Reducing input uncertainties in damage-based finite element models of nuclear containment structures through Bayesian inference

Numerical modeling plays a key role in assessing the safety and performance of reinforced concrete nuclear containment buildings (NCBs). However, uncertainties arising from material heterogeneity, boundary imperfections and modeling assumptions can significantly affect predictive accuracy. This paper presents a Bayesian updating (BU) framework for reducing such uncertainties in finite element simulations. The methodology employs a Markov-Chain Monte-Carlo (MCMC) algorithm to infer posterior parameters distributions based on experimental evidence. Its implementation is first illustrated through a simply supported beam under a central load, highlighting the influence of the algorithmic settings (number of chains, sample size, proposal scale, and measurement noise) on posterior convergence and uncertainty reduction. The framework is then applied to the PACE-1450 mock-up, a reinforced concrete representative structural volume of NCBs, using a surrogate model to assimilate observed crack counts. Results show significant variance reduction for key parameters, an improved agreement with experimental data and an enhanced predictive capability of damage-based numerical models.