Combining Machine Learning and Digital Rock Physics Methods for Multiscale Simulation of Permeability and Porosity in a Carbonate Core Plug

Carbonate reservoir rocks reveal heterogeneity at several length scales ranging from microns to centimeters. The general strategy to simulate rock properties in carbonate samples is based on a multiscale imaging approach. First, the whole core plug sample is scanned at a coarse scale with a resolution around 20 μm to detect the main texture variabilities. Then, few millimeters subsets representing each visually identified texture are extracted and scanned with a resolution of 1 μm. Numerical simulations of permeability can be implemented at this stage as most of pore network is revealed. The last step consists on using the simulated properties at fine scale to produce an effective property for the whole core plug. In order to fill the gap between scales, we propose to use a machine learning method based on the U-Net architecture to characterize quantitatively image textures and segment the core plug voxel by voxel based on the identified classes. Then, rock properties simulated at subset scale are upscaled to the whole core plug by averaging methods. To illustrate our proposed workflow, we implemented it to upscale the porosity and permeability properties for a carbonate rock sample from an oilfield reservoir in United Arab Emirates.