Quantification of mineral reactivity using machine learning interpretation of micro-XRF data

Julie J. Kim, Florence T. Ling, Dan A. Plattenberger, Andres F. Clarens, Catherine A. Peters

Research output: Contribution to journalArticlepeer-review

8 Scopus citations

Abstract

Accurate characterizations of mineral reactivity require mapping of spatial heterogeneity, and quantifications of mineral abundances, elemental content, and mineral accessibility. Reactive transport models require such information at the grain-scale to accurately simulate coupled processes of mineral reactions, aqueous solution speciation, and mass transport. In this work, millimeter-scale mineral maps are generated using a neural network approach for 2D mineral mapping based on synchrotron micro x-ray fluorescence (μXRF) data. The approach is called Synchrotron-based Machine learning Approach for RasTer (SMART) mapping, which reads μXRF scans and provides mineral maps of the same size and resolution. The SMART mineral classifier is trained on coupled μXRF and micro-x-ray diffraction (μXRD) data, which is what distinguishes it from existing mapping tools. Here, the SMART classifier was applied to μXRF scans of various sedimentary rock samples including consolidated shales from the Eagle Ford (EFS1), Green River (GRS1), Haynesville (HS1), and New Albany (NAS1) formations and a syntaxial vein from the Upper Wolfcamp formation. The data were obtained using an x-ray microprobe at beamline 13-ID-E at the Advanced Photon Sources. Individual mineral maps generated by the SMART classifier well-captured distributions of both dominant and minor phases in the shale rocks and revealed EFS1 and GRS1 to be carbonate rich shales, and NAS1 and HS1 to be sulfide rich shales. The EFS1 was further characterized for its trace mineral abundances, grain sizes, trace element composition, and accessibility. Approximately 4.4 wt% of the rock matrix were found to be pyrite, with a median grain size of 3.17 μm in diameter and 62% of the grains predicted to be smaller than 4 μm. Quantifications of trace elements in pyrite revealed zinc concentrations up to 4.2 wt%, along with minor copper and arsenic copresence. Mineral accessibility was examined by contact with other phases and was quantified using a new type of image we are calling an adjacency map. Adjacency analyses revealed that of the total pyrite surface present in the EFS1, 28% is in contact with calcite. The adjacency maps are useful for quantifying the likelihood that a mineral could be exposed to fluids after dissolution of a contacting reactive phase like calcite. Lastly, pooling data from different samples was demonstrated by training a classifier using two sets of coupled μXRF-μXRD data. This classifier yielded an overall accuracy of >96%, demonstrating that data pooling is a promising approach for applications to a wide suite of rock samples of different origin, size, and thickness.

Original languageEnglish (US)
Article number105162
JournalApplied Geochemistry
Volume136
DOIs
StatePublished - Jan 2022
Externally publishedYes

All Science Journal Classification (ASJC) codes

  • Environmental Chemistry
  • Pollution
  • Geochemistry and Petrology

Keywords

  • Machine learning
  • Mineral classifier
  • Reactive transport modeling
  • SMART mineral mapping
  • Shale
  • Synchrotron XRF

Fingerprint

Dive into the research topics of 'Quantification of mineral reactivity using machine learning interpretation of micro-XRF data'. Together they form a unique fingerprint.

Cite this