Speaker
Description
Accurate characterization of porous media at the pore scale is fundamentally challenged by two critical limitations: the scarcity of core data available only at discrete well locations, and the high spatial heterogeneity inherent in rock formations that renders small, randomly sampled sub-images non-representative of bulk core properties. This work introduces PCP-GAN, a tailored multi-conditional Generative Adversarial Network (cGAN) framework, designed to synthesize geologically accurate pore-scale images with precise and simultaneous control over multiple petrophysical properties.
The unified cGAN framework was trained on an integrated dataset of thin section imagery derived from four distinct geological depths (1879.50 m to 1943.50 m) within a marine carbonate formation. By simultaneously utilizing both sample depth and porosity as conditional inputs, the model was forced to learn both universal pore network principles and the unique, depth-specific geological characteristics of the sequence. This conditioning enabled the model to accurately capture a wide spectrum of pore architectures, ranging from high-porosity grainstone fabrics to complex, low-porosity crystalline lithologies with anhydrite mineral inclusions.
PCP-GAN demonstrated high precision in property generation, achieving an R-squared value of 0.95 for porosity control across all tested geological conditions, with mean absolute errors consistently below 0.02. Beyond quantitative metrics, visual fidelity analysis confirmed high mineralogy accuracy, specifically, the model successfully preserved features critical to geological interpretation, such as dolomite grain boundaries, angular crystal morphology, and the sharp delineation of non-porous anhydrite patches in the crystalline samples (Figure below). Furthermore, comprehensive morphological analysis confirmed that the generated images preserved critical pore network characteristics, including the average pore radius, specific surface area, and tortuosity, within standard geological tolerances.
Crucially, we developed a validation framework to benchmark the representativeness of the generated images against laboratory-measured core data (porosity and permeability). Optimized synthetic images were selected based on a dual-constraint error metric. These generated images exhibited a combined property deviation (dual-constraint error) of only 2–12% from the core targets. This performance stands in contrast to the high spatial variability observed in the real rock, where randomly extracted sub-images from the same cores showed significantly higher property deviations, ranging from 36–570%. This remarkable improvement indicates that the framework successfully addresses the core representativeness challenge in digital rock physics.
This breakthrough ability to produce synthetic rock images that are quantitatively more representative of bulk formation properties than natural, randomly sampled sub-volumes offers a powerful new tool. It significantly enhances the reliability and applicability of digital rock physics modeling and is a critical advancement for characterizing sparse-data environments relevant to energy storage, carbon capture and storage, and sustainable groundwater resource management.
| Country | United Kingdom |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
