Conditional Generative Adversarial Networks for Subsurface Scenario-Based Uncertainty

Abstract Subsurface modeling workflows are critical for multi-million-dollar decisions in various subsurface resources, yet they are significantly impacted by inherent uncertainties. While realization-based uncertainty is commonly addressed, scenario-based uncertainty, which accounts for different conceptual geological models and their parameters, is often overlooked due to its time-consuming and iterative nature. Recently, generative artificial intelligence models, such as variational autoencoders, generative adversarial networks, and diffusion models, have shown promise in subsurface modeling and history matching, primarily for realization-based uncertainty quantification. However, their application to scenario-based uncertainty remains limited, often relying on modifications to loss functions that introduce training instabilities. We propose a novel workflow for quantifying scenario-based uncertainty in subsurface modeling using a conditional Wasserstein generative adversarial network with a gradient penalty (WGAN-GP), integrated with conditional batch normalization and a projection discriminator. Our approach implicitly learns label conditioning, eliminating the need for explicit conditioning loss terms and promoting more stable training and simpler parameter tuning. The proposed model supports multiple continuous conditioning labels, including object orientation, proportion, and width, and employs a trigonometric approach for parameterizing orientation. We demonstrate our workflow on a two-dimensional synthetic channel reservoir, incorporating it into a data assimilation framework using an ensemble smoother with multiple data assimilation to condition well locations and derive conditioning labels. Quantitative and qualitative checks confirm the model's ability to accurately reproduce and interpolate various geological model parameters, while maintaining geological consistency and realization-based uncertainty.