Deep Learning and Multimodality Imaging to Improve Shale Fabric Characterization
Abstract
A significant challenge in analyzing shale source rock samples is obtaining high resolution images through non-destructive imaging techniques. 2D imaging techniques offer high spatial resolution but are often destructive, while 3D imaging methods preserve the sample at the expense of resolution and contrast. Multimodality shale imaging techniques seek to assimilate low- and high-resolution imaging data to produce images comparable in quality to destructive techniques while also preserving samples for further tests and experimentation such as flow transport tests for pore network connectivity, permeability, and chemical species absorption. In this study, a 30 μm diameter Vaca Muerta shale cylindrical sample was imaged at the nanoscale using transmission X-ray Microscopy (TXM) (non-destructive) and FIB-SEM (destructive). Acquired images had similar cross-sectional pixel resolution. TXM/SEM slice image pairs were identified using structural similarity index (SSIM), invariant moments, and mean squared error. These paired images were then aligned using standard image registration methods to produce the final dataset of aligned dual-modality images. We use these registered pairs to train three image translation models: baseline linear and nonlinear filters, image-to-image convolutional neural networks (CNNs), and conditional generative adversarial networks (CGANs) to map from input TXM images to FIB-SEM images. The models were trained on image patches sampled from the full-slice images, and randomized transforms were applied to the input images to further augment the amount of available data. Predicted SEM images were assessed using Peak Signal-to-Noise Ratio (PSNR) and SSIM metrics. The image-to-image CNN and CGAN models showed significantly improved contrast and image quality, which in turn aids in organic matter segmentation and mineral identification. Overall, our results show that it is possible to overcome many difficulties posed by non-destructive imaging by coupling dual-modality imaging with image translation models. We believe this approach opens up many potential future avenues in characterizing the microstructural fabric of shale.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMMR13B0066A
- Keywords:
-
- 1832 Groundwater transport;
- HYDROLOGY;
- 5104 Fracture and flow;
- PHYSICAL PROPERTIES OF ROCKS;
- 5139 Transport properties;
- PHYSICAL PROPERTIES OF ROCKS;
- 5199 General or miscellaneous;
- PHYSICAL PROPERTIES OF ROCKS