Combining Image Editing and SinGAN for Conditional Sedimentary Facies Modeling

Traditional sedimentary facies modeling using generative adversarial networks (GANs) usually requires extensive datasets for network training. However, obtaining large datasets that align with reservoir depositional characteristics is often complex and costly. This letter introduces a conditional generative adversarial network (CSinGAN) based on a single training image. CSinGAN does not use conditional data in the training phase. In the model generation stage, conditional facies simulation is achieved by adjusting the intermediate model using image editing techniques. The conditional realizations of the three sets of training images successfully matched the well data. The variogram function and connectivity function indicate that CSinGAN can generate heterogeneous structures that conform to the statistical characteristics of the training image. We used multiscale sliced Wasserstein distance to verify that the realizations of the CSinGAN outperform the classical multipoint geostatistical algorithms. This study demonstrates the viability of using a single training image in GANs for conditional sedimentary facies modeling.