GSA Annual Meeting in Denver, Colorado, USA - 2016

Paper No. 197-1
Presentation Time: 8:05 AM


DYAR, M. Darby, Planetary Science Institute, 1700 East Fort Lowell, Suite 106, Tucson, AZ 85719-2395,

Spectral features, software, and techniques used for identification of major phases in planetary science applications have remained largely unchanged for decades, but such stasis is already being impacted by artificial intelligence algorithms that can learn without explicit instructions. These models may take us beyond our ability to simply relate spectral features to crystal structures. Major atomic emission, absorption energies, and stretching/bending modes are well-characterized and understood for most mineral groups, but subtle changes in those spectral peaks resulting from complexities of cation substitutions can be exploited by machine learning (ML). Minor features that are more difficult to relate to specific transitions/modes are generally neglected or unappreciated in existing paradigms, yet ML techniques can utilize information from those portions of the spectra as well. ML methods make use of a combination of both conspicuous and subtle features that can be rich in diagnostic information.

Critical for these new ML approaches will be creation of improved reference datasets that augment and mitigate some of the inadequacies of existing databases, which may have impure and/or poorly characterized samples that can confound automated approaches. Moreover, there are very few adequate spectral data on mixtures. Remedying these shortcomings will require major effort but the potential benefits are enormous. ML approaches including manifold alignment can assist with uniting disparate data and identifying/removing useless outliers.

Several types of spectroscopy have been shown to benefit from ML approaches. Baseline/continuum removal can now be automated and optimized for any data set or application. Fe X-ray absorption studies have long used the pre-edge feature in the FeKα absorption edge for Fe3+/ΣFe measurements, yet ML-identified features in the main edge produce far more quantitative results. Automated whole spectrum matching approaches dramatically improve the accuracy of mineral identification in Raman spectroscopy over single peak matching methods. Multivariate analysis is required for predicting laser-induced breakdown spectra in geologic materials. Finally, ML derivations of olivine and pyroxene compositions/ratios from reflectance outperform existing approaches.

  • GSA_2016_future_092716.pdf (5.2 MB)