A Physics-Constrained Surrogate Model for Multi-Hazard Collapse Assessment of Buildings Under Post-Fire Concurrent Wind-Earthquake Loading

Abstract

Conventional structural design frameworks assess natural hazards as statistically independent phenomena, a practice that can lead to significant underestimation of risk for structures subjected to sequential or concurrent hazards. The generation of probabilistic fragility functions under such cascading loads, particularly for post-fire seismic events, presents a computational barrier for standard non-linear dynamic analysis. To address this barrier, this study introduces a comprehensive computational framework centered on a physics-constrained neural network (PCNN) to serve as a high-fidelity surrogate model. The framework first uses a non-linear 12-degree-of-freedom structural model to generate a baseline dataset of collapse times under post-fire, concurrent wind-earthquake loading via the computationally efficient endurance time (ET) method, confirming that wind effects are negligible under ambient conditions and that the framework correctly identifies this hazard hierarchy without prior labeling, while fire and seismic parameters dominate. This dataset is subsequently used to train the PCNN, which is validated to achieve exceptional predictive accuracy (R2= 0.991), performing on par with a state-of-the-art Random Forest model while enforcing physical constraints. A feature importance analysis confirmed that structural collapse is dominated by fire intensity (≈55%) and initial structural period (≈45%). The validated PCNN is then applied to demonstrate the framework’s capability, rapidly generating fragility curves that quantify the catastrophic effect of fire on seismic resilience. This analysis reveals that a severe 800 °C localized fire reduces the structure’s median collapse capacity by 94.7%, thereby establishing the proposed framework as a successful template for tackling complex, non-linear problems in multi-hazard engineering.