MR003-0019
Three-Dimensional Source Rock Image Reconstruction through Multimodal Imaging

Monday, 14 December 2020
Poster
Anthony R Kovscek, Stanford University, Energy Resources Engineering, Stanford, CA, United States, Timothy Anderson, Stanford University, Electrical Engineering, Stanford, CA, United States, Bolivia Vega, Stanford University, Energy Resources Engineering Department, Stanford, CA, United States, Jesse Mckinzie, University of Wyoming, Laramie, United States, Yuhang Wang, University of Wyoming, Laramie, WY, United States and Saman A Aryana, The University of Wyoming, Department of Chemical Engineering, Laramie, WY, United States
Abstract:
Sample imaging over a range of different modalities offers a set of advantages and limitations that dictate the scope and quality of their outcomes. Tradeoffs are often created among them. Non-destructive modalities, such as X-ray based techniques, allow preservation of the sample, but may fall short on image contrast and resolution. Many destructive modalities, such as the combination of ion beam milling and electron microscopy, produce high-resolution, high-contrast images, at the cost of being two dimensional and preventing future use of the sample. There is no single image modality that combines being non-destructive, and achieve high-resolution, and high contrast information directly from full rock volumes.

We present an approach for predicting 3D source rock image volumes using only 2D paired multimodal images and unpaired 3D images. Our dataset consists of a transmission X-ray microscopy (TXM) reconstructed 3D image volume and focused ion beam (FIB) milled scanning electron microscopy (SEM) image sequence acquired for a Vaca Muerta oil window shale sample. The FIB-SEM images are paired with 2D cross sections of the TXM image volume to create an aligned 2D multimodal image dataset. We train a variant of a generative adversarial network (GAN) deep learning model to predict high-contrast/high-resolution images (similar to FIB-SEM) from low-contrast input images (TXM). Our model employs novel regularization methods to enforce continuity and connectivity between slices, enabling reconstruction for 3D image volumes. The synthesized rock volumes are evaluated based on their structural and flow properties. These flow properties are established using an LBM framework with a combined halfway bounce-back and Maxwellian diffusive reflection for boundaries. Results demonstrate a potential pathway to enhance 3D image volumes by training machine learning models with paired 2D data and unpaired 3D data guided by flow behavior. With further refinement, our method offers an approach for bridging some of the current limitations of imaging modalities and improve their capabilities for source rock characterization.

<< Previous Abstract | Next Abstract >>