@article{Wu2026, 
author = {Xue-Chao Wu and Wen-Yao Fan and Shi-Jie Peng and Xin-Hua Xu and Feng Deng and Xi-Xi Du},
title = {Stochastic simulation and finely characterization of heterogeneous reservoirs based on MSGAN and BicycleGAN},
year = {2026},
journal = {Petroleum Science},
volume = {23},
number = {9},
pages = {5334-5358},
keywords = {Reservoir modeling, Generative adversarial networks, Heterogeneous geological patterns, Data augmentation, Conditional simulation},
url = {https://www.sciopen.com/article/10.1016/j.petsci.2026.07.009},
doi = {10.1016/j.petsci.2026.07.009},
abstract = {Subsurface reservoirs can be constructed based on Generative Adversarial Networks (GANs), but some challenges cannot be overlooked. For example, both prior geological patterns and knowledge are hardly quantified so that GANs cannot be trained without sufficient samples. Meanwhile, the heterogeneous patterns are difficult to characterize, with less consideration of multiple geological constraints, resulting in poor modeling accuracy and reliability. Therefore, we integrated Multi-Stage GAN (MSGAN) and BicycleGAN to finely characterize heterogeneous reservoirs. Specifically, MSGAN can extract and quantify geological patterns based on multiscale representations of pyramid structure. Through concurrent training and parameter inheritance, these simulated prior samples share the same feature space with the Training Image (TI) so that they can support BicycleGAN's training. The bijective consistency between latent codes and output modes is established, and reservoir characterization is achieved through the nonlinear mapping between multiple conditioning data and output modes. Both synthetic cases and field application were introduced to verify the applicability of the proposed modeling framework. During data augmentation based on MSGAN, different simulations preserve a better generation diversity, and the complex geological pattern fitting with multiple geological constraints can be reproduced by BicycleGAN. Each conditional simulation is similar to TI in terms of spatial variability, channel connectivity and spatial structures, significantly reducing modeling uncertainties.}
}