Beam Prediction Method for Antenna Arrays in Radio Astronomy via Artificial Neural Networks

With the SKA-Low construction underway, we explore the potential of using artificial neural networks to generate embedded element patterns (EEPs) at fine frequency channels. Exploiting the existing FEKO-simulated database of EEPs, we trained a prediction model that generates 512 EEPs (for 256 antennas with two polarizations) in only 1.28 sec. To perform the model evaluation, we derived specific performance metrics that are based on simulation tolerance threshold and consider the level of error that would have a potential impact on observations. High accuracy in EEP prediction is achieved, with significant improvements in computational efficiency compared to the full-wave simulations.