Improving seismic fault mapping through data conditioning using a pre-trained deep convolutional neural network: A case study on Groningen field
Seismic fault interpretation is a crucial and indispensable step in reservoir exploration that requires substantial time. As a result, much research has been dedicated to applying deep learning in this venture. Deep learning has shown significant progress in the identification of seismic faults. How...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Published: |
Elsevier B.V.
2022
|
| Online Access: | https://www.scopus.com/inward/record.uri?eid=2-s2.0-85127071545&doi=10.1016%2fj.petrol.2022.110411&partnerID=40&md5=163d700ea6710c065c7de00a7579fb64 http://eprints.utp.edu.my/33058/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Seismic fault interpretation is a crucial and indispensable step in reservoir exploration that requires substantial time. As a result, much research has been dedicated to applying deep learning in this venture. Deep learning has shown significant progress in the identification of seismic faults. However, its applicability has been hindered given the lack of appropriate labelled fault data and poor seismic imaging. The deep convolutional neural network (DCNN) employed in this study is a cutting-edge deep learning technique for image improvement and identification. In this study, detecting seismic faults and improving seismic imaging using a pre-trained DCNN is proposed using the Groningen field that has complexity in accurately imaging sub-salt geological structures as a case study. The presence of salt in the field causes wave attenuation. The fault mapping procedure is considered a segmentation of the 3D seismic problem and trains an encoder-decoder architecture, using a Deep Residual U-net, to generate a fault probability volume. A decent fault prediction result is achieved on the Groningen seismic volume. The next step of this study involved seismic data conditioning, where DCNN is applied to denoise volumes to improve visualisation, significantly below the salt structure. The DCNN, when used to improve seismic imaging, achieved a signal-to-noise ratio (SNR) of 30.2, which is almost quadruple that of the original volume. Faults were mapped on the conditioned volume using DCNN, resulting in an improved seismic fault probability volume. This technique achieved distinctively interpreted faults illustrating the significant improvements DCNN brings to the seismic imaging process. In interpreting new seismic volumes with poor imaging and low signal-to-noise ratio, caused by a change in seismic frequency and amplitude propagating through an attenuating medium like the Groningen field, researchers and geophysicists may apply the DCNN for volume conditioning before mapping seismic faults using neural networks. © 2022 Elsevier B.V. |
|---|