Disciplines across the geosciences are beginning to implement machine learning (ML) algorithms, but the field of spectroscopy has remained stubbornly dedicated to peak-fitting techniques for quantitative analyses. Paradigm shifts will occur as this field broadens to utilize ML techniques, despite issues preventing widespread use. Foremost among these is the precedent that only specific spectral features can be correlated to specific compositional information, and that other regions of the spectra are useless. A second major issue is the lack of deep spectral libraries with adequate metadata from which algorithms can learn. Both regression and classification algorithms benefit from ML approaches, providing information on key spectral regions that contribute to understanding the variables of interest. Computational approaches have additional benefits for error estimation, providing root mean square errors that are more interpretable than chi-squared estimates.

Examples of ML advantages are presented. X-ray absorption spectroscopy has relied upon the pre-edge region and interpolation between known Fe²⁺ and Fe³⁺ peak positions in end-member minerals. ML approaches using the entire XANES/main-edge  range produce improved valence state predictions with well-quantified accuracy. For laser-induced breakdown spectroscopy, multivariate approaches such as partial least squares regression provide elemental analyses that are far more accurate than those based upon single emission lines of individual chemical species. Moreover, they have led to discoveries of previously unreported emission lines for specific elements. Finally, asteroid classification has long relied upon the Bus-DeMeo methods that use an empirical method for characterizing telescopic spectra and often require manual inspection of visible-near-IR spectra. ML approaches trained on meteorite spectra show superior performance in classifying asteroids using Earth-based observations.

All such promising approaches require fundamental data to support prediction accuracy. Efforts to collect, collate, and preserve reference data are ongoing but still fall short for most applications due to the poor state of metadata/documentation and lack of diversity in databases. As a result, the capabilities and sophistication of ML approaches for solving geochemical and mineralogical problems currently far outstrip the quantity and caliber of databases on which to train them.  This disparity will only grow until the importance of acquiring fundamental data, such as mineral mixtures with known volumetric abundances or standards with known compositions, is recognized and supported by researchers and funding agencies.