Initial Major Element Quantification Using SuperCam Laser Induced Breakdown Spectroscopy
Résumé
SuperCam uses Laser Induced Breakdown Spectroscopy (LIBS) to collect atomic emission spectra from targets up to ~7 meters from the Perseverance rover. Due to the complexity of LIBS physics and the diversity of geologic materials, we use an empirical approach to major element (SiO2, TiO2, Al2O3, FeOT, MgO, CaO, Na2O, K2O) quantification, based on a suite of 1198 SuperCam laboratory spectra of 334 standards, including the rover calibration targets. SuperCam LIBS spectra are pre-processed by subtracting "dark" (passive/non-LIBS) spectra, denoising, continuum removal, instrument response correction, conversion to radiance, and wavelength calibration. For quantification, the spectra are masked to remove noisy sections of the spectrum and normalized by dividing signal in each spectrometer by the total signal from that spectrometer. We also found that the additional preprocessing steps of peak binning and/or per-channel standardization improved the results in some cases. These data are used to train multivariate regression models, with parameters optimized using cross-validation to avoid overfitting. We considered a variety of regression algorithms including Partial Least Squares (PLS), Least Absolute Selection and Shrinkage Operator (LASSO), Ridge, Elastic Net, Support Vector Regression (SVR), Random Forest (RF), Gradient Boosting Regression (GBR), Local Elastic Net, and blended sub-models. Models were selected based on test-set performance, accuracy of predictions of the onboard calibration targets, comparison of Mars and laboratory spectra, and the geochemical plausibility of Mars results. In some cases we found that the average of the predictions of several algorithms gave better results than any single method. Accuracy of predictions is estimated as the root mean squared error of prediction (RMSEP) for the test set. As additional spectra are collected from Mars, we continue to validate and improve upon this initial SuperCam elemental quantification. Areas of investigation include calibration transfer, probabilistic regression methods, and regression models for additional elements.Figure 1: Test set predictions vs actual compositions for each major element. Perfect predictions would fall on the line. RMSEP measures the accuracy of the model in wt.%.