ABSTRACT
Internal erosion in gap-graded soils is a major risk faced by earth dams and levees. However, the understanding of internal erosion mechanisms, especially at the particle scale, is still limited. This research couples the Discrete Element Method (DEM) with Computational Fluid Dynamics (CFD) to simulate the internal erosion (suffusion) process in gap-graded soil samples and further train Artificial Intelligence (AI)/Deep Learning (DL) algorithms to identify subtle patterns and anomalies related to internal erosion initiation. A time-lapse micro-structure visualisation approach is introduced using 3D voxelization of soil elements under internal erosion. Particle-scale parameters such as particle and flow velocity, number of contacts, contact forces etc are extracted from the CFD-DEM simulations throughout the internal erosion process forming time-series tensors used to train the AI models. The 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 together with the variations and anomalies throughout the erosion process. Using Sequential Training framework, anomalies within the data are detected by Convolutional Autoencoder models to identify the locus and time of internal erosion initiation. In addition, the micro-mechanisms during the initiation of internal erosion such as fine particle migration and contact loss are investigated. The 3D voxelization approach for internal erosion micro-mechanism visualisation can be integrated with advanced imaging techniques (e.g., micro-computed tomography) for early detection of internal erosion in future.
