• Mccanta, Molly C.
• Lanzirotti, Antonio
• Dyar, Melinda Darby
• Ytsma, Cai
• Sutton, Steve
• Steven, Cody

Abstract

XAS data from a suite of 460 experimentally equilibrated silicate glasses were used to develop a universal calibration for determining %Fe3+ and log fO2 from XAS data. Partial least squares (PLS) and Least-absolute shrinkage and selection operator (LASSO) models were created using as input data the entire spectral range (7012-7350 eV), the entire range with the 8 major elements appended, SiO2 datasets with 35-50, 50-60, 60-70, and 70-80 wt% SiO2 and 0-2, 2-4, 4-7, and 7-11 wt% Na2O+K2O. Models used 5-fold cross-validation to optimize model parameters and then trained using all available data with accuracies reported as root mean square errors (RMSE-C). In all cases, model coefficients were weighted toward using wavelengths from the pre-edge rising to the main edge. <P />Overall, LASSO performed slightly better for predicting %Fe3+ than PLS, and somewhat worse for predicting log fO2. Models with appended compositions showed no improvement. LASSO models predicting %Fe3+ had RMSE-C of 4.4 %Fe3+ when using all 460 spectra. Accuracies improved considerably for SiO2-partitioned data, with RMSE-C = 4.6, 3.5, 3.2, and 2.3 %Fe3+, respectively and variably for Na2O+K2O-partitioned data with RMSE-C = 4.9, 2.1, 5.5, and 3.6 %Fe3+, respectively. <P />PLS predictions of log fO2 using the entire spectral range in all samples had accuracies of 1.1 log units and improved to 0.9 log units with compositions added. Smaller compositional ranges greatly improved accuracies: RMSE-C values for SiO2-stratified models were 0.6, 0.4, 0.3, and 1.9 log units, but optimal accuracy came from PLS models trained on Na2O+K2O partitions, with RMSE-C = 0.4, 0.7, 0.8, and 0.0 log units. <P />Compared to other methods for assessing %Fe3+ in glasses using only the pre-edge region, these multivariate methods offer many advantages: 1) model accuracy is robustly characterized, 2) calibrations apply to wide ranges of silicate glasses, 3) no time-consuming and error-adding spline removal and curve-fitting is needed; 4) %Fe3+ and fO2 can be simply calculated using a spreadsheet or Python code, and 5) log fO2 can be directly measured rather than inferred. <P />This calibration paves the way for studies of glasses in diverse igneous systems as well as partition coefficients in coexisting mineral-glass pairs....

Topics

• impedance spectroscopy
• mineral
• Oxygen
• glass
• glass
• iron
• x-ray absorption spectroscopy