model proved to be satisfactory for analyses performed by the higher-spectral-resolution instrument, whereas complementary algorithms were necessary to achieve better results with the lower-spectral-resolution instrument.