InterPore2026

Upscaled Prediction of Key Petrophysical Properties Directly from Digital Rock Images

19 May 2026, 15:05

1h 30m

Poster Presentation (MS15) Machine Learning in Porous Media Poster

Speaker

Yuntao Jia (Beihang University)

Description

Relative permeability and capillary pressure are key petrophysical parameters for subsurface applications such as hydrogen and CO2 storage. However, obtaining them from laboratory measurements or pore-scale numerical simulations is time-consuming and computationally expensive, especially for large-size rocks. Existing artificial intelligence methods often rely on hand-crafted petrophysical descriptors, fixed-size samples, or insufficient physical constraints, which limit their adaptability and practical applicability.
In this work, we present a series of physics-informed data-driven models for predicting relative permeability and capillary pressure directly from 3D digital rock images across scales. The training and testing datasets are generated by two-phase network modeling of over 100 segmented 3D digital rock images. Primarily, a hybrid ConvLSTM-CNN model is developed for direct prediction from fixed-size digital rocks, avoiding explicit parameterization of pore-structure descriptors. Second, the framework is extended to cross-size prediction through spatial pyramid pooling and physical information embedding, in which computed tomography resolution, interfacial tension, and contact angle distribution are incorporated to improve adaptability and physical consistency. Finally, a more advanced upscaling model is constructed to predict core-scale properties directly from partial sub-volume images by embedding spatial position and length-fraction information, enabling prediction for large-size rocks from limited structural observations.
Compared with network modeling resulting, the direct-prediction, cross-size, and upscaling models achieve accuracies of about 95.0%, 95.3%, and 98.6%, respectively. The deep-learning models reduce inference time by more than 90% compared with network simulations. The upscaling model is further validated against experimental measurements on an unseen centimeter-scale core sample. When only 1/16 of the whole-core structural information is used as input, the model achieves 97.2% accuracy compared with the experimental results. These results demonstrate efficiency and accuracy of our physics-informed deep learning framework for predicting and upscaling key petrophysical properties from 3D digital rock images.