Using Neural Networks to Constrain Cosmology from Neutral Hydrogen Distributions

Upcoming measurements of the redshifted 21-cm line are poised to make the first large breakthroughs in the reionization era. By directly imaging neutral hydrogen, key constraints can be gleaned for cosmology and astrophysics, specifically ΛCDM parameters and astrophysical parameters. The amount of data produced from these observations is expected to be immense and optimal methods need to be developed in order to extract the relevant cosmological information. Here we neutral hydrogen data from the astrophysical simulations, SIMBA and IllustrisTNG, to train neural networks to perform likelihood-free inference on ΛCDM parameters. Power spectra of the simulation data is used as a summary statistic and used as training data. We examine how the information at smaller scales is related to the underlying cosmology and astrophysical models. Robustness of our method is examined by testing against different simulation techniques.