Two-Dimensional Stochastic Structural Geomodeling with Deep Generative Adversarial Networks

Abstract Structural geological modeling is aimed at finding a representation of geological units. This is a complex ill-posed problem, and the data may be sparse and of varying quality, leading to multiple geological models consistent with them. Despite continuous advances for decades in geological modeling, recent studies still show some unresolved industrial challenges. Consequently, this work aims to improve geological modeling by proposing a novel approach with a focus on improving uncertainty management. A stochastic approach is developed based on deep generative methods, namely generative adversarial networks (GANs). Thanks to a synthetic dataset, a GAN is trained to generate two-dimensional unconditional geomodels that are plausible. Subsequently, a Bayesian inversion is performed with a Metropolis-adjusted Langevin algorithm (MALA) to produce geomodels consistent with field data. The proposed approach is validated on different conditioning data. For each case, the approach is able to successfully produce a variety of geomodels that can be linked to different geological settings. The uncertainties on geological units are measured by Shannon entropy on the generated models.