IMPROVED MACHINE-LEARNING FAULT DETECTION AT THE SAN JUAN BASIN CARBONSAFE PROJECT SITE

Fault detection is crucial for site characterization in geologic carbon storage. Machine-learning fault detection is an emerging technique that has many advantages compared with conventional human fault detection, including high computational efficiency and the capability of detecting faults with tiny geologic formation displacements, etc. However, accurate migration images are also crucial for fault detection. Noisy migration images can result in false detection of faults. We improve migration images of the San Juan Basin CarbonSAFE project site using migration velocity analysis and prestack depth migration. We further remove image noise in migration images for reliable fault detection. We then perform fault detection on migration images using a nested residual neural network. Our results indicates that there are no major faults around the planned CO2 injection zone.