Uncertainty Quantification and Confidence Assessment in Time-Dependent, Multidisciplinary Simulations

Abstract

Multidisciplinary simulations are often assembled based on limited data and approximate individual and partial-physics model components. Methodologies are proposed to address the following three challenges in the context of both inverse and forward uncertainty quantification (UQ) problems: 1) computational expense of multidisciplinary simulations, 2) error accumulation across multiple models and over time, and 3) uncertainty due to the availability of limited data. For inverse problems, the development of a segmented Bayesian model calibration strategy reduces the computational effort of calibration when multiple information sources are available. Further, prediction confidence is improved by reducing the uncertainty that aggregates between coupled analyses and through time using a partitioned approach to calibrate model errors. Methodology contributions for the forward problem include an efficient global sensitivity analysis method (to support dimension reduction) that incorporates existing model calibration results and an optimization framework that balances prediction confidence and computational effort to select variable model fidelity in multidisciplinary simulations. These methods are illustrated with time-dependent, aerothermoelastic analyses of airfoils subjected to high-speed flow.