Semi-supervised rotation measure deconvolution and its application to MeerKAT observations of galaxy clusters

Context. Faraday rotation contains information about the magnetic field structure along the line of sight and is an important instrument in the study of cosmic magnetism. Traditional Faraday spectrum deconvolution methods such as RMCLEAN face challenges in resolving complex Faraday dispersion functions and handling large datasets. Aims. We developed a deep learning deconvolution model to enhance the accuracy and efficiency of extracting Faraday rotation measures from radio astronomical data, specifically targeting data from the MeerKAT Galaxy Cluster Legacy Survey (MGCLS). Methods. We used semi-supervised learning, where the model simultaneously recreates the data and minimizes the difference between the output and the true signal of synthetic data. Performance comparisons with RMCLEAN were conducted on simulated as well as real data for the galaxy cluster Abell 3376. Results. Our semi-supervised model is able to recover the Faraday dispersion for extended rotation measure (RM) components, while accounting for bandwidth depolarization, resulting in a higher sensitivity for high-RM signals, given the spectral configuration of MGCLS. Applied to observations of Abell 3376, we find detailed magnetic field structures in the radio relics, and several active galactic nuclei. We also applied our model to MeerKAT data of Abell 85, Abell 168, Abell 194, Abell 3186, and Abell 3667. Conclusions. We have demonstrated the potential of deep learning for improving RM synthesis deconvolution, providing accurate reconstructions at a high computational efficiency. In addition to validating our data against existing polarization maps, we find new and refined features in diffuse sources imaged with MeerKAT.