An improved data-driven method for the prediction of elastic properties in unconventional shales from SEM images

Abstract

This paper demonstrates a quick look approach to estimate rock mechanical properties, such as Young's modulus, from SEM images of drill cuttings acquired from several unconventional plays across North and South America. Unlike traditional methods that involve extensive lab measurements or interpretive well logs, our approach leverages image-based analyses, significantly reducing the subjectivity and computational burden often encountered in previous strategies.

In our analysis, we processed SEM images from 14 plays to extract both textural and shape-based features. Textural attributes such as entropy, homogeneity, contrast, and energy provide insights into the disorder and mineralogical contrasts within the rock. There exist several shape-based features such as area, aspect ratio, circularity, solidity, extent, eccentricity, Euler number, and orientation which can describe the geometric properties of mineral constituents. We utilize both textural attributes and pore aspect ratios and integrate deep learning inputs to construct various predictive models.

Our models correlate these features with empirically measured Young's modulus values using non-parametric regression. This integrated approach has shown to provide a robust and generalizable model capable of estimating Young's modulus across a diverse set of geological formations with high reliability, even when tested against previously unseen images.

This study acknowledges that mineralogy and the juxtaposition of various minerals may be unable to fully account for the variations seen in the Young's modulus. Complex pore systems such as lenticular pores may lead to overestimation of elastic moduli. Therefore, the inclusion of both textural and shape attributes as proxies for mineralogy and their spatial arrangement addresses key controls in the mechanical behavior of rock samples, thereby enhancing the model's applicability across varied mineralogical and porosity conditions.

Our findings indicate that the combination of texture and shape analyses, coupled with machine learning techniques, can efficiently and accurately predict mechanical properties in tight rocks. This method represents a significant advancement over traditional approaches, providing a fast, non-subjective, and computationally efficient tool for preliminary rock mechanics analysis. This work underscores the potential of using SEM image analyses as a powerful tool for rapid screening and detailed rock mechanics studies, moving towards more streamlined and data-driven exploration and production strategies.