Deep learning of SHERLOC Raman spectra for facilitating Mars Astromaterial identification

Raman spectroscopy has been explored for planetary exploration, particularly through the SHERLOC instrument for Mars habitability and astromaterial search tasks. The SHERLOC and other Raman-based astromaterial search tasks mostly rely on comparative Raman peak analysis and expert interpretation with respect to known spectral peaks. It is desirable to employ machine learning (ML) and deep learning (DL) methods for high-performance automated astromaterial identification from Raman spectra. In this work, for the first time, we benchmarked three ML and two DL models for the identification of six mineral classes from SHERLOC Raman spectra. Using the expert-labeled spectra of two campaigns (Crater Floor and Upper Fan), the best-performing DL model yielded identification accuracies of 89.3%, while maintaining competitive class-wise performance under class-imbalanced conditions. The best-performing DL model was further applied to non-diagnostic and unlabeled SHERLOC Raman spectra, and the resulting predictions were broadly consistent with material characteristics reported in the literature. These results suggest that DL methods may provide supportive information for automated Raman-based astromaterial identification and spectral interpretation in planetary exploration.