Generative geomodelling: Deep Learning vs. Geostatistics

Suihong Song, Jiayuan Huang, Tapan Mukerji

Generative geomodelling aims to simulate subsurface facies distributions while honoring multiple types of conditioning data and geological knowledge. This study selects three typical multiple-point statistics (MPS) approaches—Direct Sampling (DS), Quick Sampling (QS), and SNESIM—and two Generative Adversarial Network (GAN) workflows—post-GANs perturbation and GANSim—as representatives to compare traditional geostatistics-based and deep learning-based generative geomodelling methods, based on two sedimentary reservoir scenarios. In addition to the latest GANSim enhancements—namely, the local discriminator and facies-indicator output designs—this paper further proposes injecting global feature information into intermediate layers of the generator, instead of concatenating global features with latent vectors, to improve constraint effectiveness. The geomodelling results demonstrate that GANs, especially GANSim, consistently produce geologically realistic and diverse facies models that are accurately conditioned to well facies data, global features, and facies probability maps. In comparison, MPS approaches perform well in honoring well facies and probability maps but produce facies models with significantly lower geological realism. Their conditioning effectiveness on global features is also less reliable. GANSim achieves geomodelling speeds hundreds of times faster than MPS methods. Flow simulations show that GANSim results yield more accurate and less uncertain predictions than MPS outputs. Moreover, although trained on stationary conceptual geomodels, the trained GANSim generalizes well to model large, nonstationary reservoirs by spatially varying the input global feature maps and carefully designing the conditioning probability maps, making it a powerful and flexible tool for high-fidelity conditional geomodelling.