Researchers at the Department of Energy’s Oak Ridge National Laboratory teach microscopes to drive discoveries with an intuitive algorithm, developed at the lab’s Center for Nanophase Materials Sciences, that could guide breakthroughs in new materials for energy technologies , detection and computing.
“There are so many potential materials, some of which we cannot study at all with conventional tools, that require more efficient and systematic approaches to design and synthesis,” said Maxim Ziatdinov of the Science and Technology Division. IT engineering from ORNL and CNMS. “We can use intelligent automation to access unexplored materials and create a shareable, repeatable path to discoveries that weren’t possible before.”
The approach, published in Intelligence of natural machinescombines physics and machine learning to automate microscopy experiments designed to study the functional properties of materials at the nanoscale.
Functional materials respond to stimuli such as heat or electricity and are designed to support everyday and emerging technologies, ranging from computers and solar cells to artificial muscles and shape-memory materials. Their unique properties relate to atomic structures and microstructures that can be observed with advanced microscopy. However, the challenge has been to develop efficient ways to locate regions of interest where these properties emerge and can be studied.
Scanning probe microscopy is an essential tool for exploring structure-property relationships in functional materials. The instruments scan the surface of materials with an atomically sharp probe to map structure at the nanometer scale – the length of one billionth of a meter. They can also detect responses to a range of stimuli, providing insight into fundamental mechanisms of polarization switching, electrochemical reactivity, plastic deformation or quantum phenomena. Today’s microscopes can perform a point-by-point scan of a square nanometer grid, but the process can be extremely slow, with measurements collected over several days for a single material.
“Physical phenomena of interest often only manifest in a small number of spatial locations and are related to specific but unknown structural elements. discovering, identifying these regions of effective interest is a major bottleneck,” said former ORNL CNMS scientist and senior author Sergei Kalinin, now at the University of Tennessee, Knoxville. is to teach microscopes to actively and much more efficiently search for regions with interesting physics than to search by grid.”
Scientists have turned to machine learning and artificial intelligence to overcome this challenge, but conventional algorithms require large human-coded datasets and may not save time in the end.
For a smarter approach to automation, the ORNL workflow integrates human physical reasoning into machine learning methods and uses very small datasets – images acquired from less than 1% of the sample – as a starting point. The algorithm selects points of interest based on what it learns during the experience and knowledge outside the experience.
As a proof of concept, a workflow was demonstrated using scanning probe microscopy and applied to well-studied ferroelectric materials. Ferroelectrics are functional materials with reorientable surface charge that can be exploited for computational, actuation, and sensing applications. Scientists want to understand the link between the amount of energy or information these materials can store and the structure of the local domain governing this property. The automated experiment discovered the specific topological defects for which these parameters are optimized.
“The takeaway is that the workflow was applied to hardware systems familiar to the research community and made a fundamental discovery, something previously unknown, very quickly – in this case, within hours,” Ziatdinov said.
The results were faster – by orders of magnitude – than conventional workflows and represent a new direction in intelligent automation.
“We wanted to move away from training computers exclusively on data from previous experiments and instead teach computers to think like researchers and learn on the fly,” Ziatdinov said. “Our approach is inspired by human intuition and recognizes that many material discoveries have been made through the trial and error of researchers who rely on their expertise and experience to guess where to look.”
Yongtao Liu of ORNL was responsible for the technical challenge of running the algorithm on an operational microscope at CNMS. “It’s not an out-of-the-box capability, and a lot of work goes into connecting hardware and software,” Liu said. “We focused on scanning probe microscopy, but the setup can be applied to other experimental approaches to imaging and spectroscopy accessible to a wider user community.”
The journal article is published under the title “Experimental discovery of structure-property relationships in ferroelectric materials via active learning”.
The work was supported by CNMS, which is a user facility of the DOE Office of Science, and the Center for 3D Ferroelectric Microelectronics, which is an Energy Frontier Research Center run by Pennsylvania State University and supported by the DOE Office of Science.