From Voxels to Graphs an Efficient Approach to Estimate Elastic Properties of Digital Rocks

Abstract Digital Rock Physics (DRP) enables computational simulation of elastic rock properties but remains limited by high computational cost and challenges in upscaling. To overcome these constraints, we propose a graph-based AI surrogate modeling framework that transforms 3D digital rock volumes into unstructured and weighted graphs, where nodes represent grains (or pores), and edges encode spatial connectivity. Using a dataset of 30,000 subsamples from five sandstone formations, we generate weighted graphs through a custom graph builder leveraging computer vision techniques. These graphs preserve the rock's complex geometry and topology more effectively than traditional voxel grids. A Graph Neural Network (GNN) is trained to predict bulk and shear moduli from these graph representations, achieving high accuracy on test data. The model also demonstrates strong generalization to new rock samples without additional training, outperforming conventional methods that require re-training. Moreover, the graph framework enables the upscaling of elastic property predictions from small sub volumes to full core plugs. This approach introduces a scalable, resolution-independent, and compact alternative to traditional DRP, addressing its main bottlenecks while preserving the fidelity of rock property predictions.