Machine-Learning Techniques Characterize Source-Rock Images at the Pore Scale

Abstract

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 208632, “Reconstruction and Synthesis of Source-Rock Images at the Pore Scale,” by Timothy Anderson, SPE, Stanford University. The paper has not been peer reviewed. Nanoimaging techniques for characterizing pore-scale structure of shales trade off between high-resolution/high-contrast sample-destructive imaging modalities and low-contrast/low-resolution sample-preserving modalities. Acquisition of nanoscale images also is often time-consuming and expensive. In the complete paper, the author introduces methods for overcoming these challenges in image-based characterization of the fabric of shale source rocks using deep-learning models. Introduction A promising application of data-driven scale-translation techniques is nanoscale imaging. This application is important for studying shales because of the importance of nanoporosity in shale gas storage. Nanoimaging techniques, however, present specific challenges and can result in small image data sets that do not allow for accurate quantification of petrophysical or morphological properties. Consequently, data translation and generation both offer many opportunities to assimilate multiple nano- and microscale modalities and to overcome limitations of nanoimaging systems. The author proposes an image-based characterization work flow (Fig. 1) for data-driven scale translation that uses deep-learning image synthesis and translation models to assimilate multimodal, multiscale, and data-scarce source-rock images for predicting petrophysical and morphological properties. A central challenge in source-rock characterization addressed in this work is the reconstruction of 3D volumes when only 2D training images are available. Image-translation models are presented for reconstruction of 3D image volumes from 2D training data and a porous media image-synthesis algorithm that generalizes to 3D grayscale and multimodal volume generation from 2D training data. The complete paper describes the translation and synthesis models, applies these models to source-rock image data sets, and discusses extensions and future directions for the introduced work flow.