Transformer-based scalable DNN series for interferometric imaging with R2D2

Modern radio interferometers produce observations at unprecedented spatial scales, posing significant computational challenges for deep learning-based image reconstruction. The recently introduced R2D2 paradigm has demonstrated accurate high-dynamic-range radio-interferometric imaging with robust uncertainty quantification. Still, it has so far been limited to moderate 512 × 512 image sizes under telescope-specific training. In this work, we introduce a scalable R2D2 framework for radio-interferometric imaging up to 4096 × 4096 image dimensions.
To enable supervised learning at this scale, we develop a scalable synthetic radio-interferometric image generation pipeline that composes realistic source morphologies into statistically diverse scenes across multiple image dimensions through linear source composition, geometric and morphological transformations, and background modelling. These images are subsequently used to simulate telescope-specific radio interferometric observations under diverse observational configurations for R2D2 training. Building upon this dataset, we develop a hierarchical hybrid architecture that combines a Swin Transformer encoder with the U-WDSR reconstruction network. The Swin Transformer progressively compresses large images into hierarchical multi-scale feature representations that capture long-range spatial dependencies, while U-WDSR exploits these representations to recover fine-scale image structures within the iterative R2D2 reconstruction process. This hierarchical design substantially reduces the computational burden of large-scale imaging while maintaining high reconstruction quality. Extensive experiments demonstrate the scalability, computational efficiency, and generalisation capability of the proposed framework, providing a practical step toward learned image reconstruction for future SKA-scale radio-interferometric imaging.