Séminaire IPAG

---

Solving Inverse Problems in Astronomy with Invertible Neural Networks

jeudi 27 juin 2024 - 11h00

Victor Ksoll - U. Heidelberg/ Center for astronomy
--- 
Astronomy is host to a wide variety of so called inverse problems. Given a system, which is described by a set of fundamental physical parameters that give rise to a corresponding set of observable quantities, the inverse problem refers to the task of recovering the underlying parameters of the system from the set of observables. In many cases the mapping from physical parameters to observables is very well understood via e.g. sophisticated simulations, whereas the inverse problem is much more difficult. Due to an inherent loss of information in the forward mapping, the inverse problem is often subject to degeneracy, i.e. different sets of physical parameters giving rise to similar sets of observables, such that the recovery of the underlying physical parameters from the observations becomes complicated. Invertible Neural Networks (INNs) are a family of deep learning models, which are particularly well-suited to solving inverse problems. Introducing a set of latent variables to capture the information loss of the forward process, INNs can estimate the full posterior distribution of the target physical parameters given an observation, allowing them to capture and highlight the degeneracies that may occur in a given inverse problem. We are developing INNs for a variety of problems in astronomy, including e.g. the recovery of physical parameters of stars from photometric observations, the characterisation of stellar spectra or the 3D reconstruction of dust in star-forming cloud cores from dust emission maps. In this talk, I will introduce INNs as inverse problem solvers and provide an overview of the astronomical subjects that we have tackled with INNs in our work so far.

---

Hôtes : Isabelle Joncour

Salle Manuel Forestini, 414 rue de la piscine, 38400 Saint Martin d'Hères