A Segmentation-Independent Workflow for Digital Rock Property Prediction Using Pore-Structure-Preserving Generative Adversarial Networks and Synthetic Data: Porosity Case Study

Summary Traditional digital rock physics (DRP) workflows rely heavily on image segmentation to distinguish pore and matrix spaces. This reliance introduces substantial uncertainty that propagates through property calculations and compromises the quality of training data for machine learning (ML) models. Segmentation inconsistencies between methods and operators introduce significant variability in computed properties, undermining model reliability and reproducibility. This study presents a segmentation-independent, machine-learning-based workflow capable of predicting rock properties directly from real CT images using entirely synthetic training data with deterministic labels. This work serves as a proof-of-concept focused on pore-structure preservation and is validated on 2D porosity prediction in sandstone samples, chosen because it enables rigorous end-to-end validation without confounding factors from indirect property measurements. Synthetic binary rock images with controlled porosity are generated using a modified Wasserstein generative adversarial network with gradient penalty (WGAN-GP). These binary images are translated into realistic grayscale counterparts using a cycle-consistent generative adversarial network (CycleGAN), bridging the gap between idealized training data and the complex textures of real CT scans. A convolutional neural network (CNN) trained on these CycleGAN-generated images can then predict rock properties from real CT scans without any segmentation. The proposed workflow achieves excellent performance when applied to Bentheimer sandstone CT data. A synthetic data-driven setup, where both training and validation are performed exclusively on synthetic binary images, shows strong predictive accuracy when tested on unseen real CT slices from the same formation [R² =0.958, mean absolute error (MAE) = 0.56%]. External validation across seven different sandstone formations, beyond the model’s training domain, demonstrates successful generalization in five cases (R² =0.925, MAE = 1%) after lightweight preprocessing, while highlighting limitations of generalization beyond the training domain. Porosity, specific perimeter, Euler characteristic density, and hydraulic radius were compared before and after CycleGAN translation to verify pore-structure preservation. The observed conservation of pore geometry and connectivity supports the applicability of this workflow to other flow-related properties. This work represents a fundamental shift in DRP methodology. It bypasses segmentation-induced uncertainty, addresses sampling bias through controlled synthetic data generation, and enables property prediction when physical samples are unavailable or cost-prohibitive. While demonstrated here through porosity for rigorous validation, the segmentation-independent workflow is inherently property-agnostic and opens the door to scalable, reliable, and largely automated DRP workflows.