Thesis defense of Kevin Schmidt

Radio interferometers achieve the highest resolutions at the cost of sparse data coverage. Incompletely sampled sky distributions in Fourier space result in noise artifacts in the source reconstructions. Established cleaning software is often time-consuming and lacks reproducibility. In this work, I propose a novel cleaning strategy for radio interferometer data based on convolutional neural networks to adjust current analysis strategies to the new telescope standards. This deep learning-based approach will allow a straightforward application that generates reproducible results with short reconstruction times. The newly developed simulation chain enables the simulation of Gaussian radio galaxies and mimics observations by radio interferometers. By iterative adjustments, complexity is increased, ending up with a simulated data set comparable to MOJAVE archive data. In parallel, the deep learning framework radionets, capable of uncertainty estimates, is built to analyze large data samples with comparable characteristics. The improved reconstruction technique will allow scientists to focus more on their scientific analysis and omit a vast workload on data cleaning tasks. Various evaluation techniques are created to quantify the trained deep learning models’ reconstruction quality. Furthermore, the reconstruction performance is assessed on input data with different noise levels by comparing the resulting predictions with the simulated source distributions. Source orientations and sizes are well reproduced, while the recovered intensities show substantial scatter, albeit not worse than existing methods without fine-tuning. Finally, all improvements are combined to train a deep learning model suitable to evaluate MOJAVE observations.

Back