Comparison of popular Generative Adversarial Network flavours for fluvial reservoir modelling

This paper presents a commparative study of popular generative adversarial network flavours for fluvial reservoir modelling. Fluvial reservoir contains complex sedimentary architecture and facies heterogeneity. To better represent the real complexity of fluvial reservoir, we used a process-based model, FLUMY, to generated a low NTG fluvial dataset. Low NTG systems have less amalgamation and more complex sand-body connectivity. We simplify this dataset by reducing the number of facies. Thus, the training dataset for testing different GANs is composed of three-facies process-based 2D realizations. Candidates in this study include DCGAN, WGAN, WGAN-gp and PatchGAN. Visual comparison of the GANs generations found that some GANs have difficulties in learning certain key geolgocial features from this training dataset. PatchGAN outperforms in the aspect of channel geometry and facies placement. However, there are still some geologically unrealistic featues remaning. For example, the 'closed channel' pattern. Future efforts in GAN-based modelling would benefit from tackling complex geological systems.