Emulation of HI 21-cm signal from the epoch of reionization using vision transformers

Abstract

The Epoch of Reionization (EoR) refers to the era when the first luminous sources started to emit the UV radiation and ionized the neutral hydrogen (HI) present in the intergalactic medium (IGM). This era corresponds to one of the major phase transition of the Universe. Studying the redshifted 21-cm signal coming from the neutral hydrogen in the IGM can provide the solid answers to essential questions about EoR such as exact duration of these epochs, formation and properties of the ionizing sources, the morphology of ionized regions and its evolution. Several ongoing radio experiments are working on the detection of the redshifted 21-cm signal, including LOFAR, HERA, MWA, and uGMRT, and the next generation radio interferometers, such as the SKA, will focus on the era of Cosmic Dawn and Epoch of Reionization (CD-EoR). We need to build forward models of the redshifted 21-cm signal that can interpret the observations made by these radio experiments. The simulations that are generally used for forward modeling are computationally expensive when it comes to rerunning them for a large CD-EoR parameter space. This creates a need to develop an emulator that is computationally cheap and can be employed for exploration of the EoR parameter space. While most existing approaches focus on emulating summary statistics of the 21-cm signal, such methods lose information compared to emulating the 21-cm signal itself. Emulation of the EoR 21-cm field suffers from the fundamental problem of modeling features or fields which have a large dynamic range (e.g. 4-5 orders of magnitude variation). The previous efforts of emulation of the 21-cm signal faced limitations because they didn’t take all length scales into account due to the bias of the model. This led to the underprediction of large and small-scale features depending on the model architecture. Furthermore, these models use fixed astrophysical parameters; hence, they cannot be used for any kind of inference exercise. Taking the drawbacks of these previous model architectures into account, we developed an emulator using Vision Transformers (ViTs) that takes in the entire 3D dark matter density and halo field as inputs, along with the three EoR parameters to predict the hydrogen neutral fraction field. The multi-head self-attention mechanism of the transformer makes it easier to capture the long-range dependencies, making it a perfect candidate for this task. The trained and validated model can quickly and accurately produce the neutral fraction fields for a given set of EoR parameters.