Reservoir History Matching Using Mirror Generative Adversarial Network

Abstract History matching critically determines reservoir forecast reliability. Ensemble-based methods relying on Gaussian assumptions (e.g., ES-MDA, EnKF) cannot effectively address parameter estimation in non-Gaussian spaces, while unidirectionally constrained deep learning approaches like conditional generative adversarial networks (CGAN) exhibit mapping inconsistencies between geological parameters and production responses. A production-informed mirror generative adversarial network (MirrorGAN) integrating production-to-geology and geology-to-production mappings is proposed to resolve these challenges. This bidirectional mapping enforces geological plausibility through bidirectional consistency checks, achieving automated history matching in heterogeneous reservoirs. The proposed model comprises two integrated modules designed to transform production data into corresponding reservoir property maps. First, a cascaded image generator is constructed to generate reservoir property distributions consistent with production behavior based on dynamic response data. Subsequently, a production data reconstruction module converts the generated property maps back into production data, ensuring high consistency with the original inputs. The proposed model was validated using a synthetic dataset comprising 4,000 samples generated through numerical reservoir simulations. Benchmark tests across multiple reservoir models demonstrated that the MirrorGAN achieved a 36.25% improvement in capturing fine-scale geological heterogeneity compared to CGAN. The generated geological models also reduced the mean absolute percentage error (MAPE) of production history matching by 57.75%, confirming the effectiveness of the proposed framework. Further analysis indicates that the proposed model maintains satisfactory predictive performance even with a limited amount of training data. Moreover, as the volume of training data increases, the prediction accuracy improves consistently, demonstrating the model's strong robustness and generalization capability. These findings indicate that the model effectively overcomes the limitations of traditional data-driven methods in modeling non-Gaussian geological parameter spaces and resolves the inherent "illusion of consistency" problem in unidirectional machine learning models. The proposed approach provides a novel pathway for reservoir parameter inversion and uncertainty analysis, addressing the deficiencies of Gaussian-based ensemble methods and one-way mapping neural network inversion techniques, thereby offering reliable technical support for the formulation of efficient reservoir development strategies.