Improving the training performance of generative adversarial networks with limited data: Application to the generation of geological models

Abstract

In the past years, there is a growing interest in the applications of Generative Adversarial Networks (GANs) to generate geological models. Although GANs have proven to be an effective tool to learn and reproduce the complex data patterns present in some geological models, some challenges still remain open. Among others, a well-noticed problem is the need for a large number of samples to ensure high-quality training, which can be prohibitively expensive in some cases. As an attempt to offer a (possibly partial) solution to the aforementioned challenge, in this study, we investigate the feasibility and effectiveness of a zero-centered discriminator regularization technique for improving the performance of a GAN. Additionally, we evaluate an adaptive data augmentation technique to overcome the potential issue of limited training data, for the purpose of generating geologically feasible realizations of hydrocarbon reservoir models. Our findings demonstrate that a combination of the two techniques lead to notable performance improvements of a GAN. Particularly, it is observed that using the adaptive data augmentation technique in a GAN can yield similar results to those obtained by the GAN with a much larger dataset.