Quantifying Model Selection Uncertainty in Structural Analysis: Methodology and Application

With increasing focus on complex engineering systems under rare events, computational models are critical for predictions due to the scarcity or absence of data. However, selecting an appropriate model can be challenging. Using a single model without available test calibration could result in significant bias in performance predictions. A case study of a steel moment frame reveals substantial discrepancies between predictions given different model selections, particularly under increasing ground motion intensities. To address this challenge, this paper proposes a model averaging methodology based on the logic tree approach to incorporate multiple models and capture their full spectrum of potential outcomes. A combination of dimensionality reduction techniques and Bayesian inference is employed to determine the weight of each model candidate. The weight is represented as a Dirichlet prior, which is determined by incorporating expert knowledge and structural component testing results. This approach enables the assignment of a more informative prior even when no observational data is available, while also leaving room for updating the prior if data becomes available. By leveraging model averaging and uncertainty propagation techniques to address model uncertainty in structural analysis, a response interval is then generated for the quantity of interest. This methodology provides a more comprehensive, informative, and feasible assessment to address model selection uncertainty to improve predictive capabilities and enhance the ability of engineers to treat design in a probabilistic manner. In its present formulation, however, the approach is applicable to precollapse response regimes where finite response quantities can be defined and does not explicitly address collapse‐state realizations.