Machine learning-based upscaling of rock permeability from pore scale to core scale: effect of training dataset size and sub-core volumes

Abstract

Permeability characterizes the capacity of porous formations to conduct fluids, thereby governing the performance of carbon capture, utilization, and storage (CCUS), hydrocarbon extraction, and subsurface energy storage. A reliable assessment of rock permeability is therefore essential for these applications. Direct estimation of permeability from low-resolution CT images of large rock samples offers a rapid approach to obtain permeability data. However, the limited resolution fails to capture detailed pore-scale structural features, resulting in low prediction accuracy. To address this limitation, we propose a convolutional neural network (CNN)-based upscaling method that integrates high-precision pore-scale permeability information into core-scale, low-resolution CT images. In our workflow, the large core sample is partitioned into sub-core volumes, whose permeabilities are predicted using CNNs. The upscaled permeability at the core scale is then determined through a Darcy flow solver based on the predicted sub-core permeability map. Additionally, we examine the optimal sub-core volume size that balances computational efficiency and prediction accuracy. This framework effectively incorporates small-scale heterogeneity, enabling accurate permeability upscaling from micrometer-scale pores to centimeter-scale cores.