Automated Grain Segmentation and Mineral Classification in Rock Thin Sections

This technical report presents a machine learning solution for the automated characterization of minerals and grains in rock thin sections. Traditional petrographic analysis is time-consuming and subjective, often leading to operator bias. The proposed machine learning-based image analysis methods aim to enhance the efficiency of identifying individual grains and delineating mineralogy in thin sections. The methodology includes data preparation using QEMSCAN mineral maps, multi-scale grain segmentation, and mineral classification through a UNET model. The results demonstrate the capability to generate accurate mineral maps and grain boundaries from optical images, significantly improving the speed and reliability of petrographic analysis. The report discusses the challenges faced, including the limitations of current machine learning approaches in accurately classifying certain minerals, particularly clay. Overall, the findings indicate that the machine learning workflow can expedite the analysis process while maintaining a level of accuracy that, although not yet equivalent to experienced petrographers, offers substantial advantages for high-throughput screening.