Authors : Ahmed Merzoug, Lei Liu, Michael Pyrcz
Abstract
Optimal subsurface development decision-making relies on the calculation of an accurate ensemble of subsurface feature realizations that integrate all available sources of information. Generative artificial intelligence (genAI) has been proposed as a solution for building subsurface realization ensembles that may exceed geostatistical methods for conditioning not only to local measurements but also to qualitative geological concepts based on the complicated depositional, preservation, and alternation history of subsurface resource accumulations. GenAI models such as variational autoencoders, generative adversarial networks, and denoising diffusion probabilistic models have been applied, but these models are often checked by ocular inspection or limited statistics. We propose an extended genAI model checking protocol with novel metrics, diagnostics, and plots to expand current geostatistical model checking methods, known collectivity as minimum acceptance checks. This genAI model checking protocol includes checks for global univariate distribution, local and global uncertainty models, spatial continuity, local data exactitude, and local dynamic response. To demonstrate this expanded genAI model checking protocol, we design and train three diverse genAI methods on sequential Gaussian simulation realization-based training images for different variogram ranges. The results show that GANs passed all checks, DDPMs had some limitations in reproducing local uncertainty, whereas VAEs failed most of the checks. Our enhanced workflow allows for a quantitative measure of the performance of genAI models. It allows practitioners to select suitable genAI models for their subsurface modeling workflows and to have improved confidence in the development decision based on these genAI realizations.
31 авг 2024
