Machine learning methods applied to combined Raman and LIBS spectra: Implications for mineral discrimination in planetary missions

The combined analysis of geological targets by complementary spectroscopic techniques could enhance the characterization of the mineral phases found on Mars. This is indeed the case with the SuperCam instrument onboard the Perseverance rover. In this framework, the present study seeks to evaluate and compare multiple machine learning techniques for the characterization of carbonate minerals based on Raman-LIBS (Laser-Induced Breakdown Spectroscopy) spectroscopic data. To do so, a Ca-Mg prediction curve was created by mixing hydromagnesite and calcite at different concentration ratios. After their characterization by Raman and LIBS spectroscopy, different multivariable machine learning (Gaussian process regression, support vector machines, ensembles of trees, and artificial neural networks) were used to predict the concentration ratio of each sample from their respective datasets. The results obtained by separately analyzing Raman and LIBS data were then compared to those obtained by combining them. By comparing their performance, this work demonstrates that mineral discrimination based on Gaussian and ensemble methods optimized the combine of Raman-LIBS dataset outperformed those ensured by Raman and LIBS data alone. This demonstrated that the fusion of data combination and machine learning is a promising approach to optimize the analysis of spectroscopic data returned from Mars. This study examined Raman and LIBS spectroscopy and their combined data for mineral quantification. Despite LIBS models showed higher dispersion, they exhibited superior precision in RMSEP compared to Raman models. Furthermore, LIBS was effective in detecting extreme concentrations. Spectral parameter-based models consistently outperformed PCA-based models, yielding lower RMSEP and higher R2 values. Notably, the combination of Raman-LIBS data significantly improved the results, especially when employing Gaussian and neural network models. This advancement holds implications for planetary exploration missions. image