Deep Learning for Connectivity Identification in Random Subsurface Flows: A Methodological Workflow for Early Solute Arrival Time Quantification

Abstract Identification of preferential flow paths in heterogeneous subsurface environments is key to assess early solute arrival times at environmentally sensitive targets. We propose a novel methodology that leverages the information contained in preferential flow paths to quantify early arrival times and their associated uncertainty. Our methodology is based on a two‐stage approach that combines Convolutional Neural Networks (CNN) and Multi‐Layer Perceptron (MLP) techniques. The CNN is used to identify preferential flow paths, the MLP being employed to map tortuosity of these paths and key geostatistical parameters of conductivities therein onto early arrival times. As such, our approach provides novel insights into the relationship between the geostatistical characterization of conductivities along preferential flow paths and early arrival times. The effectiveness of the approach is exemplified on synthetic two‐dimensional (randomly) heterogeneous hydraulic conductivity fields. In this context, we assess three distinct CNN architectures and two MLP architectures to determine the most effective combination between these to reliably and effectively quantifying preferential flow paths and early arrival times of solutes. The resulting framework is robust and efficient. It enhances our ability to assess early solute arrival times in heterogeneous aquifers and offers valuable insights into connectivity patterns associated with preferential flow paths therein.