Scalable Precision Imaging in Radio Astronomy: from Learned denoisers on GPU to Science (SPIRALS)

射电天文学未来几十年雄心勃勃的科学目标引发了新一代望远镜的发展，目标是以比目前仪器高得多的精度(即分辨率和灵敏度)成像天空。要使这些望远镜具有预期的锐利视力，需要成像算法能够将无线电干涉测量数据转换为目标精度的图像，同时具有健壮性(即包括校准和不确定性量化功能)，并最终可扩展到亿级数据量。“射电天文学中可扩展的精确成像：从图形处理器上的可学习去噪器到科学”(SPRALS)工作方案旨在设计变革性的深度学习方法来应对这一挑战，并将其应用于尖端科学案例，从探测星系团中的光晕和遗迹，到详细绘制射电星系的形态图，从Meerkat望远镜的观测，到使用事件视界望远镜(EHT)数据的黑洞成像。这些去噪器将被集成到一个并行算法结构中，以定义一种新的图像形成算法，该算法同时具有精度、鲁棒性和可伸缩性的能力。将为最新和最高效的高性能计算硬件技术，主要是大规模的GPU系统，设计并映射一个并行的Python软件实现。其次，作为学习去噪器的一个副产品，将开发一种低成本的“后处理器”来增强遗留图像。在处理科学案例之前，将利用未来平方公里阵列望远镜和深天气阵列望远镜概念的模拟以及Jansky甚大阵列和Meerkat望远镜的实际数据，特别侧重于在整个视场具有复杂和微弱结构的漫反射辐射的宽带成像，对算法和软件进行验证，最大可达兆兆字节。

The ambitious science goals in radio astronomy for the next decades have triggered the development of a new generation of telescopes targeting imaging the sky with much higher precision (i.e. resolution and sensitivity) than current instruments. Endowing these telescopes with their expected acute vision requires image formation algorithms capable of transforming radio interferometry data into images at target precision, while being robust (i.e. including calibration and uncertainty quantification functionalities), and ultimately scalable to exascale data volumes. The "Scalable Precision Imaging in Radio Astronomy: from Learned denoisers on GPU to Science" (SPIRALS) work programme aims to design transformative deep learning methodology to address this challenge, and apply it on cutting-edge science cases, from the detection of halos and relics in galaxy clusters and detailed morphology mapping of radio galaxies from surveys of the MeerKAT telescope, to black hole imaging with Event Horizon Telescope (EHT) data.In detail, firstly, artificial neural networks will be trained as simple "denoisers" encapsulating advanced learned physical models of the both radio sky and interfering observation effects. These denoisers will be integrated into a parallel algorithmic structure to define a new image formation algorithm with simultaneous capability for precision, robustness, and scalability. A parallel Python software implementation will be designed for, and mapped onto the latest and most efficient high performance computing hardware technologies, primarily large scale GPU systems. Secondly, as a by-product of the learning of denoisers, a low-cost "post-processor" will be developed to enhance legacy images. Before addressing the science cases, algorithms and software will be validated up to Terabyte image size, using both simulations from the future Square Kilometre Array (SKA) telescope and from the Deep Synoptic Array (DSA-2000) telescope concept, and real data from the Jansky Very Large Array (JVLA) and MeerKAT telescopes, with particular focus on wideband imaging of diffuse emission with complex and faint structure across the field of view.