This work demonstrates sparse surrogate construction followed by dimensionality reduction and surrogate enabled Bayesian inference of model inputs to achieve data-informed uncertain predictions of an expensive computational model. We will rely on polynomial chaos (PC) surrogates as flexible functional representation of uncertain inputs and outputs. High-dimensionality is tackled by Bayesian compressed sensing leading to a sparse set of polynomial bases and allowing efficient global sensitivity analysis and dimensionality reduction.
Talk given by Cosmin Safta.