Finite Element Model Calibration Using Genetic Algorithm and Bayesian Optimization

Resumo

Calibration techniques are essential for applying finite element models (FEM) in structural assessment and monitoring, given the uncertainties inherent to modeling. Among indirect calibration approaches, metaheuristic algorithms stand out for their ability to explore large search spaces and avoid local minima, but they require many costly model evaluations. To address these limitations, surrogate-based approaches have emerged, approximating the behavior of the objective function and substantially reducing the number of physical model evaluations needed to obtain suitable solutions. Bayesian Optimization (BO), for example, employs a probabilistic model to estimate the target function, enabling not only efficient optimization but also providing quantitative insights into parameter sensitivity and prediction uncertainty. In this work, the performance of a widely used metaheuristic, the Genetic Algorithm (GA), is compared with that of BO in the calibration of a synthetic bridge model, using modal data with added noise. The results demonstrate BO’s versatility, competitive accuracy under noisy conditions, and markedly lower computational cost, while also indicating that hybrid strategies can further improve precision.