Variational Bayesian model updating using normalizing flows

Abstract

We investigate the use of normalizing flows to approximate transport maps from tractable reference densities to complex Bayesian posterior distributions for Bayesian model updating. A Gaussian process (GP) surrogate with active sampling is used to provide a differentiable target density for optimizing the transport map. While results show normalizing flows can capture multimodal behavior in a simple example, further work is needed to refine the active sampling strategy and enable mode identification in the GP surrogate for robust multimodal density approximation.

Full Paper