'Self-driving' microscopes discover shortcuts to new materials

Researchers at the Department of Energy's Oak Ridge National Laboratory are training microscopes to make discoveries using an intuitive algorithm created at the lab's Center for Nanophase Materials Sciences, which might lead to advances in novel materials for energy, sensing, and computing.

"There are so many potential materials, some of which we cannot study at all with conventional tools, that need more efficient and systematic approaches to design and synthesize," stated Maxim Ziatdinov of ORNL's Computational Sciences and Engineering Division and the CNMS. "We can use smart automation to access unexplored materials as well as create a shareable, reproducible path to discoveries that have not previously been possible."

The method, which was published in Nature Machine Intelligence, combines physics and machine learning to automate microscopy studies aimed at studying the functional features of materials at the nanoscale.

Functional materials are materials that respond to stimuli like heat or electricity and are designed to enable a wide range of technologies, from computers to solar cells to artificial muscles and shape-memory polymers. Their distinct characteristics are linked to atomic structures and microstructures visible using sophisticated microscopy. However, the challenge has been to develop efficient ways to locate regions of interest where these properties emerge and can be investigated.

Scanning probe microscopy is an important method for investigating structure-property interactions in functional materials. Instruments use an atomically sharp probe to map out the structure of materials at the nanoscale scale, or one billionth of a meter. They may also detect reactions to a variety of stimuli, revealing basic principles like as polarization switching, electrochemical reactivity, plastic deformation, and quantum events. Today's microscopes can scan a nanometer square grid point by point, but it's a time-consuming procedure that can take days for a single substance.

"The interesting physical phenomena are often only manifested in a small number of spatial locations and tied to specific but unknown structural elements. While we typically have an idea of what will be the characteristic features of physical phenomena we aim to discover, pinpointing these regions of interest efficiently is a major bottleneck," explained former ORNL CNMS scientist and lead author Sergei Kalinin, now at the University of Tennessee, Knoxville. "Our goal is to teach microscopes to seek regions with interesting physics actively and in a manner much more efficient than performing a grid search."

To overcome this challenge, scientists have turned to machine learning and artificial intelligence, but traditional algorithms require massive, human-coded datasets and may not save time in the end.

The ORNL methodology mixes human-based physical reasoning into machine learning approaches for a more intelligent approach to automation, and it starts with relatively tiny datasets (pictures from less than 1% of the sample). The algorithm chooses locations of interest depending on what it learns in the experiment and outside knowledge.

Scanning probe microscopy was used to illustrate a process on well-studied ferroelectric materials as a proof of concept. Ferroelectrics are reorientable surface charge functional materials that may be used in computation, actuation, and sensing. Scientists want to know how the quantity of energy or information these materials can store is related to the local domain structure that governs it.The automated experiment identified the topological faults for which these values are optimum.

"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," Ziatdinov explained.

The results were hundreds of times faster than traditional procedures, indicating a new trend in intelligent automation.

"We wanted to move away from training computers exclusively on data from previous experiments and instead teach computers how to think like researchers and learn on the fly," said Ziatdinov. "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."

The technical barrier of getting the algorithm to operate on an operating microscope at the CNMS was handled by Yongtao Liu of ORNL. "This is not an off-the-shelf capability, and a lot of work goes into connecting the hardware and software," Liu explained. "We focused on scanning probe microscopy, but the setup can be applied to other experimental imaging and spectroscopy approaches accessible to the broader user community."

The journal article is published as "Experimental discovery of structure-property relationships in ferroelectric materials via active learning."
Previous Post Next Post