Signatures of warm dark matter in the cosmological density fields extracted using Machine Learning

Abstract

We aim to construct a machine-learning approach that allows for a pixel-by-pixel reconstruction of the intergalactic medium (IGM) density field for various warm dark matter (WDM) models using the Lyman-alpha forest. With this regression machinery, we constrain the mass of a potential WDM particle from observed Lyman-alpha sightlines directly from the density field. We design and train a Bayesian neural network on the supervised regression task of recovering the optical depth-weighted density field $Δ_τ$ as well as its reconstruction uncertainty from the Lyman-alpha forest flux field. We utilise the Sherwood-Relics simulation suite at $4.1\leq z \leq 5.0$ as the main training and validation dataset. Leveraging the density field recovered by our neural network, we construct an inference pipeline to constrain the WDM particle masses based on the probability distribution function of the density fields. We find that our trained Bayesian neural network can accurately recover within a $1σ$ error $\geq 85\%$ of the density field pixels from a validation simulated dataset that encompasses multiple WDM and thermal models of the IGM. When predicting on Lyman-alpha skewers generated using the alternative hydrodynamical code Nyx not included in the training data, we find a $1σ$ accuracy rate $\geq 75\%$. We consider 2 samples of observed Lyman-alpha spectra from the UVES and GHOST instruments, at $z=4.4$ and $z=4.9$ respectively and fit the density fields recovered by our Bayesian neural network to constrain WDM masses. We find lower bounds on the WDM particle mass of $m_{\mathrm{WDM}} \geq 3.8$ KeV and $m_{\mathrm{WDM}} \geq 2.2$ KeV at $2σ$ confidence, respectively. We are able to match current state-of-the-art WDM particle mass constraints using up to 40 times less observational data than Markov Chain Monte Carlo techniques based on the Lyman-alpha forest power spectrum.