A MACHINE LEARNING APPROACH TO PREDICTING MULTI-COMPONENT MINERAL COMPOSITIONS FOR SPACECRAFT X-RAY DIFFRACTION INSTRUMENTS

To better understand the mineral forming environments, geologic history, and potential past or present habitability of Mars, previous studies [1,2] exploited the relationships between ionic radii and unit-cell dimensions to estimate mineral compositions on the Martian surface. Due to the complexity of the multi-component systems, we limited the number of chemical elements to 3 or fewer. We have expanded our approach by incorporating a machine learning framework, Label Distribution Learning (LDL) [3,4], to predict complex, multi-element mineral compositions. LDL is a novel framework for classification problems where every observation is associated with several labels and a value for each representing the degree to which that label describes the observation. Since mineral groups (e.g. feldspars, olivines, pyroxenes) differ in their compositional components we train multiple models using the unit-cell dimensions of samples in the dataset as measured by an XRD instrument to predict the chemical composition of the sample in terms of its ionic components. We have found these LDL models to exceed in predictive performance the analogous methods used in previous studies and present here some of our findings and plans for future work.

References:

[1] Morrison et al. (2018) American Mineralogist, 103(6), 857-871

[2] Morrison et al. (2018) American Mineralogist, 103(6), 848-856

[3] X. Geng. (2016) IEEE Transactions on Knowledge and Data Engineering (IEEE TKDE)

[4] X. Geng at al. (2013) IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI)