Artificial neural network model for scour prediction at pile groups

Pile groups are commonly employed to support pile caps in intricate piers of bridges and marine structures. Scouring around pile groups results from the erosive action of flowing water, posing a threat to the structure’s stability. The three dimensional nature of the flow and the complexity of sediment transport processes amplify the scour mechanism’s intricacy, complicating scour depth modelling. While numerous scour depth prediction regression models exist in the literature, their reliability is limited when dealing with the highly complex scour process around pile groups, especially when based on a constrained dataset. This study delves into alternative approaches, utilizing machine learning techniques, specifically artificial neural networks (ANNs), coupled with various training algorithms. Analysis of large datasets compiled from existing studies indicates that ANN models outperform regression models. Among the ANN models considered, conventional Feedforward Backpropagation (FFBP) demonstrated superior performance compared to Radial Basis Function (RBF) and Generalized Regression Neural Network (GRNN) models. The study’s models were further compared with several prominent regression models in the literature, all based on extensive datasets. The findings consistently show better prediction performance in the present study. Additionally, a sensitivity analysis identified flow intensity and the approach flow Froude number as the most critical parameters influencing relative scour depth. Relative spacing between piles in line with the flow and relative number of piles are the least sensitive parameters that affect relative scour depth.