3D Multiclass Digital Core Models via microCT, SEM-EDS and Deep Learning

¹ Schlumberger Moscow Research, Reservoir Potential Department, 125171 Moscow, Russia
² NRNU MEPHI, ICIS, 115409 Moscow, Russia

Abstract

We describe an integrated methodology for constructing a 3D multiclass model of a rock sample, based on X-ray microtomography (microCT) and quantitative evaluation of minerals (QEMSCAN) by automated SEM-EDS (Scanning Electron Microscopy, Energy Dispersive Spectroscopy). We focus on building an automated operator-independent workflow, allowing to distinguish between voxels featuring substantially different physical properties, such as void, quartz, denser and less dense clay aggregates. The workflow is demonstrated using a set of five ⌀8 mm Berea sandstone miniplugs. For each miniplug, a ~4000³ voxel microCT image is acquired. Next, each miniplug is cut into smaller pieces, and the 45 resulting polished surfaces are subjected to the QEMSCAN analysis, producing ~4000² pixel mineral maps. Each mineral map is automatically spatially registered with the corresponding microCT image using an in-house surface-based algorithm. Further, the ground truth images for the supervised multiclass segmentation are constructed from the mineral maps. We compare 3D and 2D convolutional neural network (CNN) architectures with the baseline Naïve Bayes classifier, which is roughly equivalent to the approaches commonly used in practice today. We find that supervised CNN-based segmentation is fairly stable, despite microCT image quality non-uniformness and achieves higher quality scores compared to feature based and baseline approaches.