Advanced CO₂ Interpretation from 4D Sleipner Seismic Images using Swin-Unet3D

Luping Qu, Kristopher A. Innanen

The interpretation of CO₂ is important for monitoring the storage status of CO₂ and the risk of leakage in Carbon Capture, Utilization, and Storage (CCUS). Traditional manual interpretation of imaging dataset, while informative, is labor-intensive and often suffers from inconsistency over the extended periods of monitoring. This inconsistency largely stems from the inevitable evolution of seismic acquisition and processing technologies, as well as the subjectivity inherent in manual interpretation methods. 3D convolutional neural networks (CNNs) have seen considerable applications in object detection within seismic imaging, achieving notable success. Yet, their design constraints, specifically the limited size of convolutional kernels, have resulted in an inherent limitation in capturing long range dependencies within the data. While Vision Transformers (ViT) excel in learning such long-distance dependencies, they are burdened by a high parameter count and struggle with local dependency information in data-scarce scenarios.In response to these challenges, this study introduces the Swin-Unet3D model, innovatively adapted for CO₂ sequestration monitoring. This model reimagines voxel segmentation in geological imaging as a sequence-to-sequence prediction task. Its novel feature extraction sub-module is a hybrid architecture that combines the strengths of both Convolution and ViT. This parallel structure ensures comprehensive learning of both global andlocal dependency information within the image. The model, which is trained, validated,and tested using the Sleipner CO₂ storage project’s time-lapse dataset spanning from 1984to 2010, marks an improvement in CO₂ interpretation.