Uncertainty-aware fragility modeling of urban building exteriors subjected to hurricane-induced windborne debris with conditional generative adversarial nets

Building exteriors are highly susceptible to windborne debris damage during hurricanes, a concern that is growing in urgency amid climate change. Fragility functions that estimate the probability of exterior failure as a function of hurricane intensity are commonly used to assess this risk. However, developing the fragility function often requires numerous calls of expensive physics-based models. To address this challenge, this study proposes a machine learning-based approach that leverages conditional generative adversarial networks (cGAN) and its generative capability to efficiently and accurately develop fragility functions. cGAN model is introduced to model the complex high-dimensional distribution of the maximum impact momentum of windborne debris on the building exterior conditioned on hurricane intensity measures. To ensure the accuracy and reliability of cGAN, a stochastic simulation-guided training strategy is proposed that treats data preparation as an augmented stochastic simulation problem. This approach enhances exploration of the input space and enables the formulation of a novel training performance metric tailored toward fragility modeling. Once trained, the cGAN model can rapidly generate extensive damage realizations under various hurricane scenarios, which facilitates the construction of fragility functions. Furthermore, Monte Carlo dropout is introduced to efficiently quantify the uncertainty of the fragility function established using cGAN. Such uncertainty provides end users with a measure of confidence in the derived fragility estimates. An illustrative example is presented to validate the efficacy of the proposed method. Results show that the proposed performance metric is effective in guiding the selection of a reliable cGAN model that provides satisfactory damage and fragility estimates, and the proposed uncertainty quantification method can help provide rational confidence in the estimates.