Multi-Scale Investigation of Shale Using an Integrated SEM and Machine Learning Approach

Abstract

Description This study presents a novel workflow for obtaining nanoscale rock property information from SEM images of shale over a large surface area- up to 25 × 25 mm- compared to 1 × 1 mm area capability of current technologies. The techniques involved here make use of machine learning algorithms and multiscale imaging in order to map the surface of the sample at low resolution, and then use targeted high resolution imaging to compute and upscale porosity and organic matter fractions over the whole sample. Applications Acquiring a high resolution SEM mosaic of an entire plug sample surface is impractical and storage-intensive, while capturing high resolution SEM information on isolated areas can raise questions about the representativeness of the measurements compared to the whole sample. This study incorporates machine learning algorithms in order to map the various rock fabrics present in the large polished region at low resolution and determine the distribution and abundance of each texture within the sample. The algorithms then recommend optimum areas required to image at high resolution to characterize the sample. Subsequent high resolution imaging in order to characterize the abundance of porosity and organic matter within each identified fabric allows for upscaling of these properties to determine the fraction of porosity and organic matter over the entire surface of the sample. Result and Conclusions The workflow presented can reduce the amount of imaging time and data storage required by several orders of magnitude while still characterizing the full heterogeneity of a shale sample. The procedure was validated by using the upscaling method on multiple regions, then subsequently imaging those regions in their entirety at high resolution for segmentation. Comparison of the two methods provided a good match between high resolution data and upscaling method. Technical Contributions -Prior studies have indicated evidence of different computed porosity from SEM image analysis compared to bulk laboratory methods. Characterizing a more representative area will help resolve these differences. -This method can be conducted much more rapidly than previous methods with no measurable loss of accuracy. -This method greatly reduces the volume of image data created while still characterizing heterogeneity.

AAPG Datapages/Search and Discovery Article #90259 ©2016 AAPG Annual Convention and Exhibition, Calgary, Alberta, Canada, June 19-22, 2016