Machine learning-aided deterministic, partially probabilistic, and fully probabilistic seismic resilience assessment methods for highway bridges

Highway bridges are critical lifelines vulnerable to seismic hazards, yet balancing computational efficiency and modeling fidelity in their resilience assessment remains a persistent challenge. This study develops a machine learning (ML)-aided framework integrating three multi-fidelity methods—deterministic (DT), partially probabilistic (PP), and fully probabilistic (FP)—to enable rapid seismic resilience quantification for highway bridges. Four ML algorithms are rigorously optimized and compared, with Random Forests emerging as the most effective for predicting engineering demand parameters (EDPs) such as column drift ratios, bearing displacements, and joint movements. The Random Forests-based surrogate models, publicly shared via Zenodo, significantly reduce computational costs while maintaining accuracy. A case study reveals that DT methods, while computationally lean, underestimate restoration time particularly under strong excitations due to the neglection of uncertainties in structural damage evaluation and restoration model parameters. The FP method integrates uncertainties in damage and restoration, achieving the highest fidelity but with computational costs and technical requirements. The PP method balances accuracy and efficiency by probabilistically evaluating damage while using deterministic restoration models. The hierarchical DT-PP-FP approach provides practitioners with adaptable tools for diverse precision, data availability, and resource constraints, advancing seismic resilience assessment of bridges through ML-driven efficiency and probabilistic rigor.