A stable deep adversarial learning approach for geological facies generation

Simulation of geological facies in an unobservable volume is essential in various geoscientific applications. Given the complexity of the problem, deep generative learning is a promising approach to overcome the limitations of traditional geostatistical simulation models, in particular their lack of physical realism. This research studies the application of generative adversarial networks and variational Bayesian inference for the conditional simulation of channelized reservoirs in subsurface volumes. In this paper, we review generative deep learning approaches, in particular adversarial networks and stabilization techniques that aim to facilitate their learning. We also study the problem of conditioning these models to observations by a Bayesian variational approach, comparing a conditional neural network model to a Gaussian mixture model. The proposed approach is tested on 2D and 3D simulations generated by the Flumy model based on stochastic modeling of sedimentary deposition processes. Morphological metrics are used to compare the proposed method with different generative adversarial network architectures. The results indicate that by using recent stabilization techniques, generative adversarial networks can efficiently sample complex target data distributions.

https://www.sciencedirect.com/science/article/pii/S0098300424001213#d1e4764