Mapping Dark Matter to the Lyman-α Forest with Neural Networks

by Dr Bhaskar Arya (IIT-Kanpur)

Friday 10 Oct 2025, 14:30 → 15:30 Asia/Kolkata

304A

The Lyman-α forest is a powerful probe of the intergalactic medium (IGM) and the underlying matter density. Traditional semi-analytic approaches, such as the lognormal model, have been previously used to efficiently generate mock spectra. In this regard, I will first discuss the IGM parameter estimates obtained using the approximation. We show that the model is able to recover thermal history reasonably well at z ∼ 2.5. However, the model struggles to robustly recover the key parameter, the hydrogen photoionization rate, due to its simplified treatment of IGM physics and density evolution. To overcome this challenge, we explore a machine learning–based framework that employs neural networks trained on simulation data to directly predict Lyman-α spectra from the underlying dark matter distribution. By learning non-linear mappings beyond the lognormal approximation, the network captures subtle features in the transmitted flux. We demonstrate the efficacy of this approach in accurately predicting the spectra at a fixed cosmology and discuss the path forward in emulating the Ly-α forest for various cosmological and dark matter models.