Parth Nayak (LMU Munich)
Title: LyαNNA: A Deep Learning Field-level Inference Machine for the Lyman-α Forest
Abstract: The Lyα forest, a congregation of absorption lines in the observed spectra of distant quasars imprinted by the intervening intergalactic medium (IGM), is a powerful probe of cosmic structures and thermodynamic properties of the universe at cosmological redshifts of z ~ 2 - 6. The inference of those characteristics from the Lyα forest conventionally relies on established summary statistics of the transmitted flux field, however, this inevitably leads to loss of some relevant parts of the information carried by the field. I will talk about a deep learning approach we have developed (described in this paper 2311.02167) in order to help remedy this problem for large future datasets. This framework consists of a 1D ResNet that aims to extract all the pertinent features directly from the field and to optimally compress them into statistical vectors, thereby facilitating Bayesian inference. We observe factors-of-a-few improvement in the posterior constraints on the IGM thermal parameters over the most commonly used summary statistics and their combinations, when employing our machinery. I will also briefly touch upon the ongoing efforts to include observational and modelling systematic uncertainties in the data for creating a more robust pipeline. Our case study demonstrates the untapped potential of deep learning techniques for inference from large and complex datasets at the field level in astrophysics and cosmology.