Generalized Conditioning of Generative Artificial Intelligence for History Matching Subsurface Models

초록

Resource engineering depends heavily on subsurface modeling and history-matching for multimillion-dollar decision-making, where reducing uncertainty requires the integration of diverse data. While generative artificial intelligence (GenAI) has shown promise for parameterizing complex subsurface models into low-dimensional representations that preserve geological realism, current approaches suffer from limited generalizability. Specifically, GenAI models are trained on preconditioned subsurface models, which limits them to specific well configurations. This approach requires complete retraining when new hard data are acquired, which is expensive, especially for large models. We introduce a novel, generalizable approach that fundamentally redefines this workflow. Our method trains the primary GenAI model on unconditioned subsurface models and then connects it with a smaller inference network that handles the conditioning to specific well data. The combination of the GenAI model and inference network is then used for history-matching using ensemble smoother multiple data assimilation. If the data change, only the inference network is retrained, allowing broader application of GenAI for history-matching, which is independent of hard measurements. Using a Wasserstein generative adversarial network with a gradient penalty as our core GenAI model, coupled with a fully connected inference network, we demonstrate our workflow on a three-dimensional fluvial channel reservoir case study, where the measurement and well number/location changes with time. Results confirm that our approach efficiently handles updated well information while maintaining geological consistency, providing a more practical solution for real-world subsurface modeling challenges where data configurations evolve over time.

키워드