Rapid Post-Earthquake Assessment of Bridge Portfolios Using Damage-State Augmentation

Rapid portfolio-level post-earthquake bridge damage assessment is essential for restoring transportation networks and supporting emergency decisions, yet the current approaches face two persistent barriers. High-fidelity nonlinear analyses are computationally prohibitive at portfolio scale, while data-driven classifiers have been hindered by scarce and severely imbalanced damage-state labels, especially for severe damage. To address these gaps, this study presents a computational model consisting of two coupled components: a recurrent neural network surrogate for efficient response prediction and a gradient boosting-based classifier for damage-state identification. The surrogate learns the mapping from ground-motion features and bridge parameters to curvature-related response quantities, which are then used as physically meaningful features for classification. The classifier prioritizes the informative samples based on the predictive uncertainty and incorporates imbalance-mitigation strategies during training. Across five representative bridge types, the proposed model achieves test accuracies of 87.5%-96.5% for type-specific training and 83.1%-89.8% for joint cross-type training, demonstrating strong cross-type generalization.