Time-lapse data matching using neural networks with multiple reflections

Shang Huang, Daniel O. Trad

Time-lapse seismic is a complex problem in reservoir monitoring because near-surface noise, poor subsurface illumination, and the inherent weak amplitude of reservoir changes affect the quality of the interpretation. Because of its importance, geophysicists have tried many approaches to solve these problems. In this chapter, I apply deep learning methods to address these challenges. Firstly, a stacked long short-term memory (SD-LSTM) neural network adapts the near-surface baseline data to the near-surface monitor data. This assumes that differences from the near-surface are not due to changes in the reservoir but differences in the seismic experiment (acquisition and processing). A U-Net is then followed to work on differences between monitor and baseline images to suppress noise on a large scale. Furthermore, the energy from surface multiples is added during migration and shot record generation in the forward modelling step to increase subsurface illumination.A double-difference method is applied to the predicted and observed data to give a final difference. The results show that the SD-LSTM can anticipate and mitigate noise in the monitor data. The final difference between baseline and monitor models has suppressed significant noise after combining SD-LSTM, U-Net and surface multiples. The proposed method is also tested in a field dataset, DAS VSP data from the CaMI FRS project, with extended bidirectional SD-LSTM and convolutional neural networks (CNN). The output provides meaningful information and prediction for CO₂ injection migration within a tar-get area, which matches the CO₂  DocumentGo Back