Determining the systemic redshift of Lyman α emitters with neural networks and improving the measured large-scale clustering

We explore how to mitigate the clustering distortions in Lyman α emitter (LAE) samples caused by the misidentification of the Lyman α (⁠Lyα⁠) wavelength in their Lyα line profiles. We use the Lyα line profiles from our previous LAE theoretical model that includes radiative transfer in the interstellar and intergalactic mediums. We introduce a novel approach to measure the systemic redshift of LAEs from their Lyα line using neural networks. In detail, we assume that for a fraction of the whole LAE population their systemic redshift is determined precisely through other spectral features. We then use this subset to train a neural network that predicts the Lyα wavelength given an Lyα line profile. We test two different training sets: (i) the LAEs are selected homogeneously and (ii) only the brightest LAE is selected. In comparison with previous approaches in the literature, our methodology improves significantly the accuracy in determining the Lyα wavelength. In fact, after applying our algorithm in ideal Lyα line profiles, we recover the clustering unperturbed down to 1cMpch−1⁠. Then, we test the performance of our methodology in realistic Lyα line profiles by downgrading their quality. The machine learning technique using the uniform sampling works well even if the Lyα line profile quality is decreased considerably. We conclude that LAE surveys such as HETDEX would benefit from determining with high accuracy the systemic redshift of a subpopulation and applying our methodology to estimate the systemic redshift of the rest of the galaxy sample