Predictive capabilities of data-driven machine learning techniques on wave-bridge interactions

Abstract

To explore coastal bridge safety subjected to extreme waves during coastal natural hazards, numerical simulations that combine finite element methods and experimental data have been recognized as effective in computing wave-induced loads on coastal bridges. However, the structural design and performance assessment for bridge networks require laborious efforts and massive computational resources to account for uncertain scenarios. To provide reliable wave force estimation tools and facilitate the associated risk assessment, this study performs a hydrodynamic experiment on the wave-bridge interactions and develops data-driven Long-Short-Term-Memory (LSTM) Machine Learning (ML) models for time series forecasting of wave forces. Specifically, a 1:30 scale bridge superstructure specimen is used for the wave test in the wave channel. Different solitary wave and regular wave conditions are tested. Time histories of wave profiles, wave-induced forces, and pressures are measured and served as a dataset basis for the training of LSTM models. High-performance LSTM prediction models are developed through the tuning of different hyperparameters. The well-trained models have high accuracy and could predict the wave force time series based on the excitation wave profiles in seconds. It is envisioned that LSTM models could provide more reliable estimations with the development based on more data sources, providing a fast path for structural design, analysis, and maintenance.