Machine Learning and Digital Rock Physics Approaches for Multiscale Characterization of Rock Properties

Characterizing accurately rock properties at the core scale is essential for effective reservoir scale modelling. In particular, in carbonate rocks posing unique challenges due to their inherent heterogeneities across multiple scales. While standard core analysis methods provide precise laboratory measurements, in many cases they fail capturing pore-scale variability within core plug samples. Digital rock physics (DRP) has emerged, in the last decades, as a powerful method addressing this gap, utilizing X-ray computed tomography (CT), micro-CT, and numerical simulations to analyse rock properties. DRP has been used widely to estimate rock properties such as porosity, permeability, and elastic moduli in carbonate and siliciclastic rocks. Nevertheless, there remains no standardized workflow for numerically characterizing rock properties in carbonates.
This study proposes three innovative applications leveraging computer vision and machine learning methods. The first application focuses on analyzing X-ray CT data to classify core sample textures. By modeling CT data, extracting representative textural descriptors, and employing the Kohonen method—an unsupervised classification technique—this approach identifies and categorizes primary textures within core sample images. The second application aims to interpolate rock properties obtained from core plug laboratory measurements along core samples. This approach exploits the continuity of properties like porosity and density observed in three-dimensional X-ray CT images by using a Convolutional Neural Network (CNN) system to interpolate these properties along the cores.
The third application introduces a novel multiscale method for simulating permeability and porosity in heterogeneous carbonate samples using 3D X-ray CT images. This approach uniquely incorporates a quantitative description of heterogeneity through machine learning-based texture classification. The texture classification results are then applied to scale up simulations of rock properties from fine to coarse scales. Finally, the proposed methods are demonstrated using two carbonate samples from a Middle East carbonate oilfield reservoir.