Hybrid Surrogate Modeling of a Quay Wall: An Automatically Tuned Framework Integrating Long Short-Term Memory and Feedforward Neural Networks

This study presents the development of a hybrid surrogate model combining Long Short-Term Memory (LSTM) and Feedforward Neural Network (FNN) architecture to predict the maximum bending moment in sheet pile walls. The model integrates static input parameters with wall deformation data as sequential input, enabling accurate mapping of complex geotechnical behavior. To identify the most effective architecture, an extensive hyperparameter tuning process was conducted, investigating various activation functions and network configurations using automated search techniques. The optimized model demonstrates excellent predictive performance, achieving a coefficient of determination R² of 0.999. These results highlight the potential of hybrid neural networks as efficient and reliable tools for surrogate modeling in geotechnical applications.