Title: Machine learning galaxy properties from 21 cm lightcones: impact of network architectures and signal contamination

Abstract: Imaging the cosmic 21 cm signal will map out the first billion years of our Universe. The resulting 3D lightcone (LC) will encode the properties of the unseen first galaxies and physical cosmology. Unfortunately, there is no obvious summary statistic to use when interpreting this non-Gaussian image, and the commonly-used power spectrum may waste valuable information. Here we build on previous work using Convolutional Neural Networks (CNNs) to infer astrophysical parameters directly from 21 cm LC images. Guided by the properties of LCs, we combine recurrent layers characterizing evolution along the redshift axis with 2D convolutional layers characterizing local correlations in the sky-plane. Such Recursive Neural Networks (RNNs) are known for efficiently learning temporal correlations in sequential data. Using a large database of simulated cosmic 21 cm LCs, we confirm that RNNs outperform previously-used CNNs in recovering UV and X-ray galaxy properties, reducing the mean squared parameter estimation error by factors of $\sim 2 - 8$. We also corrupt the cosmic signal by adding noise expected from a 1000 h integration with the Square Kilometre Array, as well as excising a foreground-contaminated ''horizon wedge''. Parameter prediction errors increase when the NNs are trained on these contaminated LC images, though recovery is still good even in the most pessimistic case (with $R^2 \ge 0.5 - 0.95$). However, we find no notable differences in performance between network architectures on the contaminated images. We argue this is due to the size of our dataset, highlighting the need for larger datasets and/or better data augmentation in order to maximize the potential of NNs in 21 cm parameter estimation.