Extreme-scale precision imaging in radio astronomy

基于孔径合成技术的下一代射电干涉仪由于具有上级角分辨率和灵敏度，有望取得突破性进展，为天文学的基本问题带来答案。然而，干涉仪的感测策略仅提供原始天空的不完整线性信息。从模糊和噪声的接收数据中恢复干净的图像形成了复杂的不适定逆成像问题。此外，新射电望远镜的巨大规模，如平方公里阵列（SKA），也带来了与目标规模相称的巨大数据流，前所未有的精度和灵敏度。SKA产生的宽带图像立方体将达到1 PB的大小。为了满足如此强大的设备的能力，图像处理管道中的每个部分都需要调整。目前，最优化被认为是设计天文成像反卷积算法的最有前途的框架之一。目标函数可以看作是保真度项和正则化项的总和。由于对正则化项的分面处理的相关研究，可以首先将大数据量划分为小的和重叠的块。然后并行处理成为可能，并且可扩展性大大提高。另一个值得注意的趋势是将深度学习框架引入天文成像，以利用神经网络的可扩展性和并行性。一个令人兴奋的进展是使用神经网络来近似正则化算子并集成到图像恢复过程中。这个博士项目将从扩展当前的优化天文成像算法开始，并利用神经网络的力量来提高重建图像的分辨率和动态范围。然后，并行性和可扩展性将被调查和优化，旨在扩大这些算法，以满足下一代无线电干涉仪的要求。在此基础上，将考虑天文成像中的定标问题和不确定度评定问题。该项目的最终目标是实施这些优化的算法并将其部署在生产HPC系统上，以微调实施和算法，并使其具有针对真实的应用程序的性能和可扩展性。实现这一点的一个关键特性是将异构计算引入到实现中。除了CPU节点，还有一系列计算硬件，如GPU、FPGA、DSP等，其可以集成到该过程中以进一步提高处理速度。此外，射电天文成像算法也适用于解决医学成像问题。因此，用于评估和优化我们的算法和实现的数据集将在天文成像中测试结果后扩展到医学成像。最后，如果可能的话，我将把在极端规模天文成像和HPC系统方面的经验进一步应用到其他领域，例如设计用于天文成像的特定处理单元和优化HPC机器上的大型销售渲染问题。

Because of superior angular resolutions and sensitivities, next-generation radio interferometers based on aperture synthesis are expected to make breakthroughs and bring answers to essential questions in astronomy. However, the sensing strategy of interferometers only provides incomplete linear information of the original sky. Recovering the clean image from the blur and noisy received data forms a complicated ill-posed inverse imaging problem. Additionally, the gigantic scales of the new radio telescopes, such as Square Kilometres Array (SKA), also bring enormous data flows, commensurate to the target scale, unprecedented precision and sensitivity. The wide-band image cubes generated by SKA will reach the size of 1 Petabyte. To meet the capabilities of such powerful equipment, every section in the image processing pipeline needs to be tuned. Nowadays, optimisation is suggested to be one of the promising frameworks in designing deconvolution algorithms for astronomical imaging. And the objective function can be seen as the sum combination of fidelity term and regularisation term. Thanks to relevant research into faceted processing of regularisation terms, the large data volumes can be divided into small and overlapped blocks first. Then parallel processing becomes possible, and the scalability is greatly improved. Another noticeable trend is introducing deep learning frameworks into astronomical imaging to take the advantages of scalability and parallelism in neural networks. An exciting progress is using neural networks to approximate regularisation operators and integrate into image recovery process. This PhD project will start by extending current optimisation astronomical imaging algorithms and leverage the power of neural networks to improve the resolution and dynamic range of the reconstructed images. Then, the parallelism and scalability will be investigated and optimised, aiming to scale these algorithms up to meet the requirements of next-generation radio interferometers. Feeding into this, the calibration problems and uncertainty qualification problems in astronomical imaging will be considered. The ultimate goal of this project is implementing those optimised algorithms and deploying them on production HPC systems to fine tune the implementations and algorithms and making them performant and scalable for real world applications. One key feature in enabling this is the introduction of heterogeneous computing into the implementations. Besides CPU nodes, there is a range of computing hardware, such as GPUs, FPGAs, DSPs, etc., that can be integrated into this process to further improve processing speeds. Additionally, the algorithms for radio astronomical imaging are also applicable in solving medical imaging problems. As such, the data sets used to evaluate and optimise our algorithms and implementations will be extended to medical imaging after the results have been tested in astronomical imaging. Lastly, I will further apply the experience in extreme scale astronomical imaging and HPC system to other fields if possible, such as designing specific processing units for astronomical imaging and optimising large sale rendering problems on HPC machines.