AI powers self-driving science

Researchers at the Department of Energy's Oak Ridge National Laboratory and the University of Tennessee, Knoxville, took another step toward the autonomous, or self-driving, research laboratory of the future in a study that used artificial intelligence to guide data collection in real time.

"We started the experiment and then sat back and let the AI drive," said Marshall McDonnell, an ORNL research software engineer. "We let the AI decide based on each piece of incoming data what to do next and how to get the most useful information in the most efficient way."

The research team used the Cornell High Energy Synchrotron Source (CHESS) at Cornell University in New York to map residual strain within a 3D printed metal component commonly used in turbines and aerospace machinery. The data gathered can be used to pinpoint vulnerabilities and strengthen weak spots.

Researchers typically conduct the primary data analysis in such experiments only after completion, which limits the ability to adjust data collection and to zero in on key details in real time. The team wanted to find out whether an AI model could analyze that data minute by minute and guide the beamline accordingly.

"Experiment steering represents a fundamental shift in how we use large-scale scientific facilities," said Chris Fancher, an ORNL senior scientist. "Rather than following a fixed experimental plan, researchers can dynamically guide measurements based on results as they are collected, maximizing the knowledge gained from each experiment, ultimately accelerating discovery across a wide range of scientific disciplines."

Michela Taufer, a professor of computer science at UTK, and Stephen DeWitt, an ORNL senior computational scientist, helped connect the CHESS beamline with ORNL's Interconnected Science Ecosystem (INTERSECT) and Distributed INTERSECT Active Learning (DIAL) program. The team established that connection through the National Science Data Fabric, a National Science Foundation-supported suite of software tools and data repositories that centralizes various aspects of the modern scientific workflow, enabling ORNL scientists to conduct the study remotely from Tennessee.

INTERSECT and its programs, such as DIAL, serve as a common computational infrastructure for autonomous workflows across scientific disciplines at ORNL.

The workflow set up by the research team enabled DIAL to read incoming data in real time and adjust measurements to collect those of greatest interest.

"This autonomous experiment at CHESS is the payoff of years of investment in connecting disparate facilities for autonomous science," DeWitt said. "By connecting the unique capabilities across ORNL and our partner institutions, not only do we have a more efficient ability to characterize materials, but we have a new test bed to develop the next generation of applied math and AI methods for autonomous experiments."

The results of the beamline study will be discussed in a forthcoming webinar and published later.

The study exemplifies the goals of the Genesis Mission, a national initiative by DOE's 17 national laboratories to build the world's most powerful scientific platform. The initiative seeks to accelerate discovery science, strengthen national security and drive energy innovation by connecting computing, data and experimental facilities into a unified system. Programs that support this effort include the American Science Cloud, a secure, federated and science-optimized cloud environment that integrates DOE's world-leading computing and experimental facilities, data resources and high-performance networks.

Besides DeWitt, Fancher and McDonnell, the research team included ORNL's Lance Drane, Konstantin Pieper and Viktor Reshniak; UTK's Jack Marquez and Kin Hong Ng; Valerio Pascucci and Giorgio Scorzelli at the University of Utah; Amy Gooch at ViSOAR LLC; and Amlan Das, Keara Soloway, Werner Sun and Rolf Verberg at CHESS.

Support for this research came from the DOE Office of Science Advanced Scientific Computing Research program and from the National Science Foundation.

UT-Battelle manages ORNL for DOE's Office of Science, the single largest supporter of basic research in the physical sciences in the United States. DOE's Office of Science is working to address some of the most pressing challenges of our time. For more information, visit https://energy.gov/science. - Matt Lakin