.png){kind=icon}
.png){kind=icon}
Return
Porosity prediction based on a structural modeling deep-learning method
.png){kind=icon}
PRE
AI
DOI：10.1190/GEO2024-0035.1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The seismic inversion method is a pivotal approach for acquiring porosity parameters. The performance of inversion depends on several factors, such as algorithm stability, the precision of the rock-physics model, and the quality of seismic data. Some deep-learning techniques successfully address the aforementioned challenges of physical approximations in a data-driven manner, offering an intelligent approach to establish the intricate nonlinear relationship between seismic data and porosity. However, these methods mainly follow a trace-by-trace inversion paradigm that fails to consider the structural attributes of geologic strata, thereby overlooking the inherent spatial continuity of the seismic data. We extend the conventional intelligent prediction algorithm from 1D to 2D and develop a quantitative porosity prediction method based on structural modeling deep learning (SMDL). The designed framework simulates the generalized seismic inversion, including an inversion processor (IP) and a forward processor (FP). First, the modeling of multivariate information from well logs and seismic horizon data generates 2D seismogram and porosity models, which serve as the training data set. Subsequently, we introduce an improved TransUNet in the IP to map seismograms to porosity. The predicted porosity is then input into an FP based on an improved UNet to simulate the seismogram. Finally, the network system uses the mean square error within the processor to achieve a dual-scale constraint for intelligent porosity prediction tasks. To ascertain the feasibility and robustness of SMDL, we conduct validation experiments using the SEAM model. In addition, we apply SMDL to field data from the South China Sea, demonstrating that the SMDL can further improve the lateral continuity and accuracy of predicted results. Moreover, horizon slices of predicted results based on different methods are extracted, and the porosity interpretation slice of SMDL shows high congruence with actual geologic knowledge and drilling situations.
Keywords:
PRESTACK SEISMIC DATA
JOINT ESTIMATION
NEURAL-NETWORKS
INVERSION
IMPEDANCE
RESERVOIR
SATURATION