Seismic inverse modeling method based on generative adversarial networks

• Published in: Journal of petroleum science & engineering Vol. 215; p. 110652
• Main Authors: Xie, Pengfei, Hou, Jiagen, Yin, Yanshu, Chen, Zhangxin, Chen, Mei, Wang, Lixin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2022
• Subjects: Generative adversarial network; MCMC; Reservoir Modeling; Seismic Inversion; Variogram Modeling; Reservoir Modeling; MCMC; Seismic Inversion; Variogram Modeling; Generative adversarial network
• ISSN: 0920-4105; 1873-4715
• DOI: 10.1016/j.petrol.2022.110652

Seismic inverse modeling is a common method in reservoir architecture characterization associated with geology. The conventional seismic inversion method is difficult to combine with complicated and abstract knowledge on geological modes, and its uncertainty is difficult to be assessed. In this paper, an inversion modeling method based on a wasserstein generative adversarial networks with gradient penalty (WGAN-GP) is introduced to integrate geology, well data and seismic data. A WGAN-GP is a generation model algorithm based on generative adversarial networks (GANs) that extracts spatial structure and abstract features of a training image. A gradient penalty function is added to an original loss function of GANs to improve the robustness. After assessment of the loss function, variograms and connectivity functions, the trained network is applied to seismic inversion simulation. In inversion, an optimal model is selected by the Metropolis with Markov chain Monte Carlo algorithm. Results show that the trained network can reproduce a thousand models containing millions of grid cells with a specific mode similar to a training image in 1 s. The inversion models conform to well data with 100% accuracy and have an efficient correspondence with prior seismic data. A whole inversion process completes 360,000 iterations in 4 h. The optimal inversion model has a subequal Root Mean Square Error (RMSE) with the true model and visually resembles a channel. With the proposed method, geological knowledge has a stable characterization in model realizations. •An inversion workflow is proposed to bridge a gap between geophysics data and geological knowledge.•The workflow is composed of a machine learning algorithm and a post-stack seismic inversion.•The WGAN-GP algorithm has a stable characterization on geological patterns.•Markov chain Monte Carlo (MCMC) sampling with GPUs speeds up the inversion.•The approach provides a way to decrease uncertainty and deal with the high-dimensional problem.