Physics-enhanced machine learning surrogate modelling for real-time seismic fragility assessment of reinforced concrete bridges

Seismic fragility assessment is crucial for disaster management and emergency response, yet practical large-scale applications face two main challenges: (1) accurately identifying critical limit states of reinforced concrete (RC) pier columns, and (2) rapidly assessing seismic demands of structures. This study proposes a physical knowledge-enhanced machine learning (PEML) framework for seismic fragility assessment of regular bridges. Unlike traditional Physics-Informed Neural Networks (PINN), which rely on explicit mathematical formulations that are often unavailable for RC components, the PEML framework integrates qualitative physical trends and empirical relationships governing RC column behavior into a data-driven modeling process. By explicitly accounting for uncertainties and their interdependencies, an integrated probabilistic limit state model is developed. Furthermore, nonlinear time-history analysis is streamlined into a Deep Neural Network (DNN) for real-time seismic demand prediction. A case study demonstrates that the PEML-DNN framework achieves 90 % accuracy with computational efficiency 300 times greater than traditional nonlinear time-history analysis methods. This approach provides a unique, scalable, and physically informed solution for regional-scale seismic fragility assessment of bridges, supporting rapid and probabilistically robust decision-making.