Machine Learning Mid-Infrared Spectral Models for Predicting Modal Mineralogy of CI/CM Chondritic Asteroids and Bennu

Planetary surfaces can be complex mixtures of coarse and fine particles that exhibit linear and nonlinear mixing behaviors at mid-infrared (MIR) wavelengths. Machine learning multivariate analysis can estimate modal mineralogy of mixtures and is favorable because it does not assume linear mixing across wavelengths. We used partial least squares (PLS) and least absolute shrinkage and selection operator (lasso), two types of machine learning, to build MIR spectral models to determine the surface mineralogy of the asteroid (101955) Bennu using OSIRIS-REx Thermal Emission Spectrometer (OTES) data. We find that PLS models outperform lasso models. The cross-validated root-mean-square error of our final PLS models (consisting of 317 unique spectra of samples derived from 13 analog mineral samples and eight meteorites) range from ∼4 to 13 vol% depending on the mineral group. PLS predictions in vol% of Bennu's average global composition are 78% phyllosilicate, 9% olivine, 11% carbonates, and 6% magnetite. Pyroxene is not predicted for the global average spectrum, though it has been detected in small amounts on Bennu. These mineral abundances confirm previous findings that the composition of Bennu is consistent with CI/CM chondrites with high degrees of aqueous alteration. The predicted mineralogy of two previously identified OTES spectral types vary minimally from the global average. In agreement with previous work, we interpret OTES spectral differences as primarily caused by relative abundances of fine particulates rather than major compositional variations.