Predicting Olivine Composition Using Raman Spectroscopy Through Band Shift and Multivariate Analyses
MetadataShow full item record
Olivine minerals control many of the properties of Earth’s upper mantle. Additionally, olivines affect rheology, they may store hydrogen, and they are diagnostic of crystallization temperature. Olivine’s presence in meteorites and terrestrial bodies makes the study of this mineral group crucial to the planetary science community. Olivine composition can be characterized using Raman spectroscopy. This in situ technique will be used to characterize Martian minerals in upcoming missions such as ExoMars and Mars 2020. Raman spectra of 93 olivines were acquired on Bruker’s 532 nm Senterra spectrometer. Of these samples, 25 were also run on Bruker’s BRAVO and Senterra (785 nm) spectrometers. Raman spectra of the olivine group minerals in the solid solution between forsterite (Mg2SiO4) and fayalite (Fe2SiO4) have a high intensity doublet between 800 and 880 cm-1. Historically, the band shift of these two peaks was utilized to predict the Mg/Fe contents (Kuebler et al., 2006; Foster et al., 2007; Gaisler and Kolesov, 2007; Mouri and Enami, 2008; Yasuzuka et al., 2009; Ishibashi et al., 2011), though these studies used different instruments and only limited data sets. This thesis compares the band shift method for understanding olivine composition with a more novel method using partial least squares (PLS), the least absolute shrinkage and selection operator (lasso), and least angle regression (LARS). To evaluate the accuracy of each univariate model, both the R2 value of each fit (on a plot of composition versus peak centroid) and the cross-validated root mean squared error (RMSE-CV) of each model were used. Internal cross-validation of each data set was the most accurate %Fo predictor. For example, the model for which all data were acquired on the BRAVO spectrometer (under identical operating conditions) predicts %Fo content best for other data acquired on the BRAVO. For this reason, previous studies may appear deceptively accurate because they did not use cross-validation nor evaluate the accuracy of predictions on “unseen data” (data not included within the model). Aggregated data sets from multiple instruments show excellent performance and can be generalized to other instruments for which calibrations are not available, such as the upcoming Raman Mars instruments. The most accurate %Fo predictions that avoid instrument bias result from an aggregated model produced through PLS or lasso multivariate analyses. A model that isolates the olivine doublet is also suggested instead of utilizing the entirety of the spectrum. Overall, this thesis demonstrates that multivariate analyses are superior to univariate methods for prediction of olivine composition from Raman spectroscopy. Multivariate analyses that use multiple instrument data avoid instrument bias and leverage multiple aspects of the spectra. Recommended PLS and lasso models with the smallest errors are listed in the appendix of this thesis for the use of future workers.