Simulation and Sensor Optimization for Multifidelity Dynamics Model Calibration

Abstract

This paper studies a multifidelity resource optimization methodology for simulation and data collection in the calibration of dynamics model parameters. Nonlinear dynamics systems require high-fidelity modeling; however, expensive high-fidelity simulations cannot be used in Bayesian model calibration: Markov chain Monte Carlo sampling requires thousands of runs to calculate the posterior distributions. The use of low-fidelity modeling is also not appropriate, because it leads to significant error in calibration. A multifidelity approach is pursued, where a limited number of high-fidelity runs are used to improve a low-fidelity-based surrogate model, resulting in stronger physics-informed priors for calibration with experimental data. Two optimization approaches are studied: the first selects high-fidelity model runs that maximize the information gain in the modeling stage within the multifidelity approach, and the second selects the number and location of the sensors in the experimental setup to maximize the information gain from the experiments. The proposed methodologies are illustrated for a curved panel subjected to acoustic and nonuniform thermal loading. Two models of different fidelity are employed to calibrate the damping and model errors. The optimization methodology considers two complicating factors: 1) the damping behavior is input-dependent, and 2) the sensor uncertainty is affected by temperature.