Core plug porosity prediction using microtomography, supervised labeling, and a shifted window transformer

Recent advances in machine learning have enabled the automatic analysis of microtomography (µCT) images, facilitating more efficient rock property identification. This study aims to predict the experimentally measured open porosity of reservoir rocks using µCT images of standard core plugs. A dataset of 136 core plugs was collected, including 49 sandstone and 87 carbonate samples. Open porosity was experimentally determined using gas volumetry. The core plugs (30 ± 1 mm in height and diameter) were scanned using µCT with a resolution of 34.6–38.0 µm, producing 16-bit image stacks. The dataset consisted of 100,232 images (64,119 carbonate and 36,113 sandstone). To label the images, we introduced a supervised method called Segmentation of Unresolved Pores via Experimental Reference (SUPER), which segments dark voxels to match the experimentally measured open porosity, adapting to each sample’s characteristics. Three shifted window (Swin) transformer models were trained: a universal model and specialized models for sandstone and carbonate. The models used transfer learning with ImageNet weights, followed by fine-tuning. Testing confirmed that specialized models outperformed the universal model. This highlights that training an ensemble of models adapted to specific rock types leads to better performance than a single general model for porosity prediction. A key challenge arose with sandstones, especially fine-grained types, where small pores merged due to resolution limitations. Future work should improve image resolution and feed detailed images into the model. The method has potential for full-scale core scans and early porosity assessment in raw core plugs, including fragile reservoirs with oil or bitumens.