MRI: Development of Heterogeneous Edge Computing Platform for Real-Time Scientific Machine Learning

该项目旨在开发一个异构边缘计算平台，用于极端边缘的实时科学机器学习（HEC-SML）。这样的平台将允许实时分析和控制光学，扫描探针和透射电子显微镜。将计算放在边缘——尽可能靠近数据源——在将数据发送到高性能计算设施时，可以避免延迟和带宽挑战，并将实现实时数据分析，从而通过创造性的探究进行科学实验。HEC-SML的发展将导致显微镜、机器学习（ML）和异构计算概念的融合。在显微镜，先进的控制系统将使新的成像模式；个性化医疗方法；以及对功能和量子材料的发现和理解。在机器学习中，HEC-SML将激励策略的设计来施加物理约束，并开发优化方法来训练更有效的算法。这些不同领域的联合研究创造了新的目标，推动了每个学科的变革进步。HEC-SML支持的研究将促进并培训能够解决无线通信、医疗保健监控、先进制造和多代理自治系统的边缘计算挑战的科学家。该工具还将使几个学生的课程沿着这一领域的融合发展。该项目将吸引生物科学、材料科学、机器学习和异构计算领域的跨学科研究人员，并带来新的研究热点。高性能计算（HPC）在计划计算、并行计算和分布式计算方面取得了巨大的进步。实验显微术要求以与实验相关的时间尺度（纳秒-分钟）处理高速的大量数据。HPC设施不能满足这些需求，因为它们不是为专用网络和计算而设计的，而且通常没有用于低延迟计算的异构计算节点。HEC-SML将是一种用于显微镜实时分析和控制的专用仪器。一项技术创新是将HPC与显微镜放在一起，以实现低成本、可重构、专用的高速网络。有了这个，我们将开发一个集中的边缘计算平台，用于信号处理，数据减少和ML工具，用于许多显微镜模式。该仪器将为光学、扫描探针和透射电子显微镜的实时分析和控制提供一个平台。HEC-SML将为细胞计数和分选以及能量转换、传感和量子材料的表征和操作提供新的能力。HEC-SML将为显微镜和其他数据密集型科学实验提供交钥匙解决方案。此外，HEC-SML将推动个性化医疗的变革性进步；传感、能量转换、量子材料；使用实验数据的科学和物理知识ML方法；共同设计异构计算和ML，用于低延迟数据减少、科学信号处理和控制系统。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project aims to develop a Heterogeneous Edge Computing platform for real-time Scientific Machine Learning (HEC-SML) at the extreme edge. Such a platform will allow for real-time analysis and control of optical, scanning probe, and transmission electron microscopy. Putting computation at the edge –- as close to the data source as possible -– circumvents latency and bandwidth challenges when sending data to high-performance computing facilities and will enable real-time data analysis to conduct scientific experiments with creative inquiry. The development of HEC-SML will lead to the convergence of microscopy, machine learning (ML), and heterogeneous computing concepts. In microscopy, advanced control systems will enable new imaging modalities; methods for personalized medicine; and discovery and understanding of functional and quantum materials. In ML, HEC-SML will motivate the design of strategies to impose physics constraints and develop optimization methods for training more efficient algorithms. Combined research in these disparate fields creates new objectives that motivate transformative advances in each discipline. The research enabled by HEC-SML will enable advances to and train scientists that can address edge computing challenges for wireless communication, healthcare monitoring, advanced manufacturing, and multi-agent autonomous systems. The instrument will also enable several student curriculum advances along this convergence of fields. This program will engage interdisciplinary researchers in biological sciences, materials science, machine learning, and heterogeneous computing and lead to novel research thrusts. High-performance computing (HPC) has made tremendous advances in scheduled, parallelized, and distributed computing. Experimental microscopy requires that voluminous data at high velocity is processed in timescales relevant to the experiment (nanoseconds-minutes). HPC facilities cannot meet these needs as they are not designed for dedicated networking and computation and typically do not have heterogeneous compute nodes for low-latency computation. HEC-SML will be a purpose-built instrument for real-time analysis and control of microscopy. A technical innovation is co-locating HPC with microscopy to enable low-cost, reconfigurable, dedicated high-speed networking. With this, we will develop a centralized edge computing platform for signal processing, data reduction, and ML tools for many microscopy modalities. The instrument will provide a platform for real-time analysis and control of optical, scanning probe, and transmission electron microscopy. HEC-SML will provide new capabilities for counting and sorting cells as well as for the characterization and manipulation of materials for energy conversion, sensing, and quantum materials. HEC-SML will provide a turn-key solution for microscopy and other data-intensive scientific experiments. Furthermore, HEC-SML will enable transformative advances in personalized medicine; sensing, energy conversion, and quantum materials; scientific and physics-informed ML methods using experimental data; and codesign of heterogeneous computing and ML for low-latency data reduction, scientific signal processing, and control systems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.