Physics-informed W-Net GAN for the direct stochastic inversion of fullstack seismic data into facies models

Predicting the spatial distribution of geological facies in the subsurface from fullstack geophysical data is a main step in the geo-modeling workflow for energy exploration and environmental tasks and requires solving an inverse problem. Generative adversarial networks (GAN) have shown great potential for geologically accurate inverse modeling, although with limitations in computational costs and in accounting for uncertainty in the prediction of facies-dependent properties. To overcome this limitation, we propose a GAN architecture for multivariate inverse modeling, which is able to learn the physics-based mapping between facies and seismic domains, and account for the spatial uncertainties of the facies and elastic properties. In a single training stage, the network models a distribution of realistic facies patterns solving a seismic inversion problem, based on the observed data, and learned features. The method is first demonstrated on 2-D application examples, and then applied for the inversion of a 2-D seismic section extracted from the Norne field (Norwegian North Sea). The results show that through fast training, the proposed GAN can model facies distributions fitting the observed data, reproducing the prior facies patterns and the data uncertainty, while honoring the physics of the system under investigation.