PI: Harry Bevins

Project Title: Beam Modelling and Emulation for REACH

Background: One of the most pressing open questions in observational cosmology concerns the nature of the first stars and galaxies. The Cosmic Dawn and Epoch of Reionization (300 million to one billion years after the Big Bang) mark the transformation of the Universe from a sea of neutral hydrogen into the ionized, galaxy-rich cosmos observed today. A powerful probe of these epochs is the sky-averaged or global 21cm signal from neutral hydrogen, which encodes information about the thermal and ionization state of the intergalactic medium as the first luminous sources formed. The Radio Experiment for the Analysis of Cosmic Hydrogen (REACH), currently collecting commissioning data at the Karoo Radio Observatory in South Africa, aims to detect this signal across the 50–135 MHz. A fundamental challenge is accurately characterising the antenna’s frequency-dependent gain pattern (the ‘beam’): the observed antenna temperature is a convolution of the true sky temperature with the antenna gain, so any inaccuracy in the assumed beam can prevent a detection of the signal or even lead to a false detection. Existing analyses must assume a perfectly-known, fixed beam — an assumption that breaks down due to manufacturing tolerances and environmental changes over time.

DiRAC-Enabled Progress: Using 3.22 million CPU core-hours on the DiRAC Data Intensive system at Cambridge (CSD3), we have generated over 2,000 high-fidelity electromagnetic (EM) simulations of the REACH antenna beam pattern using a Method of Moments (MoM) code. Each simulation models the full frequency-dependent gain across the 50–135 MHz band at 1 MHz resolution, for a set of physical parameters describing the antenna geometry and its surroundings. We have assembled multiple training datasets of increasing complexity, systematically varying:

• Antenna blade height (800–1200 mm) and ground plane height (1000–1500 mm)
• Real and imaginary components of the soil permittivity beneath the antenna
• Lateral position and rotation of the ground plane

These datasets are now being used to train neural network beam emulators. Figure 1 shows an example antenna mesh and the corresponding simulated beam patterns on the hemisphere above REACH, illustrating the dramatic change in morphology across the REACH band with frequency. We have developed a prototype neural network emulator capable of reproducing beam patterns at approximately −20 dB accuracy, and development is ongoing as we expand the training data and emulator architecture towards a production run.

Figure 1. (a) Method of Moments (MoM) electromagnetic simulation mesh of the REACH dipole antenna mounted above its large, serrated ground plane (~33 m diameter). Meshes like this, generated with GMSH on CSD3, are used to define the geometry for each of the 2,000+ beam simulations produced to date. (b) and (c) Simulated gain patterns of the REACH antenna at 60 MHz and 135 MHz, projected on to the hemisphere above REACH (zenith at center, horizon at edge; colour scale in dB relative to peak gain).

Outlook and Future Work: Two papers are in preparation arising from this work. The first will present the beam emulation framework and demonstrate its performance in the REACH analysis pipeline. The second will use the training data to construct a non-parametric basis set for beam reconstruction within the REACH pipeline. The production-run emulator will enable simultaneous fitting of the antenna gain, astrophysical foregrounds, and the 21cm cosmological signal for the first time.