Machine learning applications on lunar meteorite minerals: From classification to mechanical properties prediction

Abstract

Amid the scarcity of lunar meteorites and the imperative to preserve their scientific value, non-destructive testing methods are essential. This translates into the application of microscale rock mechanics experiments and scanning electron microscopy for surface composition analysis. This study explores the application of Machine Learning algorithms in predicting the mineralogical and mechanical properties of DHOFAR 1084, JAH 838, and NWA 11444 lunar meteorites based solely on their atomic percentage compositions. Leveraging a prior-data fitted network model, we achieved near-perfect classification scores for meteorites, mineral groups, and individual minerals. The regressor models, notably the K-Neighbor model, provided an outstanding estimate of the mechanical properties—previously measured by nanoindentation tests—such as hardness, reduced Young’s modulus, and elastic recovery. Further considerations on the nature and physical properties of the minerals forming these meteorites, including porosity, crystal orientation, or shock degree, are essential for refining predictions. Our findings underscore the potential of Machine Learning in enhancing mineral identification and mechanical property estimation in lunar exploration, which pave the way for new advancements and quick assessments in extraterrestrial mineral mining, processing, and research.