Radio interferometric imaging is an ill-posed inverse problem because the sky brightness distribution is sampled sparsely and irregularly in the Fourier (uv) domain. Incomplete uv-coverage contributes to imaging artefacts, uncertain flux recovery, and distortions of source morphology, particularly for faint and extended emission. Many machine-learning approaches operate on dirty or reconstructed images, after instrumental effects and information loss have already been incorporated into the image representation. Learning directly from complex visibilities offers an alternative: it allows the model to represent the measurement domain continuously and to infer unsampled Fourier components under a learned prior before conventional image formation.
We present the adaptation and high-performance-computing evaluation of a transformer-conditioned neural field for continuous visibility-domain modeling. Sparse complex visibilities, together with their uv coordinates, are encoded by a Transformer into a latent representation. This representation conditions a multilayer perceptron through feature-wise linear modulation. The resulting neural field can be queried at arbitrary Fourier coordinates and therefore does not restrict the reconstructed visibility function to a fixed grid.
We evaluate the approach using simulated observations designed to reproduce key characteristics of LOFAR HBA data. The network is implemented and tested on a multi-GPU HPC system. Our study examines model fidelity and generalization to unsampled coordinates and the computational limits imposed by the number of input visibilities and target image resolution. We also investigate strategies for compressing large visibility datasets into tractable latent representations while retaining information relevant across multiple spatial scales.
We will present preliminary visibility-modeling results, computational and memory-scaling benchmarks, and the methodological developments required to extend the approach to real interferometric observations. These results identify both the potential of continuous visibility representations and the remaining challenges associated with data volume and realistic instrumental sampling.