A new AI algorithm called R2D2 is learning to clean up blurry radio telescope images faster and more accurately than any existing method. The problem is that next-generation radio telescopes—like the Square Kilometre Array (SKA) and the next-generation Very Large Array (ngVLA)—produce enormous, messy datasets. The classic CLEAN algorithm, which has been the workhorse of radio astronomy for decades, cannot keep up. It fails to deliver the combination of precision, speed, scalability, and robustness needed for modern four-dimensional imaging (two spatial dimensions, plus frequency and polarisation). Without a better computational imaging method, these billion-euro telescopes will not reach their full visual acuity. If R2D2 succeeds, it will transform how astronomers study cosmic magnetism, dark matter, dark energy, and the evolution of stars and galaxies. The project will demonstrate the algorithm on real data from four major observatories, including real-time all-sky imaging with LOFAR and ultra-fast sub-array imaging with SKA. This is fundamental science with no immediate practical application—but past advances in computational imaging have unexpectedly improved medical scanners, satellite imaging, and autonomous navigation.
View original technical description
Synthesis imaging by radio interferometry (RI) is a powerful technique in astronomy that leverages antenna arrays to observe the sky with high resolution and sensitivity. A new generation of RI telescopes is currently emerging. Their science goals range from studying cosmic magnetism, dark matter, and dark energy, to understanding the structure and evolution of stars and galaxies. In this new era, the mathematical inverse problem for the formation of images from observed data is extremely challenging, with bespoke computational imaging (CI) algorithms required to deliver a new regime of joint precision, efficiency, scalability, and robustness in image formation. We refer to this four-fold target capability as RI4. Current CI methodology for RI, from the famous CLEAN algorithm that has served the field for decades, to recent algorithms leveraging the latest developments in optimisation and deep learning theory, fail to satisfy the RI4 demand, calling for a new paradigm. In other words, in a context where major investments are being made to develop RI telescopes, a commensurate effort is warranted towards a CI paradigm that endows such instruments with an optimal visual acuity. A new CI algorithm for RI image formation was recently proposed by the project leads, dubbed R2D2, which can be understood as an AI version of CLEAN. R2D2 forms an image from observed data as a series of residual images, iteratively estimated as outputs of Deep Neural Networks (DNNs) taking the previous iteration’s estimate and associated data residual as inputs. Our vision and research hypothesis is that R2D2 seeds a transformative paradigm towards the target RI4 capability. Building on early R2D2 developments for 2-dimensional monochromatic intensity imaging, the project’s aim is to deliver RI4 for full 4-dimensional wideband polarisation all-sky imaging. The research is organised in two work packages, each with specific objectives: work package 1 (Methodology & Software) will develop R2D2’s full 4-dimensional functionality, implement a parallel resource-adaptive Python code, and validate RI4 capability in simulation; work package 2 (Case Studies) will demonstrate R2D2 for flagship case studies, in both simulated and real data settings, namely ultra-fast sub-array imaging with the Square Kilometre Array (SKA) and the next-generation Very Large Array (ngVLA), on-the-fly calibration-free imaging with the Deep Synoptic Array (DSA-2000), and real-time all-sky imaging with the LOw Frequency ARray (LOFAR). The project is defined in partnership with the radio astronomy institutes and observatories operating these instruments.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know