Abstract

Radio interferometric imaging underpins much of modern radio astronomy, yet its computational cost increasingly limits scientific return for current facilities and poses a major challenge for the Square Kilometre Array (SKA), which will be the largest radio interferometric array ever constructed. A dominant bottleneck lies in the gridding and degridding operations that enable FFT -based imaging from irregularly sampled visibilities, particularly for wide-field, high-resolution observations.

In this talk, I present my contributions to improving the accuracy and efficiency of interferometric imaging, spanning both optimal gridding methods and wide-field imaging algorithms. My work on gridding focuses on directly minimising image-domain error rather than relying on analytic kernel assumptions, while my wide-field imaging developments reduce the computational cost of imaging large sky areas at high resolution. These methods have been adopted in widely used imaging software and have enabled high-fidelity, low-frequency, wide-field images previously considered impractical, including benchmark LOFAR results at arcsecond resolution.

Building on this foundation, I discuss ongoing and future work aimed at fundamentally reducing imaging cost through AI-driven approaches. By combining numerical optimisation, compact analytic representations, and machine-learning–based amortisation, this research seeks to deliver imaging methods that are both scientifically optimal and computationally scalable. More ambitiously, neural operators offer the prospect of bypassing gridding entirely, learning the visibility-to-image mapping directly from irregularly-sampled data without intermediate gridding, FFT , or correction steps.