Abstract

Inferring model parameters from observational data of a physical system is the setup for many inverse problems. Solving these kinds of problems can give key insight into the state of a system for quantities that are not directly observable, such as material properties. In this talk, we discuss a population-based perspective on solving inverse problems where the data available comes from a collection of physical systems and we are interested in characterising the (indirectly observable) properties of these systems at a distributional level. We call this: calibrating priors from indirect data. Furthermore, we show how this can be accomplished while concurrently learning ML-based surrogates which capture the behaviour of the physical systems of interest.