Abstract: Learning data shifts in spectroscopy via structured normalizing flows
Since landing at Gale Crater in 2012, the ChemCam laser-induced breakdown spectroscopy (LIBS) instrument onboard the Mars rover Curiosity has obtained spectral measurements of thousands of rock and soil analysis targets. The compositions of the major elements may be predicted using models trained on samples with known compositions measured by a laboratory instrument under Mars-like atmospheric conditions. However, laboratory measurements and rover measurements on identical sets of calibration targets still display some notable differences, prompting development of an Earth-Mars correction. Currently, this correction is computed using the ratios of Mars and laboratory spectra on a few of the calibration targets, but this calculation is sensitive to the small amount of data and may not generalize well on new samples. In this work, we explore the Earth-Mars spectral difference in a probabilistic framework by investigating how the spectral probability densities differ on Mars and Earth. On a structured latent space obtained by standard dimension reduction methods, we construct a composition operator with Normalizing Flows that learns how the latent spaces for Mars spectral data and Earth spectral data differ in distribution. We arrive at a structured approach for learning spectral data shifts between Earth and Mars that leverages all of the available measurements and provides a richer and potentially more robust and interpretable Earth-Mars correction.
