FLOW THROUGH POROUS MEDIA: DATA-DRIVEN MODELLING USING DEEP NEURAL NETWORKS

Abstract

The Multihead-ConvNet architecture is a remarkable neural network design that has proven to be highly effective in predicting fluid flow in 3D digital rock images. Its main strength lies in its ability to extract spatial correlations between the morphology of the porous medium and the fluid velocity field, enabling fast and accurate predictions. In this approach, 3D binary pictures are used as input data, and the network is trained by extracting simple geometrical information that captures the key characteristics of the porous medium. By training a deep neural network with this data, the model becomes well-suited for simulating flow through porous materials, eliminating the need for computationally intensive numerical simulations. One of the most significant advantages of this method is its incredible speed. The trained network can provide correct flow field predictions in less than a second, which is a significant improvement compared to the time-consuming nature of traditional numerical simulations. This breakthrough allows for a substantial reduction in computational costs, enabling researchers and practitioners to obtain results in a fraction of the time previously required. The U-Net architecture, coupled with 3D binary pictures and simple geometrical information extraction, represents a significant advancement in the field of fluid flow prediction in porous media. Its speed and accuracy make it a highly efficient alternative to traditional numerical simulations, empowering researchers and practitioners with a powerful tool for analyzing and understanding fluid behavior in complex porous materials.