Multi-Fidelity Information Fusion for Structural Dynamics Model Calibration
Absi, Ghina Nakad
:
2019-04-02
Abstract
This dissertation develops a novel approach for fusing information from physics models of different levels of fidelity in the Bayesian estimation of system parameters. In order to balance computational effort and accuracy, the proposed method first builds a surrogate model using low-fidelity physics simulations. Then it uses a small number of high-fidelity physics simulations to improve the surrogate model, and uses the improved surrogate for calibration with experimental data. This multi-fidelity strategy facilitates computational efficiency, in surrogate training as well as in Bayesian calibration. Furthermore, the improvement of the surrogate model with high-fidelity results before calibration with experimental data provides stronger, physics-informed priors for the calibration quantities. This is particularly useful when limited experimental data are available, and a reliable, but fast model is needed for calibration.
The multi-fidelity calibration method is extended to the calibration of input-dependent system parameters, where the hyper-parameters of the functional relationships between the input and the parameters are estimated. This extension also takes into consideration the effect of the input on the uncertainty in the sensor measurement.
The multi-fidelity approach is optimized in two ways to maximize the information gain: (1) selecting the high-fidelity simulations to improve the surrogate of the low-fidelity model (simulation optimization); and (2) selecting the experimental sensor configuration (i.e., number and locations of the sensors). The proposed methodology is illustrated for the estimation of damping parameters of a fuselage panel close to the engine in a hypersonic aircraft, which is subjected to acoustic and thermal loading.