SMART mineral mapping: Synchrotron-based machine learning approach for 2D characterization with coupled micro XRF-XRD

A Synchrotron-based Machine learning Approach for RasTer (SMART) mineral mapping was developed for the purpose of training a mineral classifier for characterization of millimeter-sized areas of rock thin sections with micron-scale resolution. An Artificial Neural Network (ANN) was used to extract relationships between coupled micro x-ray fluorescence (mu XRF) data for element abundances and micro x-ray diffraction (mu XRD) data for mineral identity. Once trained, the resulting classifier, i.e., the SMART mineral mapper, can identify minerals using only mu XRF data. This is the real value of this machine learning approach because mu XRF data are relatively fast to collect and interpret whereas mu XRD data take longer to collect and interpret. Training and testing of an initial mapper were done with 192 coupled mu XRF-mu XRD data points sampled from a 0.25 mm(2) area of a shale from the Eagle Ford formation, which was scanned with 2 mu m resolution. All data used in this work were obtained from the Advanced Photon Source synchrotron beamline 13-ID-E at Argonne National Laboratory. Three minerals were mapped in the Eagle Ford rock sample, for which there were 8 elements characterized. In the testing phase, the minerals were correctly classified with accuracy of 97 % and higher. The trained SMART mapper was applied for self-similar upscaling by mapping a 14 mm(2) scan of the Eagle Ford sample. Generated maps captured micro-scale features characteristic of the stratified texture of the rock, and the identified minerals agreed well with bulk XRD analysis of the powdered rock. The SMART mapper was also applied to a scan of a 6 mineral mixture of known composition to demonstrate ability to distinguish minerals of similar chemistry. The trained SMART mapper is transferable to scans from other x-ray microprobes because of the mu XRF data normalization that accounts for sample-and beamline-specific properties like thickness, detector configuration, and incident energy.

Automated summary

The SMART mineral mapping approach developed in this study combines synchrotron-based machine learning, micro-scale XRF and XRD data analysis, and neural network techniques to accurately identify minerals in rock thin sections. The trained classifier was able to achieve a high accuracy rate in identifying minerals based solely on mu XRF data, demonstrating the efficiency and transferability of the SMART mapper in mineral mapping and characterization.