A Dual-Generative Adversarial Network Framework With Dynamic Prior Updating for History Matching Under Limited Knowledge

Summary

Automatic history matching (AHM) is essential for providing reliable numerical models for subsurface reservoir description and evaluation, yet conventional approaches depend critically on high-quality prior models. When prior knowledge deviates from actual geological conditions, it leads to unreliable matching results, posing a significant challenge in the field. To address this issue, we propose a dual generative adversarial network (dual-GAN) history-matching framework, where the first GAN is used for facies modeling and the second GAN is used for property modeling, allowing for dynamic updates to the prior knowledge of both facies and properties. At the core of this framework is the introduction of the conditional Wasserstein GAN (CWGAN) as the geological modeling engine, combined with the ensemble smoother with multiple data assimilation (ESMDA) algorithm. During the iterations, this framework jointly quantifies and updates both the Gaussian noise and the key conditional parameters that drive geological model generation. The method is validated using both Gaussian and non-Gaussian channelized models. The results demonstrate that the proposed method accurately reconstructs permeability fields, achieves excellent history matching for production data, and provides accurate predictions of production data at future timesteps. It proves especially effective in non-Gaussian channelized cases, where it closely approximates the true parameter distribution and correctly identifies high-permeability channels. This work offers a robust and generalizable paradigm for history matching under limited or poor prior knowledge.