A Hybrid Convolutional Neural Network–Transformer Neural Network Model for Permeability Prediction in Rock Porous Media

Abstract

Abstract. Permeability is a crucial property indicating the fluid transport capability of rock porous media (RPM). This study introduces ConViT, a novel deep learning model that synergistically integrates vision transformer (ViT) and convolutional neural networks (CNNs) for precise permeability estimation from binarized RPM images. Its core innovation is a gated positional self-attention (GPSA) module that effectively fuses local convolutional features with global contextual attention. Evaluated on a comprehensive porous materials dataset, ConViT demonstrated superior performance, achieving a coefficient of determination (R2) of 0.9807 on the test set, significantly outperforming other benchmark architectures. To enhance interpretability, EigenCAM was employed to identify the image regions most influential to predictions. Furthermore, the model's practical utility was validated through a transfer learning application, successfully predicting the permeability of tight sandstone cast thin sections from the Shaximiao Formation with an accuracy of 82.04%. This work provides a robust, interpretable, and end-to-end solution for RPM permeability evaluation.