Simulating radio observational image data with deep generative models

Abstract

Radio astronomical surveys have reached a regime in which the volume, complexity, and scientific value of the data place increasing demands on methods for simulation, calibration, and automated analysis. In particular, the development and validation of such methods require realistic synthetic data that reproduce the statistical and morphological properties of observed radio sources. In this thesis, I investigated deep generative modelling approaches for the synthesis of realistic radio images and survey data with the aim of providing controllable and efficient simulation tools for radio astronomy. I first developed diffusion models trained on observations from the LOFAR Two-Metre Sky Survey and the VLA FIRST survey to generate realistic images of individual radio galaxies. By conditioning the models on peak flux values or morphological class labels, I was able to preserve intensity information and control source morphology during sampling. The resulting models produced high-quality images that closely matched the statistical properties of the training data. In particular, classifier-based evaluations confirmed that the generated images reproduced morphological classes reliably. Moreover, I demonstrated that high-fidelity generation was achievable with only 25 sampling steps, representing a substantial improvement in sampling efficiency for radio astronomical applications. Building on this, I developed software for machine learning-assisted simulations of realistic LOFAR observations. For this purpose, I used a diffusion model to synthesize individual radio galaxies with controlled size and assembled them into a synthetic sky model based on a simulated input catalogue obtained from existing software. This sky model was then converted into visibilities, realistic noise was added, and the data were deconvolved to produce simulated sky maps. I showed that the resulting observations reproduced key properties of real LOFAR data, including the flux and size distributions of sources as well as the overall sensitivity of the maps. The developed framework is flexible and may be adapted to other radio instruments. Finally, I extended these ideas by implementing a latent diffusion model for the synthesis of larger radio survey map cutouts. By operating in a compressed latent representation, the model enabled the generation of large images with good quality. I encoded source information on position, brightness, and size from the public source catalogue as conditioning context, and I trained the model additionally for image inpainting to enable sequential sampling for large map generation. To account for the high dynamic range of the data, I also implemented a custom pixel scaling function. This allowed me to generate realistic LOFAR cutouts with precise control over source positions and fluxes, and to construct maps of arbitrary size by combining multiple samples in a consistent manner. Overall, I showed that diffusion-based generative models provide a powerful and flexible approach to producing realistic synthetic radio data. The methods developed in this dissertation contribute to the simulation of radio survey observations and to the development of analysis methods for current and future radio telescopes.