Optimisation of bridge scour depth prediction with neural networks: a comprehensive review Available to Purchase

A comprehensive review was conducted to evaluate neural network (NN) approaches as advanced alternatives for accurately predicting bridge pier scour depth. From a summary of 74 studies, it was found that multi-layer perceptrons, convolutional neural networks, recurrent neural networks and hybrid models (e.g. adaptive neuro-fuzzy inference systems) more suitable for managing complex, dynamic scour processes. Case studies state that NN-based models reduce the root mean square error (RMSE) by 27–58% and achieve an average R² of more than 0.9 – a value that is superior to conventional processes. One such case is an evolutionary radial basis function neural network that was trained with the help of genetic algorithms. This system managed to reduce the RMSE by 32% in just one model and by 58% in Hydraulic Engineering Circular 18. Artificial neural network/particle swarm optimisation hybrids can reduce the average absolute error by 45%. This review also emphasises the significance of hybrid frameworks, which integrate physical concepts into data-driven methods to be more interpretable and dynamic. The bibliometric analysis shows that this sphere of research is increasing and was predetermined by the development of computational power and access to big data.