In a demonstration led by researchers at the Cornell High Energy Synchrotron Source (CHESS), in collaboration with Oak Ridge National Laboratory (ORNL), the University of Tennessee and the University of Utah, data collected at a CHESS beamline were analyzed in real time at ORNL (800 miles away in Tennessee) and new measurement instructions were automatically sent back to the instrument on the Cornell campus. The process repeated continuously until the experiment was complete, marking a significant step toward synchrotron experiments that adapt as they run, make more efficient use of valuable beamtime and accelerate scientific discovery.
The demonstration took place at CHESS's Structural Materials Beamline (SMB), where scientists use high-energy X-rays to probe the internal stresses of engineering materials. Understanding those stresses is critical for developing stronger, more reliable components for manufacturing, transportation, energy systems and other industries. “At these instruments, time is a scarce resource,” said Amlan Das, staff scientist at the CHESS beamline where the experiment took place. “Researchers often have only a limited amount of time to collect the data they need. The question we're asking is: How can we make the best possible use of every x-ray photon?”
Image
In a typical structural materials experiment, researchers select hundreds or even thousands of measurement locations across a sample before data collection begins. Because they can't know in advance where the important features will appear, they often collect far more data than ultimately proves necessary. This is a reliable approach but consumes significantly more beamtime than an experiment that could adapt on the fly.
The challenge extends well beyond structural materials. Researchers at CHESS and other synchrotron facilities have spent years developing autonomous measurement strategies in which an AI-based decision-maker - an “autonomous agent” - evaluates incoming data to determine where the next measurement will be most informative, concentrating effort where the material is changing and skipping regions that offer little new information. CHESS researchers have previously used the approach to accelerate the discovery of new organic molecules and functional thin films. Adapting it to structural materials could shrink some experiments from several days to a single day while preserving the quality of the resulting data.
Earlier this spring, CHESS researchers demonstrated autonomously driven measurements for residual stress mapping for the first time. Those measurements were led by the Materials Solutions Network at CHESS (MSN-C), a program sponsored by the Air Force Research Laboratory (AFRL) and drew on years of work streamlining how these complex measurements are collected and analyzed.
“The long-term vision of the MSN-C program includes a strategic focus on measurement efficiency through automation,” said Paul Shade, Senior Research Materials Engineer and Principal Investigator at AFRL, and program manager for the MSN-C program. “These recent efforts at developing autonomous workflows are well aligned with that vision and offer a promising glimpse into the future”
This approach is powerful because not every location within a material is equally informative. Researchers compare the challenge to exploring unfamiliar terrain: Some regions reveal little, while others contain important structural changes that deserve closer examination. By identifying areas of uncertainty as data are collected, the autonomous workflow focuses measurements where they are most likely to improve scientific understanding. “The experiment becomes adaptive,” Das said. “Instead of following a predetermined path, the system can respond to what it's observing.”
The new collaboration extended that success further, allowing the autonomous agent to not be in the same location as the instrument. During a recent demonstration a stainless-steel sample was measured at SMB while data were simultaneously transmitted to ORNL's INTERSECT autonomous workflow platform through the National Science Data Fabric (NSDF), a collaborative cyberinfrastructure project designed to connect scientific facilities, computing resources and researchers across institutions.
“This connection now means that the autonomous agent can make decisions based on information from other facilities and modalities. Just like human experimenters, we want the agent to be as well informed as possible.”, said Das.
As new measurements arrived, an active-learning system analyzed the incoming data and identified the locations most likely to yield new information about the sample. Those recommendations were routed back to CHESS, where the beamline adjusted its measurements accordingly, a continuous feedback loop connecting the instrument, analysis software and researchers in real time.
Image
"CHESS has invested heavily in AI readiness, building the cyberinfrastructure that connects our instruments and data pipelines with external computing resources," said Werner Sun, CHESS IT Director. "This collaboration with NSDF and ORNL demonstrates how those capabilities let us incorporate advances from our partners into experiments that are not only autonomous but are also scalable and reproducible in the long run."
For CHESS, the demonstration builds on a broader effort to make experiments faster, more flexible and more accessible to users. Over the past several years, researchers have steadily expanded automation at the Structural Materials Beamline, streamlining everything from sample alignment to data collection and analysis. Arthur Woll, director of MSN-C, said the demonstration drew on several capabilities developed at CHESS, including the beamline's state-of-the-art energy-dispersive diffraction (EDD) detector system; automated data-analysis pipelines that process measurements in real time; laser metrology tools that precisely map a sample's position within the experimental hutch; and beamline control software that coordinates instruments and workflows.
“Autonomous experiments depend on a lot of pieces working together,” Woll said. “You need the instrumentation, the data processing, the controls and the communication between all of those systems. What's exciting is seeing those pieces come together in a way that allows the experiment to respond in real time.”
The implications reach far beyond a single beamline. Researchers envision experiments that examine more samples, explore more complex materials and collect higher-value data within the same beamtime. Their workflows coordinate sophisticated experiments across multiple institutions, combining expertise in instrumentation, computing, data science and materials research.
“None of this happens in isolation,” Das said. “This project brought together experts from several organizations, each contributing a different piece of the puzzle. That's what made this demonstration possible.”
This collaboration was made possible through many efforts, including those of: Chris Budrow & Diwakar Naragani (MSN-C), Marshall McDonnell, Chris Fancher, Stephen DeWitt, Lance Drane, Konstantin Pieper, Viktor Reshniak (ORNL), Michela Taufer, Jack Marquez, Kin Hong Ng (UTK), Amy Gooch (VisOAR), Valerio Pascucci, Giorgio Scorzelli (UUtah).
Additional CHESS contributions came from Rolf Verberg and Keara Soloway, whose beamline integrations connected the data acquisition system, data analysis software and strain-map outputs to the active-learning loop.
CHESS is supported by the National Science Foundation (NSF), the National Institutes of Health and New York state; MSN-C is supported by the Air Force Research Laboratory. The National Science Data Fabric is supported by the NSF, and ORNL's INTERSECT initiative is supported by the U.S. Department of Energy.