Complementary Theory and Data Driven Analysis for Sky Surveys
Computationally intensive simulations make systematic uncertainties hard to quantify when compared to real data. Guided by unsupervised domain adpation in machine learning, we show how we build an end-to-end differentiable pipeline that is capable of accelerating and correcting simulations by learning from large survey data sets. As an application to astrophysics, we build an emulator to stellar spectra simulations, which we complement with an unsupervised generative network learned from spectroscopic survey data. The machine learning pipeline is then used for accurate stellar parameters estimation and to show the potential of the method to discover new spectral features associated to elemental abundances.