Unsupervised Machine Learning Workflow for Identifying Micro-Resistivity Borehole Image Features

Abstract The interpretation of micro-resistivity borehole image data is essential for geomechanical and geological analysis in the oil and gas industry but is often hindered by time-consuming manual processes. This study presents a novel unsupervised clustering workflow that leverages advanced deep learning models, including DINO for image feature extraction and LaMa for inpainting. The workflow combines several steps: borehole image inpainting, extracting features, normalizing, reducing dimensions, and clustering. Our findings show that this approach significantly improves the accuracy of borehole image restoration, as confirmed by metrics like RSNR, SSIM, and VSA. Furthermore, the workflow accurately identifies borehole features, such as breakouts and fractures validated by ARI, NMI, and F1 scores. The workflow also demonstrates strong adaptability across different borehole images. The significance of this research lies in its potential to enhance borehole image interpretation, offering a scalable and efficient solution that enhances operational decision-making and geomechanical model calibration.