GANSim‐3D for Conditional Geomodeling: Theory and Field Application. Issue 7 (21st July 2022)

• Record Type: Journal Article
• Authors: Song, Suihong; Mukerji, Tapan; Hou, Jiagen; Zhang, Dongxiao; Lyu, Xinrui
• Is Part Of: Water resources research. Volume 58:Issue 7(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-07-21
• Subjects: facies modeling -- geological reservoir -- GANs -- karst cave -- process‐mimicking geomodeling; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2021WR031865
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: We present a Generative Adversarial Network (GAN)‐based 3D reservoir simulation framework, GANSim‐3D, where the generator is progressively trained to capture geological patterns and relationships between various input conditioning data and output earth models, and is thus able to directly produce multiple 3D realistic and conditional earth models from given conditioning data. Conditioning data can include 3D sparse well facies data, probability maps, and global features, such as facies proportion. The generator only includes 3D convolutional layers, and once trained on a data set consisting of small‐size data cubes, it can be used for geomodeling of 3D reservoirs of large arbitrary sizes by simply extending the inputs. To illustrate how GANSim‐3D is practically used and to verify GANSim‐3D, a field karst cave reservoir in Tahe area of China is used as an example. The 3D well facies data and 3D probability map of caves obtained from geophysical interpretation are taken as conditioning data. First, we create training, validation, and test datasets consisting of 64 × 64 × 64‐size 3D cave facies models integrating field geological patterns, 3D well facies data, and 3D probability maps. Then, the 3D generator is trained and evaluated with various metrics. Next, we apply the pretrained generator for conditional geomodeling of two field cave reservoirs of size 64 × 64 × 64 and 336 × 256 × 96, respectively. The produced reservoir realizations prove to be diverse, consistent with field geological patterns and field conditioning data, and robust to noise in the 3D probability maps. Each realization with 336 × 256 × 96 cells only takes 0.988 s using 1 GPU. Key Points: Generative Adversarial Network (GAN)‐based 3D reservoir simulation framework (GANSim‐3D) produces multiple 3D realistic conditional reservoir models of large arbitrary sizes from various conditioning data Application in field karst cave reservoirs shows robust performance of GANSim‐3D Geomodeling with GANSim‐3D is quite fast: <1 s for one realization with 336 × 256 × 96 cells