Deep learning of SHERLOC Raman spectra for facilitating Mars Astromaterial identification.

Area of Science:

• Planetary Science
• Astrobiology
• Spectroscopy

Background:

• Raman spectroscopy is crucial for planetary exploration, particularly for identifying potential biosignatures and materials on Mars using instruments like SHERLOC.
• Current astromaterial identification relies heavily on expert analysis of comparative Raman spectra, which can be time-consuming and subjective.

Purpose of the Study:

• To benchmark machine learning (ML) and deep learning (DL) models for automated identification of mineral classes from SHERLOC Raman spectra.
• To evaluate the performance of DL models in identifying astromaterials under class-imbalanced conditions.

Main Methods:

• Benchmarking three ML and two DL models using expert-labeled SHERLOC Raman spectra from Mars missions (Crater Floor and Upper Fan campaigns).
• Applying the best-performing DL model to unlabeled and non-diagnostic spectra to assess its predictive consistency.

Main Results:

• The top-performing DL model achieved 89.3% accuracy in identifying six mineral classes.
• The DL model demonstrated competitive class-wise performance even with imbalanced datasets.
• Predictions on unlabeled spectra aligned with known material characteristics from existing literature.

Conclusions:

• Deep learning models show significant potential for high-performance, automated astromaterial identification from Raman spectra.
• DL methods can enhance spectral interpretation and support future planetary exploration missions.
• This study provides a foundation for integrating AI-driven spectral analysis in the search for extraterrestrial materials.