Second Target Station workshop explores AI-aided future for neutron experiments

Plenary speaker Tony Hey, chief data scientist at Rutherford Appleton Laboratory in England, discusses artificial intelligence applications for large-scale facilities at the STS workshop. Credit: ORNL/Genevieve Martin

Plenary speaker Tony Hey, chief data scientist at Rutherford Appleton Laboratory in England, discusses artificial intelligence applications for large-scale facilities at the STS workshop. Credit: ORNL/Genevieve Martin

Currently the world’s most powerful, pulsed accelerator-based neutron source, the Spallation Neutron Source (SNS) at the US Department of Energy’s Oak Ridge National Laboratory (ORNL) will be upgraded and expanded in the next decade to increase beam brightness by 25 times. This will offer richer research capabilities and enable breakthrough discoveries in many areas of materials research and development.

Specifically, the Second Target Station (STS) will deliver a new target for the production of neutrons and a suite of instruments up to 1,000-times higher performing than existing instruments.

Scientists of many disciplines convened at ORNL’s “Science at the Second Target Station” workshop on Dec. 9-10 to discuss the research opportunities that will be enabled by the STS. No matter the scientific domain, one clear benefit of the STS will be the influx of data provided by the brighter beam (more neutrons per pulse), greater range of wavelengths, and advancements in instrumentation.

Scientists who conduct research at SNS are grappling with how they will effectively analyze all this new information-and they are not alone.

“Neutron sources across the world are highly oversubscribed,” said Hans Christen, director of the ORNL Neutron Scattering Division. “We can’t afford inefficiencies in our experimental approaches or missing information that may be hidden inside our data sets. I am convinced that modern data analytics, including artificial intelligence (AI), can help us get more out of our neutrons.”

Research facilities are exploring reliable ways to automate and enhance data collection and analysis through AI and machine learning as data gushes from sophisticated instruments and computing hardware permits unprecedented amounts of storage.

“Instead of going home with a USB stick’s worth of data, researchers have to analyze terabytes of data,” said Tony Hey, chief data scientist at the Science and Technology Facilities Council’s Rutherford Appleton Laboratory in England, describing synchrotron experiments. A terabyte is about the storage capacity of most modern laptops.

In his plenary session “AI for Science at Large-Scale Experimental Facilities,” Hey described that hardware advances like accelerated computing processors known as GPUs and developments in machine learning algorithms like neural networks have spurred recent applications of AI for scientific research.

Supercomputer help for STS data

Considering ORNL is home to not only world-class neutron science facilities but the world’s most powerful supercomputer for scientific research-Summit, a 200-petaflop, hybrid CPU-GPU machine designed with AI in mind-and future home of the Frontier exascale supercomputer, ORNL is a unique place to leverage computing resources for collecting and analyzing experimental data.

“It’s a very exciting time for Oak Ridge and the national labs because the Second Target Station is going to produce experiments from a range of sciences with more sophisticated data and more of it because of its beamline intensity,” Hey said. “And you have supercomputers and, next, an exascale computer.”

More than 300 ORNL researchers use AI to expedite scientific breakthroughs.

“AI will become increasingly important in both the analysis of data and to guide experiments,” Christen said. “Because of ORNL’s strengths and expertise in materials science, neutron scattering, and computing, this is an almost ideal setting to accelerate such interdisciplinary approaches.”

Hey relayed his own efforts at Rutherford Appleton Laboratory to integrate experimental neutron, light source, and microscopy facilities with computing and data clusters.

“We’re trying to put these places together to do better, more efficient science, which is what you’re trying to do here,” he said. “Oak Ridge is an example of what all you can do with computing resources.”

The STS instruments and the data they produce will be critical for maintaining US leadership in neutron science and advancing scientific inquiry. Nondestructive and neutrally charged, neutrons are particularly useful for probing biological, magnetic, dynamic, and other complex materials, including those used in new computing devices, engines, batteries, and industrial and consumer products.

“Scientists need the unique capability of the Second Target Station for certain types of research. The STS will open many new directions in a diversity of scientific fields,” said Yang Zhang (YZ), associate professor at the University of Illinois Urbana-Champaign (UIUC) and a former ORNL Clifford G. Shull Fellow in the ORNL Neutron Sciences Directorate. “With this workshop, we’re asking, ‘What are the game-changing scientific problems we can solve with STS that couldn’t be addressed before?'”

How can AI be used at STS?

YZ, who combines computational modeling and neutron-based experiments in the study of “molecules, materials, and machines” in his Z Lab at UIUC, led a town hall discussion on how AI and data science methods could be utilized at the STS.

“Roughly speaking, there are two kinds of applications of AI to neutron science,” YZ said. “First, how can we use AI to help optimize data collection or design new instruments that are deeply integrated with AI? Maybe AI can help with data collection by doing things like experimental planning, measurement optimization, real-time anomaly detection, feature enhancement, etc. …

“Second, how can AI help us with data analysis and modeling? This is even harder and requires organic integrations with specific domain knowledge.”

Town hall panelists included Hey and ORNL researchers Changwoo Do, an instrument scientist specializing in small-angle neutron scattering (SANS) and spin echo-which are techniques expected to provide new insights into biological membranes and thin films at the STS; Timmy Ramirez-Cuesta, neutron scattering team leader for chemical spectroscopy; Alan Tennant, Physical Sciences Directorate initiative leader for quantum materials; and Mathieu Doucet, a computer scientist who streamlines the SNS user experience as part of the Scientific Computing and Software Engineering Group.

Panelists shared their thoughts on the potential for AI to improve scientific outcomes at the STS-including accelerating the experiment and data collection process, refining experimental strategies, and revealing new scientific insights through enhanced data collection and analysis-as well as the possible challenges.

Do said he became interested in neutron scattering applications of AI because of the time sensitivity in SANS experiments. “SANS is about how fast we can measure,” he said. “The challenge is knowing what [AI] algorithms will work. We have been collecting scattering images and now have over 100,000 data files, so maybe we can find a way to utilize that data to apply machine learning and AI to accelerate data acquisitions.”

Tennant credited the use of neutrons combined with machine learning and high-performance computing for new insights into the physics of glass formation. “You can use [machine learning] to model experimental strategies and look at multimodal data together, which is very difficult to do otherwise,” he said.

Panelists also addressed concerns about over-automation and a dependence on AI for making scientific discoveries.

“AI is meant to be supportive of experiment, rather than replacing it,” said Hey, who leads a scientific machine learning group at his home institution that curates experimental facility data to understand how AI and machine learning can be validated on existing datasets.

“Just as computer simulation has been developed in the past, we can do the same thing with AI,” Do said. “We can develop confidence in the data. In the meantime, we have to develop close collaboration with computer scientists. We need to combine our expertise.”

Through collaborations like the “Science at the Second Target Station” workshop, input from scientists-at ORNL and beyond-who are navigating research in the era of data-driven science and AI will be critical to next-generation facilities like the STS.

As Hey said when asked during the town hall to summarize his thoughts on AI for science in a single word: “Potential. I think there’s great potential for advances in science…so my word would be potential.”

/Public Release. View in full here.