3D Geological Image Synthesis from 2D Examples Using Generative Adversarial Networks

Abstract

Summary Recently, Generative Adversarial Networks (GAN) have been proposed as a potential alternative to Multipoint Statistics (MPS) to generate stochastic fields from a large set of training images. A difficulty for all the training image based techniques (including GAN and MPS) is to generate 3D fields when only 2D training data sets are available. In this paper, we introduce a novel approach called Dimension Augmenter GAN (DiAGAN) enabling GANs to generate 3D fields from 2D examples. The method is simple to implement and is based on the introduction of a random cut sampling step between the generator and the discriminator of a standard GAN. Numerical experiments show that the proposed approach provides an efficient solution to this long lasting problem.