In this thesis, we study the procedures for calibrating numerical models of structural systems to match the experimental data. Two possible approaches are discussed in more detail - the response surface method and the Bayesian approach. Using the response surface method, we approximate the response with a function of chosen parameters and search for values that minimize the error between the approximation and the experiments. The Bayesian approach, however, is based on the assumption that parameter values are a random variable with a certain probability distribution and searches for the most probable solutions. Two examples are considered: a simple beam and a structure consisting of a steel frame and a steel-reinforced concrete slab supported on a corrugated steel plate; experimental results are available for the latter. Numerical analyses are performed with the commercial software Ansys. Calibration of the numerical model by both methods is performed using our own computer code written in the Python programming language. It turns out that successful calibration with a response surface is only possible if the initial numerical model is a good approximation of the real system. However, satisfactory results when calibrating a numerical model with moderate model error can also be achieved with the Bayesian approach. Regardless of the chosen method, calibration of a numerical model should be used only after discretization and model error is minimized.
|
