Rapid Synthesis of Porous Media Volumes with Deep Learning Generative Models

Abstract:

Deep generative models have been successful in many domains, including image style transfer, speech audio synthesis, and dialogue generation. In energy sciences, these models have been applied to applications from geologic history matching to multimodal image synthesis. An area of interest when studying source rocks at the pore scale is using limited imaging data to generate multiple realizations of a rock’s pore structure and to quantify uncertainty in petrophysical properties. One type of deep generative model, generative adversarial networks (GANs), has been used to synthesize microscale rock image volumes. While the performance of these models are promising, GAN-based approaches have two major shortcomings: 1) the inability to use likelihood-based methods for image inpainting or denoising, and 2) the limitation of training on fixed image volumes.

In this work, we present an alternative approach to synthesize porous media volumes using generative flow models, a likelihood-based model fundamentally different from GANs. We train generative flow models on SEM and FIB/SEM images and use these models to produce multiple realizations of the input training rock samples. We present results for generating sandstone, carbonate, and shale image volumes and show that the topological and flow properties of the generated samples match. We also demonstrate the ability to synthesize 3D image volumes from only 2D training data. This capability has applications for synthesizing nanoscale shale volumes, where the only imaging data available is often a small number of 2D microscopy images. We further show the ability of these models to synthesize multimodal imaging data and apply these models to data inpainting, where we predict or repair one imaging modality (e.g. electron microscopy) based on imaging data obtained from another modality (e.g. computed tomography). Overall, generative flow models offer a promising new direction for synthesizing, statistically characterizing, and processing source rock images. With further study, the approach presented here can advance characterization methods for source rocks from unconventional formations.