Prediction of Effective Elastic Moduli of Rocks using Graph Neural Networks

Understanding and predicting rock mechanical properties, such as elastic moduli, plays a crucial role across a range of geoscience and engineering fields including energy resources engineering (Sone and Zoback, 2013; Madhubabu et al., 2016), geotechnical engineering (Zhang et al., 2021), and seismology (Byerlee and Brace, 1968). These properties govern how rocks react to in-situ stresses, shaping the macroscopic behaviors of geological structures. Accurate characterization of these properties is therefore instrumental in predicting phenomena such as seismic wave propagation, reservoir behavior, and slope stability. Fundamentally, these macroscopic behaviors originate from the intricate features at the microscopic scale. Specifically, the effective elastic moduli of rocks are influenced by three main factors: (1) the composition of pores and minerals, (2) the properties of these constituents, and (3) the geometric details of the rock's microstructure (Mavko et al., 2020). While the first two factors can be relatively straightforwardly measured and characterized, capturing the complexity of the geometric details often presents a substantial challenge. Historically, traditional micromechanics homogenizations and rock physics relied on empirical relationships or theoretical models based on idealized microstructures to estimate rock properties (Li and Wang, 2008; Han et al., 1986; Mindlin, 1949; Hill, 1952).