Abstract
One of the challenges of astrophysics and cosmology is to study complex non-linear processes using an often limited number of multi-component observations. This task is made all the more difficult by the fact that the physical modeling of these processes is not always complete, which implies relying solely on available observations, without any prior training stage. In this seminar, we look at how to build efficient low-dimensional models that take into account the physical character and regularity of the processes under study. These maximum entropy models, built from scattering transform representations, can be constructed directly from observational data.We thendiscuss how these tools can be used to develop new types of component separation, enabling us in particular to estimate the statistics, and thus build a model, of unknown processes from multicomponent observations.