Data-driven fragility assessment of bridges using Machine-Learning techniques

Accurate assessment of bridge vulnerability to seismic events is essential for ensuring structural resilience and guiding effective risk management strategies. This study investigates the use of Machine Learning (ML) models to predict the main parameters defining seismic fragility curves—namely, the median seismic demand (μ), dispersion (β), and ultimate displacement capacity (du)—based on a limited number of structural input variables. A case study involving 25 simply supported bridges along the A19 highway in Sicily serves as the basis for a comprehensive dataset, generated through nonlinear static analyses. Fragility functions are constructed via the Cloud Analysis technique. Several supervised learning algorithms, including Gaussian Process Regression (GPR), Support Vector Regression (SVR), and ensemble-based approaches, are trained to infer fragility parameters from structural features. Among the tested models, GPR achieves the highest accuracy and generalization performance. The fragility curves derived from ML predictions show strong consistency with those obtained from traditional analytical methods, validating the approach. The proposed methodology enables efficient seismic vulnerability screening of bridge stocks with a considerable reduction in computational demand.