Modeling natural convection in porous media using convolutional neural networks

Abstract

Deep learning has become increasingly prevalent in a wide range of engineering contexts. In this work, we tried to make a connection between the groundwater engineering community and the field of deep learning. Natural convection in porous media is usually simulated using common numerical modeling tools with high computational costs. In this work, we aim to use supervised learning in input-output pairs (porous media characteristicsheat map distribution) in an image regression task, employing an encoder-decoder convolutional neural network (ED-CNN) to develop a meta-model that is able to predict the distribution of heat map resulting from a natural convection process in porous media or to estimate the characteristics of the porous domain when the heat map distribution is given. In order to achieve this objective, a training data set of samples is prepared using Comsol Multiphysics numerical modeling and is trained with the proposed encoder-decoder CNN. We also employed several evaluation metrics such as root mean squared error (RMSE), coefficient of determination (ܴଶ-score) to assess the robustness of the developed network. We observed promising results in both approaches, as well as accuracy and speed, indicating the network's relevance in a variety of groundwater engineering applications to come in the future.