Physics-guided machine learning fragility assessment of self-centering-braced RC bridges using CatBoost demand surrogates

This study proposes a physics-guided machine learning framework for the seismic fragility assessment of reinforced-concrete bridges equipped with self-centering braces (SCBs), with conventional restrainers used as the benchmark system. Incremental Dynamic Analyses (IDAs) were performed in OpenSeesPy to build a demand-capacity database covering eight engineering demand parameters (EDPs). Eight machine learning models, including artificial neural networks, random forest, XGBoost, LightGBM, CatBoost (CTB), support vector regression, k-nearest neighbors, and linear regression, were optimized using Optuna and compared in terms of predictive performance. Among them, CTB provided the best overall accuracy, with cross-validated mean R² values greater than 0.97 for all EDPs. The CTB-based demand predictions were combined with lognormal capacity models to define probabilistic exceedance labels for different damage states. These labels were then used to train a physics-guided neural network (PGNN) for fragility curve generation. The PGNN incorporates monotonicity and damage-state ordering constraints into the loss function, together with gradient guidance derived from SHAP and ICE analyses, to produce smooth, monotonic, and non-crossing fragility curves. Component-level fragilities were assembled into system-level fragility curves using a series-parallel OR/AND formulation. At the system level, the median Sa(1.0 s) for Slight damage increased from 0.066 g for the restrainer bridge to 0.219 g for the SCB bridge (+229.5%), while the Moderate-state medians differed by only 1.9%. For Extensive damage, the SCB median intensity was 7.7% lower than that of the restrainer bridge. At Sa(1.0 s) = 1.5 g, the SCB configuration showed 15.6% and 24.2% higher exceedance probabilities at the Extensive and Complete damage states, respectively. Overall, the proposed CTB-PGNN framework reduces the limitations of fixed-form regression models and provides physically consistent fragility estimates, offering a useful tool for retrofit prioritization and performance-based seismic design.