Automatic Whole-Spectrum Matching Techniques for Identification of Pure and Mixed Minerals using Raman Spectroscopy

Abstract:

In situuse of Raman spectroscopy on Mars is planned for three different instruments in the next decade. Although implementations differ, they share the potential to identify surface minerals and organics and inform Martian geology and geochemistry. Their success depends on the availability of appropriate databases and software for phase identification. For this project, we have consolidated all known publicly-accessible Raman data on minerals for which independent confirmation of phase identity is available, and added hundreds of additional spectra acquired using varying instruments and laser energies.

Using these data, we have developed software tools to improve mineral identification accuracy. For pure minerals, whole-spectrum matching algorithms far outperform existing tools based on diagnostic peaks in individual phases. Optimal matching accuracy does depend on subjective end-user choices for data processing (such as baseline removal, intensity normalization, and intensity squashing), as well as specific dataset characteristics. So, to make this tuning process amenable to automated optimization methods, we developed a machine learning-based generalization of these choices within a preprocessing and matching framework. Our novel method dramatically reduces the burden on the user and results in improved matching accuracy.

Moving beyond identifying pure phases into quantification of relative abundances is a complex problem because relationships between peak intensity and mineral abundance are obscured by complicating factors: exciting laser frequency, the Raman cross section of the mineral, crystal orientation, and long-range chemical and structural ordering in the crystal lattices. Solving this un-mixing problem requires adaptation of our whole-spectrum algorithms and a large number of test spectra of minerals in known volume proportions, which we are creating for this project. Key to this effort is acquisition of spectra from mixtures of pure minerals paired with a standard reference (diamond); the ratios of their peak areas are independent of laser energy and normalization techniques and are analogous to molar absorptivities. With these test data and our advanced algorithms, quantitative estimates of mineral modes can be accurately derived from Raman spectra of mixtures.