GAN learning complex fluvial facies distribution from process-based modelling

Summary This work shows the advantages and disadvantages of modelling complex geological models using generative adversarial networks (GAN). A process-based model, FLUMY, is used to create the training dataset. Compared to previous work in this area, this dataset contains varied geo-body geometry, asymmetrical channel sinuosity and irregular meander morphology. In short, this training dataset is closer to the real complexity of fluvial reservoir. The results indicate our GAN can capture complex multi-facies distribution, their relationships, and facies geometry. However, the GAN generated realizations contain many geologically unrealistic features. In this paper, we list two types of unrealistic features, named 'mislabelling' and 'incorrect channel-levee relationship'. Two proposed methods are proved that they can reduce the amount of the unrealistic features. Embedding one-hot encoder in GAN can cure the 'mislabelling' issue. Multi-discriminator strategy is helpful to assist GAN to learn spatial relationships among different facies better.