Optimization methods for numerical model calibration in structural engineering

Abstract

The finite element method (FEM) is widely used in civil engineering to simulate structural behavior. To address uncertainties, model calibration using experimental modal data may be employed, allowing for the adjustment of hard-to-measure properties to enhance the model’s dynamic response. This study evaluates the performance of two metaheuristic algorithms, Particle Swarm Optimization (PSO) and Genetic Algorithm (GA), alongside Bayesian Optimization (BO). A simply supported beam was modeled in ANSYS to generate pseudo-experimental modal data, including ten natural frequencies and corresponding mode shapes, with uncertainties represented by five design variables. The search spaces were defined based on literature values, and initial points were generated using Latin Hypercube Sampling (LHS). The optimization process for PSO and GA was conducted in MATLAB, while BO was implemented in Python, interfacing with ANSYS to iteratively refine model parameters and minimize discrepancies in natural frequencies and modal assurance criterion (MAC) values between the numerical and reference models. Results indicate that PSO demonstrated better performance among the metaheuristic algorithms, achieving a more accurate cost function value than BO. However, BO delivered strong results with a restricted number of evaluations, which makes it particularly advantageous for complex problems. Both methodologies are applicable; however, BO stands out for its computational efficiency.