Reducing Time-Lapse Seismic Data Acquisition Cost with Deep Learning

Abstract

Abstract Time-lapse (4D) seismic is one of the most important tools for monitoring of subsurface reservoir production. However, repeated monitoring data is time-consuming and expensive. To reduce monitoring data acquisition cost and the subsequent 4D data processing, we developed three novel deep learning powered algorithms to solve challenging problems including sparse monitoring seismic data reconstruction, simultaneous source shooting monitoring data deblending, and rapid subsurface property estimation directly from very sparse pre-migration seismic data. These algorithms exploit the subsurface information buried in the baseline data and then transfer the learned knowledge to the monitoring data, which narrows down the sampling space of deep learning network to circumvent the network generalization challenge for optimal network inferencing. Customized network architectures and special learning strategies were designed to overcome specific challenges in each individual application. For sparse data reconstruction, we adopted a self-supervised learning approach combined with two-pass learning workflow to improve computation efficiency. A two-stage network was developed to enhance the deblending accuracy while preventing the non-uniqueness issue. A unique deep learning method was proposed to estimate the subsurface property change due to production or other activities such as CO2 sequestration directly from a few pre-migration seismic gathers without being contaminated by strong artifacts that are often observed using conventional imaging or inversion algorithms to deal with sparse gathers. The numerical experiments demonstrated the accuracy and efficiency of these deep learning algorithms on both synthetic data and field data.