Cost-effective CO₂ monitoring strategies powered by deep learning

Time-lapse (4D) seismic is an essential and reliable tool for monitoring geologically sequestered CO 2 in the subsurface. Unfortunately, 4D seismic is expensive due to repeated data acquisition and the complex procedures involved in 4D seismic data processing. To reduce the cost of time-lapse monitoring of injected CO 2 plume bodies, we have proposed three novel strategies powered by deep learning. Each of these strategies is driven by customized deep learning networks and specialized training schemes to meet both the requirements of monitoring accuracy and the constraints of data availability. These strategies address challenging problems at various stages of time-lapse seismic monitoring, ranging from seismic data acquisition to data processing. Specifically, the first strategy was designed to reconstruct sparsely acquired seismic data into a complete dense data set without losing critical subsurface information. The resulting reconstructed dense data set can then be fed into conventional seismic processing modules without suffering from the aliasing issue. Consequently, the first strategy does not reduce the cost of 4D data processing. The second strategy attempts to bypass time-consuming conventional physics-based seismic data processing procedures by leveraging the powerful nonlinear mapping capabilities of neural networks, which allows us to use sparser data for CO 2 monitoring in situations where data reconstruction is unfeasible due to a lack of coherence in these sparser data sets. Therefore, this strategy reduces data processing cost but is more sensitive to noise. The third strategy is a unique workflow to further cut monitoring costs by replacing 3D seismic data with multiple 2D lines combined with optional point seismic measurements to reconstruct full 3D seismic image volumes or 3D subsurface property models for economical CO 2 monitoring. While this strategy features the most cost-effective data acquisition schemes, it is important to note that high-wavenumber structural information between sparse 2D lines cannot be fully recovered.