The 21-cm signal of neutral hydrogen is a sensitive probe of the Epoch of Reionization (EoR) and Cosmic Dawn. Currently operating radio telescopes have ushered in a data-driven era of 21-cm cosmology, providing the first constraints on the astrophysical properties of sources that drive this signal. However, extracting astrophysical information from the data is highly non-trivial and requires the rapid generation of theoretical templates over a wide range of astrophysical parameters. To this end emulators are often employed, with previous efforts focused on predicting the power spectrum. In this work we introduce 21cmGEM - the first emulator of the global 21-cm signal from Cosmic Dawn and the EoR. The smoothness of the output signal is guaranteed by design. We train neural networks to predict the cosmological signal using a database of ~30,000 simulated signals which were created by varying seven astrophysical parameters: the star formation efficiency and the minimal mass of star-forming halos; the efficiency of the first X-ray sources and their spectrum parameterized by spectral index and the low energy cutoff; the mean free path of ionizing photons and the CMB optical depth. We test the performance with a set of ~2,000 simulated signals, showing that the relative error in the prediction has an r.m.s. of 0.0159. The algorithm is efficient, with a running time per parameter set of 0.16 sec. Finally, we use the database of models to check the robustness of relations between the features of the global signal and the astrophysical parameters that we previously reported.
تحميل البحث