Monitoring is an important task of any carbon sequestration process to ensure efficient trapping by examining the behaviors of injected plume bodies after injection, which includes inferring some indicative properties that reflect the property contrast between CO2 and the hosting environment from recorded seismic signals. In this study, we introduce a novel deep learning algorithm to establish the nonlinear mapping between the depth domain property contrast and the time domain seismic response to CO2 injection and plume body migration using a fully processed baseline data and a monitoring dataset. The trained network is able to instantaneously estimate the subsurface property change for any new monitoring dataset without conventional velocity model building and imaging. A deep learning architecture with a new multi-branch design with different filtering sizes is implemented for better feature extraction off the dipping events within the seismic gathers. In addition, a customized binary cross-entropy loss function is used to tackle the imbalanced training labels. We demonstrate the successful application of this deep learning algorithm on mapping the velocity contrast change of CO2 plume bodies from shot gathers synthesized based on the original Sleipner field datasets.

This content is only available via PDF.
You can access this article if you purchase or spend a download.