A CNN-based approach for upscaling multiphase flow in digital sandstones

In digital rock physics, finding the trade-off between image resolution and field of view and, at the same time, maintaining the accuracy of flow prediction in a low-resolution large-scale model is a challenging task. Multiphase flow characteristics are strongly dependent on image resolution. Therefore, the upscaling process based on pore-scale simulations is crucial for predicting low-resolution model properties regarding the referred trade-off. This study has disclosed an upscaling method taking advantage of convolutional neural networks (CNNs) and downsampling techniques. The method has been applied to a set of multi-scale images taken from Fontainebleau sandstone. High-resolution images are downsampled and then labelled with their properties to train CNNs. Trained CNNs are then used to predict the upscaled properties of low-resolution samples. The predicted upscaled properties by CNNs are assigned to computational grids of the reconstructed low-resolution model in a continuum-scale simulator. By performing multiphase flow simulations of the mentioned model and various assigned properties, the resultant dynamic behaviours from low-resolution and CNN upscaled properties are compared with the dynamic behaviour of the high-resolution case. Results showed a satisfying match between the dynamic behaviour of the upscaled model through CNNs and high-resolution properties with a significant reduction of computational cost and time.

Automated summary

The study utilizes convolutional neural networks and downsampling techniques to train high-resolution images for predicting and assigning upscaled properties to low-resolution models, resulting in satisfactory outcomes in multiphase flow simulations.