Speaker
Description
Multiple-line intensity mappings can trace large-scale structure from now up to reionization, when the first galaxies formed. LIM 3D tomography provides valuable insights into cosmological structure growth, the intergalactic medium, and the properties and environment of ionizing sources. These mappings need updated modeling and inference methods for LIM cosmology because of the large scale and temporal range, highly non-Gaussian measured LIM signals, and large quantity of data. Our group in Heidelberg has developed a framework to both create mock simulations of large-scale tomographic LIM mappings (such as Lyman-alpha, H-alpha, and 21cm) using fast semi-numerical (approx. hydrodynamical) simulations coupled with scaling relations and to apply simulation-based inference to LIM tomography. As highlighted with 21cm tomography, in this talk I will show that a convolutional summary network coupled to a conditional invertible network through a physics-inspired latent representation performs optimal and extremely fast posterior inference of astrophysical and cosmological parameters. Our method's ability to detect non-Gaussian information makes it a viable alternative to power spectra as a summary statistic for LIM inference. We further investigate the generation of simulated LIM volumes using advanced machine learning, in particular generative networks, and discuss the potential synergy with our previous tomographic mocks.
