University of Illinois researchers are part of a $15M institute developing real-time artificial intelligence to accelerate discovery in data-driven science
Today, the National Science Foundation (NSF) announced its launch of the $15M Accelerated AI Algorithms for Data-Driven Discovery (A3D3) Institute, as part of its $75M investment in five new Harnessing the Data Revolution Institutes across the U.S. Researchers at the new Institutes will tackle some of society’s most pressing fundamental questions at the frontiers of science and engineering. The primary mission of the A3D3 Institute is to lead a paradigm shift in the application of real-time artificial intelligence at scale to advance scientific knowledge and accelerate discovery.
A3D3 is a multi-disciplinary institute led by principal investigators (PIs) at the California Institute of Technology, Duke University, Massachusetts Institute of Technology, Purdue University, University of California, San Diego, University of Illinois Urbana-Champaign, University of Minnesota, University of Washington and University of Wisconsin-Madison. The A3D3 Institute brings together multidisciplinary teams of researchers, engineers, and educators, including Deming Chen, Abel Bliss Professor of Engineering in Electrical and Computer Engineering and Coordinated Science Laboratory and director of the Xilinx Center of Excellence and Mark Neubauer, a professor in the Department of Physics, affiliate professor in the Department of Electrical and Computer Engineering and faculty affiliate at the National Center for Supercomputing Applications (NCSA). Neubauer is a founding member of the NCSA Center for AI Innovation and core faculty member in the Illinois Center for Advanced Studies of the Universe (ICASU).
Big-Data Science Meets the AI Revolution
The demand for computing resources in the next generation of data-intensive scientific experiments will outstrip the capabilities of existing infrastructure. Achieving the science goals of these experiments and creating new opportunities for data-driven discovery will require a radical rethinking of science-driven methodologies to achieve fast signal processing and data analysis, coupled to existing and planned cyberinfrastructure.
According to Neubauer, who is co-PI of the A3D3 Institute, artificial intelligence (AI) has emerged as a solution for rapidly processing large amounts of complex scientific data, enabling new insights into nature at scales ranging from fundamental particles, to the human brain, to supernova and supermassive black holes.
“Clever combination of innovative AI algorithms and new hardware architectures for accelerated computation is leading to a revolution in the way we analyze data,” says Neubauer. “Together they provide highly-performant, configurable data analysis capabilities that allow fundamentally new scientific questions to be posed and answered. Tightly coupling the design of AI algorithms with new processor technologies such as in real-time systems is the key idea behind the A3D3 institute. We believe that A3D3 will be the nexus for real-time AI applied to frontier science.”
To address the new challenges raised by the AI algorithms and hardware co-design, the institute plans to leverage new high-level synthesis tools being developed in Chen’s research group. High-level synthesis is a novel compiler technique that enables a greater design abstraction, so domain scientists or engineers can use high-level languages (such as C++ or Python) to program customizable accelerators, including FPGAs, for AI algorithm acceleration.
“High-level synthesis can significantly improve design productivity compared to the conventional design process where the customized accelerators had to be designed using hardware description languages,” says Chen. “Our current effort to accelerate large and complex AI algorithms through high-level synthesis has opened up new opportunities for novel hardware design and programming, which fit into the institute’s overall mission very well.”
To take full advantage of fast AI, A3D3 will target fundamental problems in three fields of science: high energy physics, multi-messenger astrophysics, and systems neuroscience. A3D3 will work closely within these domains to develop customized AI solutions to process large datasets in real time, significantly enhancing their discovery potential. Through dedicated outreach efforts to big-data experiments, A3D3 will empower scientists with new tools to deal with the coming scientific data deluge. A key goal of A3D3 is to develop the institutional knowledge that is essential for real-time applications of AI in any scientific field.
Developing new AI Capabilities in High Energy Physics
Neubauer is a member of the ATLAS group, one of two experiments at the Large Hadron Collider (LHC) at CERN in Switzerland that discovered the Higgs boson, leading to the 2013 Nobel Prize in Physics. The experiments continue to search for new physics using LHC data, which is increasing in volume and complexity.
“The pursuit of new physics at the LHC will benefit greatly from synergies with the A3D3 Institute’s research,” says Neubauer. “At the LHC, the challenge of processing data is daunting. With future aggregate data rates exceeding 1 petabit per second, the data rates at the LHC exceed those of all other scientific devices in the world. The aim of A3D3 is to build a series of tools that will enable the real-time processing of this data deluge using AI. A3D3 aims to enable advanced analyses, such as anomaly detection, and particle reconstruction on all collisions – occurring at a rate of 40 million times per second!”
Answering the Call for New AI tools in Multi-Messenger Astrophysics
Within the field of multi-messenger astrophysics, A3D3 is working to integrate AI to promptly and efficiently process data collected from telescopes, neutrino detectors, and gravitational- wave detectors, to quickly identify astronomical events from the most violent phenomena in the cosmos. Rapid identification of transient astrophysical phenomena is a critical element of the broader multi-messenger astrophysics effort, which aims to achieve a coordinated observation and interpretation of these phenomena using multiple types of cosmic "messenger" signals (electromagnetic radiation, gravitational waves, neutrinos and cosmic rays) to develop a more complete understanding of the physical properties and processes of the universe.
Collaborative research efforts between A3D3 and ICSAU members, as well as joint activities such as workshops bringing together faculty from physics, astronomy, astrophysics, computer science, and mathematics, will further develop and strengthen interdisciplinary activities at the University of Illinois and across the broader scientific community.
“I am very excited about new collaborations between the A3D3 institute and ICASU in multi- messenger astrophysics,” says Illinois Physics Professor and ICASU Director Nicolás Yunes. “ICASU’s mission is precisely to establish new collaborations of this type and to promote interdisciplinary work to address big questions about our universe. Combining the expertise and resources of these two centers opens the window to potentially revolutionary discoveries in astrophysics.”
Applying AI Data Solutions in Neuroscience
In systems neuroscience, A3D3 is working to discover the computations that brain-wide neural networks perform to process sensory and motor information during behavior. In this pursuit, A3D3 will develop and implement high-throughput and low-latency AI algorithms to process, organize, and analyze massive neural datasets in real time. These real-time analyses will enable new approaches to probing brain function such as causal, closed-loop manipulations. Applying powerful AI methods to systems neuroscience will significantly advance our ability to analyze and interpret neural activity and its relationship to behavior. In this effort, A3D3 will collaborate with the Illinois-IBM Discovery Accelerator Institute (IIDAI).
Chen, who is the Hybrid Cloud Theme co-lead within IIDAI, notes, “The leading research theme of IIDAI is to develop and prototype a secure, smart, and highly performant hybrid cloud system that would be able not only to accelerate sophisticated and large-scale workloads efficiently, such as brain-wide neural networks, but also to preserve privacy and maintain integrity of all personal or other sensitive data in the system. I see a lot of synergies between the two institutes across various scientific domains and cloud computing and look forward to exploring opportunities for close collaboration in the future.”
This project is supported by the National Science Foundation Grant no. 2117997.