De-risking overburden and caprocks for CO2 storage using machine-learning seismic fault attributes

Abstract

Fault and fracture geometries, densities, and distributions play a critical role in assessing and mitigating risks associated with potential CO₂ storage sites in sedimentary basins, particularly saline aquifers. To enhance fault detection in 3D seismic data, we have developed, trained, and deployed a lightweight machine learning segmentation algorithm. This deep learning model, trained on synthetic seismic data, generates fault scores—pixel-scale classifications ranging from 0 to 1—where higher values indicate a greater likelihood of structural discontinuities. These fault scores are used to derive a fault density attribute, which summarizes the expected fault network distribution along seismic sections. Our workflow is computationally efficient and provides interpreters with valuable insight into the lateral and vertical distribution of faults. We apply this methodology to a 3D seismic survey of the Smeaheia area, Norway, covering the N-S trending Vette Fault and sections of the Øygarden Fault Complex (ØFC). Fault mapping was conducted at the reservoir level, as well as in the caprock and overburden. The detected fault patterns at the top of the Draupne Formation, the presumed caprock unit in the region, and fault pattern at the Top Cromer Knoll Group, align well with manual interpretations. Additionally, in the footwall of the deep-crustal ØFC, we identify faults extending to the seafloor, suggesting that a non-negligible fault density may be present within the caprock. Our results are compared with 3D variance and 3D semblance seismic attributes, further validating the efficacy of our approach.