Spectroscopic Characterization of Impactites and a Machine Learning Approach to Determine the Oxidation State of Iron in Glass-Bearing Materials

We investigated a suite of impact glass-bearing rocks using a multi-analytical approach including visible-near-infrared diffuse reflectance spectroscopy, Mossbauer spectroscopy, and powder X-ray diffraction. In order to better understand and interpret the obtained results, we built a database containing physical, chemical, and spectroscopic information on glasses and glass-bearing materials using new results from this study and published works. We used the database to explore systematic relationships between parameters of interest and finally we applied several machine learning algorithms (support vector machine, random forests, and gradient boosting) to test the possibility to regress the oxidation state of iron from chemical and spectroscopic information. Our results show that even small amounts of mafic crystalline phases have a big influence on the spectral features of glass-bearing rocks. Samples without mafic crystalline inclusions show the typical spectrum of glasses (two broad and shallow bands roughly centered around 1,100 and 1,900 nm) with minor variations due to bulk chemistry. We described a non-linear relationship between average reflectance (average reflectance value between 500 and 1,000 nm), FeO + TiO2 content, grain size, and Fe3+/Fe-TOT. We tested the relation for the finer grain size (0-25 mu m), and we qualitatively assessed how it is affected by grain size, Fe3+/Fe-TOT, and crystal content. Finally, we developed a machine learning pipeline to regress the Fe3+/Fe-TOT of glass-bearing materials using the proposed database. Our machine learning calculations give satisfactory results (MAE: 0.0321) and additional data will enable the application of our computational strategy to remotely acquired data to extract chemical and mineralogical information of planetary surfaces.