Data‐Driven Detection of Internal Erosion Initiation in Gap‐Graded Soils: Combining Particle‐Scale CFD‐DEM Simulation With 3D Convolutional Autoencoder

Abstract

Internal erosion in gap‐graded soils poses significant risks to water‐retaining structures such as earth dams. However, its underlying mechanisms at the particle scale remain poorly understood. This study couples the discrete element method (DEM) with computational fluid dynamics (CFD) to simulate internal erosion in gap‐graded soil assemblies and employs data‐driven techniques to detect early‐stage erosion. Particle‐scale parameters, such as contact forces, particle velocity and fluid velocity, are extracted from the transient CFD‐DEM simulations. These features are transformed into multi‐dimensional voxel‐based tensors representing the particle–fluid interactions, which are used to train deep learning models. Autoencoder models with 3D convolutional neural network (CNN) layers as encoder and decoder are developed to investigate the micro‐scale patterns within the particle‐fluid assembly. Through sequential training techniques, the temporal evolution of anomalies is captured, enabling identification of the initiation point of internal erosion. The results reveal how microscale behaviours, such as particle motion, contact forces and contact number, contribute to macroscale erosion processes. The outcome of this research can inspire further research into AI‐based early detection techniques for internal erosion in earth dams.