Quantifying seismic response of helical pile-supported bridges using machine learning

This thesis investigates the seismic performance of bridge systems using a combination of numerical simulations and machine learning techniques to assess structural responses and predict damage, considering soil-structure interaction. A finite element model (FEM) is developed in OpenSees to simulate the seismic behavior of a pile-soil-bridge system and is validated against shake table test data. The bridge model, consisting of four piers supported by pile groups, is subjected to two earthquake records, El Centro and Tianjin, at varying intensity levels. The results showed good agreement with the experimental data, with the El Centro earthquake inducing higher acceleration and moment responses, while the Tianjin earthquake caused greater displacement responses, underscoring the impact of earthquake characteristics on bridge behavior. A fragility study is then conducted for a three-span bridge supported by helical piles in cohesive soil. Latin Hypercube Sampling (LHS) is employed for addressing material uncertainty and Incremental Dynamic Analysis (IDA) is used to derive the fragility curves. A total of 6,600 nonlinear time history analyses are conducted on 15 bridge samples subjected to 22 ground motion records. The key factors influencing damage at piers and helical piles are identified. Consequently, regression analysis revealed that longer spans increased drift, while damping ratio and pile spacing significantly affected settlement. Helical piles are found to be more vulnerable to ductility demands than to settlement, marking them as a critical component in the system. Machine learning algorithms are then applied to classify damage states based on the dataset obtained from numerical analyses. Among the twelve models evaluated, CatBoost and LightGBM provided the best classification performance, while traditional models like Linear Discriminant Analysis and Naive Bayes struggled with class imbalance. The findings highlight the effectiveness of machine learning in seismic damage prediction, despite the challenges posed by imbalanced datasets.