Unified Bayesian Networks for Uncertain Inputs and Partial Model Ensembles

The purpose of this project is to develop a framework for probabilistic reasoning where the interaction of computational models is governed by a Bayesian network, and to demonstrate its potential in the assessment of thermo-structural failure probabilities for vehicles operating at hypersonic speeds. Current approaches in Uncertainty Quantification (UQ) focus mainly on the solution of direct or inverse problems. This separation poses limitations to tackle realistic problems in probabilistic reasoning, e.g. determining output sensitivities and optimal experimental design for complex multi-physics simulators. The goal of this project is to go beyond traditional approaches in UQ, integrating numerical solvers into frameworks for probabilistic reasoning. We leverage Bayesian networks to combine uncertain inputs, models of various fidelities and expert opinion with experimental data. We also use advanced response surface analysis through multi-element surrogates and multi-fidelity estimators to speed up online inference through belief propagation.

(View full image)