DGGV-E-Publikationen

Titel: Automated heavy mineral analysis of silt-sized sediment by artificial-intelligence guided Raman Spectroscopy

Autoren:
Nils Keno Lünsdorf1, Jan Ontje Lünsdorf3, Gábor Újvári2, Hilmar von Eynatten1

Institutionen:
1Georg-August-Universität Göttingen, Department of Sedimentology and Environmental Geology, Göttingen, Germany; 2Institute for Geological and Geochemical Research, Research Centre for Astronomy and Earth Sciences, Budapest, Hungary; 3Insterburger Strasse 2, 26127, Oldenburg

Veranstaltung: GeoKarlsruhe 2021

Datum: 2021

DOI: 10.48380/dggv-qw19-8q13

Zusammenfassung:
Compositional data on heavy minerals is fundamental in sedimentary provenance analysis. Typically, this data is gathered by optical microscopy and more recently, by mineral chemical analysis (MLA, QEMSCAN) or Raman micro-spectroscopy. In silt-sized sediments optical microscopy is unfeasible. We introduce a systematic and highly efficient approach to assess the heavy mineral composition in fine grain-size fractions (10-30 µm and 30-62 µm) by Raman micro-spectroscopy.

The approach starts with a web-application that creates and visualizes large mosaic images from which arbitrary objects can be selected for training and inference of a region-based convolutional neural network (R-CNN). Here, mineral grains are automatically selected by passing the tiles of a mosaic image of the sample slide into the R-CNN. For each detected grain a polygon is computed from which positional and optical parameters are derived. Using this polygon data, the measurement parameters at the Raman spectrometer are individually set to account for varying Raman scattering intensities and irradiation resistivity. After the compositional data is obtained, Raman spectra are evaluated and further single-grain geochemical methods (ICPMS, EMPA) can be applied to the identified and referenced grains (e.g. U-Pb dating of zircon).

The method was tested on 13 samples from three loess profiles from Germany and Hungary. About 100.000 minerals were analyzed and provenance signals demonstrate clear contrasts between the sections. Being automated, this approach allows for analyzing large sample numbers with higher precision (i.e. counting statistics) on silt-sized materials, thus opening new avenues in sedimentary provenance analysis.