DiffSim: Denoising diffusion probabilistic models for generative facies geomodeling
Minghui Xu, Suihong Song, Tapan Mukerji
Abstract
Constructing high-resolution and realistic geomodels plays an important role in the decision-making processes of earth resources exploration and other sustainability strategies like subsurface carbon dioxide sequestration. Generative models have shown great promise in geomodelling as they are able to embed abstract geological knowledge. Therefore, we explore the capabilities of denoising diffusion models, new emerging generative methods, to learn the complex and high-dimensional data distribution of subsurface facies geomodels. The experiments on a synthetic channel data set illustrate the effectiveness of unconditional diffusion models in guaranteeing spatial patterns, data distribution, and diversity. Importantly, the models produce realizations free from artifacts that would contradict geological authenticity. In addition, we also test conditional diffusion models to create realistic facies models while conditioning to well facies data.