Pore-scale modelling and implementation of micro x-ray Computed Tomography (μxCT) images have become a reliable method to predict the petrophysical properties of rocks. However, the prediction of porosity and investigation of permeability anisotropy is highly reliant on image resolution. The trade-off between image resolution and field of view makes properties prognostication of large-scales low-resolution models challenging. To estimate the properties of single-phase flow accurately, implementing the upscaling method is vital to correlate different scales. We have combined Convolutional Neural Networks (CNNs) with the downsampling technique as an upscaling approach and then carried it out on multiscale images of Fontainbleau sandstone. By employing pore-scale simulation, the normalized downsampled high-resolution images are labelled with their porosity, permeability in X-direction (kx), Y-direction (ky), and Z-direction (kz), used to train the CNNs. The upscaled properties of low-resolution images are predicted using trained CNNs. A continuum-scale simulator has been used to simulate single-phase flow in different directions to observe the anisotropy effect and compute the overall values of upscaled properties. The results suggest that the represented upscaling method can significantly close the properties of the low-resolution to the same properties of the high-resolution model while enormously reduces the computational time and cost.
- Convolutional neural networks
- Digital rock physics
- Single-phase flow