Machine learning techniques for the estimation of limit state thresholds and bridge-specific fragility analysis of R/C bridges

Based on past earthquake events, bridges are the most critical and usually the most vulnerable components of road and rail transport systems, while bridge damage is related to substantial direct and indirect losses. In view of this, the need for direct and reliable assessment of bridge vulnerability has emerged, and several methodologies have been developed using probabilistic analysis for the derivation of fragility curves. A new framework for the derivation of bridge-specific fragility curves is proposed herein, introducing machine learning techniques for a reliable estimation of limit state thresholds of the most critical component of the bridge system (which in standard -ductility based- design is the piers), in terms of a widely used engineering demand parameter, i.e. displacement of control point. A set of parameters affecting the seismic capacity and the failure modes of bridge piers is selected, including geometry, material properties, and reinforcement ratios for cylindrical piers. Training and test sets are generated from multiple inelastic pushover analyses of the pier component, and Artificial Neural Networks (ANN) analysis is performed to derive closed-form relationships for the estimation of limit state thresholds. The latter are compared with closed-form relationships available in the literature, highlighting the effect of machine learning techniques on the reliable estimation of bridge fragility curves for all damage states.