Rock mechanical properties can be acquired using well logs, or laboratory-based measurements. This paper describes an approach to determine these properties in near real-time using the analyses of SEM images acquired from limited samples, such as drill-cuttings. Prior image-based approaches rely on image segmentation and assignment of mechanical properties to the mineral constituents, followed by computational evaluation of the multi-mineral domain for mechanical properties. While this has shown promising results in the past, there is a high degree of subjectivity related to the assignment of mechanical properties to the segmented minerals within the image and can be computationally prohibitive.
It is commonly known that mechanical properties, such as Young’s modulus, depend not only on the mineralogy, but also the orientation and juxtaposition of the minerals spatially. Our workflow relies on extracting this information from SEM images. In our study, we utilize SEM images from 16 unconventional plays across North America. We process the images to acquire textural information and quantify textural attributes such as entropy (disorder), homogeneity, contrast (a proxy for mineralogy), and energy. We then relate measured values of Young’s modulus in each of these 16 plays with the textural attributes using non-parametric regression resulting in a unified, easily generalizable model that performs robustly when tested against previously unseen images.
Our work exploits the information content in images to reliably estimate the Young’s modulus in tight rocks. This is a promising result because the textural attributes chosen to serve as a proxy for mineralogy and for their spatial distribution, both of which control the mechanical behavior of the rock sample. One drawback of the approach is that the use of textural attributes is insufficient to address the effect of lenticular pore systems and overestimates the Young’s modulus in such cases. However, with a minor modification, we were able to address this deficiency and reliably predict Young’s modulus in a wide range of formations with varying mineralogy, porosity values and pore shapes.
Using texture analyses, we were able to rapidly predict Young’s modulus with little to no subjective input and/or prohibitive computational needs. While prior approaches were promising and successful, it appears that texture alone contains sufficient information to aid in interpreting images for mechanical properties. With this work, we demonstrate that images contain a wealth of information that can easily be exploited for screening or quicklook analyses using machine learning.
