Deep Learning the Intergalactic Medium using Lyman-alpha Forest at $ 4 \leq z \leq 5$
Unveiling the thermal history of the intergalactic medium (IGM) at $4 \leq z \leq 5$ holds the potential to reveal early onset HeII reionization or lingering thermal fluctuations from HI reionization. We set out to reconstruct the IGM gas properties along simulated Lyman-alpha forest data on pixel-by-pixel basis, employing deep Bayesian neural networks. Our approach leverages the Sherwood-Relics simulation suite, consisting of diverse thermal histories, to generate mock spectra. Our convolutional and residual networks with likelihood metric predicts the Ly$α$ optical depth-weighted density or temperature for each pixel in the Ly$α$ forest skewer. We find that our network can successfully reproduce IGM conditions with high fidelity across range of instrumental signal-to-noise. These predictions are subsequently translated into the temperature-density plane, facilitating the derivation of reliable constraints on thermal parameters. This allows us to estimate temperature at mean cosmic density, $T_{\rm 0}$ with one sigma confidence $δT_{\rm 0} \sim 1000{\rm K}$ using only one $20$Mpc/h sightline ($Δz\simeq 0.04$) with a typical reionization history. Existing studies utilize redshift pathlength comparable to $Δz\simeq 4$ for similar constraints. We can also provide more stringent constraints on the slope ($1σ$ confidence interval $δ{\rm γ} \lesssim 0.1$) of the IGM temperature-density relation as compared to other traditional approaches. We test the reconstruction on a single high signal-to-noise observed spectrum ($20$ Mpc/h segment), and recover thermal parameters consistent with current measurements. This machine learning approach has the potential to provide accurate yet robust measurements of IGM thermal history at the redshifts in question.