Interpretable machine learning based tsunami bridge fragility assessment

Coastal bridges are frequently subjected to tsunami hazards due to the recent climate change and global warming phenomena. Risk assessment of these bridges is important for proper disaster management and mitigation. Surrogate Tsunami Demand Models (STDM) are useful relationships for simplifying the process of development of fragility curves. The current study proposes a novel methodology to develop STDMs based on Machine Learning (ML) for the risk assessment of a typical coastal bridge due to tsunami load, instead of the conventional Monte Carlo Simulation (MCS) method. In this study, twelve machine learning models are employed for developing ML based Surrogate Tsunami Demand Model (MLSTDM). Tsunami fragility functions are then created using the best performing models. It is found that the proposed MLSTDMs can accurately estimate the fragility with fewer simulations when compared to MCS method. Further a SHAP (SHapley Additive exPlanations) analysis is used to interpret the performance of surrogate ML models in predicting tsunami responses.