Bayesian structural model updating under prior misspecification using conditional normalizing flows

This study proposes a likelihood-free Bayesian structural model updating framework that is robust under prior misspecification. The approach employs Conditional Normalizing Flows to learn posterior distributions directly from simulated structural responses. To mitigate the degradation caused by imperfect engineering priors, a proposal-based training strategy and prior-proposal correction are introduced to ensure consistent posterior inference. Benchmark and nonlinear structural case studies demonstrate that the method accurately recovers multimodal posteriors and enables efficient structural model updating using high-dimensional acceleration time-history data.