Hybrid surrogate models of quay walls

The use of hybrid surrogate modelling techniques for the life-cycle management of anchored quay walls was investigated. A synthetic dataset was generated using a simplified structural model based on classical earth pressure theory, representing a range of geometrical and hydraulic boundary conditions. Two neural network (NN) architectures were developed and compared: (a) a baseline feedforward neural network (FNN) using static input features and (b) a hybrid model combining bidirectional long short-term memory (BiLSTM) layers with dense layers (BiLSTM–FNN), which incorporates sequential displacement data. Both models were tuned across multiple trials with varying architectures, activation functions and learning rates. The final architectures were deployed in supervised learning to train surrogate models. The BiLSTM–FNN model outperformed the FNN, achieving significantly lower validation loss and superior predictive accuracy, but at a higher computational cost. This modelling approach provides an effective tool for estimating internal structural forces such as maximum bending moments, thereby supporting predictive maintenance and optimised design. The results demonstrate the potential of hybrid NN architectures within digital twin frameworks for port infrastructure, contributing to enhanced resilience and more efficient resource use.