Beyond fragility: physics-driven neural surrogates for seismic resilience prediction of bridges

Traditional fragility-based methods are rigorous, but they can be computationally intensive and difficult to scale to large bridge inventories, particularly when resilience assessments must propagate fragility outputs through functionality and recovery models for time-dependent decision support. This study presents a physics-driven neural surrogate framework that complements fragility-informed workflows by directly predicting a bridge-level seismic resilience index as a continuous system metric. Using pre-1971 concrete box-girder bridges as a case study, we generate a simulation-informed dataset from high-fidelity nonlinear time-history analyses in OpenSees, covering 1,600 bridge-ground motion scenarios. A multilayer perceptron (MLP) model is trained with systematic hyperparameter tuning over loss functions, optimizers, network depth, and regularization. The final MLP achieves over 97% prediction accuracy and outperforms baseline ensemble learning models. By learning directly from physics-based simulations, the proposed surrogate enables rapid and scalable resilience estimation, supporting retrofit prioritization, emergency planning, and resilience-informed design in seismically active regions.