Automatic reconstruction of geological reservoir models based on conditioning data constraints and BicycleGAN

Abstract

For geological reservoir units with different sizes of pore spaces and relatively stronger anisotropic heterogeneities, Generative-Adversarial-Network-based (GAN) modeling methods can overcome limitation of numerical-simulation-based ones and support finely representation of nonstationary models. However, due to conditioning data weak constraints, GANs might ignore local detail features, with artifacts and noise during simulation. Some pre-processing strategies are strictly limited by the conditioning data distribution patterns, and corresponding processes should be updated frequently when new data are introduced, with relatively complicated operation. Therefore, the BicycleGAN framework was introduced in this paper. Based on bijective consistency between the latent vector and output, an image-to-image translation task with different dimensions is established, and multiple networks were coupled to realize finely representation of spatial attributes. Specifically, the mapping between conditioning data and output is established through the U-Net architecture, which not only local detail information is considered, but also reduce the impact of conditioning data distribution patterns. Meanwhile, the mapping between latent and actual test time distribution is also established by an encoding function to ensure simulation authenticities. In addition, a joint loss function combined with conditioning loss and prior loss is defined to ensure reconstruction accuracy of different facies types. Four kinds of categorical and continuous training images were selected to verify the network simulation performances. Results show that reconstruction accuracy for facies types of conditioning data is almost consistent with those of the references, keeping similarities in terms of spatial variability, connectivity, structural similarity, and facies type reproductions. Meanwhile, for the saved BicycleGAN, inputting different kinds of conditioning data once only, and a variety of simulations can be obtained rapidly, thereby realizing conditioning reconstruction of geological reservoir models.