Physics-Guided Machine Learning for Predicting Gas Permea-bility of Standard Carbonate Core Plugs from Low-Resolution Microtomography Image Stacks

This study presents a physics-guided workflow for predicting the gas permeability of carbonate reservoirs directly from low-resolution microtomography (µCT) imagery. Standard core plugs were scanned at 34.6–36 µm/voxel, and a total of 52,327 grayscale d aggregation against experimental plug-scale measurements. The grayscale images and log-transformed permeability labels were used to train a Swin Transformer model, pre-trained on ImageNet. Two models were developed independently: one using harmonic-mean aggregation and the other using the bottleneck approach. Both models demonstrate stable convergence despite the highly skewed data distribution. The harmonic-mean model achieved R² = 0.904 on the validation set, while the bottleneck model yielded R² = 0.879. Although the higher R² reflects a closer fit to the overall trend, the bottleneck model, in blind testing on ten independent samples (0.4–2300 µm² × 10⁻³), reduced the MAE from 165 to 104 µm² × 10⁻³ (−37 %) and the RMSE from 255 to 140 µm² × 10⁻³ (−45 %) relative to the harmonic-mean model. The method provides a fast and interpretable permeability prediction based solely on raw µCT slices, without requiring image segmentation or 3D reconstruction. The proposed approach demonstrates robust performance across a wide range of standard carbonate plugs and effectively captures permeability trends even in the presence of structural heterogeneity. While samples with extremely large fractures or vugs can introduce local inconsistencies in labelling due to the limitations of slice-based estimation, these cases are rare and can be systematically addressed in future work. Overall, the results highlight the strong potential of physics-guided machine learning to accelerate digital core analysis and provide reliable, image-driven permeability predictions for complex carbonate reservoirs.