DeepShape. II. Source separation for radio weak-lensing measurements using deep learning

For wide field radio observations, all sources within the primary beam contribute to the measured visibilities, and at high source densities, they can also overlap in the image plane. Isolating individual sources, therefore, becomes non-trivial. Accurate source separation is critical for shape measurement, yet traditional approaches either rely on simple model-based separation in the image domain or on computationally intensive joint modelling in visibility space, underscoring the need for scalable and robust methods. Our aim was to develop an end-to-end framework that enables accurate shape measurement in crowded radio fields by explicitly addressing source separation. The goal is to isolate individual galaxy contributions from wide field visibility data, remove contamination from neighbouring sources, and provide a scalable pipeline suitable for next-generation radio weak-lensing surveys. We extended the DeepShape framework to wide field observations by combining faceting-based source isolation with plug-and-play deconvolution. Reconstructed facets were then deblended using a variational autoencoder-inspired network augmented with multi-head attention, followed by shape estimation using the original shape measurement network. Our results demonstrate that accurate shape measurement in crowded wide field radio data is feasible within a scalable end-to-end framework. While source entanglement leads to increased uncertainty relative to isolated galaxies, the recovered shapes exhibit multiplicative biases at the ten per cent level, comparable to state-of-the-art visibility-plane approaches. Crucially, the per-galaxy inference time is approximately 0.85 seconds, delivering an order-of-magnitude speed-up and establishing the practicality of the method for next-generation radio weak-lensing surveys.