Multi-condition controlled sedimentary facies modeling based on generative adversarial network

Abstract

Obtaining reliable sedimentary facies models is critical for geology, which primarily involves inferring conditional distributions from sparse wells. Multiple-point simulation (MPS) captures complex geological patterns by simulating multi-point probabilities, but it is sensitive to the stationarity of training images, and the limited training image is difficult to reflect the uncertainty of the prior geological models. Generative Adversarial Networks (GANs) have recently been proposed to spontaneously simulate multiple geological models that behave differently, generating realizations that respect spatial observations (hard data), with impressive results. In this paper, we explore a more flexible approach to conditioning synthetic realizations beyond hard data, allowing stylistic control over the realizations. We construct a conditional generative adversarial network consisting of a deep generator and a multi-scale discriminator as the conditional mapping model, and obtained realizations with improved visual quality. Based on this, style codes, which are numerical codes representing geological properties, are introduced to extend the existing mapping model. The resulting extended generative model can synthesize images that respect the hard data and exhibit the desired geological style. In addition, we collect and construct a dataset, which contains abundant and realistic fluvial patterns to reflect adequate prior geological knowledge. Extensive experiments have been carried out on this dataset, the results show that the generated images reflect absolute responses to style codes while maintaining the matching rate of 94.78% to the hard data, which can provide new ideas for conditional geological modeling.