Using data-consistent inversion to build population-informed priors for Bayesian inference
Please login to view abstract download link
Bayesian inference is an essential tool for building predictive digital twins as it incorporates information from observational data to reduce uncertainties in the underlying computational models. However, data from an individual physical asset may be sparse, requiring highly-informative priors for Bayesian inference; in practice, the physical knowledge needed to construct such priors may not be available. Consequently, this work presents a novel approach for leveraging data from a population of related assets to construct informative Bayesian priors. Specifically, we use data-consistent inversion to estimate population-informed priors that improve inference on an individual asset. We demonstrate this approach on a computational mechanics application. Numerical examples show that utilizing population-informed priors significantly increases the Kullback–Leibler divergence, i.e. the information gain from the posterior to the prior, in comparison to standard prior specification. These results are complemented with theory for linear-Gaussian inference that establishes the conditions under which using our approach is guaranteed to improve posterior estimates of uncertainty.
