Physics-Informed Machine Learning to Identify Features in LCROSS NIR Data

Introduction: The main goal of the Lunar Crater Observation and Sensing Satellite (LCROSS) mission was to determine the presence and composition of ice in a permanently shadowed region (PSR). Near the lunar poles, permanently PSRs provide stable cold traps where temperatures are consistently below 120 K, allowing for the accumulation and long-term retention of volatile species, including water ice. LCROSS targeted one such PSR near the Moon’s South Pole, the Crater Cabeus (84.9°S35.5°W) with average temperatures of about 37 K, making it a prime location to detect ice deposits.
The LCROSS mission executed a controlled impact experiment. It was comprised of two distinct impact events. The first, the spent ATLAS V upper-stage Centaur rocket from the co-manifested Lunar Reconnaissance Orbiter (LRO). Four minutes later, the LCROSS shepherding-spacecraft (S-S/C) struck the Moon; the velocity of both impacts was roughly 2.5km/s. The S-S/C transmitted real-time observations from approximately four minutes before the Centaur’s impact until nearly one second prior to the spacecraft’s own collision. Descending through the Centaur’s impact ejecta plume allowed visible and near-infrared (NIR) spectrometer and cameras to acquire data to characterize the plume’s morphology, evolution, and composition. Colaprete et al. applied a stepwise linear fitting approach to the LCROSS NIR spectra, modeling total absorbance as a linear combination of the manually selected species-specific spectral contributions, using a reduced chi-squared analysis to assess goodness-of-fit.