Characterizing silicate materials via Raman spectroscopy and machine learning: Implications for novel approaches to studying melt dynamics

Silicate melt characteristics impose dramatic influence over igneous processes that operate, or have operated on, differentiated bodies: such as the Earth and Mars. Current understanding of these melt properties, such as composition, primarily comes from investigations on their volcanic byproducts. Therefore, it is imperative to innovate on modalities capable of constraining melt information in environments where a reliance on laboratory methods is severed. Recent investigations have turned to Raman Spectroscopy and amorphous volcanics as a suitable pairing for exploring these ideas. Silicate glasses are a proxy for igneous melts; and Raman spectroscopy is a robust analytical technique capable of operating in-situ. Existing calibrations for retrieving geochemical information from such samples using their Raman data are extremely underdeveloped, with only a handful of approaches available. Here, two supervised machine learning algorithms; Partial Least Squares (PLS) and Least Absolute Shrinkage & Selection Operator (LASSO) are employed with Raman spectroscopy to quantify geochemical information in volcanic glasses and tephra, while also qualifying the underlying atomic mechanics that drive Raman signal variability. This approach establishes a foundation for future explorations into new-age modeling technologies for geoscience experiments. Chapter I’s PLS geochemical model predicted the concentrations of oxide constituents in synthetic silicate glasses (SiO₂, Na₂O, K₂O, CaO, TiO₂, Al₂O₃, FeO_T, MgO) with increased accuracy and applicability over currently available offerings. The study presents the largest and most diverse sampling suite yet utilized to produce such models. Chapter II highlights the limitations to PLS and LASSO based strategies for constraining iron (Fe)-redox information in glasses but uncovers their ability to accurately predict glass structural parameters like polymerization (NBO/T). Chapter III yielded accurate predictions of tephra concentrations from various mixed sediment samplings using PLS and LASSO calibrations. Spectra parameterizations highlighted that tephra signatures are unique enough to be readily distinguished from more crystalline profiles using Raman spectroscopy and machine learning. PLS and LASSO technologies are shown to be suitable, yet immature, avenues for unraveling the geochemical underpinnings of the Raman collections made in this work and help set the stage for future applications to Raman data from planetary missions such as the Perseverance Rover.