| Abstract |
Characterization of subsurface reservoirs often requires geological facies models to identify areas with favorable rock properties. With the development of computing power, deep learning approaches, such as generative adversarial networks (GAN), have become widely used for simulating complex geological models. However, training of the GAN typically requires a large quantity of training data for updating neural parameters. This process is generally performed using traditional geostatistical methods based on multiple-point statistics, object-based, or process-based models to build the training data. In this study, a method based on a GAN is proposed in a new implementation where the GAN is trained using a training image, a conceptual model from which the statistics of the geological patterns can be extracted. The training image is first down-sampled to different scales, and the generator and the discriminator are trained alternately for each scale. The training process is implemented from the coarsest to the finest scale to progressively learn the spatial statistics from the training image. The proposed GAN is applied to simulate the two-dimensional Lena River delta and three-dimensional Descalvado aquifer analog model, in which complex geological patterns and structures from the training image are successfully learned and reproduced by the GAN. The gradual deformation method is further applied to iteratively condition the realizations by the generator to observed data, in an optimization workflow. The optimization scheme is implemented many times to obtain multiple independent models that all match the observed data. |
, Darío Graña 
