Direct mapping from macroscopic attributes to seismic robustness: A trustworthy machine learning framework for regional highway bridges

For rapid post-earthquake screening of bridge networks, this study develops a trustworthy machine learning framework to address the computational bottleneck of high-fidelity simulations. We propose a direct mapping from macroscopic structural attributes to a continuous seismic robustness index (RI). By performing random sampling from a uniform distribution over the spectral acceleration ( S a ) domain, fragility functions are converted into deterministic continuous labels for supervised learning. To prevent information leakage inherent in grouped bridge datasets, leakage-free GroupKFold cross-validation is adopted. Results reveal that conventional random partitioning spuriously inflates the coefficient of determination R 2 ≈ 0.95 , whereas Extreme Gradient Boosting (XGBoost) achieves a realistic R 2 of 0.78 for cross-group extrapolation. Physical consistency is assessed primarily through fixed-profile RI – S a monotonicity: an S a -monotone constrained XGBoost model eliminates non-physical increases of predicted RI with rising demand while retaining accuracy comparable to the unconstrained baseline. In leave-one-class-out extrapolation tests on unseen single-span bridges, performance degradation correlates with domain shift quantified via Jensen–Shannon divergence and Population Stability Index, enabling proactive identification of extrapolation risks. From an engineering perspective, the framework establishes interpretable structure–performance relationships with explicitly defined applicability boundaries. It supports high-throughput batch inference and achieves strong portfolio-level binary screening performance when the continuous RI is thresholded at 60 (F1-score = 0.94), providing a verifiable baseline for regional emergency resource allocation. Finally, using the National Bridge Inventory, the constrained model is applied to screen 3536 highway bridges in Los Angeles County under the 1994 Northridge earthquake scenario with USGS ShakeMap S a ( T = 1.0 ) . Following a two-stage applicability check—intensity support examination and structural shift review—macroscopic inventory data are converted into post-earthquake spatial screening maps that illustrate the regional workflow.