Functional materials hold unique properties tied to their atomic structures and microstructures, which support modern and emerging technologies ranging from computers to solar cells to artificial muscles. However, designing efficient methods for locating the regions of interest where these unique properties emerge has been challenging, with many experiments relying on lengthy point-by-point grid searches or trial and error. Researchers at Oak Ridge National Laboratory (ORNL) have now developed a “smarter” way to investigate the properties of functional materials, enabling self-driving microscopes to quickly pinpoint regions of interest using advanced active learning algorithms.
Conventional machine learning algorithms often require large, human-coded datasets that add time to “teach” the machine prior to the experiment. The ORNL approach instead allows the microscope to “learn on the fly” during the experiment, similar to how a human research would use their own intuition and reasoning to guide their actions, according to Maxim Ziatdinov, of ORNL’s Computational Sciences and Engineering Division and Center for Nanophase Materials Sciences. This human-based physical reasoning workflow relies only on very small datasets – images acquired from less than 1% of the sample – as a starting point to select points of interest. The method saves time by avoiding a painstaking grid search, which could take days to complete for a single material, and by allowing the algorithm to continuously learn as it goes.
The researchers demonstrated their new machine learning workflow using scanning probe microscopy to examine already well-studied ferroelectric materials. The automated experiment discovered specific topological defects in the materials for which parameters such as energy and information storage are optimized. Additionally, the “smart” workflow was orders of magnitude faster than conventional workflows in producing results that could lead to new discoveries in material functionality. This research was published in Nature Machine Intelligence.
“The takeaway is that the workflow was applied to material systems familiar to the research community and made a fundamental finding, something not previously known, very quickly – in this case, within a few hours,” said Ziatdinov.
While the method required is not an “off-the-shelf” capability and required substantial effort to connect the software and hardware to make the experiment possible, according to first author Yongtao Liu, the researchers believe the setup could be applied to more techniques other than scanning probe microscopy, such as experimental imaging and spectroscopy approaches, opening up more paths to discovery for functional materials.
Photo: A smart approach to microscopy and imaging developed at Oak Ridge National Laboratory could drive discoveries in materials for future technologies. Credit: Adam Malin/ORNL, U.S. Dept. of Energy